Datasets:

llmunlearn

/

unlearn_dataset

like 1

Follow

Unlearning of LLMs 6

Introduction {#S1} ============ The protozoan parasite *Neospora caninum* causes considerable economic losses to cattle industry worldwide \[[@R1]\]. One of the definitive hosts of *N. caninum*, which is responsible for shedding of environmentally resistant oocysts, is the domestic dog \[[@R6]\]. In dogs, neosporosis can cause severe clinical manifestations, including myositis-polyradiculoneuritis, encephalomyelitis and dermatitis \[[@R7], [@R12]\]. Worldwide reports of clinical and subclinical infections in all hosts were summarized by Dubey et al. \[[@R1]\] and Dubey and Schares \[[@R2]\], but there was no mention of neosporosis in Serbia. Recently, occurrence of *N. caninum* specific antibodies in both cattle and dogs has been confirmed in Serbia as well \[[@R3], [@R9], [@R13], [@R17]\] and published in local journals. The aim of this study was to investigate current serological status of dairy cattle and dogs in Vojvodina (Northern Province of Serbia), with regard to the possible risk factors. Material and methods {#S2} ==================== Studied area {#S3} ------------ Vojvodina is a northern (45°15′ N 19°50′ E) province of the Republic of Serbia which occupies 21,506 km^2^ of the state territory \[[@R5]\]. The major part of the province's territory consists of fertile plains with the Danube, Tisa and Sava rivers dividing it in to three regions: Bačka, Banat and Srem. The climate of Vojvodina is moderately continental, characterised by hot, dry summers, cold winters and relatively low rainfall. Animals and sample collection {#S4} ----------------------------- Blood samples were collected from 356 dairy cows from both commercial farms (109 cows) and smallholdings (247 cows) in Srem, Banat and Bačka region ([Figure 1](#F1){ref-type="fig"}), during the 2009--2013 period. Samples from 271 cows were obtained on the farm by jugular venipuncture and 85 samples were collected at the abattoir. Among these were 74 samples from cows with a history of various reproductive disorders and 197 samples from reproductively healthy cows. Medical history could not be obtained for 85 abattoir samples. Prior to sampling, minimum recommended size of the sample was calculated, using Win Episcope 2.0 software \[[@R16]\]. Announcement of the Statistical Office of the Republic of Serbia on the number of cattle \[[@R14]\] and expected prevalence of 17.3% \[[@R13]\] that would ensure a 95% confidence interval and produce an error of 5% were used as input data for this calculation. Minimum recommended size of 220 animals was obtained. Therefore, a sample of 356 animals would not only provide an unbiased estimation of the prevalence of *N. caninum*-specific antibodies in the dairy cows of Vojvodina but would also reduce the error to 3.92%.Figure 1.Dots on the map of Vojvodina represent localities where blood samples were taken from cows or, in the case of blood collected at the abattoir, places of cows' origin. Black dots refer to localities where seropositive cows were found, while white dots stand for localities where there were no positive samples. Figures represent the number of sampled cows from specific locality and number of positive ones (between parentheses). Blood samples were obtained from 99 dogs during the 2008--2013 period, from various sources on several locations in Srem, Banat and Bačka ([Figure 2](#F2){ref-type="fig"}). None of the sampled dogs, including six farm dogs, belonged to the same farms as cows from this study. All canine blood samples were collected from cephalic vein. The dogs were clinically examined prior to sampling and for each dog a record sheet, which included information on breed, gender, age, previous health problem and diet, was filled. For stray dogs, only gender and age (estimated by competent veterinarian) could be obtained.Figure 2.Geographical distribution of localities where blood samples were collected from hunting, farm and stray dogs. Numbers outside parentheses represent the dogs from the specific category that was tested on the given locality. Numbers between parentheses refer to the positive ones. Sample preparation procedure was identical for both cows and dogs: after being left to clot, blood was centrifuged (3000 rpm for 10 min) and separated sera were stored at −20 °C until examination. Serological testing {#S5} ------------------- Two types of serological tests were used to detect *N. caninum* antibodies in cow sera. Initially 100 sera were assayed using the commercial competitive ELISA test kit (cELISA, VMRD Inc., Pullman, USA). The remainder 256 cow and all dog sera were examined with an indirect fluorescent antibody test (IFAT) using reagents marketed by VMRD. Both tests were performed according to the manufacturer's instructions. Cow sera, examined with ELISA, that presented inhibition percentages equal to or higher than 30% were considered positive. For IFAT, a recommended cut-off of 1:200 for cows and 1:50 for dogs was used. Dog sera that exhibited positive reaction at 1:50 were serially diluted until negativity was reached. Statistical analysis {#S6} -------------------- Seroprevalences and their confidence intervals, for both cows and dogs, were calculated using Quantitative Parasitology 3.0. \[[@R11]\]. For the statistical analysis of the possible effects of different factors (origin, farm type and history of reproductive disorders in cows and utilisation, breed, gender, age, origin and feeding habits in dogs) on the occurrence of anti-*N. caninum* antibodies the chi-squared test was used at a significance level of 95% (*p* ≤ 0.05). Where appropriate chi-squared test was replaced by Fisher's exact test or unconditional exact test. All these tests were computed using the above mentioned statistical software. Odds ratio (OR) was calculated using Win Episcope version 2.0 software \[[@R16]\]. Positive OR (OR \> 1), with 95% confidence interval (CI) that does not overlap with the null value of 1 and with *p* ≤ 0.05, was considered statistically significant \[[@R15]\]. Confidence interval and *p*-value for OR were calculated using online MedCalc software version 12.6 \[[@R8]\]. Results {#S7} ======= Antibodies to *N. caninum* were found in 15.4% (55/356, CI 95%:12.0--19.6) of cow sera. Seven positive sera were detected by ELISA and the remaining 48 by IFAT ([Table 1](#T1){ref-type="table"}). Among the risk factors evaluated, only cows originating from smallholdings had significantly greater odds (OR = 5.28, CI 95%: 2.0--13.6, *p* = 0.0006) of being seropositive in comparison to the group of farm cows.Table 1.Prevalence of seropositive cows according to the geographical origin, farm type and health status.ExaminedPositiveELISAIFATTotalELISAIFATTotalRate (%)CI 95 (%)*p*-Value*Geographical origin*Srem10566609913.67.1--24.10.766Banat31781093161917.411.4--25.6Bačka591221814232714.910.3--20.9*Farm type*Commercial farms33761091454.61.5--10.40.0001[a](#TFN1){ref-type="fn"}Smallholdings671802476445020.215.4--25.8 *Reproductive disorders*Yes22527409912.26.3--21.50.148No01971970373718.813.9--24.8nd7878572910.65.5--19.2Total1002563567485515.412.0--19.6nd = No data regarding medical history could be obtained[^1] *N. caninum* antibodies were found in 17.2% (17/99, CI 95%: 10.8--26.2) of dogs, with titres of 50 in 15 dogs, 100 in two and 200 in one dog. Out of 17 seropositive animals, 14 (14/71, 19.7%, CI 95%: 11.2--30.9) came from the group of hunting dogs, one (1/22, 4.5%, CI 95%: 0.2--22.2) was a stray dog and two (2/6, 33.3%, CI 95%: 6.3--72.9) belonged to a small group of farm dogs. Of all the risk factors evaluated statistical difference (*p* = 0.02) was observed in the occurrence of *N. caninum*-specific antibodies in dogs from Bačka versus dogs from Banat ([Table 2](#T2){ref-type="table"}).Table 2.Prevalence of anti-*Neospora caninum* antibodies in hunting, stray and farm dogs according to the breed, gender, age and region of origin.HuntingStrayFarmTotal*n*Positive%*n*Positive%*n*Positive%*n*Positive%*Breed*Purebreed661421.20002150681522.1Mongrel5002214.541253126.5*Gender*Female32515.61317.7215047714.9Male39923.19004125521019.2*Age (years)*≤11317.76002002114.8\>1, ≤5451124.41317.74250621422.6\>51321.430000016212.5*Origin*Srem1022000000010220Banat1915.32214.521504337Bačka421126.20004125461226.1[a](#TFN2){ref-type="fn"}Total711419.72214.56233.3991717.2*n* = Number of examined dogs; % = prevalence[^2] In 42 hunting dogs with known feeding habits 11 (26.2%, CI 95%: 14.9--41.6) were seropositive. No statistical differences (*p* = 0.4) were found between the dogs fed commercial food (2/12, 16.7%, CI 95%: 2.1--48.4), homemade food (8/22, 36.4%, CI 95%: 17.2--59.3) and the mixture of two diets (1/8, 12.5%, CI 95%: 0.6--50.0). Seropositive purebred dogs were present as follows: Dachshund (*n* = 1), English Springer Spaniel (*n* = 1), English Pointer (*n* = 1), German Shorthaired Pointer (*n* = 4), German Hunting Terrier (*n* = 1), Weimaraner (*n* = 1), Setter (*n* = 1), German Wirehaired Pointer (*n* = 2), Labrador Retriever (*n* = 1), Small Munsterlander Pointer (*n* = 1) and German Shepherd (*n* = 1). Most (92 of 99) dogs were clinically normal and seven dogs had clinical signs which could not be related to neosporosis. Discussion {#S8} ========== The results of this study corroborate previous findings \[[@R3], [@R9], [@R13], [@R17]\] of *N. caninum* antibodies in both cattle and dogs from the territory of Vojvodina. The prevalence of antibodies in dairy cattle from our study (15.4%) was higher than the one found by Gavrilović et al. \[[@R3]\]. Similar to our study, their sample consisted of both aborting and randomly sampled cows, from both commercial and smallholding farms. Nevertheless, they found only 4.6% (23/500) of cows to be seropositive, which could be partly due to the small proportion of aborting cows in the total sample and the fact that they restricted their research to the south district of Banat. Different commercial tests were utilized in these studies (Gavrilović et al. used ELISA manufactured by IDEXX and Svanova, while we used IFAT and ELISA manufactured by VMRD) which may have also influenced the results. Vidić et al. \[[@R17]\] and Savović et al. \[[@R13]\] reported occurrence of *N. caninum* antibodies in 3.7% (5/132) and 17.3% (9/52) of dairy cows, respectively. Although small size of the study samples and the fact they consisted entirely of cows with reproductive disorders make these prevalences difficult to compare with, we referred to them in the interest of data transparency. With regard to the risk factors, significant difference was observed between seroprevalences in cattle from commercial and from small farms in our study. Probable reason for these observations is the fact that biosecurity measures on commercial farms are generally superior to those on small family farms which are more likely to have dogs, the potential source of infection. Furthermore, cattle from smallholdings are usually associated with extensive production and often allowed to graze, which is, according to some authors \[[@R4], [@R10]\], one of the risk factors associated with seropositivity in cattle. Gavrilović et al. \[[@R3]\] also observed that all cows from seropositive commercial farms in their study had access to the outdoor paddocks which resulted in higher herd prevalence in commercial farms (50.0%), in comparison to smallholdings (22.64%). On individual level, however, they found no significant difference between commercial (4.67%) and small (4.50%) farms. Prevalence of *N. caninum* antibodies in dogs varies based on many factors, including age, origin and diet \[[@R1]\]. Pavičić et al. \[[@R9]\] were first to detect *N. caninum* antibodies in four out of 31 (12.9%) dogs from Vojvodina. Examined dogs were mostly hounds, originating from Srem and Banat region. In the present study, we found antibodies in 17 out of 99 dogs (17.2%). Samples were collected from all three regions of Vojvodina and higher prevalence was observed in dogs from Bačka (26.1%) in comparison to dogs from Banat (7%). Other factors, such as breed, age, diet or utilisation did not appear significant. However, since the sample size was relatively small and samples originated from various sources and were collected over several years, no definitive conclusion can be made. Nevertheless, we recorded findings for the benefit of future studies. Presently, there are no reports of confirmed clinical neosporosis in any host in Serbia, including cattle and dogs. In conclusion, our study confirmed the presence of *N. caninum* antibodies in general population of dairy cows in Vojvodina, with a higher prevalence detected in animals originating from smallholding farms. Specific antibodies against *N. caninum* were also found in dogs from the same province, with slightly higher prevalence observed in dogs from Bačka. To the Ministry of Education, Science and Technological Development of the Republic of Serbia which supported this study with Grants No. TR31034 (PhD grant for Ljiljana Kuruca) and No. TR31084. **Cite this article as**: Kuruca L, Spasojević-Kosić L, Simin S, Savović M, Lauš & Lalošević V: *Neospora caninum* antibodies in dairy cows and domestic dogs from Vojvodina, Serbia. Parasite, 2013, **20**, 40. [^1]: Difference between these prevalences was statistically significant (*p* ≤ 0.05). [^2]: Difference between these prevalences was statistically significant (*p* ≤ 0.05).

\section{Introduction} Observations of exoplanets will be limited to disc-integrated measurements for the foreseeable future. This is true whether a planet can be spatially resolved from its host star (direct imaging) or not (as in current studies of short-period transiting planets). Spectra with long integration times yield invaluable information about the spatially and temporally averaged composition and temperature-pressure profile of the atmosphere. Time-resolved photometry, on the other hand, tells us about the weather, climate, and spatial inhomogeneities of the planet. The time-variability of a planet occurs on two timescales, rotational and orbital\footnote{These are one and the same for synchronously rotating planets.}, and yields different information depending on whether it is observed in reflected or thermal light. \emph{Thermal phases} inform us about the diurnal heating patterns of the planet: the day-side temperature, the night-side temperature, and the hottest local time on the planet \citep{Cowan_2008}. Depending on whether it has an atmosphere, such observations can constrain a body's rotation rate, as well as its average Bond albedo, thermal inertia, emissivity, surface roughness and wind velocities \citep{Spencer_1990, Cowan_2011}. \emph{Rotational variations in thermal emission} are caused by inhomogeneities in the planet's albedo and thermal inertia. This has been studied for minor solar system bodies, where it can be used to break the degeneracy between albedo markings and shape \citep[eg,][]{Lellouch_2000}. \emph{Reflected phases} are a measure of the disk-integrated scattering phase function, telling us ---for example--- about clouds and oceans on the planet, especially when combined with polarimetry \citep{Williams_2008, Mallama_2009, Robinson_2010, Zugger_2010}. \emph{Rotational variations at reflected wavelengths} can identify the rotation rate of an unresolved planet \citep{Palle_2008}. Once the rotation rate has been determined, one can constrain the albedo markings on a world \citep{Russell_1906}, indicating surface features like continents and oceans \citep{Ford_2001, Cowan_2009, Oakley_2009, Fujii_2010}. Finally, the spatial distribution of landmasses can be inferred, and the planet's obliquity can be estimated if diurnal variations are monitored at a variety of phases \citep{Cowan_2009, Oakley_2009, Kawahara_2010}. In this paper, we study rotational (a.k.a.\ diurnal) variability of Earth's poles at visible wavelengths. At these wavelengths, the observed flux consists entirely of reflected sunlight. Earthshine ---the faint illumination of the dark side of the Moon due to reflected light from Earth--- was first explaied in the early 16$^{\rm th}$ century by Leonardo Da Vinci, and has been used more recently to study the reflectance spectrum and cloud cover variability \citep{Goode_2001, Qiu_2003, Palle_2003, Palle_2004, Montanes_2007}, vegetation ``red edge'' signature \citep{Woolf_2002, Montanes_2005, Seager_2005, Hamdani_2006, Montanes_2006} and the effects of specular reflection \citep{Woolf_2002, Langford_2009} for limited regions of our planet. More recently, \cite{Palle_2009} measured the disc-integrated \emph{transmission} spectrum of Earth by observing the Moon during a lunar eclipse. Brief snapshots of Earth obtained with the Galileo spacecraft have been used to study our planet from afar \citep{Sagan_1993, Geissler_1995} and numerical models have been developed to anticipate how diurnal variations in disc-integrated light could be used to characterize Earth \citep[][Robinson et al.\ 2011]{Ford_2001, Tinetti_2006, Tinetti_2006b, Palle_2008, Williams_2008, Oakley_2009, Fujii_2010, Kawahara_2010, Zugger_2010, Robinson_2010}. In \cite{Cowan_2009} we presented two epochs of rotational variability for disk-integrated Earth seen equator-on. We performed a principal component analysis (PCA) of the 7 optical wavebands to identify the eigencolors of the variability. There were two components that accounted for more than 98\% of the color variations seen. The two-dimensionality of the PCA indicated that three major surface types were necessary to explain the observed variability. The dominant eigencolor was red, which we identified as being primarily sensitive to cloud-free land. A rotational inversion of the red eigenprojection yielded a rough map of the major landforms of Earth: the Americas, Africa-Eurasia and Oceania, separated by the major oceans: the Atlantic and Pacific. One concern with the diurnal light curve inversion of \cite{Cowan_2009} and \cite{Oakley_2009} is the unknown obliquity of the planet: there is no good Bayesian prior for the obliquities and rotation rates, except that they will be slightly biased towards prograde rotation \citep{Schlichting_2007}. Numerical experiments have shown that a pole-on viewing geometry might complicate retrieval of a planet's rotational period \citep{Palle_2008}, but once that periodicity is identified, it is possible to create albedo maps of the planet, although without knowledge of the planet's obliquity one will not know what latitudes those maps correspond to \citep{Cowan_2009}. Idealized numerical experiments show that ---in principle--- the obliquity can be extracted if one observes diurnal variability at a variety of phases \citep{Kawahara_2010}, but it is not yet clear how well such a technique would work for a cloudy planet with unknown surface types. Finaly, Earth's climate is sensitive to the latitudinal configuration of continents and few percent changes in insolation (neither of which will be well constrained for exoplanets) leading to bifurcations between temperate and snowball climates \citep{Voigt_2010}. Given an extrasolar Earth analog, how can we use optical photometry to distinguish between the two branches of this positive feedback loop? In this paper we report and analyze disk-integrated observations of Earth's polar regions obtained from the Deep Impact spacecraft as part of the EPOXI mission of opportunity. In \S~2 we present the observations; in \S~3 we discuss the color variability; in \S~4 we make longitudinal profiles of these colors; in \S~5 we introduce a simple model for snowball Earth and compare its diurnal color variations to those of Earth's polar regions; we summarize our results in \S~6 and state the implications of this study for mission planning in \S~7. \section{Observations} The EPOXI\footnote{The University of Maryland leads the overall EPOXI mission, including the flyby of comet Hartley 2. NASA Goddard leads the exoplanet and Earth observations.} mission reuses the still-functioning Deep Impact spacecraft that successfully observed comet 9P/Tempel~1. EPOXI science targets include several transiting exoplanets and Earth \emph{en route} to a flyby of comet 103P/Hartley~2. The first round of EPOXI Earth observations were taken from a vantage point very near the Earth's equatorial plane \citep[][Robinson et al. 2010a; Livengood et al. 2011]{Cowan_2009}. In the current paper, we focus on the later polar observations, summarized in Table~\ref{observations}. As with the equatorial observations, they were obtained with Earth near quadrature (phase angle $\alpha=\pi/2$), a favorable phase for directly imaging exoplanets, since the angular distance between the planet and its host star is maximized\footnote{Strictly speaking, this is only true for a planet on a circular orbit.}. \begin{table*}[htb] \begin{minipage}{126mm} \caption{\bf EPOXI Earth Observations} \label{observations} \begin{tabular}{@{}llcccccc} \hline Name &Start of Observations& Starting & Sub-Observer& Sub-Solar& Dominant&Phase$^{3}$ & Illuminated Fraction\\ & & CML$^{1}$ & Latitude & Latitude&Latitude$^{2}$ && of Earth Disc$^{3}$\\ \hline {\bf Earth1: Equinox} &2008 Mar 18, UTC 18:18:37 & $150^{\circ}$ W & $1.7^{\circ}$~N&$0.6^{\circ}$~S & 0.5$^{\circ}$~N&57$^{\circ}$ & 77\%\\ {\bf Earth5: Solstice} &2008 Jun 4, UTC 16:59:08 & $150^{\circ}$ W & $0.3^{\circ}$~N&$22.7^{\circ}$~N & 13$^{\circ}$~N&77$^{\circ}$ & 62\%\\ {\bf Polar1: North} &2009 Mar 27, UTC 16:19:42& $152^{\circ}$ W & $61.7^{\circ}$~N&$2.6^{\circ}$~N & $34^{\circ}$~N&87$^{\circ}$ & 53\%\\ {\bf Polar2: South} &2009 Oct 4, UTC 09:37:11 & $59^{\circ}$ W & $73.8^{\circ}$~S& $4.3^{\circ}$~S&$39^{\circ}$~S & 86$^{\circ}$ & 53\%\\ \hline \end{tabular} \medskip $^{1}$The CML is the Central Meridian Longitude, the longitude of the sub-observer point.\\ $^{2}$The dominant latitude is that expected to contribute the most photons, assuming a uniform Lambert sphere.\\ $^{3}$The planetary phase, $\alpha$, is the star--planet--observer angle and is related to the illuminated fraction by $f = \frac{1}{2}(1+\cos\alpha)$. \end{minipage} \end{table*} Deep Impact's 30~cm diameter telescope coupled with the High Resolution Imager \citep[HRI,][]{Hampton_2005} recorded images of Earth in seven 100~nm wide optical wavebands spanning 300--1000~nm, summarized in Table~\ref{bandpasses}. \begin{table}[htb] \caption{\bf EPOXI Photometric Bandpasses} \label{bandpasses} \begin{center} \begin{tabular}{ccc} \hline Waveband & Cadence & Exp. Time\\ \hline 350~nm & 1 hr & 73.4~msec\\ 450~nm & 15 min & 13.3~msec\\ 550~nm & 15 min & 8.5~msec\\ 650~nm & 15 min & 9.5~msec\\ 750~nm & 1 hr & 13.5~msec\\ 850~nm & 15 min & 26.5~msec\\ 950~nm & 1 hr & 61.5~msec\\ \hline \end{tabular} \end{center} \end{table} Although the EPOXI images of Earth offer spatial resolution of better than 100~km, we mimic the data that will eventually be available for exoplanets by integrating the flux over the entire disc of Earth and using only the hourly EPOXI observations from each of the wavebands, producing seven light curves for each of the two observing campaigns, shown in Figures~\ref{polar1_lightcurves} and \ref{polar2_lightcurves}. The photometric uncertainty in these data is exceedingly small: on the order of 0.1\% relative errors. Details of the observations and reduction will be presented in Livengood et al. (2011). After performing aperture photometry, the measured disk-integrated flux of Earth as seen from the Deep Impact spacecraft has units of specific flux [W m$^{-2}$ $\mu$m$^{-1}$]. We then apply the following steps: 1) We multiply by the HRI filter bandwidth of 0.1~$\mu$m to convert from specific flux to flux [W m$^{-2}$]. The detailed bandpass shapes are not important, provided that we use the same 0.1~$\mu$m top-hat bandpasses in computing the solar flux in step 3). 2) We divide the flux observed from the spacecraft by $(R_{\oplus}/r)^{2}$ to obtain the disc-averaged flux from Earth at the top of the atmosphere, where $R_{\oplus}$ is Earth's radius and $r$ is the spacecraft range. 3) We use a Kurucz model\footnote{Ideally, one would obtain a spectrum of the planet's host star with the same instrument used for imaging the planet, but this was not possible here for technical reasons.} for the solar specific flux at 1~AU, and convert it to flux in each of the HRI wavebands using the 0.1~$\mu$m top-hat bandpasses. 4) Dividing the result of steps 2) and 3) by each other, we obtain the top-of-the-atmosphere reflectance of the planet at the observed phase. 5) We further divide the reflectance by the scaled Lambert phase function \citep[$f(\alpha)=\frac{2}{3\pi}\left(\sin\alpha + (\pi-\alpha)\cos\alpha\right)$;][]{Russell_1916}, thus obtaining the planet's apparent albedo, $A^{*}$. The precise definition of $A^{*}$ is given in Equation~5; for now it is sufficient to think of it as the average albedo of the planet, weighted by the illumination and visibility of various regions at that moment in time. Unlike the observations presented in \cite{Cowan_2009}, the viewing geometry is sufficiently different for the two polar observations that we treat them separately. In particular, a single exoplanet could be observed by the same stationary observer with the two viewing geometries of \cite{Cowan_2009} since the sub-observer latitude was essentially equatorial at both epochs (see Table~1). Although the sub-stellar latitude varies with phase for non-zero obliquities, the sub-observer latitude is constant, provided that one can neglect precession (for more details on viewing geometry, see \S~\ref{viewing_geometry}). By contrast, the two time series presented in the current paper have sub-observer latitudes of 62$^{\circ}$~N and 74$^{\circ}$~S, respectively. \begin{figure}[htb] \includegraphics[width=84mm]{Polar1_unfitted_lightcurves.eps} \caption{{\bf North:} Light curves obtained in the seven HRI-VIS bandpasses by the EPOXI spacecraft when it passed above Earth's equatorial plane on March 27, 2009. The bottom-right panel shows changes in the bolometric albedo of Earth. The sub-observer longitude at the start of the observations is $152^{\circ}$W, the sub-observer latitude is $61.7^{\circ}$N throughout. The relative peak-to-trough variability ranges from 16\% (450~nm) to 10\% (750~nm).} \label{polar1_lightcurves} \end{figure} \begin{figure}[htb] \includegraphics[width=84mm]{Polar2_unfitted_lightcurves.eps} \caption{{\bf South:} Light curves obtained in the seven HRI-VIS bandpasses by the EPOXI spacecraft when it passed below Earth's equatorial plane on October 4, 2009. The bottom-right panel shows changes in the bolometric albedo of Earth. The sub-observer longitude at the start of the observations is $59^{\circ}$W, the sub-observer latitude is $73.8^{\circ}$S throughout. The relative peak-to-trough variability ranges from 11\% (350~nm) to 20\% (650~nm).} \label{polar2_lightcurves} \end{figure} \begin{figure}[htb] \includegraphics[width=84mm]{average_albedo_spectrum.eps} \caption{24-hour average broadband spectra of Earth as seen from the Deep Impact spacecraft as part of the EPOXI mission.} \label{average_albedo_spectrum} \end{figure} \begin{table*}[htb] \begin{minipage}{126mm} \caption{\bf EPOXI Earth Observations: Apparent Albedos} \label{apparent_albedos} \begin{tabular}{@{}lccccccc} \hline Name & 350~nm & 450~nm & 550~nm & 650~nm & 750~nm & 850~nm & 950~nm\\ \hline {\bf Earth1: Equinox} & 0.482 (0.016) & 0.352 (0.014) & 0.268 (0.010) & 0.256 (0.010) & 0.264 (0.014) & 0.290 (0.017) & 0.230 (0.015)\\ {\bf Earth5: Solstice} & 0.462 (0.021) & 0.346 (0.017) & 0.268 (0.011) & 0.257 (0.009) & 0.270 (0.014) & 0.298 (0.019) & 0.230 (0.017)\\ {\bf Polar1: North} & 0.575 (0.028) & 0.464 (0.023) & 0.373 (0.015) & 0.368 (0.012) & 0.388 (0.012) & 0.425 (0.013) & 0.355 (0.012)\\ {\bf Polar2: South} & 0.590 (0.022) & 0.475 (0.022) & 0.383 (0.023) & 0.372 (0.023) & 0.382 (0.022) & 0.414 (0.023) & 0.351 (0.017)\\ \hline \end{tabular} \medskip The number in parentheses represents the root-mean-squared (RMS) time-variability of the apparent albedo in that waveband. For example, the Earth1 350~nm time series had a mean apparent albedo of 0.482 and an RMS variability of 0.016, or 3\%. \end{minipage} \end{table*} The time-averaged spectra for the EPOXI polar and equatorial observations are shown in Figure~\ref{average_albedo_spectrum}, and the apparent albedo values are listed in Table~\ref{apparent_albedos}. The literature abounds with vaguely-defined ``reflectance'' measurements: as a result of differing definitions of reflectance used by different EPOXI team members, the albedos reported in \cite{Cowan_2009} were about $2/3$ of the correct value. (Note that this uniform offset had no impact on the color variations and analysis presented in that paper.) We present the corrected values of apparent albedo in Table~3 and Figure~3. The major features of the time-averaged broadband albedo spectrum of Earth are: 1) a blue ramp shortward of 550~nm due to Rayleigh scattering, 2) a slight rise in albedo towards longer wavelengths due to continents, and 3) a steep dip in albedo at 950~nm due to water vapor absorption. Significantly, apart from a 20-30\% uniform offset, the polar and equatorial observations have indistinguishable albedo spectra. \section{Determining Principal Colors} As in \cite{Cowan_2009} we assume no prior knowledge of the different surface types of the unresolved planet. Our data consist of 25 broadband spectra of Earth for each of two viewing geometries. For the equatorial observations \citep{Cowan_2009}, we found substantial variability in all wavebands (though the near-IR wavebands exhibited the most variability, leading to the dominant red eigencolor). The polar observations also show variability at all wavebands (Table~2), but as we argue below, the intrinsic cause of this variability is not necessarily the same surface types rotating in and out of view. The multiband, time-resolved observations of Earth can be thought of as a locus of points occupying a 7-dimensional parameter space (one for each waveband). Principal component analysis (PCA) allows us to reduce the dimensionality of these data by defining orthonormal eigenvectors in color space (a.k.a.\ eigencolors). Quantitatively, the observed spectrum of Earth at some time $t$ can be recovered using the equation: \begin{equation} \label{pca} A^{*}(t, \lambda) = \langle A^{*}(t, \lambda)\rangle + \sum_{i=1}^{7} C_{i}(t) A_{i}(\lambda), \end{equation} where $\langle A^{*}(t, \lambda)\rangle$ is the time-averaged spectrum of Earth, $A_{i}(\lambda)$ are the seven orthonormal eigencolors, and $C_{i}(t)$ are the instantaneous projections of Earth's colors on the eigencolors. The terms in the sum are ranked by the time-variance in $C_{i}$, from largest to smallest. Insofar as the color variations are dominated by the first few terms of the sum, the locus does not occupy the full 7-dimensional color space, but a more restricted manifold. The dimensionality of the manifold is one fewer than the number of surface types rotating in and out of view. E.g., a two-dimensional locus (a planar manifold) requires three surface types; a three-dimensional locus requires four surface types, etc. The general problem of estimating the pure surface spectra based on the morphology of such a locus of points is beyond the scope of this paper. It is a form of spectral unmixing, and is an area of active research in the remote sensing community \citep[e.g.,][]{LeMouelic_2009}. In practice, there are two different ways to perform PCA, which may give quantitatively different results. The analysis can be run using the covariance of the data, ${\rm Cov}(X,Y)= E[(X-E[X])(Y-E[Y])]$, where $E[X]$ is the expected value of $X$; or it can be run using the correlation of the data, ${\rm Corr}(X,Y)= {\rm Cov}(X,Y)/(\sigma_{X}\sigma_{Y})$, where $\sigma_{X}$ is the standard deviation of $X$. The correlation matrix is a standardized version of the covariance matrix; this is useful when the measured data do not all have the same units, since division by the standard deviation renders them unitless. When the data are unitless to begin with, as is the case for our albedo measurements, running covariance-PCA is preferable \citep[e.g.,][]{Borgognone_2001}. In \cite{Cowan_2009} we used covariance-PCA, and we continue to do so here\footnote{For this paper, we use the Interactive Data Language (IDL) routine PCOMP (with the /COVARIANCE keyword set) to perform principal component analysis, while in \cite{Cowan_2009} we used the IDL routine SVDC, which performs singular value decomposition.}. \subsection{Testing PCA on Simulated Data} \label{numerical_experiment} Although PCA is a mathematically and numerically robust technique for analyzing patterns in data, interpreting its results can be ambiguous. In particular, we would like to verify to what extent there is a one-to-one correspondence between the eigencolors output by the PCA and real surfaces on Earth. To this end, we test the PCA routine on a suite of simulated data produced by the Virtual Planetary Laboratory's validated 3D spectral Earth model (details can be found in Robinson et al. 2011). The simulations used here were designed to closely mimic the Earth1 EPOXI observations taken in March 2008. We run five different versions of the VPL 3D Earth model: 1) Standard: this model is an excellent fit to the EPOXI Earth1 observations; the remaining models are identical, but in each case a single model element has been ``turned off'': 2) Cloud Free; 3) No Rayleigh Scattering; 4) Black Oceans; 5) Black Land. We show the results of this experiment in Appendix I; here we simply state our conclusions: 1) PCA successfully determines the dimensionality of the color variability and therefore the minimum number of different surface types contributing to color variations. In particular, $n$-dimensional variations require $n+1$ surface types (N.B. we count clouds as a surface type). 2) Rayleigh scattering is important in determining the time-averaged broadband colors of Earth, but does not significantly affect its rotational color variability. 3) Cloud-free land surfaces, which are red, contribute a red eigencolor to the diurnal variability. The presence of relatively cloud-free land (deserts) near the equator explains why the rotational map of the red eigencolor \citep[Figure~10 in][]{Cowan_2009} successfully identified the major landforms and bodies of water on Earth. 4) Oceans are essentially a null surface, contributing neither to the broadband colors of Earth, nor to the time-variability of those colors, except insofar as the presence of oceans corresponds to a shortage of land. 5) In the absence of land, the variability is gray, due to large-scale inhomogeneities in cloud cover. 6) PCA necessarily outputs orthogonal eigencolors and a good deal of Earth's variability is due to clouds. Therefore, if the first eigencolor is red, then the second eigencolor may be blue even if there is no blue surface rotating in and out of view; this is an improvement on the interpretation of \cite{Cowan_2009}. \subsection{Results of PCA for Polar Observations} In Figures~\ref{polar1_variability} \& \ref{polar2_variability} we show the eigenvalue spectra for time-variations in the 7 eigencolors identified by the PCA of the EPOXI polar observations. The eigenvalue for a given component is the projection of the data's variance onto that eigenvector; we plot here the square root of the eigenvalues, which is a measure of the RMS variability of the data projected onto an eigenvector. The variability has been normalized in the figures such that the sum of the variability for all seven components is unity. By definition, the low-order principal components have the largest variance. For the North observation, there are two eigencolors that dominate the color variations of Earth: the third eigencolor contributes only $\sim4$\% of the planet's color variability. As in \cite{Cowan_2009}, this means that the colors of Earth populate a two-dimensional plane rather than filling the entire seven-dimensional color-space, and this requires at least three surface types. The southern observation, on the other hand, is dominated by a single eigencolor (the second eigencolor contributes to variability at the $<10$\% level). This means that ---for the 24 hours of observations--- the colors of the planet populated a one-dimensional line in the seven-dimensional color volume, requiring only two surface types. \begin{figure}[htb] \includegraphics[width=84mm]{Polar1_variability.eps} \caption{{\bf North} Normalized variability in the 7 eigencolors of Earth's North polar regions, based on EPOXI observations taken on March 27, 2009. The color variations of Earth during these observations are well described as a combination of components 1 \& 2.} \label{polar1_variability} \end{figure} \begin{figure}[htb] \includegraphics[width=84mm]{Polar2_variability.eps} \caption{{\bf South} Normalized variability in the 7 eigencolors of Earth's South polar regions, based on EPOXI observations taken on October 4, 2009. The color variations of Earth during these observations are well described by component 1.} \label{polar2_variability} \end{figure} The eigencolors (the $A_{i}(\lambda)$ from Equation~\ref{pca}) are shown in Figures~\ref{polar1_eigenspectra} \& \ref{polar2_eigenspectra}. The raw eigencolors are ---by definition--- orthogonal and normalized ($\sum_{j=1}^{7}A_{i}^{2}(\lambda_{j})=1$), and this is how we presented them in \cite{Cowan_2009}. Here we have instead scaled the eigencolors by their associated eigenvalues ($\sum_{j=1}^{7}A_{i}^{2}(\lambda_{j})=v_{i}$, where $v_{i}$ is the eigenvalue of the $i$'th component), so the dominant components exhibit larger excursions from zero. The eigenprojections ($C_{i}(t)$ from Equation~\ref{pca}) are shown in Figures~\ref{polar1_eigenprojections} \& \ref{polar2_eigenprojections}. The standard deviation of an eigenprojection corresponds to the variance or eigenvalue of that component. By definition, the low-order eigenprojections have the largest deviations from 0. Note that in \cite{Cowan_2009} we instead plotted the normalized eigenprojections ($\sum_{k=1}^{25}C_{i}(t_{k})^{2}=1$), which made it easier to compare the shapes of the eigenprojections but masked their relative importance. The North polar observations are dominated by two eigencolors. At first glance, the two eigencolors are identical, only offset in the vertical direction, but they are (by construction) orthogonal. The more important of the two is blue, in that it is most non-zero at short wavelengths and nearly independent of what is going on at long wavelengths; the second eigencolor is red: it is most non-zero at long wavelengths and is largely insensitive to variability in blue wavebands. Based on the findings presented in \S~\ref{numerical_experiment}, we may infer that clouds and continents are rotating in and out of view as seen from this vantage point. Furthermore, cloud-related variability appears to be more important here than it was for the equatorial observations, which had a dominant red eigenvector followed by a blue, rather than vice-versa. The South polar observations are dominated by a single, gray eigencolor. Snow and clouds both have gray optical albedo spectra, so either may be contributing to the photometric variability. The absence of an important red eigencolor is due to the relative dearth of continents in the southern hemisphere. The second eigencolor is two orders of magnitude down in variance, or one order of magnitude in variability. It indicates that red and blue surfaces are trading places as the world turns ($A_{2}(\lambda)$ is positive at short wavelengths, negative at long wavelengths, and zero in between), but the forced orthogonality of the eigencolors makes this interpretation ambiguous. \begin{figure}[htb] \includegraphics[width=84mm]{Polar1_on_the_hour_eigenspectra.eps} \caption{{\bf North} Spectra for the eigencolors of northern Earth, as determined by PCA. The two dominant eigencolors are the bold solid and dotted lines. The eigenspectra have been normalized by their eigenvalues, so the dominant components exhibit larger excursions from zero.} \label{polar1_eigenspectra} \end{figure} \begin{figure}[htb] \includegraphics[width=84mm]{Polar2_on_the_hour_eigenspectra.eps} \caption{{\bf South} Spectra for the eigencolors of southern Earth, as determined by PCA. The two dominant eigencolors are the bold solid and dotted lines. The eigenspectra have been normalized by their eigenvalues, so the dominant components exhibit larger excursions from zero.} \label{polar2_eigenspectra} \end{figure} \begin{figure}[htb] \includegraphics[width=84mm]{Polar1_on_the_hour_eigenprojections.eps} \caption{{\bf North} Contributions of northern Earth's eigencolors, as determined by PCA, relative to the average Earth spectrum. The observations span a full rotation of the planet, starting and ending with the spacecraft directly above $152^{\circ}$~W longitude, the North Pacific Ocean.} \label{polar1_eigenprojections} \end{figure} \begin{figure}[htb] \includegraphics[width=84mm]{Polar2_on_the_hour_eigenprojections.eps} \caption{{\bf South} Contributions of southern Earth's eigencolors, as determined by PCA, relative to the average Earth spectrum. The observations span a full rotation of the planet, starting and ending with the spacecraft directly above $59^{\circ}$~W longitude, in the South Atlantic Ocean.} \label{polar2_eigenprojections} \end{figure} \section{Rotational Mapping} \label{viewing_geometry} In this section we address how to infer the longitudinal color inhomogeneities of the unresolved planet based on time-resolved photometry. Note that this is in principle an independent question from that of identifying surface types on the planet (\S~3). One could try to infer the surface types on a planet without knowing or caring about their spatial distribution; or one could simply produce longitudinal color maps while remaining agnostic about what these tell us about surfaces (where ``surface'' here includes clouds). In practice, however, the two are intimately tied: a planet only exhibits rotational variability if it has a variegated surface \emph{and} substantial spatial inhomogeneities in the distribution of these surfaces. \subsection{Cloud Variability} As in \cite{Cowan_2009}, we wish to estimate disk-integrated cloud variability, as this imposes a limit on the accuracy of any rotational maps we create\footnote{Note that diurnal cloud variability is not necessarily an obstacle for PCA, since that analysis is not predicated on a periodic signal.}. After 24 hours of rotation the same hemisphere of Earth should be facing the Deep Impact spacecraft, so the integrated brightness of the planet's surface should be nearly identical, provided one has accounted for the difference between the sidereal and solar day, as well as changes in the geocentric distance of the spacecraft and in the phase of the planet as seen from the spacecraft. Even after correcting for all known geometric effects, the observed fluxes at the start and end of our observing campaigns differ by $\Delta A^{*}/\langle A^{*}\rangle =$3--6\% and 0.4--1\% for the North and South polar observing campaigns, respectively. We attribute this discrepancy to diurnal changes in cloud cover. Our 24-hour polar observation cloud variability of 4\% and 1\%, respectively is comparable to our estimate of cloud variability from previous EPOXI Earth observations \citep[2\% and 3\%,][]{Cowan_2009} and somewhat smaller than estimates from Earthshine observations. For example, \cite{Goode_2001} and \cite{Palle_2004} found day-to-day cloud variations of roughly 5\% and 10\%, respectively. Although we are still very much in the realm of small number statistics, it is conceivable that Earthshine observations over-estimate diurnal changes in cloud cover: under-estimating night-to-night calibration errors would lead to over-estimating day-to-day cloud variability. Depending on the size of the telescope and cloud meteorology for a given planet, either photon counting or cloud variability can dominate the error budget for rotational inversion. For the purposes of rotational mapping, we adopt effective Gaussian errors in the apparent albedo of $\sigma_{A^{*}}=0.01$, comparable to the actual day-to-day cloud variability on Earth. \subsection{Normalized Weight and the Dominant Latitude} Using the formalism of \cite{Cowan_2009}, the visibility and illumination of a region on the planet at time $t$ are denoted by $V(\theta, \phi, t)$ and $I(\theta, \phi, t)$, respectively, where $\theta$ is the latitude and $\phi$ is the longitude on the planet's surface. $V$ is symmetric about the line-of-sight, is unity at the sub-observer point, drops as the cosine of the angle from the observer and is null on the far side of the planet from the observer; $I$ is symmetric about the star--planet line, is unity at the sub-stellar point, drops as the cosine of the angle from the star and is null on the night-side of the planet. Following \cite{Fujii_2010} and \cite{Kawahara_2010}, we define the normalized weight, \begin{equation} W(\theta, \phi, t) = \frac{V(\theta, \phi, t) I(\theta, \phi, t)} {\oint V(\theta, \phi, t) I(\theta, \phi, t) d\Omega}, \end{equation} which quantifies which regions of the planet are contributing the most to the observed light curve, under the assumption of diffuse (Lambertian) reflection. $W$ can be thought of as the smoothing kernel for the convolution between an albedo map, $A(\theta,\phi)$ and an observed lightcurve, $A^{*}(t)$. To first order, the character of a lightcurve can be understood in terms of the shape and location of the weight function. The normalization (denominator) of the weight is a simple function of phase: \begin{equation} \oint{V(\theta,\phi,t) I(\theta,\phi,t) d\Omega} = \frac{2}{3} \left[\sin\alpha + (\pi-\alpha)\cos\alpha \right]. \end{equation} As the planet rotates, $W$ sweeps from East to West. The width (in longitude) of the weight determines the longitudinal resolution achievable by inverting diurnal light curves: a broad $W$ at full phase leads to a coarser map than the slender $W$ of a crescent phase. This of course neglects the practical issues of inner working angles and photon counting noise. The peak of $W$ lies half-way between the sub-stellar and sub-observer points and corresponds to the location of the glint spot. The latitude of the glint spot may change throughout an orbit: the sub-observer latitude is fixed in the absence of precession, but the sub-solar latitude exhibits seasonal changes for non-zero obliquity. The peak of the weight is the area of the planet that contributes the most to the observed disc-integrated light curve. E.g., a polar sub-observer latitude and an equatorial sub-stellar latitude would yield a weight with a maximum at mid-latitudes. In detail, $W$ is also tempered by the usual $\sin\theta$ dependence of $d\Omega$ (ie: there is more area near the equator than near the poles). The dominant latitude (the latitude where the most photons would originate from in the case of a uniform Lambert sphere) is therefore not simply the peak of $W$, but is rather the average $\theta$, weighted by $W$: \begin{equation} \theta_{\rm dom}(t) = \oint{W(\theta, \phi, t) \theta d\Omega}= \frac{\oint{V I \theta d\Omega}}{\oint{V I d\Omega}}. \end{equation} In Table~1 we list the dominant latitude for the four EPOXI observations. Significantly, the dominant latitude is temperate for the ``polar'' observations, despite the exotic viewing geometry. This simple argument explains why the time-averaged colors of the polar EPOXI Earth observations are so familiar: most of the photons will not originate from the snowy and icy regions of Earth. That being said, the mid-latitudes probed by the polar observations are significantly more cloudy than the tropics \citep[yearly mean cloud cover in the tropics is 25--50\%, while at 45$^{\circ}$~S cloud cover is 75--100\%; see Figure~6a of][]{Palle_2008}. As shown in Appendix~I, clouds contribute a uniform (gray) increase in albedo of 20--50\% between the cloud-free and standard VPL models, so a latitudinal difference in cloud cover is a natural explanation for the observed difference in albedo. Furthermore, polar snow and ice necessarily contribute more to the polar than the equatorial EPOXI observations. For example, the Earth1 and Polar1 EPOXI observations were both obtained at the same time of year (March 2008 and 2009, respectively), so we may meaningfully ask how the different viewing geometries affect the contribution from snowy regions. If the global-mean snowline for March lies at $55^{\circ}$N, we find that 2\% of the weight is in snow-covered regions for the equatorial observation, while this fraction is 16\% for the polar observation. In general, for global mean snow lines between 50--60$^{\circ}$N, the snow-covered regions contribute 7--9 times more to the polar observation than to the equatorial observation. But such an argument is unlikely to work for the Polar2 observation, which probes the relatively land-free southern oceans, and the time-averaged Polar1 and Polar2 broadband colors are indistinguishable. We therefore believe the increased weight of clouds to be the main source of the 20--30\% greater absolute value of $A^{*}$ in the polar observations compared to the equatorial observations. \begin{figure}[htb] \includegraphics[width=84mm]{polar_earth_map.eps} \caption{\emph{Top:} Land coverage map for modern Earth. The colored lines indicate important latitudes for the EPOXI North (blue) and South (red) polar observations. The dotted lines show the sub-solar latitudes; the dashed lines show the sub-observer latitudes; the solid lines show the dominant latitudes, which are expected to contribute the most to the lightcurves. \emph{Bottom:} The longitudinal land coverage profiles for the EPOXI Polar observations.} \label{polar_earth_map} \end{figure} In the top panel of Figure~\ref{polar_earth_map}, we show a map of land coverage on Earth and indicate the sub-solar, sub-observer, and dominant latitudes for the North and South polar observations. The bottom panel of the figure shows the longitudinal land fraction profiles for the two polar observations, obtained by integrating the 2-D map by the weight function dictated by viewing geometry. It indicates the location of the major landforms probed by the polar observations and can be compared to the longitudinal profiles of eigencolors presented in the following section. \subsection{Lightcurve Inversion} The flux ratio primarily depends on the planet's orbital phase, the observer--planet--star angle, and the ratio $(R_{p}/a)^2$. We define the apparent albedo, $A^{*}$, as the ratio of the flux from the planet divided by the flux we would expect \emph{at the same phase} for a perfectly reflecting Lambert sphere \citep[see also][]{Qiu_2003}. This amounts to the average albedo of the planet, weighted by $W$: \begin{equation} A^{*}(t, \lambda) = \oint{W(\theta, \phi, t) A(\theta, \phi, \lambda) d\Omega} = \frac{\oint V I A d\Omega}{\oint V I d\Omega}. \end{equation} A uniform planet would have an apparent albedo that is constant over a planetary rotation in the $P_{rot} \ll P_{orb}$ limit; a uniform Lambert sphere would further have a constant apparent albedo during the entire orbit. For non-transiting exoplanets, the planetary radius may be unknown, in which case $A^{*}$ can only be determined to within a factor of $R_{p}^{2}$, with a lower limit on the radius obtained by setting $A^{*}=1$. Lightcurve inversion means inferring $A(\theta,\phi, \lambda)$ from $A^{*}(t, \lambda)$. If observations only span a single rotation, or if a planet has zero obliquity, one can only constrain the longitudinal variations in albedo. The visibility, $V(\theta,\phi,t)$, and illumination, $I(\theta,\phi,t)$, can be expressed compactly in terms of the locations of the sub-observer and sub-stellar points: \begin{equation} \begin{array}{ll} V(\theta,\phi,t) = &\max[\sin\theta \sin\theta_{\rm obs} \cos(\phi - \phi_{\rm obs}) \\ & + \cos\theta \cos\theta_{\rm obs}, 0]\\ I(\theta,\phi,t) = & \max[\sin\theta \sin\theta_{\rm star} \cos(\phi-\phi_{\rm star}) \\ & + \cos\theta \cos\theta_{\rm star}, 0], \end{array} \end{equation} where $\phi_{\rm obs}(t) = \phi_{\rm obs}(0) - \omega_{\rm rot}t$ is the sub-observer longitude, $\theta_{\rm obs}$ is the constant sub-observer latitude, $\phi_{\rm star}(t)$ and $\theta_{\rm star}(t) = \arccos[\cos(\xi_{0}+\omega_{\rm orb}t-\xi_{\rm obl})\sin\theta_{\rm obl}]$ are the sub-stellar longitude and latitude; $\omega_{\rm rot}$ and $\omega_{\rm orb}$ are the rotational and orbital angular velocities of the planet, $\xi_{0}$ is the initial orbital position of the planet, $\xi_{\rm obl}$ is the orbital location of northern summer solstice, and $\theta_{\rm obl}$ is the planet's obliquity. It is non-trivial to compute $\phi_{\rm star}(t)$ over a sizable fraction of an orbit, requiring a numerical integration or use of the equation of time. For the current application, however, the planet's rotation period is much shorter than its orbital period, so it is sufficient to assume that $\theta_{\rm star}$ is constant and $\phi_{\rm star}(t)$ advances linearly at one revolution per solar day (as opposed to the sidereal day used in computing $\phi_{\rm obs}(t)$). We use Horizons\footnote{http://ssd.jpl.nasa.gov/horizons.cgi} to compute the relative positions of the Deep Impact spacecraft, Earth and the Sun at the start of the various EPOXI campaigns. Both of the EPOXI polar observations were obtained with a viewing geometry very close to quadrature\footnote{For small obliquities, polar observations imply that the planet is in a nearly face-on orbit, and therefore permanently at quadrature. But there is no reason to assume low obliquity for terrestrial planets, so polar observations need not be made at quadrature.}. The weight function therefore has a width of 90$^{\circ}$ in longitude, indicating that we would need a model with 8 longitudinal slices of uniform albedo to achieve Nyquist sampling of the rotational lightcurve, in the absence of specular reflection \citep[for more discussion on slice vs sinusoidal longitudinal profiles see][]{Cowan_2008, Cowan_2009}. An 8-slice model with variable phase offset (prime meridian) would have 9 model parameters; we instead use sinusoidal maps with terms up to fourth order, which also have 9 model parameters but which converge better \citep{Cowan_2008}. The best-fit reduced $\chi^{2}$ of these models are somewhat lower than unity because the $\sigma_{A^{*}}=0.01$ ``error bars'' that we use are much larger than the point-to-point scatter in the lightcurves. We estimate uncertainties in our longitudinal eigencolor maps with a Monte Carlo test. Using our adopted photometric error of $\sigma_{A^{*}}=0.01$, we generate 10,000 statistically equivalent instances of the observed lightcurves assuming Gaussian, uncorrelated errors. We run the same PCA and lightcurve inversion on each of these mock data sets and take the standard deviation in the resulting maps to be the uncertainty in our fiducial maps. Note that the rotational inversion may be performed directly on the lightcurves shown in Figures 1 \& 2, and independently of the PCA described in \S~3, yielding albedo maps of Earth in various wavebands. Instead, we combine PCA and lightcurve inversion as we did in \cite{Cowan_2009} and produce longitudinal maps of the dominant eigencolors, shown in Figures~\ref{polar1_eigen_maps} \& \ref{polar2_eigen_maps}. Based on the numerical experiments of \S~3.1, we expect both dominant eigencolors in Figure~\ref{polar1_eigen_maps} to be tracking clouds and snow-covered land, while the red eigencolor is also sensitive to cloud-free land. Since clouds are more prevalent at these latitudes, the red eigencolor does not faithfully locate the major landforms, as it did in \cite{Cowan_2009}. The southern polar observation (Figure~\ref{polar2_eigen_maps}) shows a broad maximum in the gray eigencolor at a longitude of 90$^{\circ}$~W, roughly corresponding to the location of Patagonia and the Antarctic Peninsula. Since snow-covered land is essentially indistinguishable from clouds at these poor spectral resolutions, we must remain agnostic about the source of this variability. \begin{figure}[htb] \includegraphics[width=84mm]{Polar1_eigen_maps.eps} \caption{{\bf North:} Longitudinal profiles of the two dominant eigencolors of Earth based on the lightcurves in Figure~\ref{polar1_lightcurves}, the eigencolors shown in Figure~\ref{polar1_eigenspectra}, and the known phase and rotational period of Earth.} \label{polar1_eigen_maps} \end{figure} \begin{figure}[htb] \includegraphics[width=84mm]{Polar2_eigen_maps.eps} \caption{{\bf South:} Longitudinal profiles of the dominant eigencolor of Earth based on the lightcurves in Figure~\ref{polar2_lightcurves}, the eigencolors shown in Figure~\ref{polar2_eigenspectra}, and the known phase and rotational period of Earth.} \label{polar2_eigen_maps} \end{figure} In general, the cloudy mid-latitudes keeps the eigencolor maps from faithfully identifying major landforms (eg, compare Figures~\ref{polar1_eigen_maps} and \ref{polar2_eigen_maps} to Figure~\ref{polar_earth_map}). \section{Albedo Model of Snowball Earth} Since the polar regions of Earth are largely covered in snow and ice, it is worth asking if one might confuse a habitable planet like Earth with a snowball planet \citep[ie: one caught in the cold branch of a snow-albedo positive feedback loop. See, for example, ][and references therein]{Tajika_2008}. In this section we describe a toy model for the reflectance of such a snowball planet, and compare the resulting photometry to the EPOXI polar observations. Note that \cite{Vasquez_2006} presented \emph{bolometric} (white light) diurnal lightcurves for a model snowball Earth, but these are not useful for the current comparison. \subsection{Geography} The geography of our snowball Earth model is shown in Figure~\ref{continents_map}. We use the same idealized paleogeography for the Sturtian glaciation ($\sim$750~Mya) as \cite{Pierrehumbert_2005}. Paleomagnetism only constrains the magnetic \emph{latitude} of continents, however, and we are at liberty to choose any longitudinal distribution. The diurnal variability of the planet is determined solely by its longitudinal geography, however. If the continents are spread out uniformly in longitude, for example, the planet would not exhibit any rotational variability in the absence of heterogeneous cloud cover. We instead adopt the opposite limit of a single mega-continent. This will tend to exaggerate the amplitude of the diurnal variations in apparent albedo, but the changes in color should be robust. Assuming that sea-level was not grossly different from today, and that only trace amounts of continent-formation has occurred in the intervening 750 Myrs, continents should cover $\sim 25$\% of the planet, as today. At first sight, assuming constant water levels during a global glaciation seems inconsistent. However, it is only icecaps (ice on land) that significantly change water levels, while the geological evidence for snowball Earth episodes instead require that the \emph{oceans} be frozen at the equator. In fact, this criterion effectively shuts down the planet's hydrological cycle, so the polar ice caps are not very different from the present day. In any case, the high latitudes of a snowball Earth should be covered in snow, regardless of whether the underlying regions are continent or ocean. Therefore, the precise fraction of land vs ocean does not directly impact the observed diurnal variability. For a snowball planet, we assume that both oceans and continents are covered in snow at latitudes greater than 30$^{\circ}$. Closer to the equator, we assume that continents are bare, dry land due to the very low precipitation; while oceans are covered in blue glacier ice that flows towards the equator \citep{Goodman_2003}. \begin{figure}[htb] \includegraphics[width=84mm]{continents_map.eps} \caption{The layout of surface types in our Snowball Earth model. Regions within 60$^{\circ}$ of the poles ---both continents and oceans--- are covered in snow. The tropics are dry (bare land or glacier ice) due to negative net precipitation.} \label{continents_map} \end{figure} \subsection{Albedo Spectra} Climate models of Snowball Earth are concerned with albedo only insofar as it modulates the energy budget of the planet. That is to say, they care about the Bond albedo, $A_{\rm B}$, the fraction of incident energy that is reflected back into space. We, on the other hand, are concerned with the appearance of the planet as seen from the outside, $A^{*}(t,\lambda)$. It is beyond the scope of this paper to make a detailed snowball earth spectral model. We therefore assume diffuse (a.k.a. Lambert) reflection and use a wavelength-dependent albedo, $A(\lambda)$, that is simply a function of location rather than a bidirectional reflectance distribution function. A hard Snowball Earth will not have appreciable expanses of liquid water to contribute to glint. That being said, other elements, notably clouds, are not strictly Lambertian \citep[eg,][]{Robinson_2010}. For a cloud-free and airless model planet, the albedo spectrum at each point on the planet is simply determined by the albedo spectrum of the surface type at that location. Reflection from clouds and Rayleigh scattering from air molecules complicates this picture, however. We treat the albedo from these semi-transparent media as follows: \begin{equation}\label{combine_albedo} A(\lambda) = 1 - (1-A_{\rm Ra}(\lambda))(1-A_{\rm cl}(\lambda))(1-A_{\rm surf}(\lambda)), \end{equation} where $A_{\rm Ra}$ is the effective albedo due to Rayleigh scattering, $A_{\rm cl}$ is the albedo due to clouds, and $A_{\rm surf}$ is the albedo of the surface type at that point on the planet. This simple expression captures the essential behavior of clouds and Rayleigh scattering: they always increase the effective albedo of a region, but the effect is most pronounced for a dark underlying surface. \begin{figure}[htb] \includegraphics[width=84mm]{surface_albedo_spectra.eps} \caption{The broadband optical albedo spectra for the components of our snowball Earth model. The Bond albedos of the surfaces are estimated using a Solar spectrum and integrating to 5~$\mu$m (snow: 0.8; glacier ice: 0.6; thin cloud: 0.3, dry land: 0.3, Rayleigh scattering: $<0.1$).} \label{surface_albedo_spectra} \end{figure} Our model has 5 elements, each with a distinctive albedo spectrum, shown in Figure~\ref{surface_albedo_spectra}. We use spectra for dry land and snow from Robinson et al. (2011). The snow albedo spectrum we use is for medium grained snow, while the cold, dry climate of a snowball Earth would create small-grained snow, as seen in Antarctica \citep{Hudson_2006}. There is no perceptible difference in the broadband albedo spectra of these two kinds of snow at optical wavelengths, however. We use the empirical albedo spectrum for blue glacier ice from \cite{Warren_2002}. To mimic the thin clouds expected on a frozen planet with reduced hydrological activity, we take a generic cloud spectrum (Robinson 2009, priv. comm.) and divide the albedo by 2. (Using thicker clouds increases the Bond albedo of our model planet but does not significantly change the color variability.) We distribute the clouds on the planet using a snapshot of cloud maps from a snowball Earth general circulation model \citep{Abbot_2010}. Note that this model was run using the same idealized geography shown in Figure~12 and thus offers a good estimate of the spatial ---and in particular longitudinal--- variations in cloud cover. We estimate the disc-integrated effect of Rayleigh scattering by comparing the Standard and Rayleigh-Scattering-Free VPL models (described in Appendix~I) and using Equation~\ref{combine_albedo}. \begin{figure}[htb] \includegraphics[width=84mm]{time_averaged_spectrum.eps} \caption{Time-averaged albedo spectrum of Snowball Earth. The vertical bars show the RMS variability in each band. Relative peak-to-trough variability ranges from 42\% (450~nm) to 6\% (950~nm).} \label{time_averaged_spectrum} \end{figure} Our snowball Earth model has a time-averaged Bond albedo of 70\%, which is self-consistent with the snow-albedo feedback\footnote{Despite its name, a snowball planet's albedo does not equal that of snow (80\%). The cold, dry atmosphere keeps the tropical land bare ($A_{B}=30\%$) and exposes glacier ice ($A_{B}=60\%$) near the equator.}: \cite{Wetherald_1975} used a 70\% albedo to induce their ``White Earth'' solution in a 2-D model. \cite{Pierrehumbert_2005} ran GCMs of a hard snowball Earth and maintained the snowball state with albedos of 60--67\%. \cite{Chandler_2000}, on the other hand, ran snowball Earth global climate models (GCMs) with bolometric albedos of 20--40\%; \cite{Vasquez_2006} have an average Bond albedo of approximately 50\% for the frozen Earth. \subsection{Time Variability} For directly-imaged exoplanets, the albedo cannot be determined independently of the planet's radius: photometry of reflected light will constrain the quantity $A R_{p}^{2}$. Therefore, while the general agreement in $A_{B}$ between our toy model and self-consistent simulations is encouraging, one cannot in general use the absolute albedo of an exoplanet as a diagnostic. Multiband observations will tell us about the colors of the planet, however, and with sufficiently high cadence observations it will be possible to measure the variations in apparent albedo due to the planet's rotation. We compute the variations in apparent albedo for an equatorial observer and a planet at quadrature. The time-averaged spectrum of our snowball Earth model (Figure~\ref{time_averaged_spectrum}) is completely different from that of the modern Earth, regardless of viewing geometry (Figure~\ref{average_albedo_spectrum}). The flat spectrum shortward of 650~nm is due to snow and glacier ice, which are so reflective at these wavelengths as to make Rayleigh scattering imperceptible. The drop in albedo at longer wavelengths is also driven by snow and glacier ice. For our snowball model, the shortest wavelengths exhibit the most variability, as shown by the vertical bars in Figure~\ref{time_averaged_spectrum}. This is because bare land and glacier ice exhibit the largest contrast at short wavelengths, while at longer wavelengths they both have near-IR albedos of $\sim$40\%. This is in stark contrast to the case for the modern Earth, which exhibits variability at all wavebands. We conclude that ---given high-quality photometry--- the modern-day Earth could not be mistaken for a snowball planet, regardless of the viewing geometry. \section{Summary} We presented time-resolved, disc-integrated observations of Earth's polar regions from the Deep Impact spacecraft as part of the EPOXI Mission of Opportunity. These complement the equatorial views presented in \cite{Cowan_2009}. We found that both of the polar observations have broadband colors similar to the equatorial Earth, but with uniformly higher albedos. We explained this in terms of the 2-D weight function for disc-integrated observations of Earth, which was most sensitive to the tropics for the equatorial observations, and most sensitive to mid-latitudes for the polar observations. We performed PCA on a suite of simulated rotational multi-band lightcurves from NASA's Virtual Planetary Laboratory 3D Earth model. We found that PCA correctly indicates the number of different surfaces rotating in and out of view. We found that while the red eigencolor consistently tracks cloud-free land, a blue eigencolor only tracks oceans when clouds are entirely absent from the simulation. In the general (cloudy) case, a blue eigencolor is simply tracking cloud inhomogeneities. Gray eigencolors, when they are present, track large cloud patterns and/or snow-covered land. We also performed PCA on the EPOXI polar observations. Comparing these eigencolors to known surface types on Earth, we establish that the variability seen in the North EPOXI Polar observation is due to clouds, continents and oceans rotating in and out of view; the lack of large cloud-free land (ie: deserts) at the latitudes probed by these observations keep us from being able to faithfully extract the positions of major northern landforms. The South polar observation, on the other hand, was characterized by gray variability due to a large cloud pattern in the south oceans. Lastly, we constructed a simple reflectance model for a snowball Earth, and found that both the time-averaged broadband colors and diurnal color variations of a Snowball Earth (gray snow $+$ near-IR roll-off; variability at blue wavelengths) would be distinguishable from the modern-day Earth (near-UV Rayleigh ramp $+$ gray clouds; variability at all wavelengths), regardless of viewing angle. \section{Discussion} We listed the possible constraints one could get from time-resolved photometry in Section~1, here we briefly review the three forms of photometric characterization used in this paper in the context of measurements from planned space telescopes. We have considered time-resolved, multi-band optical photometry, which could be obtained with a space-based high-contrast imaging mission. On its own, such a telescope could discover nearby exoplanets, determine their approximate orbits, and characterize their time-averaged colors and time-variability of colors. The \emph{time-averaged colors} of a planet will be the most accessible observable. We have shown in this paper that the broadband colors would be sufficient to distinguish between the modern day Earth and a snowball Earth. On modern-day Earth, polar ice and snow don't contribute significantly to the time-averaged albedo of Earth ---even with polar viewing geometries--- because of the glancing angle of sunlight at those latitudes. The coldest regions of a planet receive the least sunlight, and therefore contribute correspondingly little to the disk-integrated properties of the planet. The \emph{time-variability of the colors} are harder to obtain, requiring shorter integration times and therefore a larger telescope, all other things being equal. On the other hand, variability measurements are more robust to contamination from exo-zodiacal light. We have shown that modern Earth ---regardless of viewing angle--- exhibits photometric variability at all wavelengths (RMS variability within a factor of 2 for all wavebands), while Snowball Earth varies 7 times more at short wavelengths than at long wavelengths. A more subtle analysis of the time variability may even allow us to distinguish between the equatorial and polar viewing geometries of Earth, because clouds play a larger role at mid-latitudes. From an equatorial vantage point, the dominant eigencolor is red, followed by blue; for the polar geometries, the ordering of the red and blue eigencolors is flipped, or there is a single dominant grey eigencolor. If the planet's radius can be estimated, then its albedo can be put on an absolute scale and one can estimate its Bond albedo. Transiting planets have very well characterized radii, but nearby earth analogs will almost certainly \emph{not} be transiting. Instead, a radius estimate will require an additional large space mission: either an infrared high-contrast imaging telescope, or a space based astrometry mission. In the first case, thermal and reflected photometry can be combined to estimate the planet's radius. If thermal photometry is obtained at a variety of phases, then the efficiency of heat transport to the planet's night-side may be estimated \citep{Cowan_2011b}, and the systematic uncertainty will be $\sim2$\% due to the unkown efficiency of latitudinal heat transport \citep{Cowan_2011}. The uncertainty in the radius will therefore likely be dominated by the ---known--- uncertainties in thermal and reflected photometry. In the second case, the star's astrometric wobble provides a mass measurement for the planet; by assuming a planetary density, one can estimate the planet's radius. The dominant source of uncertainty here is the planet's composition: given a mass, a planet's radius may vary by 50\% \citep[eg,][]{Charbonneau_2009, Batalha_2011}, leading to absolute albedo estimates only valid to within a factor of 2. Transiting planet surveys will likely reduce these systematic uncertainties by providing an empirical mass-radius relation for planets across a wide range of masses. It is not clear to what extent the \emph{Kepler} mission \citep[eg,][]{Borucki_2011} will help define the mass-radius relation: although the vast majority of \emph{Kepler} candidates are likely to be \emph{bona fide} planets \citep{Morton_2011}, most will not have mass estimates. The smaller, and better characterized, radius uncertainty would therefore most likely come from combining optical and infrared photometry, rather than from a mass measurement. The \emph{Bond albedo} would allow us to better distinguish between the equatorial ($A_{B}\approx 0.3$) and polar ($A_{B}\approx 0.4$) EPOXI observations, or between a snowball planet ($A_{B}\approx 0.7$) and a temperate one. In general, this quantity would be very useful in determining a planet's energy budget and would go a long way towards constraining its habitability. \acknowledgments This work was supported by the NASA Discovery Program. We thank D.S.~Abbot and R.T.~Pierrehumbert for providing us with cloud maps of snowball Earth. N.B.C.\ acknowledges many useful discussions with S.G.~Warren about snowball Earth, and thanks W.~Sullivan for encouraging him to complete his astrobiology research rotation. E.A.\ is supported by a National Science Foundation Career Grant.

\section{Introduction} GW170729 was observed on July 29, 2017 by the Advanced LIGO~\cite{TheLIGOScientific:2014jea} twin detectors. Its detection was announced in~\cite{LIGOScientific:2018mvr} as part of GWTC-1, the gravitational-wave (GW) transient catalog of compact binary coalescences (CBCs)~\cite{GWOSC:GWTC}. As reported in~\cite{LIGOScientific:2018mvr}, GW170729 was emitted during the coalescence of two stellar-mass black holes (BH). It was observed during the offline analysis of the detection pipelines GstLAL~\cite{Messick:2016aqy} and PyCBC~\cite{Usman:2015kfa,Nitz_2017}, that search for signals from CBC events, as well as cWB~\cite{Klimenko:2015ypf}, a pipeline tuned to search for transient signals whose frequency increases with time. A number of reasons make GW170729 unique among the binary BHs (BBHs) presented in GWTC-1. At a measured source-frame total mass of $\sim 85 M_{\odot}$ and a distance of $\sim 3$Gpc (median values), it is likely the most massive and distant BBH. Additionally, it is one of only two GW events that show evidence for nonzero spins with an effective spin of $(0.11-0.58)$ at the 90\% credible level~\cite{LIGOScientific:2018mvr}. Finally, it is the only event for which the more flexible, non-CBC-specific cWB search returns a lower false alarm rate than the CBC-specific GstLAL and PyCBC template-based searches. In spite of these, Ref.~\cite{LIGOScientific:2018jsj} concludes that GW170729 is consistent with the population of the other BBH detections \footnote{An additional candidate claimed in~\cite{Zackay:2019tzo}, if confirmed, would also correspond to a binary with a non zero effective spin.}. The fact that GW170729 is the most massive BBH found so far makes it a good candidate to observe the effects of higher-order waveform modes. The GW emission from BBHs can be described as a superposition of GW modes $h_{\ell,m}$ as $h\equiv h_+ - i h_{\times}=\sum_{\ell,m} Y_{\ell,m}^{-2}(\iota,\varphi)h_{\ell,m}(t,\vec{\theta})$ \cite{Thorne:1980ru}. Here $h_+$ and $h_\times$ are the two GW polarizations, $Y_{\ell,m}^{-2}$ denote spin-2 weighted spherical harmonics \cite{sharm}, which depend on the location $(\iota,\varphi)$ of the observer around the binary, while the modes $h_{\ell,m}$ depend on the masses and spins of the binary, denoted by $\vec{\theta}$. During most of the inspiral stage, $h$ is dominated by the quadrupole modes, $({\ell,m})=(2,\pm 2)$. The rest, known as higher-order modes, grow in strength during the merger and ringdown stages, their impact being larger for highly asymmetric and nearly edge-on binaries \cite{Pan:2011gk,Varma:2014jxa,Bustillo:2016gid,Pekowsky:2012sr,Varma:2016dnf,Bustillo:2015qty,Graff:2015bba}. Finally, for more massive BBH systems the inspiral emission moves out of the sensitive band of advanced detectors, while sensitivity to the merger-ringdown increases and so does the impact of higher-order modes \cite{VanDenBroeck:2006ar,Bustillo:2015qty,Graff:2015bba}. Standard detection and parameter estimation of BBH events is usually performed using GW templates without the higher-order mode content of the signals~\cite{2018arXiv181205121M,DalCanton:2017ala,TheLIGOScientific:2016pea,LIGOScientific:2018mvr}. Reference~\cite{LIGOScientific:2018mvr} studied the fact that the event is recovered with higher significance by the flexible unmodeled, but less sensitive, search than the template-based searches~\cite{CalderonBustillo:2017skv}. By performing injections of signals without higher-order modes or spin-precession, it was argued that the difference in the measured significances is in fact not unlikely. It was shown that $\sim 4\%$ of the injected signals were recovered with a higher significance from cWB than PyCBC. At the same time, the presence of strong higher-order modes in the GW signal can potentially lead to biased parameter estimation if they are omitted in the waveform templates \cite{Littenberg:2012uj,Varma:2016dnf,Varma:2016dnf,Bustillo:2015qty,Graff:2015bba}. So far all reported events are consistent with nearly equal-mass, face-on BBHs, a fact that has prevented such biases, as shown in \cite{Abbott:2016wiq,Abbott:2016apu,Kumar:2018hml} for the case of events observed during the first observation run of Advanced LIGO. Even in this case, usage of models with higher-order modes can improve the accuracy of parameter estimation \cite{Littenberg:2012uj,Graff:2015bba,London:2017bcn,CalderonBustillo:2018zuq}. Consequently, a reanalysis of GW150914 and GW170104 events using models with higher-order modes obtains modestly tighter parameter constraints with respect to previous analyses \cite{Kumar:2018hml}. In this paper we present a detailed investigation into the properties of GW170729. We carry out a parameter estimation analysis similar to the one in~\cite{LIGOScientific:2018mvr} in order to study the effect of spin-precession, higher-order modes, and spin priors on inferences drawn about GW170729. We make use of a more extended set of CBC waveform models belonging to three distinct waveform model families: phenomenological~\cite{Ajith:2007qp,Ajith:2009bn,Santamaria:2010yb,Husa:2015iqa,Khan:2015jqa,London:2017bcn,Hannam:2013oca}, effective-one-body~\cite{Taracchini:2013rva,Bohe:2016gbl,Pan:2013rra,Cotesta:2018fcv,Pan:2011gk,Babak:2016tgq}, and numerical relativity~\cite{2013PhRvL.111x1104M,2016CQGra..33t4001J,2017CQGra..34v4001H,2015PhRvL.115l1102B,2017PhRvD..95j4023B,2017PhRvD..96b4058B}. This set includes two new improved spin-aligned waveform models that include the effect of higher-order waveform modes. We gauge the importance of a physical phenomenon, namely spin-precession and higher-order modes, by comparing posterior densities for various source parameters obtained through analyses using waveform models with and without that physical phenomenon included. While we find no indication of spin-precession, higher-order modes have a distinct impact on the posterior density for various source parameters. We find that CBC waveform models that include higher-order modes result in posterior distributions for the mass ratio of the binary that are shifted away from unity, resulting in more support for unequal masses than originally concluded in ~\cite{LIGOScientific:2018mvr}. In particular, we find a highest probability density (HPD) interval of the mass ratio of $(0.3-0.8)$ at the 90\% credible level, while the corresponding upper limit of the HPD interval on the mass ratio without higher-order modes is $\sim 0.96$. This improved measurement, obtained using waveform models that include more physical effects, shows that GW170729 is not consistent with the merger of two equal-mass BHs at the $90\%$ level. At the same time, models with higher-order modes lead to marginally less support for positive effective spin $\chi_{\rm{eff}}$~\footnote{The effective spin parameter is defined as the sum of the mass-weighted projections of the component spins along the orbital angular momentum and it is conserved to at least the second post-Newtonian order~\cite{Racine:2008qv}.} and binary orientations where the orbital angular momentum points along or away from the line of sight. In particular, we find that the probability that the effective spin is positive is reduced from $99\%$ when higher-order modes are omitted to $94\%$ when they are included in the waveform models. We obtain consistent results when we use various CBC waveform models from different waveform families to describe the data. We thus argue that our conclusions are robust against systematic errors in the waveform models. As the source-frame mass of the more massive BH is close to the proposed mass upper limit due to pair instability\footnote{It has been suggested that pulsational pair instability supernovae will result in no BH remnants with masses above $\sim 50 M_{\odot}$ as the remnant is disrupted during the explosion, e.g.~\cite{Woosley_2017,Marchant:2018kun}.}~\cite{2002ApJ...567..532H} and the posterior for the binary mass ratio favors unequal masses, we further investigate the possibility of second generation (2g) merger~\cite{davide2g,Fishbach:2017dwv}. In a 2g merger scenario, the primary BH is the remnant of an earlier BBH merger. As such, it is expected to be more massive than its companion in GW170729 resulting in unequal binary masses, and to have a relatively large spin magnitude. We contrast this scenario to a first generation (1g) merger scenario which favors comparable component masses. We reanalyze the data using two different priors tailored to 1g and 2g mergers and calculate the Bayes Factor (BF) of the 2g versus the 1g hypotheses. We find a BF of 4.7:1(1.4:1) in favor of the 2g scenario when using waveforms with (without) higher-order modes. This value favors the 2g model, but not decisively so, in agreement with the results of~\cite{LIGOScientific:2018jsj}. This question has been earlier and independently addressed by Kimball et al.~\cite{Kimball:2019mfs} and our results are in agreement with their results. Finally, we compare the signal reconstruction obtained with CBC waveform models to a morphology-independent reconstruction~\cite{Cornish:2014kda}. We quantify the consistency between the CBC and the generic reconstruction by computing the noise-weighted overlap between the two. We find broad consistency between the two with overlap values typical of what is expected for this mass range and signal strength~\cite{TheLIGOScientific:2016uux}. This result is only minimally affected by the inclusion of higher-order modes. Posterior samples from all our analyses are available in~\cite{170729SamplesRelease}. The rest of the paper presents our analysis and conclusions in detail. Section~\ref{analysis} describes the analysis we carry out including the CBC waveform models, the priors, and the generic analysis. Section~\ref{CBCresults} presents results derived under the CBC waveform models and posterior densities for the source parameters. Section~\ref{BWresults} gives the results of the generic analysis and how they compare to the CBC-specific analysis. Finally, Sec.~\ref{conclusions} presents our main conclusions. \section{Analysis} \label{analysis} In this section we describe the details of the analysis we perform including the data, waveform models, and inference approaches we use. Our analysis follows closely and builds off of the work originally presented in~\cite{LIGOScientific:2018mvr}. We employ two complementary approaches: one is based on waveform models constructed specifically to describe compact binary coalescences, while the other uses a more flexible waveform model that can capture unexpected signal morphology. We describe both analyses in the following. \subsection{Data and setup} We use the publicly available LIGO and VIRGO strain data for GW170729 from the Gravitational Wave Open Science Center~\cite{GWOSC,Vallisneri:2014vxa}. The LIGO strain data have been post-processed to subtract several sources of instrumental noise~\cite{Driggers:2018gii,Davis:2018yrz} and calibrated as described in~\cite{LIGOScientific:2018mvr}. In particular we analyze $4$s of strain data centered at the GW170729 trigger time. The analysis covers a bandwidth from $f_{\rm low}=20$~Hz with the upper frequency cutoff set to $f_{\rm high}=1024$~Hz for waveform models without higher-order modes and $f_{\rm high}=2048$~Hz when higher-order modes are included. For masses typical of GW170729 (a detector-frame total mass of 120$M_{\odot}$ and mass ratio of 0.5) this upper frequency cutoff ensures that the analysis includes up to at least the $\ell=5$ ringdown harmonic of a compact binary merger, the highest frequency mode available in the waveform models we use. We assume that the noise in the three detectors is Gaussian and stationary. The power spectral density (PSD) of the noise is obtained from the same $4$s of on-source data with the technique described in~\cite{Littenberg:2014oda}. Specifically, a model consisting of a cubic spline and a number of Lorenzians is used to obtain posterior samples for the PSD from which a median PSD value is computed separately for each frequency bin. This median PSD is used in the estimation of the likelihood function in {\tt LALInference}, {\tt RIFT}, and {\tt BayesWave} and we use the same PSD as~\cite{LIGOScientific:2018mvr} that is publicly available in~\cite{GW170729PSD}. \subsection{ CBC waveform models} \begin{table*}[] \centering \begin{tabular}{c|ccccccccc} \hline \hline Waveform Model &&& Spin Dynamics &&& Modes ($\ell,|m|$) &&& Algorithm \\ \hline {\tt IMRPhenomPv2}~\cite{Hannam:2013oca} &&& Precessing &&& (2,2) &&& {\tt LALInference}\\ {\tt IMRPhenomD}~\cite{Husa:2015iqa,Khan:2015jqa} &&& Aligned &&& (2,2) &&& {\tt LALInference}\\ {\tt IMRPhenomHM}~\cite{London:2017bcn} &&& Aligned &&& (2,2),(2,1),(3,3),(3,2),(4,4),(4,3) &&& {\tt LALInference}\\ {\tt SEOBNRv3}~\cite{2014PhRvD..89f1502T,2014PhRvD..89h4006P} &&& Precessing &&& (2,2) (2,1) &&& {\tt LALInference}\\ {\tt SEOBNRv4}~\cite{Bohe:2016gbl} &&& Aligned &&&(2,2) &&& {\tt LALInference}\\ {\tt SEOBNRv4HM}~\cite{Cotesta:2018fcv} &&& Aligned &&& (2,2), (2,1), (3,3), (4,4), (5,5) &&& {\tt LALInference}\\ NR HM ~\cite{2013PhRvL.111x1104M,2016CQGra..33t4001J,2017CQGra..34v4001H} &&& Aligned &&& ($\ell \leq 4$,$|m|\leq\ell$) &&& {\tt RIFT}\\ NR/{\tt NRSur7dq2}{} HM ~\cite{2015PhRvL.115l1102B,2017PhRvD..95j4023B,2017PhRvD..96b4058B} &&& Aligned &&& ($\ell \leq 4$,$|m|\leq\ell$) &&& {\tt RIFT}\\ NR ~\cite{2013PhRvL.111x1104M,2016CQGra..33t4001J,2017CQGra..34v4001H} &&& Aligned &&& ($\ell=2$,$|m|\leq\ell$) &&&{\tt RIFT}\\ NR/{\tt NRSur7dq2}{} ~\cite{2015PhRvL.115l1102B,2017PhRvD..95j4023B,2017PhRvD..96b4058B} &&& Aligned &&& ($\ell=2$,$|m|\leq\ell$) &&& {\tt RIFT}\\ \hline Wavelets~\cite{Cornish:2014kda} &&& Flexible &&& Flexible &&&{\tt BayesWave}\\ Chirplets~\cite{Millhouse:2018dgi} &&& Flexible &&& Flexible &&& {\tt BayesWave}\\ \end{tabular} \caption{List of waveform models we use to model the GW signal. The second and third columns indicate the spin dynamics and higher-order content $(\ell, |m|)$ of each model in the coprecessing frame respectively. The fourth column gives the algorithm we use with each model. The bottom horizontal line separates the CBC-specific models and the morphology-independent models. } \label{table:pe-table} \end{table*} The top half of Table~\ref{table:pe-table} lists the CBC waveform models we use; these models describe the inspiral, merger, and ringdown signal from the coalescence of two BHs as predicted by General Relativity (GR). All CBC waveforms we use can be divided into three main families: (i) phenomenological models ({\tt IMRPhenom}), effective-one-body models ({\tt SEOBNR}), and numerical relativity (NR). The first family is based on results of post-Newtonian theory~\cite{Blanchet:2014zz} to compute the inspiral phase and a phenomenological approach to describe the merger, aided by calibration to EOB-NR hybrid waveforms~\cite{Hannam:2013oca,Husa:2015iqa,Khan:2015jqa}. The second family uses the effective-one-body approach~\cite{Buonanno:1998gg,Buonanno:2000ef}, which is based on a resummation of post-Newtonian results to describe the inspiral, and uses calibration to NR simulations for the late-inspiral and merger~\cite{Zilhao:2013hia,Loffler:2011ay,Lovelace:2010ne,2013PhRvL.111x1104M}. Both families describe the ringdown employing results of BH perturbation theory~\cite{Barausse:2011kb,Taracchini:2014zpa}. The NR waveforms are obtained by solving the full non-linear Einstein equations and are subject only to numerical errors~\cite{ 2010RvMP...82.3069C}. Besides the waveform family, models also differ on whether they include the effect of spin-precession~\cite{Apostolatos:1994mx,Kidder:1995zr} and higher-order modes, as indicated in Table~\ref{table:pe-table}. From the {\tt IMRPhenom} and {\tt SEOBNR} families we use one aligned-spin and one spin-precessing model without higher-order modes as well as a spin-aligned model with higher-order modes. None of the models from these two waveform families currently include both precessing spins and higher-order modes. In particular, from the {\tt IMRPhenom} family, we use: the spin-precessing {\tt IMRPhenomPv2}{}~\cite{Hannam:2013oca} and the spin-aligned {\tt IMRPhenomD}{}~\cite{Husa:2015iqa,Khan:2015jqa} models. From the {\tt SEOBNR} family we use: the spin-precessing {\tt SEOBNRv3}{}\footnote{{\tt SEOBNRv3}{} includes the modes $(\ell, |m|) = (2,2), (2,1)$ in the coprecessing frame, the coordinate system for which the $z$-axis is instantaneously aligned with the Newtonian angular momentum, see~\cite{Buonanno:2002fy}.}~\cite{2014PhRvD..89f1502T,2014PhRvD..89h4006P} and the spin-aligned {\tt SEOBNRv4}{}\footnote{In particular, we use the reduced order model implementation of {\tt SEOBNRv4}{} which is computationally less expensive~\cite{Purrer:2014fza,Purrer:2015tud}.}~\cite{Bohe:2016gbl} models. As far as higher-order modes are concerned, we use {\tt IMRPhenomHM}{}~\cite{London:2017bcn}, a spin-aligned model of the {\tt IMRPhenom} family that includes the $(\ell, |m|) = [(2,2),(2,1),(3,3),(3,2),(4,4),(4,3)]$ higher-order modes and {\tt SEOBNRv4HM}{}~\cite{Cotesta:2018fcv}, a spin-aligned model of the {\tt SEOBNR} family that includes the $(\ell, |m|) = [(2,1),(3,3),(4,4),(5,5)]$ higher-order modes. Posterior samples obtained with {\tt IMRPhenomPv2}{} and {\tt SEOBNRv3}{} have already been made publicly available by the LIGO-Virgo Collaborations and we use them directly~\cite{GW170729Samples}. The NR simulations we use include a total of 763 spin-aligned and 625 spin-precessing simulations~\cite{2013PhRvL.111x1104M,2016CQGra..33t4001J,2017CQGra..34v4001H}. We optionally augment the list of NR simulations with waveforms computed using {\tt NRSur7dq2}{}~\cite{2015PhRvL.115l1102B,2017PhRvD..95j4023B,2017PhRvD..96b4058B}, a surrogate model directly based on NR. The surrogate model we use is valid for mass ratio $0.5\le q\le1$ and dimensionless spin magnitude $\chi\le0.8$; however, the analysis performed results in a full posterior due to the inclusion of the NR waveforms that cover the remaining region. For both NR-related analyses, we include results with higher-order modes $(\ell\leq4, |m|\leq\ell)$ and with only the $(\ell=2, |m|\leq\ell)$ modes. When using the NR simulations we also assume that the spins remain aligned to the orbital angular momentum (no spin-precession). \subsection{Priors} Our analysis employs the following priors for the source parameters. The detector-frame component masses $m_1, m_2$ are assumed to be uniform between $10M_{\odot}$ and $200M_{\odot}$ with $m_1>m_2$, while the mass ratio $q\equiv m_2/m_1$ is restricted to be above 0.125. The sky location and orientation of the binary, as well as the directions of the component spins are uniform on the unit sphere. The distance is uniform in volume with a maximum cut off of $7$Gpc, while the time and phase of arrival are uniform. We have verified that the mass and distance prior ranges encompass the entire region where the posterior distribution has non-negligible support. For the magnitude of the dimensionless component spins $\chi_i, \,i\in \{1,2\},$ we make different choices in order to investigate how this affects the posterior. The first prior is uniform-in-$\chi$ up to 0.99 for both spin-aligned and spin-precessing waveform models. The second prior is uniform-in-$\chi_z$, where $\chi_z$ is the spin projection along the axis perpendicular to the orbital plane, with the restriction that the spin magnitude is below 1. For spin-precessing model, the in-plane $\chi_x$ and $\chi_y$ components are also uniform; in that case this prior is sometimes referred to as `volumetric' prior as it corresponds to the spin vector being uniformly distributed within the unit sphere. Finally, we also use priors targeted toward CBCs where the primary component is the product of a past merger. We study two cases. The 1g case uses a spin prior that is uniform-in-$\chi$ for both component spins and a mass ratio prior that favors equal masses \begin{align} p_{1g}(q) \propto \left\{ 1 + \exp{\left[ -k\left( q-q_0 \right) \right]} \right\}^{-1}, \end{align} where $k=20$ and $q_0=0.8$. This prior choice was motivated in~\cite{salvoprior}. In the case of a 2g merger, the primary is expected to be more massive than the secondary binary component. We use a Gaussian mass ratio prior with a mean of $0.5$ and a standard deviation of $0.2$, motivated by Fig. 2 of~\cite{davide2g}. In the 2g case, the primary is also expected to be spinning more rapidly. We therefore use a prior where $\chi_1$ is distributed according to a Gaussian centered at $0.7$ with a width of $0.1$~\cite{davide2g,Fishbach:2017dwv}. The priors of secondary spin magnitude and both spin directions are uniform. \subsection{{\tt LALInference} and {\tt RIFT}} Given a waveform model and a set of prior choices, we compute the joint multidimensional posterior distribution of the source parameters. For fast-to-evaluate waveform models we use the publicly available software library {\tt LALInference}~\cite{Veitch:2014wba,lalinference_o2} to directly sample the posterior distribution. This approach computes the likelihood exactly at various points of the parameter space, but in order to obtain enough independent samples, millions of likelihood evaluation are required. This is prohibitive for models that are slow to evaluate, such as NR. In these cases we use {\tt RIFT}~\cite{2015PhRvD..92b3002P,2018arXiv180510457L}. {\tt RIFT}'s three-stage algorithm first evaluates the likelihood on a dense grid; then approximates it via interpolation; and then uses Monte Carlo integration to produce the full posterior distribution. The number of grid points used for this particular NR-only analysis is 63,000, and the number of added surrogate points for the NR/{\tt NRSur7dq2}{} grid was 40,000; this brought the total number of points for the NR/{\tt NRSur7dq2}{} to 103,000. For context, {\tt RIFT} in general calculates the marginalized likelihood on thousands grid points in parallel. For each marginalized likelihood, which has fixed intrinsic parameters, we evaluate the likelihood at $\approx10^6$ different extrinsic parameters. Even though the number of evaluations are orders of magnitude larger than for {\tt LALInference}, the overall wallclock time is considerably lower because the likelihood evaluations are faster and done in parallel, see \cite{2019arXiv190204934W} for details. However, due to grid limitations (discreteness and limited range), {\tt RIFT} does not sample both extrinsic and intrinsic parameters jointly from the full posterior distribution. Instead it marginalizes over all extrinsic parameters to calculate the likelihood and posterior for just the intrinsic parameters. All results obtained using {\tt LALInference} marginalize over the same detector calibration amplitude and phase uncertainty as in~\cite{LIGOScientific:2018mvr} and publicly available in~\cite{GW170729CalEnv} using the method described in~\cite{TheLIGOScientific:2016wfe,SplineCalMarg-T1400682}. All {\tt RIFT} results assume perfect calibration; this choice was shown to not affect the intrinsic binary parameters~\cite{Vitale:2011wu}. \subsection{{\tt BayesWave}} Finally, we also use a minimal-assumptions analysis that does not make use of CBC-specific waveform models. We use {\tt BayesWave}~\cite{Cornish:2014kda}, a publicly-available algorithm~\cite{bayeswave} that does not explicitly assume that the signal is a CBC\footnote{While {\tt BayesWave} does not assume an explicit signal morphology, it does assume that the signal is elliptically polarized, that it propagates at the speed of light, and that there is no phase decoherence during the propagation.}, and instead models it through a linear combination of basis functions, either sine Gaussian (known as Morlet Gabor) wavelets, or ``chirplets''~\cite{Millhouse:2018dgi}, as listed in the bottom half of Table~\ref{table:pe-table}. The latter are sine Gaussians modified with a linearly evolving frequency. {\tt BayesWave} relies on a transdimensional sampler~\cite{10.1093/biomet/82.4.711} to explore the multidimensional posterior of the parameters of the wavelets/chirplets (frequency, time, phase, amplitude, quality factor, and possibly the frequency derivative) as well as the number of wavelets/chirplets in the linear combination. We then compare the signal reconstruction obtained with the morphology-independent models of {\tt BayesWave} and with CBC waveform models. Broad agreement between the wavelet reconstruction and {\tt IMRPhenomPv2} was established in~\cite{LIGOScientific:2018mvr} and we here perform the same test for waveform models that include higher-order modes. We also quantify the level of consistency through the detector network overlap \cite{Apostolatos:1995pj}, defined as \begin{equation} {\cal{O}}_{N} \equiv \frac{(h_1,h_2)_{N}}{\sqrt{(h_1,h_1)_{N}(h_2,h_2)_{N}}} \end{equation} where $(h_1,h_2)_{N}$ denotes the inner product over the network defined by \begin{equation} (h_1,h_2)_{N} = \sum_{i}^{n} (h_1^i, h_2^i) \end{equation} where $i$ sums over all the detectors in the network, and $(h_1^i,h_2^i)$ is the inner product in an individual detector defined by \begin{equation} (h_1^i,h_2^i)\equiv 4 \Re \int_{0}^{\infty} \frac{\tilde{h}_1^i(f)\tilde{h}^{i*}_2(f)}{S_n^i(f)}df. \end{equation} In the above, $\tilde{h}_1^i(f)$ denotes a signal reconstruction sample computed with CBC models and $\tilde{h}_2^i(f)$ is a reconstruction sample computed with {\tt BayesWave}. Finally, $S_n^i(f)$ denotes the PSD of the detector. The superscipt $i$ denotes the quantities as they appear in the the $i^{\rm{th}}$ detector. \section{CBC-model-based analysis} \label{CBCresults} In this section we present results on inference with CBC waveform models. We study how the posteriors for the various source parameters are affected by the inclusion of spin-precession and higher-order modes in the waveforms. We also study the effect of spin priors and waveform systematics on the validity of our conclusions. \subsection{Higher-order modes} \label{2HM} The importance of higher-order modes on a GW signal observed in the detectors depends on both extrinsic parameters, such as the inclination of the binary, and intrinsic parameters, such as the mass ratio. In general, signals from edge-on and asymmetric binaries include more power in higher-order modes. To study the effect of higher-order modes on GW170729 we analyze the data with waveform models both with and without higher-order modes. Table~\ref{table:par-table} gives the median and 90\% symmetric credible interval and/or HPD (highest probability density interval) for various source parameters from multiple waveform models and the two spin priors. \begin{table*}[] \centering \begin{tabular}{cccccccccccccc} \hline \hline Parameter &$m_{1} (M_{\odot})$ & $m_{2} (M_{\odot})$ & $M (M_{\odot})$ &$q$ &$\chi_{\rm{eff}}$ & $\chi_{\rm{p}}$ & SNR & $D_L(\text{Gpc})$ & $\vert \cos{\theta_{JN}}\vert$\\ \hline {\tt IMRPhenomPv2 ($\chi$)} & $51.0^{+13.9}_{-12.4}$& $31.9^{+9.3}_{-9.6}$& $83.7^{+13.0}_{-12.0}$& $0.62^{+0.36}_{-0.23}$& $0.35^{+0.22}_{-0.23}$& $0.42^{+0.34}_{-0.29}$& $10.7^{+0.4}_{-0.4}$& $2.85^{+1.31}_{-1.28}$& $0.83^{+0.17}_{-0.40}$ \\ {\tt IMRPhenomPv2 ($\chi_z$)} & $52.0^{+15.9}_{-11.6}$& $33.2^{+9.9}_{-9.7}$& $85.8^{+12.4}_{-12.4}$& $0.64^{+0.34}_{-0.25}$& $0.41^{+0.21}_{-0.21}$& $0.58^{+0.29}_{-0.29}$& $10.7^{+0.4}_{-0.4}$& $2.96^{+1.30}_{-1.38}$& $0.84^{+0.16}_{-0.42}$ \\ {\tt IMRPhenomD ($\chi$)} & $50.5^{+13.5}_{-11.2}$& $32.5^{+10.0}_{-8.8}$& $82.8^{+13.4}_{-12.6}$& $0.65^{+0.32}_{-0.23}$& $0.34^{+0.21}_{-0.23}$& -& $10.7^{+0.4}_{-0.4}$& $2.75^{+1.26}_{-1.41}$& $0.80^{+0.20}_{-0.46}$ \\ {\tt IMRPhenomD ($\chi_z$)} & $50.8^{+12.8}_{-11.5}$& $34.5^{+9.7}_{-9.2}$& $85.5^{+13.9}_{-11.4}$& $0.68^{+0.29}_{-0.25}$& $0.42^{+0.19}_{-0.21}$& -& $10.8^{+0.4}_{-0.3}$& $2.90^{+1.32}_{-1.48}$& $0.79^{+0.21}_{-0.46}$ \\ {\tt IMRPhenomHM} ($\chi$) & $57.0^{+12.6}_{-12.0}$& $29.5^{+9.2}_{-9.0}$& $86.0^{+12.2}_{-12.0}$& $0.52^{+0.26}_{-0.21}$& $0.27^{+0.23}_{-0.27}$& -& $11.1^{+0.4}_{-0.4}$& $2.15^{+1.19}_{-1.15}$& $0.70^{+0.30}_{-0.42}$ \\ {\tt IMRPhenomHM} ($\chi_z$) & $58.1^{+12.5}_{-13.4}$& $32.0^{+9.5}_{-8.9}$& $89.7^{+13.4}_{-12.9}$& $0.55^{+0.26}_{-0.21}$& $0.36^{+0.21}_{-0.23}$& -& $11.1^{+0.4}_{-0.4}$& $2.30^{+1.24}_{-1.22}$& $0.68^{+0.32}_{-0.46}$ \\ {\tt SEOBNRv3 ($\chi$)} & $49.5^{+13.2}_{-10.8}$& $35.3^{+8.9}_{-8.6}$& $85.0^{+13.8}_{-12.8}$& $0.72^{+0.28}_{-0.22}$& $0.38^{+0.23}_{-0.22}$& $0.45^{+0.32}_{-0.31}$& $10.8^{+0.4}_{-0.4}$& $2.86^{+1.38}_{-1.38}$& $0.78^{+0.22}_{-0.47}$ \\ {\tt SEOBNRv4 ($\chi$)} & $50.8^{+12.4}_{-11.8}$& $33.4^{+9.6}_{-9.8}$& $83.9^{+13.7}_{-13.3}$& $0.66^{+0.30}_{-0.25}$& $0.35^{+0.20}_{-0.24}$& -& $10.8^{+0.4}_{-0.3}$& $2.77^{+1.30}_{-1.45}$& $0.79^{+0.21}_{-0.46}$ \\ {\tt SEOBNRv4 ($\chi_z$)} & $51.7^{+12.4}_{-12.6}$& $35.2^{+9.4}_{-9.6}$& $86.4^{+14.9}_{-12.7}$& $0.68^{+0.29}_{-0.24}$& $0.42^{+0.21}_{-0.26}$& -& $10.8^{+0.4}_{-0.3}$& $2.97^{+1.46}_{-1.43}$& $0.79^{+0.21}_{-0.47}$ \\ {\tt SEOBNRv4HM} ($\chi$) & $55.2^{+10.2}_{-12.2}$& $29.8^{+9.9}_{-9.5}$& $84.6^{+12.5}_{-11.3}$& $0.54^{+0.31}_{-0.20}$& $0.25^{+0.26}_{-0.26}$& - & $11.0^{+0.4}_{-0.4}$& $2.30^{+1.36}_{-1.17}$ & $0.72^{+0.28}_{-0.39}$ \\ {\tt SEOBNRv4HM} ($\chi_z$) & $54.8^{+11.1}_{-12.1}$& $32.8^{+9.8}_{-9.2}$& $87.5^{+12.2}_{-11.7}$& $0.60^{+0.32}_{-0.21}$& $0.34^{+0.26}_{-0.24}$& - & $11.0^{+0.4}_{-0.4}$& $2.66^{+1.38}_{-1.31}$ & $0.74^{+0.26}_{-0.41}$ \\ \end{tabular} \caption{Parameters of GW170729 obtained with various waveform models and two spin priors, uniform-in-$\chi$ (labelled $\chi$) and uniform-in-$\chi_z$ (labelled $\chi_z$). We quote median values and 90\% credible intervals for the source-frame primary mass, the source-frame secondary mass, the source-frame total mass, the effective spin $\chi_{\rm{eff}}$, and the effective precession parameter $\chi_{\rm{p}}$~\cite{Schmidt:2014iyl}. For the mass ratio we quote the median value and the 90\% HPD. All masses are given in the source frame assuming the cosmological parameters of~\cite{Ade:2015xua} to convert luminosity distance to redshift. The effective precession parameter $\chi_{\rm{p}}$ is absent in the spin-aligned models. } \label{table:par-table} \end{table*} \begin{figure*}[] \includegraphics[width=\columnwidth,clip=true]{corner_mtotal_source_q_chi_eff-IMRP.pdf} \includegraphics[width=\columnwidth,clip=true]{corner_mtotal_source_q_chi_eff-SEOB.pdf} \caption{Corner plots for the posterior densities of the total binary mass in the source frame, the mass ratio, and the effective spin parameter for the different waveform families. We show results obtained with CBC waveforms of the {\tt IMRPhenom} family (left) and the {\tt SEOBNR} family (right). Results from both waveform families suggest that the inclusion of higher-order modes in the models results in evidence for more unequal-mass binaries and less support for nonzero effective spin. The total mass of the binary remains unchanged.} \label{fig:intrinsic_IMREOB} \end{figure*} We start by discussing the binary intrinsic parameters, in particular masses and spins. In Fig.~\ref{fig:intrinsic_IMREOB} we show multi-dimensional corner plots for the posterior densities of the source-frame total mass $M$, the mass ratio $q$ and the effective spin parameter $\chi_{\rm{eff}}$ with waveform models of the {\tt IMRPhenom} (left) and the {\tt SEOBNR} family (right). For these figures and all the figures of this subsection we show results with the uniform-in-$\chi$ spin prior. Within each family, the spin-precessing model is given with a solid line, the spin-aligned model with a dashed line and the spin-aligned model with higher-order modes with a dotted line. In all cases we find that the inclusion of higher-order modes does not have a large effect on the total mass measurement. The mass ratio of the system and the effective spin posteriors are both shifted. In particular we find that waveforms with higher-order modes consistently provide more support for unequal-mass systems and smaller effective spins. For the mass ratio we find that the 90\% HPD interval is $(0.31-0.78) [(0.34-0.85)]$ with higher-order modes and $(0.42-0.97) [(0.41-0.96)]$ without them when using the spin-aligned {\tt IMRPhenomHM} [{\tt SEOBNRv4HM}] and {\tt IMRPhenomD} [{\tt SEOBNRv4}] models respectively. We conclude that GW170729 is not consistent with an equal-mass merger at the 90\% level at least. The combination of unchanged total mass but lower mass ratio means that the primary mass of GW170729 is inferred to be larger than previously measured. Reference~\cite{LIGOScientific:2018jsj} studied the population of the 10 detected BBHs and concluded that no more than 1\% of BHs in BBHs are expected to be above $45 M_{\odot}$. We find that our updated primary mass measurement is not at odds with this conclusion. In particular, we find that the probability that $m_1$ is lower than $45 M_{\odot}$ is 17\% using {\tt IMRPhenomPv2} and reduced to 6\% with {\tt IMRPhenomHM}. Given that we have detected 10 BBHs, it is not unlikely that the true $m_1$ of one of them is at the sixth posterior percentile. A more detailed population analysis in needed to quantify this statement, but this is beyond the scope of this paper. We also find that higher-order modes result in less support for a positive effective spin in GW170729. Reference~\cite{LIGOScientific:2018mvr} reported that for GW170729 $\chi_{\rm{eff}} \sim (0.11- 0.58)$ at the 90\% credible level using combined posterior samples between {\tt IMRPhenomPv2} and {\tt SEOBNRv3}. Interestingly, this credible interval does not include zero, suggesting that at the 90\% level GW170729 has a nonzero effective spin. The inclusion of higher-order modes slightly changes this picture as we now find that the corresponding 90\% credible intervals no longer exclude zero: $\chi_{\rm{eff}} \sim(-0.01-0.49)$ with {\tt IMRPhenomHM}{} and $\chi_{\rm{eff}} \sim(-0.02-0.50)$ with {\tt SEOBNRv4HM}{}. The effective spin parameter is still probably positive with the probability of $\chi_{\rm{eff}}>0$ being $94\%$ with higher-order modes, which is slightly reduced from the corresponding probability of $99\%$ when higher-order modes are not taken into account. Overall higher-order modes cause the 95\% lower limit for $\chi_{\rm{eff}}$ to shift by $-0.10$ for the {\tt IMRPhenom} family, $-0.13$ for the {\tt SEOBNR} family. \begin{figure*}[] \includegraphics[width=\columnwidth,clip=true]{q_systematics.pdf} \includegraphics[width=\columnwidth,clip=true]{chieff_systematics.pdf} \caption{Effect of waveform systematics. We show the mass ratio posterior (left) and the effective spin posterior (right) computed with different CBC waveform models that include higher-order modes (dotted lines) and models that do not include higher-order modes (dashed). Small differences between the posteriors from different waveform approximants are present, but these differences are much smaller than the effect of higher-order modes for both parameters. } \label{fig:systematics} \end{figure*} Despite the broad consistency between results obtained with different waveform families, it is still possible that our results are partly affected by systematic uncertainties in the waveform models. To address this in Fig.~\ref{fig:systematics} we plot again the mass ratio (left) and effective spin (right) posterior densities for the {\tt IMRPhenom} and {\tt SEOBNR} waveform models with (dashed lines) and without (dotted lines) higher-order modes. We find very small differences between both all the dotted and all the dashed lines and in particular between the new waveform models {\tt IMRPhenomHM} and {\tt SEOBNRv4HM} that include higher-order modes. More importantly, we find a clear separation between the dotted lines, i.e. the posteriors that include higher-order modes, and the dashed lines, i.e. the posteriors that do not include higher-order modes. As a further test, in Fig.~\ref{fig:intrinsic_NR} we compare posteriors for $q$ and $\chi_{\rm{eff}}$ computed with {\tt SEOBNRv4HM} (red line) with NR waveforms (magenta lines), as well as with NR/{\tt NRSur7dq2}{} (green lines). Compared to Fig.~\ref{fig:intrinsic_IMREOB} we omit the total mass posterior as {\tt RIFT} did not compute source-frame quantities. In order to perform a fair comparison, all posteriors have been computed with {\tt RIFT}, while for technical reasons we cannot use {\tt IMRPhenomHM} with {\tt RIFT}. We find excellent agreement between NR with higher-order modes, NR/{\tt NRSur7dq2}{} with higher-order modes, and {\tt SEOBNRv4HM}. This shows that {\tt SEOBNRv4HM} is as accurate as NR waveforms in describing GW170729. Moreover, the agreement between {\tt SEOBNRv4HM} and {\tt IMRPhenomHM} in Fig.~\ref{fig:systematics} suggests that the latter {\tt IMRPhenom} waveform is also highly accurate for the event studied here. While these posteriors are broadly consistent with those obtained in Fig.~\ref{fig:systematics}, we find a disagreement in results obtained with {\tt LALInference} and {\tt RIFT} with the same waveform model for the mass ratio at the $7\%$ level. The nature of this difference and further numerical estimates are described in Appendix~\ref{RIFTtroubles}. Overall, Figures~\ref{fig:systematics} and~\ref{fig:intrinsic_NR} suggest that despite minor differences between the waveform models considered here, our main conclusions are robust. \begin{figure}[] \includegraphics[width=\columnwidth,clip=true]{corner_q_chi_eff-NR-NRSur.pdf} \caption{Corner plots for the posterior densities of the mass ratio and the effective spin parameter for the NR waveform family with (magenta lines) and without (green lines) {\tt NRSur7dq2}{} and {\tt SEOBNRv4HM} (red line) computed with {\tt RIFT}. As before, we use dashed (dotted) lines for posteriors with (without) higher order modes. We observe excellent agreement between NR and NR/{\tt NRSur7dq2}{}, confirming the high accuracy of the NR surrogate model. Additionally, we find very good agreement between the NR analysis and {\tt SEOBNRv4HM} when both are used with the same inference code, {\tt RIFT}. } \label{fig:intrinsic_NR} \end{figure} \begin{figure}[] \includegraphics[width=\columnwidth,clip=true]{MatchedFilterSNR.pdf}\\ \includegraphics[width=\columnwidth,clip=true]{theta_jn_distance_log.pdf} \caption{Posterior density for the matched-filter SNR (top) and distance-inclination (bottom) for various waveform models. We find that both within the {\tt IMRPhenom} and the {\tt SEOBNR} waveform families, the waveform model that includes higher-order modes returns the highest value of matched-filter SNR. This higher SNR translates to lower distance and inclination angles closer to 90 degrees. } \label{fig:extrinsic} \end{figure} We now turn to the binary extrinsic parameters. Figure~\ref{fig:extrinsic} shows the matched-filter network signal-to-noise ratio (SNR) posterior density on the top and the two-dimensional posterior density for the luminosity distance and the inclination on the bottom. The network SNR is defined as the square root of the squared sums of the matched-filter SNR in each interferometer, calculated as $\rho = (h,d)/\sqrt{(h,h)}$, where $d$ is the data and $h$ the signal model. The inclination is defined as the angle between the \emph{total} angular momentum vector of the binary, whose direction we treat as fixed, and the line of sight. We present results for waveform models of the {\tt IMRPhenom} (blue) and the {\tt SEOBNR} (red) families. The SNR depends on both the intrinsic loudness of the data and the agreement between the signal and the template $h$. Since the data are common for all analyses, a larger SNR indicates a better agreement/overlap between the data and the template. While we find that within both families spin-precession has a minimal impact on the SNR, waveforms with higher-order modes report slightly larger SNR values, see the top panel of Fig.~\ref{fig:extrinsic}. This suggests that their inclusion leads to a marginally better fit of the data. The data, though, include both the GW170729 signal and a random realization of Gaussian noise, so a better fit of the data does not necessarily imply that waveforms with higher-order modes recover a larger fraction of the GW signal. We quantify the impact of higher-order modes on GW170729 by computing the Bayes Factor in favor of {\tt IMRPhenomHM} compared to {\tt IMRPhenomD}. We find a BF of 5.1:1. While it favors the model with HM, this BF is consistent with the fact that the HM waveforms are able to extract marginally more SNR from the data\footnote{The BF is related to the SNR through $\log \rm{BF} \propto 1/2\, \rm{SNR}^2$. Therefore an SNR increase of $\sim 0.2$ compared to $\sim 10.8$ (see Table~\ref{table:par-table}) would result in a BF of $\sim 8$. This suggests that the measured BF of $5$ is consistent with the SNR increase due to HM.}. Moreover, we emphasize that the BF is not the same as the odds ratio in favor of higher-order modes, which quantifies our degree of belief that higher-order modes are present in the signal. The odds ratio is the BF times the prior odds in favor of the presence of higher-order modes. The latter is formally infinite within GR, as the theory of gravity unequivocally predicts that higher-order modes are present in all CBC signals. The BF presented here only quantifies if higher-order modes are a necessary feature of the models in order to describe the data, and not whether we believe that they exist in general. Regarding the bottom panel of Fig.~\ref{fig:extrinsic}, we find that waveforms with higher-order modes result in less support for face-on/off binary orientations. This observation, coupled to the fact that we see more support for unequal masses and lower spins, see Fig.~\ref{fig:intrinsic_IMREOB}, suggests that higher-order modes lead to more support for sources that are intrinsically of lower amplitude\footnote{We have also verified this by computing the posterior of the intrinsic loudness (defined as the product of the SNR and the distance) with and without higher-order modes. We find that higher-order modes lead to larger probability for intrinsically quieter sources than the quadrupole templates.}. This in turn leads to a posterior distribution for the luminosity distance that is shifted to lower values, as compared to analyses without higher-order modes, as also seen in the bottom panel of Fig~\ref{fig:extrinsic}. Overall, we find that the inclusion of higher-order modes induces small but noticeable shifts in the parameter posteriors. Specifically, the matched-filter SNR increases, the mass ratio posterior obtains more support for unequal masses, and the effective spin parameter is more consistent with lower values; parameter measurements are given in Table~\ref{table:par-table}. The general consistency between the waveform model families we study here shows that our conclusions are robust against waveform systematics. \subsection{Spin prior} \begin{figure*}[] \includegraphics[width=\columnwidth,clip=true]{q_prior_IMR.pdf} \includegraphics[width=\columnwidth,clip=true]{chieff_prior_IMR.pdf}\\ \caption{Effect of spin prior on the mass ratio and effective spin of GW170729. As expected, we find that the mass ratio is minimally affected by the spin prior. The effective spin, however, is shifted to larger values with the uniform-in-$\chi_z$ prior, resulting in increased evidence for nonzero spins. We show results obtained with {\tt IMRPhenom} waveform models, but obtain similar results with {\tt SEOBNR} and NR waveform models as well.} \label{fig:prior} \end{figure*} Besides the CBC waveform models, posterior measurements are also affected by prior choices, in particular the spin prior~\cite{salvoprior}. To test the effect of the spin prior, we reanalyze the data this time assuming a uniform-in-$\chi_z$ prior, where $\chi_z$ is the spin component perpendicular to the orbital plane. The results are presented in Fig.~\ref{fig:prior} for the mass ratio (left) and the effective spin (right) and for waveforms of the {\tt IMRPhenom} family. We have verified that we obtain qualitatively similar results when using {\tt SEOBNR} and NR waveforms models. Due to computational constraints we have only checked results with {\tt SEOBNRv3} and the uniform-in-$\chi$ prior, as computed in~\cite{LIGOScientific:2018mvr}. We find that the spin prior has a minimal effect on the mass ratio posterior. This is expected as the correlation between mass ratio and effective spin is mostly present in the inspiral phase of a CBC. High-mass systems, such as GW170729 are instead dominated by the merger and ringdown in the LIGO sensitive frequency band. In this case little correlation exists between mass ratio and effective spin~\cite{Ng:2018neg}, and changing the spin prior doesnt affect the mass ratio posterior. The effective spin parameter, on the contrary, is directly affected by the choice of the spin prior, and clear differences are visible. The uniform-in-$\chi_z$ prior favors larger spin magnitudes than the uniform-in-$\chi$ prior. As a result, the effective spin posterior is shifted to larger values. The median and 90\% credible interval for the effective spin is $0.41^{+0.21}_{-0.21}$ under the uniform-in-$\chi_z$ prior and $0.35^{+0.22}_{-0.23}$ under the uniform-in-$\chi$ prior using the {\tt IMRPhenomPv2} waveform model. Additional spin measurements for other waveform models are presented in Table~\ref{table:par-table}. \subsection{Second generation merger} \begin{figure*}[] \includegraphics[width=\columnwidth,clip=true]{MassRatio1g2g.pdf} \includegraphics[width=\columnwidth,clip=true]{chieff1g2g.pdf} \caption{Posterior density for the mass ratio (left) and the effective spin (right) under the 1g prior (solid lines) and the 2g prior (dashed lines) for various waveforms. The black lines show the prior density distributions for each parameter. We also show the uniform-in-$\chi$ posterior and prior from Fig.~\ref{fig:prior} for comparison. We find that the 1g-2g prior has a minimal effect on the effective spin parameter, but affects the mass ratio considerably. All 2g runs have strong support for unequal masses, as expected from the prior.} \label{fig:2G} \end{figure*} Finally, we study the possibility of a 2g merger. In that scenario, GW170729 is created in a dense environment such as a nuclear or a globular cluster and its primary mass is the product of a previous merger of two BHs~\cite{davide2g,Fishbach:2017dwv,DiCarlo:2019pmf}. In that case, the mass ratio of the system is expected to be closer to 2:1 (as we find when we use waveform models with HM) and the spin magnitude of the primary should be close to 0.7 (the typical spin of the remnant BH after the merger of two equal-mass, nonspinning BHs)~\cite{Gonzalez:2006md,Berti:2007fi,Buonanno:2007sv,Campanelli:2006gf,Baker:2003ds} \footnote{Though a 2g merger scenario provides a simple way to produce a highly spinning BH, such systems could also be produced in certain astrophysical scenarios, e.g~\cite{Bouffanais:2019nrw}.}. We repeat our analysis with two more priors tailored to the cases of a 1g and a 2g merger scenario. Table~\ref{table:par-table-2g} gives the median and 90\% symmetric credible interval and/or HPD for various source parameters obtained with the 1g and 2g priors. Figure~\ref{fig:2G} shows the effect of this 2g prior on the mass ratio (left) and the effective spin (right) of the binary when using {\tt IMRPhenomD} and {\tt IMRPhenomHM}. We have verified that we obtain similar results with {\tt SEOBNR} and NR waveform models. As expected from the priors we have selected, 2g runs show strong support for unequal masses. This support is even more evident for waveforms with higher-order modes, as anticipated from the results of the previous section. The effective spin parameter is similar, with 1g runs showing more support for nonzero binary components. To further quantify the prior effect, we calculate the Kullback-Leibler (KL) divergence of posterior against prior~\cite{kullback1951}. We find a KL divergence for the mass ratio in the 1g case that is $\sim10$ times smaller than the KL divergence in 2g or default prior scenarios, implying that we do not extract much information by applying the 1g mass ratio prior. For the effective spin, we find that the KL divergence in both priors are comparable as expected because the prior on the effective spin does not change as drastically as the prior on the mass ratio. We also compute the 2g-vs-1g BF. We find 4.7:1 (1.4:1) for waveforms with (without) higher-order modes. We conclude that there is not enough support for the hypothesis that the GW170729 primary needs to be the product of a past merger in order to explain the data's properties. Note that both 2g and 1g models have the same number of parameters, but different distributions in mass ratio and primary spin. The resulting BFs are therefore only affected by how well each model fits the data. \begin{table*}[] \centering \begin{tabular}{cccccccccccccc} \hline \hline Parameter &$m_{1} (M_{\odot})$ & $m_{2} (M_{\odot})$ & $M (M_{\odot})$ &$q$ &$\chi_{\rm{eff}}$ & $\chi_{\rm{p}}$ & SNR & $D_L(\text{Gpc})$ & $\vert \cos{\theta_{JN}}\vert$\\ \hline {\tt IMRPhenomPv2 (1g)} & $43.9^{+8.3}_{-6.7}$& $38.0^{+7.0}_{-6.3}$& $82.1^{+12.9}_{-11.6}$& $0.88^{+0.12}_{-0.14}$& $0.39^{+0.20}_{-0.22}$& $0.47^{+0.31}_{-0.31}$& $10.7^{+0.4}_{-0.4}$& $3.08^{+1.17}_{-1.47}$& $0.83^{+0.17}_{-0.42}$\\ {\tt IMRPhenomPv2 (2g)} & $55.1^{+11.3}_{-12.6}$& $30.2^{+8.0}_{-7.2}$& $85.5^{+12.1}_{-11.2}$& $0.55^{+0.23}_{-0.20}$& $0.37^{+0.18}_{-0.19}$& $0.48^{+0.25}_{-0.29}$& $10.7^{+0.4}_{-0.4}$& $2.83^{+1.19}_{-1.31}$& $0.83^{+0.17}_{-0.40}$\\ {\tt IMRPhenomD (1g)} & $43.8^{+7.9}_{-7.2}$& $38.2^{+6.9}_{-6.4}$& $81.9^{+13.9}_{-11.2}$& $0.88^{+0.12}_{-0.13}$& $0.38^{+0.18}_{-0.20}$& -& $10.8^{+0.4}_{-0.4}$& $2.99^{+1.26}_{-1.43}$& $0.80^{+0.20}_{-0.44}$\\ {\tt IMRPhenomD (2g)} & $53.7^{+11.0}_{-11.6}$& $30.2^{+7.9}_{-7.0}$& $84.0^{+13.3}_{-11.8}$& $0.56^{+0.21}_{-0.19}$& $0.34^{+0.17}_{-0.20}$& -& $10.8^{+0.4}_{-0.3}$& $2.69^{+1.23}_{-1.31}$& $0.79^{+0.21}_{-0.44}$\\ {\tt IMRPhenomHM (1g)} & $45.5^{+10.0}_{-7.9}$& $37.9^{+6.7}_{-6.4}$& $83.6^{+13.3}_{-11.7}$& $0.84^{+0.16}_{-0.16}$& $0.38^{+0.19}_{-0.20}$& -& $10.9^{+0.4}_{-0.5}$& $2.83^{+1.17}_{-1.41}$& $0.80^{+0.20}_{-0.38}$\\ {\tt IMRPhenomHM (2g)} & $58.2^{+10.5}_{-10.1}$& $29.7^{+7.1}_{-7.4}$& $87.6^{+12.7}_{-11.2}$& $0.51^{+0.17}_{-0.16}$& $0.30^{+0.18}_{-0.21}$& -& $11.1^{+0.4}_{-0.4}$& $2.18^{+1.18}_{-1.05}$& $0.69^{+0.31}_{-0.43}$\\ {\tt SEOBNRv4 (1g)} & $44.3^{+8.5}_{-7.2}$& $38.7^{+7.0}_{-6.9}$& $83.0^{+14.2}_{-11.6}$& $0.88^{+0.12}_{-0.14}$& $0.39^{+0.21}_{-0.20}$& -& $10.8^{+0.4}_{-0.4}$& $2.98^{+1.34}_{-1.50}$& $0.79^{+0.21}_{-0.45}$\\ {\tt SEOBNRv4 (2g)} & $53.9^{+11.5}_{-11.1}$& $31.2^{+8.3}_{-7.5}$& $85.1^{+14.3}_{-11.4}$& $0.57^{+0.21}_{-0.20}$& $0.34^{+0.19}_{-0.20}$& -& $10.8^{+0.4}_{-0.4}$& $2.68^{+1.28}_{-1.36}$& $0.78^{+0.22}_{-0.46}$\\ {\tt SEOBNRv4HM (1g)} & $45.1^{+8.4}_{-7.7}$& $38.5^{+7.8}_{-6.3}$& $83.7^{+14.3}_{-12.0}$& $0.87^{+0.13}_{-0.14}$& $0.38^{+0.20}_{-0.19}$& -& $10.9^{+0.4}_{-0.4}$& $2.94^{+1.35}_{-1.43}$& $0.78^{+0.22}_{-0.47}$\\ {\tt SEOBNRv4HM (2g)} & $56.6^{+9.3}_{-9.9}$& $30.4^{+6.2}_{-7.7}$& $86.8^{+11.5}_{-11.0}$& $0.54^{+0.16}_{-0.16}$& $0.29^{+0.19}_{-0.18}$& -& $11.1^{+0.4}_{-0.4}$& $2.43^{+1.21}_{-1.04}$& $0.72^{+0.28}_{-0.42}$\\ \end{tabular} \caption{Parameters of GW170729 obtained with various waveform models and the 1g and 2g priors. We quote median values and 90\% credible intervals for the primary mass, the secondary mass, the total mass, the effective spin $\chi_{\rm{eff}}$, and the effective precession parameter $\chi_{\rm{p}}$. For the mass ratio we quote the median value and the 90\% HPD. All masses are given in the source frame. The effective precession parameter $\chi_{\rm{p}}$ is absent in the spin-aligned models. } \label{table:par-table-2g} \end{table*} \section{Morphology-Independent analysis} \label{BWresults} The studies presented in the previous section relied on specific waveform models for the signal emitted during a CBC as predicted by GR. We here follow a more generic approach and analyze GW170729 in a morphology-independent way that does not explicitly assume it is a CBC described by the currently available waveform models. We use {\tt BayesWave} to reconstruct the signal and then compare this reconstruction to the one obtained with CBC models. The comparison is shown in Fig.~\ref{fig:bayeswave}, where we plot the whitened strain as a function of time. At each time, the shaded region denotes the 90\% credible interval of the reconstruction using CBC waveform models (blue) and {\tt BayesWave} (orange). The left panel is obtained with the wavelets, while the right panel is obtained with chirplets. The top row is made with the CBC waveform model {\tt IMRPhenomPv2} which includes the effect of spin precession, but not higher-order modes. The bottom row uses {\tt IMRPhenomHM} which assumes that the spins remain aligned with the orbital angular momentum, but includes higher-order modes. \begin{figure*}[] \includegraphics[width=\columnwidth,clip=true]{bw_Wavelet_li_IMRPhenomPv2_comp.pdf} \includegraphics[width=\columnwidth,clip=true]{bw_Chirplet_li_IMRPhenomPv2_comp.pdf} \includegraphics[width=\columnwidth,clip=true]{bw_Wavelet_li_IMRPhenomHM_comp.pdf} \includegraphics[width=\columnwidth,clip=true]{bw_Chirplet_li_IMRPhenomHM_comp.pdf} \caption{Signal reconstruction comparison for GW170729. We plot the 90\% credible interval of the whitened strain data as a function of time for each of the LIGO detectors computed with CBC waveform models (blue) and {\tt BayesWave} (orange). The top plots show {\tt IMRPhenomPv2} while the bottom plots show {\tt IMRPhenomHM}. The left plots use the wavelet model of {\tt BayesWave}, while the right plots use the chirplet model. The $\textrm{x}$ axis represents the time in seconds from the nearest integer GPS time before the event. The $\textrm{y}$ axis represents the strain amplitude whitened using a filter which is the inverse Amplitude Spectral Density (ASD) of the noise in the detector. The units are in multiples of the standard deviation of the noise. The generic signal reconstruction is consistent with the CBC signal reconstruction both when the latter includes higher-order modes and when it does not.} \label{fig:bayeswave} \end{figure*} As discussed in~\cite{LIGOScientific:2018mvr}, {\tt BayesWave} can sometimes reconstruct features that are not present in the CBC reconstructions as can be seen in the right hand side panels of Fig.~\ref{fig:bayeswave}, for example around t=0.27s on the right panel. Unlike CBC model waveforms, wavelet-based models are not limited to a physically motivated waveform morphology. As a result, the {\tt BayesWave} sampler can sometimes pick up random coherence between nearby noise samples. However, these outlying wavelets do not point to any potential additional features in the waveform. In fact, they are absent in the 50\% credible intervals of the reconstruction, implying that they have a low significance. Similar outliers were observed in {\tt BayesWave} analyses applied to simulated signals added to real data. We find broad agreement between the CBC reconstruction and the {\tt BayesWave} reconstruction in all cases. In particular, the 90\% credible intervals obtained with the two methods overlap for all waveform models and {\tt BayesWave} basis functions. The agreement suggests that the omission of higher-order modes does not degrade the reconstruction enough to leave a coherent residual detectable by {\tt BayesWave}. This conclusion is in agreement with the results of the previous section, as well as the corresponding reconstruction plot in~\cite{LIGOScientific:2018mvr}. \begin{figure*}[] \includegraphics[width=\columnwidth,clip=true]{imrpv2_imrphm_li_maxl_wavelet_overlaps.pdf} \includegraphics[width=\columnwidth,clip=true]{imrpv2_imrphm_li_maxl_chirplet_overlaps.pdf} \caption{Overlap between the {\tt BayesWave} and the CBC reconstruction for GW170729. We plot the overlap histogram between 1000 random {\tt BayesWave} waveform samples and the ML CBC waveform obtained with {\tt IMRPhenomPv2} (blue) and {\tt IMRPhenomHM} (green). The left panel uses {\tt BayesWave}'s wavelet model, while the right panel uses the chirplet model. The solid and dashed vertical lines represent the overlap of the MBW reconstruction with the ML CBC waveform, with the blue dashed and solid green lines representing the overlaps computed using {\tt IMRPhenomPv2} and {\tt IMRPhenomHM} respectively. As is described in the text, we expect these overlaps to be higher than those computed using individual {\tt BayesWave} samples.} \label{fig:overlap} \end{figure*} In order to make this statement quantitative, we draw 1000 samples from the {\tt BayesWave} posterior and compute their overlap with the maximum likelihood (ML) CBC waveform from the analyses using {\tt IMRPhenomPv2} and {\tt IMRPhenomHM}. The posterior distribution for the overlap is given in Fig.~\ref{fig:overlap} for wavelets (left) and chirplets (right). We find overlaps typically between $0.6-0.9$. This large spread in the overlaps is a result of the inherent flexibility in wavelet-based analyses causing a large variance in the reconstructed signal morphology. Therefore, unlike CBC waveform samples which are motivated by a physical theory, a single {\tt BayesWave} sample is not constrained by physical reconstruction considerations, other than propagation at the speed of light. Instead, the median {\tt BayesWave} waveform (MBW), defined as being the median across the sample waveforms at each time or frequency step, represents a collective estimate across samples, assumed to represent ``the wisdom of the crowd". The MBW is a stable estimate of the reconstruction as it is relatively immune to the stochastic fluctuations in the variable dimensional sampler. Each of the four vertical lines in Fig.~\ref{fig:overlap} represent the overlap values between the MBW and the ML CBC waveform. They are summarized in Table~\ref{table:overlaps} where we also show results obtained using {\tt SEOBNR} waveform models. The ML CBC waveform has in general a higher agreement with the MBW waveform than with each of the individual samples. \begin{table}[] \centering \begin{tabular}{l |c c} \hline &Wavelets&Chirplets\\ \hline {\tt IMRPhenomPv2}&$0.88$&$0.90$\\ {\tt IMRPhenomD}&$0.87$&$0.89$\\ {\tt IMRPhenomHM}&$0.89$&$0.90$\\ {\tt SEOBNRv4}&$0.87$&$0.90$\\ {\tt SEOBNRv4HM}&$0.88$&$0.89$ \end{tabular} \caption{Overlaps between the median {\tt BayesWave} reconstruction and the maximum likelihood CBC waveform with different waveform models.} \label{table:overlaps} \end{table} We find that waveforms both with and without higher-order modes achieve large overlaps with the MBW reconstruction, in the range of $0.87-0.9$, consistent with expectations. In fact, Ref.~\cite{TheLIGOScientific:2016uux} showed that for masses and SNRs typical of GW170729 the expected overlap between simulated signals and their median reconstructions is in the $0.85-0.9$ range at the 1-$\sigma$ level, similar to what we obtain here. The small remaining disagreement between the ML CBC reconstruction and the MBW reconstruction is due to the fact that, unlike modeled analyses, {\tt BayesWave} is only sensitive to excess signal that stands out and above the detector noise. This means that it is less sensitive than CBC analyses in the lower frequencies, $<40$ Hz. We find that if we increase the low frequency cut off in the overlap calculation to $40$ Hz, the overlaps improve by $\sim 0.07$ for each pair of {\tt BayesWave} and CBC waveforms. Besides the good reconstruction, we find that waveforms with higher-order modes lead to similar overlaps with the {\tt BayesWave} reconstruction to waveforms without higher-order modes. We perform the Kolmogorov--Smirnov test for the two overlap distributions for each panel of Fig.~\ref{fig:overlap} and find only 0.048 and 0.017, which implies that both {\tt IMRPhenomPv2} and {\tt IMRPhenomHM} reconstruct the data comparably well. This confirms that GW170729 is consistent with a CBC and that the higher-order modes are not strong enough to lead to a degradation of the signal reconstruction if neglected. \section{Conclusions} \label{conclusions} We analyze the publicly available strain data for GW170729, the highest mass and most distant confirmed GW detection by the LIGO and Virgo detectors. In particular we investigate the effect of higher-order modes and spin priors on the inference of the source parameters. We find that higher-order modes leave small but noticeable effects, while spin priors affect the spin measurements as anticipated. We find that the inclusion of higher-order modes in the models leads to changes in the estimates for the mass ratio, the effective spin, and the SNR. Our updated parameter measurements imply decreased support for equal binary component masses and nonzero effective spin. In particular we conclude that the mass ratio is $(0.3-0.8)$ at the 90\% credible level, a value that excludes equal masses. We also find that the 90\% credible interval for the effective spin parameter has changed from $(0.11- 0.58)$ as reported in~\cite{LIGOScientific:2018mvr} to $(-0.01-0.50)$, which now marginally includes zero. The effective spin parameter still has a 94\% probability of being positive. Consistent with these findings, we compute the BF in favor of the presence of higher-order modes, and find it to be 5.1:1. Moreover, their omission does not dramatically change the measured parameters, which would happen if they were strong \cite{Varma:2016dnf,Bustillo:2015qty,Graff:2015bba}. This conclusion is also consistent with the fact that current matched-filter searches for CBCs have a reduced efficiency toward signals with strong higher harmonics \cite{Harry:2017weg,Capano:2013raa} and this event was indeed observed in both the GstLAL and PyCBC searches. We argue that the observed changes in parameter measurements are not due to systematic errors in the CBC waveform models. We compare results obtained with different waveform models with and without higher-order modes, including waveforms computed with NR. That leads us to believe that both the increased support for unequal masses and the decreased support for nonzero spin are robust conclusions. We emphasize that the fact that the evidence for higher-order modes is weak does not contradict the fact that waveforms with higher-order modes lead to improved parameter measurements. In fact, accurate modeling of relevant physical effects can improve parameter measurements. This can be because said physical effect is present even if it is too weak to unequivocally detect, or because it helps exclude regions of the parameter space for which that effect would be larger than what we observe. Similar shifts in the posteriors (though in the opposite direction for the distance and inclination) where in fact observed in~\cite{Kumar:2018hml} when reanalyzing GW150914 with waveforms that include higher-order modes. We augment the analysis using CBC waveform models with a morphology-independent analysis using {\tt BayesWave}. We find broad agreement between the CBC analysis and the generic analyses regardless of whether the CBC model uses higher-order modes or not. We quantify this conclusion in terms of the overlap between the CBC and the generic reconstruction, which we find to be $\sim 0.9$, consistent with expectations for signals of this mass and SNR~\cite{TheLIGOScientific:2016uux}. Posterior samples from all analyses are available at~\cite{170729SamplesRelease}. \acknowledgements We thank Christopher Berry, Thomas Dent, Tristano DiGirolamo, Bhooshan Gadre, Roland Haas, Nathan Johnson-McDaniel, Riccardo Sturani, and Aaron Zimmerman for helpful comments and suggestions. This research has made use of data, software and/or web tools obtained from the Gravitational Wave Open Science Center (https://www.gw-openscience.org), a service of LIGO Laboratory, the LIGO Scientific Collaboration and the Virgo Collaboration. LIGO is funded by the U.S. National Science Foundation. Virgo is funded by the French Centre National de Recherche Scientifique (CNRS), the Italian Istituto Nazionale della Fisica Nucleare (INFN) and the Dutch Nikhef, with contributions by Polish and Hungarian institutes. The Flatiron Institute is supported by the Simons Foundation. JCB acknowledges support from Australian Research Council Discovery Project DP180103155. CJH, KKYN and SV acknowledges the support of the National Science Foundation and the LIGO Laboratory. LIGO was constructed by the California Institute of Technology and Massachusetts Institute of Technology with funding from the National Science Foundation and operates under cooperative agreement PHY-1764464. MH was supported by Science and Technology Facilities Council (STFC) grant ST/L000962/1 and European Research Council Consolidator Grant 647839. Parts of this research were conducted by the Australian Research Council Centre of Excellence for Gravitational Wave Discovery (OzGrav), through project number CE170100004. ROS and JL are supported by National Science Foundation (NSF) PHY-1707965 and PHY-1607520. PS acknowledges support from the NWO Veni grant no. 680-47-460 and the Science and Technology Facilities Council grant ST/N000633/1. MH acknowledges support from the Swiss National Science Foundation (SNSF) Grant IZCOZ0\_177057. LS was supported by the Young Elite Scientists Sponsorship Program by the China Association for Science and Technology (2018QNRC001), and partially supported by the National Science Foundation of China (11721303), and XDB23010200. The GT authors gratefully acknowledge the NSF for financial support from Grants No. PHY 1806580, PHY 1809572, and TG-PHY120016. Computational resources were provided by XSEDE and the Georgia Tech Cygnus Cluster. The RIT authors gratefully acknowledge the NSF for financial support from Grants No. PHY-1607520, No. PHY-1707946, No. ACI-1550436, No. AST-1516150, No. ACI-1516125, No. PHY-1726215. This work used the Extreme Science and Engineering Discovery Environment (XSEDE) [allocation TG-PHY060027N], which is supported by NSF grant No. ACI-1548562. Computational resources were also provided by the NewHorizons, BlueSky Clusters, and Green Prairies at the Rochester Institute of Technology, which were supported by NSF grants No.\ PHY-0722703, No.\ DMS-0820923, No.\ AST-1028087, No.\ PHY-1229173, and No.\ PHY-1726215. The SXS authors at Caltech acknowledge the Sherman Fairchild Foundation, and NSF grants PHY-1708212 and PHY-1708213. GL, NA, and AG are supported by NSF PHY-1606522, PHY-1654359, PHY-1654359. Computations were done using the orca cluster supported in part by PHY-1429873 and by Cal State Fullerton. AG is also supported in part by Nancy Goodhue-McWilliams. % The authors are grateful for computational resources provided by the LIGO Laboratory and supported by National Science Foundation Grants PHY-0757058 and PHY-0823459. Some of the computational work for this manuscript was also carried out on the computer cluster {\tt Vulcan} at the Max Planck Institute for Gravitational Physics in Potsdam. Plots have been made with {\tt matplotlib}~\cite{Hunter:2007} and {\tt corner}~\cite{corner}.

Background {#Sec1} ========== Human Immunodeficiency Virus (HIV) continues to be a major world public health problem, having claimed more than 35 million lives so far. In 2017, approximately 940,000 people died from HIV-related causes globally \[[@CR1]\]. Sub-Saharan Africa is the most affected region, with 25.8 million people living with HIV in 2017. It accounts for two-thirds of the global total of new HIV infections \[[@CR2]--[@CR4]\]. Tanzania has about 1.4 million people living with HIV, with adult prevalence of 4.7%. In 2015, 54,000 people were newly infected with HIV mainly between the age group between 15 and 49 years, and 36,000 people died from an AIDS-related illness in the same year \[[@CR5], [@CR6]\]. Of all people living with HIV at the moment globally, it is estimated that only 60% know their status. The remaining 40% or over 14 million people need to access HIV testing services \[[@CR7]\]. The World Health Organization (WHO), and Tanzanian National Guideline for Management of HIV and AIDS recommend routine HIV testing in all health facilities \[[@CR8], [@CR9]\]. Through provider initiated testing and counselling, people living with HIV might be identified early and can then be enrolled into the management system \[[@CR10]\]. Routine HIV screening testing at the emergency department (ED) provides an opportunity for early recognition of HIV positive patients \[[@CR11], [@CR12]\]. Identifying them early will provide the possibility of initiating treatment early so as to stop the progression to AIDS. Further more, the identification of HIV negative individuals will provide an opportunity of linkage to HIV prevention services. Studies in other countries have shown that the routine HIV testing at the ED is feasible with a very good acceptance rate and with good yield \[[@CR12], [@CR13]\]. Those diagnosed HIV positive have been enrolled in ongoing follow-up for their continuous care \[[@CR14]\]. However, as a emergency medicine is very new field in Tanzania, there has been no study done in Tanzania looking the feasibility of routine HIV testing and its yield in emergency departments. The first full capacity Emergency Department was opened in 2010 in at the main referral hospital in Dar es salaam. In this ED, HIV testing occurs on a provider-initiated basis, but it is generally performed only on medically ill patients for the diagnostic purpose, and not as part of the routine ED Screening. The study aimed to determine the feasibility and yield of HIV screening among otherwise healthy adult trauma patients as these trauma patients present non-medically ill individuals who would otherwise not have attended hospitals and would therefore not have opportunity to have an ED provider initiated testing. With the goal towards assessing the possibility of creating an ED protocol that would allow every patient attending the ED in Tanzania to be counseled for taking the HIV test. Methods {#Sec2} ======= Study design {#Sec3} ------------ This was a prospective cohort study of adult trauma patients presenting to the Emergency Medicine Department of Muhimbili National Hospital from March 2017 to September 2017. Study area {#Sec4} ---------- Muhimbili National Hospital (MNH) is a tertiary referral hospital located in Dar es salaam, Tanzania. The hospital has a bed capacity of 1500 with around 1000 to 1200 admissions per week, operating 24 h in 7 days of the week. The MNH-EMD is the first public ED in the country and was opened in 2010. It is staffed with emergency physicians, postgraduate students in emergency medicine training program, medical officers, critical care nurses and nursing officers. The MNH-EMD is the first entry point to the hospital for most of the patients attending MNH, and sees around 200 patients daily. Acutely ill patients are received at ED, resuscitated and stabilized before being disposed to the appropriate ward. Currently HIV testing at the ED is based on physician discretion and is usually done on only sick medical patients. Participants {#Sec5} ------------ All adult (≥ 18 years old) trauma patients attending the EMD-MNH were eligible for the study. Patients were excluded if they had altered mental status, could not speak English or Kiswahili, those who are known to have HIV infection as evidenced by an HIV treatment card, pregnant women already attending Antenatal clinics, and hemodynamic instability---as defined by the treating physician. Study procedures {#Sec6} ---------------- The research assistant frequently and actively went through the entire department (resuscitation rooms and low acuity areas) during 12 h of day clinical shifts of week days to identify trauma patients with the help of nurses in the respective area. The patient registry was also reviewed to identify trauma patients present the department. All adult trauma patients meeting the inclusion criteria were offered a chance to participate in the study after they had received initial treatment and stabilization. The research assistant explained the study, and patients reviewed and signed a consent form to be involved in the study (Consent for the study did not mean consent for testing.) Participants then received pre- testing HIV counseling provided by an EMD nurse who had training in HIV counseling and testing. Participants were informed that no extra charges would be incurred for the test, and that their treatment plan will not be affected by their decision to be tested. Patient privacy and confidentiality were carefully observed before and after testing regardless of the test results. Structured case report forms were completed, documenting demographics, acceptance of testing, results and readiness to receive results. Testing {#Sec7} ------- Testing was done as per the 2016 National Guidelines for HIV and AIDS Management in Tanzania. Finger prick blood sample was collected then tested in the EMD for HIV using rapid immunochromatographic assay (*SD*-*BIOLINE HIV*-*1/2 Alere, Abbott Laboratories, Illinois USA*) as the first testing tool. Those with a negative SD BIOLINE test were considered free of HIV. If the *SD*-*BIOLINE HIV*-*1/2* test positive, a second gold standard immunochromatographic (*Uni*-*Gold™ HIV, Trinity Biotech Plc, Ireland*) test was used as the confirmatory test. Those with *Uni*-*Gold HIV* test positive were considered to have HIV infection. A negative *Uni*-*Gold HIV* test was considered inconclusive; in this case the blood sample was taken to the central pathology laboratory for HIV antibody and antigen testing and results were received the same day. Post testing counseling {#Sec8} ----------------------- Post testing counseling was done before test results were provided. Patients were asked about their readiness to receive the test result. For those who were ready to receive test results, the results were provided. Those who were not ready to receive the test results were counseled to come back to the department at any moment that they felt ready to receive the test results, and were also told they can always go to the nearby health facility of their choosing to be tested for HIV again whenever they are ready. Those who were willing to receive their results and were positive for HIV were provided a special referral form and instructed to attend Counseling and Treatment Centers (CTC) and other HIV treatment centers for further assessment and care after their discharge from the hospital. Provider-selected testing {#Sec9} ------------------------- For each enrolled patient, the Wellsoft electronic health record (Wellsoft Corporation, Somerset, USA) was reviewed to determine if the treating physician had ordered an HIV test. The proportion of individuals tested for HIV, and positive cases detected by the EMD physician, were compared with that detected by routine testing during the study. Feasibility {#Sec10} ----------- Feasibility was assessed by noting any problems with availability of the test, difficulty of administering the test or assessing results, or any adverse events. Follow up {#Sec11} --------- HIV positive patients were contacted by telephone 2 months after their discharge to ask if they had made a visit to CTC as they were instructed. Data analysis {#Sec12} ------------- Data from *REDCap* (*Version 7.2.2, Vanderbilt, Nashville, TN, USA*) was exported to Statistical Package for Social Science (SPSS version 22.0 IBM Ltd, Carolina, USA) for analysis. Descriptive statistics; median, proportion, interquartile range \[IQR\] and counts are reported. Proportion of trauma patients accepting HIV testing was calculated as enrolled adult trauma patients who accepted testing divided by the total number of adult trauma patients enrolled in the study. The yield was the proportion positive tests among those who accepted testing and was calculated by dividing the adult trauma patients who tested positive for HIV by the total number of adult trauma patients who accepted testing. Results {#Sec13} ======= During the study period 2848 trauma patients presented to the EMD, among them 326 (11.45%) met the inclusion criteria, gave written consent and hence enrolled in the study (Fig. [1](#Fig1){ref-type="fig"}).Fig. 1Study Enrollment Demographic characteristics {#Sec14} --------------------------- A total of 326 adult trauma patients were enrolled, Median age was 33 (IQR 25--42 years) and 248 (76%) were male. Total of 136 (48%) were referred from other hospitals to MNH. Approximately half of patients (160) were married, 173 (53%) were unemployed and 224 (69%) had no formal education beyond primary school Table [1](#Tab1){ref-type="table"}.Table 1Demographic characteristics of trauma patients approached for an opt out HIV screening attending at EMD of MNHVariableFrequency, N (%)Sex Male248 (76%) Female78 (24%)Age groups Median age (IQR)33 (25--42)  18--31145 (44%)  32--41100 (31%)  42--5351 (16%)  54--6619 (6%)  \> 6611 (3%)Referral status Referred136 (48%) Not referred190 (58%)Marital status Married160 (49%) Single154 (47% Divorced5 (2%) Widowed7 (2%)Occupation Public or private servant102 (31%) Unemployed173 (53%) Petty trader32 (10%) Student15 (5%) Others4 (1%)Education level Primary education224 (69%) Secondary education73 (22%) College/University education21 (6%) No formal education8 (3%) Acceptance and readiness to receive test results {#Sec15} ------------------------------------------------ Of 326 patients enrolled and counseled for HIV testing, 250 patients (77%) accepted testing; 69% of all males accepted testing while all females (100%) accepted testing. Married and single patients were similar in their willingness to be tested. Acceptance was high (79%) among those with primary education. Among the 250 patients tested, 247 (98.8%) were ready to receive their test results. One male and two females were not ready to receive their HIV test results; all 3 were single, and had not gone beyond primary school. All three had been referred from other hospitals to MNH (Tables [2](#Tab2){ref-type="table"} and [3](#Tab3){ref-type="table"}).Table 2Distribution by gender, age, marital status and referral status of trauma patients who accepted HIV screening and were ready to receive the test resultsVariableNumberAccepted Testing for HIVReady to receive test resultsTest results positiveOverallNn/N (%)n/N (%)n/N (%)Total346250 (77.0%)247 (98.8%)14 (5.6%)Sex Male248172 (69.0%)171 (99.0%)9 (5.2%) Female7878 (100%)76 (97.0%)5 (6.4%)Age group 18--31145109 (76%)107 (98%)5 (4.6%) 32--4110078 (78%)77 (99%)5 (6.4%) 42--535141 (80%41 (100%)1 (2.4%) 54--661915 (79%)15 (100%)2 (13.3%) \> 66117 (64%)7 (100%)1 (14.3%)Marital status Married160127 (79%)127 (100%)5 (3.9%) Single154115 (75%)112 (97%)8 (7.0%) Divorced53 (60%)3 (100%)0 Widowed75 (71%)5 (100%)1 (20.0%)Referral status Referred190152 (80%)149 (98%)12 (7.9%) Not referred13698 (72%)98 (100%)2 (2.0%) Table 3The distribution by occupation and education level of trauma patients who accepted the HIV screening and were ready to receive the test resultsNumberAccepted testing for HIVReady to receive test resultsTest results positive*Occupation*Public or private servant10277 (75%)75 (97%)6 (7.8%)Unemployed173137 (79%)136 (99%)6 (4.4%)Petty trader3224 (75%)24 (100%)2 (8.3%)Student158 (53%)8 (100%)0Others44 (100%)4 (100%)0*Education level*Primary education224176 (79%)173 (98%)14 (8.1%)Secondary education7357 (78%)57 (100%)0College/university education2112 (57%)12 (100%)0No formal education85 (63%)5 (100%)0 Yield {#Sec16} ----- Out of 250 patients who agreed to be tested, 14 (5.6%) tested positive for HIV 9 (5.23%) were male and 5 (6.41%) were female (Fig. [1](#Fig1){ref-type="fig"}, Tables [2](#Tab2){ref-type="table"} and [3](#Tab3){ref-type="table"}). Self-reported CTC attendance {#Sec17} ---------------------------- Among the 12 HIV positive patients that received results, 6 patients (50%) reported having made visits to CTC and were on ARVs; 3 patients (25%) reported to have not gone to CTC. The remaining 3 (25%) were could not reached; either their phone number were not reachable, or the person answering stated it was the wrong number (Fig. [1](#Fig1){ref-type="fig"}). HIV testing as part of regular ED care {#Sec18} -------------------------------------- Of 250 patients tested for HIV during the study, none were tested for HIV as part of their regular care in the emergency department. Therefore, all 14 HIV positive patients would not have been identified by the physician who was providing care for these trauma patients. Feasibility {#Sec19} ----------- During the entire period of study, all patients who agreed to be tested received the test as per National guidelines. The testing kits were available whenever needed, results were provided on time and there were no adverse events throughout study period. Discussion {#Sec20} ========== In this study we found that routine HIV testing in an emergency department in Tanzania is feasible and acceptable, majority of patients in the study were willingly tested and ready to receive their results. We found an overall HIV positivity rate of 5.6% among patients who participated in the study. The majority of enrolled patients were young. This was due to the fact that the study was confined to trauma patients. Younger age people are at greater risk of having trauma because of their daily activities \[[@CR15], [@CR16]\]. Previous studies have also shown that younger age groups are also more likely to have HIV \[[@CR13], [@CR17]\]. However, we chose trauma patients for this study primarily because we wanted to test the feasibility and acceptability of a routine HIV testing in an otherwise healthy population and compare the results with what providers in our department normally do regarding HIV testing in such patients. We found that routine HIV screening at the EMD for diagnosing people living with HIV, but unaware of their status is feasible, as evidenced by more than 75% of those counseled agreeing to be tested. Mass education and training has helped to raise the awareness of HIV and AIDS in Tanzania and that likely contributed to a good acceptance rate. While a previous emergency department study in the US have shown a high overall acceptance rate as in our study \[[@CR23]\], this is not always the case. A study done in Singapore showed a very low acceptance rate which was 21% \[[@CR18]\]. All female trauma patients approached for routine HIV testing accepted; similar studies done in different health care settings in different parts of the world have shown that the acceptance rate tends to be higher in women \[[@CR17]\]. In our study, the unanimous acceptance rate by women may be due to the fact that women have more visits to health facilities, including antenatal and postnatal clinics where HIV education is commonly provided. Almost all trauma patients who accepted testing were ready to receive their test results. This is similar to studies in other sub-Saharan countries: In Uganda, only 4% of patients did not appear for their HIV test result \[[@CR19]\]. Another study from Zambia showed that all patients who accepted HIV testing were ready to receive their test results \[[@CR17]\]. A study done in Uganda showed that of all those who agreed to be tested, only one person did not return for test results \[[@CR14]\]. Even among those that tested positive, we found that more than 85% of those tested positive for HIV were ready to receive their test results. Similarly, studies done in Uganda did not show any difference in readiness to receive test results among those who tested positive for HIV and those who tested negative \[[@CR19], [@CR20]\]. This readiness found in all these studies, including ours, was probably due to proper pre- and post-test counseling. Of those patients that accepted HIV screening, almost 6% tested positive which is similar to HIV prevalence in the country; the Tanzanian National Guidelines for HIV and AIDS management reports the HIV prevalence to be 4.8% \[[@CR21]\]. The yield was much higher than in similar studies done in emergency departments in USA \[[@CR13], [@CR22]\]. Another study done in Singapore showed the yield to be much lower than in the this study (0.18%) \[[@CR18]\]. And a study in the USA found that routine testing identified 1.9% new cases of HIV \[[@CR23]\]. In our study, the proportion of positive HIV tests was slightly higher nationally the prevalence of HIV infection in Tanzania in women is 6.3% and 3.9% for men \[[@CR8]\]. The higher prevalence of HIV in females is thought to be due to biological differences, which increase the chance of females acquiring the virus. This study did not show a very large difference between males and females, but not all males agreed to be tested. One of the important goals for ED HIV screening is to link newly HIV diagnosed patients to specialized HIV centers (CTC) for continuous care. In this study, 50% of those who were HIV positive and received their results went to CTC, and at follow up reported they were already on treatment. Three others who were reached (25% of positive results) did not go to CTC. This may be due to the fact that the study enrolled trauma patients and they might not have recovered fully, preventing them from attending CTC. However, similar findings were reported in the study done in United States; only 56% of newly HIV diagnosed patients attended the specialized centers for HIV \[[@CR13]\]. Other studies in similar settings in Uganda, Zambia and in Singapore have shown that most of the patients who have tested positive in routine HIV or PITC adhered to instructions to enroll in a clinic \[[@CR14], [@CR17], [@CR18]\]. In our department, most HIV screening is done to assist with diagnosis of a medical disease, and so non-suspicious patients rarely get tested. In this study we found that none of the patients that were included in the study would have been tested by the treating physician and so none of the new cases would have been identified. A study of routine HIV testing done in Emergency Department in USA showed that out 9 patients diagnosed with HIV, only 2 of them would have been diagnosed by the physician \[[@CR13]\]. In another study conducted in USA most patients with newly diagnosed cases of HIV identified in the ED through routine testing were not admitted to the hospital, and thus would not be identified in a regular hospital care \[[@CR23]\]. Limitations {#Sec21} ----------- This was a single center study, which reduces its generalizability. The study excluded children, pregnant women, unconscious patients and those with hemodynamic instability. Three of the positive patients were could not be reached, and three did not receive result limiting the ability to determine compliance with follow up care. Conclusion {#Sec22} ========== Routine HIV screening among adult trauma patients being seen in an ED in Tanzania is feasible, acceptable and helps to identify new cases of HIV positive patients who can then enroll for continuous HIV and AIDS care. Future studies should focus on factors that can increase acceptance of testing, and improve the linkage to CTC among HIV positive patients identified at the ED. Conceptualization: JR HS SK EW. Data curation: JR HS SK. Formal analysis: JR HS SK. Funding acquisition: JR. Methodology: JR HS SK EW. Project administration: JR. Supervision: HS SK. Validation: JR HS SK. Writing-original draft: JR. Writing review and editing: JR HS SK EW. All authors read and approved the final manuscript. Acknowledgements {#FPar1} ================ The author would like to thank Dr. Brittany L. Murray, our research assistant, and study participants for making this project a success. Competing interests {#FPar2} =================== The authors declare that they have no competing interests. Availability of data and materials {#FPar3} ================================== The dataset supporting the conclusions of this article is available from the authors on request. Consent to publication {#FPar4} ====================== Not applicable. Ethics approval and consent to participate {#FPar5} ========================================== This study was conducted after approval from the MUHAS IRB and permission from MNH. All patients consented to participate in the study. Funding {#FPar6} ======= This was a non-funded project; the principal investigators used their own funds to support the data collection and logistics. Publisher's Note {#FPar7} ================ Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

1. Introduction {#sec1-molecules-23-02805} =============== Transition metal complexes composed of multidentate ligands based upon tellurium are far less explored than those of sulfur and selenium, which might be due to the lack of easily available ligands containing tellurium, as well as the historical misconception that organotellurium compounds were extremely malodorous and toxic \[[@B1-molecules-23-02805],[@B2-molecules-23-02805],[@B3-molecules-23-02805],[@B4-molecules-23-02805],[@B5-molecules-23-02805],[@B6-molecules-23-02805],[@B7-molecules-23-02805],[@B8-molecules-23-02805]\]. Up to the 1970s, studies of coordination chemistry were mostly restricted to monodentate ligands. The first bidentate telluroether ligand **I** (Te, P-type) was described by Gysling and Luss in 1984 \[[@B9-molecules-23-02805]\], whereas the first tridentate telluroethers **II** (N,Te,N-type), and **III** (P,Te,P-type) were reported by Singh et al. \[[@B10-molecules-23-02805]\] as well as Lin and Gabbaï ([Scheme 1](#molecules-23-02805-sch001){ref-type="scheme"}) \[[@B11-molecules-23-02805]\]. In preceding work, we reported on the synthesis of bis(6-diphenylphosphino­acenaphth-5-yl)telluride (**IV**) \[[@B12-molecules-23-02805]\], which also holds potential as a tridentate telluroether ligand. Herein, we describe the reaction of **IV** with (CO)~5~MnBr and (CO)~5~ReCl, giving rise to the formation of three octahedral transition metal carbonyl complexes with this ligand. 2. Results and Discussion {#sec2-molecules-23-02805} ========================= The reaction of bis(6-diphenylphosphinoacenaphth-5-yl)telluride (6-Ph~2~P-Ace-5-)~2~Te (**IV**) with (CO)~5~ReCl in THF at 65° for 7 days afforded an 1:1 isomeric mixture of the complex *fac/mer*-(6-Ph~2~P-Ace-5-)~2~TeRe(Cl)(CO)~3~ (*fac*-**1/***mer*-**1**), which slowly interconvert in solution. The progress of the reaction was followed by ^31^P NMR spectroscopy. Upon cooling to r.t. solely *fac*-**1** precipitates from the reaction mixture and was isolated in 45% isolated yield ([Scheme 2](#molecules-23-02805-sch002){ref-type="scheme"}). The analogous reaction of **IV** with (CO)~5~MnBr in THF at 65° occurred within 32 h and gave rise to the isomeric mixture of *fac*/*mer*-(6-Ph~2~P-Ace-5-)~2~TeMe(Br)(CO)~3~ (*fac*-**2**/*mer*-**2**) and as was already observed for the Te-Re complex *fac*-**1**, *fac*-**2** precipitates solely from the reaction mixture upon cooling to r.t. and was isolated in 47% yield ([Scheme 2](#molecules-23-02805-sch002){ref-type="scheme"}). In both complexes *fac*-**1** and *fac*-**2**, the telluroether **IV** serves as bidentate ligand. In the solid-state, *fac*-**1** and *fac*-**2** are reasonably stable toward moist air and show no signs of decomposition even after prolonged times of storage. Unfortunately, both complexes show only poor solubility in the most common solvents used for NMR spectroscopy. In CDCl~3~ solution, both *fac*-**1** and *fac*-**2** are in equilibrium with their related *meridional* complexes *mer*-**1** and *mer*-**2**, which can be inferred by ^31^P-NMR spectroscopy ([Scheme 3](#molecules-23-02805-sch003){ref-type="scheme"}). The ^31^P-NMR spectrum of a freshly prepared solution of *fac*-**1** in CDCl~3~, shows almost exclusively two chemical shifts at δ = −0.9 and −25.0 ppm of equal intensity, which are assigned to the P atom that coordinates to the Re atom and the P atom that engages in interaction with the Te atom (the assignment is based upon the comparison with *fac*-\[Re(Cl)(CO)~3~(PPh~2~C~10~H~6~PPh~2~)\] \[[@B13-molecules-23-02805]\], (\[(Ph~2~P(Me~2~pz)~2~)Re(CO)~3~Br\] and \[(Ph~2~P(Me~2~pz))Re(CO)~4~Br (pz = pyrazole) \[[@B14-molecules-23-02805]\]). However, the formation of *mer*-**1** can already be detected even from the freshly prepared solution of *fac*-**1**, showing chemical shifts at δ = 14.7 and −26.4 ppm, which increase in intensity over time ([Figure 1](#molecules-23-02805-f001){ref-type="fig"}). It should be noted that the chemical shifts of *fac*-**1** consist of singlets, whereas the doublets were observed for *mer*-**1** with a coupling constant of *J*(^31^P-^31^P) = 11.6 Hz. In contrast, the Te--Mn complex **2** shows significantly faster formation of *mer*-**2** (δ (^31^P) = 50.0 and −28.0 ppm) from a freshly prepared solution of *fac*-**1** (δ (^31^P) = 36.2 and −25.6 ppm, [Figure S2](#app1-molecules-23-02805){ref-type="app"}d). When the reaction of **IV** with (CO)~5~MnBr was repeated with a longer reaction time of 7 days, a different product, namely (6-Ph~2~P-Ace-5-)~2~TeMn(Br)(CO)~2~ (**3**) formed, which was obtained in 40% isolated yield ([Scheme 2](#molecules-23-02805-sch002){ref-type="scheme"}). The formation of **3** can be rationalized by the liberation of CO from *fac*-**2** and/or *mer*-**2**. In **3**, the telluroether serves as tridentate ligand. Notably, a similar reactivity of *fac*-**1** was not observed. The solubility of **3** in the most common solvents is slightly higher than those of *fac*-**1** and *fac*-**2**, which allowed the acquisition of full set of NMR data. The ^31^P-NMR spectrum (CDCl~3~) of **3** shows a broad signal at *δ* = 59.7 ppm (ω~1/2~ = 110 Hz), whereas the ^125^Te-NMR spectrum exhibits a singlet at *δ* = 753.1 pm that is slightly high-field shifted in comparison to **IV** (*δ* = 704.4 ppm) \[[@B12-molecules-23-02805]\]. The molecular structures of *fac*-**1**, *fac*-**2**, and **3** are shown in [Figure 2](#molecules-23-02805-f002){ref-type="fig"}, [Figure 3](#molecules-23-02805-f003){ref-type="fig"} and [Figure 4](#molecules-23-02805-f004){ref-type="fig"}. Selected bond parameters are collected in [Table 1](#molecules-23-02805-t001){ref-type="table"}. The spatial arrangement of the Re and Mn atoms is octahedral. In *fac*-**1** and *fac*-**2**, only one P atom coordinates to the Re and Mn atoms, whereas the other P atoms engage in a through-space interactions with the Te atoms. Consistent with -the covalent radii of Mn and Re, the Mn--Te bond length of *fac*-**2** (2.599(1) Å) and **3** (2.546(1) Å) are smaller than the Re--Te bond length of *fac*-**1** (2.733(1) Å). The Mn--Te distances are in between values reported for \[Mn(CO)~4~TePh\]~2~ (mean 2.674(2) Å) \[[@B15-molecules-23-02805],[@B16-molecules-23-02805],[@B17-molecules-23-02805]\] and \[Cp′(CO)~2~Mn\]~2~TeMes^+^ (2.44 Å) \[[@B18-molecules-23-02805]\]. Likewise, the Mn--P bond length of *fac*-**2** (2.343(1) Å) and **3** (2.341(1) Å) is smaller than the Re--P bond length of **1** (2.466(1) Å). The average Mn--P distances are comparable with those of \[Mn(Cl)(CO)~3~(κ^1^-PPh~2~C~10~H~6~PPh~2~)\] (2.328(1) Å) \[[@B13-molecules-23-02805]\] and *^i^*^Pr^PN^H^PMn(CO)~2~Br (2.263 (1) Å) \[[@B19-molecules-23-02805]\]. The Re(1)--C(5)/C(6) bond distances of *fac*-**1** (2 × 1.930(2) Å) and the Mn(1)--C(5)/C(6) bond distances of *fac*-**2** (1.817(6), 1.819(5) Å) are very similar to those of \[ReCl(CO)~3~(PPh~2~C~10~H~6~PPh~2~)\] (mean 1.952(16) Å)\[[@B13-molecules-23-02805]\] and *fac*-\[MnCl(CO)~3~(PPh~2~C~10~H~6~PPh~2~)\] (mean 1.815(2) Å) \[[@B20-molecules-23-02805]\]. Due to the *trans*-effect of the P atom, the Re(1)--C(7) bond distance of *fac*-**1** (1.951(2) Å) and the Mn(1)--C(7) distance of *fac*-**2** (1.915(7) Å) are slightly longer. The Mn(1)--C(5) and Mn(1)--C(6) bond lengths of **3** (1.800(4), 1.811(5) Å) are in average very similar than that of *fac*-\[Mn(CO)~3~(dppbz)Br\] (mean 1.807(4) Å; dppbz = 1,2-(diphenylphosphino)benzene) \[[@B21-molecules-23-02805]\]. In the carbonyl region of the IR spectrum of *fac*-**1** three strong absorption bands at $\widetilde{v}$ = 2018, 1928, and 1882 cm^−1^ are visible, respectively, consistent with a facial arrangement of three carbonyl groups \[[@B12-molecules-23-02805],[@B22-molecules-23-02805]\]. The IR spectrum of **3** reveals asymmetric and symmetric stretching vibrations at $\widetilde{v}$ = 1936 and 1852 cm^−1^, which compare well with those of (^i^PrPONOP)Mn(CO)~2~Br (1866, 1946 cm^−1^) and ^i^PrPNHPMn(CO)~2~Br (1824, 1916 cm^−1^) \[[@B19-molecules-23-02805]\]. 3. Conclusions {#sec3-molecules-23-02805} ============== The present study demonstrates the ability of bis(6-diphenylphosphinoacenaphthyl-5-)telluride (6-Ph~2~P-Ace-5-)~2~Te (**IV**) \[[@B12-molecules-23-02805]\] to act as a bidentate and tridentate ligand toward transition metals carbonyls. The reaction of **IV** with (CO)~5~ReCl and (CO)~5~MnBr provided three complexes, namely *fac*-(6-Ph~2~P-Ace-5-)~2~TeM(X)(CO)~3~ (*fac*-**1**: M = Re, X = Cl; *fac*-**2**: M = Mn, X = Br) and (6-Ph~2~P-Ace-5-)~2~TeMn(Br)(CO)~2~ (**3**). In solution, the facial complexes, *fac*-**1** and *fac*-**2**, are engaged in a reversible equilibrium with their meridional complexes, *mer*-(6-Ph~2~P-Ace-5-)~2~TeM(X)(CO)~3~ (*mer*-**1**: M = Re, X = Cl; *mer*-**2**: M = Mn, X = Br). 4. Experimental Section {#sec4-molecules-23-02805} ======================= *General.* Reagents were obtained commercially (Sigma-Aldrich, Taufkirchen, Germany) and were used as received. Dry solvents were collected from an SPS800 mBraun solvent system. Bis(6-diphenylacenaphth-5-yl)telluride, (6-Ph~2~P-Ace-5-)~2~Te (**IV**) was prepared according to literature procedures \[[@B12-molecules-23-02805]\]. ^1^H-, ^13^C-, ^31^P-, and ^125^Te-NMR spectra were recorded at room temperature using a Bruker Avance-360 and a Bruker Avance-200 spectrometer and are referenced to tetramethylsilane (^1^H, ^13^C), phosphoric acid (85% in water) (^31^P), and dimethyltelluride (^125^Te). Chemical shifts are reported in parts per million (ppm), and coupling constants (*J*) are given in Hertz (Hz). The FTIR spectra were recorded on Thermo Scientific Nicolet^TM^ iS10. The ESI-MS spectra were obtained with a Bruker Esquire-LC MS. Dichloromethane/acetonitrile solutions (c = 1 × 10^−6^ mol L^−1^) were injected directly into the spectrometer at a flow rate of 3 µL min^−1^. Nitrogen was used both as a drying gas and for nebulization with flow rates of approximately 5 L min^−1^ and a pressure of 5 psi. Pressure in the mass analyzer region was usually about 1 × 10^−5^ mbar. Spectra were collected for 1 min and averaged. The nozzle-skimmer voltage was adjusted individually for each measurement. *Synthesis of (6-Ph~2~P-Ace-5-)~2~TeRe(CO)~3~Cl* (*fac*-**1**): Re(CO)~5~Cl (9.20 mg, 25.4 µmol) was added to a solution of (6-Ph~2~P-Ace-5-)~2~Te (20.0 mg, 24.9 μmol) in THF (5 mL). The mixture was continually stirred at 65 °C for 7 d. The product precipitated upon cooling to room temperature as colorless microcrystals. (12.5 mg, 11.3 µmol, 45%; Mp. 197 °C (dec.)). Due to the low-solubility and the equilibrium of *fac*-**1** and *mer*-**1** in solution, reliable ^13^C- and ^125^Te-NMR spectra could not be obtained. *fac*-**1**: ^1^H-NMR (200.1 MHz, CDCl~3~) *δ* = 7.99 (d, ^3^*J* = 7.9 Hz, 1H), 7.42 (m, 18H), 7.18 (dd, *J* = 7.9, 3.4 Hz, 4H), 6.71 (d, ^3^*J* = 7.4 Hz, 1H), 6.22 (d, ^3^*J* = 7.5 Hz, 1H), 3.50 (s, 4H), 3.45--3.15 (m, 4H) ppm. ^31^P-NMR (81.0 MHz, CDCl~3~) *δ* = −0.9 (s), −25.0 (s) ppm. ESI-MS (CH~2~Cl~2~/CH~3~CN 1:10, positive mode): *m*/*z* = 1073.4 (C~51~H~38~O~3~P~2~ReTe) for \[(6-Ph~2~P-Ace-5-)~2~Te-Re(CO)~3~\]^+^. FTIR: $\widetilde{v}$ (CO) = 2018, 1928, 1882 cm^−1^. *mer*-**1**: ^31^P-NMR (81.0 MHz, CDCl~3~) *δ* = 14.7 (d, *J*(^31^P-^31^P) = 11.6 Hz), −26.4 (d, *J*(^31^P-^31^P) = 11.6 Hz) ppm. *Synthesis of (6-Ph~2~P-Ace-5-)~2~TeMn(CO)~3~Br* (*fac*-**2**): Mn(CO)~5~Br (17.5 mg, 64.4 μmol) was added to a solution of (6-Ph~2~P-Ace-5-)~2~Te (50 mg, 62.3 μmol) in THF (10 mL). The mixture was continually stirred at 65 °C for 32 h. The product precipitated upon cooling to room temperature as orange microcrystals (30 mg, 29.4 μmol, 47%; Mp. 185 °C (dec.)). Due to the low-solubility, the equilibrium between *fac*-**2** and *mer*-**2** in solution and the concomitant formation of **3** reliable ^13^C- and ^125^Te-NMR spectra could not be obtained. *fac*-2: ^31^P-NMR (81.0 MHz, CDCl~3~) *δ* = 50.4 (s, br), 39.9 (s, br), −27.4 (d, *J*(^31^P-^31^P) = 7.2 Hz) ppm. *mer*-2: ^1^H-NMR (200.1 MHz, CDCl~3~) *δ* = 7.95 (d, ^3^*J* = 7.2 Hz, 2H), 7.80--6.85 (m, 26H), 3.41 (m, 8H) ppm. ^31^P-NMR (81.0 MHz, CDCl~3~) *δ* = 50.0 (s, br), −27.9 (d, *J*(^31^P-^31^P) = 7.2 Hz) ppm. ESI-MS (CH~2~Cl~2~/CH~3~CN 1:10, positive mode): *m*/*z* = 943.1 (C~51~H~38~O~3~P~2~MnTe) for \[(6-Ph~2~P-Ace-5-)~2~Te--Mn(CO)~3~\]^+^. FTIR: $\widetilde{v}$ (CO) = 2011, 1939, 1895 cm^−1^. Synthesis of (6-Ph~2~P-Ace-5-)~2~TeMn(CO)~2~Br (3). Mn(CO)~5~Br (7.00 mg, 25.5 μmol) was added to a solution of (6-Ph~2~P-Ace-5-)~2~Te (20.0 mg, 24.9 μmol) in THF (5 mL). The mixture was continually stirred at 65 °C for 7 d. The product precipitated upon cooling to room temperature as yellow microcrystals (10.0 mg, 10.0 μmol, 40%; Mp. 150 °C (dec.)). ^1^H-NMR (360.3 MHz, CD~2~Cl~2~) *δ* = 8.03 (s, 4H, H~o~), 7.60 (d, *J* = 7.1 Hz, 2H, H~4~), 7.52 (s, 6H, H~m+p~), 7.35--7.19 (m, 8H, H~m'+p'~+ H~8~), 7.19--7.05 (m, 6H, H~o'~+ H~4~), 7.00 (m, 2H, H~7~), 3.50--3.34 (m, 8H, H~1+2~). ^13^C-NMR (90.6 MHz, CD~2~Cl~2~) *δ* = 150.8 (s, C~c~), 150.2 (s, C~d~), 139.9 (t, *J* = 4.0 Hz, C~b~), 139.0 (s, C~4~), 136.8 (t, *J* = 5.1 Hz, C~a~), 136.2 (s, C~7~), 135.7 (t, *J* = 18.1 Hz, C~i~), 135.7 (d, *J* = 18.0 Hz, C~i'~), 134.5 (t, *J* = 4.9 Hz, C~o~), 133.1 (t, *J* = 4.6 Hz, C~o'~), 129.8 (s, C~m~), 129.0 (s, C~m'~), 127.8 (t, *J* = 4.4 Hz, C~p~), 127.5 (t, *J* = 4.2 Hz, C~p'~), 127.0 (d, *J* = 17.5 Hz, C~6~), 119.5 (t, *J* = 3.9 Hz, C~3~), 119.4 (s, C~8~), 104.6 (t, *J* = 7.0 Hz, C~5~), 29.8 (s, C~1~), 29.7 (s, C~2~). ^31^P-NMR (81.0 MHz, CD~2~Cl~2~) *δ* = 59.7 (s, br). ^125^Te-NMR (113.7 MHz, CD~2~Cl~2~) *δ* = 753.1 (s). ESI-MS (CH~2~Cl~2~/CH~3~CN 1:10, positive mode): *m*/*z* = 859.4 (C~50~H~36~MnO~2~P~2~Te) for \[\[(6-Ph~2~P-Ace-5-)~2~Te--Mn(CO)~2~\]--2CO\]^+^. FTIR: $\widetilde{v}$ (CO) = 1936 (s), 1852 (s) cm^−1^. X-ray crystallography. Single crystals of *fac*-**1**, *fac*-**2**, and **3** were grown by slow evaporation of CH~2~Cl~2~ solutions that contained small amounts of *n*-hexane. Intensity data of *fac*-**1**, *fac*-**2**·CH~2~Cl~2~, and **3**·2 CH~2~Cl~2~ were collected at 100 K on a Bruker Venture D8 diffractometer with graphite-monochromated Mo-Kα (0.7107 Å) radiation. All structures were solved by direct methods and refined based on F^2^ by use of the SHELX program package as implemented in WinGX \[[@B23-molecules-23-02805]\]. Strongly disordered solvent molecules were accounted using the SQUEEZE routine for **1** and **2** \[[@B24-molecules-23-02805]\]. All non-hydrogen atoms were refined using anisotropic displacement parameters. Hydrogen atoms attached to carbon atoms were included in geometrically calculated positions using a riding model. Crystal and refinement data are collected in [Table 2](#molecules-23-02805-t002){ref-type="table"}. [Figure 1](#molecules-23-02805-f001){ref-type="fig"}, [Figure 2](#molecules-23-02805-f002){ref-type="fig"}, [Figure 3](#molecules-23-02805-f003){ref-type="fig"} and [Figure 4](#molecules-23-02805-f004){ref-type="fig"} were created using DIAMOND \[[@B25-molecules-23-02805]\]. Crystallographic data (excluding structure factors) for the structural analyses have been deposited with the Cambridge Crystallographic Data Centre. Copies of this information may be obtained free of charge from The Director, CCDC, 12 Union Road, Cambridge CB2 1EZ, UK (Fax: +44-1223-336033; e-mail: <deposit@ccdc.cam.ac.uk>). The Deutsche Forschungsgemeinschaft (DFG) is gratefully acknowledged for financial support. **Sample Availability:** Samples of *fac*-**1**, *fac*-**2**, and **3** are available from the authors. The following are available online. ###### Click here for additional data file. T.G.D. carried out the experimental work, the spectroscopic characterization and participated in writing the manuscript. As Co-Investor, E.H. was involved in the interpretation of the results and also participated in writing the manuscript. As Crystallographer, E.L. carried out the X-ray structure analyses. As Principal Investigator, J.B. was responsible for writing the publication. The authors declare no competing financial interest. Figures, Schemes and Tables =========================== {#molecules-23-02805-sch001} {#molecules-23-02805-sch002} {#molecules-23-02805-sch003} {#molecules-23-02805-f001} {#molecules-23-02805-f002} {#molecules-23-02805-f003} {#molecules-23-02805-f004} molecules-23-02805-t001_Table 1 ###### Experimental interatomic distances \[Å\] and angles \[°\] of *fac*-**1**, *fac*-**2**, and **3**. *fac*-1 (M = Re) *fac*-2 (M = Mn) 3 (M = Mn) ------------------------------------------------- -------------------- --------------------- --------------------- **Bond Lengths and Angles** M(1)--Te(1) 2.733(1) 2.599(1) 2.546(1) M(1)--P(1) 2.466(1) 2.343(1) 2.293(1) M(1)--P(2) \- \- 2.277(1) M(1)--X(1) 2.487(1) 2.498(1) 2.540(1) M(1)--C(5) 1.930(2) 1.819(5) 1.800(4) M(1)--C(6) 1.930(2) 1.817(6) 1.811(4) M(1)--C(7) 1.951(2) 1.915(7) \- C(5)--O(5) 1.142(2) 1.127(7) 1.146(6) C(6)--O(6) 1.121(2) 1.108(7) 1.037(5) C(7)--O(7) 1.141(2) 0.993(8) ^a^ \- P(1)--M(1)--P(2) \- \- 170.8(1) P(1)--M(1)--C(7) 171.9(1) 171.2(2) \- X(1)--M(1)--C(6) 176.2(1) 173.9(2) 176.3(1) Te(1)--M(1)--C(5) 174.0(1) 175.9(2) 175.6(1) ***peri* Region distances** P(1)···Te(1) 3.363(1) 3.466(1) 3.247(2) P(1)···Te(2) 3.111(1) 3.141(1) 3.429(1) ***peri* Region bond Angles** P(1)--C(18)--C(19)/P(2)--C(48)--C(49) 122.9(1)/119.5(1) 124.6(3)/118.8(3) 124.3(3)/121.0(3) C(10)--C(19)--C(18)/C(40)--C(49)--C(48) 130.2(1)/129.0(1) 129.8(4)/129.9(4) 130.3(4)/129.8(4) Te(1)--C(10)--C(19)/Te(1)--C(40)--C(49) 127.6(1)/122.0(1) 129.6(3)/124.1(3) 123.2(3)/126.1(3) ∑ of bay angles 380.7(3)/370.5(3) 380.0(10)/372.8(1) 377.8(1)/376.9(1) Splay angle ^b^ 20.7(3)/10.5(3) 20(10)/12.8(1) 17.8(1)/17.4(1) **Out-of-Plane Displacement** Te(1) ^c^ 0.198(1)/−0.626(1) −0.191(1)/−0.390(1) 0.0899(3)/−0.666(1) P(1) −0.203(1) 0.097(1) −0.160(1) P(2) 0.278(1) 0.263(1) 0.543(1) **Central Acenaphthene Ring Torsion Angles** C:(13)--(14)--(19)--(18)/(43)--(44)--(49)--(48) 175.5(1)/177.9(1) −179.3(6)/−173.3(5) 179.4(4)/169.4(5) C:(15)--(14)--(19)--(10)/(45)--(44)--(49)--(40) 178.7(1)/176.2(1) 178.5(5)/−177.5(5) 179.4(4)/178.6(5) ^a^ This value might be erroneously low due to unresolved disorder. ^b^ Splay angle: sum of the three bay region angles −360°. ^c^ *fac*-**1**, *fac*-**2**, and **3** show a transoid out-of-plane displacement. molecules-23-02805-t002_Table 2 ###### Crystal data and structure refinement of *fac*-**1**, *fac*-**2**, and **3**. *fac*-1 *fac*-2·CH~2~Cl~2~ 3·2 CH~2~Cl~2~ ----------------------------------------- -------------------------- ------------------------------- ------------------------------- Formula C~51~H~36~ClO~3~P~2~ReTe C~52~H~38~BrCl~2~MnO~3~P~2~Te C~52~H~40~BrCl~4~MnO~2~P~2~Te Formula weight, g mol^−1^ 1107.99 1106.11 1163.03 Crystal system Triclinic Monoclinic Triclinic Crystal size, mm 0.09 × 0.08 × 0.06 0.10 × 0.08 × 0.05 0.09 × 0.08 × 0.08 Space group *P* $\overline{1}$ *P*2~1~/*n* *P* $\overline{1}$ *a*, Å 12.2258(3) 10.2073(2) 12.542(5) *b*, Å 14.3531(3) 22.4724(5) 12.738(5) *c*, Å 14.8486(3) 19.7629(5) 15.516(5) α, ° 78.509(1) 90 100.490(5) β, ° 77.874(1) 101.421(1) 100.078(5) γ, ° 89.216(1) 90 100.930(5) *V*, Å^3^ 2495.33(9) 4443.5(2) 2336(2) *Z* 2 4 2 ρ~calcd~, Mg m^−3^ 1.475 1.653 1.653 μ (Mo *Kα*), mm^−1^ 3.163 2.077 2.089 *F*(000) 1080 2200 1156 θ range, deg 2.31 to 29.57 2.29 to 27.56 2.38 to 30.12 Index ranges −16 ≤ h ≤ 16 −13 ≤ h ≤ 12 −17 ≤ h ≤ 17 −19 ≤ k ≤ 19 −29 ≤ k ≤ 29 −17 ≤ k ≤ 17 −20 ≤ l ≤ 20 −25 ≤ l ≤ 25 −21 ≤ l ≤ 21 No. of reflns collected 275,626 120,248 202,873 Completeness to *θ*~max~ 99.8% 99.8% 99.8% No. indep. Reflns 13,983 10,262 13,719 No. obsd reflns with (*I* \> 2*σ*(*I*)) 13,702 8891 11,512 No. refined params 532 553 556 GooF (*F*^2^) 1.074 1.042 1.054 *R*~1~ (*F*) (*I* \> 2*σ*(*I*)) 0.0148 0.0559 0.0553 *wR*~2~ (*F*^2^) (all data) 0.0385 0.1690 0.1529 Largest diff peak/hole, e Å^−3^ 0.844/−0.846 1.724/−1.561 1.955/−1.665 CCDC number 1861602 1861603 1861604

Background {#Sec1} ========== Currently, Descemet stripping automated endothelial keratoplasty (DSAEK) is the most frequently performed surgical treatment for corneal stromal edema due to corneal endothelial dysfunction. In Japan, postoperative complications of cataract surgery or laser iridotomy are the major causes for corneal endothelial dysfunction along with a small number of cases of Fuchs dystrophy. After Kobayashi et al. reported non-Descemet's stripping automated endothelial keratoplasty (nDSAEK) \[[@CR1]--[@CR3]\], nDSAEK has become the mainstream technique to treat corneal edema in Japan. Graft dislocation is the most common postoperative complication either after DSAEK or nDSAEK \[[@CR4]\]. Graft dislocation is usually treated immediately with air injection into the anterior chamber (AC) or with resurgery to suture the graft to achieve centration. On the other hand, cases of spontaneous reattachment of dislocated donor graft after DSAEK have been reported \[[@CR5], [@CR6]\]. We have experienced several cases of nDSAEK graft detachment which did not spontaneously reattach, including cases with extensively dislocated graft and others with persisting folds and wrinkles in the transplanted graft. Here we report a case of spontaneous reattachment of dislocated graft after nDSAEK that had occurred during follow-up with good visual prognosis. Case presentation {#Sec2} ================= The patient was a 54-year-old male without any systemic diseases. He underwent PKP in 2001 for keratoconus in his left eye, and received another PKP in May 2014 due to opacity caused by graft rejection. After the second PKP, the graft was in good condition but visual acuity declined because of progressed cataract. Subsequently, phacoemulcification with intraocular lens implantation (PEA + IOL) was carried out in May 2015. Six months later, graft edema appeared due to bullous keratopathy \[Fig. [1](#Fig1){ref-type="fig"}\] and the best corrected visual acuity (BCVA) was 0.2.Fig. 1The cornea before nDSAEK. **a** The whole graft showed edema due to corneal endothelial dysfunction after cataract surgery. **b** Corneal stromal edema and corneal opacity were observed Under local anesthesia, nDSAEK was performed in May 2016. We prepared a temporal scleral tunnel incision 5.0 mm wide and used a precut donor graft of 8.5 mm. After making four venting incisions, we performed inferior iridectomy with a vitreous cutter. The nDSAEK button was pulled into the AC using a double glide technique \[[@CR7]\] with an AC maintainer, a Busin glide spatula (Moria Japan, K.K., Tokyo, Japan), and gripping forceps 23 G (Eye Technology Ltd., Essex, UK). Graft apposition was adjusted and interface fluid between the recipient cornea and the donor graft was expressed through corneal venting incisions. Air was slowly injected into the AC using a syringe equipped with a 30 G needle. However, the air injection could not achieve sufficient tamponade effects. Because the surgery was carried out under local anesthesia and the patient appeared quite stressed and tense, we suspected that this might have caused the high vitreous pressure. Slit lamp microscopy confirmed sufficient graft attachment immediately after the surgery and the operation was completed. No intraoperative complications were noted. However, the graft almost completely detached 3 h later except a temporal part. Air was reinjected on the following day and the attachment was improved although approximately 1/5 of the graft remained detached in the inferior nasal area. We determined that air reinjection would not further improve the condition and decided to follow him up, which was also the patient's wish for not having any more operations immediately. Nine days after the nDSAEK, graft dislocation progressed further \[Fig. [2](#Fig2){ref-type="fig"}\]. The area of detachment extended to almost 3/4 of the graft and the BCVA was 0.2.Fig. 2The cornea 9 days after the nDSAEK. **a** Graft dislocation extended to about 3/4 of the whole graft and edema was spotted across the cornea. **b** Edema in the corneal stroma was found and the cornea appeared cloudy Unexpectedly, spontaneous graft attachment was observed at 1 month after the nDSAEK and the corneal edema had almost subsided leaving only a mild edema in the superior temporal cornea. His vision improved subjectively and objectively, and the BCVA was 1.0 \[Fig. [3](#Fig3){ref-type="fig"}\]. Three months after the nDSAEK, the graft remained in position without any observed complications and the cornea kept its transparency \[Fig. [4](#Fig4){ref-type="fig"}\]. The patient has been followed up ever since.Fig. 3The cornea 1 month after the nDSAEK. **a** Corneal edema was resolved and the whole graft appeared transparent. **b** No graft detachment was notedFig. 4The cornea 3 months after the nDSAEK. **a** The whole graft remained transparent. **b** The graft did not dislocate and the cornea kept transparency Discussion {#Sec3} ========== We experienced a case of a large nDSAEK graft detachment that had spontaneously reattached 1 month after surgery. Our case demonstrated that a close observation without immediate additional operations could be an option for a large but partial graft detachment in the early postoperative period after nDSAEK. Fuchs corneal dystrophy is the leading cause for corneal endothelial dysfunction in Europe and North America and with Fuchs dystrophy, the recipient Descemet's membrane and endothelium must be removed to eliminate the guttae that distort vision. In Japan, cataract surgery and argon laser iridotomy are the main factors for corneal endothelial dysfunction and there is no need to strip the Descemet's membrane. Therefore, nDSAEK has been widely adopted for the treatment of corneal edema. Graft dislocation is a postoperative complication that occasionally occurs after DSAEK and nDSAEK. Reportedly, the decision to remove the Descemet's membrane or not does not make a remarkable difference in the frequency of postoperative graft dislocation \[[@CR1]--[@CR3]\]. When graft dislocation is observed postoperatively, air injection into the AC, reoperation for graft centration, and graft suturing for insufficient attachment are the common procedures performed immediately to deal with the dislocation. Morishige et al. reported that the duration of bullous keratopathy influences the visual prognosis after DSAEK. They concluded that 1 year after the onset of bullous keratopathy, postoperative visual prognosis becomes poor due to cicatrization \[[@CR8]\]. Their finding suggests that early intervention for graft dislocation is necessary to prevent visual loss. On the other hand, spontaneous graft reattachment has been reported. Hayes and colleagues reported 12 cases of graft dislocation after DSAEK that were followed up without resurgery \[[@CR6]\]. In their study, spontaneous reattachment occurred between postoperative 5 days to 7 months. In the 2 cases that required a longer period for reattachment (3 and 7 months), corneal transparency could not be achieved. On the contrary, corneal transparency was achieved in the 2 cases observed with spontaneous reattachment within a shorter period (5 days and 20 days) \[[@CR6]\]. Spontaneous graft reattachment does not occur commonly and it is difficult to predict its occurrence. Its mechanism is also unknown. We suspected that the mechanism of spontaneous reattachment may be related to the following: 1) The graft observed with dislocation is always pushed toward the side of the host cornea by the aqueous humor flow and ocular pressure. 2) The aqueous humor between the graft and the host cornea tends to be absorbed by the pumping function of the endothelial cells. Moreover, graft reattachment after DSAEK is considered to share a similar mechanism with the spontaneous reattachment in cases exhibiting detachment of the Descemet's membrane after cataract surgery, which are considered rare \[[@CR9]\]. In the reported case, graft attachment did not appear to have any problem immediately after the nDSAEK, however, graft detachment started 3 h later. We considered 3 possible causes for the graft detachment: 1) The posterior surface of the cornea could have been flattened by the site of the corneal suture since the patient had previously received PKP. 2) The corneal graft of 8.5 mm in diameter might have been relatively too large. 3) Because the operation was performed under local anesthesia and the vitreous pressure was high, the air injection might have been insufficient and caused the fluid to remain between the host cornea and the donor graft. Likewise, graft detachment could occur in various cases. Although persistent graft detachment can cause poor visual outcome due to corneal fibrosis, the decision between follow-up and immediate reoperations should be made carefully on a case-by-case basis considering the additional burden on patient's physical condition or possible complications such as iris damage, primary graft failure, and infections. Conclusion {#Sec4} ========== In conclusion, our case has demonstrated that even in eyes with a large but partial graft detachment after nDSAEK, the graft detachment can be monitored up to a month for spontaneous reattachment when further intervention appears to be difficult. Although good visual outcomes were obtained in our case, further investigations on the mechanism of spontaneous reattachment, how often it occurs, and in what cases it is most likely to occur will be necessary. AC : Anterior chamber BCVA : Best corrected visual acuity nDSAEK : Non-Descemet stripping automated endothelial keratoplasty PKP : Penetrating keratoplasty We thank Ms. Reiyo Tahara for her professional manuscript editing service. Funding {#FPar1} ======= None. Availability of data and materials {#FPar2} ================================== All data are shown in the figures. AT carried out substantial parts of the work including diagnosing the patient, drafting the manuscript, and finalizing it. MF participated in the treatment by performing PKP, PEA+IOL, and nDSAEK, and following up the patient. KS and AK participated in the treatment of the patient. YS helped finalize the manuscript and gave the final approval of the version to be published. All the authors have read and approved the final manuscript. Ethics approval and consent to participate {#FPar3} ========================================== This study was approved by the Kindai University Ethics Committee and adhered to the tenets of the Declaration of Helsinki. Consent for publication {#FPar4} ======================= Written informed consent was obtained from the patient for publication of this case report and any accompanying images. Competing interests {#FPar5} =================== The authors declare that they have no competing interests. Publisher's Note {#FPar6} ================ Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Background {#Sec1} ========== Over the last decade medical education has moved away from a traditional didactic, system with lecture-based teaching to more contemporary self-directed learning supported by case-based discussions stimulated by problem-based curricula. More recently, electronic media, such as web-based online supporting material, live streaming of lectures and downloadable applications have emerged as valuable teaching tools \[[@CR1]--[@CR3]\]. Yet, despite making use of modern platforms, one problem remains the same: these are often passive ways of learning, lacking interactive or integrated assessment elements (of students learning). It is now acknowledged that to maximise medical students learning, in terms of effectiveness and retention, learning should be active e.g. problem based, preferably interactive (especially for deep learners) and clinically relevant and that it is an unavoidable truth that assessment drives learning \[[@CR4]--[@CR10]\]. The use of simulated (modelled) patients, in virtual learning environments, is becoming common-place in both undergraduate and post-graduate medical teaching \[[@CR11]\]. Simulations enable commonly encountered and potentially serious conditions to be explored, in a safe and non-threatening environment, with instant feedback on performance. The use of simulated patients adds to the richness of a learning experience and places theory in a clinical context \[[@CR12]\]. The author's aim therefore was to exploit local expertise in the Medical School (MS) at the University of Sheffield (UoS) in cardiovascular modelling, clinical cardiology and e-learning, to create an environment, incorporating both interactive and assessment elements, to assist undergraduate medical students to develop a good understanding of key physiological and pathophysiological concepts relating to the cardiovascular system. The environment used easy to understand cardiovascular system (CVS) models, up-to-date clinical information and realistic, simulated, clinical cases, all based on the needs of the current medical curricula. The science of the cardiovascular system and the art of practicing cardiology are knowledge and skills that medical students and junior doctors must master. Cardiac disease is common, being among the most frequent reasons for a patient to present at hospital, with symptoms as varied as chest pain, breathlessness, palpations or dizziness. The practice of cardiology is complex, and one of the most rapidly growing subspecialised disciplines in all of hospital medicine, with a huge associated panoply of imaging modalities, interventional techniques and electrical devices. Such complexity suggests the requirement for hi-fidelity learning modalities. Whilst there are a wide range of on-line case-based teaching platforms offered by journals e.g. British Medical Journal (BMJ) and learned societies such as the European Society of Cardiology (ESC) (<http://learn.escardio.org/Default.aspx>), there are no tools such as those developed in dermatology by Wahlgren et al. (2006) that target undergraduates \[[@CR13], [@CR14]\]. The work by Wahlgren et al. (2006) developed a computerised interactive case simulation system, where the "student selects a patient and proposes questions for medical history, examines the skin, and suggests investigations, diagnosis, differential diagnoses and further management" based on authentic cases with images from real patients \[[@CR13]\]. Whilst students reported this facilitated their learning, there was no actual assessment of their learning. More specifically to cardiovascular science, Dassen et al. \[[@CR15]\], developed the "CircAdapt" mathematical model for first year medical students in Maastrict, which was a lumped parameter model (LPM) describing the hemodynamic interaction between the left and right ventricle, simulating the pulmonary and systemic circulations. This model enabled students to interact with model parameters, such as heart rate or peripheral resistance to mimic different physiological states, such as exercise or hypovolemic shock. This study did not report any assessment of learning, integration with actual clinical cases or provide student feedback on the model. Many approaches exist to develop mathematical models of the circulation, governed by energy, mass and momentum conservation. Lumped parameter, or zero-dimensional, circulation models are often described in terms of hydraulic-electrical analogues; just as in an electrical circuit the voltage gradient determines the flow of current, the flow of blood in the cardiovascular system is determined by the pressure gradient and so a LPM can be used to simulate physiology and pathophysiology of the cardiovascular system, furthering understanding. The benefits of an LPM approach is that they typically provide simple, elegant and quick solutions and require relatively few model parameters which can typically be related directly to concepts familiar to the clinician. Considering this context, an interactive model-based environment, which explains the fundamentals of cardiovascular physiology and pathophysiology, placed in a relevant clinical context for undergraduates, with built in assessment, could help the transition from passive student to responsible clinician. Aims {#Sec2} ---- To incorporate a LPM into an e-learning environment (ELE) to act as a tool for understanding CVS physiology, map the progression of key disease processes and allow students to classify severity of disease, which will all enable a greater understanding of abnormal cardiovascular physiology, in a clinical context which is appealing to students. Methods {#Sec3} ======= Lumped parameter model {#Sec4} ---------------------- This study used a LPM of the human circulation, similar to that used by Fischer et al. \[[@CR16]\], but modelled on computer, as a physiological teaching tool (see Fig. [1](#Fig1){ref-type="fig"}) \[[@CR17]\]. This model was selected for its simplicity; each element could be easily understood providing insight into the dynamics of the CVS, it was validated and had been used in earlier work to model heart failure severity \[[@CR18], [@CR19]\]. According to Shi \[[@CR20]\] LPMs "assume a uniform distribution of the fundamental variables (pressure, flow and volume) within any particular compartment (organ, vessel or part of vessel) of the model at any instant in time" and are based upon ordinary differential mathematical equations. Other, more complex, distributed parameter models in 1-, 2- or 3-D recognise the variation of these parameters in space and use partial differential mathematical equations to describe arterial pressure and flow. Thus the different dimensional levels can be considered in the following way; a LPM can be thought of as representing a time-dependent function as a single region in space, a 1-D model as a line modelling various points along the length (x axis) with time, a 2-D model as a plane modelling various points along length (x axis) and width (y axis) with time and a 3-D model extends across all three planes modelling various points along length (x), width (y) and depth (z axis). A LPM can elegantly represent the human heart and peripheral circulation as an electrical analogue, with a capacitor (C) representing the elastic property of the large arteries (total arterial compliance) and a resistor (R) representing the frictional loss in the smaller vessels (systemic vascular resistance). The pumping chambers are represented by variable capacitors and the valves by diodes, ensuring a unidirectional flow of current representing blood flow. To model hypertension (HTN) the resistance can be increased, to model haemorrhagic shock (HS) the volume of blood in the system can be reduced and to model heart failure (HF), the contractile performance of the left ventricle can be reduced. Such models are validated and freely available from the online CellML model repository and can be run using OpenCell software (<https://www.cellml.org/tools/downloads/opencell>) which is, similarly, freely-available.Fig. 1Lumped parameter model of the (left) heart, with left atrium (Ela), left ventricle (Elv), aortic valve (CVao) and mitral valve (CVmi) and systemic loop (circulation), with resistor (Rv) and capacitor (Cv). Note P and Q, denote pressure and flow respectively which can be measured at these components The medical school {#Sec5} ------------------ The University of Sheffield Medical School (UoS MS) has around 250 students in each of the 5 years of undergraduate study, divided into 4 phases, phase 1 "introductory clinical competency" e.g. medical sciences, phase 2 "basic clinical competency" e.g. clinical attachments, phase 3 "extended clinical competency" e.g. women and children's health and phase 4 "advanced clinical competency" e.g. shadowing junior doctors. The UoS MS curriculum was reviewed, identifying key areas such as blood pressure control, the cardiac cycle and defining cardiac output, which the model was able to help explain. Relevant core clinical problems such as breathlessness and chest pain, were also identified to complement the basic science and also inform the interactive cases. Content development {#Sec6} ------------------- The clinical topics, HF, HTN and HS were chosen as a focus of the model and to form a basis for the ELE. Up-to-date clinical guidelines on these topics were reviewed and, along with previous work, used as the basis for the models, to ensure that the outputs of the model (blood pressure (BP), stroke volume (SV) and heart rate (HR)) accurately represented different stages of each disease as described by the guidelines (see Tables [1](#Tab1){ref-type="table"} and [2](#Tab2){ref-type="table"}) \[[@CR19], [@CR21]--[@CR23]\]. This had the additional benefit of helping the students become familiar with clinical guidelines, categorisation of disease types and severity criteria. To ensure timeliness when engaging with the environment and to minimise possible difficulties in using the model, outputs were pre-computed before exposing the environment to the students (i.e. the model was not solved in real time when using the environment). The model results were created using baseline parameters (representing normal human physiology) and the input variables were then tuned (as mentioned above) to match those typically found in patients in each stage of each disease. For example mild, moderate and severe HTN were simulated by gradually increasing the systemic resistance R (see Table [1](#Tab1){ref-type="table"}). Input variables and model outputs were displayed to the student in both tabular and graphical formats, the latter resulting in a left ventricle pressure-volume loop (see Fig. [2](#Fig2){ref-type="fig"}).Table 1The input variables used to model the various pathophysiological processes with the LPMInputsPathological ProcessLVEmaxRCVolumeTmmhg/mlmmhg.s/minml/mmhg(ml)(s)Normal2.50.0941.757001HypertensionStage 12.51.1941.2657001Stage 22.71.3741.1657001Accelerated2.71.6141.1657001Heart FailureMild21.0941.757001Moderate1.41.0941.6557000.857Severe0.81.1941.5557000.75HaemorrhageStage I2.80.9941.752000.6Stage II3.21.0341.745000.5Stage III31.0341.737000.429Stage IV31.0181.728500.375*C* Total arterial compliance, *LVEmax* Maximal left ventricular elastance, *R* systemic vascular resistance, *T* duration of each heartbeat (i.e. 60/heart rate), *Volume* circulating blood volume Table 2The output parameters produced by the LPM for the various pathophysiological processesOutputsPathological ProcessGuideline Derived DefinitionBlood Pressure (mmHg)Cardiac OutputHeart RateSystolicDiastolic(l/min)(bpm)NormalN/A12080560HypertensionStage 1140--159/90--99 mmHg14080560Stage 2\>160/\>100 mmHg16090560Accelerated\>180/\>120 mmHg180100560Heart FailureMildLVEF \<50%120804.560ModerateLVEF \<40%11070470SevereLVEF \<30%100603.580HaemorrhageStage I\<15% volume loss120805100Stage II15--30% volume loss100704120Stage III30--40% volume loss80603.5140Stage IV50% volume loss60503160*Bpm* beats per minute, *l/min* litres per minute, *LVEF* left ventricular ejection fraction Fig. 2Screen shot of the environment displaying lumped parameter results in mild hypertension. The left hand side displays pressure traces in the atria, ventricle and systemic arterial system in mild hypertension (red), with comparison with a healthy adult (green) and ECG in black. The table in the top right provides a comparison of the physiological and modelled parameters in a healthy adult (green), hypertension (red) and % change from baseline (yellow). The lower right figure shows these changes as a pressure volume loop. At the bottom users can choose to view the results from models of mild, moderate or severe hypertension Three virtual cases were developed for each pathology to map the model information and help the students place the physiology and model results in clinical context. Each comprised a brief history, examination findings and previous investigations where relevant. Questions were posed to the user on diagnosis, treatment, pathophysiology, severity of disease and how disease was represented within the model. Building the environment {#Sec7} ------------------------ The environment was built using a slide presentation platform (Microsoft Powerpoint), published as a PDF and hosted on Minerva; UoS MS's Managed Learning Environment (<https://www.sheffield.ac.uk/polopoly_fs/1.145927!/file/Minerva_eportfolio.pdf>). Minerva is a web-based central resource for students and staff fulfilling many functions relating to the delivery and management of the undergraduate medical curriculum. These include information about the structure of the curriculum, timetables, course handbooks, learning materials and resources, clinical placements, student selected choice, integrated learning activities, personal portfolios, assessments and news items. It also provides access to the core curriculum database. The environment comprised a general introduction, modules on basic cardiovascular physiology and the three areas of pathology sub-divided into aetiology, epidemiology, pathophysiology (Fig. [3](#Fig3){ref-type="fig"}), treatments (Fig. [4](#Fig4){ref-type="fig"}), interactive model results (Fig. [2](#Fig2){ref-type="fig"}) and virtual clinical cases (Figs. [5](#Fig5){ref-type="fig"} and [6](#Fig6){ref-type="fig"}). For each disease, questions were built into the text, both on the model parameters (for example how could one model HTN using the LPM) and then on the disease itself (for example how a patient with HF would be treated).Fig. 3Bar chart comparing cardiovascular science knowledge before and after viewing the learning environment using Student's T-test (black: first year medical students, white: students from all years) Fig. 4Screen shot of the electronic learning environment (ELE) demonstrating information given for each of the three clinical conditions. This includes: definition, categorisation, pathophysiology, epidemiology, aetiology At the bottom of the screen "Back" and "Proceed" buttons allow sequential navigation through the environment. The navigation tab on the left allows users to jump between areas of the ELE Fig. 5Screen shot of the electronic learning environment (ELE) demonstrating information given for each of the three clinical conditions. This includes clinical assessment, investigation, treatment and sequalae. At the bottom of the screen "Back" and "Proceed" buttons allow sequential navigation through the environment. The navigation tab on the left allows users to jump between areas of the ELE Fig. 6Screen shot of the electronic learning environment displaying a virtual clinical case of hypertension. The history, examination, past history and investigation outcome are provided as an introduction followed by a choice of 4 questions Assessing and launching the environment {#Sec8} --------------------------------------- ### Assessment {#Sec9} To capture the end-users' opinion of the environment, a questionnaire was developed, exploring the perceived relevance of the topic, the manner of delivery and the fidelity of the environment itself. Similarly, following initial exposure to a small focus group (see below) a series of 6 questions were developed based on information contained within the environment testing knowledge of the physiology and pathophysiology of each condition. The questions were based on either a true/false or extended-matching question (EMQ) format familiar to the students, as used at UoS MS (see Additional file [1](#MOESM1){ref-type="media"}). ### The learning environment exposed {#Sec10} The exposure of the ELE involved three distinct elements. Phase one, in November 2012, a small self-selected focus group of 9 students (from all years) were recruited in response to an advert on Minerva and given a 90-min interactive tutorial, following which they were supported by the authors in use of the environment. The feedback from this session was used to revise the environment. Phase two, the environment was demonstrated to 28 first year medical students following a routine lecture on the cardiovascular system (in December 2012), and the environment was then released to this group in January 2013, for a period of 2 weeks, after they had finished the cardiovascular taught module and before starting their first clinical placements. Feedback was gained on the environment in terms of content, delivery, style and presentation with a free text option (see Additional file [2](#MOESM2){ref-type="media"}) and the students also tested their knowledge of the CVS, based on information contained within the environment, both before and after use (see Additional file [1](#MOESM1){ref-type="media"}). This cohort was independent of the focus group (many of whom were close to graduation). Use of the environment was optional and other than the demonstration in the lecture, no other assistance was given. Phase three, Finally, the environment was made available to the entire UoS MS. Over seventy students from all years viewed the environment, for a period of 2 weeks, gave feedback and were assessed on their learning. Statistics {#Sec11} ---------- Performance on the questions in Additional file [1](#MOESM1){ref-type="media"}, before and after exposure to the ELE was compared with statistical analysis using SPSS statistics, Version 21 (IBM, USA), using a two-tailed t-test with a confidence interval set at 95%. Results {#Sec12} ======= The questionnaire (see Additional file [2](#MOESM2){ref-type="media"}) provided quantitative data on the utility of the ELE whereas qualitative data was acquired (see Table [3](#Tab3){ref-type="table"}) from free text boxes. Quantitative assessment of students' learning was based on 6 questions covering the content of the environment. These questions were directly derived from the cardiovascular medicine component of the UoS MS undergraduate curriculum. There was a statistically significant increase in their performance after exposure to the environment, with a 33% and 22% increase in the number of questions answered correctly, *p* \< 0.0001 and *p* \< 0.001 respectively (see Fig. [7](#Fig7){ref-type="fig"}).Table 3The demographic data of the 3 cohorts of students and their responses to the questionnaire about the e-learning environmentGroupFocus Group (*N* = 9)First Year (*N* = 28)All Medical School (*N* = 76)Age21--2344%18--2064%18--2056%Cardiology^a^18%36%32%Physical^b^sciences22%36%29%Question No.^c^% Agree/ Strongly AgreeMean Likert score% Agree/ Strongly AgreeMean Likert score% Agree/ Strongly AgreeMean Likert score4884.2934.1883.951004.4934.4944.361004.1934.1884.17884.1783.9773.88663.6783.8713.79663.6854.1884.110883.7934.4944.411774.0783.9823.912884.1934.3944.2131004.2854.1854.0141004.21004.5914.215884.21004.4944.1^a^Interested in a career in cardiology or cardiovascular sciences^b^Did you study the physical sciences and^c^see questionnaire for a full list of questions Fig. 7Screen shot of the electronic learning environment displaying a virtual clinical case of hypertension. The answer to a question appears (grey bubble) when the user clicks on this item During phase one comments from the focus group centred on minor (typographical) errors, a requirement for more information about the pressure-volume (PV) loops used to display the modelled results (many students were unfamiliar with these), a request for an increased number of clinical cases as these were well-received and, finally, a more interactive approach to Q&A scenarios making the answer to each question available via a single click (this version showed the answers to all five simultaneously on a separate page). The students reported that they found navigation intuitive, that it successfully communicated the importance of the physical principles of the cardiovascular system and was useful for demonstrating the effects of common cardiovascular pathologies. Suggestions included making links between pages more obvious, using colour-coding to indicate disease severity and using tabs for different pathologies. These suggestions were adopted in order to improve the next iteration of the ELE. Free text comments included -- *"Excellent interactive tool"*, *"Nice the way it integrates both physiology and pathologies of the heart"* and *"CRT = means capillary refill time but then also cardiac resynchronisation therapy, this is confusing"*. Table [3](#Tab3){ref-type="table"} displays the questionnaire results. During phase two first year medical students were given access to the environment for a period of 2 weeks. They engaged well with the environment, requiring minimal guidance; none of the authors were contacted during this period for advice on how to use the environment. As the students enjoyed the clinical cases, and commented they would like more, an additional ten were added for phase 3 evaluation. The students reported that too many abbreviations were used, many of which they had not encountered before. These were explained in full or removed from the phase 3 version. Free text comments include *"Helped me revise the physiology of heart contraction effectively"*, *"Realistic interpretation of medicine when we finally graduate"* and *"Too much information on each page"*. During phase three the UoS MS as a whole was given access to the environment, via Minerva, for a period of 2 weeks. Comments included *"Similar resources for respiratory and neurology would be great!*" , *"An excellent and extremely useful resource, thank you!"* and *"...could do with being less busy, rigid and dark."* Discussion {#Sec13} ========== This is the first study to demonstrate that a LPM can be embedded in a cardiovascular ELE, positively received by medical students and exposure to which appears to improve knowledge of specific areas of cardiovascular pathology, namely hypertension, heart failure and haemorrhage. Other simulations exist for medical students studying cardiovascular science and cardiology, for example how to examine patients and reading of ECG's, but as noted by Owen & Wong most simulations are used as an introduction, not a tool for repetitive practice \[[@CR24]--[@CR28]\]. Such tools allow students to increase their confidence and experience in particular skills, like diagnosing a cardiac murmur or recognising a particular cardiac arrhythmia. Our work was aimed specifically at increasing medical students knowledge of cardiovascular physiology and pathology using an ELE based upon a simple LPM and then applying this to the diagnosis and management of relatively low fidelity simulated patients e.g. text descriptions only. Many studies, including this one, lack a direct comparison to pre-existing practice. For example student performance may improve following exposure to a simulation, but they are seldom randomised to one methodology or another. Only two studies were identified as having done this. Firstly, Sverdup et al. (2010) randomised students to traditional bedside teaching versus computer simulated heart sounds, and indeed found no difference in performance between third year medical students at Oslo University Medical School \[[@CR29]\]. Kern et al. (2011) used students from previous years as controls and provided simulated patients in addition to, not instead of, traditional training and so it is difficult to tease apart the influence of the simulations, even though performance was improved \[[@CR24]\]. Indeed, as demonstrated by Lavaranos, simply introducing an intensive traditional bed-side cardiology course may be sufficient to improve learning, rather than hi-fidelity simulations \[[@CR30]\]. Further questions arise related to ELE effectiveness such as how much is retained in the short, medium and long term and how performance and confidence in clinical practice following graduation is altered? This is highlighted by the study by Vulkanovic-Criley \[[@CR31]\], who demonstrated that even following the completion of clinical training in cardiology, qualified doctors may perform no better than their medical student counterparts, despite the exposure to an ELE comprising of case vignettes of patients with cardiovascular pathology. Many studies, including this one, were conducted at a single academic institution, with only one study using multiple centres; this comprehensive study was undertaken by Vulkanovic-Criley et al. \[[@CR31]\], who tested over 860 participants, including 318 medical students, across multiple centres, on cardiac physiology, auditory skills, visual skills and integration of both auditory and visual skills using computer graphic animations and assessments of virtual cases. However, this was not designed to improve knowledge of skills e.g. it was not an ELE, but merely test them. Intriguingly, they found that whilst cardiac examination skills improved during medical school (from first to third year), they did not improve or differ significantly following the third year of medical school, following qualification or indeed across any group of doctors, regardless of seniority or speciality. Only heart specialists, cardiologists, had significantly better cardiac examination skills compared to the rest of the groups. Petrusa et al. \[[@CR32]\] document the introduction of a 4 year multimedia curriculum in cardiology, into 4 medical schools, at significant cost, comprising of "ten interactive, patient-centered, case-based modules focused on the history, physical examination, laboratory data, diagnosis, and treatment" and whilst 80% of the students rated the system highly, a similar number to this study, there was no comparison with previous teaching delivery and so there is an implicit assumption that this is better, without any measurement, either of students performance or indeed how highly the old system was rated. The most similar previous work to this study is Dassen et al. \[[@CR15]\], who adopted a similar approach based on the 'CircAdapt' model, for first year medical students in Maastrict, but are yet to publish outcome data on how students respond to, or learn from, such a model. Such information is critical to this type of model, or environment development, as without user-feedback or evidence of improvement following use, the models cannot be regarded as valid tools for teaching physiology. Furthermore, whilst the Dassen method involves a more complex but stand-alone model, it also requires background information to be provided during face-to-face teaching. The ELE described here has the advantage that it does not require real-time teacher-student supervision and was tested by the whole medical school. It also includes an integrated Q&A component to test the students understanding of the model, CV physiology and pathophysiology. The simulated clinical cases play an important role in bringing this information together. On the other hand, the Dassen method allows the model to be run in real-time; if sensible physiological boundaries were set this could also be done for our ELE, enabling students to choose their own combinations of vascular resistance, vascular compliance and cardiac performance. A delicate balance had to be achieved between adding detail to the model (with the risk of bewildering those students with little underpinning knowledge of physics) or embedding the model so deeply within the environment that its purpose was lost (see Figs. [5](#Fig5){ref-type="fig"}, [6](#Fig6){ref-type="fig"} and [7](#Fig7){ref-type="fig"}). It is interesting to consider; 'would the ELE be as useful if the model was removed?' The interactive clinical cases and the outputs of their respective models helped maintain a sense of realism and a clinical relevance and enabled the students to test their knowledge and understanding in a safe environment. Consequently we believe that a good balance was reached in this respect and consider that the model provided a platform on which to build the environment, allowing insight into the processes underlying the pathology without which the experience would have been poorer. The process of development was necessarily an iterative one, drawing on previous experience of such environments within the department. The main lesson learned from previous environments, was that the target audience and purpose of environment must be identified a *priori*, otherwise a seemingly excellent environment could be developed that was subsequently found to be either too complex for undergraduate medical students or contained too much basic science for cardiac specialists. At each stage feedback from user-directed improvements ensured that the environment was relevant, appropriate and interesting. Feedback was generally positive, with criticisms limited to presentation, rather than the detailed content or concept as a whole. Similarly, the environment was endorsed by both cardiology academics and medical educationalists at the UoS Medical School, who deliver cardiovascular science lectures and design the curriculum, respectively. The collaborative effort including an e-learning technologist, cardiovascular modellers and a cardiologist with an interest in education, meant that the three key areas of the developmental process, identification of what could be modelled, what was relevant to model and the method of delivery of model output were discussed contemporaneously rather than retrospectively. This led to a streamlined development process. As presented in the results, there was a significant improvement in the number of questions answered correctly after exposure to the ELE, suggesting that the students benefited from the experience. Furthermore, an improvement in baseline results (before exposure) can be seen, between iterations of the model. Notably, the 2nd iteration was available to only first year medical students but the 3rd iteration was accessible to all year groups; it would be expected that their level of clinical knowledge, and therefore their performance, would be much higher. From a technical perspective it became challenging to maintain a structured design of the ELE in its final iteration, which contained over 200 pages. However the user was not required to interact with each page sequentially and no criticisms were recorded in this regard. Choosing how and what to display from the range of possible outputs of the model required a fine balance. Although the students had limited exposure to the concept of PV loops, they had had some exposure to the significance of measures such as SV, BP and HR and the PV loop is an excellent way of graphically demonstrating relationships between cardiovascular parameters. Abbreviations commonly understood by the focus group, proved not to be immediately transparent to first year medical students. As a result, nearly all of the abbreviations were removed from the second iteration of the software with a key provided to aid clarity. Ideally, students would be granted unlimited access to the environment, both in terms of time and usage, but due to time constraints the students were restricted to an access period of 2 weeks after which the link was removed. In addition the software was not downloadable. It is possible that this may have limited uptake and interaction. The model chosen was a simple LPM, but there are many more complex models of the heart and peripheral circulation available; these include 2D, 3D as well as other more complex 0D representations. Whilst the inclusion of a more complex 3D model might yield more complex outputs there is a danger that the software would become slow and unwieldy and ultimately, 3D solutions may not add anything over and above the simple model. In the current environment the results of the model were pre-computed. The model was run with a range of typical inputs and the user had access to a pre-specified set of results. In the future, it would be possible to allow the students to run the model in real time, offering infinite possibilities but boundaries would have to be set to prevent results that are physiologically improbable or impossible. The assessment questions were provided in a format familiar to medical students through the routine MBChB audit of teaching quality. As seen in Table [3](#Tab3){ref-type="table"}, it is notable that whilst all users were of a similar age, less than 30% either had a predilection for cardiology or had studied physical sciences at either pre-university or undergraduate level, demonstrating just how accessible the ELE was without necessarily needing prior knowledge or indeed a special interest. Future work {#Sec14} ----------- The simulated patients could be further developed, allowing history-taking, examination findings, results from investigations and treatment options, all interacting with the model \[[@CR33]--[@CR35]\]. Other opportunities involve tailoring the environment to the British Cardiology Society (BCS) curriculum for cardiologists in training or collaborations with colleagues from other specialties and scientific disciplines for the development of similar environments, in respiratory medicine for example. This study focused on cardiology which is perceived as one of the most interesting specialities. The general approach has potential to be extended to focus on specialties which students and doctors struggle to engage with, neurology for example, but evidence suggests performance across these subjects is actually similar \[[@CR36], [@CR37]\]. One could foresee such an environment being downloadable to a smart phone or tablet device, which may increase engagement and perhaps allow users to edit the environment and upload improved and updated versions. Customisation could even enable them to create their own simulated patients with questions to test their peers. Comparison of performance before and after using the model and comparing this with performance before and after lectures on the same subject would be a useful comparison of teaching methods. The effects of therapies on cardiovascular performance e.g. administration of IV fluids, b-blockers, or use of biventricular pacemakers could also be included. Finally, one could envisage the development of a series of informal teaching aids with students downloading an environment, carrying out beta-testing of the software, and then uploading it for others to use. Clinicians and academics also acknowledge the utility and future possibilities of such environments and as such, there is enthusiasm locally to integrate these types of applications into the undergraduate curriculum. Limitations {#Sec15} ----------- The questionnaire used to assess the ELE was not validated, but was based upon the UoS MS curriculum, which in turn was incorporated into the material on the cardiovascular system for the ELE used in this study and so the results must be viewed in that regard. Only 15 students provided free text responses and so the generalisability of this form of feedback is limited. No attempt was made to investigate potential impact on examination performance or to investigate if exposure to the model encouraged more medical students to choose a period of research in the cardiovascular sciences. There were also many questions embedded in the environment for which responses were not recorded and so it would have been useful to record performance during usage to see whether knowledge improved during progression through the environment. There are some possible confounding variables, for example it was not feasible to restrict access to the ELE during testing and so it is possible some students treated the assessment like an open book, the testing reflecting their ability to access information in the ELE not whether they were able to recall and retain the same information from the ELE. Also, whilst knowledge of the cardiovascular system appeared to improve, ultimately it is understanding of the cardiovascular system that educators seek to enhance, as noted by Dullo and Chaudhary \[[@CR38]\], especially as some students may understand less than they know. Finally, the sample size was small, not randomised, conducted at a single institution, not powered to detect changes in knowledge following exposure to the ELE and was also not directly compared to other learning modalities, or indeed, the students existing one. Conclusions {#Sec16} =========== LPMs are valid components of CVS teaching tools when used in conjunction with a novel and interactive electronic learning environment (ELE). ELEs enable students to interact with, and manipulate, physiological processes that occur in response to disease. Rather than simply reading how the physiological and pathophysiological processes happen, users can study this interactively. In this study, we discovered that students found using this LPM embedded in this particular ELE an interesting and positive experience. Even from a solid foundation of traditional teaching input, considerable improvements were made in students' knowledge during their assessment. This work combines some exciting and developing fields in medical education, namely modelling, e-learning and simulated patients and illustrates the benefits that can be achieved with engagement of educationalists, clinicians and modellers. Additional files ================ {#Sec17} Additional file 1:Questions to test the users knowledge of the cardiovascular system based on information contained in the e-learning environment, assessed before and after exposure. (DOCX 13 kb) Additional file 2:Learning environment questionnaire to explore the perceived relevance of the topic, the manner of delivery and the fidelity of the environment itself. (DOCX 14 kb) BMJ : British Medical Journal BP : Blood pressure CV : Cardiovascular CVS : Cardiovascular system ECG : Electrocardiogram ELE : e-learning environment EMQ : Extended matching questions ESC : European Cardiology Society HF : Heart failure HR : Heart rate HS : Haemorrhagic shock HTN : Hypertension LPM : Lumped parameter model MS : Medical School Q&A : Question and answer SV : Stroke volume UoS : University of Sheffield **Electronic supplementary material** The online version of this article (10.1186/s12909-017-1058-1) contains supplementary material, which is available to authorized users. Not applicable Funding {#FPar1} ======= This work was supported by a grant from the Virtual Physiological Human (VPH) Network of Excellence (NoE): Grant Agreement Number 223920 and the Engineering and Physical Sciences Research Council (EPSRC): Grant Number: R/125661--1. The funders paid no role in the design of the study or collection, analysis, and interpretation of data or in writing the manuscript. Availability of data and materials {#FPar2} ================================== All parameters used for the model, how to access the model used it self, the e-learning platform, the questions used to assess knowledge and the questionnaire used to access the model are freely available, either in the paper or from links given in the paper. The idea for this study was conceived by DRW. DRW, YS and MB created the e-learning platform, with assistance from JF and AN. DRW and MB delivered the e-learning platform and analysed the data from both the questions and questionnaire. DRW wrote the first draft of the paper, in consultation with PVL. All authors read and approved the final manuscript. Ethics approval and consent to participate {#FPar3} ========================================== Ethical approval was granted by the University of Sheffield Medical School (Reference Number 012395) and students gave their verbal consent to participate in the study, as no identifiable data would be used and this was part of routine teaching audit. Consent for publication {#FPar4} ======================= Not applicable Competing interests {#FPar5} =================== The authors declare that they have no competing interests. Publisher's Note {#FPar6} ================ Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

\section{Introduction} Multi-thermal-phase gas configurations are often encountered in astrophysics. For example, interstellar gas is known to consist of several thermally distinct components whose temperatures are in the range $10^2-10^6$\,K, (e.g, McKee \& Ostriker 1977). Gaseous halos of $L^*$ galaxies constitute cool ($10^4$\,K) gas condensations embedded in a hot ($10^6$\,K) volume-filling virialized gas (e.g., Mo \& Miralda-Escude 1996, Chelouche et al. 2007 and references therein). The intracluster medium (ICM) of nearby clusters shows rich structure of cool ionized filaments embedded in a hot X-ray emitting material and extending to large scales (e.g., Conselice et al. 2001, Fabian et al. 2003). A similar structure is also observed around radio-loud quasars (RLQs) where a considerable fraction of the mass of the gaseous nebulae lies in cool clumps of gas (e.g., Crawford \& Fabian 1992, Fu \& Stockton 2006). More recently, it has been realized that many astrophysical outflows are also multiphase with cool condensed material embedded in a hotter and more dilute ambient medium. These systems include: planetary nebulae (e.g., Meaburn et al. 1992), stellar winds (e.g., Bouret et al. 2005), and quasar flows (e.g., Das et al. 2005). The physical mechanisms responsible for multiphase gas configurations in different systems are quite diverse. A partial list includes thermal instability and compression by radiative shocks in hot material, and evaporation and gravitational binding of cool material. While our understanding of the complex physics leading to multiphase structures is incomplete, it is well established that such systems are a wide-spread phenomenon. Recent studies suggest that some multiphase systems seem to suffer from a heat "deficit". One example is that of highly ionized X-ray outflows from the central regions of Seyfert galaxies that are believed to be thermally driven and are seen to reach high velocities despite the effect of adiabatic cooling (e.g., Chelouche \& Netzer 2005). An analogous case to that are the large scale narrow line region (NLR) outflows in active galactic nuclei (AGN) where the effect of adiabatic cooling is even more pronounced (e.g., Everett \& Murray 2007). These discrepancies may be alleviated if a yet unidentified heating source is lurking in those systems which would balance adiabatic losses. Another example is that of cooling flows in which the time for radiative cooling of the hot component appears to be short compared to the Hubble time, yet the gas does not cool significantly. Various explanations have been put forward to explain the cooling flow problem including thermal conduction (e.g., Zakamska \& Narayan 2003), $PdV$ work done by expanding radio bubbles that are inflated by (recently activated) AGN (e.g., Fabian et al. 2003 and references therein; see also Mathews et al. 2006), and the dissipation of gravitational energy via cool condensations falling through the hot gas (Murray \& Lin 2004; see also Dekel \& Birnboim 2007). In this paper we wish to address the general problem of the heating/cooling balance in multiphase media around sources with time-varying ionizing fluxes. Flux variability is a characteristic of many astronomical objects and of active (accreting) systems in particular [see e.g., Geha et al. (2003) for the case of quasars and Zezas et al. (2007) for a study of X-ray binaries]. While the examples given in this paper focus on quasar-related phenomena, we emphasize that the general processes outlined here may apply to other systems such as X-ray binaries and variable stars. This paper is organized as follows: In \S2 we present a general analytic formalism that describes the radiative forcing of clouds which causes them to pulsate and emit acoustic waves into their surrounding medium. Absorption of acoustic radiation and several heat transport mechanisms are also discussed. We address a few specific systems to which this physical mechanism may apply in \S3. Numerical calculations confirming some aspects of our analytic approximations are also presented. Our conclusions follow in \S4. \section{The Emission and Dissipation of Acoustic Waves} \subsection{General Setup} Here we consider a generic configuration of a multiphase medium by assuming dense and cool gas condensations (clouds) with time-averaged temperature $T_c$ embedded in a hot, dilute, volume-filling medium at temperature $T_h$ (see Fig. 1). We assume that the two phases are, in a time-average sense, in pressure equilibrium such that \begin{equation} \rho_h=\rho_c \frac{T_c}{T_h}, \end{equation} where $\rho_c,~\rho_h$ are the densities of the cool and hot component, respectively. Characterizing such an equilibrium requires knowledge of the relevant heating and cooling mechanisms in each phase. These processes are complicated, but we note that, generally, cooling is $\propto \rho^2$; hence the cooling time would be shorter for the dense and cool phase compared to the dilute and hot one. In a steady-state thermal balance, this is also true for the heating rates. Below we show that a fraction of the photon energy which is absorbed by the cool gas can be redirected to heat the hot ambient medium by means of acoustic waves. Such a process naturally occurs when the cool phase is photo-heated by a time-varying radiation field (e.g., that of a quasar). Clouds distributed throughout the volume of the system would change their pressure due to the varying radiative heating and would therefore pulsate. As such, they would act as bells (or monopole loudspeakers) emitting acoustic radiation into their environment. Unlike Murray \& Lin (2004), our model for the heating of the volume-filling hot medium does not require differential velocities between the phases to operate, and does not necessarily lead to the grinding of the cool phase. \begin{figure} \plotone{f1.eps} \figcaption{A schematic (not-to-scale) view of the multiphase medium considered in this work. Clouds (dark circles) are embedded in a hot dilute medium and are roughly in pressure equilibrium with it (see text). The clouds are assumed to be spherical with a radius $R_c$, density, $\rho_c$, and temperature, $T_c$ (see inset). The entire volume is illuminated by an ionizing source of varying flux (upper-right corner) that radiatively forces the clouds to pulsate and emit acoustic waves into their surrounding medium.} \label{geometry} \end{figure} \subsection{Emission} Consider an isolated spherical cloud which is exposed to a source of ionizing flux that varies with time $t$ (Fig. 1). Consider also a specific mode of oscillation such that the flux, \begin{equation} F(t)=F_0+\delta F e^{i\omega t}, \end{equation} where $t_\omega = 2\pi \omega^{-1}$ and $\delta F$ is the amplitude of variation. We further assume that \begin{equation} {\rm max} ( t^c_{\rm cool}, t^c_{\rm heat}) < t_\omega \ll {\rm max} ( t^h_{\rm cool}, t^h_{\rm heat}) \end{equation} with $t^c_{\rm cool},~t^c_{\rm heat}~(t^h_{\rm cool},~t^h_{\rm heat})$ being the cooling time and heating time of the cool (hot) phase, respectively (see \S 3 for the discussion of specific systems). In this regime, the cloud reacts instantaneously to the varying flux by changing its temperature, and hence its pressure (we check the validity of this condition for specific systems in \S 3). So, for example, as the level of ionizing flux rises, more energetic electrons are injected into the gas thereby increasing its temperature. The cloud, being over-pressured compared to its environment, would then expand into the ambient medium. For large enough flux variations ($\delta F/F\sim 1$), the cloud's temperature and pressure ($P$) variations satisfy $\Delta T_c/T_c \sim \Delta P/P\sim 1$ (see \S 3 for detailed photoionization calculations). In this case, the cloud will expand at roughly its sound speed, $c_c$. We neglect any time dependence of $c_c$ in the present analysis as this is proportional to $\sqrt{\Delta P/P}$ which is of order unity. We note that since $t_\omega \ll t^h_{\rm cool}$, the hot phase will not react to flux variations of this duration and its properties may be considered constant over $t_\omega$. The last condition may be relaxed if the hot component is in a range of parameter space where $\partial {\rm ln}(T)/\partial {\rm ln} (F)=0$, as would be the case for gas at the Compton temperature. In what follows, we shall work in the long wavelength regime, namely, $t_\omega \geq R_c/c_h$ where $R_c$ is the cloud radius (see below for a more general definition of $t_\omega$ which takes into account the finite cooling time of the cloud). At much higher frequencies a phase develops between the gas pressure and velocity (as they become orthogonal on small scales) so that the cycle-averaged radiation is small. Acoustic flux is often over-estimated in the literature. When a cloud $-$ or a piston $-$ expands subsonically (indeed $c_c\ll c_h$ by construction in our case), the ambient medium in its immediate environment would be effectively incompressible and would pulsate {\it with} the cloud not producing sound (i.e., compression) waves. Sound waves {\it are} produced once the surrounding medium has not had time to respond to the changing volume of the region interior to it which occurs at a distance $\sim c_ht_\omega$ from the cloud. This distance marks the transition to the radiative zone (see appendix). From the continuity condition, the velocity at that location \begin{equation} \dot{\xi} = c_c \left (1+ \frac{ c_h t_\omega}{R_c}\right )^{-2}. \end{equation} Hence, for large enough $t_\omega$, $\dot{\xi}\ll c_c$. The acoustic luminosity radiated away by a single pulsating cloud would therefore be \begin{equation} L_{\rm acoustic}=2\pi R_c^2 \rho_h c_c^2 c_h \left ( 1+ \frac{ c_h t_\omega}{R_c}\right )^{-2}. \end{equation} Clearly, for large enough $t_\omega$, $L_{\rm acoustic}$ becomes very small. It is interesting to compare the total energy emitted over one cycle to the energy stored in the hot gas within a shell defined by the surface of the cloud at a distance $r=R_c$ and $r=R_c+c_ht_\omega$. As $\dot{\xi}$ declines rapidly with $r$, we can write \begin{equation} \frac{L_{\rm acoustic} t_\omega}{4/3\pi R_c^3 \rho_h c_c^2}\propto \frac{R_c}{c_h t_\omega}. \end{equation} Hence, for long enough wavelengths, this ratio is $\ll 1$ and most of the $PdV$ work done by the cloud by expanding into the ambient medium is returned to it during the compression phase (e.g., Leighton 1997 and the appendix). This is also why typical loudspeakers are rather inefficient in producing sound converting, on average, only $\sim$1\% of the electric power to audible acoustic power. The ratio of acoustic energy emitted by all the clouds over the cooling timescale of the hot phase (from adiabatic or radiative losses) and the thermal energy stored in the hot phase, $E_h$, is \begin{eqnarray} \displaystyle \frac{E_{\rm acoustic}}{E_h} & = & \frac{3}{2} \frac{\epsilon_V}{1-\epsilon_V} \left ( \frac{c_c}{c_h} \right )^2 \frac{c_h t^h_{\rm cool}}{R_c} \left ( 1+ \frac{c_h t_\omega}{R_c} \right )^{-2} \\ \nonumber \displaystyle & \sim & \epsilon_V \left ( \frac{c_c}{c_h} \right )^2 \frac{R_c}{c_h t_\omega} \frac{t^h_{\rm cool}}{t_\omega} \end{eqnarray} with $\epsilon_V$ being the volume filling factor of the cool phase. The last expression is valid in the limit $c_h t_\omega /R_c \gg 1$ and $\epsilon_v \ll 1$. Clearly, a pre-requisite for effective acoustic heating is that $E_{\rm acoustic} \geq E_h$. This is achieved for short $t_\omega$. Nevertheless, $t_\omega$ cannot be arbitrarily short since it must satisfy \begin{equation} t_\omega \geq t_\omega^0={\rm max} \left ( \frac{R_c}{c_h}, \frac{\rho_c c_c^2}{ \Lambda (T,\rho_c,F)} \right ) \end{equation} where $\Lambda$ is the net cooling/heating rate per unit volume averaged over half a cycle. If the converse is true then the clouds' thermal state would change little with time and radiative forcing would be less efficient (this is shown to be the case in \S3). For $t_\omega > t_\omega^0$, $E_{\rm acoustic}/E_h \propto t_\omega^{-2}$, hence $t_\omega^0$ is the mode which is most efficient in converting photon flux to acoustic flux. It is interesting to compare the acoustic luminosity emitted by the cloud to its photo-heating rate \begin{equation} \frac{L_{\rm acoustic}}{4\pi/3 R_c^3 \Lambda} \simeq \frac{R_c}{c_ht_\omega } \left ( \frac{c_c}{c_h} \right )^2 \frac{ \rho_c c_c^2}{\Lambda t_\omega} \leq \left ( \frac{ c_c}{c_h} \right )^2. \end{equation} Clearly, only a small fraction of the photo-heating luminosity is transformed to acoustic luminosity (e.g., at most $\sim$1\% will be converted to acoustic luminosity for $T_c=10^4$\,K and $T_h=10^6$\,K). For the case in which the emitted acoustic radiation is fully absorbed in the system then acoustic heating would overcome the cooling of the hot phase once \begin{equation} \frac{\epsilon_V}{1-\epsilon_V} \frac{\Lambda(\rho_c)}{\Lambda ( \rho_h)} \left ( \frac{c_c}{c_h} \right )^2 \sim 1, \end{equation} where we have used equation 9 and our assumption of pressure equilibrium. If the ratio of the cooling rates for the hot and cold phases is $\sim (\rho_h/\rho_c)^2$ (as would be the case if, for example, the thermal properties of the hot phase are determined primarily by photo-heating; see \S 3) then we require that the mass fraction of the cool component, $\epsilon_M$, satisfies \begin{equation} \epsilon_M \sim \frac{\rho_c}{\rho_h}\epsilon_V \sim 1. \end{equation} This brings us to an important conclusion: {\it the mass of the cool phase should be at least comparable to that of the hot phase for acoustic heating to be effective}. We discuss this result more quantitatively in \S 3.1-3.3. It is possible to analytically estimate the typical size of clouds that can give rise to significant acoustic heating. A natural scale comes from the requirement for a minimal $t_\omega^0$ which occurs for $R_c/c_h=\rho_c c_c^2/\Lambda$ (see equation 8 and the numerical results of \S3). This gives the following order-of-magnitude estimate \begin{equation} {R_c} \simeq \frac{t^c_{\rm cool}}{l/c_h}l, \end{equation} i.e., $R_c/l$ is of order the ratio of the cooling time of the cool phase to the sound-crossing timescale of the hot phase. A specific case worth mentioning is that of adiabatically cooling/heating, thermally-driven systems (such as X-ray and NLR quasar outflows). Here, the thermal state of the hot medium satisfies \begin{equation} P\sim \Lambda \frac{l}{c_h}. \end{equation} That is, the time it take for the entire volume to heat (or cool) is of order the dynamical time. Using this property of such systems, we can readily estimate estimate $R_c$ to be, \begin{equation} R_c \sim \left ( \frac{c_c}{c_h} \right )^4l = \left ( \frac{\rho_h}{\rho_c} \right )^2l \end{equation} [in this case, the required number of clouds in the system is roughly $(\rho_c/\rho_h)^5$]. We emphasize, however, that this provides only a rough estimate for $R_c$ and that, in fact, there is a range of cloud sizes which can induce significant acoustic heating, as we demonstrate for a few specific systems in \S3. \subsection{Absorption} The emission of acoustic energy by a cool cloud does not guarantee that the energy will dissipate over the relevant spatial scales required to effectively heat up the volume. Dissipation of acoustic energy in the hot medium occurs via absorption (and conductance) as well as by steepening of acoustic waves into weak shocks. The latter process is negligible in our case since the clouds pulsate subsonically and the wave amplitude is quickly diminished in the spherical case (c.f., Stein \& Schwartz 1972). The relevant dissipation length scale due to absorption of radiation is (e.g., Mathews et al. 2006) \begin{equation} l_d\simeq \mu^{-1}P c_h \left ( \frac{t_\omega}{2\pi} \right) ^2, \label{ldiss_a} \end{equation} where $\mu$ is the {\it effective} viscosity (including both the effects of viscosity and conduction; e.g., Fabian et al. 2005). Here we take the usual Braginskii-Spitzer (e.g., Spitzer 1962) value for high temperature solar composition gas $\sim 10^{-16} \eta T_h^{5/2}~{\rm g~cm^{-1} s^{-1}}$ (e.g., Lang 1999) which is applicable to e.g., the ICM for $\eta=10^{-2}-1$ (Narayan \& Medvedev 2001, Fabian et al. 2005; but see Ettori \& Fabian 2000 for possibly much lower values). Due to the lack of observational constraints on the conductivity of other systems (e.g., quasar outflows), we shall assume values similar to those of the ICM. Rewriting equation 7 with dissipation included, we obtain a generalized condition for the importance of acoustic heating, \begin{equation} \frac{E_{\rm acoustic}^{\rm dissipated}}{E_h}=\frac{E_{\rm acoustic}}{E_h} \left ( 1-e^{-l/l_d} \right ) \gtrsim 1. \end{equation} For $l_d \gg l$, absorption is inefficient and most of the acoustic energy (a fraction $\sim 1-(l/l_d)$ of it) will escape the system. Absorption is efficient over the entire volume for $l_d \sim l$. For $l_d \ll l$, absorption is very efficient yet is localized to the environment of the clouds. In this case several additional conditions are required to effectively heat up the entire volume. These are discussed below. \subsubsection{The case of $l_d \ll l$} For volumetric heating to be important, we require many clouds which are homogeneously distributed throughout the volume of the system and/or the presence of some heat transport mechanism. For clouds distributed randomly in space (as indeed seems to be observationally supported; see \S 1), the characteristic distance between clouds is $\simeq \epsilon_V^{-1/3}R_c$. If no heat transport mechanism is operating then we require \begin{equation} \epsilon_V^{1/3}\frac{l_d}{R_c} \geq 1 \end{equation} so that dissipation of acoustic waves due to all clouds occurs throughout the volume. A much less restrictive case is that in which heat conduction and/or advection are present. In this case, even if the above condition is not satisfied, heat could still be distributed over the entire volume. Heat conduction is due to free electrons whose velocity is, $v_e\simeq \sqrt{3k_BT_e/m_e}\simeq 6\times10^3 \sqrt{T_h/10^6\,{\rm K}}~{\rm km~s^{-1}}$. The propagation of electrons through a plasma takes the form of a random walk with some mean free path, $\lambda_e$. The largest uncertainty arises in estimating $\lambda_e$, which depends on the strength and configuration of the magnetic field, $B$, as well as on the pressure and density of the (hot) medium. If $B=0$ then $\lambda_e = \lambda_e^{B=0} \simeq 10^4T_h^2/(\rho_h/m_H)$\,cm (where $m_H$ is the mass of a hydrogen atom; see Cowie \& McKee 1977). When $B>0$ the medium becomes anisotropic with $\lambda_e$ being of order the Larmor radius across magnetic field lines. For a detailed discussion of these issues see Lazarian (2006, and references therein). Here we assume that $10^{-2}<\lambda_e/\lambda_e^{B=0}<1$ (Narayan \& Medvedev 2001). Hot electrons must traverse a distance of order the typical cloud separation over $t_{\rm cool}^h$, so that heat could be uniformly distributed throughout the volume. This results in the following condition for efficient conduction: \begin{equation} \frac{\epsilon_V^{-2/3} R_c^2}{v_et_{\rm cool}^h \lambda_e} \lesssim 1. \end{equation} In addition to conductance, heat may be advected by e.g., turbulent motion in the gas (we do not consider here the additional heating by the dissipation of turbulent energy; e.g., Chelouche \& Netzer 2005). For a Kolmogorov-type transonic turbulence, the velocity of eddies of size $l_e$ is $v_e=c_h(l_e/l)^{1/3}$ and for efficient advection we therefore require \begin{equation} \frac{\epsilon_V^{-1/3} l^{1/3} R_c^{2/3}}{c_h t_{\rm cool}^h} \lesssim 1. \end{equation} \subsection{Caveats} We have outlined a general analytic framework for studying the emission of acoustic waves by pulsating cool, dense clouds and the dissipation of such waves in a hot surrounding medium. We have not addressed the issue of cloud longevity, which could be important as clouds are prone to various hydrodynamic instabilities such as Kelvin-Helmholtz and Rayleigh-Taylor, as well as evaporation/condensation. Our understanding of these processes is rather limited in the general astrophysical context as well as in the specific systems considered below, and we make no attempt to further address this here. That said, the ubiquity of multiphase structures in astrophysical systems implies that such configurations are, at least statistically, quasi-steady-state phenomena. A further caveat concerns the role of magnetic fields in determining the propagation of sound waves in the medium. Better estimates for the magnitude of (magneto-)acoustic heating will require not only detailed numerical simulations but also a better understanding of the individual systems to which this process may apply. \begin{figure} \plotone{f2.eps} \figcaption{Time-dependent calculations of the temperature of a cloud which is exposed to a time varying flux of a sinusoidal form (denoted by a dotted line; see legend). All quantities are shown relative to their mean values. $t_\omega$ was varied in the range $(0.1-100)\times t^c_{\rm cool}$ as calculated from the steady-state solution. As expected, the cloud thermal state follows the flux variations for $t_\omega \gg t^c_{\rm cool}$. A phase develops and the temperature variation amplitude is somewhat lower for $t_\omega \sim t^c_{\rm cool}$. The cloud's temperature variation is quenched for $t_\omega \ll t^c_{\rm cool}$. These calculations justify our approximation of the heating/cooling rates by their steady-state values (see \S 3).} \label{time_d} \end{figure} \section{Applications to Astrophysical Systems} \begin{figure*} \includegraphics[width=7in]{f3.eps} \figcaption{Contour level plots of the ratio of acoustic heating to cooling as a function of $\epsilon_V$ and $R_c$ (normalized to the size of the system, $l$) for the systems considered in this paper (\S3.2-3.4). For any $R_c$, it is assumed that there exists a mode with $t_\omega = t_\omega^0$ for which acoustic heating is evaluated. Contours with values $\geq 1$ indicate regions in the parameter space where acoustic heating dominates over cooling in the hot phase. Three astrophysical systems are considered: X-ray outflows in Seyfert I galaxies ({\it left}), extended NLR outflows in Seyfert galaxies ({\it middle}), and cooling flows ({\it right}). Color coding of contours is the same for all panels. We note that the contours' shape near $\epsilon_V=1$ is somewhat uncertain due to the relatively little volume occupied by the hot material and the importance of absorption and scattering of acoustic waves from the cool phase in this case. Clearly, for all systems there exist a wide range in the parameter space where acoustic heating is important (or even dominant; see text). This occurs when the cool gas mass is comparable to or larger than that of the hot component. Smaller clouds are more effective in heating the system as long as $R_c/c_h$ is greater than their cooling time. For quasar outflows, the results for the cloud size and the minimal $\epsilon_V$, where acoustic heating becomes important, are consistent with analytic scalings (see \S 3) and some observational constraints (see text).} \label{phase_diagram} \end{figure*} In this section we examine how acoustic heating can be important in astrophysical environments. It is not the purpose of this paper to cover all possible cases but rather to study the applicability of the process to a few specific systems. In all cases considered below there is either an active ionizing source with a varying flux or one has recently been active. Here we consider examples pertaining to quasar outflows and to cooling flows. To this end, we have taken fiducial parameter values from the literature concerning the physical properties of the cool and hot phase but note the appreciable uncertainties associated with them. These parameter values include rough estimates for $T_c,~T_h$, and $\rho_c$. Estimates for the total size of the system, $l$, and its cooling time, $t_{\rm cool}^h$ are also taken from the literature. The characteristic values of $\epsilon_V$ and $R_c$ are very poorly constrained in those systems and are therefore treated as free parameters. A major difficulty when estimating the response of the cloud to radiative forcing is to properly assess the net heating/cooling function $\Lambda$ (equation 8) which depends on whether the gas is in photoionization and thermal equilibrium at all times (as would be the case if the heating and cooling timescales are the shortest in the problem) or whether its thermal state depends on its history in some intricate way. To keep the problem tractable, we estimate the net heating/cooling function by the value of the cooling function under steady-state conditions (in which case its values match those of the heating function). For solar metalicity, optically-thin gas exposed to a type-I quasar continuum (Sazonov et al. 2005), the heating rate does not have a strong dependence on the temperature, varying by a factor of order unity in the range $10^4<T_c<10^6$\,K, so that an adequate approximation for our purpose is \begin{equation} \Lambda (T,\rho)\simeq 10^{-23}(\rho_c/m_H)^2~{\rm erg~s^{-1}~cm^{-3}}. \end{equation} This value shall be used to obtain an order of magnitude estimate for the heating/cooling timescale of the cloud (equation 8). To check the validity of the above approximation for $\Lambda$ in cases where strict thermal equilibrium is not justified, we have calculated several time-dependent photoionization models using {\sc cloudy} c07.02 (Ferland et al 1998) and assuming a light-curve, of a fixed amplitude, varying the mode frequency, i.e., $t_\omega$. Ischoric conditions are assumed for the calculations since $c_c/c_h \ll 1$ and we focus on the threshold of the long wavelength regime (see below). Calculations were carried out for the case of $\rho_c/m_H=10^5\,{\rm cm^{-3}}$ and placing the cloud at 1\,pc from the quasar so that its steady-state temperature $-$ if it were exposed to a constant flux $F$ $-$ is $\sim 8\times 10^4$\,K (corresponding roughly to X-ray outflows temperatures; see \S 3.1). The results, shown in Figure. 2, indicate that our estimate for the net cooling function (i.e., $t^c_{\rm cool}$) using steady-state values is indeed reasonable. In particular, the cloud does not react to the varying flux level of the source for $t_\omega \ll t^c_{\rm cool}$ and radiative forcing would be inefficient. The cloud temperature does follow the light-curve for $t_\omega \sim t^c_{\rm cool}$, albeit with a phase, and its temperature fluctuations, $\Delta T_c$, are almost as large as those obtained for $t_\omega \gg t^c_{\rm cool}$ (where little or no phase is present since the cloud reacts instantly to the source's flux level). We note here that $\Delta T_c$ depends not only on $\Delta F$ but also on $\partial {\rm ln}(T)/\partial {\rm ln} (F)$: in particular, for the chosen ionizing continuum, large flux variations are expected for $T_c \sim \,{\rm a~few}\times 10^4-10^5$\,K. The proper treatment of marginally stable or thermally unstable regions (Krolik \& Kriss 2001) is beyond the scope of this paper. In the above, we have not addressed the possibility of changes to the spectrum of the ionizing source which are known to occur (e.g., Netzer et al. 2003) and could influence the gas temperature even when the total integrated flux remains unchanged. Such a case can be easily accommodated within our general formalism (\S 2) by considering $F$ to be the relevant ionizing flux level (e.g., that just above the Lyman edge for $T_c \sim 10^4$\,K gas or around soft X-ray energies for $T_c \sim 10^5$\,K gas). As spectral variations of active sources are not well constrained by observations and their inclusion will add little to our understanding of the general effect, we choose to not address this issue here. Quasars are known to vary considerably ($\delta F/F \sim 1$) over a wide range of timescales, from minutes to years (e.g., Giveon et al. 1999, Netzer et al. 2002 and references therein). While current observations do not directly constrain the power of such variations on the timescales and wavebands most relevant to our study, it is reasonable to assume that the ionizing flux of quasars varies considerably over the entire relevant frequency range. This motivates us to adopt the following approach: we assume that for any choice of system parameters (e.g., cloud size, density, and temperature), there exists a mode in the quasar light-curve oscillating with $t_\omega \simeq t_\omega^0$ and having a large amplitude of flux variation ($\delta F/F \sim 1$) . Our calculations indicate that, for all systems considered here, uniformly distributed heating by acoustic waves is possible so that one (or more) of equations $17-19$ is always satisfied when the more restrictive, equation 16 is. Generally, smaller clouds (shorter $t_\omega$) occupying a large fraction of the volume give rise to more efficient heating. Nevertheless, as clouds cannot be infinitely small, on account of their cooling time being longer than $t_\omega$ (equation 8), there is a intermediate range of cloud sizes with a total volume above a certain threshold, which is most efficient in heating the ambient medium. The details depend on the system in question as we now discuss. \subsection{X-ray Outflows in Seyfert 1 Galaxies} The parameters for X-ray gaseous outflows from Seyfert 1 galaxies which we adopt here are: $T_c=8\times 10^4\,{\rm K},~T_h=10^7\,{\rm K},~\rho_c/m_H=3\times 10^5~{\rm cm^{-3}},~l=3\,{\rm pc}$ and $t_{\rm cool}^h=10^4$\,years (Chelouche \& Netzer 2005). Considering equation 16, there is a large range of the parameter space in which acoustic heating is considerable and can balance adiabatic cooling (see Fig. 3). In particular, such heating would be effective - that is, comparable to radiative heating - for $\epsilon_V \gtrsim 10^{-2}$, i.e., once the mass in the cool phase is comparable or larger than that of the hot phase (c.f. equation 11). Cloud sizes should be in the range $10^{14}-10^{15}~{\rm cm}$ ($R_c/l \sim 10^{-4}$; see equation 14) for acoustic heating to be effective. Such values are consistent with full occultation of the X-ray source in Seyfert 1 galaxies (e.g., Chelouche \& Netzer 2005). We note that if $\epsilon_V\simeq 0.1$ (i.e., $\epsilon_M \gtrsim 10$) then acoustic heating may even dominate over photo-heating resulting in a hot-phase temperature that would exceed the Compton temperature. \subsection{Extended NLR Outflows in Seyfert Galaxies} Little is known about the physical properties of NLR flows, and the existence of a hot medium is mainly inferred from the dynamics of the low temperature, optically-detected gas. We take fiducial values of $T_c=10^4\,{\rm K},~T_h=10^6\,{\rm K}$ and consider multiphase gas on $\sim 30$\,pc scales (e.g., Das et al. 2005, Everett \& Murray 2007). For luminosities typical of type-I Seyfert galaxies, the implied density of the hot phase on such scales is of order $10\,{\rm cm^{-3}}$ (Everett \& Murray 2007; see also Chelouche \& Netzer 2005) hence $\rho_c/m_H=10^3~{\rm cm^{-3}}$. Here, the dynamical time is of order the adiabatic-cooling time which is of order $10^5$\,years (Everett \& Murray 2007). In this case we find a wide range in parameter space where acoustic heating can be important. Everett \& Murray (2007) quote values for the filling factor of the cool phase as high as $10^{-2}$. Figure 3 shows that, in this case, the emission of acoustic energy by clouds with $R_c \sim 10^{16}~{\rm cm}$ could contribute significantly to the heating-cooling balance of the hot phase. Clouds of this size cannot be individually resolved by current observations yet similar cloud sizes have been inferred from density and column density measurements of UV absorber properties on similar scales (e.g., Kraemer et al. 2001 and references therein). Arguments applying to X-ray outflows (equation 18,19) apply here too. \subsection{Cooling Flows in Galaxy Clusters} Observations of RLQs show the presence of intrinsically bright O\,II and O\,III emission line regions with apparent sizes of a few$\times 10$\,kpc that consist of numerous unresolved cool clouds (e.g., Fu \& Stockton 2006 and references therein). It is thought that at least some RLQs and radio galaxies inhabit the cores of galaxy clusters that possess massive cooling flows toward their centers (e.g., Crawford \& Fabian 1992). In this case, the optical emission is thought to originate from cool clouds embedded within the hot medium (Crawford \& Fabian 1992). It is not known whether such multiphase gas configurations are a property of all cooling flows in galaxy clusters; nor is it clear if all cooling flow clusters which do not cool effectively contain active or recently active nuclei (though they do seem to have a $>50$\% duty-cycle for jet activity; Fabian \& Sanders 2006). Nevertheless, we can use typical cloud parameters inferred from optical observations and ambient medium parameters deduced from X-ray studies: $T_c=2\times 10^4\,{\rm K},~T_h=10^7\,{\rm K},~\rho_c/m_H=10~{\rm cm^{-3}},~l=5\times 10^4\,{\rm pc}$ and $t_{\rm cool}^h=7\times 10^9$\,years (see e.g., Dunn \& Fabian 2006 and Fu \& Stockton 2006). Figure 3 shows that for $\epsilon_V\gtrsim 10^{-3}$ (indicating comparable mass in the hot and cool phase; c.f. Fu \& Stockton 2006), clouds of size $\sim 0.01-1$\,pc could efficiently balance radiative cooling in those systems. Interestingly, similar cloud sizes have been deduced by Hamann et al. (2001) for the clouds seen in absorption toward quasars. Larger values of $\epsilon_V$ (hence $\epsilon_M$) would be required for efficient acoustic heating if the combined duty-cycle of intracluster active sources is much smaller than unity (Shen et al. 2007). Given the uncertainties associated with the duty-cycle of quasars in cooling flow environments and the properties of the cooling flow itself (e.g., concerning its multiphase structure), it is difficult to assess just how important acoustic heating is, in the form considered here. \section{Conclusions} In this paper we have presented a simple model for estimating the effect of acoustic heating by radiatively-forced pulsating clouds on the thermal state of a hot ambient medium. We find that smaller clouds are more efficient acoustic emitters as long as their cooling time is longer than the oscillation period. Acoustic heating is significant when the mass of cool clouds is comparable to, or larger than that of the hot phase. In particular, for large enough mass in cool gas, acoustic heating may even dominate over photo-heating. This implies the possible existence of higher temperature gas around sources with variable flux compared to non-variable objects with similar characteristics. This work demonstrates that heating of multiphase photoionized gas may depend not only on the mean flux of the source but also on deviations from it, as well as on the structure of the medium itself. We show that there exists a range of clouds sizes which is most efficient in emitting acoustic energy. Applying this model to quasar outflows and cooling flows, we find that acoustic heating could be important in these cases if the typical size of cool clouds is of order $10^{-4}$ the size of the system. Better understanding of the multiphase nature of photoionized environments in terms of the relevant densities, cloud sizes, and cooling rates, as well as the properties of the ionizing source (i.e., its power spectrum of variations) are required to properly assess this effect in the broader astrophysical context. \acknowledgements This work is supported by NASA through a Chandra Postdoctoral Fellowship award PF4-50033. It is a pleasure to thank David Bowen, Shane Davis, and Bill Mathews for a careful reading and debugging of an earlier version of this paper. We thank P. Goldreich, A. K\"{o}nigl, and R. Langer for fruitful discussions, H. Netzer for valuable comments, and G. Ferland and collaborators for the latest version of {\sc cloudy}. Many enjoyable auditioning hours of both "monopole" and dipole loudspeakers with David Bowen and Matt Kleban are greatly appreciated.

\section{Introduction} Einstein gravity explains the observed universe fairly well. It has however some theoretical drawbacks, most important of which is the problem of initial singulariry\cite{HE,Wald}. This problem is usually interpreted to imply the limit of applicability of Einstein gravity. There are two possible ways to approach this problem. One is to modify the theory of gravity in the classical framework. The most popular modified theory is the higher-curvature gravity(HCG) theories \cite{Nariai,NT,EM}. Another is to quantize gravity. Quantum gravity has a long history but its completion seems to require still a long time. HCG is also required in quantum field theory in curved spacetime\cite{BD}, string perturbation theory\cite{Zwie}. Recently HCG is applied to cosmology to explain, e.g. inflation\cite{Staro} or expansion of the present universe\cite{CDTT} which seems to be in the stage of an accelerated expansion\cite{Boom,SN,CMB}. In the former a term proportional to the square of the scalar curvature is added to the Lagrangian density of Einstein gravity\cite{Staro,NO} and in the latter a term proportional to the inverse of the scalar curvature is added\cite{CDTT}. In the first order Palatini formalism, a wider class of HCG's is also applied to cosmology to explain the present expansion of the universe.\cite{1st}\\ Thus the reality of HCG is increasing. In this work we adress the problem in the canonical formalism of $f(R)$ type HCG and propose a consistent formalism. The curvatures contain the second order time derivatives of the metric which in HCG cannot be removed by partial integration as in Einstein gravity. Thus HCG is a theory with higher order time derivatives(HDT). Standard procedure for the canonical formalism of HDT has been given by Ostrogradski\cite{Ost}. In the usual method of canonical formalism of gravity in terms of ADM variables, this method however is not applicable directly. The reason is that the curvatures depend on the time derivatives of the lapse function and the shift vector so that they obey field equations leading to the breaking of general covariance. This problem is resolved by the method of Buchbinder and Lyakhovich(BL)\cite {BL,EKKSY} which generalizes the choice of the new generalized coordinate corresponding to the time derivatives of the original generalized coordinates. Applied to the HCG, the new generalized coordinetes corresponding to the time derivatives of the metric was chosen to be the extrinsic curvature. The time derivatives of the lapse and shift are absorbed in the time derivatives of the extrinsic curvature. However, in this method, a change of the original generalized coordinates induces a change of Hamiltonian as will be explained below, which is not the case in theories without higer order time derivatives. For example, in the case of FRW spacetime, the Hamiltonian is different whether we use the scale factor or its logarithm as the generalized coordinate. We propose a canonical formalism which resolves this problem. We combine the advantageous points of both the method of Ostrogradski and BL. We choose the extrinsic curvature as the new generalized coordinate as in BL and define the momenta canonically conjugate to them similarly to Ostrogradski's method. The extrinsic curvature is the Lie derivative of the three metric which reduces to the time derivative in flat space. The resulting Hamiltonian is shown in fact invariant under the tarnsformations of the original generalized coordinates which do not change the 3-dimensional metric of the hypersurface of constant time. In section 2, we explain how the Hamiltonian changes under the transformation of the original generalized coordinate in the method of BL. Section 3 is devoted to the presentation of a new canonical formalism for a HCG of $f(R)$-type. In section 4 we demonstrate the invariance of the Hamiltonian under the two kinds of transformations of the generalized coordinates, the three metric. Summary and discussion are given in section 5. \section{A problem in the method of Buchbinder and\\ Lyakhovich} In this section we show that in the method of BL the Hamiltonian changes under the transfornation of the generalized coordinates using a simple model. Let us consider a system described by a Lagrangian $$ L=L(q^i,\dot{q}^i,\ddot{q}^i) \eqno(2.1) $$ in which case the new generalized coordinates are defined as $$ Q^i\equiv \dot{q}^i. \eqno(2.2) $$ In the method of BL, the Lagrangian is modified using the Lagrange multiplier method so that the definition (2.2) is derived from variational principle. Denoting the modified Lagrangian as $L^{*}$, it is given as $$ L^{*}\equiv L+p_{i}(\dot{q}^i-Q^i),\ \ \ L=L(q^i,Q^i,\dot{Q}^i). \eqno(2.3) $$ The multipliers $p_{i}$ are the momenta canonically conjugate to $q^i$. The canonical formalism is obtained by the Legendre transformation starting from $L^{*}$. Momenta $P^i$ canonically conjugate to $Q_{i}$ are given by $$ P^i\equiv {\partial L^{*}\over \partial\dot{Q}_{i}}={\partial L\over \partial\dot{Q}_{i}}. \eqno(2.4) $$ Then the Hamiltonian $H^{*}$ is given as $$ H^{*}\equiv p_{i}\dot{q}^i+P_i\dot{Q}^i-L^{*} =p_{i}Q^i+P_i\dot{Q}^i-L. \eqno(2.5) $$ Now we make transformations of the generalized coordinates: $$ \phi^i=f^i(q^j),\ \ \ {\rm or}\ \ \ q^i=g^i(\phi^j). \eqno(2.6) $$ Momenta conjugate to $\phi^i$ are denoted as $\pi_{i}$, new generalized coordinates as $\Phi_{i}$ and momenta canonically conjugate to them as $\Pi_{i}$. These variables are not uniquely related to old variables. Since $Q^i$ are related to $\dot{q}^i$ through the variational principle, their relations to $\Phi^i$ are not fixed a priori. The solutions are, however, $Q^i=\dot{q}^i$, it is natural to assume the transformation of $Q^i$ to be the same as $\dot{q}^i$. Differentiation leads to $$ \dot{q}^i={\partial g^i\over \partial\phi^j}\dot{\phi}^j\equiv {\partial g^i\over \partial\phi^j}\Phi^j, \eqno(2.7) $$ so we have the following transformation $$ Q^i={\partial g^i\over \partial\phi^j}\Phi^j. \eqno(2.8) $$ Similarly, the momenta $p_{i}$ are Lagrange multipliers, so that their transformations are also not fixed a priori. We fix their transformations by requiring that the modified Lagrangians are the same: $$ L^{*}(q^i,Q^i,\dot{Q}^i)=L^{*}(\phi^i,\Phi^i,\dot{\Phi}^i) =\bar{L}(\phi^i,\Phi^i,\dot{\Phi}^i)+\pi_{i}(\dot{\phi}^i-\Phi^i). \eqno(2.9) $$ Then we have $$ p_{i}(\dot{q}^i-Q^i)=\pi_{i}(\dot{\phi}^i-\Phi^i). \eqno(2.10) $$ Using the first equation of (2.7) and (2.8) in (2.10), we have $$ \pi_{i}={\partial g^j\over \partial\phi^i}p_{j}. \eqno(2.11) $$ The Hamiltonian, obtained by the Legendre transformation of the right hand side of (2.9), is expressed as $$ \bar{H}^{*}\equiv \pi_{i}\dot{\phi}^i+\Pi_i\dot{\Phi}^i-\bar{L}^{*} =\pi_{i}\Phi^i+\Pi_i\dot{\Phi}^i-\bar{L}. \eqno(2.12) $$ Since $L=\bar{L}$, the change of the Hamiltonian is $$ \Delta H^{*}=\bar{H}^{*}-H^{*}=\Pi_{i}\dot{\Phi}^i-P_{i}\dot{Q}^i. \eqno(2.13) $$ Now $$ \Pi_{i}\equiv{\partial \bar{L}\over \partial \dot{\Phi}^i} ={\partial L\over \partial \dot{\Phi}^i}=P_{j}{\partial g^j\over\partial\phi^i} \eqno(2.14) $$ and $$ \dot{\Phi}^i ={\partial^2 f^i\over \partial q^j\partial q^k}\dot{q}^kQ^j +{\partial f^i\over \partial q^j}\dot{Q}^j. \eqno(2.15) $$ Using (2.15) and (2.16) in (2.13), we have $$ \Delta H^{*}={\partial^2 f^i\over \partial q^k\partial q^j}Q^kQ^j\Pi_{i}. \eqno(2.16) $$ Thus the Hamiltonian changes under the transformation of the original generalized coordinates and reasonable transformation of other canonical variables. If the Hamiltonian represents the energy of the system, which is often the case, this is unreasonable physically. The difference depends on the variables characteristic to higher derivative theory. We note that if the variables transform to make the Hamiltonian invariant, the Lagrangian is changed. In order to resolve this problem, we propose a new canonical formalism which can be interpreted to be the generalization of the Ostrogradski's method. \section{A canonical formalism in terms of Lie derivatives} The idea is simple. As noted above, the problem of Ostrogradski's method was resolved by the method of BL where the choice of the generalized coordinates corresponding to the time derivatives of the original generalized coordinates is extended to almost arbitrary functions of these variables. In the application to gravity, the extrinsic curvature is chosen as the new generalized coordinates. It is the Lie derivatives of the metric which is the generalized coordinates. Lie derivatives reduce to the time derivatives for flat space. So this choice can be thought of the simple generalization of the Ostrogradski's choice. Thus we move to the canonical formalism by the Ostrogradski transformation contrary to the Legendre transformation with the modified Lagrangian as in the method of BL. Ostrogdadski transformation is carried out as follows. Let us consider a system described by a Lagrangian containing the $n$-th time derivatives of the generalized coordinates $q^i\;(i=1,\ldots,N)$. $$ L=L(q^i,\dot{q}^i,\ldots,q^{i(n)}) \eqno(3.1) $$ Take the following variation of the action $$ \delta S\equiv \int_{t_{1}+\delta t_{1}}^{t_{2}+\delta t_{2}}L(q^i+\delta q^i,\ldots, q^{i(n)}+\delta{q}^{i(n)})dt -\int_{t_{1}}^{t_{2}}L(q^i,\ldots,q^{i(n)})dt \eqno(3.2) $$ where $\delta q^i$ are decomposed as the sum of the variations of the function $q^i(t)$, which we denote by $\delta^{*}q^i$, $\delta^{*}q^i\equiv (q+\delta q)(t)-q(t)$, and those due to the change of the time , $q(t+\delta t)-q(t)=\dot{q}^i\delta t$: $$ \delta q^i=\delta^{*}q^i+\dot{q}^i\delta t \eqno(3.3) $$ where the second terms are assumed to contribute only near the end points. Then the above variation is written, retaining only the first order terms in small quantities, as $$ \delta S =\int_{t_{1}+\delta t_{1}}^{t_{1}}L(q^1,\ldots, q^{i(n)})dt +\int_{t_{2}}^{t_{2}+\delta t_{2}}L(q^1,\ldots, q^{i(n)})dt +\delta^{*}\int_{t_{1}}^{t_{2}}Ldt. \eqno(3.4) $$ Using the approximation $$ \int_{t_{k}}^{t_{k}+\delta t_{k}}L(q^1,\ldots, q^{i(n)})dt =L\left(q^1(t_{k}),\ldots, q^{i(n)}(t_{k})\right)\delta t_{k}, \ \ (k=1,2) \eqno(3.5) $$ the first two terms in (3.4) become the following $$ \left[L\delta t\right]_{t_{1}}^{t_{2}}\equiv \delta S_{1}. \eqno(3.6) $$ Rearranging the sum in the third term of (3.4), we have $$ \delta S_{2}\equiv \delta^{*}\int_{t_{1}}^{t_{2}}Ldt =\left[\delta F\right]_{t_{1}}^{t_{2}}+\delta^{*}S_{2} \eqno(3.7) $$ where $$ \delta F=\sum_{i=1}^N\Biggl[\sum_{s=0}^{n-1}\Biggl\{\sum_{r=s+1}^n (-1)^{r-s-1}D^{r-s-1}\Bigl({\partial L\over \partial(D^rq^i)}\Bigr)\Biggr\} \delta^{*} q^{i(s)}\Biggr] \eqno(3.8) $$ and $$ \delta^{*} S_{2}=\sum_{i=1}^N\int_{t_{1}}^{t_{2}}\sum_{s=0}^{n}(-1)^sD^s\Bigl( {\partial L\over \partial(D^sq^i)}\Bigr)\delta^{*}q^idt, \eqno(3.9) $$ where $D$ represents the time derivatine, i.e. $\displaystyle D\equiv {d\over dt}$. $\delta^{*}S_{2}$ vanishes if we require the variational principle. The new generalized coordinates are taken as $$ q_{s}^i\equiv D^sq^i \eqno(3.10) $$ and the momenta canonically conjugate to these cordinetes are defined by the coefficients of the variations of these coordinates in $\delta F$ $$ p_{i}^s\equiv \sum_{r=s+1}^n\left[(-1)^{r-s-1}D^{r-s-1}\Bigl({\partial L\over \partial D^rq^i}\Bigr)\right]. \eqno(3.11) $$ The Hamiltonian is defined as $(-1)\times$ the coefficient of $\delta t$ in $L\delta t+\delta F$ after using (3.3) $$ H=\sum_{i=1}^N\sum_{s=0}^{n-1}p_{i}^{s}Dq^{i}_{(s)}-L. \eqno(3.12) $$ Note that for $s=n-1$, (3.11) has a simple expression $$ p_{i}^{n-1}={\partial L\over \partial\dot{q}_{n-1}^i}. \eqno(3.13) $$ Thus the Ostrogradski transformation is a generalization of the Legendre transformation. It is noted that the highest order derivatives need not be the same for all $i$. In this work we investigate the higher-curvature gravity(HCG) of $f(R)$ type in which the Lagrangian density is given by a function of the scalar curvature $$ {\cal L}=\sqrt{-g}f(R) \eqno(3.14) $$ where $g\equiv \det g_{\mu\nu}$. The action is given as usual: $$ S=\int_{t_{1}}^{t_{2}}Ldt,\ \ \ L=\int{\cal L}d^3x \eqno(3.15) $$ We adopt the ADM variables $N({\bf x},t)$, $N^i({\bf x},t)$ and $h_{ij} ({\bf x},t)$ with respect to a hypersurface $t=constant(\Sigma_{t})$ as the generalized coordinates. In terms of these variables, the scalar curvature is expressed as follows $$ R =2N^{-1}h^{ij}\partial_{0}K_{ij}-U-2N^{-1}\Delta N -2N^{-1}(N^k\partial_{k}K+2N^{i;j}K_{ij}). \eqno(3.16) $$ Here $K_{ij}$ is the extrinsic curvature of $\Sigma_{t}$ $$ K_{ij}={1\over 2N}\left(\partial_{0}h_{ij}-N_{i;j}-N_{j;i}\right), \eqno(3.17) $$ and $$ U\equiv 3K_{ij}K^{ij}-K^2-\tilde{R}, \eqno(3.18) $$ where $K\equiv h^{ij}K_{ij}$ and $\tilde{R}$ is the scalar curvature of $\Sigma_{t}$. The determinant $g$ is expressed as $-N^2h$ with $h\equiv\det h_{ij}$. In this case $\delta F$ takes the following form $$ \begin{array}{ll} \displaystyle \int d^3x\Biggl[& \displaystyle {\partial {\cal L}\over \partial(\partial_{0}h_{ij})}\delta^*h_{ij} -\partial_{0}\left({\partial {\cal L}\over \partial(\partial_{0}^2h_{ij})}\right)\delta^*h_{ij} +{\partial {\cal L}\over \partial(\partial_{0}\partial_{k}h_{ij})}\delta^{*}\partial_{k} h_{ij} +{\partial {\cal L}\over \partial(\partial_{0}^2h_{ij})}\delta^{*}\partial_{0}h_{ij} \\[5mm] &\displaystyle +{\partial {\cal L}\over\partial(\partial_{0}N)}\delta^{*} N +{\partial{\cal L}\over \partial(\partial_{0}\partial_{k}N^i)}\delta^{*} \partial_{k}N^i +{\partial L\over\partial(\partial_{0}N^i)}\delta^{*} N^i\;\Biggr]. \end{array} $$ From this expression, it appears that $N$ and $N^i$, which represent the choice of the coordinate system, have the momenta canonically conjugate to them. In that case, they are allowed to be only the solutions of dynamical equations and the general covariance is broken. However, the time derivatives, $\partial_{0}^2h_{ij}$, $\partial_{0}N$, $\partial_{0}N^i$ and $\partial_{0}\partial_{k}N^i$ are involved only through $\partial_{0}K_{ij}$. It is also noted that time derivatives $\partial_{0}h_{ij}$ and $\partial_{0}\partial_{k}h_{ij}$ are involved through $\partial_{0}K_{ij}$, $K_{ij}$ and $\partial_{k}K_{ij}$. Thus $\delta F$ reduces to the integration of $$ \begin{array}{l} \displaystyle {\partial{\cal L}\over\partial(\partial_{0}K_{ij})}\delta^{*}K_{ij} +\displaystyle \Biggl[{\partial{\cal L}\over \partial(\partial_{0}h_{ij})} -\partial_{0}\left({\partial{\cal L}\over \partial(\partial_{0}^2h_{ij})}\right) -\partial_{k}\left({\partial{\cal L}\over \partial(\partial_{0}\partial_{k}h_{ij})}\right)\\[7mm] \hspace*{4cm}\displaystyle -{\partial{\cal L}\over \partial(\partial_{0}K_{kl})}{\partial K_{kl}\over \partial h_{ij}} +\partial_{m}\left({\partial{\cal L}\over \partial(\partial_{0}K_{kl})}{\partial K_{kl}\over \partial(\partial_{m}h_{ij})}\right)\Biggr]\delta^{*}h_{ij} \end{array} \eqno(3.19) $$ where we use the relations, e.g. $$ {\partial(\partial_{0}K_{kl})\over \partial(\partial_{0}N)}={\partial K_{kl}\over \partial N}. $$ From this expression, if we adopt $K_{ij}$ as the new generalized coordinate, instead of $\partial_{0}h_{ij}$, time derivatives of $N$ and $N^i$ are absorbed in that of $K_{ij}$ and the restriction on $N$ and $N^i$ disappears. So $K_{ij}$ is taken as the new generalized coordinate as in the method of BL and is denoted as $Q_{ij}$. The momenta canonically conjugate to $h_{ij}$ and $Q_{ij}$, $p^{ij}$ and $\Pi^{ij}$ respectively, are taken as the coefficients of their variations in (3.19), according to Ostrogradski: $$ p^{ij}={\partial{\cal L}\over \partial(\partial_{0}h_{ij})} -\partial_{0}\left({\partial{\cal L}\over \partial(\partial_{0}^2h_{ij})}\right) -\partial_{k}\left({\partial{\cal L}\over \partial(\partial_{0}\partial_{k}h_{ij})}\right) -{\partial{\cal L}\over \partial(\partial_{0}K_{kl})}{\partial K_{kl}\over \partial h_{ij}} +\partial_{m}\left({\partial{\cal L}\over \partial(\partial_{0}K_{kl})}{\partial K_{kl}\over \partial(\partial_{m}h_{ij})}\right) \eqno(3.20\rm a) $$ and $$ \Pi^{ij}={\partial{\cal L}\over\partial(\partial_{0}Q_{ij})}. \eqno(3.20\rm b) $$ Hamiltonian $H$ is defined as the coefficient of $(-1)\times\delta t$ in $L\delta t+\delta F$ after using (3.3). Then Hamiltonian density is given by $$ {\cal H}=p^{ij}\partial_{0}h_{ij}+\Pi^{ij}\partial_{0}Q_{ij}-{\cal L}. \eqno(3.21) $$ It is noted that $K_{ij}$ is half the Lie derivative of $h_{ij}$ along the normal to the hypersurface $\Sigma_{t}$. The righthand sides of (3.20) are rewritten explicitely in terms of the Lie derivatives as follows $$ \left\{\begin{array}{l} p^{ij}=-\sqrt{h}\left[f'(R)Q^{ij}+h^{ij}f''(R){\cal L}_{n}R +N^{-2}\partial_{k}NN^kf'(R)h^{ij}\right] \\[3mm] \Pi^{ij}=2\sqrt{h}f'(R)h^{ij} \end{array}\right. \eqno(3.22) $$ where the scalar curvature is expressed as $$ R=2h^{ij}{\cal L}_{n}Q_{ij}+Q^2-3Q_{ij}Q^{ij}+\tilde{R}-2\Delta(\ln N). \eqno(3.23) $$ It is seen from (3.22) that $\Pi^{ij}$ has only the trace part and is written $$ \Pi^{ij}={1\over d}\Pi h^{ij}\ \ \ {\rm with}\ \ \ \Pi=2d\sqrt{h}f'(R) \eqno(3.24) $$ where $d$ is the dimension of space. Converting this relation, the scalar curvature is expressed as $$ R=f'^{-1}(\Pi/2d\sqrt{h})\equiv \psi(\Pi/2d\sqrt{h}). \eqno(3.25) $$ \section{Invariance of the Hamiltonian} In this section, we demonstrate that the Hamiltonian defined in the previous section is invariant under the transformation of the generalized coordinates induced by (1) the general coordinate transformation on the hypersurface $\Sigma_{t}$ and (2) the transformation of the metric $h_{ij}$ to its function $G_{ij}(h_{kl})$ such as the transformation of the scale factor $a(t)$ to its logarithm $\phi(t)\equiv \ln a(t)$. \subsection{Invariance under the coordinate transformation on $\Sigma_{t}$} Let us consider the following general coordinate transformation on $\Sigma_{t}$ $$ \begin{array}{ccl} x^0&\rightarrow&\bar{x}^0=x^0\\[3mm] x^i&\rightarrow&\bar{x}^i=f^i(x^1,x^2,x^3). \end{array} \eqno(3.26) $$ Under this transformation, $h_{ij}$ transforms as a second rank tensor, $N$ a scalar and $N^i$ a vector. On the other hand the momenta transform as contravariant tensor densities $$ \left\{\begin{array}{l} \bar{p}^{ij}=\displaystyle {\partial (x)\over \partial (\bar{x})}{\partial \bar{x}^i\over \partial x^k} {\partial \bar{x}^j\over \partial x^l}p^{kl} \\[5mm] \bar{\Pi}^{ij}=\displaystyle {\partial (x)\over \partial (\bar{x})}{\partial \bar{x}^i\over\partial x^k} {\partial \bar{x}^j\over \partial x^l}\Pi^{kl} \end{array}\right. \eqno(3.27) $$ where a overbar represents the transformed quantity and $\partial (x)/\partial (\bar{x})$ is the Jacobian of the transformation. The transformed Hamiltonian density is defined as in (3.21) $$ \bar{\cal H} =\bar{p}^{ij}\partial_{0}\bar{h}_{ij}+\bar{\Pi}^{ij}\partial_{0}\bar{Q}_{ij} -\bar{\cal L}. \eqno(3.28) $$ The time derivatives and the Lie derivatives do not effect the transformation properties so that $Q_{ij}$, $\partial_{0}h_{ij}$ and $\partial_{0}Q_{ij}$ transform as second rank tensors. Since the Lagrangian density is the scalar density, so is also the Hamiltonian density $$ \bar{\cal H}={\partial (x)\over \partial (\bar{x})}{\cal H}. $$ Therefore we have the desired result $$ \bar{H}=\int_{t_{1}}^{t_{2}}\bar{\cal H}dtd^3\bar{x} =\int_{t_{1}}^{t_{2}}{\partial (x)\over \partial (\bar{x})}{\cal H}dt {\partial (\bar{x})\over \partial (x)}d^3x=H. \eqno(3.29) $$ This result is expected since the coordinate system we begin is not specified. Thus what we have shown is essentially the consistency of our method. \subsection{Invariance under the transformation to the function of the metric} Next consider the transformation of the generalized coordinates $$ h_{ij}\rightarrow \phi_{ij}=F_{ij}(h_{kl})\ \ \ {\rm or}\ \ \ h_{ij}=G_{ij}(\phi_{kl}). \eqno(3.30) $$ Under this transformation, three dimensional space is unchanged so that the new generalized coordinate $Q_{ij}={\cal L}_{n}h_{ij}/2$ is unchanged and so is the momentum canonically conjugate to it $\Pi^{ij} =\partial{\cal L}/\partial(\partial_{0}Q_{ij})$. These are expressed in terms of the transformed quantities as $$ \left\{\begin{array}{l} \displaystyle Q_{ij}={1\over2}{\cal L}_{n}G_{ij}(\phi_{kl}) \\[5mm] \Pi^{ij}=2\sqrt{h}f'(\psi)G^{ij}(\phi_{kl}). \end{array}\right. \eqno(3.31) $$ Then using (3.22), (3.31) and the relation $$ \delta h_{ij}={\partial G_{ij}\over \partial \phi_{kl}}\delta\phi_{kl}, \eqno(3.32) $$ the first term of $\delta F$, (3.19), takes the following form $$ -\sqrt{h}\left\{f'(\psi)Q^{ij} +G^{ij}f''(\psi){\cal L}_{n}\psi +N^{-2}\partial_{k}NN^kf'(R)h^{ij}\right\}{\partial G_{ij}\over \partial\phi_{kl}} \delta^{*}\phi_{kl} +2\sqrt{h}f'(\psi)G^{ij}\delta^{*}Q_{ij}. \eqno(3.33) $$ The momentum $p_{\phi}^{ij}$ canonically conjugate to $\phi_{ij}$ is defined as the coefficient of $\delta^{*}\phi_{ij}$: $$ p_{\phi}^{ij}=-\sqrt{h}\left\{f'(\psi)Q^{kl} +G^{kl}f''(\psi){\cal L}_{n}\psi +N^{-2}\partial_{k}NN^kf'(R)h^{ij}\right\}{\partial G_{kl}\over \partial\phi_{ij}}. \eqno(3.34) $$ From (3.22) and (3.33), we have the following relation $$ p^{ij}={\partial F_{kl}\over \partial h_{ij}}p_{\phi}^{kl}. \eqno(3.35) $$ Transformed Hamiltonian density ${\cal H}_{\phi}$ is defined as $$ {\cal H}_{\phi}\equiv p_{\phi}^{ij}\dot{\phi}_{ij}+\Pi^{ij}\dot{Q}_{ij}-{\cal L}. \eqno(3.36) $$ The change of the Hamiltonian density is given by $$ \Delta {\cal H}\equiv {\cal H}_{\phi}-{\cal H} =p_{\phi}^{ij}\dot{\phi}_{ij}-p^{ij}\dot{h}_{ij}. \eqno(3.37) $$ Using (3.32) and (3.34), it is shown that $\Delta {\cal H}$ vanishes. This result should be necessary if the Hamiltonian has something to do with the energy. Application to the FRW spacetime, in which case the transformation is from the scale factor $a(t)$ to its function $G(a)$, e.g., $\ln a(t)$, is straightforward. \section{Summary and discussions} We proposed a canonical formalism of HCG by combining those of Ostrodradski and BL using the Lie derivatives instead of the time derivatives. Invariance of the Hamiltonian is shown under the transformation of generelized coordinates preserving the hypersurface $\Sigma_{t}$ which is lacking in BL. In fact, the transformation properties in HCG have not been addressed. The result is important in HCG where the conformal transformation to Einstein frame is an often used technique\cite{equiv}. The problem with this transformation is which of the metric is physical, i.e. the observed one. Many criteria have been proposed to select the physical one, however the problem has not been settled\cite{MS}. However the conformal transformation depends on the curvature, so does on the momenta from the viewpoint of canonical formalism. It is not obvious whether the canonical equations of motion in both frame is equivalent. Also it would be necessary to show that the Poisson brackets defined in each frame are consistent. If the answers to these problems are negative, equivalence of both frames is broken and the canonical quantization leads to different quautum theories. The results of these problems will be reported separately. Finally we comment on quantum gravity. It is well known that there are two versions. One is the quantum theory of gravitons based on the duality of gravitational wave and graviton and has the same footing as the ordinary quantum theory of matter and radiation. This version, however, requires a background spacetime. If the quantum theory is the fundamental theory in physics, the background spacetime should also be determined quantum theoretically. It is the subject of quantum cosmology which is the second version of quantum gravity. In this version no guiding principle such as duality of particles and waves has been known. The usually adopted procedure is the canonical quantization. For this, the classical theory is a prerequisite. Such a theory may not be the Einstein gravity but a kind of HCG. If this were the case, investigations of the early stage of the univerese, or string theory might give some clues. The strings feel tidal force, which necessarily leads to HCG. Observational informations are generally more important. These informations may be brought about from future cosmological observations.

Introduction ============ Endoscopic biliary stenting (EBS) is a standard treatment for malignant biliary obstruction. However, EBS placement is associated with several adverse events, including pancreatitis, liver abscess, and migration of the stent. Although the incidence is very rare, rupture of a pseudoaneurysm (PA) has been also reported as a potentially life-threatening complication after EBS and requires prompt endoscopic and/or endovascular treatment ([@B1], [@B2]). We herein report two cases of severe biliary bleeding from a PA that developed after the endoscopic placement of a biliary self-expandable metallic stent (SEMS). Case Reports ============ Case 1 ------ A 72-year-old woman who had been undergoing chemotherapy for the past year and had since chosen best supportive care was admitted to our hospital for the treatment of a biliary obstruction due to the metastasis of her gastric cancer. Endoscopic retrograde cholangiopancreatography (ERCP) demonstrated multiple stenoses within the hilar and common bile duct ([Fig. 1a](#g001){ref-type="fig"}), and 2 SEMSs (a 10 mm×8 cm, uncovered Niti-S™ Large Cell D-type and a 10 mm×6 cm fully covered Supremo, both from TaeWoong Medical, Gimpo, South Korea) were placed ([Fig. 1b](#g001){ref-type="fig"}) after a forceps biopsy of the lower bile duct. The histology of the biopsy sample confirmed moderately to poorly differentiated adenocarcinoma, compatible with the metastasis of gastric cancer. Fourteen days later, the patient developed a fever, jaundice, and anemia for five days. Contrast-enhanced computed tomography (CT) revealed a PA at the upper edge of the biliary stent, 16×11 mm in size ([Fig. 2a](#g002){ref-type="fig"}). Angiography was urgently performed, and a PA identified at the branch of the right hepatic artery was embolized with multiple coils ([Fig. 2b and c](#g002){ref-type="fig"}). {#g001} {#g002} The patient subsequently demonstrated no evidence of hemobilia or cholangitis, and her bleeding was controlled. However, she died of gastric cancer progression two months after the coil embolization. Case 2 ------ A 70-year-old man with a widely spreading biliary stricture underwent ERCP ([Fig. 3a](#g003){ref-type="fig"}), a biliary forceps biopsy, and naso-biliary drainage placement. Tissue obtained by the biopsy revealed invasive tubular adenocarcinoma that led to the clinical diagnosis of cholangiocarcinoma. Subsequent gadoxetate-sodium-enhanced magnetic resonance imaging (MRI) revealed multiple metastases, indicating a need for metallic stent placement but not surgical resection. ERCP was performed. After biliary balloon dilation (Hurricain™ RX Catheter, 8 mm; Boston Scientific, Marlborough, USA), 2 SEMSs (both uncovered Niti-S™ Large Cell D-type stents, 10 mm×8 cm and 10 mm×6 cm) were placed in a partial stent-in-stent formation in the left and right hepatic ducts to the lower common bile duct ([Fig. 3b](#g003){ref-type="fig"}). {#g003} One week after the biliary stenting, the patient developed a fever that lasted a few days along with epigastric pain and melena. Esophagogastroduodenoscopy (EGD) revealed no evidence of bleeding, including from the major papilla. Two days after EGD, the patient developed hematemesis, and enhanced CT was performed to search for abnormal vascular lesions. CT revealed a PA arising from the posterior superior branch of pancreatoduodenal artery ([Fig. 4a](#g004){ref-type="fig"}) and protruding into the stent. This PA had not been detected by the previous session of enhanced MRI. Emergency angiography was performed, and the aneurysm was fully embolized with multiple coils until the blood inflow was arrested ([Fig. 4b and c](#g004){ref-type="fig"}). {#g004} The patient showed no further signs of gastrointestinal bleeding, and systemic chemotherapy was initiated with gemcitabine (1,000 mg/m^2^/week, day 1, 8, 15 in 4 weeks). However, two months later, the patient died from hematemesis, probably due to the recurrence of biliary PA. Discussion ========== Endoscopic biliary stenting is widely applied for the treatment of malignant biliary strictures, and a SEMS is often used in unresectable cases. Biliary stenting is associated with several complications, including pancreatitis, cholangitis, cholecystitis, perforation, and hemorrhaging, which develop in 7-24% of stent cases ([@B2]-[@B5]). PA is a rarely reported complication after biliary stenting; however, rupture of a PA may be a life-threatening event. Although the actual incidence is not known with accuracy, a PubMed keywords survey during 2003 and 2018 using "pseudoaneurysm" and "biliary stent," revealed 18 cases with ruptured PA in association with biliary stenting ([Table](#t001){ref-type="table"}). The cause of PA is not clear. These PAs arise from various levels of arteries (i.e. right or left hepatic artery, gastroduodenal artery, and pancreatoduodenal artery) depending on their lesions. Interestingly, PA has developed even in cases with benign strictures and plastic stent placements ([@B6]-[@B19]), and the duration from the biliary stenting reportedly ranges from five days to two years ([@B6], [@B11]). In malignant cases, the causal arteries are often involved in the tumors; therefore, their running direction and fragility are altered. Even a superficial forceps biopsy can cause pulsative arterial bleeding ([@B20]). In addition to the mucosal necrosis due to cancer invasion, luminal compression by the stents and the effects of chemotherapies before and after the stenting are considered risk factors for PA rupture. The PA was located at the severely narrowed hepatic hilar portion, which corresponded to the central part of the SEMS, thus suggesting the presence of high compression on the arterial wall due to the expansion of the metallic stent. In addition, some reported cases (including the two present cases) have had cholangitis before PA rupture ([Table](#t001){ref-type="table"}), suggesting that cholangitis may also be a cause of the development of PA. ###### Reported Cases of Pseudoaneurysm Ruptured after the Billiary Stenting (literature Review). Case no. Ref. no. Age Gender Disease Treatment for cancer Stent kinds diameter Stenosis Part Prior Colangitis Pseudoaneurysm   Rupture of pseudoaneurysm ---------- --------------- ----- -------- -------------------------------------- ---------------------- ------------- ---------------- --------------- ------------------ ---------------- ---------- -- --------------------------- ---------- -------------- ------ ----- 1 6 47 F lymphoma CRT SEMS+ PS unknown CBD presence RHA ND melena 2 years TAE, surgery none no 2 7 62 F hilar cholangiocarcinoma CRT PS 10Fr HH presence LHA 20 fever, jaundice 1 month TAE none no 3 8 68 F hilar cholangiocarcinoma ND PS 8.5Fr HH presence RHA ND melena, jaundice 20 days TAE none no 4 9 70 M extrahepatic cholangiocarcinoma CT SEMS 10mm CBD presence RHA 9×6 abdminal pain, jaundice 9 months TAE none no 5 10 48 M pancreatic cancer CRT → CT SEMS 10mm CBD absence RHA ND melena, jaudice 8 months TAE none no 6 11 72 M pancreatic cancer CRT SEMS 10mm CBD presence RHA ND hemetamesis, melena 5 days none yes yes 7 82 F pancreatic cancer none SEMS 10mm CBD absence PAPDA ND hemetamesis 20 days TAE none no 8 80 M gallbladder carcinoma CT SEMS 10mm CBD absence RHA ND hemetamesis, melena 6 months TAE none no 9 12 51 M pancreatic cancer ND SEMS unknown CBD ND GDA ND melena 76 days TAE none no 10 65 M gallbladder carcinoma ND SEMS unknown CBD ND GDA ND melena 15 days TAE none no 11 72 M hilar cholangiocarcinoma ND SEMS unknown HH ND RHA ND hemetamesis 152 days TAE, surgery none no 12 13 60s M pancreatic cancer CRT → CT SEMS 10mm CBD absence RHA 8 jaundice, shock 6 months TAE none no 13 14 78 M pancreatic cancer CRT SEMS unknown CBD ND GDA ND melena, shock 4 months TAE none no 14 15 75 M pancreatic cancer ND SEMS 10 mm CBD absence GDA 8 melena, abdominal pain 1 month TAE none no 15 16 47 M post-surgical bile leaks NA PS 11.5Fr and 7Fr NA absence RHA ND melena ND TAE none no 16 17 78 F biliary stone NA PS unknown NA presence RHA 13×10 hemetameis 1 year TAE yes no 17 18 56 M obstructive jaundice after hepatitis NA PS unknown CBD absence LHA 11.6×9.7 hemetamesis 13 days TAE none no 18 19 78 F benign biliary stenosis NA PS unknown B3 presence LHA 3 jaudice, fever 14 days TAE none no 19 Current cases 72 F obstructive jaundice by metastasis CT SEMS 10 mm HH, CBD presence RHA 16×11 jaundice, fever 14 days TAE none no 20 70 M extrahepatic cholangiocarcinoma CT SEMS 10 mm CBD presence PSPDA 8×7 epigastric pain, melena 10 days TAE yes yes CBD: comon hepatic duct, CRT: chemoradiotherapy, CT: chemotherapy, ND: not described, NA: not applicable, SEMS: self-expandable metallic stents, PS: plastic stent , RHA: right hepatic artely, LHA: left hepatic artely, PSPDA: posterior superior pancreaticoduodenal artery, GDA: gastroduodenal artery, TAE: transarterial embolization \*Duration: duration from stenting to bleeding The treatment of a ruptured biliary PA is usually conservative and can include blood transfusion, transarterial embolization (TAE), and surgery. To date, no consensus has been reached concerning the priority of TAE and surgical treatments. However, when taking the physical invasiveness and the physical condition of patients with an advanced stage of cancer into consideration, TAE is viewed as the first therapeutic choice for most cases. In fact, most reported cases have been treated with intravascular therapy and have shown good outcomes ([Table](#t001){ref-type="table"}). Among the 19 reported TAE-treated cases, including the present ones, only 2 cases (case 16 and case 20 in [Table](#t001){ref-type="table"}) showed recurrences. Yasuda et al. ([@B17]) reported a 78-year-old case (case 16) accompanied by an arterial-biliary fistula that developed re-rupture of the PA after the initial session of TAE. In that case, chemotherapy had not been performed, but a biliary stent had been placement for a year without the removal of common bile duct (CBD) stones. The authors speculated that the accompanying cholangitis and CBD stones were associated with the PA formation. In our case, the administration of two courses of systemic chemotherapy with gemcitabine may have affected the fragility and neovascularity of the biliary wall. The accumulation of more cases is needed for the further assessment of the risk factors for recurrence. At present, continuing follow-up with enhanced CT is recommend after stent placement for malignant biliary obstruction. In cases of postsurgical biliary PA, SEMS placement is reported to be an effective treatment ([@B21], [@B22]). Cases of post-stenting PA all develop biliary bleeding, so a covered SEMS might be a useful tool for achieving pressure hemostasis and may be a viable treatment choice. However, as mentioned, the pressure from a self-expanding stent may conversely induce the subsequent development of a PA. Therefore, the utility of SEMS treatment requires further validation. The current study reported two cases of a ruptured PA following metallic stent placement for the treatment of a malignant biliary stricture. Thus far, no effective preventative strategy for PA formation has been described. However, post-stenting imaging screening can reduce the incidence of PA rupture by prophylactic TAE. Rupture of a biliary PA can be a life-threatening event; however, TAE appears to be an effective hemostatic treatment. **The authors state that they have no Conflict of Interest (COI).** [^1]: Correspondence to Dr.　Hiroyuki Matsubayashi, <h.matsubayashi@scchr.jp>

\section*{Acknowledgment} Our work on the superheating field was urged upon us by Hasan Padamsee, who recognized both the importance of firm estimates of the theoretical maximum performance for niobium cavities, and the confusion in the SRF field at that time about potential new materials. We also thank Alex Gurevich for extensive consultation and collaboration, particularly on the new work on laminates in Section~(\ref{sec:laminates}). Much of that section was inspired by our conversations with him and tests his independently derived thoughts and conclusions on the subject. DBL and JPS were supported by NSF DMR-1312160, Sam Posen is supported by the United States Department of Energy, Offices of High Energy Physics. Fermilab is operated by Fermi Research Alliance, LLC under Contract No. DE-AC02-07CH11359 with the United States Department of Energy. GC acknowledges partial support by the EU under REA Grant Agreement No. CIG-618258. ML was supported by NSF Award No. PHY-1416318, and DOE Award No. DE-SC0008431. This work was supported by the U.S. National Science Foundation under Award OIA-1549132, the Center for Bright Beams. \section{Basic theory of the superheating field} \label{sec:basictheory} The superheating field $H_{\mathrm{sh}}$ is set by the competition between magnetic pressure (imposed by the external magnetic field), the energy cost to destroy superconductivity, and the attractive force due to the zero-current boundary condition at the interface. In Ginzburg-Landau theory, the ratio $\kappa = \lambda/\xi$ of the penetration depth $\lambda$ to the coherence length $\xi$ determines many properties of superconductors. In particular, $\kappa < 1/\sqrt{2}$ and $\kappa>1/\sqrt{2}$ are associated with type-I and type-II superconductivity, respectively. In the flux-line lattice of type-II superconductors, both the vortex supercurrent and magnetic field are confined to a tube of radius $\lambda$. The superconductivity is destroyed (the density of superconducting electrons vanishes) over a smaller vortex core of radius $\xi$. Within GL theory, $H_{\mathrm{sh}}(T)/H_c(T)$ depends on materials properties only through the parameter $\kappa$, which is independent of temperature. A careful calculation using linear stability analysis~\cite{transtrum11} shows that $H_{\mathrm{sh}}$ plateaus at about $0.75 H_c$ in the large $\kappa$ limit, and diverges as $\kappa^{-1/2}$ for $\kappa \ll 1$. \subsection{Simple arguments for the superheating field} \label{subsec:simpleArgument} We now give simple arguments and pictures to estimate the superheating field of superconductors (see e.g.~\cite{liarte16}). The main idea is to compute the work necessary to push magnetic field onto the superconductor through an energy barrier set by the magnetic energy, and compare the result with the condensation energy. It is worth noting that there are important qualitative differences between these simple arguments and the actual linear stability analysis of the GL free energy. We will return to these issues when we discuss the effects of anisotropy in Sec.~(\ref{subsec:anisotropy}), and discuss them further in the full publication~\cite{liarte16}. \begin{figure}[!h] \centering \includegraphics[width=0.9\linewidth]{Sec/Fig/fig1.pdf} \caption{(On the left) Illustration of a superconductor occupying the half-space $x>0$, and subject to an applied magnetic field $\mathbf{H}$ that is parallel to the $z$ axis. ``SC'' stands for superconductor. (On the right) Approximate shape of a superconducting RF cavity in the regions of high magnetic fields. As in the flat case, the magnetic field that is generated by the accelerating beam (and excited by an external RF source, driving the operating/accelerating mode) is parallel to the interior surface of the cavity. \label{fig:geometry}} \end{figure} Consider a superconductor occupying the half-space $x>0$, and subject to an applied magnetic field $\mathbf{H}$ that is parallel to its surface, along the direction $z$. We illustrate this geometry on the left side of Fig.~(\ref{fig:geometry}), where ``SC'' stands for superconductor. Note that the superconductor region extends to infinite in the positive and negative $y$ and $z$ directions, and in the positive $x$ direction; there are no `corners' in this geometry~\footnote{The absence of corners is an important limiting factor in our approach, for corners typically facilitate field penetration in real samples of arbitrary shapes. Modern RF cavities have an approximate cylindrical shape in the region of high magnetic fields (see right side of Fig.~(\ref{fig:geometry})), with no corners, so such geometric considerations become unimportant.}. Let us start with the argument for the superheating field of a type-I superconductor. For small external magnetic fields, the order parameter does not vanish at the vacuum-superconductor interface. However, if we push a slab of magnetic field onto the superconductor (just enough to make the order parameter vanish at the interface), we will destroy superconductivity over a length scale of order $\xi$. The work per unit area that is necessary to push magnetic energy onto the superconductor is set by the magnetic pressure and the penetration length; it is given approximately by $[H_{\mathrm{sh}} / (4\pi)] H_{\mathrm{sh}} \, \lambda$ in cgs units. To estimate the superheating field, we compare this work with the condensation energy per unit area $[{H_c}^2 / (8\pi)] \, \xi$, resulting: \begin{eqnarray} \frac{H_{\mathrm{sh}}}{ H_c} \approx 2^{-1/2} \, \kappa^{-1/2}. \label{eq:typeI_estimate} \end{eqnarray} Equation (\ref{eq:typeI_estimate}) should be compared with the small-$\kappa$ limit of the exact result using Ginzburg-Landau theory~\cite{transtrum11}: $H_{\mathrm{sh}} / H_c \approx 2^{-1/4} \, \kappa^{-1/2}$. In type-II superconductors, field penetration occurs via vortex nucleation, and the superheating field is set by the magnetic pressure that is necessary to push a vortex through a surface barrier onto the superconductor% \footnote{\label{foot:Yogi} Note that this argument is {\em not} related to Yogi's `vortex line nucleation'~\cite{Yogi1977,saito04} estimate of $\ensuremath{H_\mathrm{sh}}$. The latter, developed to analyze impressive experimental data, was qualitatively incorrect~\cite{transtrum11}. In particular, its estimate for the metastable limit $\ensuremath{H_\mathrm{sh}}$ for large $\kappa$ went {\em below} $H_{c1}$, which makes no sense. This misled the SRF field for years into ignoring the potential importance of higher $\kappa$ materials.}. There are two steps to this penetration. First, the core of the superconducting vortex (of radius $\sim\xi$) must penetrate into the surface, at a cost given by the core volume times the condensation energy. Second, this newly penetrated vortex must fight past an attractive force toward the surface due to the boundary conditions at the surface, which is usually estimated~\cite{bean64} by the attraction to an `image vortex'. Below we discuss the superheating field estimated from the initial penetration of the vortex. (Bean and Livingston's original estimate~\cite{bean64} of the superheating field starts (somewhat arbitrarily) at a distance $x=\xi$ after this initial penetration, and focuses on the effects of the attractive longer-range force.) Figure~(\ref{fig:vorticesXY}) illustrates the penetration of a vortex core (red disk) onto a superconductor occupying the half-space $x>0$. The magnetic work per unit length to push the vortex core onto the superconductor is given approximately by the condensation energy (per unit length): \begin{eqnarray} \frac{ H_{\mathrm{sh}} }{4\, \pi} \frac{\Phi_0}{\pi \lambda^2} \, 4 \, \lambda \, \xi \approx \frac{{H_c}^2}{8\,\pi} \pi \xi^2, \label{eq:typeII_estimate_a} \end{eqnarray} where $ H_{\mathrm{sh}} / ( 4\, \pi )$ is the magnetic pressure, $\Phi_0$ is the fluxoid quantum, $\pi \lambda^2$ is the vortex area in the $xy$ plane, $4 \, \lambda \, \xi$ is approximately the area that is associated with the region of field penetration (area of the orange box in Fig.~(\ref{fig:vorticesXY}); it is the amount of the area of the vortex that penetrates the superconductor when a vortex core is pushed inside), and $\pi \xi^2$ is the area of the vortex core. Using $\Phi_0 = 2 \, \sqrt{2} \, \pi H_c \, \lambda \, \xi$ in Eq.~(\ref{eq:typeII_estimate_a}): \begin{eqnarray} \frac{H_{\mathrm{sh}}}{ H_c} \approx \frac{ \sqrt{2} \, \pi }{32} \approx 0.14, \label{eq:typeII_estimate_b} \end{eqnarray} independent of $\kappa$. \begin{figure}[!h] \centering \includegraphics[width=0.9\linewidth]{Sec/Fig/fig2.pdf} \caption{Illustrating the penetration of a vortex core into a type-II superconductor (from~\cite{liarte16}). We estimate the superheating field from the work necessary to push a vortex core a distance $x\sim\xi$ into the superconductor. The vortex then must fight past an attractive force to a depth $x\sim\lambda$ to destroy the Meissner state. \label{fig:vorticesXY}} \end{figure} How does this estimate compare with the field estimated from the attractive force, and with the true answer? The true answer, given below in Section~(\ref{subsec:linearstability}), is about five times larger: $H_{\mathrm{sh}} / H_c \approx 0.75$. Bean and Livingston's estimate of the superheating field due to the attractive force to the image vortex is $H_{\mathrm{sh}}/ H_c = 0.71$, of the same form as our estimate $0.14$ but larger and closer to the true estimate. We present the calculation of the field necessary to introduce the core primarily due to its simplicity, and also because it motivates our analysis of anisotropic superconductors in Section~(\ref{subsec:anisotropy}). One should think of these two contributions as being sequential rather than serial: first the core must penetrate, and then the vortex must fight the longer-range attraction to enter the bulk. (It is interesting and convenient that these two fields are of the same scale.) The GL calculation in Section~(\ref{subsec:linearstability}) of course incorporates both the initial core penetration and the longer range attractive force, together with cooperative effects of multiple vortices entering at the same time. Note that, while the {\em field} needed to push the vortex core into the superconductor is roughly comparable to that needed to push the vortex past the attractive long-range potential, the two contributions contribute very differently to the total {\em energy barrier} to flux penetration. Energy is force times distance: the two forces are comparable but the Bean-Livingston force acts on a scale longer by a factor $\kappa = \lambda/\xi$ than our core nucleation, and will dominate the barrier height. Finally, note that in practice the dominant mechanisms for vortex nucleation that set the superheating field will not involve straight vortices penetrating all along their lengths (as in our calculation above) or, even more impressively, arrays of straight vortices cooperatively pushing their way through the surface barrier (Section~(\ref{subsec:linearstability}) below). We expect that disorder and flaws (discussed in Section~(\ref{subsec:disorder})) will lead to localized intrusions of single vortex loops into the material (Fig.~(\ref{fig:fluxTube})). \subsection{Linear stability calculation of the superheating field} \label{subsec:linearstability} In this section we have seen that the superheating field arises in a bulk superconductor due to the competing effects of magnetic pressure and the destruction of superconductivity. Using relatively simple arguments, we derived the qualitative dependence of this field on $\kappa$. We now describe a more rigorous calculation of the superheating field using a linear stability analysis. Linear stability analysis is commonly used in a variety of pattern formation problems\cite{langer1980instabilities, brower1983geometrical, sharma1998pattern, sivashinsky1983instabilities, bodenschatz2000recent, ben1984boundary}. For type II superconductors, the transition from the Meissner state to the mixed state is triggered by fluctuations of a critical wavelength that spontaneously break the transverse symmetry of the bulk sample, which when coupled to the inhomogeneous depth dependence of the Meissner state, make the superheating transition a challenging application of this method. We here describe this calculation using the Ginzburg-Landau theory for concreteness, although the basic procedure could be extended to other theories as we discuss below. Our presentation follows closely the procedure described in \cite{transtrum11}, however, the calculation has a long history in the literature\cite{gennes65, galaiko66, kramer68, fink69, christiansen69, kramer73, chapman95, dolgert96}. The Ginzburg-Landau free energy for a superconductor occupying the half space $x>0$ in terms of the magnitude of the superconducting order parameter $f$ and the gauge-invariant vector potential $\mathbf{q}$ is given by \begin{align} \mathcal{F}[f,\mathbf{q}] & = \int_{x>0}d^{3}r\Big\{\xi^{2}(\nabla f)^{2}+\frac{1}{2}(1-f^{2})^{2} +f^{2}\mathbf{q}^{2} \nonumber \\ & \quad +(\mathbf{H}_{a}-\lambda\nabla\times\mathbf{q})^{2}\Big\}, \end{align} where $\mathbf{H}_{a}$ is the applied magnetic field (in units of $\sqrt{2}H_{c}$). We take the applied field to be oriented along the $z$-axis $\mathbf{H}_{a} = (0,0,H_a)$, and the order parameter $f=f(x)$ to depend only on the distance from the superconductors surface. Assuming that the order parameter is real and parameterizing the vector potential as $\mathbf{q}=(0,q(x),0)$ fixes the gauge. The Ginzburg-Landau equations that extremize $\mathcal{F}$ with respect to $f$ and $\mathbf{q}$ are \begin{equation} \xi^{2}f''-q^{2}f+f-f^{3}=0,\qquad \lambda^{2}q''-f^{2}q=0, \label{eq:1DGL} \end{equation} and with our choices $ H=\lambda q' $, where primes denote derivatives with respect to $x$. With appropriate boundary conditions\cite{tinkham96,transtrum11} these equations can be solved numerically to characterize the Meissner state. For a given solution $(f,\mathbf{q})$ we next consider the second variation of $\mathcal{F}$ associated with small perturbations $f\rightarrow f+\delta f$ and $\mathbf{q}\rightarrow\mathbf{q}+\delta\mathbf{q}$ given by \begin{eqnarray} \delta^{2}\mathcal{F} & = & \int_{x>0}d^{3}r\Big\{\xi^{2}(\nabla\delta f)^{2}+ 4f\delta f\mathbf{q}\cdot\delta\mathbf{q}+f^{2}\delta\mathbf{q}^{2}\nonumber \\ & & (3f^{2}+\mathbf{q}^{2}-1)\delta f^{2}+\lambda^{2}(\nabla\times\delta\mathbf{q})^{2}\Big\}. \label{eq:d2F} \end{eqnarray} If the expression in Eq.~(\ref{eq:d2F}) is positive for all possible perturbations, then the solution is (meta) stable. Since the solution $(f,\delta\mathbf{q})$ depends only on the distance from the boundary (and is therefore translationally invariant along the $y$ and $z$ directions), we can expand the perturbation in Fourier modes parallel to the surface. As shown in Ref.~\cite{kramer68}, we can restrict our attention to perturbations independent of $z$ and write \begin{align} & \delta f (x,y) =\delta\tilde{f}(x)\cos ky, \nonumber \\ & \delta\mathbf{q}(x,y) =(\delta\tilde{q}_{x}\sin ky,\delta\tilde{q}_{y}\cos ky,0), \label{eq:fouriermodes} \end{align} where $k$ is the wave-number of the Fourier mode. The remaining Fourier components (corresponding to replacing $\cos \to \sin$ and vice-versa in Eq.~(\ref{eq:fouriermodes})) are redundant as they decouple from those given in Eq.~(\ref{eq:fouriermodes}) and satisfy the same differential equations derived below. After substituting into the expression (\ref{eq:d2F}) for the second variation and integrating by parts, we arrive at \begin{widetext} \begin{align} &\delta^{2}\mathcal{F}=\int_{0}^{\infty}dx\left(\begin{array}{ccc} \delta\tilde{f} & \delta\tilde{q}_{y} & \delta\tilde{q}_{x}\end{array}\right) \left(\begin{array}{ccc} -\xi^{2}\frac{d^{2}}{dx^{2}}+q^{2}+3f^{2}+\xi^{2}k^{2}-1 & 2fq & 0\\ 2fq & -\lambda^{2}\frac{d^{2}}{dx^{2}}+f^{2} & -\lambda^{2}k\frac{d}{dx}\\ 0 & \lambda^{2}k\frac{d}{dx} & f^{2}+\lambda^{2}k^{2}\end{array}\right)\left(\begin{array}{c} \delta\tilde{f}\\ \delta\tilde{q}_{y}\\ \delta\tilde{q}_{x}\end{array}\right).\label{eq:d2FOperator} \end{align} \end{widetext} This matrix operator is self-adjoint, and the second variation will be positive definite if its eigenvalues are all positive. In the eigenvalue equations for this operator, the function $\delta\tilde{q}_{x}$ can be solved for algebraically. The resulting differential equations for $\delta\tilde{f}$ and $\delta\tilde{q}_{y}$ are \begin{eqnarray} -\xi^{2}\delta\tilde{f}''+(3f^{2}+q^{2}-1+\xi^{2}k^{2})\delta\tilde{f}+2fq\delta\tilde{q}_{y} &=& E\delta\tilde{f}, \nonumber \\ & & \label{eq:eigen1}\end{eqnarray} and \begin{equation} -\lambda^{2}\frac{d}{dx}\left[\frac{f^{2}-E}{f^{2}+\lambda^{2}k^{2}-E}\delta\tilde{q}_{y}'\right] +f^{2}\delta\tilde{q}_{y}+2fq\delta\tilde{f} = E\delta\tilde{q}_{y}, \label{eq:eigen2} \end{equation} where $E$ is the stability eigenvalue. Note that by decomposing in Fourier modes, the two-dimensional problem is transformed into a one-dimensional eigenvalue problem. Numerically, it can be solved by the same methods as the Ginzburg-Landau equations\cite{transtrum11}. The stability eigenvalue will depend on the solution of the Ginzburg-Landau equations, i.e., the applied magnetic field $H_{a}$, and the Fourier mode $k$ under consideration. The superheating field is found by varying both the applied magnetic field and Fourier mode until the smallest eigenvalue first becomes negative. The wave-number of the destabilizing fluctuations are therefore found simultaneously with $\ensuremath{H_\mathrm{sh}}$ and denoted by $k_c$. Values of $\ensuremath{H_\mathrm{sh}}$ and $k_c$ were calculated in Ginzburg-Landau theory for a wide range of $\kappa$ in references\cite{transtrum11, dolgert96} along with analytic estimates. The results are summarized in Figure~(\ref{fig:HshGLT}). \begin{figure} (a) \par\smallskip \centering \includegraphics[width=0.9\linewidth]{Sec/Fig/fig3a.pdf} \\ (b) \par\smallskip \centering \includegraphics[width=0.9\linewidth]{Sec/Fig/fig3b.pdf} \caption{\label{fig:HshGLT} \textbf{Superheating Field in Ginzburg-Landau Theory} (from~\cite{transtrum11}). \textbf{(a)} A numerical estimate of $\ensuremath{H_\mathrm{sh}}$ in Ginzburg-Landau theory over many orders of magnitude of $\kappa$ was found in reference\cite{transtrum11} (black solid line), along with a large-$\kappa$ expansion (red dashed line). A Pad{\'{e}} approximation for small $\kappa$ was derived in reference\cite{dolgert96}(blue dotted-dashed line). \textbf{(b)} The linear stability calculation also yields the wavenumber of the destabilizing fluctuation $k_c$ (black solid line). This first becomes nonzero at $\kappa_c\approx 1.1495$ where it empirically behaves like $1.2\sqrt{\kappa - \kappa_c}$ (blue dotted-dashed line). Large-$\kappa$ estimates for $k_c$ were also derived in reference~\cite{transtrum11} (red dashed line).} \end{figure} For small $\kappa$, the critical fluctuation occurs with wavenumber $k_c =0$ while for large $\kappa$, $k_c > 0$. Interestingly, the transition to nonzero $k_c$ occurs at some critical $\kappa_c$ that is distinct from the type-I/type-II boundary ($\kappa = 1/\sqrt{2}$). Estimates of $\kappa_c$ vary in the literature from 0.5\cite{kramer68} to 1.13($\pm 0.05$)\cite{christiansen69}. Estimates of $\kappa_c$ from solving Eqs.~(\ref{eq:eigen1})and (\ref{eq:eigen2}) range from 1.10\cite{fink69} to 1.1495\cite{transtrum11} (our high-accuracy result). The linear stability approach described in this section could be extended to other geometries as was done for the case of a superconducting film separated from a bulk superconductor by a thin insulating layer in reference~\cite{posen15}. More complicated theories of superconductivity can also be solved using our methods by replacing the Ginzburg-Landau free energy with the appropriate analog, such as the Eilenberger formalism described in more detail in Section~(\ref{subsec:eilenberger}). \section{Conclusions} \label{sec:conclusions} This article attempts to make a case for thoughtful, broad efforts to identify and estimate fundamental physical limits of materials in parallel with experimental efforts. Our estimates for the superheating field of pure materials~\cite{catelani08,transtrum11} had a significant impact in the SRF community, we understand, not because it promised that Nb$_3$Sn could be run at twice the field of Nb. Rather, we pointed out that the `vortex line nucleation' model was incorrect (footnote~\ref{foot:Yogi} on page~\pageref{foot:Yogi})). This model, created by and initially trusted in the SRF community, had provided discouraging estimates for high $\kappa$ materials, misleading the field for some years. We also note that our controlled, theoretically grounded calculations allow one to explore questions like materials anisotropy that cannot be gleaned reliably from phenomenological models (or, indeed, from our own qualitative arguments, Section~(\ref{subsec:anisotropy})). We understand that many in the SRF community are skeptical of the use of bulk (or thick films) of new materials with higher $\kappa$, even though the theoretical estimates suggest significantly improved $\ensuremath{H_\mathrm{sh}}$ as well as lower cooling requirements. We too were concerned until recently that the smaller coherence lengths might make the metastable state more susceptible to vortex penetration. But we believe that our calculation of the effects of disorder within a particular model clearly indicates that the reliability of the new materials increases so rapidly away from $\ensuremath{H_\mathrm{sh}}$ that the effects of lower coherence length should not be a major concern. One must always make choices of where to focus resources (laminates versus bulk materials, coated copper versus niobium~\cite{russo09}; an interesting review has been recently published in this special issue~\cite{feliciano16}) -- but informed choices may involve consultation with experts outside the SRF field. We are also guardedly pessimistic about the utility of thin laminates in increasing performance. (a)~We are concerned that experimentalists continue to be inspired by the very high parallel fields sustainable by isolated thin films~\cite[Section 9]{feliciano16} (footnote~\ref{foot:IsolatedLayer} on page~\pageref{foot:IsolatedLayer}), even though these high fields are irrelevant in GHz applications~\cite{Hc1WrongRef} and likely also practically inaccessible in any AC application. (b)~Dangerous vortices in thin films will not typically reside parallel and inside the films, but penetrate in and out of the film via pancake vortices, whose motion dissipates heat. The maximum fields reachable without flux penetration for ideal thin films are rather similar to those of the bulk material~\cite{posen15}. The flux penetration needed to reach higher fields produces heating per cycle comparable to that for copper cavities~\cite{Hc1WrongRef}. (c)~We explore extensively the suggestion~\cite{Gurevich2006} that the insulating layers in laminates may act to trap flux from defects, keeping flux from entering the bulk. Here the key issue is whether the nucleated pancake vortices escape from the vicinity of their parent flaw, or annihilate with their partners at the end of each cycle. Modeling these as sources for pancake vortices, we agree that neither thermal diffusion nor disorder are dangerous, but that interactions between vortices, and between vortices generated at separate defects, might allow for escape and unacceptable heating -- warranting experimental caution and further theoretical study. Finally, a word to our colleagues outside the accelerator community. This paper is a collaboration of SRF experts from the accelerator community (Posen, Liepe) and condensed matter theorists, and we draw heavily on conversations with both theorists and experimentalists inside the community (Hasan Padamsee, Alex Gurevich, Nicholas Valles, Takayuki Kubo, Kenji Saito). These domain-specific experts have enormous experience in the challenges and issues relevant for the field; we were told that superheating fields, higher $\kappa$ materials, anisotropy, disorder, and laminates were the key questions, and have been guided into studying these in the correct limits and focusing on the right issues. We can testify that this teamwork has led to both excellent condensed-matter physics and efficient, targeted research to improve SRF performance. \section{Experiments} \label{sec:expt} \subsection{High Power pulsed RF experiments} Some of the earliest measurements showing $\ensuremath{H_\mathrm{sh}}>H_c$ for niobium were reported by Renard and Rocher based on DC magnetization measurements. Yogi et al. performed a more systematic study at RF frequencies on samples of Sn, In, Pb, and alloys, in order to cover a range of $\kappa$ values~\cite{Yogi1977}. Analysis of their data resulted in the vortex line nucleation model discussed in footnote~\ref{foot:Yogi}. Noting that measurements of the RF critical field have shown inconsistency, Campisi used a very high power RF source at SLAC to very quickly ramp up the fields in cavities \cite{Campisi1985}. The goal of these high power RF measurements is to reduce the influence of defects by outpacing the thermal effects they cause. Campisi performed high power RF measurements on Nb, Nb$_3$Sn, and Pb cavities. Hays and Padamsee performed similar measurements on these materials at Cornell \cite{Hays1997}. The niobium results are reproduced in Fig.(\ref{fig:pulsed}), showing fairly reasonable agreement with the expected superheating field close to $T_c$ \footnote{$T_c$ assumed to be 9.2 K for Valles' data.}, but then diverging at lower temperatures. \begin{figure}[!h] \centering \includegraphics[width=\linewidth]{Sec/Fig/fig4.pdf} \caption{Pulsed measurement of the maximum field of superconducting niobium cavities from Valles (symbols), compared with estimates of the theoretical maximum possible superheating field (colored ranges). All measurements show good agreement with $\ensuremath{H_\mathrm{sh}}$ at high temperatures. The cavity baked to removed HFQS degradation (red squares) also shows good agreement at low temperatures. DC flux penetration measurements (green triangles) show good agreement with $\ensuremath{H_\mathrm{sh}}$ as well. \label{fig:pulsed}} \end{figure} After these experiments were performed, new preparation techniques were developed for niobium cavities, including a recipe involving electropolishing and a bake at 120 C. This recipe was found to avoid the ``high field $Q$-slope'' (HFQS) degradation mechanism that occurs in niobium cavities at peak fields of approximately 100 mT \cite{Visentin1998,Kneisel1999}. Experiments by Valles show that pulsed measurements of unbaked niobium produced curves that diverged from the expected $\ensuremath{H_\mathrm{sh}}$ near the expected onset field of HFQS. However, after the bake was performed, the data agreed very well \cite{Posen2015b}. The $H_{c1}$ and $\ensuremath{H_\mathrm{sh}}$ curves plotted in the figure were calculated from niobium material parameters that were extracted from measurements of $R_s$ vs $T$ and $f$ vs $T$ via the SRIMP Matthis-Bardeen code \cite{Halbritter1970a,Halbritter1970}. The baked curve has a lower $\ensuremath{H_\mathrm{sh}}$ due to the change in the mean free path after the bake, which in turn affects $\kappa$. \subsection{DC flux penetration measurements by N. Valles.} Valles also performed measurements of the superheating field of unbaked niobium using a DC probe to avoid the effects of HFQS. Using a superconducting solenoid, he applied a DC field to the exterior of a niobium cavity operating at low fields. A sudden decrease in the quality factor of the cavity indicated that flux from the magnet had penetrated to the interior cavity surface. The penetration field extracted from measurements of the applied field agreed well with the expected superheating field for unbaked niobium, as shown in Fig.~(\ref{fig:pulsed}) \cite{Posen2015b}. \section{Beyond Ginzburg-Landau: Eilenberger, anisotropy, and disorder} \label{sec:fancier} The isotropic Ginzburg-Landau analysis of Section~(\ref{sec:basictheory}) is a trustworthy estimate for the superheating field only for ideal surfaces of single-band superconductors with cubic symmetry near the superconducting transition temperature $T_c$. In this section we pursue three topics that introduce new physics to this calculation. First, superconducting RF cavities are usually run at temperatures significantly lower than $T_c$; niobium cavities, with $T_c \sim 9$K are usually run at $T=2$--$4$K in working accelerators. In Section~(\ref{subsec:eilenberger}) we review calculations of the superheating fields that use Eilenberger theory, which is valid at lower temperatures, presenting the analytic results~\cite{catelani08} at large $\kappa$. Our estimates suggest that these Eilenberger corrections to GL are quantitatively important at operating temperatures, but not large. Second, many of the potential new superconductors have rather anisotropic crystal structures and electronic properties; if the superheating field has significant anisotropy, this could motivate single-crystal or controlled growth conditions to control surface orientations in cavities. In Section~(\ref{subsec:anisotropy}) we review calculations~\cite{liarte16} which show that this anisotropy will be small near $T_c$; we also discuss conflicting results for the anisotropy of multi-band superconductors (like MgB$_2$) at low temperatures. Third, in Section~(\ref{subsec:disorder}) we estimate the effects of disorder and flaws in these materials, presenting both qualitative and simple quantitative estimates of the effects of defects and dirt in locally lowering the barrier to magnetic flux penetration and thus lowering the effective superheating field. \subsection{Eilenberger theory for lower temperatures} \label{subsec:eilenberger} The Ginzburg-Landau approach to superconductivity is generally accurate near the critical temperature $T_c$, but usually the accuracy of its prediction worsen as temperature is lowered below $T_c$. A basic example of its failure is given by the temperature dependence of the order parameter $\Delta$: according to GL theory, $\Delta(T)$ behaves as $\sqrt{1-T/T_c}$, in agreement near $T_c$ with the microscopic BCS theory. The latter, however, predicts that at low temperatures the order parameter is temperature-independent up to exponentially small corrections. For our purposes, the limited validity of GL theory implies that the dependence of the superheating field on $\kappa$ discussed in Sec.~(\ref{sec:basictheory}) cannot be assumed to be quantitatively accurate at the low temperatures at which RF cavities are usually operated. This motivates us to consider a more general approach, valid at arbitrary temperature. For low-$T_c$ superconductors, the coherence length $\xi_0 = \hbar v_F/2\Delta_0$ is much longer than the Fermi wavelength; here $\xi_0$ is the zero-temperature coherence length for a clean superconductor with zero-temperature order parameter $\Delta_0$ and Fermi velocity $v_F$. Thanks to the separation in length scales (or equivalently, the separation in energy scales between $\Delta_0$ and the much larger Fermi energy), these superconductors can be modeled using the so-called \textit{quasiclassical} approach, reviewed for example in Refs.~\cite{chandrasekhar2008,rammer2007}. This powerful approach is quite flexible, permitting in principle to include effects such as Fermi surface anisotropy and impurity scattering (we will comment on the latter at the end of this section). This come at the price of having to calculate various Green's functions from which physical quantities such as the order parameter and the current can be obtained. Such calculations are usually much more involved that those of the GL approach. It was shown by Eilenberger~\cite{eilenberger1968} that one can arrive at an expression for the thermodynamic potential as functional of order parameter $\Delta(\boldsymbol{r})$ and vector potential $\boldsymbol{A}(\boldsymbol{r})$, similar to the GL functional, once the quasiclassical equations for the Green's function have been solved. While a general solution is not possible, for the case of a clean superconductor with spherical Fermi surface we developed in Ref.~\cite{catelani08} a perturbative approach valid for large $\kappa$. Then the thermodynamic potential $\Omega$ is \begin{align} \Omega &= \nu \int\!d^3r \left\{\frac{1}{3}\left(\boldsymbol{\nabla}\times\boldsymbol{A}-\boldsymbol{H}_a\right)^2 + \Delta^2 \log \left(\frac{T}{T_\mathrm{c}}\right) \right. \nonumber \\ & \quad + \!\int\!(dn)\!\left.\left[\frac{\Delta^2}{\omega_n} - 2 \left(\sqrt{\Omega^2_n +\Delta^2} -\omega_n \right) \right. \right. \nonumber \\ & \quad \left. \left. +\frac{1}{\kappa_0^2}\frac{\sqrt{\Omega^2_n+\Delta^2}}{4\Omega^2_n} \left(\boldsymbol{n}\cdot\boldsymbol{\nabla} s^{(0)}\right)^2\right]\right\} . \label{tp} \end{align} In this expression $\nu$ is the density of states at the Fermi energy, lengths are in units of the zero-temperature penetration depth $\lambda_0$, \begin{equation} \frac{1}{\lambda_0} =\frac{8\pi}{3} \left(\frac{2\pi\xi_0}{\Phi_0}\right)^2 \nu \Delta_0^2 \, , \end{equation} the vector potential is in units of $\Phi_0/2\pi\xi_0$ with $\Phi_0$ the magnetic flux quantum, $\kappa_0 = \lambda_0/\xi_0$, and $\boldsymbol{n}$ is the unit vector on the Fermi surface. We also use the short-hand notations \begin{equation} \int\!(dn) = 2\pi T \sum_n \int \frac{d\boldsymbol{n}}{4\pi}\, , \qquad \Omega_n = \omega_n - i \boldsymbol{n}\cdot\boldsymbol{A} \, , \end{equation} with $\omega_n = 2\pi T(n+1/2)$, $n=0,\,1,\,2,\ldots$, the fermionic Matsubara frequencies, and \begin{equation} s^{(0)} = \frac{2\Delta}{\sqrt{\Omega_n^2 + \Delta^2}} \, . \end{equation} The thermodynamic potential in Eq.~(\ref{tp}) reduces to the GL one near $T_c$% \footnote{In considering the limit $T\to T_c$ in Ref.~\cite{catelani08}, a prefactor was missed in Eq.~(29) and consequently Eq.~(31), which should read respectively: $\kappa_{GL} = 2\pi T\sqrt{2/3\zeta}\, \kappa_0 \approx 1.50 \kappa_0$ and $\xi(T)=\sqrt{2/3} [\Delta_0/\Delta(T)] \xi_0$ in the notation of that work.} and it can be used to find the superheating field at arbitrary temperature in the regime $\kappa_0 \gg 1$. The calculation of $H_\mathrm{sh}$ proceeds in the same manner as in the GL approach, by studying the stability against small perturbation of the local minima of $\Omega$. This study was performed in Ref.~\cite{catelani08} at leading order in $\kappa_0 \to +\infty$. The ratio $H_\mathrm{sh}/H_c$ between superheating and critical field can be calculated analytically at $T=T_c$ and $T=0$: \begin{equation} \frac{H_\mathrm{sh}^\infty}{H_c}(T_c) \simeq 0.745 \, , \qquad \frac{H_\mathrm{sh}^\infty}{H_c}(0) \simeq 0.840 \end{equation} where we use the $\infty$ symbol in the superscript to indicate that these are leading-order results. Interestingly, the zero-temperature ratio is almost 13~\% larger than the near-$T_c$ one, indicating that naive extrapolation to low-temperatures of the GL result underestimates the superheating field. At arbitrary temperature, the $H_\mathrm{sh}^\infty/H_c$ ratio can be found numerically and is shown in Fig.~(\ref{fig:eilen}). Note the non-monotonic dependence of $H_\mathrm{sh}^\infty/H_c$ on temperature, which leads the superheating field to acquire its largest value $H_\mathrm{sh}^\infty \simeq 0.843 H_c(0)$ at $T\simeq 0.04 T_c$. \begin{figure} \begin{center} \includegraphics[width=0.9\linewidth]{Sec/Fig/fig5.pdf} \end{center} \caption{Temperature dependence of the ratio $H_\mathrm{sh}^\infty/H_\mathrm{c}$ (from~\cite{catelani08}). Note the non-monotonic behavior at low temperatures.} \label{fig:eilen} \end{figure} It should be noted that while the Meissner state remains metastable up to $H_\mathrm{sh}$, a clean superconductor can become \textit{gapless} at a lower field $H_g$~\cite{lin2012effect}; for example at $T=0$ we have $H_g \simeq 0.816 H_c < H_\mathrm{sh}$. The field $H_g$ is relevant to applications such as superconducting cavities because as the applied field approaches $H_g$, AC losses rapidly increase. Indeed, in the presence of a gap the AC losses are in general exponentially suppressed, but this ``protection'' from losses is absent in the gapless state. The above results are restricted to the leading order in $1/\kappa_0$, which makes it possible to neglect the contributions from the last term in square brackets in Eq.~(\ref{tp}). At next to leading order, that term must be taken into account and leads to an expression for the superheating field of the form~\footnote{This formula can be obtained by extending to next-to-leading order the calculations of Ref.~\cite{catelani08} (G. Catelani, unpublished).} \begin{equation} \frac{H_\mathrm{sh}}{H_c}(T) \simeq \frac{H_\mathrm{sh}^\infty}{H_c} (T) + \frac{h(T)}{\sqrt{\kappa_0}}. \end{equation} This formula, with a weakly temperature-dependent dimensionless coefficient $h(T)$, has the same inverse square root dependence on $\kappa_0$ as the GL expression~\cite{transtrum11}. In closing this section, let us comment briefly on the effect of impurity scattering. Both non-magnetic and magnetic impurities were considered in Ref.~\cite{lin2012effect} in the limit $\kappa\to\infty$. At sufficient strength of the non-magnetic impurities scattering rate, there are some qualitative changes: the non-monotonicity of $H_\mathrm{sh}(T)$ is suppressed, and more importantly the gap remains open up to $H_\mathrm{sh}$. However, quantitatively the value of $H_\mathrm{sh}/H_c$ is changed by at most a few percent. In contrast, adding magnetic impurities strongly decreases $H_\mathrm{sh}$, similar to the well-known suppression of $T_c$ due to the pair-breaking effect of such impurities. \subsection{Anisotropic superconductors} \label{subsec:anisotropy} Layered superconductors can display highly anisotropic critical fields. The upper-critical field of BSCCO,% \footnote{The cuprate superconductors have d-wave order parameters, and hence have an anisotropic gap that vanishes along certain directions. Thus, as discussed for gapless superconductivity in Section~(\ref{subsec:eilenberger}), these likely will not be useful for sustained operations at GHz frequencies.} for instance, can vary by two orders of magnitude depending on the angle between the crystal anisotropy axis $\mathbf{c}$ and the applied magnetic field~\cite{tinkham96}. Near zero temperature, the upper critical field of magnesium diboride is about six times larger for $\mathbf{c} \perp \mathbf{B}$ than for $\mathbf{c} \parallel \mathbf{B}$ (see e.g.~\cite{kogan03,budko15}). Here we review Ref.~\cite{liarte16}, which investigates the effects of crystal anisotropy on the superheating field of superconductors, motivated partly by the opportunity of controlling surface orientation in order to achieve higher accelerating fields inside the cavity. Near the critical temperature, for the anisotropy axis $\mathbf{c}$ aligned with one of the Cartesian directions, the anisotropic formulation of Ginzburg-Landau theory~\cite{ginzburg52,caroli66,gorkov64,tilley65} provides a clean approach to study the anisotropy of the superheating field. We can use a change of coordinates and rescaling of the vector potential to turn the anisotropic GL free energy onto isotropic form, and then use previous results from Ref.~\cite{transtrum11} to calculate the superheating field anisotropy of several materials. We find that: \begin{eqnarray} H_{\mathrm{sh}}^{\mathrm{ani}} = \left\{ \begin{array}{ll} H_{\mathrm{sh}} ( \kappa_\parallel ), & \mbox{for $ \mathbf{c} \parallel z$,} \\ H_{\mathrm{sh}} ( \gamma \kappa_\parallel ), & \mbox{for $\mathbf{c} \parallel x$ or $y$,} \end{array} \right. \end{eqnarray} where the superheating field on the right hand side is the solution of the linear stability analysis for isotropic Fermi surfaces, which we discussed in Section~(\ref{subsec:linearstability}), using $\kappa = \kappa_\parallel$ and $\kappa = \kappa_\perp = \gamma \kappa_\parallel$ for $\mathbf{c}$ parallel and perpendicular to $z$, respectively. Within GL theory, $\gamma = \sqrt{m_c / m_a} = \lambda_c / \lambda_a = \xi_a / \xi_c$, with $m_i$, $\lambda_i$ and $\xi_i$ representing the effective mass, penetration depth, and coherence length along the $i$-th direction, respectively. Since $H_{\mathrm{sh}} \approx 0.75 \, H_c$ goes to a constant for large $\kappa$, we find that the superheating field is nearly isotropic for most high-$\kappa$ unconventional superconductors. On the other hand, $H_{\mathrm{sh}} \approx 0.84 \, H_c \, \kappa^{-1/2}$ for small-$\kappa$ type-I superconductors, resulting in an anisotropy of about $\gamma^{1/2}$ when $\kappa_\parallel \gamma$ is small. Figure~(\ref{fig:anisotropyDiagram}) displays a phase diagram in terms of $\kappa_\parallel$ and $\gamma$, showing the region where GL theory predicts type-I (left of the blue line), type-II (right of dark red line) and mixed (in between dark red and blue lines) superconductivity, and the regions where each asymptotic solution is expected. Note, in particular, that ${H_{\mathrm{sh}}}^\parallel / {H_{\mathrm{sh}}}^\perp \approx 1$ for MgB${}_2$. This result is valid only very near $T_c$, where the anisotropies in $\lambda$ and $\xi$ are equivalent. In the next paragraph we will use results from a two-gap BCS theory to estimate the superheating field anisotropy of MgB${}_2$ at lower temperatures. \begin{figure}[!h] \centering \includegraphics[width=0.9\linewidth]{Sec/Fig/fig6.pdf} \caption{Phase diagram of anisotropic superconductors in terms of mass anisotropy ($\gamma = \sqrt{m_c / m_a}$) and GL ($\lambda_a / \xi_a$) parameters (from~\cite{liarte16}). The superconductor is of type-I to the left of the blue line, of type-II to the right of the dark red line, and mixed in between (in the mixed phase, the SC is of type-I for $\mathbf{c}\parallel z$ and of type-II for $\mathbf{c} \perp z$). The blue and yellow regions correspond to the asymptotic solutions ${H_{\mathrm{sh}}}^\parallel / {H_{\mathrm{sh}}}^\perp \approx \gamma^{1/2}$ and ${H_{\mathrm{sh}}}^\parallel / {H_{\mathrm{sh}}}^\perp \approx 1$, respectively (within 10\% accuracy). Note that the superheating field of MgB${}_2$ is nearly isotropic near $T=T_c$. \label{fig:anisotropyDiagram}} \end{figure} Theoretical and experimental studies indicate that the assumption $\lambda_c / \lambda_a = \xi_a / \xi_c$ (vortex and vortex core have identical shapes within GL theory) is strongly violated for low-temperature MgB${}_2$, thus suggesting the use of two parameters to describe crystal anisotropy, namely $\gamma_{\lambda} = \lambda_c / \lambda_a$ and $\gamma_\xi = \xi_a / \xi_c$. Also, $\gamma_\lambda$ and $\gamma_\xi$ exhibit different temperature dependences, with $\gamma_\lambda$ decreasing and $\gamma_\xi$ increasing for decreasing temperature, respectively. Calculations from Ref.~\cite{kogan03} using a two-gap BCS model suggest that $\gamma_\lambda$ and $\gamma_\xi$ become equal only at $T_c$; near zero temperature, $\gamma_\xi \approx 6$ whereas $\gamma_\lambda \approx 1$, agreeing with some~\cite{cubitt03a,cubitt03b,budko15}, but not all (See Ref.~\cite{kogan03} and references therein) experimental estimates. \begin{figure*}[!ht] \centering (a) \includegraphics[width=0.3\linewidth]{Sec/Fig/fig7a.pdf} (b) \includegraphics[width=0.6\linewidth]{Sec/Fig/fig7b.pdf} \caption{From Ref.~\cite{liarte16}. (a) Illustrating vortex (blue disk) and vortex core (red disk) of zero-temperature MgB${}_2$ in the $ac$ plane, with the external magnetic field parallel to the normal of the plane of the figure. We have drawn $\xi_a$ about 30 times larger with respect to $\lambda_a$, so that the core becomes discernible; in the correct scale, the vortex core occupies the tiny black region in the middle of the figure. Notice that vortex and vortex core have identical shapes within GL theory. (b) To estimate the superheating field, we calculate the work to push a vortex core into the superconductor, thus destroying the Meissner state. The very different area of the black dashed boxes for $\mathbf{c} \parallel y$ (left) and $\mathbf{c}\parallel x$ lead to substantial anisotropy of the superheating field for low-temperature MgB${}_2$. \label{fig:mgb2Vortices}} \end{figure*} We can use our simple estimates of Section~(\ref{subsec:simpleArgument}) to make a qualitative prediction for the resulting anisotropy $H_{\mathrm{sh}}^{c\perp y} / H_{\mathrm{sh}}^{c\parallel y}$ in the superheating field, when $\gamma_\xi \ne \gamma_\lambda$ deviates from the single-band GL prediction. Now the anisotropic shape of the vortex and vortex core plays an important role (see Fig.~(\ref{fig:mgb2Vortices})a). When $\mathbf{c}$ is in the $xy$ plane, as in Fig.~(\ref{fig:mgb2Vortices})b, for instance, the superheating field is estimated from the work performed to push the black-dashed ``box'' into the superconductor, which can considerably vary from $\mathbf{c} \parallel y$ (left) to $\mathbf{c} \parallel x$ (right). This leads to an estimate $H_{\mathrm{sh}}^{c\perp y} / H_{\mathrm{sh}}^{c\parallel y} \approx \gamma_\xi / \gamma_\lambda$. A second estimate generalizes the Bean and Livingston argument of the longer-range vortex attraction to incorporate anisotropy, and leads to a slightly different result: $H_{\mathrm{sh}}^{c\parallel x} / H_{\mathrm{sh}}^{c\perp x} \approx \gamma_\xi / \gamma_\lambda$. Yet a third calculation, which we term ``Extended GL'', yields an almost isotropic result, and is based on a direct linear stability analysis of the anisotropic GL free energy (see Eq.~(7) of Ref.~\cite{liarte16}) assuming \emph{unconstrained} $\lambda$'s and $\xi$'s. Table~\ref{tab:mgb2Tab} summarizes our estimates of $\ensuremath{H_\mathrm{sh}}$ for the three geometries, using experimental values for $H_c$ and $\kappa$ for MgB$_2$. Note that we correct numerical discrepancies of our first estimates in the second row of the table: ``1st (corrected)''. The last column shows the maximum superheating field anisotropy according to each method. Most of the values of $\ensuremath{H_\mathrm{sh}}$ are as low as $\ensuremath{H_\mathrm{sh}} \approx 0.24$T for Nb~\cite{padamsee09}. We discuss the origin of these disparate predictions further in Ref.~\cite{liarte16}. \begin{table}[h] \centering \begin{tabular}{ | l | c | c | c | c | } \hline \multirow{2}{*}{Approach} & \multicolumn{3}{c |}{ $H_{\text{sh}}$ ( Tesla ) } & \multirow{2}{*}{Max. Anis.} \\ \hhline{~---~} & $c \parallel x$ & $c \parallel y$ & $c \parallel z$ & \\ \hline 1st estimate & $0.04$ & $0.006$ & $0.04$ & $\sim 6$ \\ 1st (corrected) & $0.2$ & $0.03$ & $0.2$ & $\sim 6$ \\ 2nd estimate (B \& L) & $1.13$ & $0.18$ & $0.18$ & $\sim 6$ \\ ``Extended GL'' & $0.21$ & $0.22$ & $0.22$ & $\sim 1$ \\ \hline \end{tabular} \footnotesize \caption{Estimates of the superheating field and maximum anisotropy of low-temperature MgB$_2$ for three geometries.} \label{tab:mgb2Tab} \end{table} Our GL arguments for the superheating field anisotropy can be trusted near $T_c$: at large $\kappa$ the superheating field anisotropy is not a reason to control surface orientation. Our arguments at lower temperature and for multi-band superconductors are more speculative. The vortex core shape will surely change for $x \sim \xi$ due to the boundary conditions at the surface; the anisotropy in the long-range attraction in multi-band materials may be different from that of a simple anisotropic GL approach. It will be important to apply linear stability analysis to more sophisticated theories, such as multi-gap BCS or strong-coupling Eliashberg theory, especially in the face of the conflicting results shown in Table~\ref{tab:mgb2Tab}. \subsection{Disorder and vortex nucleation.} \label{subsec:disorder} Niobium RF cavities are routinely operated in the metastable regime, at fields $H_{c1} < H < \ensuremath{H_\mathrm{sh}}$ above the field $H_{c1}$ where vortices in equilibrium would penetrate into the superconductor (and dissipate roughly the same energy as in a normal metal). Table~(\ref{tab:MaterialsParameters}) in Section~(\ref{subsec:thermalFluctuations}) gives $H_{c1}$ and $\ensuremath{H_\mathrm{sh}}$ for other candidate materials. For niobium this metastable regime gives us an important factor of $\sim1.6$ in field. Running in the metastable regime is crucial for utility with the higher temperature superconductors, whose $H_{c1}$ equilibrium fields are much lower than the operating fields for current Nb cavities (Table~(\ref{tab:MaterialsParameters})). It took many years of experimentation to raise operating fields of the niobium cavities to approach near to their fundamental limits. Will the new, more complex materials be fundamentally more challenging to optimize? Our preliminary experimental cavities using Nb$_3$Sn appear already to be operating above $H_{c1}$~\cite{posen15b}, but are not yet delivering anywhere near to the theoretically predicted superheating field. Just as we have been exploring the fundamental theoretical limits to the fields for ideal surfaces, in this section we explore the fundamental theoretical challenges in minimizing the effects of dirt, flaws, and defects in lowering the barriers to vortex entry. What kind of flaw or disorder fluctuation would be needed to allow vortices to enter at fields substantially lower than the superheating field? How big a damage region is needed to bypass the surface barrier to vortex entry? Damage will significantly affect the superconducting properties if the flaw or fluctuation has a characteristic length of order the coherence length $\xi$. Since the proposed candidate materials for next generation SRF cavities have shorter coherence lengths than niobium (Table~(\ref{tab:MaterialsParameters})), this potentially could imply that these new materials are more susceptible to defects and dirt. \begin{figure}[!h] \centering \includegraphics[width=0.7\linewidth]{Sec/Fig/fig8.pdf} \caption{{\bf Flux tube nucleation} allowing the penetration of a single vortex core into the superconductor occupying the half space $x>0$. The nucleation barrier at zero disorder can be estimated by computing the energy of this loop, plausibly a semicircular loop of radius $R$, and subtracting the magnetic work done by the pressure due to the external field $H$, where $H_\mathrm{c1} < H < H_\mathrm{sh}$. This figure illustrates the nucleation barrier for low fields near $H_{c1}$; at higher fields the radius becomes comparable to $\xi$. The boundary conditions at the surface of the superconductor lead to an attractive force on the vortex as discussed in Section~(\ref{subsec:simpleArgument}), in addition to the curvature energy of the vortex loop (ignored here). We can estimate the disorder needed to nucleate at a field $H$ by calculating the damage needed to lower this barrier to zero as the radius $R$ grows.} \label{fig:fluxTube} \end{figure} Figure~(\ref{fig:fluxTube}) shows a cartoon of a vortex loop entering a superconductor. Based on the discussion at the end of Section~(\ref{subsec:simpleArgument}) and the caption of Fig.~(\ref{fig:fluxTube}), at external fields far from $H_\mathrm{c1}$ and $H_\mathrm{sh}$, the energy of the vortex loop will grow in the absence of disorder until it reaches a critical radius $R_c$, at which point the energy will again decrease. This critical radius and the needed damage zone will get smaller as the field $H$ grows, vanishing at $H=\ensuremath{H_\mathrm{sh}}$. The energy per unit length of the vortex loop will have two contributions -- a curvature energy and an attractive energy between the vortex and the surface. The latter can be estimated from the attraction of a straight vortex to the `image vortex' needed to set the correct boundary conditions at the surface. This potential barrier (the major component of the superheating field) was estimated by Bean and Livingston for high $\kappa$ type-II superconductors~\cite{bean64}. The unitless Gibbs free energy per unit length $4 \, \pi \, G/(\sqrt{2} \, {H_c} \, \Phi_0)$ of a straight vortex flux line a depth $x$ inside a superconductor with external field $H$ can be written in the (London) large-$\kappa$ limit as~\cite{huebener01}: \begin{align} \label{eq:BLG1} G = \frac{\Phi_0}{4\pi} \left(H (e^{-x/\lambda} -1) - \frac{1}{2} \frac{\Phi_0}{2 \pi \lambda^2} K_0 (2 x/\lambda) + H_{c1}\right) \\ \frac{ 4 \, \pi \, G }{ \sqrt{2} \, {H_c} \, \Phi_0} = g(x) = h \left(e^{-x/\lambda} -1\right) - \frac{K_0 (2\,x / \lambda)}{2\,\kappa} + \frac{\ln \kappa}{2\,\kappa}, \label{eq:BLG2} \end{align} where $h = H/(\sqrt{2}H_c)$, and $K_\nu$ denotes the modified Bessel function of imaginary argument~\cite{gradschteyn07}, and for now $\kappa$ and $\lambda$ are the Ginzburg-Landau parameter and penetration length of the pure material, respectively. The first term is a magnetic pressure, the second term is the interaction with the `image vortex' that imposes the correct boundary conditions, and the third term is the energy per unit length of a vortex deep in the superconductor. We can estimate $R_c(H)$ by setting the derivative $dG/dx = 0$ in Eq.~(\ref{eq:BLG1}) and expanding the Bessel function for small arguments, leading to $R_c(H) \sim \xi \ensuremath{H_\mathrm{sh}}/H$. At the lowest field for vortex penetration $H_{c1}$ this expansion is unreliable; however, since $H_{c1} = \ensuremath{H_\mathrm{sh}} \log(\kappa) / \kappa$~\cite{huebener01}, the resulting estimate $R_c(H_{c1}) \approx \xi \kappa/\log(\kappa) = \lambda/\log(\kappa)$ is still quite good, as shown in Figure~(\ref{fig:energyBarrier}a). But the new materials of interest have lower critical fields $H_{c1}$ too small to be useful; we must run at fields $H$ comparable to $\ensuremath{H_\mathrm{sh}}$. Near $\ensuremath{H_\mathrm{sh}}$, $R_c \approx \xi$, again as shown in Fig.~(\ref{fig:energyBarrier}). For disorder or defects to remove this energy barrier, they will thus necessarily have to strongly affect a region of volume $\sim \xi R_c^2 \sim \xi^3$. \begin{figure}[h] (a) \par\smallskip \centering \includegraphics[width=0.9\linewidth]{Sec/Fig/fig9a.pdf} \\ (b) \par\smallskip \centering \includegraphics[width=0.9\linewidth]{Sec/Fig/fig9b.pdf} \caption{ (a)~Unitless Gibbs free energy (Eq.~(\ref{eq:BLG2})) to push a straight vortex line from a depth $\xi$ to depth $x$ into a superconductor like Nb$_3$Sn with $\kappa=26.4$, for several values of the magnetic field. The superheating field can be estimated in the large $\kappa$ limit from the condition $G^\prime(\xi)=0$, characterizing the vanishing of the surface energy barrier at $x=\xi$; Bean and Livingston's estimate gives $h_{\mathbf{sh}}=1/2$ so $\ensuremath{H_\mathrm{sh}} = H_c/\sqrt{2} \approx 0.71 H_c$, comparable to the correct large-$\kappa$ limit. Note that the peak in the barrier is at $x_c \sim \xi \ensuremath{H_\mathrm{sh}}/H$; near $H_{c1}$ it is roughly $\lambda \approx \kappa \xi/\log \kappa = \lambda/\log \kappa \approx \lambda$, but in the interesting region near $\ensuremath{H_\mathrm{sh}}$ it is near the coherence length $\xi$. (b)~The spatially-dependent critical temperature shift $\alpha=\alpha(x)$, needed to flatten the energy barrier and allow for the penetration of vortices, in our particular model with $\kappa$ for Nb$_3$Sn. This is shown for several values of $H$ in the interval $[H_{c1},\ensuremath{H_\mathrm{sh}})$. Here $H=0.6 H_c$ would duplicate the maximum possible superheating field for niobium. \label{fig:energyBarrier}} \end{figure} To make this more quantitative, one needs to identify and model the dominant mechanism for vortex nucleation. If the characteristic defect size is large compared to $\xi$ (e.g., nucleation on grain boundaries or inclusions of competing phases), one must model and control these individual defects. Clean grain boundaries are usually atomistically sharp (much thinner than $\xi$) and hence do not significantly decrease the local superconducting properties; indeed, studies of hot spots in large grain niobium cavities show no correlation with grain boundaries~\cite{eremeev06}, and using single crystals to avoid grain boundaries has not improved performance~\cite{kneisel05a,kneisel05b}. But in more complex materials, grain boundaries could be more disordered, thicker, or contaminated by impurities, and a grain boundary or grain boundary intersection with the correct orientation with respect to the surface could provide a route to entry. The effect of surface roughness on Bean and Livingston's surface barrier has been studied in Ref.~\cite{bass96}. Kubo has used the London model to investigate the effects of nano-scale surface topography on the superheating field~\cite{kubo15}. Perhaps most dangerous could be inclusions of metallic or poorly superconducting second phases, or irregularities in the surface morphology. If the characteristic defect size is small compared to $\xi$, and if the defects are uncorrelated in position, then the fluctuations in regions of order $\xi^3$ can be quantitatively estimated to linear order using the central limit theorem. This leads to Gaussian random fluctuations in the superconducting properties. For example, for alloys and doped crystals there are natural concentration fluctuations that will locally change the superconducting transition temperature, coherence length, condensation energy, and other properties. This is the traditional theoretical framework for field-theoretic calculations of the effects of disorder. Let us hypothesize a system where the critical temperature is decreased due to disorder. In the context of Ginzburg-Landau theory for a homogeneous system, a change in the critical temperature yields a change in the coefficient $\alpha=\alpha(x)$ of $\psi^2$, where $\psi$ is the superconductor order parameter~\cite{tinkham96}. The probability of a fluctuation in $\alpha(\mathbf{x})$ away from its pure value $\alpha_0$ would be proportional to \begin{equation} \label{eq:Pofalpha} \Pi\{ \alpha ({\mathbf{x}})\} \propto \exp \left( -\int (\alpha ({\mathbf{x}}) / \alpha_0 -1)^2 / (2 \sigma^2) d^3 {\mathbf{x}} \right), \end{equation} where $\sigma$ is a material-dependent constant that encapsulates the likelihood that the dirt in the material will cause a given fractional change $\alpha/\alpha_0$ in the critical temperature. The constant $\sigma$ will become larger either if there are bigger concentration fluctuations or if the material is particularly sensitive to dirt. In principle, we should now calculate the most probable three-dimensional profile $\alpha(\mathbf{x})$ needed to flatten the energy barrier and allow vortices in at a lowered field $H<\ensuremath{H_\mathrm{sh}}$, and then use $\Pi\{ \alpha ({\mathbf{x}})\}$ in Eq.~(\ref{eq:Pofalpha}) to estimate the probability per unit surface area $P(H/\ensuremath{H_\mathrm{sh}})=\Pi\{\alpha(\mathbf{x})\}$ of vortex penetration. Rather than doing this full variational calculation, we build on the Bean-Livingston model of Equation~(\ref{eq:BLG1}). In GL theory, the characteristic lengths scale as $\lambda \sim \alpha^{-1}$ and $\xi \sim \alpha^{-1/2}$. Hence we distinguish $\lambda_0$, $\xi_0$ and $\kappa_0$ for the pure material from $\lambda(x)=\lambda_0/a$, $\xi(x)=\xi_0/\sqrt{a}$ and $\kappa(x) = \lambda(x)/\xi(x)=\kappa_0/\sqrt{a}$ for the damaged region, where $a=\alpha/\alpha_0$. \begin{figure}[!h] \centering \includegraphics[width=0.9\linewidth]{Sec/Fig/fig10.pdf} \caption{{\bf Relative logarithmic reliability ${\cal P} = -(2\sigma^2 /\xi_0^3)\log(P(H/\ensuremath{H_\mathrm{sh}}))$ of vortex nucleation}, in a simple model of Gaussian random disorder, for the $\kappa$ values of the three candidate superconductors. Solid curves are ${\cal P}_{3D}$ for a semicircular vortex barrier model (Fig.~(\ref{fig:fluxTube}), Eq.~(\ref{eq:P3D})); dashed curves are $(d/\xi_0) {\cal P}_{2D}$ for pancake vortex nucleation in a 2D superconducting layer of thickness $d$ (Section~(\ref{sec:laminates})). \label{fig:logPvsH}} \end{figure} What is the minimum amount of dirt that is necessary to reduce the superheating field to a certain value? For instance, how much dirt would it take to reduce $\ensuremath{H_\mathrm{sh}}$ for Nb$_3$Sn (estimated at 0.42 T in Table~(\ref{tab:MaterialsParameters})) to $H = 0.25$ T ($\ensuremath{H_\mathrm{sh}}$ for niobium), a factor $H/\ensuremath{H_\mathrm{sh}} \sim 0.6$? One would need enough dirt `flatten' the surface barrier between% \footnote{Bean and Livingston measure the barrier starting at $x=\xi$, below which London theory is unreliable.} $\xi_0$ and $R_c(H) \approx \xi_0 \ensuremath{H_\mathrm{sh}}/H = 5 \xi_0/3$ along the $x$ direction (thus allowing for vortex penetration), as shown in the dashed line of Fig.~(\ref{fig:energyBarrier}a). In general, we are interested in finding an $x$-dependent parameter $\alpha=\alpha (x)$ that flattens the energy barrier from $x=\xi_0$ to $x=x_f$, where $x_f>\xi_0$, and is defined by $G(x_f) = G(\xi_0)$. The solution for $\alpha(x)$ is then found from the equation $G(x) = G(\xi_0)$ for $\xi_0<x<x_f$, and $\alpha (x) = \alpha_0$ for $x>x_f$, where in the left and right hand sides we use $\{\lambda(x), \xi(x)\}$ and $\{\lambda_0, \xi_0\}$ in Eq.~\eqref{eq:BLG1}, respectively. Note that we are making a rough approximation here. The magnetic fields and supercurrents surrounding the vortex line will see a spatially varying critical temperature $\alpha(\mathbf{x})$ whenever it is far from the surface, and properly measuring its energy and thus the surface attraction should include the resulting shift in energy. The depths $x$ of importance to us are of order the coherence length $\xi$, and thus these long distance fields and currents are largely cancelled by the image vortex a distance $2x$ away. The vortex will see a depth-dependent disorder, but its energy will be qualitatively well described by our model in the region $H\sim\ensuremath{H_\mathrm{sh}}$. Figure~(\ref{fig:energyBarrier}) shows $\alpha / \alpha_0$ as a function of $x/\xi$ for several values of $H$ in the interval $[H_{c1},\ensuremath{H_\mathrm{sh}})$. Apart from an overall constant given by the normalization of the Gaussian, the negative logarithm of the probability of this fluctuation as a function of the lowered entry field $H$ is \begin{align} -\log(P(H/\ensuremath{H_\mathrm{sh}})) &= \Pi\{\alpha(\mathbf{x})\} \\ &= \int (\alpha(\mathbf{x})/\alpha_0-1)^2/(2\sigma^2)d^3\mathbf{x} \\ &= \frac{\xi_0^3}{2\sigma^2} \int (\alpha(\xi_0 \mathbf{u})/\alpha_0-1)^2 d^3\mathbf{u} \\ &= \frac{\xi_0^3}{2\sigma^2} {\cal P}(H/\ensuremath{H_\mathrm{sh}}), \end{align} a measure of the relative logarithmic reliability of the superconductor to disorder-induced nucleation. Here we pull out the volume $\xi_0^3$ of the damage zone by changing variables to $\mathbf{u} = \mathbf{x}/\xi_0$. In a three-dimensional system with a semicircular vortex nucleation approximation, we can use our Bean-Livingston style methods to approximate this as a one-dimensional integral \begin{equation} \label{eq:P3D} {\cal{P}}_{3D}(H/\ensuremath{H_\mathrm{sh}}) = \int \pi u (\alpha(\xi_0 u)/\alpha_0-1)^2 du. \end{equation} For a vortex pancake nucleation event for a thin SIS film of thickness $d$ (discussed in Section~(\ref{sec:laminates})), we find \begin{equation} \label{eq:P2D} {\cal P}_{2D}(H/\ensuremath{H_\mathrm{sh}}) = (d/\xi_0) \int (\alpha(\xi_0 u)/\alpha_0-1)^2 du \end{equation} (see Fig~(\ref{fig:logPvsH})). Clearly, the relative reliability decreases rapidly as $H$ approaches $\ensuremath{H_\mathrm{sh}}$, by many orders of magnitude in this model calculation. The high-$\kappa$ calculation of Bean and Livingston cannot be simply extrapolated to niobium, but there is no reason to doubt that a similar sensitivity of the barrier to $H/\ensuremath{H_\mathrm{sh}}$ is expected. Nonetheless, niobium cavities are used in planned applications at $0.7\ensuremath{H_\mathrm{sh}}$~\cite{ILCtdr,geng11}, suggesting realistic values of disorder are tolerable in niobium. Indeed, the dependence of the barrier on $H/\ensuremath{H_\mathrm{sh}}$ is much stronger than its dependence on $\kappa$ or $\xi$. This suggests, examining Figure~(\ref{fig:logPvsH}), that the factor of five to ten change in $\xi_0$ with the new superconductors may not be so dangerous. The resulting two to three orders of magnitude smaller volume for the critical damage zone at fixed field, it would seem, could be remedied by working not at $0.8\ensuremath{H_\mathrm{sh}}$ but at perhaps $0.6\ensuremath{H_\mathrm{sh}}$ (Figure~(\ref{fig:logPvsH})). Manufacturing high-quality cavities from these new materials may be challenging. What our calculation can provide is reassurance that these materials should not be avoided because of their shorter coherence lengths. \section{Introduction} \label{sec:intro} To transfer energy to beams of charged particles, accelerators frequently use superconducting radio-frequency (SRF) cavities, devices that are capable of sustaining large amplitude electromagnetic fields with relatively small input power. The energy gain of a beam traversing a cavity is determined by the electric field amplitude along its path---a larger amplitude can reduce the number of cavities required to reach a given energy. This is especially important in high energy accelerators, which call for as many as tens of thousands of cavities \cite{ILCtdr}. It is therefore of interest to understand the mechanisms that fundamentally limit the accelerating electric field. For state-of-the-art SRF cavities that have been carefully prepared to prevent non-fundamental degradation processes such as field emission \cite{Bernard1992,Kneisel1993} and multipacting \cite{Proch1979}, studies show that the limit is not the electric field, but rather the interaction of the magnetic field with the superconducting material of the cavity walls. The fundamental limit to acceleration in SRF cavities is the {\em superheating field} $\ensuremath{H_\mathrm{sh}}$, introduced in Section~(\ref{sec:basictheory}). This article will cover ideas, methods, and results revolving around the superheating field and its dependence on the superconductor -- materials properties, anisotropy, defects and disorder, and laminates. The ideas and methods are primarily gleaned from the broader condensed matter community. In Section~(\ref{sec:basictheory}) we review computations of $\ensuremath{H_\mathrm{sh}}$ for clean systems using field theories from the 1950's derived for pure superconductors near their transition temperature~\cite{tinkham96}; in Section~(\ref{subsec:eilenberger}) we draw from more sophisticated theories from the 1960's to calculate $\ensuremath{H_\mathrm{sh}}$ at all temperatures~\cite{eilenberger1968}, and discuss the future need to use these historical theories to incorporate effects of strong coupling and electronic structure~\cite{eliashberg60} in new materials. In Section~(\ref{subsec:anisotropy}) we review the use of these methods to address the electronic anisotropy of some of the new materials. In Section~(\ref{subsec:disorder} we introduce an illustrative calculation of the effects of disorder using tools and methods developed in the 60's for disordered systems~\cite{lifshitz64,zittartzL66} and nucleation theory~\cite{langer68,DanielsS11}, providing reassurance that new materials will likely not be far more sensitive to flaws and dirt. Finally, in Section~(\ref{sec:laminates} we investigate the properties of superconducting laminates, by drawing from work from the 90's on the dynamics of `pancake vortices' in certain layered high-temperature superconductors~\cite{huebener01} (particularly BSCCO, Bi$_2$Sr$_2$Ca$_{n-1}$Cu$_n$O$_{2n+4+x}$). We frankly have two goals for this article. As discussed above, we wish to provide an introduction for the accelerator community into tools and methods from the broader condensed matter community that can help interpret current experimental challenges and guide plans for future research in optimizing materials properties for SRF cavities. But conversely, we want to provide a window for the broader condensed matter theory community into the remarkable frontiers of field, frequency, and materials preparation being explored by the SRF community. We invite their participation in melding 21st century materials-by-design tools from electronic structure theory with 20th century field theories of superconductivity, bridging the scales to address current technological challenges in the accelerator field. (Full disclosure: this article was supported in part by the Center for Bright Beams, an NSF Science and Technology Center whose mission is precisely to bring the accelerator community together with outside experts in physical chemistry, materials science, condensed matter physics, plasma physics and mathematics.) \subsection{Basic facts about superconductors: type I and II, $H_c$, $H_{c1}$, and $H_{c2}$} Normal conducting metals, such as copper, are not viable as radio-frequency cavities for long-pulse high-gradient applications. Due to their high surface resistance, these cavities dissipate too much power on the walls, which can result in melting, among other structural problems, if they are not sufficiently cooled. When subject to high accelerating fields, copper cavities are limited to short-pulse applications. In contrast, superconducting radio-frequency cavities have a much lower surface resistance, which implies low dissipation on the walls and high quality factors (of about $10^{10}$, compared to $10^4$ for copper)~\cite{padamsee09}. Taking into account the refrigerator power to keep the cavity in the superconductor state, SRF cavities are considerably more economical than copper cavities, and present huge benefits, especially for long-pulse applications. At high magnetic fields, however, high-temperature superconductor cavities can dissipate as much power as copper due to the nucleation and motion of vortices. At low enough temperature and applied magnetic field (which for now we assume to be constant in time), superconductors exhibit the Meissner effect: magnetic fields are expelled from the interior of the superconductor, exponentially decaying from the interface surface. Larger applied magnetic field can destroy this Meissner state in two ways, depending on the type of superconductor. In type-I superconductors, an abrupt phase transition takes place at the thermodynamic critical field $H_c$, above which the superconductor is in the normal state. In type-II superconductors, the situation is slightly more complicated. Magnetic flux penetration starts, via vortex nucleation, at a lower magnetic field $H_{c1}<H_{c}$. $H_{c1}$ is called the lower critical field. The transition to the normal phase takes place at the upper-critical field $H_{c2}$ ($H_{c2}>H_c$). In the intermediate range, $H_{c1}<H<H_{c2}$, the system is in the vortex lattice state~\footnote{At higher magnetic fields ($>H_{c2}$), \emph{surface} superconductivity can persist up to a third critical field, $H_{c3}$. This critical field should not be mistaken by the \emph{superheating field}, below which the system displays \emph{bulk} superconductivity and field expulsion.}. \subsection{The superheating field} \label{sec:superheatingFieldI} For these cavities during operation, the external magnetic field is parallel to the superconductor surface. In many applications, the threshold field for flux penetration onto the superconductor is not set by $H_c$ or $H_{c1}$ (for type-I and type-II superconductors, respectively); it is set by the metastability limit of the Meissner state, i.e. by the \emph{superheating} field~\cite{transtrum11,catelani08,garfunkel52,ginzburg58,kramer68,gennes65,galaiko66,kramer73,fink69,christiansen69,chapman95,dolgert96,bean64}. The Meissner state is metastable at $H_c<H<H_{\mathrm{sh}}$ for type-I superconductors, and at $H_{c1}<H<H_{\mathrm{sh}}$ for type-II superconductors. The onset of instability of the Meissner state is related to the vanishing of a surface energy barrier that prevents field penetration onto the superconductor even when $H>H_c$ or $H>H_{c1}$. The metastable Meissner state is analogous to the state of \emph{superheated} water (perhaps explaining the name ``superheating field''). Liquid water in a glass can be superheated in a microwave to a temperature above the liquid-gas transition temperature, but still remain in the liquid state due to the surface tension barrier at the liquid-gas interface, causing small vapor bubbles to contract rather than grow. Surface tension in water is analogous, for instance, to the surface tension due to the energy barrier preventing vortex nucleation in type-II superconductors. Unlike the case of water, as we argue in Section~(\ref{subsec:thermalFluctuations}), thermal nucleation of vortices occurs at relatively long time scales, suggesting that the Meissner state can be sustained in RF applications for fields as large as the superheating field. However, this scenario can considerably change when one considers the effects of disorder in the superconductor. Section~(\ref{subsec:disorder}) discusses disorder-induced nucleation of vortices. The superheating field is associated with spinodal curves where the local stability of the Meissner state is broken. This is a more precise definition that is useful for both type-I and type-II superconductors. We shall discuss calculations of the superheating field in Section~(\ref{sec:basictheory}). Our calculations there will be assuming an external field that is constant in time and ignore thermal fluctuations. We here discuss these approximations. \subsection{Why GHz is slow} \label{sec:slowGHz} Calculations of the superheating field for DC applied magnetic fields will be accurate for RF applications when the microscopic relaxation times are smaller than the time scales that are associated with changes in the fields inside the cavity. Time scales for the latter are of order of nanoseconds~\cite{padamsee09}. A version of time dependent Ginzburg-Landau theory given by Gor'kov and Eliashberg predicts the characteristic relaxation time near $T_c$: $\tau_{\mathrm{GL}}=\pi \, \hbar / [8 \,k\, (T_c-T)]$, where $\hbar$ is the Planck constant divided by $2\pi$ and $k$ is the Boltzmann constant~\cite{tinkham96}. For $T_c-T=1$K, one obtains $\tau_{\mathrm{GL}} \sim 10^{-3}\,$ns for oscillating fields parallel to the sample surfaces. Using $ \Delta \sim k \, T_c$, where $\Delta$ is the superconductor gap, we find $ \tau_{\mathrm{GL}} \sim \Delta^{-1} $ at low temperatures, which is similar to the scaling of collective modes in unconventional superfluids (see e.g. Section 23.5 of Ref.~\cite{ketterson99}). However, note that Gor'kov and Eliashberg theory is applicable to \emph{gapless} superconductors, filled with magnetic impurities and sufficient pair-breaking strength. For superconductors with a clean gap, the relaxation time is expected to be larger than $\tau_{\mathrm{GL}}$, and to scale with the inelastic phonon-scattering time $\tau_E$, which, near $T_c$ is of the order of $\sim 10^{-8}\,$s in Al and $\sim 10^{-11}\,$s for Pb~\cite{tinkham96}, due to its larger critical temperature~\footnote{A simple estimate given in Sec. 10.3 of Ref.~\cite{tinkham96}, assuming a Debye phonon spectrum and free-electron Fermi surface, gives $\tau_E$ scaling as ${T_c}^{-3}$.}. Yoo et al. measured an ultra fast electron-phonon relaxation time of $360$ fs for niobium~\cite{yoo90}. So, at GHz frequencies we may ignore the time dependence in studying the stability. \subsection{Why thermal fluctuations are small} \label{subsec:thermalFluctuations} \renewcommand{\arraystretch}{1.5} \begin{table*} \begin{center} \begin{tabular}{| c | r | r | r | r | r | r | r | r | r | r |} \hline Material & $\lambda$[nm] & $\xi$[nm] & $\kappa$ & $T_c$[K] & $H_{c1}$[T] & $H_{c}$[T] & $\ensuremath{H_\mathrm{sh}}$[T] & $F$[J/m$^3$] & $F\xi^3/k_B$ [K] \\ \hline\hline Nb & 40 & 27 & 1.5 & 9 & 0.13 & 0.21 & 0.25 & 17547 & 25009.0 \\ \hline Nb$_3$Sn & 111 & 4.2 & 26.4 & 18 & 0.042 & 0.5 & 0.42 & 99472 & 533.6 \\ \hline NbN & 375 & 2.9 & 129.3 & 16 & 0.006 & 0.21 & 0.17 & 17547 & 31.0 \\ \hline MgB$_2$ & 185 & 4.9 & 37.8 & 40 & 0.017 & 0.26 & 0.21 & 26897 & 229.1 \\ \hline \end{tabular} \caption{Representative material parameters for niobium, the traditional superconducting material used in SRF cavities, as well as candidate SRF materials that have the potential to reduce cooling costs due to their higher $T_c$. The coherence length $\xi$ is calculated using equations in Ref.~\cite{Orlando1979}. The penetration depth $\lambda$ is calculated from Eq.~3.131 in Ref.~\cite{tinkham96}. The ratio $\kappa=\lambda/\xi$ is called the Ginzburg-Landau parameter, and determines many properties of superconductors. A residual resistivity ratio of 100 was assumed for niobium. For MgB$_2$, the values of $\lambda$ and $\xi$ are experimental values given in the reference. For calculations, $H_c=\phi_0 / [\mu_0(2\sqrt{2}\pi\xi\lambda)]$ is used \cite{tinkham96}. $H_{c1}$ for Nb is found from fit to numerically computed data in Ref.~\cite{Hein1999} and [\onlinecite{Harden1963}]. $H_{c1}$ for strongly type II materials is found from Eq.~5.18 in Ref.~\cite{tinkham96}. $\ensuremath{H_\mathrm{sh}}$ is calculated using $\ensuremath{H_\mathrm{sh}}\simeq H_c(0.75+0.54 \, \kappa^{-1/2})$ \cite{transtrum11}. The condensation energy density $F$ is given by $\mu_0H_c^2 / 2$ \cite{tinkham96}. Nb data is extracted from Ref.~\cite{Maxfield1965}, Nb$_3$Sn data from Ref.~\cite{Hein1999}, NbN data from Ref.~\cite{Oates1991}, and MgB$_2$ data from Ref.~\cite{Wang2001}.} \label{tab:MaterialsParameters} \end{center} \end{table*} One key question for our purposes is whether thermal fluctuations can help activate vortices over the surface barrier. Thermal fluctuations in most superconductors (apart from the high-$T_c$ cuprate superconductors) are very small. This is due to the same approximation that makes the BCS theory of superconductors so successful. BCS theory is a {\em mean-field theory} of interacting Cooper pairs, which becomes exact when each Cooper pair interacts with an infinite number of neighbors (thus seeing the {\em mean} behavior of the system). Each Cooper pair is of radius roughly the coherence length $\xi$, so BCS theory will be valid when the density of Cooper pairs times $\xi^3$ is large. Simple estimates show that there are about $10^6$ centers of Cooper pairs within the region occupied by each pair state; a scenario where the pairs strongly overlap in space, and each pair only feels the average occupancy of the other pair states~\cite{schrieffer99}. Thermal fluctuations of vortices will be unimportant so long as the condensation energy density---the amount of energy $F$ that is necessary to destroy superconductivity over a unit volume---times $\xi^3$, is large compared to $k_B T$. Table~(\ref{tab:MaterialsParameters}) gives the characteristic temperature $T_{th} = F \xi^3 / k_B$ where fluctuations will become important, for niobium and also three candidate materials being explored for next generation accelerating cavities. Only for NbN is this characteristic temperature remotely comparable to $T_c$. We can gain further insight from an analytic calculation of $ E_v / ( k_B \, T ) $, where $E_v$ is the energy per unit length of a vortex line integrated over a coherence length $\xi$. Using results from BCS theory, the zero-temperature thermodynamical critical field is given by $H_c (0) = 2\, \sqrt{\pi} \sqrt{ \mathcal{N}(0)} \Delta$, where $\mathcal{N}(0) = m \, k_F / (2 \, \pi^2 \, \hbar^2)$ is the density of states at the Fermi energy, $\Delta$ is the superconductor gap at zero temperature, and $k_F$ is the Fermi wave number. Also, $\Delta \approx 1.76 \, k_B \, T_c$, and the coherence length $\xi_0 = \hbar v_F / (\pi \Delta)$, where $v_F$ is the Fermi velocity. Thus, \begin{eqnarray} \frac{ E_v }{ k_B \, T } \sim \frac{{H_c}^2 \xi^3}{k_B T} \approx \frac{ 1.4 }{ t } \left( \frac{ \varepsilon_F }{ \Delta } \right)^2, \end{eqnarray} where $ \varepsilon_F = \hbar^2 \, {k_F}^2 / ( 2 \, m ) $ is the Fermi energy, and $ t = T / T_c $. Since the gap is much smaller than the Fermi energy, we can neglect thermal nucleation of vortices; unlike the case of superheated water, the effects of thermal fluctuations is very small. More generally, we expect that $\tau_{\mathrm{mic}} \ll \tau_{\mathrm{cav}} \ll \tau_{\mathrm{t.n.v.}}$ within the Meissner metastable state, where $\tau_{\mathrm{mic}}$, $ \tau_{\mathrm{cav}}$, and $\tau_{\mathrm{t.n.v.}}$ correspond to time scales associated with microscopic degrees of freedom, the variation of the cavity fields, and thermal nucleation of vortices, respectively. The negligible effects of thermal fluctuations tells us that estimating the limiting superheating field of a perfectly clean surface will not be analogous to bubble formation for superheated water. Instead, we shall use {\em linear stability theory} in Section~(\ref{subsec:linearstability}) to estimate the field at which the uniform Meissner state becomes energetically unstable to an infinitesimal perturbation in the space of magnetic fields and superconducting order. A variant of critical droplet theory will appear in Section~(\ref{subsec:disorder}), where we estimate the effects of flaws and disorder in nucleating vortex penetration. \section{Laminates and vortex penetration} \label{sec:laminates} In recent years, much effort in superconducting RF has been devoted to exploring single or multiple thin films -- laminated structures hopefully tunable to optimize performance. This section is devoted to exploring possible advantages to such laminates. The work in this section relies heavily on extensive discussions and consultation with Alex Gurevich, whose work prompted most of the calculations presented. In practical terms, two of the candidate materials (Nb$_3$Sn and NbN) can be grown by deposition on Nb surfaces, so fabricating a surface layer onto a Nb cavity leverages existing expertise. Gurevich points out~\cite{Gurevich2013} that thermal conductivities of new candidate materials are often small; since the heat generated by the surface residual resistance at the surface must be conducted through the cavity, keeping the thickness of these new materials small can improve performance. (For Nb$_3$Sn, recent surface resistances have been small enough, at least at low fields, that thickness may not be an issue.) Gurevich has also proposed {\cite{Gurevich2006} separating one or more superconducting layers by insulating layers (a SIS geometry). Calculations show~\cite{posen15} that laminates do not substantially improve the theoretical maximum superheating field in AC applications beyond that of pure materials (or thick layers) for the film-insulator-bulk structure,% \footnote{\label{foot:IsolatedLayer} A free-standing superconducting layer (or a layer surrounded by insulators) with thickness small compared to the magnetic penetration depth $\lambda$ can have an enormous superheating field (since it can remain superconducting without paying most of the cost of expelling the flux). In the accelerator community, there is widespread focus on raising this `$H_{c1}$' for the superconducting film~\cite{feliciano16,lamura09} -- defined, somewhat unphysically~\cite{Hc1WrongRef} as the minimum field needed for a vortex to be stable parallel to and inside the film. But such an in-film stable vortex configuration demands magnetic flux on {\em both sides} of the film. In a GHz AC application, pushing the flux through the film twice per cycle generates unacceptable heating~\cite{Hc1WrongRef}. Besides, any such parallel vortex would be precariously unstable to formation of two vortex pancakes. A thin superconducting layer with a large magnetic penetration depth atop a lower-\ensuremath{H_\mathrm{sh}}\ layer with a small penetration depth can have modestly higher superheating fields, due to the way the bottom layer modifies the magnetic field penetration. } though adding a thin S$^\prime$ layer on the bulk S superconductor may lead to an enhancement of the energy barrier~\cite{Kubo2016a,Checchin2016}. Gurevich has suggested that the SIS geometry may have a different advantage -- reducing the impact of flux penetration. Our calculations in Section~(\ref{subsec:disordmeander}) suggest that SIS films with thickness $d$ small compared to the London penetration depth $\lambda$ will be more susceptible to vortex penetration than bulk films; the damage zone needed for vortex nucleation at fields below pure $\ensuremath{H_\mathrm{sh}}$ can be thinner by the fraction $\lambda/d$, presumably making them much more likely. Also, one would naively expect it to be harder to grow low-defect two-layer laminates than depositing a single layer or preparing a pure surface. Layers thick compared to the penetration depth would presumably behave similarly to a bulk material; vortices deeper than $\lambda$ do not `feel' the surface except insofar as other vortices penetrating the surface push them deeper. The dynamics after flux penetration will be substantially different for the SIS geometry than for a simple 3D superconducting surface. In either case, a flaw may nucleate one to several vortex entries when the field increases in one direction; some or all may be `pulled back' as the field shifts to the opposite direction. If the nucleation center flaws are rare and the vortices do not build up over time, they need not cause local heating enough to cause a quench. But since the RF cavities operate at GHz frequencies, and each flaw could (or should) generate multiple vortices per cycle, potentially billions of vortices could be introduced by a single flaw if they can escape re-annihilation. In three dimensions, a vortex penetrating at a point $(y,z)$ on the surface will grow in the $z$ direction pointing along the field as it penetrates a depth of order $x\sim\lambda$ (Fig.~(\ref{fig:semiloops}) left). If multiple vortices enter, they may push and entangle one another; as they interact with disorder in the material they may exhibit avalanches~\cite{aranson01, aranson05}. During the field reversal, the points where the vortices exit the material will be forced together along the $z$ direction (shrinking in length), and new vortices with opposite winding number will nucleate (potentially annihilating some or all of the old vortices). Even if this process is incomplete, leaving some tangle of vortex loops, it may enter a kind of limit cycle. Indeed, many periodically stressed disordered dynamical systems can enter into limit cycles at low levels of stress, with a transition to `turbulent' aperiodic behavior at a critical threshold (colliding colloids in reversing low-Reynolds number flows~\cite{corte08}, plasticity in vortex structures of superconductors~\cite{daroca11,motohashi11,okuma11}, etc). It is possible that the quench of RF cavities explores precisely this kind of dynamical phase transition, separating a local hot spot from an invading front of vortices. Apart from these brief speculations, we will not discuss three-dimensional dislocation dynamics further in this work; the remainder of this section will focus on the SIS geometry. In the two-dimensional SIS geometry, a vortex penetration event may end with the vortex trapped in the insulating layer, leaving two 2D vortices penetrating the outer superconducting film (Fig.~(\ref{fig:semiloops}) right. See also footnote~\ref{foot:IsolatedLayer}.) Such 2D vortices, called {\em pancake vortices}, have been studied in great detail~\cite{huebener01} in the context of high temperature cuprate superconductors, some of which are well described as nearly decoupled 2D superconducting sheets. A vortex pair nucleated by a defect at $(y, z)$ on the surface will separate along the $z$ direction as the field increases, be buffeted by thermal fluctuations, dirt, defects, and other vortices as they separate, and then be pulled back along the $z$ direction as the field reverses. (Some of the other vortices will be emitted by the same defect, once the initial pair departs and the resulting long-range suppression of nucleation drops, see Section~\ref{subsec:interacting}.) In this part we shall explore Gurevich's suggestion that, even after billions of cycles, this annihilation should be effective at avoiding vortex escape (presumably preventing a buildup of vortices which otherwise would lead to a quench). In Section~(\ref{subsec:impactParameter}), we introduce an ``impact parameter,'' the amount of lateral vortex separation between a vortex-antivortex pair that can be tolerated during a cycle while still expecting them to annihilate, in Section~(\ref{subsec:thermal}), we examine the expected lateral meandering distance expected from pancake vortices in an RF cycle, in Section~(\ref{subsec:disordmeander}), we examine the expected meandering due to disorder, and in Section~(\ref{subsec:interacting}), we briefly consider the effect of vortex-vortex interaction and the situation of two nearby defects. \begin{figure}[!h] \centering \includegraphics[width=0.9\linewidth]{Sec/Fig/fig11.pdf} \caption{Vortices in a bulk superconductor for semiloops (left). Vortices in thin superconducting films separated by insulators form pancakes.\label{fig:semiloops}} \end{figure} \subsection{Impact Parameter} \label{subsec:impactParameter} How far $\Delta x$ perpendicular to the field must a vortex pair migrate before their mutual attraction ceases to be strong enough to annihilate them at the end of a cycle? Figure~(\ref{fig:separatrix}) shows the trajectories for a pancake pair as they return at the end of a cycle, separated by different distances $x$ perpendicular to the external magnetic field, using the vortex interaction formulation from Ref. \cite{Clem1991}. There is a {\em separatrix} between trajectories which collide and trajectories which miss each other. We will call the value of $\Delta x$ at this separatrix the {\em impact parameter}, $x_{imp}$. For perhaps credible parameters $d=30$ nm, $\lambda=100$ nm, $\mu_0\ensuremath{H_\mathrm{sh}}=0.4$ T, $x_{imp} \sim 20$ nm. Simulations were used to evaluate $x_{imp}$ as a function of field, and the results are plotted in Fig.~(\ref{fig:wandercalc}). \begin{figure}[btp] \begin{center} \includegraphics[width=0.9\linewidth,angle=0]{Sec/Fig/fig12.pdf} \end{center} \caption{For initial separations smaller than the impact parameter $x_{imp}$, the pancakes annihilate (inner solid, dash, dot), but larger than $x_{imp}$, they can wander away (dash-dot, outer solid). $\alpha=0.2$ shown} \label{fig:separatrix} \end{figure} \begin{figure}[btp] \begin{center} \includegraphics[width=0.9\linewidth,angle=0]{Sec/Fig/fig13.pdf} \end{center} \caption{This figure compares $x_{imp}$, the maximum lateral separation resulting in impact of a vortex pair, to $\sqrt{\left<{x_{thermal}^2}\right>}$ and $\sqrt{\left<{x_{disorder}^2}\right>}$, the expected meandering distances due to thermal and disorder effects, for realistic parameters given in the text. Note that the former remains a factor of at least ten larger than the latter, suggesting that vortex escape by these mechanisms is a $10\sigma$ event. Thus neither thermal motion nor disorder is dangerous, according to our estimates, to prevent nucleated pancake vortices from annihilating with extremely high probability at the end of every cycle.} \label{fig:wandercalc} \end{figure} \subsection{Thermal Meandering} \label{subsec:thermal} The motion due to thermal fluctuations can be estimated using the Einstein equation, \begin{equation} \left<{x_{thermal}^2}\right>=2Dt \label{eq:einstein} \end{equation} where $x_{thermal}$ is the displacement in time $t$ due to thermal motion and $D$ is a diffusion constant. For one RF cycle at frequency $f$, $t=f^{-1}$. $D\approx\mu k_B T$, where $k_B$ is Boltzmann's constant, $T$ is the temperature and $\mu$ is the mobility of the vortex, given by Bardeen Stephen as $\rho_n/(H_{c2}\,\phi_0\, d)$, where $\rho_n$ is the normal state resistivity, $H_{c2}$ is the upper critical field and $\phi_0$ is the flux quantum. Solving, the wandering due to thermal motion is given by: \begin{equation} \sqrt{\left<{x_{thermal}^2}\right>}=\sqrt{\frac{2\,k_B \, T\, \rho_n}{H_{c2} \, d \, f \, \phi_0}} \label{eq:einstein2} \end{equation} For realistic parameters $T=2$ K, $\rho_n=100$ n$\Omega$m; $\mu_0H_{c2}=30$ T, $f=1.3$ GHz, $d=30$ nm, $x_{thermal} = 1.5$ nm, as shown in Fig.~(\ref{fig:wandercalc}). From these results, we can calculate the approximate expected rate of production of vortices that fail to annihilate. One expects that the distribution of final separations will be Gaussian with standard deviation $x_{thermal}$, suggesting that the number of vortices which do not annihilate will be given by the tail of the Gaussian. For example, at $H=0.8\ensuremath{H_\mathrm{sh}}$, $x_{imp}$ is about 22 nm from Fig.~(\ref{fig:wandercalc}), or about 15 standard deviations, making it extremely unlikely for vortices to escape due to thermal meandering alone. \subsection{Disorder Meandering} \label{subsec:disordmeander} \begin{figure}[!h] \centering \includegraphics[width=0.9\linewidth]{Sec/Fig/fig14.pdf} \caption{Surface disorder may cause pancake vortices to meander away from a nucleation site and build up in a film over many RF cycles. \label{fig:disorder}} \end{figure} To calculate the wandering due to surface disorder, illustrated in Fig.~(\ref{fig:disorder}), we consider a single-cell $f$=1.3 GHz niobium SRF cavity with an SIS structure using $d=30$ nm thick Nb$_3$Sn layers. Assume that the topmost S layer has a normally-distributed random array of defects over its surface. For our geometry, we divide the $L\times L$ (where $L\sim$10 cm) surface area of the cavity into $N$ $a\times a$ regions of order the pancake vortex size, where $N=L^2/a^2$. We represent the effect of these defects as lowering the local value of $B_c$ in a given region. Therefore these defects will nucleate vortex penetration, and they will attract pancake vortices in the film. At the worst of the defects, the expected value for $H_c$ is $\alpha H_{c,nominal}$, where $\alpha$ is a constant between 0 and 1. At this defect, vortices penetrate at approximately $H=\alpha \ensuremath{H_\mathrm{sh}}$. We represent the surface of the cavity with a distribution $\mathcal{H}$ of values for the reduction in the square of $H_c$. For simplicity of analysis, and since the defects are normally distributed, we will use the notation generally used in propagation of random errors. \begin{equation} \mathcal{H}=H_{c,nominal}^2(1-|(0\pm\sigma)|) \label{Bdist} \end{equation} Here is $\sigma$ is the variance of the normally distributed $H_c^2$ reduction. We use absolute values because there should be no $H_c$ values higher than the nominal value. We can find $\sigma$ using our restriction that the expected value for $H_c$ at the worst defect is $\alpha H_{c,nominal}$. To do this, we need to work with a normal distribution of the $H_c^2$ reduction in our N regions $\mathcal{N}(0,\sigma)=[1/(\sigma\sqrt{2\pi})] \exp [ x^2 / (2\,\sigma^2)]$. First, we need to find $\phi$, the value of $x$ above which lies just one half of one of our N regions (one half because the absolute value effectively doubles the number of samples in our integration). \begin{equation} \int_{\phi}^\infty \frac{1}{\sigma\sqrt{2\pi}}e^{-x^2 / (2\, \sigma^2)} \mathrm{d}x=\frac{1}{2}\left[1-\mathrm{erf}\left(\frac{\phi}{\sqrt{2}\sigma}\right)\right]=\frac{1}{2N} \label{sigmadef1} \end{equation} Next, we set the expected value of the distribution in this region to be 1-$\alpha^2$. This sets the expected value of $H_c$ to be $\alpha H_c$ for the most extreme defect (we also have to normalize for there being only one defect in our sample size). This defines $\sigma$. \begin{equation} \int_{\phi}^\infty \frac{x}{\sigma\sqrt{2\pi}}e^{-\frac{1}{2}\left(\frac{x}{\sigma}\right)^2} \mathrm{d}x=\frac{\sigma}{\sqrt{2\pi}}e^{-\frac{1}{2}\left(\frac{\phi}{\sigma}\right)^2}=\frac{1-\alpha^2}{2N} \label{sigmadef2} \end{equation} To obtain an analytical expression, instead of solving equations (\ref{sigmadef1}) and (\ref{sigmadef2}), we can approximate. First, estimate that $\phi\approx\alpha$: \begin{equation} \int_{\alpha}^\infty \frac{1}{\sigma\sqrt{2\pi}}e^{-\frac{1}{2}\left(\frac{x}{\sigma}\right)^2} \mathrm{d}x=\frac{1}{2N} \label{sigmasimp1} \end{equation} \begin{figure}[htbp] \begin{center} \includegraphics[width=0.9\linewidth,angle=0]{Sec/Fig/fig15.pdf} \end{center} \caption{Approximations to find an analytical expression for $\sigma$.} \label{fig:meandering} \end{figure} We can also use a linear approximation for the Gaussian, so that the integral can be evaluated by calculating the area of the triangle shown in Fig.~(\ref{fig:meandering}). The equation of the line evaluated at $y=0$ determines the length $\Delta x$ in the figure: \begin{equation} 0=\frac{-\sigma \alpha}{\sqrt{2\pi}}e^{-\frac{1}{2}\left(\frac{\alpha}{\sigma}\right)^2}\left(\Delta x\right)+\frac{1}{\sigma\sqrt{2\pi}}e^{-\frac{1}{2}\left(\frac{\alpha}{\sigma}\right)^2} \label{sigmasimp2} \end{equation} This gives $\Delta x=\sigma^2/\alpha$. Evaluating the integral, \begin{equation} \int_{\alpha}^\infty \frac{1}{\sigma\sqrt{2\pi}}e^{-\frac{1}{2}\left(\frac{x}{\sigma}\right)^2} \mathrm{d}x=\frac{\sigma}{2\alpha\sqrt{2\pi}}e^{-\frac{1}{2}\left(\frac{\alpha}{\sigma}\right)^2}=\frac{1}{2N} \label{sigmasimp3} \end{equation} Rearranging, \begin{equation} \frac{\alpha^2}{2\sigma^2}=\ln{\frac{\sigma N}{\alpha\sqrt{2\pi}}} \label{sigmasimp4} \end{equation} N is very large, so $\ln{\frac{\sigma N}{\alpha\sqrt{2\pi}}}\approx\ln{N}$. Solving, \begin{equation} \sigma=\frac{\alpha}{\sqrt{2\ln{N}}} \label{sigmasimp5} \end{equation} Now let us consider the behavior of the cavity at a field just above the expected vortex penetration field of the worst defect. Let $\hat{x}$ point into the film and $\hat{z}$ be aligned with the magnetic field. In this case, pairs of pancake vortices will form at the defect, move apart in $y$ due to the force exerted by the increasing magnetic field, and then move in the opposite direction in $y$ as the magnetic field direction reverses. In the mean time, they will have meandered some distance in $x$. If the meandering distance is very small compared to the impact parameter (approximately 50 nm), the pancake vortex pairs are most likely to meet again and annihilate. However, if the meandering distance is significant compared to the impact parameter, vortex pairs are likely to get lost, accumulating over the film. For this calculation, we will not do a full simulation. Rather, we will consider a path along $y$ of a vortex pancake moving from the defect at $y=0$ to the extremum $y=y_{max}$, without any movement in $x$. We will, however, integrate the force in $x$ that the pancake would experience along its path, and calculate what the expected meandering distance would be from this. The vortex will have two regions on either side of it in $x$ at a given time, one with $H_{c,1}$, and one with $H_{c,2}$. The centers of these two regions are separated by a distance $a$. The $x$-component of the force experienced by the vortex can be approximated from the gradient in condensation energy resulting from the difference in the local $H_c$. Magnetic energy density is given by $H^2/2\mu_0$. We convert this to energy by multiplying by the volume of a region, $a^2d$. \begin{equation} F_x=\frac{H_{c,1}^2-H_{c,2}^2}{2\mu_0a}a^2d \label{forcedef} \end{equation} The motion of the pancake is described by Bardeen-Stephen drag: \begin{equation} \eta v_x=F_x/d \label{BS} \end{equation} Here $v_x$ is the x-component of the velocity of the pancake and $\eta$ is the Bardeen-Stephen drag coefficient, defined by $\eta=H_{c2}\phi_0/\rho_n$, where $H_{c2}$ is the upper critical field of the film, $\phi_0$ is the flux quantum, and $\rho_n$ is the normal resistivity of the film. The meandering distance $\Delta x$ within a region is given by $\Delta x=v\Delta t$, where $\Delta t$ is the time spent in that region. For simplicity, we approximate that the vortex moves with uniform speed in $y$, spending equal time in each of the regions that it travels through, such that $\Delta t=\frac{1}{2f}\frac{a}{y_{max}}$. Using Eq. Eq.~(\ref{Bdist}) and Eq.~(\ref{forcedef}), we can describe a distribution $\mathcal{F}_x$ of forces that the pancake would experience. We observe that taking the difference between the absolute values from Eq.~(\ref{Bdist}) produces a normal distribution: \begin{equation} \mathcal{F}_x=\frac{a d H_{c,nominal}^2}{2\mu_0}(0\pm\sigma) \label{Fdist} \end{equation} Between two regions that are adjacent in $x$, the pancake will experience a single value of the distribution of force values. After it travels a distance $a$, it will encounter a new pair of regions and therefore a new force value. We can use $\Delta x=v\Delta t$ and Eq.~(\ref{BS}) to find $\sqrt{\left\langle\Delta x^2\right\rangle}$, the RMS wandering distance traveled by the vortex over a distance $a$: \begin{equation} \sqrt{\left\langle\Delta x^2\right\rangle}=\sqrt{\left\langle F^2\right\rangle}\frac{\Delta t}{d\eta} \label{x1dist} \end{equation} Multiply by the square root of the number of steps in a period $N_{steps}=\frac{2y_{max}}{a}$ to find the total RMS wandering distance $\sqrt{\left\langle\Delta X^2\right\rangle}$. Use Eq.~(\ref{Fdist}). \begin{equation} \sqrt{\left\langle\Delta X^2\right\rangle}=\frac{a d \sigma H_{c,nominal}^2}{2\mu_0}\frac{\Delta t}{d\eta}\sqrt{\frac{2y_{max}}{a}} \label{xdist1} \end{equation} Since $\Delta t=\frac{a}{v_y}$, from Eq.~(\ref{BS}) and using $F_y=\frac{\phi_0 H}{\mu_0}$, we obtain: \begin{equation} \frac{\Delta t}{\eta}=\frac{a d}{F_y}=\frac{a d \mu_0}{\phi_0 \Delta H} \label{dteta} \end{equation} Here $\Delta H$ is the difference in magnetic field across the film. If we are looking just above the penetration field, $\Delta H\approx\alpha \ensuremath{H_\mathrm{sh}} d/\lambda$. Eq.~(\ref{dteta}) can also be used to find $\frac{y_{max}}{a}$. \begin{equation} \frac{y_{max}}{a}=\frac{1}{2 f \Delta t} = \frac{\phi_0 \alpha \ensuremath{H_\mathrm{sh}}}{2 f \eta a \lambda \mu_0} \label{ymaxa} \end{equation} Use the definition of $\eta$. \begin{equation} \frac{y_{max}}{a}=\frac{\rho_n \alpha \ensuremath{H_\mathrm{sh}}}{2 f H_{c2} a \lambda \mu_0} \label{ymaxa2} \end{equation} Now substitute. \begin{equation} \sqrt{\left\langle\Delta X^2\right\rangle}=\frac{1}{2\sqrt{\ln{N}}}\frac{a^2 \lambda H_{c,nominal}^2}{\phi_0 \ensuremath{H_\mathrm{sh}}}\sqrt{\frac{\rho_n \alpha H_sh}{f H_{c2} a \lambda \mu_0}} \label{xdist2} \end{equation} We then use $H_{c,nominal}=\frac{\phi_0}{2\sqrt{2}\pi\lambda\xi}$. \begin{equation} \sqrt{\left\langle\Delta X^2\right\rangle}=\frac{\xi}{4\sqrt{2}\pi\sqrt{\ln{N}}}\frac{H_{c,nominal}}{\ensuremath{H_\mathrm{sh}}}\sqrt{\frac{\rho_n \alpha \ensuremath{H_\mathrm{sh}}}{f H_{c2} a \lambda \mu_0}} \label{xdist3} \end{equation} We can set our region size $a=\xi\approx 4$ nm. $\ensuremath{H_\mathrm{sh}}/H_{c,nominal}$ for Nb$_3$Sn is approximately 0.75. $N$ is approximately $10^{15}$, so $\frac{1}{4\sqrt{2}\pi\sqrt{\ln{N}}}$ is approximately $10^{-2}$. We also use $H_{c2}$=30 T, $\rho_n$=10 n$\Omega$m, and $H_{c,nominal}=530$ mT. For $\alpha=0.8$, these factors give $\sqrt{\left\langle\Delta X^2\right\rangle}\approx 2.1$ nm. Setting this as a standard deviation for vortex meandering and using $x_{imp}\sim$22 nm from above gives a result of 11 standard deviations, again making it extremely unlikely for vortices to escape due to disorder meandering. Varying $a$ and $\alpha$ over a reasonable range also gives a small value for the meandering distance, as seen in Fig.~(\ref{fig:var}). \begin{figure}[htbp] (a) \par\smallskip \centering \includegraphics[width=0.9\linewidth]{Sec/Fig/fig16a.pdf} \\ (b) \par\smallskip \centering \includegraphics[width=0.9\linewidth]{Sec/Fig/fig16b.pdf} \caption{The effect of varying $a$ for fixed $\alpha=0.8$ (a) and varying $\alpha$ for fixed $a=4$ nm (b) on the mean meandering distance of a pancake vortex due to disorder.} \label{fig:var} \end{figure} In addition to the impact parameter and thermal meandering distance, Fig.~(\ref{fig:wandercalc}) plots the meandering distance due to disorder using the analytical formulation above. Also plotted in the figure is a simulation of vortex pair creation and annihilation in the presence of disorder. In addition to forces from the applied magnetic field and from the randomly distributed array of circular defects with radius $a$,% \footnote{In the simulation, the defects pin vortices with force that increases linearly from the defect edge where it is zero, to the center of the defect, where it is a maximum. The maximum is set such that moving the vortex from the center of the defect to the edge would require work equal to the condensation energy of the volume of the vortex core.} the simulation considers the forces of one vortex on the other, and finds the maximum lateral separation between the pair during a cycle. \subsection{Interacting vortices, interacting defects} \label{subsec:interacting} We have presented analyses of thermal meandering and meandering of a vortex-antivortex pair due to disorder, and so far, nothing has caused a large buildup of vortices in the film. We have also performed a preliminary analysis of the nucleation of many vortices at a defect over the course of an RF cycle, all of which interact with each other. While an attractive force exists between vortices and antivortices, vortices exert a repulsive force on other vortices, and antivortices exert a repulsive force on other antivortices. These repulsive forces between similar vortex types appear to result in substantially larger lateral movement, which could lead to higher rates of non-annihilation. An example video is included as supplemental material~\cite{url:pancakevideo}; three frames of which are shown in Fig.~(\ref{fig:video}). Here the film thickness is 30 nm, parameters are appropriate for Nb$_3$Sn material, and the disorder is modeled as 60 pinning sites randomly spread over 0.4 square microns that exert a force of $F_{\mathrm{max}} \, k_i \, \rho$, where $F_{\mathrm{max}}=B_c^2 \, \mu_0 \, \xi^2 \, d \, r_{\mathrm{pin}} /\, 2$, $k_i$ is a value between 0 and 1 that is randomly chosen for each pinning site, and $\rho=r_{\mathrm{sep}}/r_{\mathrm{pin}}$ for $r_{\mathrm{sep}}<r_{\mathrm{pin}}$ and 0 for $r_{\mathrm{sep}}>r_{\mathrm{pin}}$, where $r_{\mathrm{sep}}$ is the distance between the vortex and the center of the pinning site and $r_{\mathrm{pin}}$ is the pinning site radius, (here chosen to be 2 nm). Note that the maximum horizontal meandering due to disorder and interactions is roughly eight nanometers, only a factor of three less than the impact parameter $x_{imp}$, suggesting that interactions may be much more dangerous for multiple defects over many cycles. However, further work on interactions is clearly needed; it is possible that realistic parameters for the disorder strength could decrease the meandering, and it is possible that the distribution of maximum meandering distances near a defect over multiple cycles could be sub-Gaussian. On the other hand, it is likely that two defects that both nucleate pancake vortices and are close enough together that the vortices interact could be dangerous even in the absence of disorder. \begin{figure}[htbp] \begin{center} (a)\includegraphics[width=0.7\linewidth,angle=0]{Sec/Fig/fig17a.pdf} (b)\includegraphics[width=0.7\linewidth,angle=0]{Sec/Fig/fig17b.pdf} (c)\includegraphics[width=0.7\linewidth,angle=0]{Sec/Fig/fig17c.pdf} \end{center} \caption{Frames from a preliminary simulation of pancake vortices (red) and antivortices (blue) being generated at a defect (green) in a thin film in an AC field. The repulsion between similar vortices causes a lateral spread. Note the horizontal and vertical scales differ by over a factor of 100. The the horizontal lines are actually circular representations of the disorder potential; the vortices drift sideways much less than they traverse vertically. (a)~Near zero field, (b)~Near a local maximum field, and (c)~Near a return to zero field, showing the cycle-to-cycle variation. } \label{fig:video} \end{figure}

The modern era of clinical optical coherence tomography (OCT) imaging and antiangiogenic therapy emerged and converged during the early years of the 21st century. This convergence of pharmaceutical and imaging technologies revolutionized the care of patients with exudative retinal diseases, most notably neovascular (wet) age-related macular degeneration (AMD), diabetic macular edema, proliferative diabetic retinopathy, and retinal vein occlusions. A common mediator of exudation in all these diseases turned out to be vascular endothelial growth factor (VEGF), and the development of VEGF inhibitors coincided with the commercial availability of the first commercially successful OCT instrument, known as the OCT III or Stratus OCT (Carl Zeiss Meditec, Dublin, CA, USA), a time-domain OCT (TD-OCT) instrument capable of performing six diagonal line scans at a rate of 400 A-scans per second. The first OCT descriptions of exudation performed with these Zeiss time-domain instruments provided revelatory insights as to the nature of fluid in the retina, under the retina, and under the retinal pigment epithelium (RPE) ([Figs. 1](#i1552-5783-57-9-OCT14-f01){ref-type="fig"}--[5](#i1552-5783-57-9-OCT14-f05){ref-type="fig"}). {#i1552-5783-57-9-OCT14-f01} Prior to OCT imaging, the gold standard was fluorescein angiography (FA), and to a lesser extent, indocyanine green angiography (ICGA). These dye-based imaging strategies provided two-dimensional images of neovascularization and exudation, and if an examiner was schooled in the art of stereoscopic viewing, the relative location of the neovascular lesion and the magnitude of the exudation could be appreciated, although this information was hard to share with patients and colleagues unless they too could view the images stereoscopically. Much of our understanding of these exudative processes came from descriptions provided by experts using slit-lamp biomicroscopic viewing of the fundus followed by stereoscopic viewing of dye-based angiographic images.^[@i1552-5783-57-9-OCT14-b01]^ As someone who spent his early career training residents and retina fellows on how to stereoscopically examine the fundus and view angiographic images, I was always uncertain whether the student truly saw what needed to be seen. With the development of OCT imaging, all that changed. Using TD-OCT, both ophthalmologist and patient could readily appreciate the anatomic nuances of exudation, and with each improvement in OCT technology, the image quality improved to the point where OCT is now the gold standard for imaging and managing patients with exudative macular diseases, especially when undergoing antiangiogenic therapies. TD-OCT and SD-OCT Imaging of Neovascular AMD {#s2} ============================================ The Stratus TD-OCT had a scanning rate of 400 A-scans per second and generated a standard scanning pattern that consisted of six radial, concentric, 6-mm-long B-scans centered on the fovea and offset by 30°.^[@i1552-5783-57-9-OCT14-b02]^ Image quality often depended on the cooperation of the patient and the skill of the operator, since tracking was not yet available and higher-density scans took longer to acquire. Over the past 15 years as OCT technology has evolved, TD-OCT has given way to spectral-domain OCT (SD-OCT). Spectral-domain OCT, a type of Fourier-domain OCT (FD-OCT) with scanning rates of approximately 27,000 to 70,000 A-scans per second, resulted in denser, more varied scanning patterns, better image quality, better reproducibility, and large volumetric datasets that could be viewed using both the traditional cross-sectional approach and an en face approach.^[@i1552-5783-57-9-OCT14-b03]^ Compared with the more advanced SD-OCT technology now available in the clinics, the capabilities of the Stratus TD-OCT instrument now seem primitive, yet at the time, the images obtained using TD-OCT imaging provided a view of macular neovascularization (MNV) and exudation that would forever change the way we managed patients. The B-scan image of a normal macula shows layers of differing reflectivity that approximate the major cellular layers of the retina and choroid.^[@i1552-5783-57-9-OCT14-b04]^ However, features of the choroidal anatomy were poorly visualized using TD-OCT imaging. Thus, by using TD-OCT, it was possible to obtain a crude optical biopsy of the macula and distinguish between a normal retina/RPE contour and an abnormal one, which was all that was needed for the retinal specialist to identify exudation. Moreover, one of the most commonly used quantitative parameters derived from OCT datasets was the retinal thickness measurement, which was obtained by segmenting the internal limiting membrane (ILM) and a boundary representing the RPE. This information was used to generate retinal thickness maps, and these maps proved useful for identifying and describing deviations from normal anatomy and how these deviations changed over time. However, it was the viewing of individual B-scans that provided the retina specialist with a powerful strategy for following disease progression and response to treatment. While it came as no surprise that OCT imaging was extremely useful for detecting the macular edema expected in cases of diabetes and retinal vein occlusions, it was a surprise to most retina specialists that intraretinal fluid was commonly found as an early and prominent component of the exudation associated with neovascular age-related macular degeneration (nvAMD).^[@i1552-5783-57-9-OCT14-b05]^ The classification system previously developed for nvAMD was based on leakage patterns seen on dye-based angiography, and this leakage was thought to result in the accumulation of fluid under the retina and under the RPE. This made intuitive sense since the neovascularization was thought to arise under the retina from the choroidal circulation, so why would there be fluid in the retina? In retrospect, the detection of intraretinal fluid turned out to be one of the most novel and important observations made when using TD-OCT for imaging nvAMD. At the time, one explanation for the presence of intraretinal edema was that the subretinal exudation had extended into the retina, but the extension of subretinal fluid into the retina was not commonly seen in other conditions associated with subretinal fluid such as central serous chorioretinopathy (CSCR) and retinal detachment. Moreover, eyes with early CSCR and subretinal fluid, prior to the onset of fibrosis and photoreceptor disorganization, tended to have good best-corrected visual acuity while eyes with nvAMD and similar amounts of subretinal fluid always seemed to have worse vision due to the intraretinal fluid. An alternative explanation for the intraretinal exudation was that this fluid came from established mature retinal vessels, and the exudation was in response to VEGF. After all, intraocular injections of VEGF into primate eyes had previously been shown to cause retinal edema and hemorrhages.^[@i1552-5783-57-9-OCT14-b06],[@i1552-5783-57-9-OCT14-b07]^ Thus, the presence of macular fluid came to be appreciated as a useful surrogate measure for the presence of excess VEGF. Another explanation for intraretinal fluid was that the neovascularization could be arising from within the retina, and this neovascularization could be directly responsible for the intraretinal macular fluid. In the early days of OCT, this hypothesis seemed far-fetched, but now we know that type 3 MNV (known as retinal angiomatous proliferation) can arise from within the retina and is associated with intraretinal edema.^[@i1552-5783-57-9-OCT14-b08]^ While type 3 MNV accounts for a minority of neovascular lesions and macular edema is a common feature of all MNV, the most likely explanation for the presence of intraretinal fluid is that VEGF induces exudation from the mature retinal vessels in nvAMD.^[@i1552-5783-57-9-OCT14-b09]^ Thus, lessons learned from early OCT imaging of nvAMD included an appreciation that intraretinal fluid was one of the first signs of VEGF-mediated exudation ([Figs. 2](#i1552-5783-57-9-OCT14-f02){ref-type="fig"}--[5](#i1552-5783-57-9-OCT14-f05){ref-type="fig"}). ![Late fluorescein angiographic (FA) images and time-domain optical coherence tomography (OCT) diagonal B-scans from a patient treated with intravitreal ranibizumab in the phase II study, then followed in the extension study with observation and no additional retreatments with ranibizumab.^[@i1552-5783-57-9-OCT14-b17]^ (**A**, **C**, **E**) Late FA images. (**B**, **D**, **F**) Diagonal OCT B-scans. (**A**) Baseline late FA image showing diffuse leakage. Visual acuity (VA) was 20/125. (**B**) Corresponding baseline OCT image showed cystic macular edema (*arrow*). (**C**) After eight injections over 140 days, the FA leakage had significantly diminished, and VA improved to 20/80. (**D**) Corresponding OCT B-scan showed resolution of the macular edema. (**E**) After an additional 22 months without any ranibizumab injections, VA improved to 20/32 and there was no evidence FA leakage. (**F**) Corresponding OCT B-scan showed a persistently dry macula.](i1552-5783-57-9-OCT14-f02){#i1552-5783-57-9-OCT14-f02} ![Late fluorescein angiographic (FA) images and time-domain optical coherence tomography (OCT) diagonal B-scans from a patient treated with intravitreal ranibizumab in the phase II study, then followed in the extension study with observation and retreatment with ranibizumab.^[@i1552-5783-57-9-OCT14-b17]^ (**A**, **C**, **E**, **G**, **I**, **K**) Late FA images. (**B**, **D**, **F**, **H**, **J**, **L**) Diagonal OCT B-scans. (**A**) Baseline late FA image showed diffuse leakage. Visual acuity (VA) was 20/200. (**B**) Baseline OCT image showed cystic macular edema. (**C**) Two weeks later, VA improved to 20/160, and the FA image showed significant reduction in leakage with focal areas of blocked fluorescence at the margin of the lesion consistent with hemorrhages. (**D**) Corresponding OCT B-scan with significant resolution of macular edema 2 weeks after the first injection. (**E**) After 140 days and 10 injections from baseline, VA improved to 20/100, and late FA image showed no leakage with near-complete resolution of the hemorrhages. (**F**) Corresponding OCT B-scan showed resolution of the macular edema with evidence of macular atrophy as seen by the hypertransmission of light into the choroid (*arrow*). No treatment was given. (**G**) After 10 months without treatment, VA was stable at 20/100, but the late FA image showed minimal leakage. (**H**) Corresponding OCT B-scan showed the appearance of small intraretinal cysts indicative of recurrent macular fluid. No treatment was given. (**I**) After 4 months without treatment, VA remained unchanged at 20/100 but the patient was complaining of increasing distortion. The FA image showed increased leakage. (**J**) Corresponding OCT B-scan showed increased macular edema, and a ranibizumab injection was performed. (**K**) Five months after retreatment, VA improved to 20/63 and the FA image showed no leakage. (**L**) Corresponding OCT B-scan showed minimal macular thickening nasal to the fovea. Overall, OCT-guided retreatment improved VA, decreased FA leakage, and decreased the macular fluid.](i1552-5783-57-9-OCT14-f03){#i1552-5783-57-9-OCT14-f03} Not surprisingly, OCT imaging also revealed that macular fluid accumulated in the space between the RPE and the neurosensory retina. Serous exudate appeared on OCT imaging as a homogeneous hyporeflective space, while proteinaceous exudation appeared less homogeneous and was shown to contain reflective fibrinous membranes ([Fig. 5](#i1552-5783-57-9-OCT14-f05){ref-type="fig"}). This proteinaceous material has been referred to as subretinal hyperreflective material (SHRM) or subretinal hyperreflective material exudate (SHRE) and is different from the hyperreflective fibrosis that can form in more advanced neovascular lesions.^[@i1552-5783-57-9-OCT14-b10][@i1552-5783-57-9-OCT14-b11]--[@i1552-5783-57-9-OCT14-b12]^ The SHRM/SHRE of particular concern is the material that accumulates between the retina and RPE as seen on OCT imaging ([Fig. 5](#i1552-5783-57-9-OCT14-f05){ref-type="fig"}). Neovascular lesions with SHRM/SHRE have been associated with worse visual acuity (VA) outcomes due to the formation of fibrotic scars and photoreceptor atrophy. Usually, type 2 MNV, which grows in the subretinal space, has been shown to be associated with a larger volume of subretinal fluid compared with sub-RPE lesions or type 1 MNV. While these type 1 neovascular lesions have been associated with less subretinal fluid, they are associated with detachments of the RPE known as retinal pigment epithelial detachments (PEDs) ([Figs. 4](#i1552-5783-57-9-OCT14-f04){ref-type="fig"}, [5](#i1552-5783-57-9-OCT14-f05){ref-type="fig"}), and early on, OCT imaging was shown to be a useful technique for identifying and characterizing PEDs.^[@i1552-5783-57-9-OCT14-b02],[@i1552-5783-57-9-OCT14-b13]^ While the presence of a PED and blood was shown to serve as a poor prognostic sign in neovascular AMD, an even worse development, often associated with hemorrhage, was the formation of a RPE tear. Optical coherence tomography imaging of RPE tears was shown to reveal an area of RPE discontinuity with the free edge of the RPE often curled under the PED.^[@i1552-5783-57-9-OCT14-b14]^ Adjacent to the tear, OCT imaging has revealed areas with increased choroidal reflectivity with improved visualization of choroidal vessels due to the absence of the RPE. Typically, the overlying retina was found to be intact, but could be separated from the area of atrophy by subretinal fluid. ![Late fluorescein angiographic (FA) images and time-domain optical coherence tomography (OCT) diagonal B-scans from a patient treated with intravitreal ranibizumab in the phase II study, followed in the extension study by observation and then retreatment with ranibizumab.^[@i1552-5783-57-9-OCT14-b17]^ (**A**, **C**, **E**, **G**, **I**, **K**) Late fluorescein angiographic images. (**B**, **D**, **F**, **H**, **J**, **L**) Diagonal OCT B-scans. (**A**) Baseline late FA image showed significant leakage. Visual acuity (VA) was 20/80. (**B**) Corresponding baseline OCT image showed subretinal fluid (*arrow*) and a retinal pigment epithelial detachment (PED, *arrowhead*). (**C**) After 140 days and six ranibizumab injections, VA improved to 20/50, and the FA image showed a marked reduction in leakage. (**D**) Corresponding OCT B-scan showed resolution of the subretinal fluid with a residual low-lying PED (*arrowhead*). (**E**) After 3 months without an injection, VA decreased to 20/100, and the FA image showed a subtle increase in leakage. (**F**) Corresponding OCT B-scan showed recurrent cystic macular fluid. Ranibizumab was injected at this visit. (**G**) Three months after the ranibizumab injection, VA improved to 20/80, and the FA image showed a marked decrease in leakage. (**H**) Corresponding OCT B-scan showed only small intraretinal cysts, and no injection was given. (**I**) After 6 months without treatment, VA slowly decreased to 20/100 and the patient then complained of distortion. FA image showed increased leakage. (**J**) Corresponding B-scan showed increased macular fluid. (**K**) Over the next 12 months, the patient received two more ranibizumab injections, and the VA improved to 20/50. The FA image showed decreased leakage. (**L**) Corresponding OCT B-scan showed marked improvement in the macular fluid.](i1552-5783-57-9-OCT14-f04){#i1552-5783-57-9-OCT14-f04} {#i1552-5783-57-9-OCT14-f05} In all these examples of macular fluid, whether the fluid was in the retina, under the retina, or under the RPE, OCT imaging was the ideal technique for detecting the excess fluid and determining whether it had changed. Since most of the macular fluid occurred in response to excess VEGF, the stage was now set for the convergence of an ideal imaging technique with an ideal treatment that would inhibit exudation and reduce the accumulation of fluid. OCT Imaging, the Rise of Ranibizumab, and Personalized Treatment Regimens for nvAMD {#s3} =================================================================================== At the same time that the Stratus OCT was introduced into the clinic, the first VEGF inhibitor, previously known as rhuFabV2 and now known as ranibizumab (Lucentis; Genentech/Roche, South San Francisco, CA, USA), entered into a phase I clinical trial.^[@i1552-5783-57-9-OCT14-b15]^ In this phase I single-injection, dose-escalation, safety trial, OCT imaging was not available and FA imaging was used to characterize the neovascular lesions only at baseline. From this phase I trial, a maximum tolerated dose of 0.5 mg was identified based on dose-limiting inflammation that occurred at the 1.0-mg dose. While safety and VA outcomes were encouraging, it wasn\'t until the phase II study that OCT imaging was first used at the Bascom Palmer Eye Institute. There were two phase II trials.^[@i1552-5783-57-9-OCT14-b16],[@i1552-5783-57-9-OCT14-b17]^ One phase II trial studied the safety and tolerability of monthly repeated dosing of 0.3- or 0.5-mg ranibizumab through 6 months, and the second phase II trial explored a dose-escalation regimen in which the ranibizumab dose was sequentially increased from 0.3 to 2 mg and injected as frequently as every 2 weeks in eyes with nvAMD. While both phase II trials demonstrated VA improvement associated with ranibizumab therapy, OCT imaging unambiguously demonstrated that the improvement in VA was correlated with the resolution of macular fluid (Rosenfeld PJ, unpublished data, 2003). Moreover, in an extension program at the end of the phase II studies, retreatment with ranibizumab was permitted at the investigators\' discretion once the fixed dosing was stopped. Investigators could base their retreatment on changes in VA, FA imaging, or OCT imaging. As these patients were being followed off treatment, OCT imaging identified the earliest recurrence of macular fluid in some patients even before FA imaging detected leakage and before VA was affected ([Figs. 2](#i1552-5783-57-9-OCT14-f02){ref-type="fig"}--[4](#i1552-5783-57-9-OCT14-f04){ref-type="fig"}). Some patients needed no additional treatment for up to 2 years while some patients frequently developed recurrent macular fluid. This recurrent fluid first appeared as intraretinal cystic spaces or as subretinal fluid. The fluid that was initially detected by OCT imaging, if untreated, would then increase over time, resulting in subsequent vision loss. Thus, a small amount of fluid led to even more fluid and served as a harbinger of disease progression. As the macular fluid gradually reaccumulated in these subjects in the extension study, retreatment with ranibizumab once again resulted in resolution of the macular fluid; and in those patients with decreased VA associated with the recurrent fluid, VA improved once the fluid resolved. These were the first observations of how OCT imaging of macular fluid could serve as a surrogate marker for the presence of excess VEGF, how OCT imaging unambiguously demonstrated the profound effect of anti-VEGF therapy on macular fluid, and how OCT monitoring of macular fluid might serve as a strategy to determine the appropriate dosing interval used for the clinical management of nvAMD. These data were presented at the Retina Subspecialty Day before the American Academy of Ophthalmology meeting in 2004 and at other scientific meetings in an attempt to convince skeptical colleagues that OCT imaging rather than FA imaging could be used in the management of nvAMD patients. To convincingly demonstrate the benefits of OCT imaging for managing nvAMD patients, Genentech supported an investigator-sponsored trial known as the Prospective OCT Imaging of Patients with Neovascular AMD Treated with Intra-Ocular Ranibizumab (PrONTO) study.^[@i1552-5783-57-9-OCT14-b18],[@i1552-5783-57-9-OCT14-b19]^ In the 2-year PrONTO study, nvAMD patients received three monthly injections of 0.5-mg ranibizumab and continued to receive monthly injections as long as there was persistent macular fluid based on OCT imaging. However, once OCT imaging revealed a fluid-free macula, patients were followed monthly without injection until at least one of several retreatment criteria were met. These criteria included a VA loss of at least five letters with OCT evidence of fluid in the macula, an increase in OCT central retinal thickness of at least 100 μm, new macular hemorrhage, or a new area of classic MNV identified by FA imaging. The purpose of these criteria was to determine if OCT imaging was the most sensitive strategy for detecting the recurrence of macular fluid. After the first year of the study,^[@i1552-5783-57-9-OCT14-b18]^ it was obvious that OCT imaging was able to detect the need for retreatment long before leakage was detected on FA imaging, before vision loss occurred, and before hemorrhage appeared. At the time, retina specialists questioned why the OCT threshold for retreatment was set at an increased central retinal thickness measurement of 100 μm, and the reason for this threshold was to prove a point. We wanted to show our skeptical colleagues that when a small amount of fluid appeared and it was left untreated, even more fluid and vision loss would result. Once we proved our point, the threshold for retreatment was lowered in the second year of the study. A zero tolerance strategy was developed, and the protocol was amended to allow for retreatment once any fluid was detected after a dry macula was achieved.^[@i1552-5783-57-9-OCT14-b19]^ After 2 years, the VA outcomes would turn out to be very similar to the outcomes observed following monthly dosing with ranibizumab in the phase II and phase III trials, but unlike the situation with a fixed monthly dosing regimen, these patients in the PrONTO study required approximately half the number of injections. The PrONTO study served as a model for subsequent studies investigating OCT-guided retreatment using a pro re nata (PRN) regimen for nvAMD.^[@i1552-5783-57-9-OCT14-b20],[@i1552-5783-57-9-OCT14-b21]^ However, several of the smaller studies failed to appreciate two key features of the PrONTO study: the requirement for monthly exams and the amended requirement that retreatment be given if any macular fluid was detected by OCT imaging once the macula was fluid-free. Moreover, retreatment was not decided by changes in VA, but only by changes in OCT imaging. While none of the phase III pivotal trials that led to Food and Drug Administration (FDA) approval for any of the anti-VEGF drugs used a pure OCT-guided PRN dosing regimen, monthly dosing versus OCT-guided PRN dosing has been studied in several large, well-controlled prospective clinical trials.^[@i1552-5783-57-9-OCT14-b22][@i1552-5783-57-9-OCT14-b23][@i1552-5783-57-9-OCT14-b24][@i1552-5783-57-9-OCT14-b25][@i1552-5783-57-9-OCT14-b26]--[@i1552-5783-57-9-OCT14-b27]^ The Comparison of Age-related Macular Degeneration Treatments Trials (CATT) and the Inhibition of VEGF in Age-related choroidal Neovascularization (IVAN) trial showed similar VA outcomes after 2 years when the same anti-VEGF drug was dosed monthly or discontinuously.^[@i1552-5783-57-9-OCT14-b22][@i1552-5783-57-9-OCT14-b23][@i1552-5783-57-9-OCT14-b24]--[@i1552-5783-57-9-OCT14-b25]^ However, the most convincing of these comparative trials is the pHase III, double-masked, multicenter, randomized, Active treatment-controlled study of the efficacy and safety of 0.5 mg and 2.0 mg Ranibizumab administered monthly or on an as-needed Basis (PRN) in patients with subfoveal neOvasculaR age-related macular degeneration (HARBOR) trial. The HARBOR trial used a PRN regimen similar to one employed in the PrONTO study.^[@i1552-5783-57-9-OCT14-b26],[@i1552-5783-57-9-OCT14-b27]^ The two doses of ranibizumab were compared using a monthly versus an OCT-guided PRN dosing regimen in which patients received three monthly injections followed by monthly retreatment as long as there was persistent fluid. Once the macula became fluid-free, patients were followed monthly and retreated if any macular fluid was detected by OCT imaging. This study not only showed that monthly dosing and OCT-guided PRN dosing were equivalent, but did not even rely on a fundus exam when making the retreatment decision. The totality of the data suggest that a rigorously implemented OCT-guided PRN dosing strategy appears to be as beneficial as a fixed monthly dosing regimen, and as a result, this strategy has gained global acceptance. The greatest advantage of an OCT-guided PRN dosing strategy is that it improves and maintains VA while significantly decreasing the number of injections needed. However, monthly visits are still required. Another dosing strategy for anti-VEGF drugs that has gained global acceptance is a regimen known as an OCT-guided treat-and-extend dosing.^[@i1552-5783-57-9-OCT14-b28][@i1552-5783-57-9-OCT14-b29][@i1552-5783-57-9-OCT14-b30][@i1552-5783-57-9-OCT14-b31]--[@i1552-5783-57-9-OCT14-b32]^ This regimen also relies on OCT imaging to determine when the macula has become fluid-free, but then patients continue to be injected at each subsequent visit. However, once the macula is fluid-free, the patient is injected and the visit intervals are slowly extended to determine when and if macular fluid recurs. If OCT imaging detects the recurrence of fluid at a visit, then a treatment is given and the interval is shortened. Usually, visit intervals are lengthened or shortened by 2-week intervals depending on the status of macular fluid. While this OCT-guided strategy mandates that an injection be given even when the macula is dry, the advantage over other strategies is that it minimizes the overall number of visits. To date, only two prospective studies, known as the LUcentis Compared to Avastin Study (LUCAS)^[@i1552-5783-57-9-OCT14-b28],[@i1552-5783-57-9-OCT14-b29]^ and Treat-and-Extend in AMD (TREX-AMD) study,^[@i1552-5783-57-9-OCT14-b30]^ have used this treat-and-extend strategy, and the results appear to be similar to the outcomes from larger monthly dosing trials. Despite the limited prospective clinical trial data, clinicians and patients appear to prefer the treat-and-extend dosing regimen as a compromise between frequent monthly visits and monthly injections. On occasion, OCT monitoring of neovascular lesions has shown the importance of even more frequent injections with anti-VEGF drugs when monthly dosing proved unable to prevent the growth and worsening exudation of the neovascular lesions.^[@i1552-5783-57-9-OCT14-b33]^ Thus, OCT monitoring allows clinicians to follow a "Goldilocks strategy" when managing patients, which is not too many injections, not too few injections, but just the right number of injections for that patient. OCT Imaging, the Demise of Pegaptanib, and the Rise of Bevacizumab {#s4} ================================================================== During the early to mid-2000s when OCT was being accepted into practices as an essential tool for monitoring patients with nvAMD, clinicians became empowered by their ability to monitor the real-time effectiveness of different therapeutic agents, whether it was verteporfin photodynamic therapy (PDT), intravitreal steroids, or anti-VEGF therapy. During this time, two anti-VEGF drugs for intraocular delivery were in clinical trials: ranibizumab (Lucentis, Genentech/Roche), an antibody fragment that binds all isoforms of VEGF-A^[@i1552-5783-57-9-OCT14-b34]^; and pegaptanib sodium (Macugen; OSI/Eyetech, New York, NY, USA), an aptamer that binds VEGF~165~ and larger isoforms.^[@i1552-5783-57-9-OCT14-b35]^ As mentioned previously, OCT imaging was useful in demonstrating the anatomic improvements that followed ranibizumab therapy, and selected sites performed OCT imaging during the phase III ranibizumab trials for neovascular AMD known as the Minimally Classic/Occult Trial of the Anti-VEGF Antibody Ranibizumab in the Treatment of Neovascular AMD (MARINA)^[@i1552-5783-57-9-OCT14-b36]^ and the Anti-VEGF Antibody for the Treatment of Predominantly Classic Choroidal Neovascularization in Age-Related Macular Degeneration (ANCHOR).^[@i1552-5783-57-9-OCT14-b37],[@i1552-5783-57-9-OCT14-b38]^ In these trials, ranibizumab was dosed monthly for 2 years, and the OCT images that were obtained on a subset of subjects from these studies did confirm the drying effect of ranibizumab therapy.^[@i1552-5783-57-9-OCT14-b39]^ In contrast, OCT imaging was never used in the pegaptanib trials, and after pegaptanib was approved in late 2004 and became commercially available in early 2005,^[@i1552-5783-57-9-OCT14-b40]^ clinicians used OCT imaging on patients undergoing treatment with pegaptanib. Not surprisingly, the disappointing VA outcomes following pegaptanib therapy were found to correlate with persistent macular fluid following treatment, thus empowering clinicians to make an objective assessment of the drug\'s efficacy.^[@i1552-5783-57-9-OCT14-b41]^ As a result of this discrepancy between the OCT outcomes when using pegaptanib versus other anti-VEGF drugs, OCT imaging confirmed the need to suppress all VEGF isoforms to obtain the maximum drying effect from anti-VEGF therapy. Optical coherence tomography imaging in the MARINA and ANCHOR trials with ranibizumab also revealed that the major cause of severe vision loss was not associated with persistent macular fluid or a scar; rather, severe vision loss was more commonly associated with a fluid-free macula.^[@i1552-5783-57-9-OCT14-b39]^ For the first time, there was an appreciation that macular atrophy was a major cause of vision loss in treated eyes with nvAMD, rather than the more typical cause of severe vision loss prior to anti-VEGF therapy, which was the formation of a disciform scar.^[@i1552-5783-57-9-OCT14-b42]^ The formation of macular atrophy after anti-VEGF therapy was confirmed in the CATT and IVAN trials after 2 years, and in these trials, the monthly dosing groups had a higher incidence of macular atrophy compared with the discontinuous treatment groups.^[@i1552-5783-57-9-OCT14-b25],[@i1552-5783-57-9-OCT14-b43],[@i1552-5783-57-9-OCT14-b44]^ While this observation has raised questions about whether anti-VEGF therapy is responsible for the macular atrophy in some patients or whether the atrophy appears as a feature of normal disease progression and a consequence of suppressing neovascularization,^[@i1552-5783-57-9-OCT14-b45][@i1552-5783-57-9-OCT14-b46]--[@i1552-5783-57-9-OCT14-b47]^ it has become abundantly clear that OCT imaging is an invaluable tool for not only monitoring the presence of macular fluid, but also for monitoring the anatomic integrity of the outer retina and RPE as macular atrophy evolves ([Fig. 6](#i1552-5783-57-9-OCT14-f06){ref-type="fig"}).^[@i1552-5783-57-9-OCT14-b48][@i1552-5783-57-9-OCT14-b49][@i1552-5783-57-9-OCT14-b50][@i1552-5783-57-9-OCT14-b51]--[@i1552-5783-57-9-OCT14-b52]^ {#i1552-5783-57-9-OCT14-f06} During the clinical development of ranibizumab and pegaptanib, there were not only questions about the appropriate dose, dosing interval, and long-term effectiveness of these drugs; there was also uncertainty about the safety of repeated intravitreal injections and whether patients would tolerate intravitreal injections as a long-term drug delivery strategy. Due to these concerns about the safety and tolerability of repeated intravitreal injections, we investigated a third VEGF inhibitor known as bevacizumab (Avastin; Genentech/Roche), a full-length humanized antibody against VEGF-A sharing a common clonal ancestry with an anti-VEGF murine monoclonal antibody that gave rise to ranibizumab; and like ranibizumab, bevacizumab was known to bind all the isoforms of VEGF-A.^[@i1552-5783-57-9-OCT14-b34],[@i1552-5783-57-9-OCT14-b53]^ Since intravenous bevacizumab was being developed as a therapy for colon cancer, we decided to initiate a self-funded study using the same intravenous strategy for the treatment of nvAMD, and once again, OCT imaging proved pivotal in these bevacizumab studies.^[@i1552-5783-57-9-OCT14-b54],[@i1552-5783-57-9-OCT14-b55]^ The Systemic Avastin in Neovascular AMD (SANA) study was initiated once bevacizumab was approved for colon cancer in February 2004. Twenty-four hours after the first systemic dose of bevacizumab in the first patient, OCT imaging unambiguously showed a profound benefit from intravenous bevacizumab. By May of 2005, OCT imaging had confirmed that systemic bevacizumab had a significant and sustained drying effect on the macula, and this drying effect was associated with VA improvement. However, during the course of the SANA study, the FDA added a black box warning describing the small but significant risk of thromboembolic events from the use of high-dose intravenous bevacizumab. While none of the patients in the SANA study experienced a thromboembolic event, they did experience mild hypertension that was easily controlled with oral medication. Due to these systemic safety concerns and the appreciation from the ranibizumab and pegaptanib studies that patients were able to tolerate repeated intravitreal injections, we decided to offer patients an off-label intravitreal dose of bevacizumab as salvage therapy. This dose was 500- to 1000-fold lower than the dose given intravenously, thus minimizing the systemic risk. Moreover, this intravitreal dose was capable of binding the same amount of intraocular VEGF as the intravitreal dose of ranibizumab. Once again, OCT imaging played a pivotal role in demonstrating the benefits of bevacizumab.^[@i1552-5783-57-9-OCT14-b56],[@i1552-5783-57-9-OCT14-b57]^ When first used as salvage therapy in a nvAMD patient with a lesion that was enlarging despite treatments with PDT and pegaptanib, OCT imaging showed a remarkable improvement in the amount of macular fluid after intravitreal bevacizumab, and this improvement was associated with VA improvement. Since ranibizumab had not yet been approved, and would not be commercially available for another year in July 2006, the off-label use of intravitreal bevacizumab spread globally and became the treatment of choice in 2005 due to its perceived clinical efficacy, its widespread availability, its superiority over pegaptanib, and its low cost. In July 2006 when ranibizumab was commercially available for the treatment of nvAMD, clinicians had a choice between two drugs manufactured by Genentech that had a similar genetic lineage but differed with respect to size, affinity for VEGF, and cost. With the widespread use of OCT imaging, clinicians were now empowered to see for themselves that ranibizumab and bevacizumab were superior to pegaptanib and that ranibizumab appeared similar to bevacizumab with respect to its drying effect on the macula and improvement in VA following treatment. Without OCT imaging, it is unlikely that bevacizumab and OCT-guided retreatment would have gained worldwide popularity. OCT Imaging and the Rise of Aflibercept for nvAMD {#s5} ================================================= In 2011, a fourth drug, known as aflibercept (Eylea; Regeneron Pharmaceutical, Tarrytown, NY, USA), was FDA approved for the treatment of nvAMD. Aflibercept, a fusion protein combining the Fc portion of an antibody with VEGF-binding domains from VEGF receptors 1 and 2, has an even higher affinity for VEGF compared with ranibizumab and a molecular size in between those of bevacizumab and ranibizumab.^[@i1552-5783-57-9-OCT14-b58]^ Moreover, it has the theoretical advantage of neutralizing placental growth factor, as well as all the isoforms of VEGF-A. In the phase III aflibercept trials known as the VEGF Trap-Eye: Investigation of Efficacy and Safety in Wet AMD (VIEW1 and VIEW2) trials, intravitreal injections of aflibercept every 2 months during the first year were shown to be equivalent to monthly injections of ranibizumab.^[@i1552-5783-57-9-OCT14-b59]^ In the second year of the trials, subjects received a mandatory injection every 3 months, with OCT imaging performed during the intervening monthly visits.^[@i1552-5783-57-9-OCT14-b60]^ If there was evidence of recurrent or persistent macular fluid using OCT imaging, then injections could be administered during these intervening visits. This modified-PRN regimen resulted in similar VA and OCT outcomes when aflibercept and ranibizumab were compared, but subjects receiving aflibercept required slightly fewer injections using OCT-guided therapy. Once again, OCT imaging proved valuable for detecting macular fluid and determining the need for retreatment in this prospective clinical trial setting. Since its approval, aflibercept has gained popularity as an anti-VEGF drug with the ability to increase the retreatment interval while maintaining improved VA and a fluid-free macula in nvAMD patients. The perception that aflibercept provides increased durability arises not only from the dosing interval of every 2 months used in the phase III studies, but also from numerous research papers showing that when patients are switched from ranibizumab or bevacizumab to aflibercept, there is a tendency for investigators to report improved macular anatomy and extended fluid-free intervals based on OCT imaging.^[@i1552-5783-57-9-OCT14-b61][@i1552-5783-57-9-OCT14-b62][@i1552-5783-57-9-OCT14-b63]--[@i1552-5783-57-9-OCT14-b64]^ A fifth drug, known as conbercept (Lumatin; Chengdu Kanghong Pharmaceutical Group, Chengdu, China), gained China FDA approval in 2013. Conbercept, which is a fusion protein similar to aflibercept and combines the binding domains of VEGF receptors 1 and 2 with an Fc portion of an antibody, has a molecular weight comparable to that of bevacizumab, but it is currently used only in China.^[@i1552-5783-57-9-OCT14-b65][@i1552-5783-57-9-OCT14-b66]--[@i1552-5783-57-9-OCT14-b67]^ Once again, OCT imaging correlated with the VA benefits observed during the conbercept clinical trials. The Evolution of OCT Imaging for the Management of nvAMD {#s6} ======================================================== Now that TD-OCT has been surpassed by SD-OCT in the clinics, the next step in the commercial evolution of OCT imaging for the management of nvAMD is the development of OCT angiography (OCTA).^[@i1552-5783-57-9-OCT14-b68][@i1552-5783-57-9-OCT14-b69][@i1552-5783-57-9-OCT14-b70][@i1552-5783-57-9-OCT14-b71]--[@i1552-5783-57-9-OCT14-b72]^ Optical coherence tomography angiography utilizes novel scanning patterns and algorithms to image blood flow in the retina and choroid without the use of exogenous dyes. Not only can blood flow be imaged, but also the scans provide detailed structural information comparable to or better than the images currently available from commercial SD-OCT instruments. Different OCTA platforms are being developed that use different scanning patterns and algorithms, but they can be broadly divided into SD-OCT and swept-source OCT (SS-OCT) instruments. Each platform will have certain advantages and disadvantages, but whether OCTA has advantages over conventional OCT imaging in the management of nvAMD patients remains to be determined. The ability to noninvasively visualize the appearance and growth of MNV should augment the routine OCT strategy of monitoring for the presence or absence of macular fluid when deciding to initiate anti-VEGF therapy and when to re-treat. Since exudation is the metric by which we currently measure VEGF activity and exudation can arise from both the neovascularization and the normal vasculature, there should be certain advantages of using OCTA to directly monitor changes in the neovascular lesion rather than just monitoring the level of exudation. Even if OCTA imaging does not influence our decision on when to treat, it will surely add to our understanding of disease onset, progression, and response to treatment. For example, OCTA can visualize the presence of type 1 MNV in eyes that were diagnosed with dry, intermediate AMD.^[@i1552-5783-57-9-OCT14-b70]^ The presence of nonexudative neovascular lesions raises new questions about whether we can predict when exudation will arise and if there is a role for anti-VEGF therapy even before exudation occurs. If OCTA can now replace FA and ICGA imaging for most MNV, then OCTA imaging will provide all the structural and flow information needed for the diagnosis and management of MNV. Summary of OCT Imaging and Antiangiogenic Therapies {#s7} =================================================== With respect to their development, OCT imaging and antiangiogenic therapies have been inextricably linked. If OCT imaging had not been available when anti-VEGF therapies were developed, then clinicians and patients would have been chained to fixed-dosing regimens without the ability to easily and noninvasively assess whether the exudation from the neovascularization had been adequately controlled. In the era of PDT before anti-VEGF therapy, patients underwent invasive, time-consuming FA imaging every 3 months to determine whether leakage was present and treatment was needed. When intravitreal pegaptanib was approved, injections were given per protocol every 6 weeks, but when patients lost vision and additional OCT imaging was performed, the presence of persistent or worsening macular fluid was confirmed. In the era of OCT imaging and pan-VEGF blockade, OCT imaging unambiguously confirmed the effectiveness of anti-VEGF therapy in resolving macular fluid. Optical coherence tomography imaging empowered clinicians to personalize therapy by evaluating different drugs and dosing regimens. Clinicians were able to customize therapy to either minimize the number of injections or minimize the number of visits. Without the confidence that came from OCT imaging and the ability to demonstrate the beneficial effect of anti-VEGF therapy within 24 hours of treatment,^[@i1552-5783-57-9-OCT14-b18]^ it is unlikely that intravenous and then intravitreal bevacizumab therapy for the treatment of exudative eye diseases would have been developed in 2005, and it is unlikely that the global ophthalmologic community would have been convinced to rapidly adopt intravitreal bevacizumab as a viable treatment strategy. While significant cost savings have resulted from the use of OCT compared with FA imaging,^[@i1552-5783-57-9-OCT14-b73]^ greater cost savings have resulted from the use of OCT-guided retreatment to avoid monthly dosing, and even greater cost savings have resulted from the use of off-label bevacizumab in lieu of the more expensive drugs. In the United States alone, over the past 10 years, the combined Medicare cost savings from the use of OCT^[@i1552-5783-57-9-OCT14-b73]^ and bevacizumab^[@i1552-5783-57-9-OCT14-b74]^ have exceeded \$20 billion when conservatively extrapolated from published utilization data for all forms of exudative eye diseases. Globally, four intravitreal drugs dominate the market for anti-VEGF therapy in neovascular AMD, diabetic macular edema, and retinal vein occlusions: bevacizumab, ranibizumab, aflibercept, and conbercept. Despite differences in molecular size, affinity for VEGF, and intravitreal half-life, all four drugs have shown efficacy with improved VA and resolution of macular fluid. The most likely explanation for the effectiveness of these four drugs is that they bind all the biological isoforms of VEGF-A whereas pegaptanib inhibited only VEGF~165~ and larger isoforms. Thus, this was the first example of how OCT imaging was able to distinguish between different therapeutic targets. As new therapies and drug delivery strategies are developed to treat exudative eye diseases, OCT will once again play an important role. For example, anti-VEGF drugs are being delivered through implantable devices, as topical therapy, and as oral medication. Treatments that combine intravitreal anti-VEGF therapy along with inhibitors of platelet-derived growth factor (PDGF), angiopoietin 2, and tissue factor are being investigated, and topical squalamine is being combined with anti-VEGF therapy in clinical trials as well. Optical coherence tomography continues to be an integral part of all these trials, but the phase III trials for the PDGF inhibitor known as E10030 (Fovista; Ophthotech Corp., Princeton, NJ, USA) have incorporated a novel inclusion requirement for their treatment-naïve subjects with nvAMD; that requirement is the presence of SHRM (ClinicalTrials.gov Identifier: NCT01944839, NCT01940900, NCT01940887). While improvement of VA remains the ultimate outcome that will determine the success or failure of any new treatment, there\'s so much more to the macula than VA, which measures only central foveal function. Unlike dye-based angiography, OCT imaging can identify macular abnormalities involving the photoreceptors, the RPE, and the choroid that can help explain visual function changes such as reduced reading speed, reduced microperimetric thresholds, and delayed dark adaptation.^[@i1552-5783-57-9-OCT14-b75]^ While OCT is clearly the gold standard imaging strategy for predicting the success of any antiangiogenic treatment, it is also a valuable strategy for studying a wide range of anatomic changes that correlate with reduced visual function and disease progression. In summary, OCT is held in such high regard that clinicians will never be convinced of a treatment\'s efficacy until OCT imaging demonstrates an unambiguous anatomic benefit. I thank Carmen Puliafito, MD, MBA, for his support and encouragement during the early to mid-2000s when both clinical OCT imaging and anti-VEGF therapies were being developed for clinical use. Without early access to OCT technology and his permission to pursue research into bevacizumab therapy for exudative eye diseases, the current treatment landscape might have evolved very differently. I also thank my clinical research fellows (Stephan Michels, MD, MBA, Anne Fung, MD, Andrew Moshfeghi, MD, MBA, and Geeta Lalwani, MD) and my clinical coordinators (Maria Esquiabro, Belen Rodriguez, and Cristina Lage-Rodriguez) for their exceptional talents and assistance in managing patients and conducting these successful anti-VEGF clinical trials. Supported by a grant from Carl Zeiss Meditec, Inc. (Dublin, CA, USA), the Macula Vision Research Foundation, a National Eye Institute Center Core Grant (P30EY014801), an unrestricted grant from Research to Prevent Blindness, and the Emma Clyde Hodge Memorial Foundation. Disclosure: **P.J. Rosenfeld**, Achillion Pharmaceuticals (C), Acucela (C, F), Apellis (F, I), Boehringer-Ingelheim (C), Carl Zeiss Meditec, Inc. (F, R), Cell Cure Neurosciences (C), Chengdu Kanghong Biotech (C), CoDa Therapeutics (C), Digisight (I), Genentech (C, F), GlaxoSmithKline (F), Healios K.K. (C), F. Hoffmann-La Roche Ltd. (C), MacRegen, Inc. (C), NGM Biopharmaceuticals (C), Neurotech (F), Ocata Therapeutics (C, F), Ocudyne (C), Regeneron (C), Stealth BioTherapeutics (C), Tyrogenex (C, F), Vision Medicines (C)

Background ========== The *MLL*gene located on chromosome 11 band q23 normally functions as a transcription regulator of the HOX genes \[[@B1]\] and is essential for normal mammalian development and hematopoiesis \[[@B2]\]. Chromosomal translocations involving *MLL*gene represent frequent cytogenetic abnormalities found in hematologic malignancies, occurring in 5--6% of patients with acute myeloid leukemia (AML), 7--10% of acute lymphoblastic leukemia (ALL), 60--70% of all acute leukemias in infants, and in most patients with t-AML/t-ALL secondary to therapy that is targeting topoisomerase II \[[@B3]\]. The function of the various *MLL*fusion genes \[[@B4],[@B5]\] and proteins is poorly understood but it appears that the fusion proteins disrupt the ability of wild-type *MLL*to regulate HOX gene expression, leading to leukemogenesis \[[@B6]\]. Recent studies demonstrated that the presence of *MLL*rearrangements can be associated to specific antigen \[[@B7],[@B8]\] and gene expression patterns. Pediatric patients with ALL carrying *MLL*rearrangement have been successfully distinguished from ALL and AML patients without *MLL*translocation as a distinct subgroup with a specific gene expression profile \[[@B9]\]. Different gene expression signatures for ALL and AML samples, with and without *MLL*translocation were also identified in adult patients indicating a common method to comprehensively characterize the *MLL*mutation \[[@B10]\]. However, the involvement of *MLL*gene in the onset and progression of leukemia event still remains unclear. The present study encompasses the efforts to clarify the relations between *MLL*translocation and acute leukemias in pediatric patients: we identified common MLL-related markers that are shared between leukemias with different phenotypes (translocation-related signature) and investigated the role of *MLL*aberration in acute leukemias with different lineage origin (phenotype-related signature). To this end, two independent cohorts of 140 (training set) and 78 (test set) pediatric patients with ALL/MLL+, AML/MLL+, ALL/MLL- and AML/MLL- have been inspected using gene expression profiling. Separated comparisons based on phenotype and translocation information have been applied to find differentially expressed genes using both comparison (SAM) and prediction (PAM) analyses. Each subgroup has been clearly distinguished using a final subset of 15 probes which separated the training cohort samples into phenotype-related and translocation-related signatures. The strength of our predictor was successfully validated on an independent test cohort. The identified markers have been further examined to explain their biological correlations using gene ontology inquiries. We assessed the key role of previously unexplored genes to specifically characterize *MLL*translocation as well as the impact of well-known genes in separating acute leukemia samples according to phenotype origin. Methods ======= Patients and Samples -------------------- A total cohort of 140 pediatric patients was enrolled in the training set. Bone marrow samples were collected at diagnosis from 106 and 34 patients with acute lymphoblastic and myeloblastic leukemia, respectively. The presence of *MLL*rearrangement was detected in 16/116 (13.8%) ALL patients and in 8/34 (23.5%) AML cases (Table [1](#T1){ref-type="table"}). Detailed sample information are provided in Additional File [1](#S1){ref-type="supplementary-material"}. A second cohort of 78 patients has been used as test set to validate the results from supervised analyses performed on the training set. ###### Patient cohort **N° patients** **Gender** **Median Age** --------------------- ----------------- ------------ ----------------------- **Training cohort** 140 70 M/70 F 6.2 y (0.1 -- 17) **ALL/MLL-** 90 48 M/42 F 6.51 y (0.88 -- 16.5) **AML/MLL-** 26 12 M/14 F 8.42 y (0.45 -- 17) **ALL/MLL+** 16 8 M/8 F 2.14 y (0.1 -- 14.5) **AML/MLL+** 8 2 M/6 F 3.59 y (0.16 -- 13.6) Patient distribution of the training cohort according to phenotype and *MLL*translocation. For all patients, ALL and AML diagnoses were performed by morphology, cytochemistry, cytogenetics, immunophenotype and molecular genetics. Based on the laboratory diagnosis, patients were risk stratified and enrolled in the following AIEOP protocols: LAL2000 (ALL after 2000 year), LAM2002 (AML after 2002 year), LAL95 (ALL before 2000 year), LAM92 (AML before 2002 year) and Interfant99 (ALL and AML less than 1 year old) \[[@B11]-[@B13]\]. This study was conducted after obtaining the informed consent from all patients and following the tenets of the Declaration of Helsinki and was approved by the ethics committees of the participating institutions before the initiation. Morphological classification ---------------------------- The morphological classification was performed by three independent investigators and the conclusive diagnosis for every case was reported according to the FAB criteria. In cases where consensus was not obtained, the three investigators re-analyzed the slides together in order to obtain consensus as to final diagnosis. Cytogenetic analyses -------------------- Samples were processed and cytogenetic studies were performed using the Q-banding technique. About 15--20 metaphases for each sample were acquired (with CASTI System) and analyzed in order to avoid clone loss. Chromosomes were identified and assigned according to the International System for Human Cytogenetic Nomenclature. FISH analysis was performed on interphase nuclei and where possible on metaphases, using an *MLL*probe (Vysis). Biomolecular analyses --------------------- The AIEOP MLL protocol, which screens for the fusion gene transcripts *MLL-AF4*for t(4;11) translocation, *MLL-AF10*for t(10;11), *MLL-AF9*for t(9;11) and *MLL-ENL*for t(11;19), was performed as stated in the previously reported method; briefly, total RNA was isolated using the RNAzol-B reagent (Duotech srl Milan, Italy) following manufacturer\'s instructions. One microgram of total RNA from each specimen was reverse transcribed using the Superscript reverse transcriptase (Life Technologies Milan, Italy) and random hexamers. PCR amplification was performed using AmpliTaq polymerase (Applied Biosystems) according to the BIOMED-1 protocols. An independent PCR reaction was performed with shift primers for confirmation of each positive result. The ABL housekeeping gene expression was assessed to determine the presence of amplifiable RNA and the efficacy of reverse transcriptions. After electrophoresis, the PCR reaction products were stained with ethidium bromide. RNA isolation for microarray analysis ------------------------------------- Microarray analysis was performed on each sample using Affymetrix Human Genome U133 Plus 2.0 GeneChip. Total RNA was extracted from stored frozen cells of leukaemia specimens using TRIzol (Invitrogen) followed by a purification step (RNeasy Mini Kit, Qiagen). RNA quality was assessed on the Agilent Bioanalyzer 2100 using the Agilent RNA 6000 Nano Assay kit (Agilent Technologies, Waldbronn, Germany). RNA concentration was determined using the NanoDrop ND-1000 spectrophotometer (NanoDrop Technologies, Inc., Wilmington, DE USA). The overall total RNA quality was assessed by A260/A280 ratio. Study aim and design -------------------- We performed comparisons between subgroups carrying the same lineage (Figure [1](#F1){ref-type="fig"}, ALL/MLL- vs ALL/MLL+ named \"L1\"; AML/MLL- vs AML/MLL+ named \"L3\") or the same translocation information (ALL/MLL- vs AML/MLL- named \"L2\"; ALL/MLL+ vs AML/MLL+ named \"L4\") using SAM and PAM algorithms. Combining the results from separated lineage-related comparisons (L1 and L3), we identified a specific MLL signature that is shared between ALL and AML subtypes. Matching the gene lists from translocation-specific comparisons (L2 and L4), we described a common lineage signature for samples with and without *MLL*translocation. {#F1} The results from SAM and PAM analyses have been tested on an independent cohort to separate samples according to their gene expression profiles. Gene Expression Profiling ------------------------- Microarrays analyses were performed using Human Genome U133 Plus 2.0 GeneChip Array (Affymetrix, Santa Clara, CA, USA) and specific RNA isolation method as previously described. Microarray data (CEL files) can be found online at GEO repository (Accession Number: GSE14062). Statistical analyses were performed using open-source BioConductor package \[<http://www.bioconductor.org>, ver.2.1\] for R System software \[The R Project for Statistical Computing, ver.2.7.0, <http://www.r-project.org>\]. The training and test sets were inspected using the same analysis procedures. Microarray CEL files were analyzed to evaluate potential errors on the arrays using quality control algorithms on PS intensity signals. To allow gene expression comparison, the robust multiple-array average (RMA) normalization has been performed among all arrays. Genes with low expression variation across all samples have been discarded using interquartile range (IQR\>1.15) as filtering criteria; the following analyses have been performed starting from 5130 filtered probes. Significance analysis of microarray (SAM) package has been used in the training set to find differentially expressed genes among ALL/MLL-, AML/MLL-, ALL/MLL+ and AML/MLL+ groups. Less than one false positive-rated gene was found using false discovery rate (FDR) \<1% and q-value = 0 cutoffs. Obviously genes identified by SAM analysis are not necessarily involved in class prediction \[[@B14]\]. Shrunken centroid algorithm was performed on the training set using Tibshirani\'s prediction analysis of microarrays (PAM) package. All probes but one (204069_at, one of the two *MEIS1*probes included in the prediction list) identified by PAM analysis matched the SAM results. The results from supervised analyses were further validated on an independent test set. SAM gene lists were used to perform gene ontology studies using the Database for Annotation, Visualization and Integrated Discovery (DAVID) web-based tool <http://david.abcc.ncifcrf.gov/>. The phenotype-related signature was analyzed through the use of Ingenuity Pathway Analysis (Ingenuity^®^System, <http://www.ingenuity.com>). Results ======= SAM results -- Training Set --------------------------- The training cohort of 140 samples was analyzed using Affymetrix HG-U133 Plus 2.0 gene expression chips. As previously described in the study design, four separated comparisons (labeled L1 to L4) were performed using SAM and PAM methods to identify differentially expressed probes for translocation-specific and lineage-specific groups. Table [2](#T2){ref-type="table"} summarizes the results retrieved by SAM analysis using a false discovery rate (FDR) \< 1% and q-value = 0 as statistical cut-off values. ###### SAM results *Constant part* Phenotype Phenotype Translocation Translocation ---------------------------- ------------------------------- --------------------------- -------------------------- -------------------------- *Variable part* Translocation Translocation Phenotype Phenotype *Comparison ID* L1 L3 L2 L4 *SAM comparisons* ALL/MLL(-) vs ALL/MLL(+) AML/MLL(-) vs AML/MLL(+) ALL/MLL(-) vs AML/MLL(-) ALL/MLL(+) vs AML/MLL(+) ***up***/*down ALL/MLL(-)* **1013**/740 **1378**/754 ***up***/*down AML/MLL(-)* **155**/555 **754**/1378 ***up***/*down ALL/MLL(+)* **740**/1013 **379**/601 ***up***/*down AML/MLL(+)* **555**/155 **601**/379 Total *1753* *710* *2132* *980* *(Common) Signature* **379**Translocation specific **622**Phenotype specific SAM results for comparisons between considered subgroups. Translocation-specific signature was obtained by matching deregulated probe sets from L1 and L3 comparisons, phenotype-specific signature from L2 and L4 comparisons. To identify a *MLL*translocation-specific signature commonly shared by different lineage subtypes, we compared ALL and AML samples separately. For each comparison, upregulated probes in one group are consequently downregulated in the other, and vice versa. The results from ALL/MLL- vs ALL/MLL+ (L1 comparison: 1013 upregulated probes in the former group, 740 upregulated in the latter) and AML/MLL- vs AML/MLL+ (L3 comparison: 155 overexpressed in the first group, 555 overexpressed in the second) were subsequently matched to generate a unique MLL-specific subset. A total of 379 common translocation-specific probe sets were found from L1 and L3 comparison combining 1753 and 710 probes, respectively \[see Additional File [2](#S2){ref-type="supplementary-material"}\]. The phenotype-associated signature was identified by comparing samples with and without *MLL*-mutations separately. We matched the results from ALL/MLL- vs AML/MLL- (L2 comparison: 1378 upregulated probes in the former group, 754 upregulated in the latter) and ALL/MLL+ vs AML/MLL+ (L4 comparison: 379 overexpressed in the first group, 601 overexpressed in the second), obtaining 622 shared phenotype-specific probe sets from 2132 (L2) and 980 (L4) differentially expressed markers \[see Additional File [2](#S2){ref-type="supplementary-material"}\]. SAM results were tested on 140 samples using an unsupervised hierarchical clustering method: patients were clearly distinguished in two groups according to translocation- (379 probes, Figure [2A](#F2){ref-type="fig"}) and phenotype-related (622 probes, Figure [2B](#F2){ref-type="fig"}) signatures. {#F2} PAM results -- Training Set --------------------------- SAM comparisons were reproduced in the training cohort using a prediction algorithm (PAM). We identified a total of 15 probe sets, corresponding to 14 genes, with predictive value in separating ALL and AML with and without *MLL*rearrangement (Table [3](#T3){ref-type="table"}). Fourteen out of 15 probes extracted by PAM matched SAM analyses results \[see Additional File [2](#S2){ref-type="supplementary-material"}\]. ###### PAM results **PAM comparisons** ALL/MLL- vs ALL/MLL+ ALL/MLL- vs AML/MLL- AML/MLL- vs AML/MLL+ ALL/MLL+ vs AML/MLL+ ----------------------- ---------------------- ---------------------- ------------------------------------------------------ ----------------------------------- **Prediction ID** L1 L2 L3 L4 **Class Error Rate** 0% 0% 0% 0% **Probes identified** 2 2 7 4 **Gene symbol** *MEIS1, MEIS1* *TCL1A, EBF1* *ZEB2, SRGAP2P1, AK2, TMEM30A, FAM62B, TMED2, HIPK3* *PAX5, CD72, LOC100130458, CSRP2* Class prediction analyses (PAM) performed on the training cohort of 140 samples using 5130 filtered probes. To test prediction performances, we applied unsupervised hierarchical clustering on 140 samples using PAM results. In Figure [3](#F3){ref-type="fig"}, the upper dendrogram clearly separated our samples into 2 branches according to ALL (106 samples) and AML (34 samples) phenotypes. Each branch further divided into MLL positive and MLL negative subgroups showing distinct gene expression profiles. Additionally, the probe sets retrieved by PAM analysis were associated into 2 major clusters as shown in the dendrogram on the left. A translocation-related signature was characterized by overexpression of *ZEB2*, *SRGAP2P1*, *TMEM30A*, *AK2*, *TMED2*, *HIPK3*, *FAM62B*and *MEIS1*genes in MLL positive samples both ALL and AML, while a common upregulation of *PAX5*, *CD72*, *CSRP2*, *LOC100130458*, *TCL1A*and *EBF1*genes correlated with ALL patients in a phenotype-related signature. {#F3} Validation on Test Set ---------------------- The predictor (15 probe sets) obtained by SAM and PAM results has been used for validation on a separated cohort of 78 samples. Patients were distributed as follows: ALL/MLL- n = 54, AML/MLL- n = 11, ALL/MLL+ n = 8, AML/MLL+ n = 5. The dendrogram in Additional File [3](#S3){ref-type="supplementary-material"} validates the strength of our predictor in discriminating ALL and AML with and without *MLL*rearrangements. Only one sample (PD529) has been misclassified in the independent cohort, probably due to the small sample size of AML/MLL+ subgroup. MLL Translocation-related signature ----------------------------------- ALL patients with and without *MLL*translocation have been compared (L1). Two probes for *MEIS1*gene (204969_at, 15559477_s\_at), which encodes a cofactor for HOX proteins that can accelerate Hoxa9-dependent leukemia, were shown to be strongly expressed in MLL positive samples, as previously reported \[[@B15]\]. *MEIS1*quantitatively regulates the differentiation arrest, cycling activity, in vivo progression, and self-renewal of MLL leukemia cells, thereby functioning as a critical and rate-limiting determinant of leukemia stem cell potential \[[@B16]\]. *ZEB2*(SIP1) gene, coding a zinc finger E-box binding homeobox 2 protein, was already reported as cancer activating factor. In synergy with another transcription factor (Snail), *ZEB2*represses transcription of the E-cad gene by binding E-box on E-cad promoter. Loss of E-cadherin (E-cad) triggers invasion, metastasis, and dedifferentiation in various epithelial carcinomas \[[@B17]\]. Interestingly *AK2*and *HIPK3*genes, here upregulated in *MLL*rearrangements, share the same target in Fas-mediated apoptosis pathway, an adaptor molecule (FADD) that interacts with various cell surface receptors and mediates cell apoptotic signals. Adenylate kinase 2 (*AK2*) regulates mitochondrial apoptosis through the formation of an AK2-FADD-caspase-10 (AFAC10) complex. Acting in concert with FADD and caspase-10, *AK2*mediates a novel intrinsic apoptotic pathway that may be involved in tumorigenesis \[[@B18]\]. Another Fas/FADD-interacting kinase, *HIPK3*(*PKY*), was first identified as a putative multidrug-resistant protein from studies of cancer cells. Common death receptor target of *AK2*and *HIPK3*suggests that a principle role of this kinase family is in regulating various aspects of death receptor signaling \[[@B19],[@B20]\]. The upregulation of *AK2*, *MEIS1*and *TMEM30A*(transmembrane protein 30A) genes in *MLL*-rearranged acute leukemias has also been demonstrated by Faber et al. \[[@B21]\] using gene expression profiling. The genes *SRGAP2P1*(SLIT-ROBO Rho GTPase activating protein 2 pseudogene 1), *TMED2*(transmembrane emp24 domain trafficking protein 2) and *FAM62B*(family with sequence similarity 62 (C2 domain containing) member B) shared similar upregulation in both ALL and AML with *MLL*translocation but their role in leukemogenesis remains to be explored. Expression of HOX and NG2 genes ------------------------------- HOX gene expression was evaluated in our patient cohort: *HOXA10*, *HOXA9*, *HOXA7*, *HOXA5*and *HOXA3*were found to be generally upregulated in samples carrying *MLL*translocation, both in ALL and AML, while *HOXA11*and *HOXA6*showed common overexpression in AML/MLL+ cases \[[@B16]\]. *NG2*gene, an integral membrane chondroitin sulfate proteoglycan expressed by human malignant melanoma \[[@B22]\] and leukemic \[[@B7],[@B23]\] cells, exhibited higher expression in both ALL and AML with *MLL*rearrangement similar to HOX genes (data not shown). Phenotype-related signature --------------------------- A total of 6 probe sets with predictive value were isolated using PAM analysis to separate ALL from AML samples with and without the presence of *MLL*aberration. All probes showed a marked downregulation in AML samples as well as a common involvement in enhancing the B-cell signaling pathway. *PAX5*plays a key role in regulating a number of genes identified by PAM and SAM analyses distinguishing ALL from AML phenotypes. *PAX5*gene codes for BSAP, a transcription factor expressed in the developing central nervous system, testis and cells of B lymphocyte lineage except terminally differentiated plasma cells \[[@B24],[@B25]\]. BSAP-binding sites have been identified in several genes, included in SAM list, encoding *VpreB1*\[[@B26]\], *BLK*\[[@B27],[@B28]\], *CD79a*\[[@B29]\], *CD19*\[[@B30]\], *EBF*\[[@B31]\], *CD72*\[[@B32]\], *BLNK*\[[@B33],[@B34]\] and *LEF1*\[[@B35]\]. CD72 antigen is a transmembrane glycoprotein that plays a fundamental role in B-cell activation and proliferation. It has been recently shown its function also in preventing the differentiation of naïve B-cells into plasma cells \[[@B36]\]. CD19 antigen is routinely monitored in immunophenotyping diagnosis as the main surface marker for B cell identification. The CD19 tyrosine phosphorylation is induced by CD72 ligation: the activation of B lymphocytes via CD72 resulted in recruitment and activation of PI 3-K, which was mediated by CD19. Moreover, CD19 enhances membrane IgM signaling like CD79a even known as Igα \[[@B37]\]. BLNK is a B-cell linker protein which is fundamental in addressing pro-B cell to pre-B cell transition \[[@B38]\]. Kabak et al. \[[@B39]\] demonstrated the direct recruitment of BLNK to Igα (CD79a) in signaling pathways and Imai et al. \[[@B40]\] showed that BLNK expression is a common leukogenic event in childhood B-lineage ALL similarly to the B lineage-specific polymerase encoded by *DNTT*gene. To assess our findings, we compared PAM and SAM results to a minimal self-sustained gene regulatory module for pro-B cell differentiation and proliferation proposed by Medina K. et al. \[[@B41]\]. All module genes (*FLT3*, *PU.1*, *E2A*, *BCL11A*, *IL-7R*, *EBF*, *PAX5*) except *PU.1*were included in SAM lists showing significant over-expression in MLL+/ALL group (Figure [4](#F4){ref-type="fig"}). ![**Genes involved in pro-B cell differentiation and proliferation**. Regulatory role of *PAX5*in B cell differentiation and proliferation. Up-regulated and down-regulated genes in ALL/MLL+ from supervised analyses are depicted red and green, respectively. The minimal self-sustaining gene regulatory module was obtained by Medina K. et al. \[[@B36]\] and expanded according to our results. The graph was generated using Ingenuity Pathways Analysis (IPA) software (Ingenuity^®^System, <http://www.ingenuity.com>).](1755-8794-2-36-4){#F4} *CSRP2*is a member of the CSRP (cysteine and glycine-rich protein) family of genes, encoding a group of LIM domain proteins, which may be involved in regulatory processes important for development and cellular differentiation but also in oncogenesis \[[@B42]\]. During lymphocyte differentiation, the expression of T-cell leukemia/lymphoma 1A gene (*TCL1A*) begins in pre-B cells, is downregulated in germinal centre B cells, and is silenced in memory B and plasma cells \[[@B43]-[@B45]\]. Performing comparative Gene-Set Enrichment Analysis (GSEA), Aggarwal et al. \[[@B46]\] demonstrated that the increased expression of *TCL1A*was significantly associated (P \< 0.05) with some of the most important pathways controlling B-cell lymphoma pathogenesis and heterogeneity, including B-cell receptor pathway. Gene Ontology results --------------------- We further inspected the SAM gene lists using a web-based gene ontology (GO) tool (DAVID) to understand the biological meaning behind large list of genes. Additional File [4](#S4){ref-type="supplementary-material"} summarizes the GO results (Biological Processes) for both SAM signatures ordered by FDR, defined as the median number of false positive genes divided by the number of significant genes. The analysis of lineage-specific signature (distinguishing ALL from AML) clearly shows the involvement of hematopoietic and immune response-related genes in the progress of disease, in particular leukocyte activation and hematopoietic organ development. MLL-related signature, distinguishing samples with and without *MLL*translocation, mainly includes regulatory genes involved in developmental process, biological process, apoptosis and programmed cell death. AML subgroup without MLL translocation and MEIS1 upregulated ------------------------------------------------------------ In the present study we showed *MEIS1*gene to be upregulated both in ALL and AML with *MLL*translocation. As depicted in Figure [3](#F3){ref-type="fig"}, a subgroup of AML without *MLL*mutation displayed clear overexpression for both *MEIS1*probes (as well as *HOXA9*and *HOXA5*genes, see Additional File [5](#S5){ref-type="supplementary-material"}), confirmed by supervised analysis on the 26 patients with AML/MLL-. Patient distribution is described in Table [4](#T4){ref-type="table"}: interestingly, all patients with t(8;21)(q22;q22) and t(15;17)(q22;q21) correlated with down-regulation of *MEIS1*, while 3 out of 4 patients with inv(16)(p13q22) highly expressed *MEIS1*gene. Our results confirm the observations by Grubach et al. \[[@B47]\] who determined RQ-PCR expression levels of a series of PcG genes (including *MEIS1*) and PcG-regulated genes in 126 AML patients and 20 healthy donors. ###### AML without MLL and MEIS1 upregulated **AML without MLL** ***MEIS1*-upregulated group** ***MEIS1*-downregulated group** --------------------- ------------------------------- --------------------------------- *Normal* 8 5 *inv(16)(p13q22)* 3 1 *t(8;21)(q22;q22)* 0 5 *t(15;17)(q22;q21)* 0 4 Total 11 15 Distribution of AML patients without *MLL*rearrangements with MEIS1 deregulated according to karyotype information. Discussion ========== The presence of *MLL*rearrangements in acute leukemias results in a complex number of biological modifications that still remain largely unexplained. Armstrong et al. \[[@B9]\] proposed *MLL*rearrangement positive ALL as a distinct subgroup with a specific gene expression profile. However, this signature was related only to MLL with ALL phenotype. Here we show that MLL, from ALL and AML origin, shares a signature identified by a small set of genes suggesting a common genetic disregulation that could be at the base of the mixed lineage leukemia in both phenotypes. Most *MLL*aberrations have been successfully characterized by cytogenetics, immunophenotyping, molecular biology and, recently, gene expression profiling. In the present study, a total cohort of 218 patients was analyzed by microarray approach to identify common deregulated MLL targets shared by different leukemia phenotypes and to inspect *MLL*involvement in acute leukemias with different lineage origin. We analyzed gene expression data of 140 (training set) and 78 (test set) pediatric samples carrying ALL (N = 106+62) and AML (N = 34+16), with (N = 24+13) and without (N = 116+65) *MLL*translocations. Supervised analyses on the training set identified two specific signatures according to lineage origin and MLL presence; the results have been subsequently validated on the test set. The genes *MEIS1*, *ZEB2*, *SRGAP2P1*, *TMEM30A*, *AK2*, *TMED2*, *HIPK3*and *FAM62B*showed marked up-regulation in patients with *MLL*mutation, both in ALL and AML. GO analysis revealed a general regulatory role of these genes in developmental process, biological process, apoptosis and programmed cell death. Our results confirm the primary function of *MEIS1*gene in regulating fundamental MLL leukemia-related and HOX genes processes. *MEIS1*up-regulation was also observed in a subgroup of AML patients showing inv(16)-like signature in the absence of *MLL*mutations. We further identified a phenotype-related signature that distinguishes lymphoblastic and myeloblastic acute leukemias. A set of 6 probes allowed for separation of ALL and AML with and without *MLL*mutation, including *PAX5*, *CD72*, *CSRP2*, *LOC100130458*, *EBF1*and *TCL1A*genes. We corroborate the main role of *PAX5*gene in orchestrating basic biological processes such as leukocyte and hematopoietic development \[[@B10],[@B48]-[@B50]\]. Mullighan et al. \[[@B51]\] showed that the genes regulating B-cell development and differentiation are mutated in 40% of pediatric ALL and that *PAX5*was the most frequent target of somatic mutation being altered in 31.7% of cases. Similarly, we showed that *PAX5*, *EBF1*, *CD72*and *TCL1A*are tightly correlated in the capacity to distinguish lymphoblastic and myeloblastic characteristics also in the presence of *MLL*mutations illustrating the importance of B-cell receptor signaling pathway in this subset of leukemias. We stressed the role of *PAX5*in repressing *PU.1*(NF-kappaβ) regulated reporter gene as supported by their opposite expression in B lineage cells with *MLL*rearrangements \[[@B52]\]. Deregulation events of *CSRP2*, *TCL1A, CD72*and *EBF*genes in acute pediatric leukaemia have been previously reported using gene expression profiling \[[@B43]\]. The role of *CSRP2*gene in leukemogenesis still has to be investigated. Bach et al. \[[@B53]\] demonstrated that proteins like *CSRP2*with LIM domains play important roles in embryo development and hematopoiesis. Moreover, Bégay-Muller et al. \[[@B54]\] showed that the LIM domain protein Lmo2 binds to AF6, a translocation partner of the *MLL*oncogene. CRP2 protein, encoded by *CSPR2*gene, was shown to transactivate the proximal promoter region of IL-6 \[[@B55]\] whose receptors are expressed in pediatric ALL with the t(4;11)/AF4 translocation \[[@B56]\]. The overexpression of *CSRP2*in *MLL*mutated samples distinguishing ALL and AML supports a novel role for *CSRP2*gene in leukemia development. Conclusion ========== Our study demonstrates that a small subset of genes identifies leukemias with *MLL*-specific rearrangements and clearly separates acute leukemia samples according to lineage origin. The subset includes well-known genes and newly discovered markers that allow for characterization of ALL and AML subgroups, with and without *MLL*rearrangements. Abbreviations ============= The following abbreviations have been used throughout the manuscript. *ALL*: Acute Lymphoblastic Leukemia; *AML*: Acute Myeloblastic Leukemia; *MLL*: Mixed Lineage Leukemia; *SAM*: Significance Analysis of Microarray; *PAM*: Prediction Analysis of Microarray; *FAB*: French-American-British; *RMA*: Robust Multiple-array Average; *FDR*: False Discovery Rate; *GO*: Gene Ontology. Competing interests =================== The authors declare that they have no competing interests. Authors\' contributions ======================= AZ established study aim and design, performed gene expression analyses and wrote the manuscript, MCDO performed microarray experiments, GTK and GB supervised the study and writing of the manuscript. All authors read and approved the final manuscript. Pre-publication history ======================= The pre-publication history for this paper can be accessed here: <http://www.biomedcentral.com/1755-8794/2/36/prepub> Supplementary Material ====================== ###### Additional file 1 **Patient information**. Supplementary information about the 140 patient cohort enrolled in the present study. ###### Click here for file ###### Additional file 2 **SAM and PAM results**. SAM results for *MLL*translocation-specific (379 probes) and phenotype-specific (622 probes) signatures. In addition, PAM results (15 probes) are listed in a separated column. Probe set number and gene symbol have been provided for each analysis. ###### Click here for file ###### Additional file 3 **Validation on Test Set**. Unsupervised hierarchical clustering of 78 pediatric leukaemia patients using the 15-probes predictor identified by SAM and PAM analyses. The dendrogram separates AML (red and blue labels) from ALL (green and orange) samples. Each group further divides into MLL-positive and MLL-negative samples. ###### Click here for file ###### Additional file 4 **Gene Ontology results**. Gene ontology results for phenotype-specific (379 probes) and *MLL*translocation-specific (622 probes) signatures using DAVID web-based tool. Biological processes are ordered by FDR relevance; highlighted rows depict GO terms with FDR \<10%. ###### Click here for file ###### Additional file 5 **AML patients without MLL mutation and MEIS1 upregulated**. SAM results from the analysis of 26 AML patients without *MLL*translocation. Class comparison analysis was performed between AML/MLL- patients with MEIS1-upregulated and MEIS1-downregulated (see hierarchical clustering results for PAM analyses, Figure [3](#F3){ref-type="fig"}). Probes up-regulated in one group are considered down-regulated in the other, and vice versa. ###### Click here for file Acknowledgements ================ This work was supported by Fondazione Città della Speranza, P.R.I.N. 40%, MIUR and AIRC. We gratefully acknowledge Dr. Manuela Frascella and Dr. Martina Pigazzi for biomolecular analyses, Dr. Laura Sainati and Dr. Monica Spinelli for morphological analyses, Dr. Anna Leszl for cytogenetic analyses, Dr. Colette Case for manuscript revision.

Exchange bias (EB) is demonstrated as the magnetic hysteresis loop shifts from the origin, which is usually induced by field cooling (FC) to form a unidirectional magnetic anisotropy at the interface between ferromagnetic (FM) and antiferromagnetic (AFM) phases. The EB effect was discovered in 1956 by Meiklejohn and Bean in Co/CoO nanoparticles[@b1], and has been extensively studied and explored in various systems including FM-AFM nanocomposite/bilayers[@b2], spin glass (SG)-FM structures[@b3], and more[@b4][@b5]. These systems have significant applications in ultrahigh-density magnetic recording, giant magnetoresistance, and spin valve devices[@b6][@b7][@b8]. Recently, spontaneous EB (SEB) has attracted much more interest because the exchange bias is able to occur under zero-field cooling (ZFC) conditions. Conventional exchange bias(CEB) is usually observed in systems with interface between different magnetic phases after field cooling, as long as there is net remnant magnetization of the ferromagnetic layer present on cooling to below the blocking temperature. In contrast to CEB, the spontaneous exchange bias(SEB) could be observed without the assistance of external magnetic field. Thus the SEB would be of easy to electric field control of EB devices as it eliminates the requirement of external magnetic field to create the unidirectional anisotropy[@b9][@b10]. SEB was first reported in NiMnIn alloys by Wang *et al*.[@b11] in 2011 and was attributed to a superferromagnetic (SFM) unidirectional anisotropy formed at the interface between superparamagnetic (SPM) domains and the AFM matrix. This SEB was subsequently explored in other magnetic systems and several different mechanisms were proposed. For example, Maity *et al*.[@b12] showed SEB in a nanocomposite of BiFeO~3~-Bi~2~Fe~4~O~9~ driven by the superinteraction bias coupling via superspin glass moments at the FM/AFM interface. Han *et al*. observed SEB in a NiMnGa alloy caused by irreversible growth of FM domains changing from a non-percolating to a percolating state[@b13]. Obviously, a comprehensive understanding of the SEB effect is still disputed. Of all the materials associated with SEB, off-stoichiometric Mn-rich Ni--Mn--Z (Z = Ga, Sn, In, Sb) Heusler alloys demonstrate remarkable EB owing to their underlying physics[@b11][@b13]. The Mn-Mn exchange interaction within the regular Mn sublattices and the Mn-Mn exchange interaction between the regular Mn sublattice and the Z sublattice correspond to the FM and the AFM phases, respectively[@b9][@b14][@b15]. The competition between the FM and the AFM phases can generate frustrated magnetic behavior and thereby lead to the SEB effect. In the other words, these magnetic Heusler alloys firstly undergo a ferromagnetic transition in austenitic state and then a first-order magnetically martensitic transition from an austenitic state with cubic structure to a martensitic state, another magnetization to the spin glass (canonical SG/cluster SG) or the super-paramagnet may occurred when continue to decrease temperature[@b16][@b17][@b18]. Abundant physical properties could be caused by the SPM or the SG[@b11][@b13]. Generally, the research would be more comprehensive through studying of composition-dependent properties. Considering the potential applications, there is an urgent demand for tunable SEB, especially in the read head of ultrahigh sensitivity. In this work, we report a giant SEB effect in Ni~50~Mn~38~Ga~12-x~Sb~x~ Heusler alloys. The maximum exchange bias *H*~*E*~ is 2930 Oe for Ni~50~Mn~38~Ga~10~Sb~2~ at 2 K, which is much higher than most other SEB materials. Most importantly, the *H*~*E*~ can be tuned from 230 Oe to 2930 Oe by adjusting the Sb alloy composition. Based on further analysis and the composition-temperature phase diagram of Ni~50~Mn~38~Ga~12-x~Sb~x~ (where x = 0--6), we demonstrate that this giant SEB is triggered by the crossover of superspin glass from canonical SG to cluster SG. Results and Discussions ======================= [Figure 1(a)](#f1){ref-type="fig"} shows the magnetization temperature dependence for zero field-cooled (ZFC) and field-cooled\` (FC) of Ni~50~Mn~38~Ga~12-x~Sb~x~ (0 ≤ × ≤ 6) alloys under 200 Oe applied magnetic field. Sb0, Sb2, Sb4, and Sb6 correspond to samples Ni~50~Mn~38~Ga~12~Sb~0~, Ni~50~Mn~38~Ga~10~Sb~2~, Ni~50~Mn~38~Ga~8~Sb~4~, and Ni~50~Mn~38~Ga~6~Sb~6~, respectively. It is observed that the magnetization increases as the Sb concentration increases. The ZFC curve exhibits a peak *T*~*p*~ around 100 K and an irreversibility between ZFC and FC curves occurring at *T*~*p*~, which is similar to that of of NiMnIn and NiMnGa[@b11][@b13]. This peak shifts towards higher temperatures as the Sb concentration increases, as seen at *T*~*p*~ = 75 K, 115 K, 122 K, and 128 K for Sb0, Sb2, Sb4, and Sb6, respectively. Another interesting feature is the systematic difference between the ZFC and the FC curves, with an appearance of an irreversible behavior below the *T*~*p*~. In general, the M-T behaviors of the SG and the SPM systems at lower temperature are similar under zero field cooling. However, when cooled in the presence of a magnetic field, the magnetization increases monotonically in SPM systems as the temperature decreases. Rather, the magnetization of SG systems tends to saturate or decrease with decreasing temperature. As shown in [Fig. 1(a)](#f1){ref-type="fig"}, the FC curve behavior of all samples indicates the SG feature in the alloys. To further clarify the origin of the complicated magnetic states below *T*~*P*~, the temperature dependence of the real portion (χ^/^) of the AC susceptibility was measured at a low magnetic field strength of 2 Oe ([Fig. 1b](#f1){ref-type="fig"}). The peak shifted toward higher temperatures and its value decreased as frequency increased. The shape of the peak at *T*~*P*~ became distinct for increased Sb concentration. This may be attributed to the atomic radius of Sb, which is larger than that of Ga; a larger atomic radius weakens the AFM by increasing the Mn-Mn distance. To extract additional parameters, we use the Vogel-Fulcher(V-F) relationship to analyze the SG-like behavior[@b19]. where *ω* is the measurement angular frequency, is the characteristic frequency of the SG, *E*~*a*~ is the activation energy of the SG, *k*~*B*~ is the Boltzmann constant, *T*~*f*~ is the frequency temperature defined above, and *T*~*0*~ is the V-F temperature that describes the interaction among SG clusters. The *T*~*0*~ values are very close to the *T*~*P*~ values obtained from the MT curves at 200 Oe in [Fig. 1(a)](#f1){ref-type="fig"}. *τ*~0~ (*τ*~0~ = 2*π*/*ω*~0~) is the microscopic relaxation time related to spin glass or cluster spin glass. All of the parameters were obtained by best fit and are shown in [Table 1](#t1){ref-type="table"}. These SG observations are in good agreement with results from the literature. For example, *τ*~0~ varies from 10^−11^\~10^−15^ for canonical SG, but the typical value for cluster SG compounds is around *τ*~0~ = 10^−919,^[@b20]. It is obvious from [Table 1](#t1){ref-type="table"} that all the characteristic parameters fall in the typical value ranges. According to previous reports[@b11][@b14], the SEB normally occurs in complex magnetic states of SG, AFM, and SSG. Hence, the magnetization loops M(H) for different Ni~50~Mn~38~Ga~12-x~Sb~x~ alloys have been measured at 2 K, as indicated in [Fig. 2(a--d)](#f2){ref-type="fig"}. The inset of each figure shows a magnified view of the loop shift towards the negative field axis; a pronounced shift can be seen. All the hysteresis loops are shifted along the field axis, indicating the presence of the EB effect. Note that the sample was cooled in a zero magnetic field from 400 K. The FM characteristic is more evident as Sb concertation increases, which indicates that the AFM interaction decreases as well. A clear SEB effect can be found at x = 2 and x = 4 (inset figure), but is absent at x = 0 and x = 6. The largest EB is observed at x = 2. This phenomenon may be attributed to the fact that Ni~50~Mn~38~Ga~10~Sb~2~ is a boundary of canonical spin and cluster spin. At the phase boundary, the ferroic system thermodynamically achieves a state with a flattened free energy profile, in which the polarization or magnetization can easily rotate almost without energy barrier. *H*~*E*~ and *H*~*C*~ are defined as *H*~*E*~ =  −(*H*~*L*~ + *H*~*R*~)/2 and *H*~*C*~ = −(*H*~*L*~ − *H*~*R*~)/2, respectively, where *H*~*L*~ and *H*~*R*~ are the left and right coercive fields, respectively[@b6][@b8]. *H*~*E*~ and coercivity *H*~*C*~ as functions of Sb content are shown in [Fig. 2(e)](#f2){ref-type="fig"} for the different Ni~50~Mn~38~Ga~12-x~Sb~x~ alloys. The *H*~*E*~ and *H*~*C*~ change from 230 Oe and 2930 Oe to 610.83 Oe and 1024.16 Oe, respectively, as the Sb concentration increases from 0 to 2. The *H*~*E*~ varies in the 2930 Oe and 80 Oe range upon doping at x = 6, and the coercivity (*H*~*C*~) also decreases from 2070 to 350 Oe. The EB effect for the Ni~50~Mn~38~Ga~10~Sb~2~ alloy at different magnetic fields (10 kOe to 60 kOe at 2 K) was also investigated. As shown in [Fig. 3(a)](#f3){ref-type="fig"}, *H*~*E*~ is 100 Oe when the applied magnetic field is under 20 kOe but continually grows until reaching a maximum value of 2930 Oe at the critical magnetic field of 50 kOe. This maximum occurring at the field of 50 kOe suggests the formation of the maximum SSG unidirectional anisotropy. Beyond an applied magnetic field strength of 50 kOe, the *H*~*E*~ begins to decrease. The *H*~*E*~ decrease may originate from the change of the SSG spin structure under larger applied magnetic fields. Meanwhile, *H*~*C*~ shows a non-monotonic behavior. The maximum of *H*~*C*~ occurs at 40 kOe, which is lower than that of *H*~*E*~. This indicates that interfaces associated with phase separation are not so stable in high magnetic fields because of the propagation of the SSG phase. A high magnetic field can disturb the interfacial exchange coupling and remove the exchange anisotropy, which can lead to the EB disappearing. From these observations, we can conclude that the number of SPM domains in the SSG state will decrease with increasing *H* due to more of the SSG states transforming to AFM states. This generates new interfaces with SSG unidirectional anisotropy[@b11]. The calculated *H*~*E*~ and *H*~*C*~ as a function of temperature for the Ni~50~Mn~38~Ga~10~Sb~2~ alloy are depicted in [Fig. 3(b)](#f3){ref-type="fig"}. *H*~*E*~ decreases as the temperature increases, and becomes almost zero at about 50 K. This temperature can be referred to as the EB blocking temperature (*T*~*B*~), because the EB effect vanishes above this temperature; it can be attributed to the weakening of the SSG-AFM coupling as temperature increases. The disappearance of the *H*~*E*~ value above the blocking temperature (*T*~*B*~ ≈ 50 K) is due to the fact that AFM interactions begin to dominate the SSG interactions. On the other hand, *H*~*C*~ initially increases with temperature and begins to decrease after reaching a maximum value. It is known that anisotropy of SSG decreases with increasing temperature, because the AFM is able to drag more SSG spins resulting in increasing coercivity below *T*~*B*~. To investigate the SEB training effect, the sample was cooled without field from 400 K to 2 K and ten hysteresis loops were measured, consecutively. As plotted in [Fig. 4](#f4){ref-type="fig"}, training effect is obviously observed in Ni~50~Mn~38~Ga~10~Sb~2~ alloy. The first cycle loop shows a prominent SEB effect with an asymmetric magnetization reversal characteristic. After that, the subsequent magnetization cycles tend to symmetric gradually. The central part of first, second, and tenth loops are shown in inset [Fig. 4(a)](#f4){ref-type="fig"}. As seen, the left coercive field *H*~*L*~ decreases dramatically with increased cycle of n, while *H*~*R*~ shows a qualitatively similar but by far less pronounced dependence on *n*. The different strengths of *n* dependences of *H*~*L*~ and *H*~*R*~ indicate that the left and right branches of the hysteresis follow different mechanisms of magnetization reversal. The usual experimentally observed relationship between *H*~*EB*~ and *n* can be expressed by [equation (2)](#eq3){ref-type="disp-formula"}. where *H*~*E*∞~ is the exchange-bias field at the *n*th cycle (in the limit of an infinite number of cycles). The results presented in [Fig. 4(b)](#f4){ref-type="fig"} were fitted using the [Equation (2)](#eq3){ref-type="disp-formula"}, giving *H*~*E*∞~ = 1210 Oe. This training effect was interpreted in terms of metastable magnetic disorder at the magnetically frustrated interface during the magnetization reversal process for SEB[@b21][@b22]. To further detail composition-dependent SEB evolution, we constructed the composition-temperature phase diagram of Ni~50~Mn~38~Ga~12-x~Sb~x~ (where x = 0--6). The composition-temperature phase diagram was determined by the martensite start temperature, *T*~*MS*~, detected using DSC \[[Fig. 5(a)](#f5){ref-type="fig"}\], the spin glass temperature, *T*~*0*~, determined by fitting AC lines \[[Fig. 1(b)](#f1){ref-type="fig"}\], and the block temperature, *T*~*B*~, collected by measuring the exchange bias (*H*~*E*~) versus temperature (*T*). [Figure 5(a)](#f5){ref-type="fig"} shows the DSC of the alloys, measured at 10 K/min heating and cooling rates. Large exothermic and endothermic peaks during cooling and heating were observed corresponding to direct and reverse martensitic transformations, respectively. Take Ni~50~Mn~38~Ga~12~ as an example, its characteristic martensitic transition temperatures including the austenite start and finish (*A*~*s*~ and *A*~*f*~) and martensite start and finish (*M*~*s*~ and *M*~*f*~) temperatures are 660 K, 674 K, 636 K, and 620 K, respectively. Its corresponding thermal hysteresis can become as large as 20 K, indicating that the martensitic phase transition is a first order phase transition. It also can be seen that the increase of Sb content may shift the transformation peaks to a lower temperature regime. Therefore, a partial substitution of Ga by Sb would suppress the martensitic transformation occurrence. In particular, we observed a wild martensitic transformation behavior in the DSC curve of the Ni~50~Mn~38~Ga~10~Sb~2~ alloy \[see [Fig. 5(a)](#f5){ref-type="fig"} blue circle\]. This DSC peak characteristic probably account for the presence of jerky characteristic in martensitic phase of the Ni~50~Mn~38~Ga~10~Sb~2~ alloy. It would be shown some special performance in martensitic phase[@b23][@b24][@b25]. This is in good agreement with previous AC curves ([Fig. 1](#f1){ref-type="fig"}). In addition, the composition Ni~50~Mn~38~Ga~10~Sb~2~ might be considered as co-existence of canonical spin and cluster spin, with the cluster spin being dominant. From the phase diagram shown in [Fig. 5(b)](#f5){ref-type="fig"}, the martensite transformation temperature of the Ni~50~Mn~38~Sb~x~Ga~12-x~ (x = 0--6) alloys decreases slightly as the Sb content increases. At the same time, the AFM phase remains almost unchanged while the SG phase varies with the Sb content. In addition, the pinning phase changes from canonical SG to cluster SG when *x* is larger than 3. The *T*~*B*~ holds at 50 K, because the anisotropy of SG is more related to temperature. We also considered the evolution of the magnetic state of Ni~50~Mn~38~Ga~12-x~Sb~x~ (x = 0--6) after ZFC under different temperatures, as shown in the inset of [Fig. 5(b)](#f5){ref-type="fig"}. Regions (1), (2), and (3) corresponds to (x ≤ 1), (1 \<× \< 3), and (3 ≤ × ≤ 6), respectively. The physical schematic diagram for the martensitic phase is shown in [Fig. 5](#f5){ref-type="fig"}(1--1),(2--1) and (3--1); it is a simplified schematic diagram with unfrozen nano SPM domains embedded in an AFM single domain under martensitic phase. The domain sizes increase with the increase of Sb constant. Moreover, all the domains are random. Upon cooling to *T* \< *T*~*0*~, the thermal energy becomes lower than the average energy barrier for flipping local domains. As a result, the domain flipping essentially stopped and the system is frozen. This process represents a glass transition from the martenstic state to a frozen spin glass state. The physical schematic diagram for the SG system is drawn in [Fig. 5](#f5){ref-type="fig"}(1--2),(2--2) and (3--2). The domain size is larger than that in martensitic phase, and the domain is frozen. The part of (1--2) shows canonical SG and the part (3--2) shows the cluster SG. Therefore, the Ni~50~Mn~38~Ga~10~Sb~2~ alloy can be regarded as a boundary between the canonical SG and the cluster SG. It shows the cluster SG (calculate by fitting AC lines), but it also keeps some canonical SG domain. As the temperature further decreases to *T* \< *T*~*B*,~ the anisotropy in the SG phase increases \[see dashed red circles in (2--3), (3-3)\]. More important, at the phase boundary, the system achieves a thermodynamic state with a flattened free energy profile. In this state, the polarization or magnetization is easily rotated almost without energy barrier[@b26][@b27][@b28]. Therefore, the largest *H*~*E*~ (\~2930 Oe) was obtained in the Ni~50~Mn~38~Ga~10~Sb~2~ alloy sample. Conclusion ========== In summary, SEB continuous tuning was demonstrated in Ni~50~Mn~38~Ga~12-x~Sb~x~ (x = 0--6) alloys, and the largest *H*~*E*~ (\~2930 Oe) was obtained in the Ni~50~Mn~38~Ga~10~Sb~2~ sample. Moreover, the tunable *H*~*E*~ physical mechanism was investigated. This behavior can be explained by the pinning phase from the canonical SG to the cluster SG through the change of the atomic radius between Mn-Mn by substituting Z under the off-stoichiometric Mn rich Ni-Mn-Z (Z = Sn, Ga, and Sb) alloys. The Ni~50~Mn~38~Ga~10~Sb~2~ alloy is like a phase boundary between the canonical SG and the cluster SG at lower temperatures. Thus, the largest SEB is observed in the Ni~50~Mn~38~Ga~10~Sb~2~ sample. More remarkable, the SEB decreased with increasing temperature, approaching zero around *T *= 50 K. The SEB change is due to decreasing SG phase anisotropy \[see the dashed red circles in [Fig. 5(b)](#f5){ref-type="fig"}\] as temperature increases. Our results not only open a new direction to realize the EB effect, but also indicate a way to fabricate SEB with different values and promote the application of magnetic Heusler alloys in the ultrahigh-density magnetic recording, giant magnetoresistance, and spin-valve devices. Methods ======= The polycrystalline Ni~50~Mn~38~Ga~12-x~Sb~x~ (x = 0\~6) alloys with are prepared by arc melting high-purity (99.99%) Ni, Mn, Ga, and Sb in an argon atmosphere. To achieve high composition homogeneity, the samples are wrapped in silica tubes and are vacuum annealed at 1173 K for 24 hours. We used differential scanning calorimetry (DSC) to determine the martensitic transition temperature. The magnetic properties, including magnetic hysteresis (MH) loops, magnetization-temperature (MT) curves, and AC susceptibility measurements are measured using a superconducting quantum interference device (SQUID) magnetometer (Quantum Design, MPMS-XL-5). We used the function of magnet reset in SQUID (A heater integrated into the magnet solenoid may be used to drive the wire above its critical temperature and eliminate most trapped magnetic flux.) to demagnetize the superconducting magnet before testing ZEB effect, and all the samples were first warm to 400 K and then cooled to 2 K before test. The composition-temperature phase diagram of Ni~50~Mn~38~Ga~12-x~Sb~x~ (x = 0--6) is plotted according to the magnetic properties and the EB measurements. Additional Information ====================== **How to cite this article**: Tian, F. *et al*. Giant spontaneous exchange bias triggered by crossover of superspin glass in Sb-doped Ni~50~Mn~38~Ga~12~ Heusler alloys. *Sci. Rep.* **6**, 30801; doi: 10.1038/srep30801 (2016). This work was supported by the National Basic Research Program of China (grant no. 2012CB619401), the National Science Foundation of China (grant no. 51371134, 51471125 and 51431007), the Program for Key Science and Technology Innovative Team of the Shaanxi Province (No. 2013KCT-05), and the Collaborative Innovation Center of Suzhou Nano Science and Technology. **Author Contributions** S.Y., Y.Z. and X.S. conceived the experiments and analyzed the results. F.T. conducted the experiments, analyzed the results and wrote the manuscript. K.C. and Y.Z. take part in the experiments and data analysis. R.Z., T.C., C.Z. and M.X. participated in discussion and helped modify the manuscript. All authors reviewed the manuscript. ![(**a**) The magnetization temperature dependence \[M(T)\] curves measured at 200 Oe with sequences of ZFC and FC, respectively, (**b**) Temperature dependence of the real part of the ac susceptibility measured at different frequencies with an AC magnetic field of 2 Oe, the inset (**b**) shows the correlation between the angular frequency and the blocking temperature (). The frequency () dispersion behavior of temperature () conforms to the Vogel-Fulcher relationship. The arrows indicate the direction of increasing frequencies.](srep30801-f1){#f1} {#f2} {#f3} {ref-type="disp-formula"} to the data for *n* *\>* 1.](srep30801-f4){#f4} {#f5} ###### Parameter values obtained by fitting the experimental data to [Equation (1)](#eq1){ref-type="disp-formula"}. Samples T~0~(K) τ~0~(s) E~a~/K~B~(K)   --------------------------------------------------------- --------- --------------- -------------- ---------------------- **Ni**~**50**~**Mn**~**38**~**Ga**~**12**~ 65.1 9.3 × 10^−14^ 43.8 canonical spin glass **Ni**~**50**~**Mn**~**38**~**Ga**~**11**~**Sb**~**1**~ 90.3 8.7 × 10^−12^ 72.5 canonical spin glass **Ni**~**50**~**Mn**~**38**~**Ga**~**10**~**Sb**~**2**~ 130.7 2.4 × 10^−10^ 104.1 cluster spin glass **Ni**~**50**~**Mn**~**38**~**Ga**~**9**~**Sb**~**3**~ 140.4 5.3 × 10^−9^ 158.4 cluster spin glass **Ni**~**50**~**Mn**~**38**~**Ga**~**8**~**Sb**~**4**~ 147.5 7.2 × 10^−9^ 173.3 cluster spin glass **Ni**~**50**~**Mn**~**38**~**Ga**~**7**~**Sb**~**5**~ 151.9 2.1 × 10^−8^ 191.7 cluster spin glass **Ni**~**50**~**Mn**~**38**~**Ga**~**6**~**Sb**~**6**~ 156.4 1.3 × 10^−7^ 213.6 cluster spin glass

Introduction {#cga12102-sec-0001} ============ Reproduction is a highly regulated process that requires coordination of the functions of numerous genes. Infertility affects 10--15% of couples, and a male factor is estimated to be involved in nearly half of these cases (Visser and Repping [2010](#cga12102-bib-0019){ref-type="ref"}). Approximately 600 testis‐specific protein‐coding genes have been identified. Null mutations have been introduced in nearly 400 genes associated with spermatogenesis using knockout mouse technology (Matzuk and Lamb [2008](#cga12102-bib-0013){ref-type="ref"}; Jamsai and O\'Bryan [2011](#cga12102-bib-0010){ref-type="ref"}; Massart et al. [2012](#cga12102-bib-0012){ref-type="ref"}). Information regarding genetic abnormalities in spermatogenesis derived using reverse genetics approaches is expected to help further understand male infertility. Recent advances in genetics have paved the way for the development of effective methods to study male infertility and subfertility. Accordingly, the number of "repro" mouse produced by JAX Reproductive Mutagenesis Program (Handel et al. [2006](#cga12102-bib-0008){ref-type="ref"}) reached more than a hundred, and a number of genes responsible for male infertility and subfertility have been identified (<http://reprogenomics.jax.org>). The bidirectional approaches described above are intended to study infertility caused by a single gene. However, the reproduction is temporally regulated by coordinated action of a number of genes. Our approach was also genetic, but the phenotype was polygenic. Previously, 17 strains of mice were surveyed for frequency of abnormal sperm‐head morphology (Gotoh [2010](#cga12102-bib-0004){ref-type="ref"}) and found that B10.M and B10.MOL‐TEN1 (TEN1) strains were ranked first and second, respectively, with respect to the frequency of sperm‐head morphological abnormality. Analysis of B10.M strain revealed that the sperm phenotype was heritable. Further genetic analysis identified two causative loci, *Shm1* on Chromosome 1 and *Shm2* on Chromosome 4 (Gotoh et al. [2012](#cga12102-bib-0005){ref-type="ref"}). The B10.M strain showed low fecundity. The relationship between sperm‐head morphology and male subfertility is yet to be established. In this study, we analyzed the genetic control of sperm‐head morphological abnormality of TEN1 strain. The relationship between sperm‐head morphological abnormality and male fertility of B10.M and TEN1 strains was discussed. Materials and Methods {#cga12102-sec-0002} ===================== Animals {#cga12102-sec-0003} ------- All experiments were approved by the Institutional Animal Care and Use Committee of the National Institute of Agrobiological Sciences (study identification code \#H18‐010). Animals were housed and cared for according to guidelines established by the Committee. B10.MOL‐TEN1 (TEN1) mice were from RIKEN BioResource Center (Tsukuba, Ibaraki, Japan). C3H/HeNCrlCrlj (C3H) mice were purchased from Charles River Japan (Yokohama, Japan). C57BL/10J (B10) mice were purchased from S. L. C. (Hamamatsu, Japan). B10.M/Sgn (B10.M) mice are maintained at our facility. Animals were maintained on a cycle of 12 h of light and 12 h of darkness under specific‐pathogen‐free conditions. A commercial mouse diet CRF‐1 (Charles River Japan, Yokohama, Japan) and water were provided. TEN1 strain {#cga12102-sec-0004} ----------- This strain was one of the congenic lines established by Dr Kazuo Moriwaki at the National Institute of Genetics, Japan (Mishima, Shizuoka, Japan) to define *H2* alleles on chromosome 17 in the Japanese wild mouse population (Shiroishi et al. [1982](#cga12102-bib-0016){ref-type="ref"}). The original donor was a wild non‐inbred animal (*Mus Musculus molossinus*). The wild‐derived *H2* allele was transmitted to B10 over 10 generations. Subsequently this strain has been maintained by sibling breeding. Sperm morphology test {#cga12102-sec-0005} --------------------- Sperm samples were collected from 2‐ to 3‐month‐old male mice. The sperm‐head morphology test was performed as described earlier (Gotoh [2010](#cga12102-bib-0004){ref-type="ref"}; Gotoh et al. [2012](#cga12102-bib-0005){ref-type="ref"}). Briefly, to obtain sperm for the assessment of morphology, the epididymis was removed from one side and sperms were transferred into 1‐mL phosphate‐buffered saline (PBS, pH 7.0) solution containing 0.1% glucose (w/v). One microliter of the sperm suspension was spread on a glass slide, air‐dried, fixed using ethanol, and sperm morphology was assessed by a phase‐contrast microscopy (Optiphot‐2; Nikon, Tokyo, Japan) (400 × magnification). Two independent samples, each containing at least 200 sperm cells, were analyzed. The epididymis from the second side was used for sperm counting. The organ was transferred into a conical glass grinder and was manually homogenized. The total number of spermatozoa present in the epididymis was counted by using a Neubauer hemocytometer and a phase‐contrast microscope (400 × magnification). Classification of sperm‐head morphological abnormality (Fig. [1](#cga12102-fig-0001){ref-type="fig"}) was done as described earlier (Gotoh [2010](#cga12102-bib-0004){ref-type="ref"}; Gotoh et al. [2012](#cga12102-bib-0005){ref-type="ref"}). {#cga12102-fig-0001} Animal mating for genetic analysis {#cga12102-sec-0006} ---------------------------------- The F1 hybrid males were generated by reciprocal crosses between C3H and TEN1 (CTF1; C3H female × TEN1 male, and TCF1; TEN1 female × C3H male). The F1 hybrids were then backcrossed either with C3H or with TEN1. The F2 hybrid males were generated by crosses between F1 females and F1 males. Preparation of genomic DNA {#cga12102-sec-0007} -------------------------- Mouse tails were resuspended in 300 μL Tris buffer (50 mM, pH 7.8) containing 100 mM ethylenediaminetetraacetic acid (EDTA), 100 mM NaCl, and 1% sodium dodecyl sulfate (SDS) (w/v). Proteinase K was then added to a final concentration of 500 μg/mL. The samples were shaken 1000 rpm, overnight at 56°C (Thermomixer comfort; Eppendorf, Hamburg, Germany). RNase A was then added to a final concentration of 10 μg/mL, and the samples were incubated for 1 h at 37°C. Following this, 300 μL DNA binding reagent containing 4.5 M guanidine hydrochloride, 0.5 M potassium acetate (pH 5.0), and 40 mg/mL Silica gel (Sigma‐Aldrich \#288519, St. Louis, MO, USA) was added to the samples. The samples were then mixed by shaking for 1 min to allow DNA to bind to the silica gel. Silica gel particles were retrieved and washed three times with a solution containing 80% ethanol, 10 mM potassium acetate (pH 5.0), and 20 μM EDTA. After air‐drying, bound DNA was eluted from silica gel using the TE buffer (pH 8.0). DNA concentration was calculated indirectly from the absorbance at 260 nm measured using UV/Vis spectrophotometer (Gene Quant; Pharmacia Biotech, Cambridge, England). Linkage analysis {#cga12102-sec-0008} ---------------- Genomic DNA was analyzed using microsatellite markers. Because the *H2* complex in mouse is located between the *H2‐K1* gene at 17.98 centimorgan (cM) (33996017‐34000333 bp; VEGA annotation of GRCm38) and the *H2‐T3* gene at 18.99 cM (36185572--36190287 bp), we chose polymorphic markers located around the *H2* complex. The specific markers used were D17Mit46 at 12.53 cM (25365940--25366175 bp), D17Mit22 at 17.98 cM (34333549--34333707 bp) and D17Mit139 at 27.40 cM (52659266--52659396 bp). Sequences of the primer pairs were as follows. fD17Mit, 5′‐TCCACCCCACTACCTGACTC‐3′ and 5′‐CCCTTCTGATGACCACAGGT‐3′; D17Mit22, 5′‐GGTAAGCATTAGATAGAGAG‐3′ and 5′‐TTATGATCTCCACACACGTG‐3′; D17Mit139, 5′‐AGACATGTGAGTACTGCACAGACA‐3′ and 5′‐ATGATGACATACCTCCTAGTAGTCCC‐3′. Primer pairs for microsatellite markers were purchased from Tsukuba Oligo (Ibaraki, Japan). Typically, a 10‐μL volume of each reaction mixture contained 10 ng of tail DNA, 50 mM KCl, 10 mM Tris‐HCl (pH 8.3), 2.5 mM MgCl~2~, 200 μM oligonucleotides, 200 μM dNTP, and 0.02 U AmpliTaq DNA polymerase (PerkinElmer, Norwalk, CT, USA). The cycling conditions were as follows: an initial step of denaturation for 2 min at 95°C; 45 s at 94°C (38 cycles); 1 min at 55°C; 1 min at 72°C; and a single final step of a 7‐min extension at 72°C. Polymerase chain reaction (PCR) products were separated by electrophoresis on a 4% MetaPhor agarose gel (FMC Bioproducts, Rockland, ME, USA) and visualized under UV light after staining with ethidium bromide. Male fertility analysis {#cga12102-sec-0009} ----------------------- Mature virgin C3H females aged between 8--12 weeks were used for the assay. A female mouse was mated with a male mouse. A vaginal plug was observed; 12 days later, the female mice were dissected, and the numbers of live embryos, placental remnants and implantation scars were counted according to the procedure used for the dominant lethal studies (Tezuka et al. [1978](#cga12102-bib-0018){ref-type="ref"}; Teramoto et al. [1980](#cga12102-bib-0017){ref-type="ref"}). Histology {#cga12102-sec-0010} --------- Testes from TEN1 (*n* = 5) and B10 (*n* = 5) strains were fixed in Bouin\'s fluid, and 5 μm sections of paraffin‐embedded samples were stained with periodic acid‐schiff (PAS), and hematoxylin and eosin. Each seminiferous tubule was classified into one of the 12 stages of spermatogenesis as described previously (Russel et al. [1990](#cga12102-bib-0015){ref-type="ref"}). Differences in the histology of each stage between the mouse strains were compared. Testis histology was examined by light microscopy (Optiphot‐2; Nikon, Tokyo, Japan). Statistical analysis {#cga12102-sec-0011} -------------------- Because both the B10.M and TEN1 strains are *H2* congenic strains on the background of the B10 genome, B10 was used as control. Data obtained in the present study were analyzed with routine multiple comparison tests as described earlier (Hojo et al. [2006](#cga12102-bib-0009){ref-type="ref"}; Aoyama et al. [2012](#cga12102-bib-0002){ref-type="ref"}) with slight modifications. Briefly, the data were first examined for equality of variance by the Bartlett\'s test (α = 0.05), which was followed by ANOVA when the group variance was homogeneous or by the Kruskal--Wallis test when the group variance was heterogeneous for analyzing differences among four strains or four crosses. When ANOVA was significant (α = 0.05), the Dunnett\'s test (parametric analysis) was conducted to detect differences between B10 and other strains of male mice or between C3H × B10 and other crosses α = 0.05). Differences between B10 and other strains or between C3H × B10 and other crosses were analyzed by the Dunnett‐type mean rank sum test (non‐parametric analysis, α = 0.05) if the Kruskal--Wallis test was significant (α = 0.05). Results {#cga12102-sec-0012} ======= Sperm‐head morphological abnormalities {#cga12102-sec-0013} -------------------------------------- The TEN1 strain displayed teratozoospermia. A wide variety of sperm‐head morphologies, including a shortened apical hook, the lack of the usual hook, a bent hook, a round head, a pin head, an amorphous head, and spermatozoa with ectopic attachment of the flagella were observed (Fig. [1](#cga12102-fig-0001){ref-type="fig"}). The observed variation in sperm‐head morphology was similar to that found in the B10.M strain (Gotoh [2010](#cga12102-bib-0004){ref-type="ref"}; Gotoh et al. [2012](#cga12102-bib-0005){ref-type="ref"}). In the TEN1 strain, frequencies of each abnormal type of sperm varies among individuals. The total frequency of various types of morphological sperm‐head abnormalities was found to be heritable. This feature was the same as that observed in the B10.M strain. In addition, besides sperm‐head abnormality, a considerable part of the sperm showed abnormality either in the midpiece or in tail. However, heritability of either midpiece‐abnormality or tail‐abnormality has not been established. Inheritance of the phenotype of sperm‐head abnormalities {#cga12102-sec-0014} -------------------------------------------------------- Figure [2](#cga12102-fig-0002){ref-type="fig"} shows the segregation of abnormal sperm‐head phenotypes of the TEN1 strain by genetic crosses between TEN1 and C3H strains. The mean frequencies of abnormal head morphology in the spermatozoa were 50.2% for TEN1 and 2.4% for C3H. The F1 hybrid males generated by reciprocal crosses (CTF1; C3H female × TEN1 male and TCF1; TEN1 female × C3H male) produced spermatozoa with a low frequency of abnormalities (average frequencies of 3.8% and 1.0%, respectively). We then backcrossed the F1 hybrids either with C3H or with TEN1. All of the male mice generated by backcrossing F1 hybrids with C3H produced sperm with a low frequency of abnormalities (mean, 4.1%). However, among the backcrosses of F1 hybrids to TEN1 mice, we observed segregation of the phenotypes of sperm‐head abnormalities to the high‐abnormality group and the low‐abnormality group. The frequency of males of the high‐abnormality group was 31%. In F2 hybrid mice, 18% of the males produced high levels of sperm with morphological abnormalities. The observed frequencies were found to be distributed continuously from the highest to the lowest. {#cga12102-fig-0002} Sperm counts and reproductive organs {#cga12102-sec-0015} ------------------------------------ Epididymal sperm count, body weight, and reproductive organ weight were also analyzed (Table [1](#cga12102-tbl-0001){ref-type="table"}). No significant differences were observed in the epididymal sperm count, testis weight, seminal vesicle weight, or epididymis weight between the TEN1 and B10 males. ###### Frequencies of abnormal spermatozoa, sperm count and weight of reproductive organs Strain Abnormal spermatozoa (%) No. Spermatozoa (×10^6^/epididymis) BW (g) Testis (g) SV (mg) Epididymis (mg) -------- ---------------------------------------------- ---------------------------------------------- ---------------------------------------------- ------------ --------------------------------------------- ---------------------------------------------- B10 Mean 4.8 16.3 28.2 176 356 78.5 SD 0.4 2.2 1.5 16 55 6.5 *n* 6 6 6 6 6 6 TEN1 Mean 50.2 17.5 34.6[\*](#cga12102-note-0003){ref-type="fn"} 172 330 74.4 SD 3.7 3.3 2 12 75 4.1 *n* 6 6 6 6 6 6 B10.M Mean 61.5[\*](#cga12102-note-0003){ref-type="fn"} 32.3[\*](#cga12102-note-0003){ref-type="fn"} 33[\*](#cga12102-note-0003){ref-type="fn"} 244 482[\*](#cga12102-note-0003){ref-type="fn"} 95.5[\*](#cga12102-note-0003){ref-type="fn"} SD 2.4 1.7 1.4 6 52 10.7 *n* 6 6 6 6 6 6 C3H Mean 2.3 7.1[\*](#cga12102-note-0003){ref-type="fn"} 31.8[\*](#cga12102-note-0003){ref-type="fn"} 145 236[\*](#cga12102-note-0003){ref-type="fn"} 56.2[\*](#cga12102-note-0003){ref-type="fn"} SD 0.5 2.3 1.4 27 35 4.4 *n* 6 6 6 6 6 6 Values represent mean, standard deviation (SD) and number examined. Data were statistically analyzed by Dunnett\'s test following one‐way ANOVA or Dunnett‐type test following Kruskal--Wallis test. Significantly different from control: \**P* ≤ 0.01. 2015 Japanese Teratology Society Chr 17 linkage {#cga12102-sec-0016} -------------- Because TEN1 was the *H2* congenic strain that showed a high frequency of sperm‐head abnormalities, we predicted that the gene responsible for the abnormal sperm‐head phenotype was located on Chr 17. However, analysis of this chromosome by using microsatellite markers showed no evidence of linkage of the abnormal sperm‐head phenotype to this chromosome. Male fertility {#cga12102-sec-0017} -------------- Male reproductive performances of the four strains, TEN1, B10.M, C3H and B10 were compared (Table [2](#cga12102-tbl-0002){ref-type="table"}), and no significant difference was observed in the number of ovulated oocytes, which were predicted by the number of corpora lutea, between the four groups. Significant loss of embryos was observed in all four groups in this study, on comparing the total number of corpora lutea and the total number of live embryos. The contribution of C3H females to this loss was not evident. On comparing both the mean number of live embryos and the rate of embryo survival between the four groups, B10.M males showed significantly low reproductive performance (*P* \< 0.05). No significant difference in male fertility was observed between the C3H males, B10 males and TEN1 males. ###### Comparison of male reproductive performance of four strains Cross (female × male) Mean no. embryos Mean no. live embryos Mean no. corpora lutea ----------------------- --------------------------------------------- --------------------------------------------- ------------------------ C3H × B10 Mean 8.1 7.6 10.3 SD 2.7 2.7 1.8 *n* 20 20 20 C3H × TEN1 Mean 7.7 7.3 9.8 SD 2.7 2.7 1.7 *n* 20 20 20 C3H × B10.M Mean 3.8[\*](#cga12102-note-0007){ref-type="fn"} 3.1[\*](#cga12102-note-0007){ref-type="fn"} 10.8 SD 2.9 2.5 1.2 *n* 20 20 20 C3H × C3H Mean 8.4 7.6 10.6 SD 2.3 2.5 1.7 *n* 20 20 20 Values represent mean, SD and number examined. Data were statistically analyzed by Dunnett\'s test following one‐way ANOVA or Dunnett‐type test following Kruskal--Wallis test. Significantly different from control: \**P* ≤ 0.01. 2015 Japanese Teratology Society Histology of the testis {#cga12102-sec-0018} ----------------------- The histological observations of testis between the B10 and the TEN1 mice testes was compared (Fig. [3](#cga12102-fig-0003){ref-type="fig"}). Various degenerative lesions of seminiferous epitheliums including a single‐cell necrosis, an empty space, and a severe atrophy of a tubule were occasionally observed in the TEN1 testis (Fig. [3](#cga12102-fig-0003){ref-type="fig"}B). Morphologically normal epitheliums were also observed (Fig. [3](#cga12102-fig-0003){ref-type="fig"}C). {#cga12102-fig-0003} Discussion {#cga12102-sec-0019} ========== It is difficult to predict male reproductive fertility, even in experimental animals. Several types of sperm tests, including sperm morphology, sperm count, sperm mobility and others are used to assess male fertility (Aitken [2006](#cga12102-bib-0001){ref-type="ref"}; Lewis [2007](#cga12102-bib-0011){ref-type="ref"}). However, no reliable correlation between the results of sperm tests and observed fertility has so far been established (Guzick et al. [2001](#cga12102-bib-0007){ref-type="ref"}). Our study was intended to focus on genetic regulation of sperm morphology in mice as a model system to analyze the complicated reproductive regulation in mammals. In this study, we analyzed if the teratozoospermia of the TEN1 strain is heritable, and then predicted the number of loci regulating the sperm‐head abnormal phenotype. Male fertility tests were also performed. Although reliable correlation between the results of sperm tests and observed fertility has not been established, sperm morphology, count, and motility are widely used to assess male fertility (Guzick et al. [2001](#cga12102-bib-0007){ref-type="ref"}). During our investigations on the relationship between sperm morphology and male fertility, we identified mutant strains showing high frequency of abnormal sperm morphology. Among these strains, B10.M and TEN1 were ranked first and second in the observed frequency of abnormal sperm morphology. Genetic analysis of B10.M revealed that the sperm phenotype was inherited (Gotoh [2010](#cga12102-bib-0004){ref-type="ref"}). Further quantitative trait loci (QTL) analysis and fine mapping identified that the sperm phenotype was controlled by two interactive recessive loci, *Shm1* on chromosome 1 and *Shm2* on chromosome 4 (Gotoh et al. [2012](#cga12102-bib-0005){ref-type="ref"}). The B10.M strain showed teratozoospermia, producing sperm cells displaying a wide variety of morphological abnormalities. Frequency of occurrence of each type of morphological abnormality varied among individuals. Collectively, our previous studies (Gotoh [2010](#cga12102-bib-0004){ref-type="ref"}; Gotoh et al. [2012](#cga12102-bib-0005){ref-type="ref"}) revealed that the total frequency of various types of sperm‐head morphological abnormalities was heritable. Severe male subfertility in the B10.M strain was evident. In the study reported here, we analyzed the TEN1 strain for sperm‐head morphological abnormalities and attempted to correlate the results with male fertility. The sperm phenotype of TEN1 was similar to that of B10.M. TEN1 strain displayed teratozoospermia. The types of sperm‐head abnormalities observed in TEN1 were identical to that observed in B10.M. The total frequency of the occurrence of abnormal sperm morphology was also found to be heritable. As shown in Figure [1](#cga12102-fig-0001){ref-type="fig"}, some sperm show abnormal shape not only in head but also in midpiece (Fig. [1](#cga12102-fig-0001){ref-type="fig"} E, H, and J). Some connection between head abnormalities and midpiece abnormalities was suspected. However, heritability of midpiece abnormalities has not been detected. The cause(s) of midpiece abnormalities is apparently different from the loci causing sperm‐head abnormalities. We tentatively set the threshold of high or low frequencies of sperm‐head abnormalities at and under 20%, respectively, for two reasons. First, because the maximum frequency of spermatozoa with sperm‐head abnormalities in individual mice within the F1 hybrid backcrossed with C3H group was below 20%. Second, because 98.9% of F2 animals produced from crosses between normal B6 and C3H strains showed sperm‐head abnormalities frequency of less than 20% (Gotoh and Aoyama [2012](#cga12102-bib-0006){ref-type="ref"}). The possibility of frequency of sperm‐head abnormalities of \>20% without mutation was considered to be rare. As per these criteria, segregation of the sperm phenotype in TEN1 was different from that of B10.M. In the backcross of F1 offspring to TEN1, 50% of animals were expected to express an abnormal phenotype if one recessive locus was involved. If two recessive loci were involved, the segregation rate was expected to be 25%. In this cross, segregation of high and low frequencies of the phenotype was located between the 50% and 25% lines (Fig. [2](#cga12102-fig-0002){ref-type="fig"}). In the F2 animals, the predicted segregation rate for one recessive locus was 25%, and that for two recessive loci, 6.25%. If a 20% frequency of abnormality is tentatively assumed to be the threshold, it located in the middle of 25% and 6.3% lines (Fig. [2](#cga12102-fig-0002){ref-type="fig"}). The results indicated that neither the one recessive locus model nor the two recessive loci model explained the segregation of the phenotype of TEN1 strain. At least three loci are expected to be involved in this case. QTL analysis and genetic mapping are needed to identify the loci controlling sperm‐head abnormality. The second difference was related to fertility. A complex relationship between sperm‐head abnormality and male fertility became apparent. The frequencies of sperm‐head abnormalities of B10.M and TEN1 were 61.5 ± 2.4% and 50.2 ± 3.7%, respectively (Table [1](#cga12102-tbl-0001){ref-type="table"}). Both scores were much higher than that for standard inbred strains including C3H (2.3 ± 0.5%) and B10 (4.8 ± 0.4%). B10.M males that showed high frequency of sperm head abnormality were diagnosed to be subfertile since they produced fewer live embryos (3.8 ± 2.9) on the 12th day after vaginal plug was observed when crossed with C3H females (Table [2](#cga12102-tbl-0002){ref-type="table"}). However, TEN1 males produced a normal number of live embryos (7.7 ± 2.7) irrespective of high frequency of sperm‐head abnormality observed in this strain. The significant difference in the embryo survival rates between B10.M males and TEN1 males could not be explained by 10% difference in sperm‐head abnormalities between the two strains. Qualitative difference between the spermatozoa of the two strains was not apparent from the observation by a light microscopy. Further studies are needed to elucidate the mechanisms underlying the differences in reproductive capacity of spermatozoa from the two strains. The origin of loci causing sperm‐head abnormality phenotype of TEN1 is not known. Although TEN1 is an *H2* congenic strain on the B10 background, its sperm phenotype was not linked to *H2* complex on chromosome 17. Either introgression of chromosomal fragment(s) containing the mutation from the *H2* donor animal, or occurrence of *de novo* mutation(s) during/after establishment of congenic strain could be the origin of alleles responsible for the sperm phenotype. A number of genes and loci important for male fertility have been identified (Handel et al. [2006](#cga12102-bib-0008){ref-type="ref"}; Matzuk and Lamb [2008](#cga12102-bib-0013){ref-type="ref"}). A single mutation in any one of over 400 genes or loci could result in male infertility or subfertility. The reproductive system is also regulated by a highly coordinated interaction of genes in a process known as epistasis. We found two interacting loci potentially responsible for the expression of subfertile phenotype of B10.M. Although not characterized, our findings suggest that at least three loci may be involved in determining the reproductive phenotype of TEN1. Our previous study examining sperm‐head morphology of F2 hybrid individuals generated by crossing two reproductively normal strains of mouse, C3H and C57BL/6J (B6), revealed that genetic factors other than mutations may contribute significantly toward determining the phenotype (Gotoh and Aoyama [2012](#cga12102-bib-0006){ref-type="ref"}). Reproductive function and sperm morphology of C3H and B6 strains were normal. However, approximately 20% of F2 animals showed higher levels of abnormalities than the parental strains, and approximately 1% of F2 animals showed very high (\>20%) levels of abnormal sperm‐head morphology and histological changes in testis. Therefore inter genetic interaction of "normal" alleles from the two parental strains could affect sperm‐head morphology. In other words, epistasis of normal alleles of a set of responsible genes may have caused the abnormal phenotype of sperm without the involvement of mutations. The involvement of epistasis in reproductive performance in mouse has been reported (Peripato et al. [2004](#cga12102-bib-0014){ref-type="ref"}; Flachs et al. [2012](#cga12102-bib-0003){ref-type="ref"}). Our findings further confirm the notion that the reproductive system is controlled by coordinated action of numerous genes. Histological observation found various degenerative lesions of seminiferous epithelium in the TEN1 testis. Normal epitheliums were also found. The degenerative changes were not observed through the entire tubule. In the seminiferous tubules, degree of degeneration varied widely from single cell necrosis to severe atrophy of the tubule. All necrotic cells observed were recognized clearly as germinal cells, and Sertoli cells were observed to be survived. A particular developmental stage of spermatogenesis has not been identified for the dead germ cells. Spaced observed in seminiferous epithelium were considered to be formed by the loss of syncytium of late spermatocytes and round spermatids. It seems that the degenerative changes take place occasionally or conditionally. In most of the testes of mice that were mutated, the genes specifically expressed in the spermatogenesis (Matzuk and Lamb [2008](#cga12102-bib-0013){ref-type="ref"}; Jamsai and O\'Bryan [2011](#cga12102-bib-0010){ref-type="ref"}; Massart et al. [2012](#cga12102-bib-0012){ref-type="ref"}) showed histological changes in the entire testis, resulting in male infertility. Histological degeneration of the TEN1 testis is mosaic (Fig. [3](#cga12102-fig-0003){ref-type="fig"}), and TEN1 males were fully fertile (Table [2](#cga12102-tbl-0002){ref-type="table"}). From the observation of TEN1 testis, it was considered that the degeneration was brought about by the failure of communication between germ cells and Sertoli cells. Although no morphological changes were detected in Sertoli cells even in the severely atrophied seminiferous tubules, some functional abnormality of the Sertoli cell may play a key role in the degeneration in the testis of TEN1. Identification of responsible genes is required to understand the reproductive phenotype of the TEN1. Sperm‐head abnormality in the B10.M has been shown to be controlled by two interacting loci, *Shm1* and *Shm2*. This strain also showed male subfertility. In contrast, the TEN1 strain was found to be fully fertile but showed high frequencies of sperm‐head morphological abnormality, predictably controlled by at least three loci. Identification of genes encoding loci controlling sperm‐head abnormality in both strains, and to investigate the function of the genes on male fertility will be needed to elucidate the issues of male reproduction. Disclosure {#cga12102-sec-0021} ========== Authors have nothing to disclose and no conflicts of interest. The authors acknowledge the assistance of Mr Masayuki Araki and Mr Heiichi Uchiyama with the experiments. We thank Dr Shiroishi for providing us B10.MOL‐TEN1 strain of mouse. This work was supported by the Ministry of Agriculture, Forestry, and Fisheries, Japan.

\section*{Acknowledgements} We would like to thank the members of the Dialogue Modelling Group of the University of Amsterdam for their useful discussions, and our anonymous reviewers for their insightful comments. This project has received funding from the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation programme (grant agreement No.\ 819455). \section{Possible Criteria to Distinguish Constructions} \label{sec:app-criteria} Lexicalised constructions can be classified according to multiple criteria \cite{titone1994descriptive,wray2002formulaic,columbus2013support}, including those listed below. \begin{itemize} \item \textbf{Compositionality} This criterion is typically used to separate idioms from other formulaic expressions, although it is sometimes referred to as \textit{transparency} to underline its graded, rather than binary, nature. There is no evidence, however, that the processing advantage of idioms differs from that of compositional phrases \cite{tabossi2009idioms,jolsvai2013meaning,carrol2020all}. \textit{Therefore we ignore this criterion in the current study.} \item \textbf{Literal plausibility} This criterion is typically used to discriminate among different types of idioms \cite{titone1994descriptive,titone2014time}---as compositional phrases are literally plausible by definition. \textit{Because we ignore distinctions made on the basis of compositionality, we do not use this criterion.} \item \textbf{Meaningfulness} Meaningful expressions are idioms and compositional phrases (e.g. \textit{`on my mind'}, \textit{`had a dream'}) whereas sentence fragments that break constituency boundaries (e.g., \textit{`of a heavy'}, \textit{'by the postal'}) are considered less meaningful \cite[as measured in norming studies, e.g., by][]{jolsvai2013meaning}. There is some evidence that the meaningfulness of multi-word expressions correlates with their processing advantage even more than their frequency \cite{jolsvai2013meaning}; yet expressions are particularly frequent, they present processing advantages even if they break regular phrasal structures \cite{bybee1999effect,tremblay2011processing}. Moreover, utterances that break regular constituency rules are particularly frequent in spoken dialogue data (e.g., \textit{`if you could search for job and that's not'}, \textit{`you don't wanna damage your relationship with'}). \textit{For these reasons, we do not exclude constructions that span multiple constituents from our analysis.} \item \textbf{Schematicity} This criterion distinguishes expressions where all the lexical elements are fixed from expressions ``with slots'' that can be filled by varying lexical elements.\textit{In this study, we focus on fully lexicalised constructions.} \item \textbf{Familiarity} This is a subjective criterion that strongly correlates with objective frequency \cite{carrol2020all}. Human experiments would be required to obtain familiarity norms for our target data, and the resulting norms would only be an approximation of the familiarity judgements of the true speakers we analyse the language of. \textit{Therefore, we ignore this criterion in the current study.} \item \textbf{Communicative function} Formulaic expressions can fulfil a variety of discourse and communicative functions. \citet{biber2004if}, e.g., distinguish between stance expressions (attitude, certainty with respect to a proposition), discourse organisers (connecting prior and forthcoming discourse), and referential expressions; and for each of these three primary discourse functions, more specific subcategories are defined. This type of classification is typically done a posteriori---i.e., after a manual analysis of the expressions retrieved from a corpus according to other criteria \cite{biber2007lexical}. In the BNC, for example, we find epistemic lexical bundles (\textit{`I don't know'}, \textit{`I don't think'}), desire bundles (\textit{`do you want to'}, \textit{'I don't want to'}), obligation/directive bundles (\textit{`you don't have to'}), and intention/prediction bundles (\textit{`I'm going to'}, \textit{`it's gonna be'}). \textit{We do not use this criterion to avoid an a priori selection of the constructions.} \end{itemize} \section{Extraction of Repeated Constructions} \label{sec:app-extraction} \begin{table*}[!h] \centering \resizebox{\textwidth}{!}{ \begin{tabular}{cccclccc} \toprule \textbf{Speaker} & \textbf{RI} & \textbf{RI~Utt} & \textbf{Dist} & \textbf{Turn} & $\boldsymbol{H(u)}$ & $\boldsymbol{H(c)}$ & $\boldsymbol{FE(c,u)}$ \\ \midrule A & 0 & 0 & - & [...] I think that everyone should have the same opportunities & & & \\ & & & & \ \ and \textbf{I don't think you should be} proud or ashamed of what & 4.24 & 1.90 & 1.21 \\ & & & & \ \ your you know what your situation is whether you what your & & & \\ & & & & \ \ what your race is whether you're a woman or a man whether & & & \\ & & & & \ \ you live from this pl whether you're in this place [...] & & & \\ \midrule A & 1 & 0 & 80 & I well I th I don't think it should \textbf{I don't think you should be} & 3.40 & 1.73 & 1.40 \\ \midrule A & 2 & 0 & 19 & Well yes perhaps but \textbf{I don't think you should be} like um & 3.95 & 1.06 & 2.25 \\ & & & & \ \ embarrassed about it or I think I think you should just sort of & & & \\ \bottomrule \end{tabular} } \caption{Repetition chain for the construction \textit{`I don't think you should be'} in dialogue S2AX of the Spoken BNC, annotated with repetition index (RI), repetition index in utterance (RI~Utt), and distance from previous mention (Dist; number of tokens). $H(u)$ is the utterance information content, $H(c)$ and $FE(c,u)$ are the construction's information content and facilitating effect.} \label{tab:chain-len7} \end{table*} We define a limited specific vocabulary of generic nouns that should not be considered referential. The vocabulary includes: \textit{bit}, \textit{bunch}, \textit{day}, \textit{days}, \textit{fact}, \textit{god}, \textit{idea}, \textit{ideas}, \textit{kind}, \textit{kinds}, \textit{loads}, \textit{lot}, \textit{lots}, \textit{middle}, \textit{ones}, \textit{part}, \textit{problem}, \textit{problems}, \textit{reason}, \textit{reasons}, \textit{rest}, \textit{side}, \textit{sort}, \textit{sorts}, \textit{stuff}, \textit{thanks}, \textit{thing}, \textit{things}, \textit{time}, \textit{times}, \textit{way}, \textit{ways}, \textit{week}, \textit{weeks}, \textit{year}, \textit{years}. We also find all the filled pauses and exclude word sequences that consist for more than 50\% of filled pauses. Filled pauses in the Spoken BNC are transcribed as: \textit{huh}, \textit{uh}, \textit{erm}, \textit{hm}, \textit{mm}, \textit{er}. Figure~\ref{fig:proportions} shows the proportion of tokens in an utterance belonging to constructions (referential and non-referential) and to non-construction sequences. Table~\ref{tab:chain-len7} shows a whole construction chain (from the first mention to the last repetition) for a construction of length 6. \begin{figure}[h] \centering \includegraphics[width=\linewidth]{figures/proportions.png} \caption{Proportion of tokens in an utterance that belong to referential constructions, non-referential constructions, and to non-construction sequences. The \textit{x} axis shows percentages indicating utterance positions in the dialogue relative to the dialogue length.} \label{fig:proportions} \end{figure} \section{Language Model} \subsection{Finetuning} \label{sec:app-finetuning} We finetune the \textit{`small' variant} of GPT-2 \cite{radford2019language} and DialoGPT \cite{zhang2019dialogpt} on our finetuning split of the Spoken BNC (see Section~\ref{sec:data}) using HuggingFace's implementation of the models with default tokenizers and parameters \cite{wolf2020transformers}. Dialogue turns are simply concatenated; we have experimented with labelling the dialogue turns (i.e., \textit{A: utterance 1, B: utterance 2} and found that this leads to higher perplexity. The finetuning results for both models are presented in Table~\ref{tab:finetuning}. We finetune the models and measure their perplexity using Huggingface's finetuning script. We use early stopping over 5 epochs.\footnote{The number of epochs (5) has been selected in preliminary experiments together with the learning rate (\num{1e-4}). In these experiments---which we ran for 40 epochs---we noticed that the \num{1e-4} learning rate offers the best tradeoff of training time and perplexity out of four possible values: \num{1e-2}, \num{1e-3}, \num{1e-4}, \num{1e-5}. We obtained insignificantly lower perplexity values with a learning rate of \num{1e-5}, with significantly longer training time: 20 epochs for GPT-2 and 28 epochs for DialoGPT.} Sequence length and batch size vary together because they together determine the amount of memory required; more expensive combinations (e.g., 256 tokens with batch size 16) require an exceedingly high amount of GPU memory. Reducing the maximum sequence length has limited impact: 99.90\% of dialogue turns have at most 128 words. DialoGPT starts from extremely high perplexity values but catches up quickly with finetuning. GPT-2 starts from much lower perplexity values and reaches virtually the same perplexity as DialoGPT after finetuning. For the pre-trained DialoGPT perplexity is extremely high, and the perplexity trend against maximum sequence length is surprisingly upward. These two behaviours indicate that the pre-trained DialoGPT is less accustomed than GPT-2 to the characteristics of our dialogue data. DialoGPT is trained on written online group conversations, while we use a corpus of transcribed spoken conversations between two speakers. In contrast, GPT-2 has been exposed to the genre of fiction, which contains scripted dialogues, and thus to a sufficiently similar language use. We select GPT-2 finetuned with a maximum sequence length of 128 and 512 as our best two models; these two models (which we now refer to as \textit{frozen}) are used for the adaptive learning rate selection (Section~\ref{sec:app-lr}). \begin{table*}[] \centering \resizebox{\textwidth}{!}{ \begin{tabular}{lcccccc} \toprule \textbf{Model} & \textbf{Learning rate} & \textbf{Max sequence length} & \textbf{Batch size} & \textbf{Best epoch} & \textbf{Perplexity finetuned} & \textbf{Perplexity pre-trained} \\ \midrule DialoGPT & 0.0001 & 128 & 16 & 3 & 23.21 & 7091.38 \\ DialoGPT & 0.0001 & 256 & 8 & 4 & 22.26 & 12886.92 \\ DialoGPT & 0.0001 & 512 & 4 & 4 & 21.73 & 21408.32 \\ GPT-2 & 0.0001 & 128 & 16 & 4 & 23.32 & 173.76 \\ GPT-2 & 0.0001 & 256 & 8 & 3 & 22.21 & 159.23 \\ GPT-2 & 0.0001 & 512 & 4 & 3 & 21.55 & 149.82 \\ \bottomrule \end{tabular} } \caption{Finetuning results for GPT-2 and DialoGPT on our finetuning split of the Spoken BNC.} \label{tab:finetuning} \end{table*} \subsection{Learning Rate Selection} \label{sec:app-lr} To find the appropriate learning rate for on-the-fly adaptation (see Section~\ref{sec:modelling}), we randomly select 18 dialogues $D$ from the analysis split of the Spoken BNC and run an 18-fold cross-validation for a set of six candidate learning rates: \num{1e-5}, \num{1e-4}, $\ldots$, 1. We finetune the model on each dialogue using one of these learning rate values, and compute perplexity change 1) on the dialogue itself (to measure \textit{adaptation}) as well as 2) on the remaining 17 dialogues (to measure \textit{generalisation}). We set the Transformer's context window to 50 to reproduce the experimental conditions presented in Section~\ref{sec:estimates}. More precisely, for each dialogue $d \in D$, we calculate the perplexity of our two frozen models (Section~\ref{sec:app-finetuning}) on $d$ and $D \setminus \{d\}$ (which we refer to as $ppl_{before}(d)$ and $ppl_{before}(D)$, respectively). Then, we finetune the models on $d$ using the six candidate learning rates, and measure again the perplexity over $d$ and $D \setminus \{d\}$ (respectively, $ppl_{after}(d)$ and $ppl_{after}(D)$). The change in performance is evaluated according to two metrics: $\frac{ppl_{after}(d) - ppl_{before}(d)}{ppl_{before}(d)}$ measures the degree to which the model has successfully adapted to the target dialogue; $\frac{ppl_{after}(D) - ppl_{before}(D)}{ppl_{before}(D)}$ measures whether finetuning on the target dialogue has caused any loss of generalisation. The learning rate selection results are presented in Figure~\ref{fig:lr}. We select \num{1e-3} as the best learning rate and pick the model finetuned with a maximum sequence length of 512 as our best model. The difference in perplexity reduction (both adaptation and generalisation) is minimal with respect to the model finetuned with a maximum sequence length of 128, but since the analysis split of the Spoken BNC contains turns longer than 128 tokens, we select the 512 version. Similarly to \citet{van2018neural}, we find that finetuning on a dialogue does not cause a loss in generalisation but instead helps the model generalise to other dialogues. Unlike \shortcite{van2018neural}, who used LSTM language models, we find that learning rates larger than $\num{1e-1}$ cause backpropagation to overshoot, even within a single dialogue. In Figure~\ref{fig:lr}, the bars for $\num{1e-1}$ and $1$ are not plotted because the corresponding data contains infinite perplexity values (due to numerical overflow). The selected learning rate, $\num{1e-3}$, is a relatively low learning rate for on-the-fly adaptation but it is still higher than the best learning rate for the entire dataset by a factor of 10. \begin{figure}[h] \centering \begin{subfigure}{.23\textwidth} \centering \includegraphics[width=\linewidth]{figures/gpt2-w128-bsz16-cv.pdf} \vspace{-2.5em} \caption{} \label{fig:lr-128} \end{subfigure}% \hspace{.01\textwidth} \begin{subfigure}{.23\textwidth} \centering \includegraphics[width=\linewidth]{figures/gpt2-w512-bsz4-cv.pdf} \vspace{-2.5em} \caption{} \label{fig:lr-512} \end{subfigure}% \caption{The adaptation and generalisation performance (defined in Section~\ref{sec:app-lr}) with varying learning rate.} \label{fig:lr} \end{figure} \section{Linear Mixed Effect Models} \label{sec:app-lme} As explained in §\ref{sec:stat-model} of the main paper, we fit a linear mixed effect model using facilitating effect as the response variable and including multilevel random effects grouped by dialogues and individual speakers.\footnote{We also try grouping observations only by dialogue and only by individual speakers. The amount of variance explained (but unaccounted for by the fixed effects) decreases, so we keep the two-level random effects.}. The fixed effects of the model, resulting from a backward stepwise selection procedure, are presented in §\ref{sec:stat-model}. Non-binary predictors are log-transformed, mean-centered, and scaled by 2~sd. The final model is summarised in Listing~\ref{lme-fe} and its coefficients are visualised in Figure~\ref{fig:fe-lme-summary}. We rely on the \texttt{lme4} and \texttt{lmerTest} R packages for this analysis. \begin{lstlisting}[label=lme-fe,float=*,frame=tb,caption=Linear mixed effect model for Facilitating Effect] MODEL INFO: Observations: 46399 Dependent Variable: Facilitating Effect Type: Mixed effects linear regression MODEL FIT: AIC = 99197.283, BIC = 99302.224 Pseudo-R^2 (fixed effects) = 0.084 Pseudo-R^2 (total) = 0.111 FIXED EFFECTS: ----------------------------------------------------------------------------------- Est. 2. --------------------------- ------- -------- -------- --------- ----------- ------- (Intercept) 0.704 0.683 0.725 65.527 185.698 0.000 log Utterance Position 0.046 0.026 0.066 4.556 9274.269 0.000 log Construction Length 0.098 0.084 0.111 14.396 46372.022 0.000 log Repetition Index 0.079 0.063 0.094 10.096 45082.205 0.000 log Distance -0.311 -0.328 -0.293 -34.571 46269.156 0.000 Previous Same Utterance -0.099 -0.184 -0.013 -2.262 46063.723 0.024 log Rep. Index in Utterance 0.178 0.130 0.226 7.243 45765.367 0.000 PMI -0.139 -0.154 -0.124 -18.225 45172.205 0.000 Referential 0.124 0.099 0.149 9.887 46214.616 0.000 ----------------------------------------------------------------------------------- p values calculated using Satterthwaite d.f. RANDOM EFFECTS: ------------------------------------------------ Group Parameter Std. Dev. ---------------------- ------------- ----------- Speaker:`Dialogue ID (Intercept) 0.082 Dialogue ID (Intercept) 0.090 Residual 0.701 ------------------------------------------------ Grouping variables: ----------------------------------------- Group # groups ICC ---------------------- ---------- ------- Speaker:`Dialogue ID 368 0.013 Dialogue ID 185 0.016 ----------------------------------------- Continuous predictors are mean-centered and scaled by 2 s.d. \end{lstlisting} \section{Further Results} \label{sec:app-further-results} \subsection{Same-Utterance Self-Repetitions} \label{sec:app-same-utterance} We investigate the interaction between cumulativity and recency (see §\ref{sec:results-mentions}) by focusing on densely clustered repetitions, produced by a speaker within a single utterance (the median distance between repetitions in the same utterance is 8 words; across turns it is 370.5 words). Table~\ref{tab:chain-len3} shows an example of same-utterance repetition. Repeating a construction when it has already been mentioned in the current utterance limits its facilitating effect ($\beta=-0.099, p<0.05$, \textit{95\% c.i.} -0.184:-0.013): if a portion of the utterance already consists of a construction, utterance information content will already be reduced, which in turn reduces the potential for the facilitating effect of repetitions. Nevertheless, we find \textbf{strong cumulativity effects for self-repetitions within the same utterance}: the repetition index \textit{within the current utterance} of a construction mention (i.e., how often the construction has been repeated so far in the utterance) has a positive effect on \textit{FE} ($\beta=0.178, p<0.005$, \textit{95\% c.i.} 0.130:0.226); see Figure~\ref{fig:fe-inturn}. In sum, same-utterance self-repetitions, especially those involving three or more mentions in a single utterance, can have a strong reduction effect on utterance information content. Although this may seem a simple yet very effective strategy for information rate mitigation, it is unlikely to be very effective in terms of the amount of information exchanged. Indeed, speakers do not use this strategy often in the Spoken BNC: 6.82\% of the total construction occurrences have at least one previous mention in the same utterance. \subsection{Interaction-Specificity} To distinguish interaction-specific constructions---those repeated particularly often in certain dialogues---from interaction-agnostic ones, we measure the association strength between a construction $c$ and a dialogue $d$ as the pointwise mutual information (PMI) between the two: \begin{align} \operatorname{PMI}(c,d) = \log_2 \frac{P(c|d)}{P(c)} \label{eq:pmi} \end{align} This quantifies how unusually frequent a construction is in a given dialogue, compared to the rest of the corpus. For example, for a construction to obtain a PMI score of 1, its probability given the dialogue $P(c|d)$ must be twice as high as its prior probability $P(c)$. Low PMI scores (especially below 1) characterise interaction-agnostic constructions, whereas higher PMI scores indicate that constructions are specific to a given dialogue. The probabilities in Eq.~\ref{eq:pmi} are obtained using maximum likelihood estimation over the analysis split of the Spoken BNC. PMI scores have a negative effect on \textit{FE} ($\beta=-0.139, p<0.005$, 95\% c.i. -0.154:-0.124), indicating that interaction-agnostic constructions have a stronger facilitating effect than interaction-specific ones. Figure~\ref{fig:fe-pmi} shows the \textit{FE} distributions for the most extreme cases: constructions with a PMI lower than 1 (`agnostic') and constructions that have been repeated in only one dialogue (`specific'). \begin{figure}[h] \centering \includegraphics[width=0.9\linewidth]{figures/lme-coefficients.png} \caption{Significant predictors of facilitating effect. Mixed effects linear regression, continuous predictors are mean-centred and scaled by 2 standard deviations.} \label{fig:fe-lme-summary} \end{figure} \begin{figure}[h] \centering \begin{subfigure}{.23\textwidth} \centering \includegraphics[width=\linewidth]{figures/fe-repindex-in-utt.png} \caption{} \label{fig:fe-inturn} \end{subfigure} \begin{subfigure}{.23\textwidth} \centering \includegraphics[width=\linewidth]{figures/fe-pmi.png} \caption{} \label{fig:fe-pmi} \end{subfigure} \caption{Facilitating effect against repetition index \textit{within the current utterance} (a) and facilitating effect of interaction-agnostic constructions ($\operatorname{PMI}(c,d)< 1$) vs.\ interaction-specific constructions ($\operatorname{PMI}(c,d) = \max_{c',d'} \operatorname{PMI}(c', d')$) (b).} \end{figure} \section{Computing Infrastructure and Budget} \label{sec:app-compute} Our experiments were carried out using a single GPU on a computer cluster with Debian Linux OS. The GPU nodes on the cluster are GPU GeForce 1001 1080Ti, 11GB GDDR5X, with NVIDIA driver version 418.56 and CUDA version 10.1. The total computational budget required to finetune the language model amounts to 45 minutes; obtaining surprisal estimates requires 4 hours, and selecting the adaptation learning rate requires 9 hours. \section{Background} \label{sec:background} \subsection{Constructions} \label{sec:background-constructions} This work focuses on \textit{constructions}, seen as particular configurations of structures and lexemes in usage-based accounts of natural language \cite{tomasello2003constructing,bybee2006usage,bybee2010language,goldberg2006constructions}. According to these accounts, models of language processing must consider not only individual lexical elements according to their syntactic roles but also more complex form-function units, which can break regular phrasal structures---e.g., \textit{`I~know I'}, \textit{`something out of'}. We further focus on fully lexicalised constructions (sometimes called \textit{formulaic expressions}, or \textit{multi-word expressions}). Commonly studied types of constructions are idioms (\textit{`break the ice'}), collocations (\textit{`pay attention to'}), phrasal verbs (\textit{`make up'}), and lexical bundles (\textit{`a lot of the'}). In §\ref{sec:extraction}, we explain how the notion of lexicalised construction is operationalised in the current study; Table~\ref{tab:construction-examples} shows some examples. \begin{table}[t] \centering \small \resizebox{\columnwidth}{!}{ \begin{tabular}{@{}p{0.151\textwidth} p{0.155\textwidth} p{0.134\textwidth}@{}} \toprule \textbf{SPXV} & \textbf{SAXQ} & \textbf{S9YG} \\ \midrule want to be with him & \textit{it on the television} & I bet you can \\ \textit{shit like that} & \textit{for a family} & yeah I used to \\ I can be & think that's a & \textit{go to bed} \\ to see her & \textit{the orient express} & and I love \\ and she just & one thing that & \textit{the window and} \\ I quite like & \textit{one of my favourites} & and I think it's \\ you don't like & \textit{on the television} & yeah I think so \\ and you're like & yes yeah I & \textit{the same people} \\ going to go & erm I think & is she in \\ you're going to & a really good & \textit{lock the door} \\ \bottomrule \end{tabular} } \caption{Top 10 constructions from three dialogues of the Spoken BNC \cite{love2017spoken}, sorted according to the PMI between a construction and its dialogue (§\ref{sec:stat-model}). Referential constructions in italics (§\ref{sec:extraction}). Headers correspond to the dialogues' IDs in the corpus.} \label{tab:construction-examples} \end{table} A common property of constructions is their frequent occurrence in natural language. As such, they possess what, in usage-based accounts, is sometimes referred to as `processing advantage' \cite{conklin2012processing,carrol2020all}. Evidence for the processing advantage of construction usage has been found in reading \cite{arnon2010more,tremblay2011processing}, naming latency \cite{bannard2008stored,janssen2012phrase}, eye-tracking \cite{underwood2004eyes,siyanova2011seeing}, and electrophysiology \cite{tremblay2010holistic,siyanova2017representation}. In this paper, we model this processing advantage as reduced information content and show that it can mitigate information rate throughout entire dialogues. \subsection{Information Content, Surprisal, and Processing Effort} \label{sec:background-surprisal} Estimates of information content have been shown to be good predictors of processing effort in perception \cite{jelinek1975design,clayards2008perception}, reading \cite{keller2004entropy,demberg2008data,levy2009eye}, and sentence interpretation \cite{levy2008noisy,gibson2013rational}. In these studies, information content is typically referred to as \textit{surprisal}, taken as a measure of how unpredictable, unlikely, or surprising a linguistic signal is in its context. As speakers take into consideration their addressee's processing effort \cite{clark1986referring,clark1989contributing}, their linguistic choices can often be explained as strategies to manage the fluctuations of information content over time. Surprisal-based accounts have indeed been successful at explaining various aspects of language production: speakers tend to reduce the duration of less surprising sounds \cite{aylett2004smooth,aylett2006language,bell2003effects,demberg2012syntactic}; they are more likely to drop sentential material within less surprising scenarios \cite{jaeger2007speakers,frank2008speaking,jaeger2010redundancy}; they tend to overlap at low-surprisal dialogue turn transitions \cite{dethlefs2016information}; and they produce sentences at a constant information rate in texts \cite{genzel2002entropy,qian2011topic,giulianelli2021analysing}. To measure information content we use GPT-2 \cite{radford2019language}, a neural language model. We thereby follow the established approach~\cite[e.g.,][]{genzel2002entropy,keller2004entropy,xu2018information} of using language models to estimate information content. Neural models' estimates in particular have been shown to be good predictors of processing effort, measured as reading time, gaze duration, and N400 response \cite{monsalve2012lexical,goodkind-2018-predictive,merkx2021human,schrimpf2021neural}. We further implement a simple neural adaptation mechanism, performing continuous gradient updates based on utterance prediction error; this not only leads to a more psychologically plausible model but also to the estimation of more human-like expectations \cite{van2018neural}.\looseness=-1 \section{Discussion and Conclusions} \label{sec:conclusion} Construction repetition is a pervasive phenomenon in dialogue; their frequent occurrence gives constructions a processing advantage~\cite{conklin2012processing}. In this paper, we show that the processing advantage of constructions can be naturally modelled as reduced information content and propose that speakers' production of constructions can be seen as a strategy for information rate mitigation. This strategy can explain why utterance information content is often found to decrease over the course of dialogues~\cite{vega2009looking,giulianelli2021analysing}, in contrast with the predictions of theories of optimal use of the communication channel~\cite{genzel2002entropy}. We observe that, as predicted, construction usage in English open-domain spoken dialogues mitigates the information rate of utterances. Furthermore, while constructions are produced at a stable rate throughout dialogues, their facilitating effect---our proposed measure of reduction in utterance information content---increases over time. We find that this increment is led by construction repetition, with facilitating effect being positively affected by repetition frequency, density, and by the contents of a construction. Repetitions of referential constructions reduce utterance information content more aggressively, arguably making them a more cost-reducing alternative to the shortening strategy observed in chains of referring expressions~\cite{krauss1964changes,krauss1967effect}, which instead tends to preserve rate constancy \cite{giulianelli2021dialogues}.\footnote{ Expression shortening is more efficient, however, in terms of articulatory cost. }\looseness-1 \paragraph{Relation to cognitive effort} We consider repetitions as a way for speakers to make dialogic interaction less cognitively demanding both on the production and on the comprehension side. This is not at odds with the idea that repetitions are driven by interpersonal synergies \cite{fusaroli2014dialog} and coordination \cite{sinclair-fernandez-2021-construction}. We think that the operationalisation of these higher level processes can be described by means of lower level, efficiency-oriented mechanisms, with synergy and coordination both corresponding to reduced collaborative effort. Although information content estimates from neural language models have been shown to correlate with human processing effort (cf.\ §\ref{sec:background-surprisal}), we cannot claim that our work directly models human cognitive processes as we lack the relevant human data to measure such correlation for the corpus at hand. \paragraph{Adaptive language model} Our decision to use an adaptive neural language model affects information content estimates in two main ways. On the one hand, due to their high frequency, constructions are likely to be assigned higher probabilities by this model, and therefore lower information content. We stress that we do not present constructions’ lower information content as a novel result, nor do we make any claims based on this result. As explained in §\ref{sec:preliminary-construction}, this is a precondition for our experiments on the facilitating effect of constructions, which is not determined exclusively by their information content (as empirically shown in §\ref{sec:ic-vs-fe}) but rather measures the effect of construction usage on the information content of entire utterances. On the other hand, because our model is adaptive, the probability of constructions is likely to increase as a result of their appearance in the dialogue history. Adaptation, however, also contributes to lower utterance information content \textit{overall} through the exploitation of topical and stylistic cues, as demonstrated by the lower perplexity of the adaptive model on the entire target dialogue as well as on other dialogues from the same dataset (see §\ref{sec:modelling} and Appendix~\ref{sec:app-lr}). In conclusion, while our adaptive language model assigns higher probabilities to frequently repeated tokens---as expected from a psychologically plausible model of utterance processing---it is not responsible for the discovered patterns of construction facilitating effect. In future work, the model can be improved, e.g., by conditioning on the linguistic experience of individual speakers. \paragraph{Types of dialogue} To consolidate our findings, construction repetition patterns should also be studied in dialogues of different genres and on datasets where utterance information content was not found to decrease. We have chosen the Spoken BNC for our study as it contains dialogues from a large variety of real-life contexts, which makes it a representative dataset of open-domain dialogue. In task-oriented dialogue, we expect constructions to consist of a more limited, task-specific vocabulary, resulting in longer chains of repetition and potentially more frequent referential construction usage. These peculiarities of task-oriented dialogue may influence the strength of the facilitating effect (as we have seen, facilitating effect is affected by both frequency and referentiality) but we expect our main results to still hold, as they are generally related to the processing advantage of constructions. \looseness-1 \paragraph{Relevance for dialogue generation models} Besides contributing new empirical evidence on construction usage in dialogue, our findings inform the development of more naturalistic utterance generation models. They suggest that models should be continually updated for their probabilities to better reflect human expectations; that attention mechanisms targeting contexts of different sizes (local vs.\ global) may have a significant impact on the naturalness of generated utterances; and that while anomalous repetitions (e.g., generation loops) should be prevented \cite{li2016deep,holtzman2019curious}, it is important to ensure that natural sounding repetitions are not suppressed. We expect dialogue systems that are able to produce human-like patterns of repetitions to be perceived as more natural overall---with users having the feeling that common ground is successfully maintained~\cite{PickeringGarrod2004}---and to lead to more effective communication~\cite{reitter2014alignment}. In our view, such human-like patterns can be reproduced by steering generation models towards the trends of information rate observed in humans. \section{Data} \label{sec:data} We conduct our study on the Spoken British National Corpus\footnote{\url{http://www.natcorp.ox.ac.uk}.}~\cite{love2017spoken}, a dataset of transcribed open-domain spoken dialogues containing 1,251 contemporary British English conversations, collected in a range of real-life contexts. We focus on the 622 dialogues that feature only two speakers, and randomly split them into a 70\% finetuning set (to be used as described in §\ref{sec:method}) and a 30\% analysis set (used in our experiments, as described in §\ref{sec:preliminary} and~§\ref{sec:results}). Table~\ref{tab:bnc-length} shows some statistics of the dialogues used in this study. \begin{table}[h] \centering \resizebox{\columnwidth}{!}{ \begin{tabular}{@{}lcccc@{}} \toprule & \textbf{Mean $\pm$ Sd} & \textbf{Median} & \textbf{Min} & \textbf{Max} \\ \midrule \textbf{Dialogue length (\# utterances)} & 736 $\pm$ 599 & 541.5 & 67 & 4859 \\[2pt] \textbf{Dialogue length (\# words)} & 7753 $\pm$ 5596 & 6102 & 819 & 39575 \\[2pt] \textbf{Utterance length (\# words)} & 11 $\pm$ 15 & 6 & 1 & 982 \\ \bottomrule \end{tabular} } \caption{ Two-speaker dialogue statistics, Spoken BNC.} \label{tab:bnc-length} \end{table} \subsection{Extracting Repeated Constructions} \label{sec:extraction} We define constructions as multi-word sequences repeated within a dialogue. To extract constructions from each dialogue, we use the sequential pattern mining method proposed by \citet{duplessis2017utterance,duplessis2017automatic,duplessis2021towards}, which treats the extraction task as an instance of the longest common subsequence problem \cite{hirschberg1977algorithms,bergroth2000survey}.\footnote{Their code is freely available at \url{https://github.com/GuillaumeDD/dialign}.} We modify it to not discard multiple repetitions of a construction that occur in the same utterance. We focus on constructions of at least three tokens, uttered at least three times in a dialogue by any of the dialogue participants. Repeated sequences that mostly appear as a sub-part of a larger construction are discarded.\footnote{We discard constructions that appear less than twice outside of a larger repeated construction in a given dialogue (e.g., \textit{`think of it'} vs.\ \textit{`think of it like'}).} We also exclude sequences containing punctuation marks or which consist of more than 50\% filled pauses (e.g., \textit{`mm'}, \textit{`erm'}).\footnote{The full list of filled pauses can be found in Appendix~\ref{sec:app-extraction}.} Applying the described extraction procedure to the 187 dialogues in the analysis split of the Spoken BNC yields a total of 5,893 unique constructions and 60,494 occurrences. Further statistics of the extracted constructions are presented in Table~\ref{tab:construction-stats}, and Table~\ref{tab:construction-examples} shows 10 example constructions extracted from three dialogues. For analysis purposes, we distinguish between referential and non-referential constructions. We label a construction as \textit{referential} if it includes nouns, unless the nouns are highly generic.\footnote{We define a limited specific vocabulary of generic nouns (e.g., \emph{`thing'}, \emph{`fact'}, \emph{'time'}); full vocabulary in Appendix~\ref{sec:app-extraction}.} Referential constructions are mostly topic-determined; examples are \emph{`playing table tennis'}, \emph{`a woolly jumper'}, \emph{`a room with a view'}. The remaining constructions are labelled as \textit{non-referential}. These mainly include topic-independent expressions and conversational markers, such \emph{`a lot of'}, \emph{`I don't know'}, and \emph{`yes of course'}. Our dataset consists of 5,291 referential and 55,203 non-referential construction occurrences, 1,143 and 4,750 construction forms; see Table~\ref{tab:construction-examples} for further examples. \begin{table}[h] \centering \resizebox{\columnwidth}{!}{ \begin{tabular}{@{}lccc@{}} \toprule & \textbf{Mean $\pm$ Sd} & \textbf{Median} & \textbf{Max} \\ \midrule \textbf{Construction Length} & 3.27 $\pm$ 0.58 & 3 & 7 \\[2pt] \textbf{Construction Frequency} & 4.29 $\pm$ 3.04 & 3 & 70 \\[2pt] \textbf{Constructions per Dialogue} & 325.34 $\pm$ 458.64 & 149 & 2817 \\[2pt] \ \ \ \ \small \textbf{\textit{Referential}} & \small 30.96 $\pm$ 39.75 & \small 19 & \small 346 \\[2pt] \ \ \ \ \small \textbf{\textit{Non-Referential}} & \small 296.88 $\pm$ 424.17 & \small 134.5 & \small 2530 \\[2pt] \textbf{Utterance Length} & 31.19 $\pm$ 36.19 & 21 & 959 \\ \bottomrule \end{tabular} } \caption{Construction statistics for our analysis split of the Spoken BNC. \textit{Constr.\ Frequency}: occurrences of a given construction in a dialogue. \textit{Constr.\ per Dialogue}: occurrences of all constructions in a dialogue. \textit{Utterance Length}: number of words in utterances containing a construction. The minimum is always 3 by design (§\ref{sec:extraction}). The difference between referential and non-referential is only significant for \textit{Constr.\ per Dialogue}.} \label{tab:construction-stats} \end{table} \section{Introduction} \label{sec:introduction} The repeated use of particular configurations of structures and lexemes, \textit{constructions}, is pervasive in conversational language use~\cite{tomasello2003constructing,goldberg2006constructions}. Such repetition can be understood as a surface level signal of processes of coordination~\cite{sinclair-fernandez-2021-construction} or `interpersonal synergy' between conversational partners~\cite{fusaroli2014dialog}. Speakers may use repetitions to successfully maintain common ground with their interlocutors~\cite{BrennanClark1996,PickeringGarrod2004}, because they are primed by their recent linguistic experience \cite{bock1986syntactic}, or to avoid a costly on-the-fly search for alternative phrasings~\cite[see, e.g.,][]{kuiper1995smooth}. At the same time, repetitions are also advantageous for comprehenders. Repeating a sequence of words positively reshapes expectations for those words, allowing comprehenders to process them more rapidly~\cite[for a review, see][]{bigand2005repetition}. As speakers are known to take into consideration both their own production cost and their addressee's processing effort~\cite{clark1986referring,clark1989contributing,frank2012predicting}, its two-sided processing advantage, as described above, makes construction repetition an efficient, cost-reducing communication strategy. In this paper, we investigate whether and how these information-theoretic properties of repetitions shape patterns of information rate in open-domain spoken dialogue. Information theory is the study of the conditions affecting the transmission and processing of information. To the foundations of the field belongs the noisy-channel coding theorem~\cite{shannon1948}, which states that for any given degree of noise in a communication channel, it is possible to communicate discrete signals nearly error-free up to a maximum information rate, the \textit{channel capacity}. If speakers use the communication channel optimally, they might send information at a rate that is always close to the channel capacity. This observation is at the basis of the principle of Entropy Rate Constancy~\cite[ERC;][]{genzel2002entropy}, which predicts that the information rate of speaker's utterances, measured as the utterance conditional entropy (i.e., its in-context \textit{Shannon information content} or \textit{information density}) remains constant throughout discourse. The ERC predictions have been empirically confirmed for written language production~\cite{genzel2002entropy,genzel2003variation,qian2011topic} but results on dialogue are mixed \cite{vega2009looking,doyle2015audience,doyle-frank-2015-shared,xu2018information,giulianelli2021dialogues}, with some studies suggesting a decreasing information rate over the course of dialogues~\cite{vega2009looking,giulianelli2021analysing}. We hypothesise that this decreasing trend in dialogue may be associated with construction repetition. We conjecture that speakers use construction repetition as a strategy for information rate mitigation, by padding the more information dense parts of their utterances with progressively less information dense constructions---leading to an overall decrease in information rate over the course of a dialogue. \looseness=-1 We extract occurrences of fully lexicalised constructions (see Table~\ref{tab:construction-examples} for examples) from a corpus of open-domain spoken dialogues and use a Transformer-based neural language model to estimate their contribution to utterance information content. First, we confirm that constructions indeed exhibit lower information content than other expressions and that information content further decreases when constructions are repeated. Then, we show that the decreasing trend of information content observed \textit{over utterances}---which contradicts the ERC principle---is driven by the increasing mitigating effect of construction repetition, measured as a construction's (increasingly) negative contribution to the information content of its containing utterance, what we call its \textit{facilitating effect}. In sum, our study provides new empirical evidence that dialogue partners use construction repetition as a strategy for information rate mitigation, which can explain why the rate of information transmission in dialogue, in contrast to the constancy predicted by the theory~\cite{genzel2002entropy}, is often found to decrease. Our findings inform the development of better dialogue models. They indicate, as suggested in related work \cite[e.g.,][]{xi2021taming}, that while avoiding \textit{degenerate} repetitions in utterance generation \cite{li2016deep,welleck2019neural} is an appropriate strategy, dialogue systems should not suppress \textit{human-like} patterns of repetition as these make automatic systems be perceived as more natural and more effective in conversational settings. \section{Experimental Setup} \label{sec:method} In this section, we define our information-theoretic measures and present the adaptive language model used to produce information content estimates.\footnote{Code and statistical analysis are available at \url{https://github.com/dmg-illc/uid-dialogue}.} \subsection{Information Content Measures} \label{sec:estimates} The \textit{information content} of a word choice $w_i$ is the negative logarithm of the corresponding word probability, conditioned on the utterance context~$u_{:w_i}$ (i.e., the words that precede $w_i$ in utterance~$u$) and on the local dialogue context $l$: \begin{align} H(w_i|u_{:w_i},l) = - \log_2 P(w_i | u_{:w_i},l) \label{eq:word-surprisal} \end{align} We define the local dialogue context~$l$ as the 50 tokens that precede the first word in the utterance.\footnote{Building on prior work \cite{reitter2006computational} that uses a window of 15 seconds of spoken dialogue as the locus of local repetition effects, we compute the average speech rate in the Spoken BNC (3.16 tokens/second) and multiply it by 15; we then round up the result (47.4) to 50 tokens.} We use tokens as a unit of context size, rather than utterances, since they more closely correspond to the temporal units used in previous work \cite[e.g.,][]{reitter2006computational}, and since the length of utterances can vary significantly (see Table~\ref{tab:bnc-length}). To measure the information content of a construction~$c$, we average over word-level information content values: \begin{align} H(c;u_{:c},l) = \frac{1}{|c|} \sum_{w_i \in c} H(w_i|u_{:c},l) \label{eq:construction-surprisal} \end{align} We use the same averaging strategy to compute the information content of entire utterances, following prior work~\cite[e.g.,][]{genzel2002entropy,xu2018information}: \begin{align} H(u;l) = \frac{1}{|u|} \sum_{w_i \in u} H(w_i|u_{:w_i},l) \label{eq:utterance-surprisal} \end{align} The above information content estimates target constructions and entire utterances but they do not qualify the relationship between the two. We also measure the information content change (increase or reduction in information rate) contributed by a construction $c$ to its containing utterance, which we call the \textit{facilitating effect} of a construction. Facilitating effect is defined as the logarithm of the ratio between the information content of a construction and that of its utterance context: \begin{align} {\it FE}(c;u,l) = \log_2 \frac{\frac{1}{|u|-|c|} \sum_{c \not\ni w_j \in u} H(w_j|u_{:w_i},l)}{\frac{1}{|c|} \sum_{w_i \in c} H(w_i|u_{:c},l)} \label{eq:facilitating-effect} \end{align} By definition, this quantity is positive when the construction has lower information content than its context, and negative when it has higher information content. When the utterance consists of a single construction, facilitating effect is set to 0. We can expect the values produced by our information content and facilitating effect measurements (Eq.\ \ref{eq:construction-surprisal} and \ref{eq:facilitating-effect}, respectively) to correlate: it is more likely for a construction to have a (positive) facilitating effect if its information content is low. When a construction's information content is high, the information content of its utterance context must be even greater for facilitating effect to occur. Nevertheless, perfect correlation does not follow a priori from the definition of the two measures; we will show this empirically in §\ref{sec:ic-vs-fe}. \subsection{Language Model} \label{sec:modelling} To estimate the per-word conditional probabilities that are necessary to compute information content (Eq.\ \ref{eq:word-surprisal}), we use an adaptive language model. The model is conditioned on local contextual cues via an attention mechanism \cite{vaswani2017attention} and it learns continually \cite[see, e.g.,][]{krause2018dynamic} from exposure to the global dialogue context. We use GPT-2 \cite{radford2019language}, a pre-trained autoregressive Transformer language model. We rely on HuggingFace's implementation of GPT-2 with default tokenizers and parameters \cite{wolf2020transformers} and finetune the pre-trained model on a 70\% training split of the Spoken BNC to adapt it to the idiosyncrasies of spoken dialogic data.\footnote{More details on finetuning can be found in Appendix~\ref{sec:app-finetuning}.} We refer to this finetuned version as the \textit{frozen} model. We use an attention window of length $|u_{:w_i}| + 50$, i.e., the sum of the utterance length up to word $w_i$ and the size of the local dialogue context. As a continual learning mechanism, we use back-propagation on the cross-entropy next word prediction error, a simple yet effective adaptation approach motivated in §\ref{sec:background-surprisal}. Following \citet{van2018neural}, when estimating information content for a dialogue, we begin by processing the first utterance using the frozen language model and then gradually update the model parameters after each turn. For these updates to have the desired effect, the learning rate should be appropriately tuned. It should be sufficiently high for the language model to adapt during a single dialogue, yet an excessively high learning rate can cause the language model to lose its ability to generalise across dialogues. To find the appropriate rate, we randomly select 18 dialogues from the analysis split of the Spoken BNC\footnote{This amounts to ca.~10\% of the analysis split. We use the analysis split because there is no risk of ``overfitting'' with respect to our main analyses.} and run an 18-fold cross-validation for a set of six candidate learning rates: \num{1e-5}, \num{1e-4}, $\ldots$,~1. We finetune the model on each dialogue using one of these learning rates and compute perplexity reduction (i)~on the dialogue itself (\textit{adaptation}) as well as (ii)~on the remaining 17 dialogues (\textit{generalisation}). We select the learning rate yielding the best adaptation over cross-validation folds (\num{1e-3}), while still improving the model's generalisation ability. See Appendix~\ref{sec:app-lr} for further details. \looseness=-1 \section{Preliminary Experiments} \label{sec:preliminary} In this section, we present preliminary experiments on the information content of utterances and constructions, which set the stage for our analysis of the facilitating effect of construction repetition. \subsection{Utterance Information Content} \label{sec:preliminary-utterance} \looseness=-1 Our experiments are motivated by the mixed results on the dynamics of information rate in dialogue discussed in §\ref{sec:introduction}. We thus begin by testing if the Entropy Rate Constancy (ERC) principle holds in the Spoken BNC, i.e., whether utterance information content remains stable over the course of a dialogue. Following a procedure established in prior work~\cite{xu2018information}, we fit a linear mixed effect model with the logarithm of utterance position and construction length as fixed effects (we will refer to their coefficients as $\beta$), and include multi-level random effects grouped by dialogue. \textit{For the ERC principle to hold, the position of an utterance within a dialogue should have no effect on its information content.} Instead, we find that utterance information content decreases significantly over time ($\beta\!=\!-0.119, p\!<\!0.005$, \textit{95\% c.i.}\ $-0.130\!:\!-0.108$), in line with previous negative results on open-domain and task-oriented dialogue~\cite{vega2009looking,giulianelli2021analysing}. The strongest drop occurs in the first ten dialogue utterances ($\beta\!=\!-0.886, p\!<\!0.005$, \textit{95\% c.i.}\ $-0.954\!:\!-0.818$) but the decrease is still significant for later utterances ($\beta\!=\!-0.043, p\!<\!0.005$, \textit{95\% c.i.}\ $-0.054\!:\!-0.032$). \subsection{Construction Information Content} \label{sec:preliminary-construction} Our hypothesis that construction repetition progressively reduces the information rate of utterances is motivated by the fact that constructions are known to have a processing advantage (see §\ref{sec:introduction} and §\ref{sec:background-constructions}). This property makes them an efficient production strategy, i.e., one that reduces the speaker's and addressee's collaborative effort. Before investigating if the hypothesised information rate mitigation strategy is at play, we test whether our information theoretic measures and the language model used to generate them are able to capture processing advantage: \textit{we expect our framework to yield lower information content estimates (Eq.\ \ref{eq:construction-surprisal}) for constructions than for other word sequences.} Indeed, the information content of constructions is significantly lower than that of non-construction sequences ($t\!=\!-168.82, p\!<\!0.005$, \textit{95\% c.i.}\ $-2.033\!:\!-1.987$).\footnote{We extract all 3- to 7-grams from our analysis split of the Spoken BNC, excluding all \textit{n}-grams that are equal to extracted constructions. We then sample, for each length $n$ from 3 to 7, $s_n$ non-construction sequence occurrences---where $s_n$ is the number of occurrences of $n$-tokens-long extracted constructions.\label{ftn:non-constructions}. The length distributions should match because length has an effect on \textit{S} and \textit{FE} (see §\ref{sec:construction-types}).} Constructions' information content is on average 2 bits lower than that of non-constructions. We conclude that our estimates of information content are a sensible model of the processing advantage of constructions. \subsection{Stable Rate of Construction Usage} \label{sec:stable-usage} In experiment §\ref{sec:preliminary-construction}, we confirmed that constructions have lower information content than other utterance material. A simple strategy to decrease utterance information content over dialogues (we do observe this decrease in the Spoken BNC, as described in §\ref{sec:preliminary-utterance}) could then simply be to increase the rate of construction usage. To test if this strategy is at play, we fit a linear mixed effect model with utterance position as the predictor and the proportion of construction tokens in an utterance as the response variable. Over the course of a dialogue, the increase in the proportion of an utterance's tokens which belong to a construction is negligible ($\beta\!=\!0.004, p\!<\!0.05$, \textit{95\% c.i.}\ $0.001\!:\!0.008$). Speakers produce constructions at a stable rate (see also Figure~\ref{fig:proportions} in Appendix~\ref{sec:app-extraction}), indicating that an alternative strategy for information rate reduction is at work.\looseness-1 \begin{table*} \centering \small \resizebox{\textwidth}{!}{ \begin{tabular}{cccclccc} \toprule \textbf{Speaker} & \textbf{RI} & \textbf{RI~Utt} & \textbf{Dist} & \textbf{Turn} & $\boldsymbol{H(u)}$ & $\boldsymbol{H(c)}$ & $\boldsymbol{FE(c;u)}$ \\ \midrule A & 0 & 0 & - & Drink? that was what he did yeah just just to just to know that & 5.99 & 4.73 & 0.40 \\ & & & & \ \ I he \textbf{might not be} a complete twat but just a fyi & & & \\ \midrule B & 1 & 0 & 1586 & Especially for my birthday mind you I \textbf{might not be} here for & 5.04 & 4.01 & 0.53 \\ & 2 & 1 & 14 & \ \ mine and I went what do you mean you \textbf{might not be} here? & & 2.70 & 0.90\\ \bottomrule \end{tabular} } \caption{ Repetition chain for the construction \textit{`might not be'} in dialogue SXWH of the Spoken BNC, annotated with repetition index (RI), RI in utterance (RI~Utt), and distance from previous mention (Dist; in tokens). $H(u)$ is the utterance information content, $H(c)$ and $FE(c;u)$ are the construction's information content and facilitating effect. } \label{tab:chain-len3} \end{table*} \subsection{Information Content vs.\ Facilitating Effect} \label{sec:ic-vs-fe} The facilitating effect \textit{FE} of a construction is a function of its information content and the information content of its containing utterance~(Eq.~\ref{eq:facilitating-effect}). To ensure that our estimates of \textit{FE} are not entirely determined by construction information content (cf.\ §\ref{sec:estimates}), we inspect the relation between the two measures empirically, by looking at the values they take in our dataset of constructions. We find that the Kendall's rank-correlation between \textit{FE} and information content is $-0.623$ ($p\!<\!0.005$): although this is a rather strong negative correlation, the fact that the score is not closer to $-1$ indicates that there are cases where the two values are both either high or low. We indeed find examples of constructions with high information content \textit{H} and high facilitating effect \textit{FE}: \vspace{0.4em} \\ \noindent \resizebox{\columnwidth}{!}{ \begin{tabular}{l@{\ }l@{}} A: & we'll level that right press p purchase and \\ B: & right \\ A: & \textbf{go back to} recommended \textit{\textbf{(}}$\boldsymbol{H\!=\!5.30 \ \ FE\!=\!1.65}$\textit{\textbf{)}} \ \ \ \vspace{0.8em} \\ \end{tabular} } as well cases where information content is low and facilitating effect is low or negative: \vspace{0.4em} \\ \noindent \resizebox{\columnwidth}{!}{ \begin{tabular}{l@{\ }l@{}} A: & right let's go and have a drink \\ B & yeah \\ A: & \textbf{let's go and have} a drink \textit{\textbf{(}}$\boldsymbol{H\!=\!2.10 \ \ FE\!=\!-2.21}$\textit{\textbf{)}} \vspace{0.8em} \\ \end{tabular} } These examples have been selected among occurrences with \textit{H}/\textit{FE} higher or lower than the mean \textit{H}/\textit{FE} $\pm$ sd, respectively $3.62\!\pm\!1.48$ and $0.62\!\pm\!0.73$. Further analysis shows that this is not only true for individual instances but for entire groups of constructions. In particular, although their information content is overall higher ($t\!=\!13.511, p\!<\!0.005$, \textit{95\% c.i.}\ $0.371\!:\!0.497$), referential constructions also have higher facilitating effect than non-referential ones ($t\!=\!3.115, p\!<\!0.005$, \textit{95\% c.i.}\ $0.016\!:\!0.072$). We conclude that the two measures capture different aspects of a construction's information rate profile, with facilitating effect being sensitive to both construction and utterance information content. \section{The Facilitating Effect of Construction Repetition} \label{sec:results} We now test whether constructions have a positive facilitating effect, i.e., whether they reduce the information content of their containing utterances. We present our main statistical model in §\ref{sec:stat-model}, describe the effects of \textit{FE} predictors specific to unique construction mentions in §\ref{sec:results-mentions}, and analyse differences between types of constructions in §\ref{sec:construction-types}.\looseness-1 \subsection{Method} \label{sec:stat-model} \looseness=-1 To understand what shapes a construction's facilitating effect, we collect several of motivated features that can be expected to be informative \textit{FE} predictors. We fit a linear mixed effect (LME) model using (i)~these features as fixed effects, (ii)~\textit{FE} as the response variable, (iii)~and multi-level random effects grouped by dialogue and individual speaker ID. The first predictor is \textit{utterance position}, i.e., the index of the utterance within the dialogue, which allows us to test if \textit{FE} increases over the course of a dialogue. We then include predictors that distinguish different types of repetition. Since we expect a construction mention to increase expectation for subsequent occurrences---thus reshaping their information content---we consider its \textit{repetition index}, i.e., how often the construction has been repeated so far in the dialogue. Expectation is also shaped by intervening material, so we additionally track \textit{distance}, the number of tokens separating a construction mention from the preceding one. As \textit{FE} is the interplay between a construction and its utterance context, it is important to know whether the utterance context contains other mentions of the construction. We use a binary indicator (\textit{previous same utterance}) to single out occurrences whose previous mention is in the same utterance; for these cases, we also count the number of same-utterance previous mentions (\textit{repetition index in utterance}). To explore whether \textit{FE} varies across types of expressions, we also include a binary feature indicating whether the construction is \textit{referential} or non-referential (§\ref{sec:extraction}). Finally, we keep track of \textit{construction length}, the number of tokens that constitutes a construction, and \textit{PMI}, the pointwise mutual information between a construction and its dialogue, which is essentially a measure of the construction's frequency in the current dialogue as a function of its overall frequency in the corpus, indicating the construction's degree of interaction-specificity.\footnote{ The probabilities for the PMI calculation are obtained using maximum likelihood estimation over our analysis split of the Spoken BNC.} To determine the fixed effects of the final model, we start with all the predictors listed above (the non-binary ones are log-transformed) and perform backward stepwise selection, iteratively removing the predictor with the lowest significance and keeping only those with $p\!<\!0.05$. All predictors make it into our final model, the one which best fits the data according to both the Akaike and the Bayesian Information Criterion. The full specification of the best model, with model fit statistics as well as fixed and random effect coefficients, are in Appendix~\ref{sec:app-lme}. The next two sections present our main findings; we report fixed effect coefficients~($\beta$), p-values~($p$), and 95\% confidence intervals~(\textit{c.i.}).\looseness-1 \subsection{Construction Mentions} \label{sec:results-mentions} Our first observation is that construction usage reduces \textit{utterance} information content. More precisely, we find that \textbf{facilitating effect is higher for constructions than for non-construction sequences} ($t\!=\!118.79, p\!<\!0.005$, \textit{95\% c.i.}\ $0.536\!:\!0.554$). Constructions have on average 62\% lower information content than their utterance context; the average percentage drops to 7\% for non-construction sequences.\footnote{ These are the same sampled non-construction sequences as in~§\ref{sec:preliminary-construction}. Their average \textit{FE} is $0.07 \pm 0.80$. } Figure~\ref{fig:fe-constr-nonconstr} shows the two distributions. We also observe a positive effect of utterance position on \textit{FE} ($\beta\!=\!0.046, p\!<\!0.005$, \textit{95\% c.i.}\ $0.026\!:\!0.06$); that is, \textbf{the facilitating effect of constructions increases over the course of dialogues}. While the proportion of construction tokens remains stable~(§\ref{sec:stable-usage}), their mitigating contribution to utterance information content increases throughout dialogues---perhaps since speakers are more likely to \textit{repeat} established constructions as the dialogue develops. We indeed find that \textbf{repeated constructions have stronger facilitating effect}: there is a significant difference between the \textit{FE} of first mentions and repetitions ($t\!=\!-38.904, p\!<\!0.005$, \textit{95\% c.i.}\ $-0.265\!:\!-0.239$), as shown in Figure~\ref{fig:fe-rep-first}. The information content of repetitions is on average 68\% lower than that of their utterance context; for first mentions, it is on average 42\% lower. Having observed that the mitigating contribution of constructions to utterance information content indeed increases with construction repetition, we now look at how the \textit{FE} of repetitions varies as a function of their distribution across time. On the one hand, we find that \textbf{facilitating effect is cumulative}: repeating a construction reduces utterance information content more strongly as more mentions of the construction accumulate in the dialogue (Figure~\ref{fig:fe-cumul}). The effect of repetition index (i.e., how often the construction has been repeated so far in the dialogue) is positive on \textit{FE} ($\beta\!=\!0.079, p\!<\!0.005$, \textit{95\% c.i.}\ $0.063\!:\!0.094$). On the other hand, the distance of a repetition from the previous mention has a negative effect on \textit{FE} ($\beta\!=\!-0.311, p\!<\!0.005$, \textit{95\% c.i.}\ $-0.328\!:\!-0.293$). That is, \textbf{facilitating effect decays as a function of the distance between subsequent mentions}. As shown in Figure~\ref{fig:fe-decay}, this is a fast decay effect: the most substantial drop occurs for low distance values. The large magnitude of this coefficient indicates that recency is an important factor for constructions to have a strong facilitating effect. Indeed, almost one third (31.8\%) of all repetitions produced by speakers are not more than 200 tokens apart from their previous mention. Further results showing strong cumulativity effects for self-repetitions within the same utterance can be found in Appendix~\ref{sec:app-same-utterance}. \subsection{Types of Construction} \label{sec:construction-types} In this section, we analyse factors shaping the facilitating effect of construction forms, rather than individual mentions. We focus on the length of a construction and on whether it is referential. Construction length has a positive effect on \textit{FE} ($\beta\!=\!0.098, p\!<\!0.005$, \textit{95\% c.i.}\ $0.087\!:\!0.119$): \textbf{longer constructions have stronger facilitating effect.} Table~\ref{tab:chain-len3} shows a full repetition chain for a construction of length 3; Table~\ref{tab:chain-len7} (Appendix~\ref{sec:app-extraction}) for one of length 6. Non-construction sequences display an opposite, weaker trend ($\beta\!=\!-0.019, p\!<\!0.05$, \textit{95\% c.i.}\ $-0.032\!:\!-0.005$), as measured with a linear model. A possible explanation for the positive trend of constructions is related to production cost. Longer constructions are more costly for the speaker, so for them to still be an efficient production choice, their facilitating effect must be higher. Finally, we observe that \textbf{referential constructions have a stronger facilitating effect than non-referential ones}. Our LME model yields a positive effect for referentiality on \textit{FE} ($\beta\!=\!0.124, p\!<\!0.005$, 95\% c.i $0.099:0.149$) and we find a significant difference between the \textit{FE} of the two types ($t\!=\!3.115, p\!<\!0.005$, \textit{95\% c.i.}\ $0.072\!:\!0.016$). Looking in more detail, first mentions of referential constructions have higher information content and lower \textit{FE} than first mentions of non-referential ones (\textit{H}: $t\!=\!15.435, p\!<\!0.005$, \textit{95\% c.i.}\ $1.115\!:\!0.864$; \textit{FE}: $t\!=\!-9.315, p\!<\!0.005$, \textit{95\% c.i.}\ $-0.246\!:\!-0.161$), perhaps since words in referential sequences tend to be less frequent and more context-dependent. However, when repeated, their information content drops more substantially, reproducing inverse frequency effects attested in humans for syntactic repetitions \cite{bock1986syntactic,scheepers2003syntactic}. As a result, their \textit{FE} exceeds that of non-referential constructions (\textit{FE}: $t\!=\!8.818, p\!<\!0.005$, \textit{95\% c.i.}\ $0.117\!:\!0.183$), with the information content of a repeated reference being 81\% lower than that of its utterance context. Overall, these findings indicate that although referential constructions are less frequent than non-referential ones (23.3\% vs.\ 76.7\%; see §\ref{sec:extraction}), their repetition is a particularly effective strategy of information rate mitigation. \looseness=-1

In the course of subsequent studies of the mechanism for the dependency of ISG translation on G3BP1, G3BP2 and CAPRIN1, the authors noted an error in the cloning of one of the two control constructs---ELF2-Fluc. The Fluc open reading frame (ORF) for the ELF2-luc was derived from a different plasmid than the open reading frame for the three other Fluc constructs (the GAPDH-Fluc control, and the IFITM2-Fluc and PKR-Fluc) (see Fig 4 in the original article). The authors had originally sequenced all the constructs but neglected to notice that the ELF2-Fluc ORF was from a different Fluc source, which was slightly different for the other three constructs. Since the authors were using these constructs to test translation efficiency, there was concern that even slight differences in the ORF could alter the results shown in Fig 4 and quantified in Table 1 of the original article. An ELF2-Fluc with the identical Fluc ORF as the other three constructs in Fig 4 was reconstructed, a new HuH-7 cell lines expressing this new ELF2-Fluc was created, an IFITM2-Fluc construct was created, and the experiment was repeated. New figures were made to explain these changes, which is included here as [S1 Fig](#ppat.1006295.s003){ref-type="supplementary-material"}. The reconstruction was carried out by Dr. Kuo-Chieh Liao. Although the results with the IFITM2 construct remained qualitatively as originally reported, the results with the ELF2 construct are closer to results with the original GAPDH control construct, diminishing the differences in translation efficiency between reporters driven by ISG UTRs and those driven by non-ISG UTRs. It is important to note that the result in the corrected [Fig 4](#ppat.1006295.g001){ref-type="fig"} does not contradict the previous results, but it makes the differences between controls and experimental less dramatic. The conclusions of this publication that remain valid are as follows: 1. G3BP1, G3BP2 and CAPRIN1 are novel mediators of the antiviral IFN response (Figure 1). Shown for antiviral response to DENV-2 and YFV(17D). 2. G3BP1, G3BP2 and CAPRIN1 are not required for accumulation of ISG mRNAs but are required for accumulation of ISG proteins (Figures 2 and S4). Shown for six ISGs (PKR, RIG-I, ISG15, IFITM2, STAT1, and MX1) vs no effect on two non-ISGs (GAPDH and ACTINB). 3. G3BP1, G3BP2 and CAPRIN1 are critical regulators of ISG mRNA translation (Figure 3). Data support that G12C are not required for bulk translation (measured by \[35S\] methionine incorporation and rRNA distribution in polysome gradients), but a subset of endogenous cellular mRNAs are sensitive to G12C knockdown (major shifts of polysome peak fractions for IFITM2 and PKR RNAs). 4. DENV-2 interferes with ISG mRNA translation (Figure 5). Data indicate that DENV-2 infection downregulates translation of ISG mRNAs (IFITM2 and PKR) but not that of two non-ISG mRNAs (ELF2 and GAPDH) and thus phenocopies what we observed with G12C knockdown. 5. G3BP1, G3BP2 and CAPRIN1 interact with DENV-2 noncoding sfRNA during infection (Figure 6) 6. DENV-2 sfRNA downregulates ISG mRNA translation (Figure 7 and S11). Data are consistent with results obtained in Figure 3 with G12C depletion and in Figure 5 with DENV-2 infection. 7. The G3BP1, G3BP2 and CAPRIN1-sfRNA interaction protects DENV-2 replicons from IFN-ß (Figure 8) To properly correct the original article, the authors have made substantial changes to the following sections: Results, Discussion, and Methods. All changes to the original article are detailed in a marked-up PDF included as [S1 File](#ppat.1006295.s001){ref-type="supplementary-material"}. A clean Word document version of the revised article is included as [S2 File](#ppat.1006295.s002){ref-type="supplementary-material"}. The corrected [Fig 4](#ppat.1006295.g001){ref-type="fig"} and [Table 1](#ppat.1006295.t001){ref-type="table"} are provided here. {#ppat.1006295.g001} 10.1371/journal.ppat.1006295.t001 ###### Summary of mRNA induction, protein induction and translation efficiency for each Fluc reporter construct. Results presented in [Fig 4](#ppat.1006295.g001){ref-type="fig"} were compiled and for each construct, the relative translation efficiency was calculated as the ratio of the average firefly luciferase activity induction (normalized to siGFP, untreated cells set as 1) over to the average firefly luciferase mRNA induction (normalized to siGFP, untreated cells set as 1). The fold difference between siGFP + IFN and siG12C +IFN is shown in bold in the fourth column. The p-values for the differences in mRNA and Fluc induction between siGFP + IFN-β and siG12C + IFN-β conditions are indicated: ns, non significant; \* p\<0.05; \*\* p\<0.01; \*\*\*p\<0.005. {#ppat.1006295.t001g} ----------------------------------------------------------------------------------------------- Construct Biological process siGFP + IFN-β siG12C\#1 + IFN-β --------------------------------- -------------------- -------------------- ------------------- GAPDH-Fluc mRNA induction 4.04 ± 0.49 2.63± 0.15 \* Fluc induction 3.19 ± 0.26 1.91 ± 0.30 \*\* Relative translation efficiency 0.78 0.73\ **-1.07** IFITM2-Fluc mRNA induction 43.79 ± 14.5 27.79± 7.58 (ns) Fluc induction 24.55 ± 4.53 6.12 ± 1.32 \*\*\* Relative translation efficiency 0.56 0.22\ **-2.54** PKR-Fluc mRNA induction 5.91 ± 1.06 3.85 ± 0.17 (ns) Fluc induction 5.01 ± 0.40 2.49 ± 0.32 \*\*\* Relative translation efficiency 0.85 0.65\ **-1.3** ----------------------------------------------------------------------------------------------- Supporting information {#sec001} ====================== ###### Marked-up PDF detailing changes. (PDF) ###### Click here for additional data file. ###### Clean Word version of revised article. (DOCX) ###### Click here for additional data file. ###### G3BP1, G3BP2 and CAPRIN1 depletion decreases both mRNA and protein level from reconstructed ELF2-Fluc construct but diminishes only Fluc activity from IFITM2-Fluc. G3BP1, G3BP2 and CAPRIN1 depletion decreases both mRNA and protein level from new ELF2-Fluc construct but only diminishes Fluc activity from IFITM2-Fluc. (ISRE)-driven firefly luciferase reporters under the control of ELF2 (corrected version) or, IFITM2 UTRs. HuH-7 cells stably transfected with the above constructs were treated with control siGFP or siG12C\#1 siRNAs, induced with 1000UI/ml of IFN-β for 10h and firefly luciferase mRNA and activity levels determined by quantitative real-time RT-PCR and normalized to HMBS and SDHA mRNA levels. Firefly luciferase protein levels were determined by measuring luciferase activity and normalized to total protein concentration (F-I). All results are expressed as fold induction from control, untreated cells. (Data of K C Liao and M A Garcia-Blanco) (TIF) ###### Click here for additional data file.

Introduction {#s0005} ============ Cancer-testis (CT) antigens are a class of genes that are predominately expressed in the adult testes and are overexpressed in several types of tumors [@bb0005]. Within the testes, CT antigen expression localizes to the testicular germ cells termed the spermatogonia [@bb0005]. Because of their restricted expression in spermatogonia and the presence of a blood-testis barrier, expression of CT antigens in cancer often induces a tumor-directed immune response [@bb0005]. Consequently, CT antigens were historically identified through immunologic techniques such as T-cell epitope cloning and serological expression analysis of cDNA expression libraries [@bb0005]. More recently, CT antigens have been identified through analysis of expressed sequence tags for genes exclusively expressed in testes and cancer [@bb0005]. These strategies have led to the classification of more than 200 CT antigens (CT antigen database, <http://www.cta.lncc.br>), with new CT antigens continuing to be discovered [@bb0010]. Because their classification mainly relies on tissue expression patterns, the function and immunogenic potential of the majority of CT antigens remain unknown [@bb0015]. One of the most intriguing features of CT antigens is their predominant localization to the X chromosome. In fact, almost half of all CT antigens are encoded by the X chromosome [@bb0005], and approximately 10% of all protein-coding genes on the X chromosome are CT antigens [@bb0020] ([Figure 1](#f0005){ref-type="fig"}, *bottom*). These genes, termed CT-X antigens, recently evolved in eutherian mammals through gene translocation and duplication events to produce large multi-gene families [@bb0005]. At least 38 classified CT-X antigens are clustered near the end of the long arm of the X chromosome (Xq27-Xq28; [Figure 1](#f0005){ref-type="fig"}, *top*). Whereas all currently established CT-X antigens are protein-coding genes, this region also contains several reported miRNAs. In particular, genomic duplication events resulted in the evolution of the primate-specific miRNA gene family at Xq27.3 containing miR-888, miR-890, miR-891a, miR-891b, miR-892a, and miR-892b ([Figure 1](#f0005){ref-type="fig"}) [@bb0025]. The miR-888 gene family has restricted expression in testes, and miR-888 has recently been reported to be overexpressed in cancer [@bb0030; @bb0035]. However, the function of miR-888 in cancer has not been sufficiently studied [@bb0040; @bb0045]. In an analysis of miRNA expression patterns in uterine endometrial cancer (EC), we previously identified miR-888 as highly overexpressed [@bb0035]. EC is the fourth most common cancer in women and the most common gynecological malignancy [@bb0050]. While patient outcomes have improved for most cancers over the past 10 years, survival for EC patients has alarmingly decreased [@bb0050; @bb0055]. One of the most potent tumor suppressors in the endometrium is the progesterone receptor (PR), which activates gene expression to induce differentiation, cell cycle arrest, and apoptosis [@bb0060; @bb0065; @bb0070]. In addition, PR expression is often lost in advanced endometrial tumors [@bb0075; @bb0080; @bb0085]. Therefore, characterization of the different mechanisms by which PR expression is lost in EC can potentially improve our understanding on how aggressive ECs develop. Our objective in this study was to determine whether miR-888 is a CT-X antigen and to understand its role in EC. MiR-888 is a primate-specific miRNA that evolved through gene translocation and duplication events on the X chromosome similar to other CT antigen genes. Here, we demonstrate that miR-888 expression is restricted to the testes and localizes to cells in the early stages of spermatogenesis. Using The Cancer Genome Atlas (TCGA) database, we found that miR-888 was most highly expressed in endometrial tumors with a significant association to high-grade tumors and increasing percent invasion. In addition, we describe a novel mechanism of PR inhibition in EC through miR-888. These data suggest that miR-888 functions in endometrial tumors to inhibit PR-mediated anti-proliferative signaling. We also suggest that miRNAs can potentially be classified as CT antigens, with miR-888 as the defining example. Materials and Methods {#s0010} ===================== Tissue Samples {#s0030} -------------- Endometrial tissues were obtained under informed written consent from patients undergoing hysterectomy at the University of Iowa Hospitals and Clinics. The protocol was approved by the University of Iowa Institutional Review Board (Protocol No. 200209010). A total of 44 endometrial samples was collected, which included 9 samples of benign endometrium (BE), 18 endometrioid adenocarcinomas (EAs), 9 serous adenocarcinomas (SAs), and 8 carcinosarcomas (CSs; [Table S1](#ec0005){ref-type="supplementary-material"}). Rhesus macaque testis samples were generously provided by Dr Jodi McBride of the Oregon National Primate Research Center (Beaverton, OR). In this study, animals were killed by sedation with ketamine followed by deep anesthesia with sodium pentobarbital and then exsanguination as previously reported [@bb0090]. Samples were collected under a protocol that was reviewed and approved by the Oregon Health & Science University Institutional Animal Care and Use Committee (Protocol No. IS00002803, "RNA interference therapy in a non-human primate model of Huntington\'s disease") as previously reported [@bb0090]. Endometrial tissues were obtained under informed written consent from patients undergoing hysterectomy at the University of Iowa Hospitals and Clinics. The protocol was approved by the University of Iowa Institutional Review Board (Protocol No. 200209010). A total of 44 endometrial samples was collected, which included 9 samples of benign endometrium (BE), 18 endometrioid adenocarcinomas (EAs), 9 serous adenocarcinomas (SAs), and 8 carcinosarcomas (CSs; Table S1). Rhesus macaque testis samples were generously provided by Dr Jodi McBride of the Oregon National Primate Research Center (Beaverton, OR). In this study, animals were killed by sedation with ketamine followed by deep anesthesia with sodium pentobarbital and then exsanguination as previously reported [@bb0090]. Samples were collected under a protocol that was reviewed and approved by the Oregon Health & Science University Institutional Animal Care and Use Committee (Protocol No. IS00002803, "RNA interference therapy in a non-human primate model of Huntington\'s disease") as previously reported [@bb0090]. Cell Lines {#s0035} ---------- Six EC cell lines (AN3CA, RL95-2, Hec1A, SK-UT-1B, ECC-1, and KLE), five breast cancer cell lines (MCF7, MDA-MB-231, MDA-MB-453, SKBR3, and T-47D), nine ovarian cancer cell lines (ES-2, Caov-3, Caov-4, OV-90, OVCAR-3, SK-OV-3, SW 626, TOV-112D, and UWB1.298), and three prostate cancer cell lines (DU-145, LNCaP and PC-3) were obtained from the American Type Culture Collection (ATCC, Manassas, VA) and were cultured according to their guidelines. Two EC cell lines, Ishikawa H [@bb0095] and the Hec50co subline [@bb0100] of the Hec50 cell line [@bb0105], were gifts from Dr Erlio Gurpide of New York University and were grown in Dulbecco\'s modified Eagle\'s medium supplemented with 10% FBS and 1% penicillin-streptomycin (Life Technologies, Carlsbad, CA) as we previously reported [@bb0100]. The Alva-31 cell line [@bb0110] was a gift from Dr Thomas Griffith of the University of Iowa and was grown in RPMI 1640 supplemented with 10% FBS and 1% penicillin-streptomycin (Life Technologies). It has previously been reported that the ECC-1 EC cell line originated from the 3-H-12 Ishikawa EC cell subline [@bb0115]. Short tandem repeat genotyping was performed on all of our cell lines and confirmed that our Ishikawa cells are from the 3-H-4 subline and that our ECC-1 cells are from the 3-H-12 subline. While these cell lines share a common ancestor, they have distinct genomic profiles and are used in this manuscript to demonstrate data replication within different cellular backgrounds. RNA Samples {#s0040} ----------- A panel of RNA from 20 different human tissues was obtained from Life Technologies, with each sample containing pooled RNA from three individuals. To this panel, we added our own RNA sample of pooled BE from three patients. TCGA Data Analysis {#s0045} ------------------ Clinical and miR-888 expression data for 450 uterine corpus EC (UCEC) specimens were downloaded from TCGA data portal (National Cancer Institute and National Human Genome Research Institute, accessed 31 May 2013), and clinical characteristics are summarized in [Table S2](#ec0010){ref-type="supplementary-material"}. A full list of individual miR-888 expression values in reads per kilobase per million (RPKM) and clinical information by patient barcode are available in [Table S3](#ec0015){ref-type="supplementary-material"}. MiR-888 expression data were also obtained for a panel of cancers using the University of Iowa Institute for Clinical and Translation Science Compass website (<https://research.icts.uiowa.edu/compass/>, accessed 28 February 2014). MiR-888 expression data were downloaded for breast invasive carcinoma (BRCA, *n* = 380), colon adenocarcinoma (COAD, *n* = 394), lung adenocarcinoma (LUAD, *n* = 407), ovarian serous cystadenocarcinoma (OV, *n* = 476), pancreatic adenocarcinoma (PAAD, *n* = 43), prostate adenocarcinoma (PRAD, *n* = 187), UCEC (*n* = 363), and uterine CS (UCS, *n* = 56), and individual miR-888 expression values in RPKM for each cancer are provided in [Table S4](#ec0020){ref-type="supplementary-material"}. Clinical and miR-888 expression data for 450 uterine corpus EC (UCEC) specimens were downloaded from TCGA data portal (National Cancer Institute and National Human Genome Research Institute, accessed 31 May 2013), and clinical characteristics are summarized in Table S2. A full list of individual miR-888 expression values in reads per kilobase per million (RPKM) and clinical information by patient barcode are available in Table S3. MiR-888 expression data were also obtained for a panel of cancers using the University of Iowa Institute for Clinical and Translation Science Compass website (<https://research.icts.uiowa.edu/compass/>, accessed 28 February 2014). MiR-888 expression data were downloaded for breast invasive carcinoma (BRCA, *n* = 380), colon adenocarcinoma (COAD, *n* = 394), lung adenocarcinoma (LUAD, *n* = 407), ovarian serous cystadenocarcinoma (OV, *n* = 476), pancreatic adenocarcinoma (PAAD, *n* = 43), prostate adenocarcinoma (PRAD, *n* = 187), UCEC (*n* = 363), and uterine CS (UCS, *n* = 56), and individual miR-888 expression values in RPKM for each cancer are provided in Table S4. RNA and Protein Extraction from Cells and Tissues {#s0050} ------------------------------------------------- Tissue samples were immersed in RNAlater ICE (Life Technologies) and incubated at − 20°C for at least 16 hours before RNA extraction. The mirVana miRNA Isolation Kit (Life Technologies) was used to isolate total RNA from tissues according to the manufacturer's instructions. For cultured cells, RNA and protein were isolated using the mirVana PARIS kit (Life Technologies). Quantitative Reverse Transcription--Polymerase Chain Reaction {#s0055} ------------------------------------------------------------- Taqman primer/probe sets for *hsa*-miR-888 and RNU48 (Life Technologies) were used to perform quantitative reverse transcription--polymerase chain reaction (RT-PCR) on 200 ng of RNA using the Taqman Reverse Transcription Kit followed by the 2 × PCR Master Mix with No AmpERASE UNG according to the manufacturer's instructions (Life Technologies). For PR quantification, primers for PR (5′-ATGTGGCAGATCCCACAGGAGTTT-3′, 5′-ACTGGGTTTGACTTCGTAGCCCTT-3′) and 18S (5′-AACTTTCGATGGTAGTCGCCG-3′, 5′-CCTTGGATGTGGTAGCCGTTT-3′) were designed using the PrimerQuest software from Integrated DNA Technologies (Coralville, IA) and oligonucleotides were ordered from Integrated DNA Technologies. RT was carried out using the SuperScript III First Strand Synthesis System (Life Technologies) and PCR was performed using the 2 × Power SYBR Green PCR Master Mix (Life Technologies) following the manufacturer's protocol. Quantitative PCR reactions were completed on the Applied Biosystems 7900HT Fast Real-Time PCR System. Fold change was calculated using the Comparative Ct Method (2^− ΔΔCt^) with RNU48 as the internal control for miR-888 and 18S as the internal control for PR. For the normal tissues ([Figure 2](#f0010){ref-type="fig"}*A*) and cell lines ([Figure S1](#f0035){ref-type="fig"}), data are reported as 2^− ΔCt^ × 10,000 because there was no intuitive negative control group to use for fold change comparisons. Taqman primer/probe sets for *hsa*-miR-888 and RNU48 (Life Technologies) were used to perform quantitative reverse transcription--polymerase chain reaction (RT-PCR) on 200 ng of RNA using the Taqman Reverse Transcription Kit followed by the 2 × PCR Master Mix with No AmpERASE UNG according to the manufacturer's instructions (Life Technologies). For PR quantification, primers for PR (5′-ATGTGGCAGATCCCACAGGAGTTT-3′, 5′-ACTGGGTTTGACTTCGTAGCCCTT-3′) and 18S (5′-AACTTTCGATGGTAGTCGCCG-3′, 5′-CCTTGGATGTGGTAGCCGTTT-3′) were designed using the PrimerQuest software from Integrated DNA Technologies (Coralville, IA) and oligonucleotides were ordered from Integrated DNA Technologies. RT was carried out using the SuperScript III First Strand Synthesis System (Life Technologies) and PCR was performed using the 2 × Power SYBR Green PCR Master Mix (Life Technologies) following the manufacturer's protocol. Quantitative PCR reactions were completed on the Applied Biosystems 7900HT Fast Real-Time PCR System. Fold change was calculated using the Comparative Ct Method (2^− ΔΔCt^) with RNU48 as the internal control for miR-888 and 18S as the internal control for PR. For the normal tissues ([Figure 2](#f0010){ref-type="fig"}*A*) and cell lines (Figure S1), data are reported as 2^− ΔCt^ × 10,000 because there was no intuitive negative control group to use for fold change comparisons. miRNA *In Situ* Hybridization {#s0060} ----------------------------- *In situ* hybridization was performed using double-digoxigenin (DIG) labeled 2′O-methyl locked nucleic acid (LNA)-ZEN probes (Integrated DNA Technologies) complimentary to miR-888 (5′-DIG-U/ZEN/GA[C]{.ul}UGA[C]{.ul}AGCU[T]{.ul}UUU[G]{.ul}AGU/ZEN/A-DIG-3′) along with a scrambled negative control probe (5′-DIG-C/ZEN/GUA[U]{.ul}UAUA[G]{.ul}CCG[A]{.ul}UUAA[C]{.ul}G/ZEN/A-DIG-3′), where LNA modifications are underlined. Hematoxylin and eosin (H&E) staining was performed using the Sakura Finetek DRS 601 automatic slide stainer (Sakura Finetek USA Inc, Torrance, CA). All *in situ* hybridization and H&E stains were performed on adjacent serial sections of the same tissue sample. *In situ* hybridization was performed following the protocol previously described by McLoughlin et al. with the following modifications [@bb0120]. Tissue sections were incubated with ZEN-LNA probes overnight at a hybridization temperature of 55°C and a final concentration of 1 μM. All samples were incubated with an alkaline phosphatase-conjugated antibody to DIG (Roche, San Francisco, CA) at a dilution of 1:1000 overnight at 4°C. All images were taken using the same microscope settings and differential interference contrast (Leica Microsystems, Wetzlar, Germany). Each image background was then edited to white using Adobe Photoshop (San Jose, CA). miRNA *In Situ* Hybridization Intensity Scoring {#s0065} ----------------------------------------------- Each seminiferous tubule was divided into three zones: A, B, and C. Zone A represents the outer edge of the tubule in which the Sertoli cells, spermatogonia, and primary spermatocytes are located, zone B represents the region that contained spermatids and zone C represents the lumen where mature sperm are located. For intensity scoring, 0 = no blue staining, 1 = scant blue staining, 2 = moderate blue staining and 3 = abundant blue staining. Scoring was blinded and performed by pathologist Katherine Gibson-Corey, PhD, of the University of Iowa. Each image contained one seminiferous tubule. Scoring was performed for miR-888 staining in three rhesus macaque testis samples with five images from each of three tissue sections. MiR-888 Transfections {#s0070} --------------------- Cells were plated into phenol-red free medium with 10% charcoal-stripped FBS (Life Technologies) to reduce the presence of hormones. ECC-1 or Ishikawa cells were transfected with a plasmid encoding miR-888 (Origene, Rockville, MD) or the empty vector (EV) control (Origene) using Lipofectamine 2000 (Life Technologies) according to the manufacturer's instructions. At 4 hours post-transfection, cells were treated with 5 nM estradiol (Sigma, St. Louis, MO) to induce PR expression. Cells were collected at 24 hours post-transfection for RNA and protein isolation and analysis. Western Blot {#s0075} ------------ A rabbit monoclonal antibody against PR isoform A/PR isoform B (PRA/PRB, \#3153; Cell Signaling Technology, Danvers, MA) was combined with a rabbit monoclonal antibody against PRB (\#3157; Cell Signaling Technology) and both were diluted 1:1000. The mouse monoclonal β-actin antibody (Sigma \#A1978) was used as a loading control for all Western blots at a dilution of 1:10,000. The National Institutes of Health ImageJ program was used to perform densitometry quantification and data were normalized to the mock-transfected controls. Molecular Cloning {#s0080} ----------------- An approximately 500-bp region surrounding each of the four miR-888 binding sites in the PR 3′UTR was amplified from ECC-1 genomic DNA using the Platinum PCR SuperMix, High Fidelity (Life Technologies; primer sequences in [Table S5](#ec0025){ref-type="supplementary-material"}). PCR products were purified using the PCR Purification Kit (Qiagen, Valencia, CA) and subcloned into the pGEM-T Easy vector (Promega, Madison, WI). Insert sequences were subsequently cloned into the psiCHECK2 vector and transformed into TOP10 OneShot competent *E. coli* (Life Technologies). An approximately 500-bp region surrounding each of the four miR-888 binding sites in the PR 3′UTR was amplified from ECC-1 genomic DNA using the Platinum PCR SuperMix, High Fidelity (Life Technologies; primer sequences in Table S5). PCR products were purified using the PCR Purification Kit (Qiagen, Valencia, CA) and subcloned into the pGEM-T Easy vector (Promega, Madison, WI). Insert sequences were subsequently cloned into the psiCHECK2 vector and transformed into TOP10 OneShot competent *E. coli* (Life Technologies). Dual Luciferase Assays {#s0085} ---------------------- ECC-1 or Ishikawa cells were transfected with miR-888 or the EV in combination with each individual psiCHECK2 luciferase vector. Cells were collected at 24 hours post-transfection, and *Firefly* and *Renilla* luciferase activities were measured using the Dual Luciferase Assay System (Promega) according to the manufacturer's protocol. *Renilla* luciferase activity was divided by *Firefly* luciferase activity and normalized to the EV control. Statistical Analysis {#s0090} -------------------- For the *in situ* hybridization semiquantification, a repeated measures analysis of variance (ANOVA) model was used to analyze the data, because multiple images of each sample were scored. Due to the highly skewed nature of miR-888 expression in TCGA samples, figures are expressed on the square-root scale and observations below the limit of detection are jittered slightly for display purposes. For testing miR-888 expression levels across cancer types from TCGA data, Wilcoxon rank sum tests were used. Penalized regression splines were used for the nonparametric regression model presented in [Figure 3](#f0015){ref-type="fig"}*B* [@bb0125]. PCR data involving miR-888 and PR expression, Western blot data, and luciferase assays were all analyzed using ANOVA models. Linear regression was used to assess the significance of the correlation between miR-888 and PR mRNA expression in EC tissue. All miR-888 and luciferase transfections were performed in triplicate on three separate occasions to produce nine individual replicates per experimental condition for statistical analysis. Results {#s0015} ======= MiR-888 is Selectively Expressed in the Testes During the Early Stages of Spermatogenesis {#s0095} ----------------------------------------------------------------------------------------- Previously reported deep sequencing data have shown that miR-888 expression is restricted to the testes [@bb0030]. To confirm these analyses, we obtained pooled RNA samples from 21 different normal adult human tissues and quantified miR-888 expression through quantitative RT-PCR ([Figure 2](#f0010){ref-type="fig"}*A*). As expected, miR-888 demonstrated exceptionally high expression in the testes and minimal or no detectable expression in all other tissues investigated ([Figure 2](#f0010){ref-type="fig"}*A*). The majority of tissues had very low miR-888 expression \[raw cycle threshold (Ct) value of 30--40\]. The kidney, prostate, and thymus were the only other tissues that had Ct values less than 30 (Ct ≈ 28--30), but their expression was minimal compared to that observed in the testes (raw Ct = 22). To determine the distribution of miR-888 expression within the testes, we performed miRNA *in situ* hybridizations on rhesus macaque testis tissue from three individual monkeys. H&E staining was performed on a serial section for tissue structure comparison ([Figure 2](#f0010){ref-type="fig"}*B*). The *in situ* hybridizations were performed using a probe complimentary to miR-888 ([Figure 2](#f0010){ref-type="fig"}*C*) or a scrambled probe ([Figure 2](#f0010){ref-type="fig"}*D*) as a negative control. miR-888 expression consistently localized to the outer edge of the seminiferous tubule ([Figure 2](#f0010){ref-type="fig"}*C*). Spermatogenesis occurs within the seminiferous tubules of the testes. As the germ cells mature, they move towards the center of the tubule. To assess miR-888 expression during sperm maturation, the seminiferous tubule was divided into three different zones and scored for miR-888 intensity by a trained pathologist blinded to sample identity ([Figure 2](#f0010){ref-type="fig"}*E*). Zone A encompasses the outer edge of the seminiferous tubule where spermatogonia, primary spermatocytes, and Sertoli cells are located. Zone B represents the region that typically contains the maturing spermatids, and zone C represents the lumen of the testes where mature spermatozoa are located. Using a staining intensity range of 0--3, zone A consistently showed the highest level of staining ([Figure 2](#f0010){ref-type="fig"}*F*). Thus, miR-888 expression predominantly localized to the early stages of spermatogenesis within the seminiferous tubule of the testes. MiR-888 Is Selectively Overexpressed in Endometrial Tumors {#s0100} ---------------------------------------------------------- To investigate miR-888 expression in cancer, we used TCGA database and obtained miR-888 expression for a panel of carcinomas that included BRCA (*n* = 380), COAD (*n* = 394), LUAD (*n* = 407), OV (*n* = 476), PAAD (*n* = 43), PRAD (*n* = 187), UCEC (*n* = 363), and UCS (*n* = 56; [Table S4](#ec0020){ref-type="supplementary-material"}). MiR-888 expression was significantly elevated in UCEC and UCS compared to all other cancers (*P* \< .05 for all pairwise comparisons involving UCEC and UCS; the two types were not, however, significantly different from each other). Detectable expression of miR-888 was only observed in a minority of specimens from other cancers ([Figure 3](#f0015){ref-type="fig"}*A*). We also investigated miR-888 expression across a panel of cancer cell lines originating from breast, endometrial, ovarian, and prostate tumors and found that miR-888 levels were highest in EC cell lines ([Figure S1](#f0035){ref-type="fig"}), albeit low in comparison to testes ([Figure 2](#f0010){ref-type="fig"}*A*) and individual miR-888 expressing endometrial tumors ([Figure S2](#f0040){ref-type="fig"}). The cell line data support TCGA results in that miR-888 expression was highest in EC. MiR-888 expression in endometrial tumors from TCGA data was heavily skewed, with 65% of samples having undetectable expression of miR-888 ([Figure 3](#f0015){ref-type="fig"}*A* and [Table S2](#ec0010){ref-type="supplementary-material"}). Such a specific expression pattern is characteristic of CT antigens, which are not endogenously expressed in the tumor tissue of origin [@bb0015]. UCEC tumors that had positive expression of miR-888 were substantially more likely to be grade 3 tumors than UCEC tumors with undetectable miR-888 levels ([Table 1](#t0005){ref-type="table"}, *P* = .001). Furthermore, miR-888 expression in UCEC tumors showed a positive association with percent invasion of the tumor ([Figure 3](#f0015){ref-type="fig"}*B*). Therefore, miR-888 expression in endometrial tumors correlates with a more aggressive tumor phenotype and has an expression pattern in cancer similar to other representative CT antigens. To investigate miR-888 expression in cancer, we used TCGA database and obtained miR-888 expression for a panel of carcinomas that included BRCA (*n* = 380), COAD (*n* = 394), LUAD (*n* = 407), OV (*n* = 476), PAAD (*n* = 43), PRAD (*n* = 187), UCEC (*n* = 363), and UCS (*n* = 56; Table S4). MiR-888 expression was significantly elevated in UCEC and UCS compared to all other cancers (*P* \< .05 for all pairwise comparisons involving UCEC and UCS; the two types were not, however, significantly different from each other). Detectable expression of miR-888 was only observed in a minority of specimens from other cancers ([Figure 3](#f0015){ref-type="fig"}*A*). We also investigated miR-888 expression across a panel of cancer cell lines originating from breast, endometrial, ovarian, and prostate tumors and found that miR-888 levels were highest in EC cell lines (Figure S1), albeit low in comparison to testes ([Figure 2](#f0010){ref-type="fig"}*A*) and individual miR-888 expressing endometrial tumors (Figure S2). The cell line data support TCGA results in that miR-888 expression was highest in EC. MiR-888 expression in endometrial tumors from TCGA data was heavily skewed, with 65% of samples having undetectable expression of miR-888 ([Figure 3](#f0015){ref-type="fig"}*A* and Table S2). Such a specific expression pattern is characteristic of CT antigens, which are not endogenously expressed in the tumor tissue of origin [@bb0015]. UCEC tumors that had positive expression of miR-888 were substantially more likely to be grade 3 tumors than UCEC tumors with undetectable miR-888 levels ([Table 1](#t0005){ref-type="table"}, *P* = .001). Furthermore, miR-888 expression in UCEC tumors showed a positive association with percent invasion of the tumor ([Figure 3](#f0015){ref-type="fig"}*B*). Therefore, miR-888 expression in endometrial tumors correlates with a more aggressive tumor phenotype and has an expression pattern in cancer similar to other representative CT antigens. To validate TCGA miR-888 expression data, we obtained RNA from 9 benign endometrial tissues and 35 endometrial tumor specimens that included 18 EAs, 9 SAs, and 8 CSs ([Table S1](#ec0005){ref-type="supplementary-material"}). MiR-888 was nine-fold overexpressed in endometrial tumor specimens compared to BE with a *P* value \< .01 ([Figures 3](#f0015){ref-type="fig"}*C* and [S2](#f0040){ref-type="fig"}). Within endometrial tumor subtypes, miR-888 was significantly overexpressed in EA (11-fold, *P* \< .01) and most predominantly overexpressed in CS (22-fold, *P* \< .01). Endometrial CS is a rare but very aggressive form of EC that has a very poor prognosis and limited treatment options [@bb0130]. In contrast, EA often presents at an early stage and grade and can be treated by surgical removal alone [@bb0135]. Thus, these data suggest that miR-888 is associated with an aggressive tumor phenotype. To validate TCGA miR-888 expression data, we obtained RNA from 9 benign endometrial tissues and 35 endometrial tumor specimens that included 18 EAs, 9 SAs, and 8 CSs (Table S1). MiR-888 was nine-fold overexpressed in endometrial tumor specimens compared to BE with a *P* value \< .01 ([Figures 3](#f0015){ref-type="fig"}*C* and S2). Within endometrial tumor subtypes, miR-888 was significantly overexpressed in EA (11-fold, *P* \< .01) and most predominantly overexpressed in CS (22-fold, *P* \< .01). Endometrial CS is a rare but very aggressive form of EC that has a very poor prognosis and limited treatment options [@bb0130]. In contrast, EA often presents at an early stage and grade and can be treated by surgical removal alone [@bb0135]. Thus, these data suggest that miR-888 is associated with an aggressive tumor phenotype. To determine what cell types expressed miR-888 in endometrial tumors, we performed miRNA *in situ* hybridization on EA and CS samples (EA38 and CS114; [Table S1](#ec0005){ref-type="supplementary-material"}) that had particularly high expression of miR-888. H&E staining for the EA and CS samples was performed on adjacent tissue sections as a reference ([Figure 4](#f0020){ref-type="fig"}, *A* and *B*). In EA, miR-888 expression localized to the cancerous glandular epithelium ([Figure 4](#f0020){ref-type="fig"}*C*), with minimal staining in the surrounding stroma and the absence of staining observed for the scrambled negative control probe ([Figure 4](#f0020){ref-type="fig"}*D*). CS is unique in that it contains both carcinomatous and sarcomatous components within a single tumor. It is hypothesized that the cancer was originally an endometrial carcinoma that dedifferentiated to form sarcoma-like cells [@bb0140]. MiRNA *in situ* hybridization of an endometrial CS revealed that miR-888 expression remained in the carcinomatous portion of the tumor ([Figure 4](#f0020){ref-type="fig"}, *E* and *F*). Interestingly, the carcinomatous portion of the tumor is typically the cell type functionally involved in tumor invasion and metastasis [@bb0140]. These data demonstrate that miR-888 expression is restricted to the cancerous glandular epithelium of endometrial tumors and points to the role of miR-888 in cancer cell signaling pathways. To determine what cell types expressed miR-888 in endometrial tumors, we performed miRNA *in situ* hybridization on EA and CS samples (EA38 and CS114; Table S1) that had particularly high expression of miR-888. H&E staining for the EA and CS samples was performed on adjacent tissue sections as a reference ([Figure 4](#f0020){ref-type="fig"}, *A* and *B*). In EA, miR-888 expression localized to the cancerous glandular epithelium ([Figure 4](#f0020){ref-type="fig"}*C*), with minimal staining in the surrounding stroma and the absence of staining observed for the scrambled negative control probe ([Figure 4](#f0020){ref-type="fig"}*D*). CS is unique in that it contains both carcinomatous and sarcomatous components within a single tumor. It is hypothesized that the cancer was originally an endometrial carcinoma that dedifferentiated to form sarcoma-like cells [@bb0140]. MiRNA *in situ* hybridization of an endometrial CS revealed that miR-888 expression remained in the carcinomatous portion of the tumor ([Figure 4](#f0020){ref-type="fig"}, *E* and *F*). Interestingly, the carcinomatous portion of the tumor is typically the cell type functionally involved in tumor invasion and metastasis [@bb0140]. These data demonstrate that miR-888 expression is restricted to the cancerous glandular epithelium of endometrial tumors and points to the role of miR-888 in cancer cell signaling pathways. PR Expression in EC {#s0105} ------------------- In the same subset of tumor tissues used for miR-888 quantification, we also quantified PR mRNA expression (7 benign samples of endometrium, 12 EAs, 7 SAs, and 7 CSs; [Table S1](#ec0005){ref-type="supplementary-material"}; samples not used did not have enough remaining RNA). As has been reported previously [@bb0080; @bb0145], we observed a loss of PR expression in endometrial tumors compared to BE ([Figures 5](#f0025){ref-type="fig"}*A* and [S2](#f0040){ref-type="fig"}). PR expression was 20-fold decreased in EA (*P* \< .001) and 100-fold decreased in SA relative to BE (*P* \< .0001). In CS, PR expression was essentially absent (1000-fold decreased, *P* \< .0001; [Figure 5](#f0025){ref-type="fig"}*A*). Linear regression demonstrated a statistically significant negative correlation between miR-888 and PR mRNA expression (*R* = − 0.579, *P* value \< .001; [Figure 5](#f0025){ref-type="fig"}*B*). In particular, this implies that a five-fold increase in miR-888 concentration is associated with a three-fold decrease in PR concentration (95% confidence interval: 1.7 to 5.3). In the same subset of tumor tissues used for miR-888 quantification, we also quantified PR mRNA expression (7 benign samples of endometrium, 12 EAs, 7 SAs, and 7 CSs; Table S1; samples not used did not have enough remaining RNA). As has been reported previously [@bb0080; @bb0145], we observed a loss of PR expression in endometrial tumors compared to BE ([Figures 5](#f0025){ref-type="fig"}*A* and S2). PR expression was 20-fold decreased in EA (*P* \< .001) and 100-fold decreased in SA relative to BE (*P* \< .0001). In CS, PR expression was essentially absent (1000-fold decreased, *P* \< .0001; [Figure 5](#f0025){ref-type="fig"}*A*). Linear regression demonstrated a statistically significant negative correlation between miR-888 and PR mRNA expression (*R* = − 0.579, *P* value \< .001; [Figure 5](#f0025){ref-type="fig"}*B*). In particular, this implies that a five-fold increase in miR-888 concentration is associated with a three-fold decrease in PR concentration (95% confidence interval: 1.7 to 5.3). MiR-888 Inhibits the PR in EC {#s0110} ----------------------------- Given the significant association between miR-888 and PR mRNA levels, we next analyzed the 3′UTR of PR for potential binding sites for miR-888 using TargetScanHuman 6.2 ([www.targetscan.org](http://www.targetscan.org){#ir0080}). Similar to other steroid hormone receptors, PR has a very long 3′UTR (≈ 10 kilobases) with several AU-rich elements (AREs; [Figure 6](#f0030){ref-type="fig"}*A*, *stars*), suggesting a very complex post-transcriptional regulatory mechanism. Intriguingly, the PR 3′UTR contained four binding sites for miR-888 in its 3′UTR ([Figure 6](#f0030){ref-type="fig"}*A*, *triangles*), with the first three sites having 7mer-m8 complementarity and the last site having 8mer complementarity ([Figure 6](#f0030){ref-type="fig"}*F*). To investigate the role of miR-888 regulation of PR, we chose to use the ECC-1 and Ishikawa EC cells because they had a low level of miR-888 expression ([Figure S1](#f0035){ref-type="fig"}) and are positive for PR expression [@bb0100; @bb0150]. The ECC-1 and Ishikawa EC cell lines were transfected with a plasmid containing the *miR-888* gene (*miR-888*) or an EV as a negative control, and miR-888 expression was confirmed by quantitative RT-PCR ([Figure S3](#f0045){ref-type="fig"}). Mock transfections were performed as a negative control and were used for data normalization. At 4 hours post-transfection, cells were treated with 5 nM estradiol to induce PR expression. Protein was isolated at 24 hours post-transfection, and Western blot analysis was performed for PR isoforms PRA and PRB. Only PRB was predominantly expressed in ECC-1 and Ishikawa cells, and miR-888 overexpression was capable of reducing PRB at the protein level in both cell lines ([Figure 6](#f0030){ref-type="fig"}, *B* and *C*). Western blots were quantified by densitometry using ImageJ software, and a significant decrease in PR protein was observed with miR-888 transfection, but not EV transfection, relative to the mock-transfected controls in the ECC-1 ([Figure 6](#f0030){ref-type="fig"}*D*) and Ishikawa ([Figure 6](#f0030){ref-type="fig"}*E*) cell lines. Given the significant association between miR-888 and PR mRNA levels, we next analyzed the 3′UTR of PR for potential binding sites for miR-888 using TargetScanHuman 6.2 ([www.targetscan.org](http://www.targetscan.org){#ir0085}). Similar to other steroid hormone receptors, PR has a very long 3′UTR (≈ 10 kilobases) with several AU-rich elements (AREs; [Figure 6](#f0030){ref-type="fig"}*A*, *stars*), suggesting a very complex post-transcriptional regulatory mechanism. Intriguingly, the PR 3′UTR contained four binding sites for miR-888 in its 3′UTR ([Figure 6](#f0030){ref-type="fig"}*A*, *triangles*), with the first three sites having 7mer-m8 complementarity and the last site having 8mer complementarity ([Figure 6](#f0030){ref-type="fig"}*F*). To investigate the role of miR-888 regulation of PR, we chose to use the ECC-1 and Ishikawa EC cells because they had a low level of miR-888 expression (Figure S1) and are positive for PR expression [@bb0100; @bb0150]. The ECC-1 and Ishikawa EC cell lines were transfected with a plasmid containing the *miR-888* gene (*miR-888*) or an EV as a negative control, and miR-888 expression was confirmed by quantitative RT-PCR (Figure S3). Mock transfections were performed as a negative control and were used for data normalization. At 4 hours post-transfection, cells were treated with 5 nM estradiol to induce PR expression. Protein was isolated at 24 hours post-transfection, and Western blot analysis was performed for PR isoforms PRA and PRB. Only PRB was predominantly expressed in ECC-1 and Ishikawa cells, and miR-888 overexpression was capable of reducing PRB at the protein level in both cell lines ([Figure 6](#f0030){ref-type="fig"}, *B* and *C*). Western blots were quantified by densitometry using ImageJ software, and a significant decrease in PR protein was observed with miR-888 transfection, but not EV transfection, relative to the mock-transfected controls in the ECC-1 ([Figure 6](#f0030){ref-type="fig"}*D*) and Ishikawa ([Figure 6](#f0030){ref-type="fig"}*E*) cell lines. To validate PR as a direct target of miR-888 regulation, we cloned 500-bp regions surrounding each PR 3′UTR miR-888 binding site ([Figure 6](#f0030){ref-type="fig"}*F*) into the psiCHECK2 vector downstream of *Renilla* luciferase. ECC-1 and Ishikawa cells were transfected with the miR-888 plasmid in combination with each PR-psiCHECK2 vector (sites PR1-PR4; [Figure 6](#f0030){ref-type="fig"}*F*), and luciferase activity was measured 24 hours post-transfection. *Renilla* luciferase activity was normalized to *Firefly* luciferase activity and data are reported relative to the EV transfection ([Figures 6](#f0030){ref-type="fig"}, *G* and *H*). Expression of miR-888 caused a significant reduction in luciferase activity of approximately 10% to 20% for all four binding sites in both cell lines ([Figures 6](#f0030){ref-type="fig"}, *G* and *H*). Therefore, these data support that PR is a direct target of miR-888 regulation in EC cells. Discussion {#s0020} ========== Here, we describe for the first time a miRNA, miR-888, that embodies the cardinal characteristics of a CT antigen. This is the first noncoding RNA to meet these criteria. Though CT antigens were historically identified through their ability to elicit an immune response when expressed in tumors, the term CT antigen has evolved to include a broader range of characteristics. There are more than 200 CT antigens currently classified in the CT antigen database (<http://www.cta.lncc.br>), many of which have yet to be investigated for their immunogenic potential [@bb0015]. Further, several CT antigens that were originally described as immunogenic have more recently been shown to elicit only a minimal immune response *in vivo*[@bb0015]. Therefore, even though the term antigen signifies an immunogenic molecule, not all CT antigens display this trait. A more fundamental characteristic of CT antigens is their highly restricted expression pattern in testes and cancer. Therefore, we suggest that noncoding RNAs displaying this expression pattern can be appropriately categorized as CT antigens. In 2010, Li et al. reported the evolution of a primate-specific X-linked miRNA gene family spanning a 33-kilobase region at Xq27.3 [@bb0025]. The *miR-888* gene family is located in a 3-kilobase cluster within this region and includes the miRNAs miR-888, miR-890, miR-891a, miR-891b, miR-892a, and miR-892b [@bb0025]. Rapid evolution through gene duplication events resulted in the emergence of the miR-888 family members in primates [@bb0025]. Ohno's law predicts that X-linked genes are more highly conserved across placental mammals than autosomal genes due to X chromosome inactivation and gene dosage effects [@bb0155]. However, CT-X antigen genes are often exceptions to this rule [@bb0020; @bb0160]. It has been hypothesized that recessive genes that are selectively beneficial to males become fixed more rapidly on the X chromosome than on the autosomes [@bb0165]. This makes the X chromosome a hotspot for CT-X antigens, while the remaining genes on the X-chromosome are highly conserved and follow Ohno's law [@bb0165]. MiR-888 follows the specific evolutionary pattern of CT-X antigens with its recent evolution on the X-chromosome and restricted expression in testis [@bb0025]. In addition, miR-888 has predicted targets that are known to be involved in sperm cell production and regulation within the testes [@bb0025; @bb0030]. We extend this literature by identifying miR-888 expression in cells during the early stages of spermatogenesis within the seminiferous tubules of the testes. This information supports the evolution of the *miR-888* gene locus in primates as a CT-X antigen that likely plays a role in testis cell biology. Landgraf et al. first discovered the *miR-888* gene family through small RNA library sequencing and reported highly specific expression of the miRNA family in the epididymis of the testes [@bb0030]. Belleannee et al. performed microarrays to investigate miRNA expression across different regions of the epididymis [@bb0170]. All members of the miR-888 family, except for miR-888, were found to be differentially expressed throughout the epididymis [@bb0170]. Furthermore, miR-888 expression was very low in the epididymis relative to its other family members, suggesting that it might not play a prominent role in this tissue [@bb0170]. Our data show that miR-888 expression is highly specific to the testes, with a low level of expression in the kidney, prostate, thymus, and trachea. MiR-888 has previously been detected in regions outside the testes but not in normal tissue under endogenous conditions [@bb0175; @bb0180]. The miR-888 opposing strand, miR-888\*, was detected in human oocytes, yet miR-888\* has a different miRNA seed sequence [@bb0185]. Therefore, we conclude that miR-888 expression is highly enriched in the human testes. It is also important to note that quantitative RT-PCR is a very sensitive technique and has demonstrated expression of many CT antigens in tissues in addition to testes, albeit at much lower levels [@bb0015]. Therefore, miR-888 expression across normal human tissues is consistent with observations of other CT antigens. Heterogeneous expression of miR-888 across various cancer types was observed wherein miR-888 was only expressed in a small subset of tumors. This expression pattern is characteristic of CT antigens, which are not expressed in the tumor tissue of origin and are only expressed in a portion of tumors [@bb0005]. In a panel of eight different types of tumors, miR-888 was most predominantly expressed in EC. This is particularly relevant in that similar to the testes, the endometrium is a tissue governed by extensive hormonal regulation. Association of miR-888 expression with endometrial tumors of high grade and increased invasion implies that miR-888 expression indicates a more advanced form of EC. Furthermore, miR-888 was most highly expressed in the very rare and aggressive endometrial CS, which has often metastasized at the time of diagnosis and has a poor five-year survival rate of only 20% to 35% [@bb0130]. Expression of other CT antigens has also been correlated with aggressive tumors [@bb0190]. In EC, CT antigens MAGEC1, MAGEA3, and MAGEA4 expression correlated with high-grade tumors [@bb0195], and MAGEA4 and NY-ESO were most predominantly expressed in endometrial CSs [@bb0200]. Thus, miR-888 expression in endometrial tumors is consistent with the characteristic expression patterns observed for CT antigens. One of the most important findings reported herein is the identification of PR as a direct target of miR-888. Throughout evolution, 3′UTR length has expanded, with steroid hormone receptors demonstrating the most extensive elongation [@bb0205]. The PR 3′UTR underwent recent expansion in primates to approximately 10 kilobases in length, generating new miRNA binding sites for evolutionarily novel miRNAs such as miR-888 [@bb0210]. The PR 3′UTR also contains binding sites for many other miRNAs, such as the recently validated miR-96 [@bb0215] and miR-126-3p [@bb0220], as well as the predicted regulators miR-181 and miR-26a [@bb0225]. The PR 3′UTR contains many AREs as well, rendering the mRNA highly unstable [@bb0230]. Therefore, the post-transcriptional regulation of PR is multifactorial and complex. The miR-888--dependent decrease in PR protein expression and 3′UTR luciferase reporter activity that we report here was relatively modest. Due to the AU-rich nature of miR-888, this is not surprising and is consistent with previous reports. For example, Garcia et al. report that AU-rich miRNA seeds have lower seed pairing stability and higher target abundance, rendering them less proficient [@bb0235]. In addition, a modest reduction in target protein levels by miRNAs has been reported by others previously [@bb0240; @bb0245; @bb0250]. Baek et al. performed protein arrays after knockout of miR-223 and found that a large number of targets were only reduced by about 30% [@bb0255]. The authors hypothesized that many miRNAs function through a fine-tuning mechanism rather than causing a dramatic change in protein translation of target mRNAs [@bb0255]. We suggest that miR-888 is but one of several factors using the long PR 3′UTR to regulate mRNA stability and protein translation. These different regulators likely work together through cross-talk mechanisms to cause a more pronounced and cumulative effect on PR protein expression than any one could have alone. The importance of PR in the testes has recently been investigated. Shah et al. reported that PR expression is low during the early stages of spermatogenesis and highest during late stages IV and V [@bb0260]. Here, we show an opposite staining pattern for miR-888 in Rhesus testes, with high expression in the early stages of spermatogenesis and the absence of expression in the later stages. Therefore, we suggest that miR-888 might function to keep PR low during the early stages of spermatogenesis, followed by a loss of miR-888 expression during the late stages of spermatogenesis, permitting PR to induce germ cell differentiation. Consistent with this hypothesis, Abid et al. reported that PR expression was lost in men with infertility due to nonobstructive azoospermia, in which germ cells fail to mature into spermatozoa [@bb0265]. In addition, miR-888 appeared in a screen for miRNAs differentially expressed in sperm from infertile males with oligozoospermia and asthenozoospermia [@bb0270]. MiR-888 was overexpressed in both types of male infertility, despite the fact that the *P* values did not reach statistical significance in this relatively small sample size [@bb0270]. Nevertheless, the preponderance of the literature indicates that tight regulation of PR signaling is essential for proper germ cell maturation, thus establishing a potential role of miR-888 in male fertility. Conclusions {#s0025} =========== We propose that miR-888 should be considered the first miRNA CT antigen. MiR-888 recently evolved in primates on the X chromosome and is expressed specifically during the early stages of spermatogenesis within the testes. MiR-888 is also overexpressed in cancer, specifically in endometrial tumors. Moreover, expression was only observed in a subset of endometrial tumors and was associated with high-grade and enhanced tumor invasion into the myometrium. In EC cells, we confirm PR as a direct target of miR-888 regulation. With PR well described as a tumor suppressor of the endometrium, miR-888 is a potential oncogene in EC. MiR-888 likely works in combination with other miRNAs and factors to post-transcriptionally regulate PR. The role of miR-888 in the testes is not yet fully defined, but regulation of PR by miR-888 in the spermatogenic cells of the testes is an intriguing hypothesis for future research. The following are the supplementary data related to this article.Figure S1MiR-888 expression in EC cell lines.Figure S2ΔCt values for miR-888 and PR quantification in EC tissues.Figure S3Transfection of ECC-1 and Ishikawa cells with the miR-888 plasmid results in high expression of miR-888.Table S1Endometrial Tumor PanelTable S2TCGA UCEC Clinical CharacteristicsTable S3TCGA UCEC Patient Clinical InformationTable S4MiR-888 Expression across Multiple Cancers from TCGATable S5PR Insert PCR Cloning Primers Supplementary data to this article can be found online at [http://dx.doi.org/10.1016/j.tranon.2015.02.001](10.1016/j.tranon.2015.02.001){#ir0095}. James Schappet assisted with TCGA data analyses. Michael Goodheart provided tumor specimens from the University of Iowa Department of Obstetrics and Gynecology Tumor Bank. Katherine Gibson-Corey provided pathologic analysis of testis samples. This work was supported by NIH R01CA99908 (K.K.L.) and the Department of Obstetrics and Gynecology Research Development Fund (K.K.L.). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Disclosure: D.D. and K.W.T. are co-owners of Immortagen, L.L.C. All other authors have declared that no competing interests exist. This article refers to supplementary materials, which are designated by Tables S1 to S5 and Figures S1 to S3 and are available online at www.transonc.com. {#f0005} {#f0010} {#f0015} {#f0020} {#f0025} {#f0030} ###### MiR-888 Expression by Tumor Grade Grade Percentage of Samples with Nonzero miR-888 Expression ------- ------------------------------------------------------- G1 17% G2 24% G3 36%[⁎](#tf0005){ref-type="table-fn"} *P* = .001.

\section{Introduction} Forcing the classification of uncertain instances in safety-critical applications can lead to misclassification, which can result in economic losses or increased costs. In contrast, the cost of possible errors can be reduced by abstaining from ambiguous classification, which has been used in cost-sensitive fields \cite{chen2018evidential,kang2017reliable,lin2018biomedical,wang2017fault}. The general classification rule with reject option in binary classification is shown in Eq.~(\ref{eq1}) and can be explained as follows. If the confidence score $s$ belonging to the positive class is larger than $t_2$, then example $m$ is classified as positive (+); if $s$ is not larger than $t_1$, then $m$ is classified as negative (-); otherwise, the example is not labeled (rejected, R). \begin{equation}\label{eq1} C_{t_1,t_2}(m) = \begin{cases} +, & \text{if } s(m)>t_2; \\ -, & \text{if } s(m) \leq t_1; \\ R, & \text{otherwise}. \end{cases} \end{equation} When $t_1=t_2$, Eq.~(\ref{eq1}) reverts to the traditional binary classification rule. The rejection thresholds $t_1$ and $t_2$ define the decision boundaries and the task of training the abstention classifier lies in determining the two rejection thresholds, which can be enforced by establishing different abstention models. Note that only binary classification is discussed in this paper. In the context of abstaining classification, the ideal situation is to establish a cost minimization model, which requires the costs of correct classification, misclassification, and rejection to be known. However, it is often difficult to obtain or estimate cost information in many real-world classification problems. For instance, in the diagnosis of normal and cancerous tissues, the imaging characteristics of some tissues are ambiguous, in which case it is hard to make a definitive diagnosis. Both misdiagnosis and a missed diagnosis can lead to physical and mental pain, and their costs are usually unknown and unequal. There are mainly two abstention models as follows: unconditional optimization model of the expected cost and conditional optimization model with rejection or performance constraints. Chow \cite{chow1970optimum,chow1957optimum} expanded classical Bayesian decision theory by proposing a generalized decision theory with reject option. The well-known Bayesian decision rules are based on a minimum error rate or a minimum risk function. For example, minimum risk Bayesian classification only considers class-dependent misclassification costs. In contrast, once the reject option is added, the generalized Bayesian decision theory minimizes the expected risk based on the rejection costs as well as the costs of correct and incorrect classification. Tortorell \cite{tortorella2004reducing,tortorella2000optimal} proposed the following implementable abstention model based on the generalized theory: \begin{equation} \label{eq2} \min_{t_1,t_2} \hspace{0.5em} cost(t_1,t_2), \end{equation} where \begin{align} \label{eq3} &cost(t_1,t_2)= \notag \\ &p(+)\cdot CFN \cdot fnr(t_1)+p(-) \cdot CTN \cdot tnr(t_1)+ \notag \\ & p(+)\cdot CTP \cdot tpr(t_2) + p(-) \cdot CFP \cdot fpr(t_2)+ \\ & p(+)\cdot CRP \cdot rpr(t_1,t_2) + p(-) \cdot CRN \cdot rnr(t_1,t_2), \notag \end{align} where $p(+)$ and $p(-)$ are the prior probabilities of the positive and negative classes, respectively, CFN represents the cost of a false negative, and $fnr$ denotes the ratio of false negative examples among all positive examples. Likewise for CTN, CTP, CFP, $tnr$, $tpr$, and $fpr$. CRP and CRN indicate the rejection costs for positive and negative classes, respectively, and $rpr$ and $rnr$ are the ratios of rejected examples with respect to the positive and negative classes, respectively. This abstention model requires the complete cost information including the costs of correct classification, misclassification, and rejection to be known. Pietraszek \cite{pietraszek2007use} proposed a bounded-abstention (BA) model that adds an abstention constraint and only requires knowledge of the misclassification costs. The BA model can be represented as: \begin{equation} \label{eq4} \begin{split} &\min_{t_1,t_2} \hspace{0.5em} \frac{CFN \cdot FN(t_1)+CFP \cdot FP(t_2)}{TN(t_1)+FP(t_2)+TP(t_2)+FN(t_1)}, \\ &s.t.\hspace{0.5em} rej(t_1,t_2) \leq k_{max}, \end{split} \end{equation} where $FN$ ($FP$) refers to the number of false negatives (positives); $TN$ ($TP$) represents the number of true negatives (positives); and $rej$ denotes the overall reject rate, i.e., the number of rejected examples divided by the sample size. It is useful to define a cost ratio of CFN to CFP. When the values of CFN and CFP are the same, i.e., the cost ratio is 1, model~(\ref{eq4}) is to minimize the error rate under an abstention constraint. In~\cite{vanderlooy2009roc}, the receiver operating characteristic (ROC) isometric model is employed to minimize the reject rate under class-dependent performance constraints. In the performance constraints, the class distribution and misclassification cost ratios are considered. Although the BA and ROC isometric models avoid setting complete costs (CTP/N, CFP/N, and CRP/N), misclassification costs or their cost ratio are still required. The abstention models mentioned above have the following shortcomings: \begin{itemize} \item they require cost information to be known. However, in practical problems: (a) such costs are hard to obtain or estimate; (b) they are usually dependent on classes and are asymmetric; and (c) in some cases, the cost information evolves over time, which prevents use of the trained model during the test phase. \item they only optimize a single performance metric, such as the expected cost, error rate, precision, or F-measure. The optimized metric may not be useful when practitioners evaluate the classification performance. \end{itemize} To overcome these drawbacks, a multi-objective bounded-abstention (MOBA) model is proposed that minimizes two essential metrics, namely, the false positive rate and false negative rate, under class-dependent abstention constraints. The MOBA model optimizes essential metrics, via which any simple or complicated metric can be calculated as long as the sample sizes of two classes are fixed. In addition, the MOBA model is applicable regardless of whether the costs are known or unknown. A popular evolutionary algorithm called the non-dominated sorting genetic algorithm II (NSGA-II) \cite{deb2002fast} is used to solve the multi-objective optimization (MOO) problem with constraints. The remainder of the paper is organized as follows. Section~\ref{sec2} introduces the motivation and optimization method of the proposed MOBA model, discusses the methods of selecting the best abstaining classifier under different conditions, and summarizes the advantages of the MOBA model. The results of two experiments, one with costs and one without, are presented in Section~\ref{sec4}, and the conclusions drawn from this research are summarized in Section~\ref{sec5}. \section{Proposed MOBA model} \label{sec2} \begin{table}[!htb] \centering \caption{Confusion matrix with rejection in binary classification problems} \begin{tabular}{|l|c|c|c|c|} \toprule \multicolumn{2}{|c|}{\multirow{2}{*}{\makebox[7em]{Confusion matrix}}} & \multicolumn{3}{c|}{\makebox[10.5em]{Predicted label}} \\ \cmidrule{3-5} \multicolumn{2}{|c|}{} & \makebox[3.5em]{+} & \makebox[3.5em]{-} & \makebox[3.5em]{R} \\ \midrule \multirow{2}{*}{\makebox[3.5em]{Real label}} & \makebox[3.5em]{+} & \makebox[3.5em]{TP} & FN & RP \\ \cmidrule{2-5} & \makebox[3.5em]{-} & FP & TN & RN \\ \bottomrule \end{tabular}% \label{t1}% \end{table}% In this section, the motivation that inspires to propose the MOBA model is first presented in Section~\ref{sec2.1} and then several basic concepts related to MOO problems are provided in Section~\ref{sec2.2}. Next, the NSGA-II algorithm is introduced along with the implementation details of solving the MOBA model (Section~\ref{sec2.3}). Selection of the best abstaining classifier and the advantages of the MOBA model are explained in Sections~\ref{sec2.4} and~\ref{sec2.5}. \subsection{Motivation} \label{sec2.1} Regardless of whether the expected cost, error rate, or F-measure are to be optimized, previous abstention models have always required certain essential metrics to be provided, such as the true positive/negative rate ($tpr/tnr$), false positive/negative rate ($fpr/fnr$), positive/negative predictive value ($ppv/npv$), and rejected positive/negative rate ($rpr/rnr$). For example, when minimizing the average cost in (\ref{eq3}), six essential metrics ($tpr/tnr$, $fpr/fnr$, and $rpr/rnr$) are required while when maximizing the F-measure, $ppv$ and $tpr$ are needed to be known. All the essential metrics can be calculated from the confusion matrix with rejection (Table~\ref{t1}). For example, $rpr=RP\,/\,(TP+FN+RP)$, i.e., the number of rejected positive examples (RP) divided by the number of all positive examples (TP+FN+RP). In view of this, optimizing the essential metrics is a natural and reasonable idea. When the sample sizes of two classes are fixed, the confusion matrix with rejection has four degrees of freedom. Therefore, to obtain a definite rejection classifier, four essential metrics should be determined. The MOBA model optimizes the false positive and negative rates under the constraints of the rejected positive and negative rates. A formal description of the MOBA model is as follows: \begin{align} &\min_{t_1,t_2} \hspace{0.5em} \bm{F(t)}=(F_1(\bm{t}),F_2(\bm{t}))=(fpr(t_2),fnr(t_1)), \label{eq5}\\ & s.t.\hspace{0.5em} \begin{cases} \label{eq6} rpr(\bm{t}) \leq p_{max}, \\ rnr(\bm{t}) \leq n_{max}, \\ t_1<t_2, \end{cases} \end{align} where $\bm{t}=(t_1,t_2)$ is the decision vector, $\bm{F(t)}$ represents the objective function vector, which contains two objective functions $fpr$ (related to $t_2$) and $fnr$ (related to $t_1$), and $t_1$ and $t_2$ ($t_1 < t_2$) denote the rejection thresholds. Note that the definitions of $fpr$ and $fnr$ here are different from those in Eq.~(\ref{eq3}). In this case, $fpr$ ($fnr$) denotes the ratios of false positive (negative) examples among the classified negative (positive) examples. After the two rejection thresholds have been determined, $fpr=FP\,/\,(TN+FP)$ and $fnr=FN\,/\,(TP+FN)$ can be calculated. The feasible solution set $T\subseteq \mathbb{R}^2$ is the set of decision vectors that satisfy the constraints: \begin{equation} T=\{\bm{t}\in \mathbb{R}^2| rpr(\bm{t}) \leq p_{max}, rnr(\bm{t}) \leq n_{max}, t_1 < t_2 \}, \end{equation} where $p_{max}$ and $n_{max}$ are the hyperparameters. Note that maximizing $tpr$ and $tnr$ is equivalent to minimizing $fpr$ and $fnr$ in Eq.~(\ref{eq5}). \subsection{Concepts associated with MOO problems} \label{sec2.2} Since MOO problems involve multiple conflicting objectives, the comparison relation in single-objective optimization problems is not applicable. For a given decision vector, some objectives are optimal whereas others are not, and optimizing the suboptimal objectives may degrade the optimal objectives. The partially ordering relation, i.e., the Pareto dominance, is used to compare decision vectors in MOO problems. A decision vector $\bm{a} \in T$ is said to Pareto dominate another decision vector $\bm{b} \in T$, denoted as $\bm{a}\prec \bm{b}$, if and only if (\textit{iff}): \begin{equation} \label{eq7} \forall i \in \{1,2\}, F_i(\bm{a}) \leq F_i(\bm{b}) \wedge \exists j \in \{1,2\}, F_j(\bm{a}) < F_j(\bm{b}). \end{equation} A decision vector $\bm{t}$ is non-dominated with regard to $T$ \textit{iff} there is no decision vector in $T$ that dominates $\bm{t}$. Such non-dominated solutions are referred to as Pareto-optimal. The set of Pareto-optimal solutions related to $T$ is referred to as the Pareto-optimal set (POS): \begin{equation} POS = \{\bm{t} \in T | \neg \exists \bm{t'} \in T, \bm{t'} \prec \bm{t} \}. \end{equation} The set of objective vectors corresponding to the POS is referred to as the Pareto-optimal front (POF): \begin{equation} POF = \{\bm{F(t)} \in [0,1]^2 | \bm{t} \in POS\}. \end{equation} Evolutionary algorithms (EAs) based on Pareto dominance exhibit excellent performance when solving MOO problems with few (two or three) objectives \cite{zitzler2000comparison}. EAs search the set of Pareto-optimal solutions in parallel in a single run. Popular EAs include the NSGA-II \cite{deb2002fast}, strength Pareto evolutionary algorithm 2 (SPEA2) \cite{zitzler2001spea2}, and Pareto envelope based selection algorithm II (PESA-II) \cite{corne2001pesa}. The popular NSGA-II algorithm was adopted in this study to solve the proposed MOBA model. \subsection{Evolutionary MOO of MOBA} \label{sec2.3} In this section, the NSGA-II algorithm is introduced along with the details required to optimize the proposed MOBA model. NSGA-II improves on the previous NSGA \cite{srinivas1994muiltiobjective} by developing a fast non-dominated sorting and elitism strategy. The basic framework of NSGA-II is presented in Algorithm 1. \begin{algorithm}[!htb] \caption{The basic framework of NSGA-II.} \SetAlgoNoLine \KwIn{$popsize$, the population size; $gensize$, the max generation number\\} \KwOut{$P$, the population at the end of $gensize$ generations \\} \BlankLine $P^0 \leftarrow$ pop-initialization($popsize$);\\ $t \leftarrow$ 0;\\ \While{$t< gensize$}{ $Q^t \leftarrow $ make-new-pop($P^t$);\\ $R^t \leftarrow P^t \cup Q^t$;\\ $P^{t+1} \leftarrow$ elite-preservation($R^t, popsize$);\\ $t \leftarrow t+1$;} Return $P^{t+1}$. \BlankLine \textbf{Function} make-new-pop($P^t$)\\ ${\mathcal{F}}^t \leftarrow$ non-dominated-sort($P^t$);\\ ${\mathcal{S}}^t \leftarrow$ crowding-distance-assignment(${\mathcal{F}}^t$);\\ $parent \leftarrow$ tournament-selection($P^t, {\mathcal{F}}^t, {\mathcal{S}}^t, popsize$);\\ $Q^t \leftarrow$ genetic-operation($parent$);\\ Return $Q^t$. \BlankLine \textbf{Function} elite-preservation($R^t, popsize$)\\ ${\mathcal{F}}^t \leftarrow$ non-dominated-sort($R^t$);\\ $P^{t+1} \leftarrow \varnothing$; $i \leftarrow$ 0;\\ \While{$|P^{t+1}|+|{\mathcal{F}_i^t}| \leq popsize$} { $P^{t+1} \leftarrow P^{t+1} \cup {\mathcal{F}_i^t}$;\\ $i \leftarrow i+1$; } ${\mathcal{S}_i^t} \leftarrow$ crowding-distance-assignment(${\mathcal{F}_i^t}$);\\ $P^{t+1} \leftarrow P^{t+1} \cup {\mathcal{F}_i^t}[\{{\mathcal{S}_i^t}\}_1^{popsize-|P^{t+1}|}]$;\\ Return $P^{t+1}$. \end{algorithm} \textbf{Pop-initialization} generates the initial population that includes $popsize$ individuals (chromosomes). Each decision vector $\bm{t}$ denotes an individual or a chromosome. In the MOBA problem, the two variables in each decision vector, i.e., the two rejection thresholds $t_1$ and $t_2$, are encoded with real values. Specifically, the scores $s$ of the training examples are first determined via a scoring classifier. Traditional classification methods, such as support vector machine and k-nearest neighbor, can be used as the scoring classifier \cite{fawcett2006introduction}. Let $s_{min}$ and $s_{max}$ denote the minimal and maximal ones among the scores of all training examples, respectively. Then, $t_1$ and $t_2$ are randomly generated in the range of [$s_{min}, s_{max}$] only if $t_1 < t_2$ is satisfied. In this study, $popsize$ and $gensize$ are set to 20 and 100, respectively. \textbf{Non-dominated-sort} sorts the individuals in population $P^t$ based on non-domination and outputs the front set ${\mathcal{F}}^t=\{{\mathcal{F}}_1^t, {\mathcal{F}}_2^t, \cdots, {\mathcal{F}}_n^t\}$. The individuals in each front ${\mathcal{F}}_i^t$ ($i \in \{1, 2, \cdots, n\}$) are non-dominated while the individuals belonging to ${\mathcal{F}}_i^t$ are dominated by the individuals in front ${\mathcal{F}}_j^t$ ($j<i, j \in \{1, 2, \cdots, n\}$) in the $t^{th}$ generation. The front sets are depicted in Figure~\ref{f1}, where $F_1$ and $F_2$ represent the values of two objective functions. For the detailed sort algorithm, please refer to~\cite{deb2002fast}. The objective values $fpr$ and $fnr$ are computed using the validation set. Specifically, all examples are divided into three parts consisting of the training, validation, and test sets. The training set is employed to construct a scoring classifier, and the scores of the examples in the validation set are computed using the scoring classifier. Given $t_1$ and $t_2$ (corresponding to the variables in each individual), Eq.~(\ref{eq1}) can be applied to the validation examples, thereby allowing the basic metrics in Table~\ref{t1} to be calculated. If the constraints~(\ref{eq6}) are not satisfied, $fpr$ and $fnr$ are assigned the maximum value of one, in which case the corresponding individual is not considered. \begin{figure}[!htb] \centering \includegraphics[width=0.68\textwidth]{Fig1.jpg} \caption{The illustration of front sets}\label{f1} \end{figure} \begin{figure}[!htb] \centering \includegraphics[width=0.68\textwidth]{Fig2.jpg} \caption{The illustration of the crowding distance}\label{f2} \end{figure} \textbf{Crowding-distance-assignment} computes the crowding distances of individuals in each front and outputs the spread set ${\mathcal{S}}^t=\{{\mathcal{S}}_1^t, {\mathcal{S}}_2^t, \cdots, {\mathcal{S}}_n^t\}$, where ${\mathcal{S}}_i^t$ ($i \in \{1, 2, \cdots, n\}$) stores the crowding distances of the individuals in front ${\mathcal{F}}_i^t$ in the $t^{th}$ generation. Note that since the crowding distance is assigned within each front, it is meaningless to compare the crowding distances of two individuals from different fronts. The essential idea of calculating the crowding distance is to sort all individuals in the same front based on each objective function and to average the Euclidean distances between the nearest neighbors in each dimension of the objective function space. An illustration of the crowding distance is shown in Figure~\ref{f2}. For objective function $F_1$, the distance $d_1$ of the $k^{th}$ individual is: \begin{equation} d_1 = \frac{F_1^{k+1}-F_1^{k-1}}{F_1^{max}-F_1^{min}}, \end{equation} where $F_1^{k+1}$ and $F_1^{k-1}$ are the objective values of the $(k+1)^{th}$ and $(k-1)^{th}$ individuals in dimension $F_1$, respectively, and $F_1^{max}$ and $F_1^{min}$ are the maximum and minimum values in the front, respectively. A similar approach can be used to obtain the distance of the $k^{th}$ individual in other dimensions of the objective function space. Finally, the crowding distance of the $k^{th}$ individual is the sum of the distances in all dimensions. Note that the crowding distance of the boundary individual is assigned as infinite. In each ${\mathcal{S}}_i^t$, the crowding distances are sorted in descending order. \textbf{Tournament-selection} uses a binary tournament strategy to select $popsize$ individuals from the population $P^t$ as follows. First, two individuals are randomly selected from $P^t$. Then, one of them is chosen based on two criteria, namely, the front rank and crowding distance. If the two individuals are in different fronts, the individual with the lower front rank is selected. However, if the front ranks of the two individuals are the same, then the individual with the larger crowding distance is selected to maintain solution diversity. The binary tournament selection process is performed $popsize$ times to obtain $popsize$ individuals, which then constitute the $parent$ population. \textbf{Genetic-operation} performs the simulated binary crossover (SBX)~\cite{agrawal1995simulated} and polynomial mutation~\cite{kakde2004survey} operations in the real value coded evolutionary MOO algorithm. Specifically, in SBX, two children chromosomes are generated as follows~\cite{beyer2001self}: \begin{align} y_{1,m} = \frac{1}{2}[(1-\beta_m)x_{1,m}+(1+\beta_m)x_{2,m}], \label{eq11} \\ y_{2,m} = \frac{1}{2}[(1+\beta_m)x_{1,m}+(1-\beta_m)x_{2,m}], \label{eq12} \end{align} where $y_{1,m}$ and $y_{2,m}$ ($m \in \{1,2\}$) are the $m^{th}$ variables in the two children $y_1$ and $y_2$, respectively; $x_{1,m}$ and $x_{2,m}$ are the $m^{th}$ variables of the randomly selected parents, respectively, and $\beta_m (\geq 0)$ is a sample from a random number generator having the density: \begin{equation} p(\beta) = \begin{cases} \frac{1}{2}(\eta_c+1)\beta^{\eta_c}, &\text{if } 0 \leq \beta \leq 1\\ \frac{1}{2}(\eta_c+1)\frac{1}{\beta^{\eta_c+2}}, &\text{if } \beta>1 \end{cases} \end{equation} where $\eta_c$ is the distribution index of the crossover. The distribution can be obtained from a uniformly sampled random number $u$ between (0,1): \begin{equation} \label{eq14} \beta(u) = \begin{cases} (2u)^{\frac{1}{\eta_c+1}}, & \text{if } u \leq 0.5 \\ \frac{1}{(2-2u)^{\frac{1}{\eta_c+1}}}. &\text{if } u>0.5 \end{cases} \end{equation} When generating children individuals using the SBX operation, a random number $u$ between (0,1) is obtained, $\beta$ is computed as per Eq.~(\ref{eq14}), and the child variables are obtained via Eqs.~(\ref{eq11}) and (\ref{eq12}). After two child chromosomes are obtained, the constraint $t_1<t_2$ is checked. If the constraint is not satisfied, a new random number between (0,1) is generated and the computation process is repeated. In polynomial mutation, a child chromosome is generated as follows: \begin{equation} \label{eq15} y_m = x_m+(x_m^u-x_m^l)\delta_m, \end{equation} where $y_m$ ($m \in \{1,2\}$) is the $m^{th}$ variable in the child chromosome $y$, $x_m$ is the $m^{th}$ variable of the parent chromosome $x$, and $x_m^u$ and $x_m^l$ are the upper and lower bounds, respectively. In the MOBA model, $x_m^u = 1$ and $x_m^l =0$, and $\delta_m$ follows the polynomial probability distribution \cite{kakde2004survey}: \begin{equation} p(\delta)=\frac{1}{2}(\eta_m+1)(1-|\delta|^{\eta_m}), \end{equation} where $\eta_m$ is the distribution index for mutation and $\delta$ can be calculated as: \begin{equation} \label{eq17} \delta(u) = \begin{cases} (2u)^{\frac{1}{\eta_m+1}}-1, & \text{if } u<0.5 \\ 1-(2-2u)^{\frac{1}{\eta_m+1}}, &\text{if } u \geq 0.5 \end{cases} \end{equation} where $u$ is a random number between (0,1). Similarly, to obtain a child chromosome, $u$ is generated randomly between (0,1) and $\delta$ is computed using Eq.~(\ref{eq17}). Finally, a child is generated as per Eq.~(\ref{eq15}) and the constraint $t_1<t_2$ is checked. New values of $u$ are generated until the constraint is satisfied. Finally, the offspring population $Q^t$ is produced from the chromosomes in the $parent$ population. In the experiments, the commonly used hyperparameters in the NSGA-II algorithm are set as follows \cite{deb2002fast,agrawal1995simulated}. The crossover probability is 0.9; the mutation probability is $1/v$, where $v$ is the number of decision variables, and here, $v=2$; and the distribution indexes of both the crossover and mutation operators are set to 20. \textbf{Elite-preservation} selects the first $popsize$ best individuals from the combined population $R^t$. All individuals in $R^t$ are sorted based on non-domination. The individuals in the low fronts are better solutions than those in the high fronts. If the number of chromosomes in ${\mathcal{F}_1}$ is smaller than $popsize$, the next non-dominated set ${\mathcal{F}_2}$ is considered until the size of $P^{t+1}$ is larger than $popsize$. It is assumed that when the $i^{th}$ front set is added into $P^{t+1}$, the number of individuals exceeds $popsize$. The crowding distances of the solutions in ${\mathcal{F}_i}$ are assigned and sorted, and the individuals that have large crowding distances are added into $P^{t+1}$ until the size of $P^{t+1}$ is exactly equal to $popsize$. The MOBA model requires two hyperparameters to be set, namely, the maximum abstention rates $p_{max}$ and $n_{max}$ with respect to the two classes. A larger abstention rate results in better performance, but increases the cost of dealing with rejected instances. Note that this trade-off always arises in abstaining classification applications. The hyper-parameters can be set depending on the requirements of the application (associated with the performance) and the resource limitations (associated with the reject rate). The hyper-parameter setting is presented in Sections~\ref{sec4.1} and \ref{sec4.2}. It is important to note that the abstention rates of the two classes can be set using different values. This ensures the MOBA model can be applied in situations where the class distribution is imbalanced and the error costs are asymmetric. Furthermore, another advantage of the MOBA model is that a set of classifiers is generated, instead of just one classifier. This provides the ability to select and change classifiers without incurring the expense of retraining. \subsection{Methods of selecting the best abstaining classifier} \label{sec2.4} In the MOBA model, the NSGA-II algorithm outputs multiple Pareto-optimal vectors $(t_1, t_2)$, each of which corresponds to an abstaining classifier. For a fixed abstaining classifier, test examples can be classified or rejected according to Eq.~(\ref{eq1}), and essential metrics can be calculated. With this in mind, the methods of selecting the best classifier depend on the cost information and the metrics used to evaluate performance. \begin{itemize} \item If the complete costs (CTP/N, CFP/N, and CRP/N) are known, then the solution having the minimum expected cost (Eq.~(\ref{eq3})) is theoretically optimal. Specifically, at the end of the 100th generation, 20 Pareto-optimal vectors are obtained, each of which corresponds to a confusion matrix with rejection. Hence, $tpr/tnr$, $fpr/fnr$, and $rpr/rnr$ can be computed, and correspondingly, a set of 20 expected costs can be obtained as per Eq.~(\ref{eq3}), from which the smallest one is selected. \item If the cost information is unknown, practitioners can compare the performance-abstention trade-offs of the Pareto-optimal solutions and select the best solution for a particular circumstance. For example, it is possible to preset tolerable maximum reject rates and then choose the classifier that has the best performance in terms of the accuracy, area under the ROC curve (AUC), etc. Once 20 confusion matrices with rejection are obtained, the corresponding rejected rates can be computed, and eligible Pareto-optimal vectors are selected according to the preset maximum value. Among the selected vectors, the best solution is the one that obtains a highest AUC, for example. Note that AUC is the average of $tpr$ and $tnr$ \cite{lopez2013insight}. \end{itemize} \subsection{Superiority of MOBA over previous abstaining models} \label{sec2.5} The advantages of the MOBA model are summarized as follows: \begin{enumerate} \item the MOBA model does not require set costs, and although the abstention constraints involve two hyperparameters, they are in the range of (0,1). This is distinctly different than the costs in previous models that take on unbounded real values. In the models of unconditional optimization of the expected cost Eq. (\ref{eq2}) and conditional optimization with constraints such as Eq. (\ref{eq4}), cost information is required to be provided, and unfortunately, it is usually unknown. Hence, empirical costs or cost ratios are commonly set in particular applications such as intrusion detection \cite{pietraszek2007classification} or cost models are used to evaluate statistical results (Section~\ref{sec4.1}). In such cases, costs can be set using any real value, which leads to constructing cost-dependent models. When the costs change, retraining models will take additional computation. In contrast, the MOBA model, which does not rely on cost information, only sets two abstention parameters in the range of (0,1). \item the MOBA model is robust to varying conditions and demands since a set of Pareto-optimal vectors (corresponding to a set of abstaining classifiers) are generated. As explained in Section~\ref{sec2.4}, the optimal abstaining classifier can be selected according to the demands (e.g., obtaining a maximum F-measure or a minimum error rate under fixed abstentions) or conditions (whether the costs are known). Furthermore, if the costs change over time, no retraining of the MOBA model is required as a new optimal abstaining classifier can be determined simply by recalculating the expected costs. \begin{table}[!htb] \centering \caption{Characteristics of datasets used in this study} \begin{tabular*}{\textwidth}{l@{\extracolsep{\fill}}cccc} \toprule Dataset & Instances & Positive & Negative & Attrsibutes \\ \midrule ionosphere & 351 & 126 & 225 & 34 \\ pima & 768 & 268 & 500 & 8 \\ credit-g & 1000 & 300 & 700 & 20 \\ ecoli3 & 336 & 35 & 301 & 7 \\ hepatitis & 155 & 32 & 123 & 19 \\ haberman & 306 & 81 & 225 & 3 \\ cmc & 1473 & 333 & 1140 & 9 \\ transfusion & 748 & 178 & 570 & 4 \\ Australian & 690 & 307 & 383 & 14 \\ \bottomrule \end{tabular*}% \label{t2}% \end{table}% \item compared to the BA model, the MOBA model can control the respective performance of two classes via the class-dependent abstention constraints, so the MOBA model is more applicable to imbalanced datasets or cost-sensitive problems. The BA model has an overall reject rate constraint. When datasets are imbalanced, the reject rates of two classes may be imbalanced even though the overall reject rate constraint is satisfied. When the two abstention parameters are set using the same value, the MOBA model can avoid imbalanced reject rates because of its class-dependent abstention constraints. In addition, the two abstention parameters in the MOBA model can have different values when dealing with imbalanced datasets. \end{enumerate} \section{Experimental results} \label{sec4} In the experiments, the proposed MOBA model was compared with two abstaining classification models: one considering costs (Section~\ref{sec4.1}) and the other one not (Section~\ref{sec4.2}). Table~\ref{t2} lists the datasets used in the experiments, which are available in the KEEL-dataset repository \cite{alcala2011keel}. Among these datasets, there are several ones associated with cost-sensitive classification tasks, such as \textit{pima} and \textit{credit-g}. For \textit{pima}, the task is to predict whether the patient is diabetic. The cost of missing a diabetic patient is higher than that of misdiagnosing a nondiabetic patient. \textit{Credit-g} is a dataset in the financial area, which classifies people as good or bad credit customers. It is worser to classify bad credit customers as good than the opposite case. Each dataset was divided into three distinct subsets: the training set (containing 60\% of the examples) used to generate confidence scores by training a scoring classifier, and the validation and test sets, each of which contained 20\% of the examples. The validation set was used to determine the rejection thresholds while the classification performance was evaluated via the test set. All the experimental results were obtained using MATLAB R2017a. \subsection{Comparison of the results when the costs are considered} \label{sec4.1} \begin{table*}[!htb] \centering \caption{Cost Models} \begin{tabular*}{\textwidth}{@{\extracolsep{\fill}}llllll} \toprule & CTP/N & CFP & CFN & CRP & CRN \\ \midrule CM1 & U[-10,0] & U[0,50] & U[0,50] & 1 & 1 \\ CM2 & U[-10,0] & U[0,100] & U[0,50] & 1 & 1 \\ CM3 & U[-10,0] & U[0,50] & U[0,100] & 1 & 1 \\ CM4 & U[-10,0] & U[0,50] & U[0,50] & U[0,30] & U[0,30] \\ \bottomrule \end{tabular*}% \label{t3}% \end{table*}% In this section, the results of the MOBA model that accounted for the costs are evaluated in comparison to those from the model by Tortorell in which an ROC convex hull (ROCCH) curve was constructed from the confidence scores and the rejection thresholds were determined based on the tangents of the ROCCH curve \cite{tortorella2004reducing,tortorella2000optimal,santos2005optimal}. In Tortorell's model, when the condition \begin{equation} \frac{CTN-CRN}{CFN-CRP} > \frac{CFP-CRN}{CTP-CRP} \end{equation} was not satisfied, the reject option could not be activated. That is to say, the traditional classifier without rejection could provide the minimal cost. In this experiment, the twin support vector machine (TWSVM) \cite{khemchandani2007twin} was used as the scoring classifier to generate confidence scores \cite{lin2017twin}. Four cost models \cite{lin2017twin} shown in Table~\ref{t3} were used, where U[a,b] denotes a uniform distribution over the interval [a,b]. Note that while CRP and CRN had equal values in CM4 in~\cite{lin2017twin}, in this experiment, CRP and CRN had different values in CM4 as class-dependent reject costs were considered. A Wilcoxon rank sum test \cite{lin2017twin,tortorella2004reducing} was performed to compare the two cost-related abstaining models. The details of the Wilcoxon rank sum test can be found in~\cite{lin2017twin}. In this test, 1000 cost matrices (CTP/N, CFP/N, and CRP/N) were generated for each cost model in Table~\ref{t3}. Then, for each cost matrix, the expected cost was computed as per Eq.~(\ref{eq3}). Finally, the numbers of cases where the cost of the MOBA model was lower, higher, or identical compared to the cost of Tortorell's model were counted. There were two scenarios that resulted in identical costs: 1) the costs of the compared methods were equal; or 2) for a certain cost matrix, the reject option in Tortorell's model was not activated, in which situation, no MOBA model was constructed. The hyperparameters in the MOBA model were set as follows. First, Tortorell's model was enforced and the reject rates with respect to the two classes were obtained. Then, the values corresponding to the reject rates were assigned to $p_{max}$ and $n_{max}$. The two reject rates obtained via Tortorell's model can be regarded as good candidates to avoid blindly setting the hyperparameters in the MOBA model. \begin{table*}[!htb] \centering \caption{Results of the Wilcoxon rank sum test for comparing MOBA and Tortorell's model} \begin{tabular*}{\textwidth}{l@{\extracolsep{\fill}}cccclcccc} \toprule & CM1 & CM2 & CM3 & CM4 & & \multicolumn{1}{c}{CM1} & \multicolumn{1}{c}{CM2} & \multicolumn{1}{c}{CM3} & \multicolumn{1}{c}{CM4} \\ \midrule \multicolumn{1}{l}{ionosphere} & \textbf{744} & \textbf{749} & \textbf{781} & 471 & \multicolumn{1}{l}{haberman} & \multicolumn{1}{c}{\textbf{466}} & \multicolumn{1}{c}{393} & \multicolumn{1}{c}{\textbf{523}} & \multicolumn{1}{c}{\textbf{484}} \\ & 158 & 203 & 182 & 56 & & \multicolumn{1}{c}{436} & \multicolumn{1}{c}{\textbf{559}} & \multicolumn{1}{c}{440} & \multicolumn{1}{c}{43} \\ & 98 & 48 & 37 & \textbf{473} & & \multicolumn{1}{c}{98} & \multicolumn{1}{c}{48} & \multicolumn{1}{c}{37} & \multicolumn{1}{c}{473} \\ \multicolumn{1}{l}{pima} & \textbf{459} & \textbf{597} & \textbf{524} & 412 & \multicolumn{1}{l}{cmc} & \multicolumn{1}{c}{93} & \multicolumn{1}{c}{45} & \multicolumn{1}{c}{135} & \multicolumn{1}{c}{338} \\ & 443 & 350 & 439 & 115 & & \multicolumn{1}{c}{65} & \multicolumn{1}{c}{23} & \multicolumn{1}{c}{67} & \multicolumn{1}{c}{57} \\ & 98 & 53 & 37 & \textbf{473} & & \multicolumn{1}{c}{\textbf{842}} & \multicolumn{1}{c}{\textbf{932}} & \multicolumn{1}{c}{\textbf{798}} & \multicolumn{1}{c}{\textbf{605}} \\ \multicolumn{1}{l}{credit-g} & \textbf{489} & \textbf{516} & 480 & 444 & \multicolumn{1}{l}{transfusion} & \multicolumn{1}{c}{\textbf{621}} & \multicolumn{1}{c}{\textbf{619}} & \multicolumn{1}{c}{\textbf{529}} & \multicolumn{1}{c}{\textbf{484}} \\ & 413 & 436 & \textbf{483} & 83 & & \multicolumn{1}{c}{281} & \multicolumn{1}{c}{333} & \multicolumn{1}{c}{434} & \multicolumn{1}{c}{43} \\ & 98 & 48 & 37 & \textbf{473} & & \multicolumn{1}{c}{98} & \multicolumn{1}{c}{48} & \multicolumn{1}{c}{37} & \multicolumn{1}{c}{473} \\ \multicolumn{1}{l}{ecoli3} & \textbf{738} & \textbf{531} & \textbf{522} & 438 & \multicolumn{1}{l}{Australian} & \multicolumn{1}{c}{\textbf{659}} & \multicolumn{1}{c}{\textbf{711}} & \multicolumn{1}{c}{\textbf{664}} & \multicolumn{1}{c}{\textbf{473}} \\ & 164 & 416 & 441 & 89 & & \multicolumn{1}{c}{243} & \multicolumn{1}{c}{241} & \multicolumn{1}{c}{299} & \multicolumn{1}{c}{54} \\ & 98 & 53 & 37 & \textbf{473} & & \multicolumn{1}{c}{98} & \multicolumn{1}{c}{48} & \multicolumn{1}{c}{37} & \multicolumn{1}{c}{\textbf{473}} \\ \multicolumn{1}{l}{hepatitis} & \textbf{510} & \textbf{520} & \textbf{514} & 435 & & & & & \\ & 392 & 432 & 449 & 92 & & & & & \\ & 98 & 48 & 37 & \textbf{473} & & & & & \\ \bottomrule \end{tabular*}% \label{t4}% \end{table*}% The results of the Wilcoxon rank sum test are shown in Table~\ref{t4}. Note that in each scenario (each dataset in each cost model), the three figures from top to bottom represent the numbers of lower, higher, and identical costs in the MOBA model compared to the costs obtained via Tortorell's model, respectively. In the table, it can be seen that there were a greater number of lower-cost cases for the MOBA model in almost all of the scenarios in the CM1, CM2, and CM3 cost models, while for most scenarios in the CM4 cost model, the identical-cost case was the most frequent. This is because the number of inactivated reject options increased. In the remaining two cases (i.e., in the lower- and higher-cost cases), the MOBA model provided a lower cost than Tortorell's model in the vast majority of cases. In addition, for a certain cost matrix, the reject option may not be activated in Tortorell's model, whereas the rejection could still be enforced in the MOBA model. This is an advantage of the MOBA model over Tortorell's model; that is, the MOBA model is not subject to a certain cost matrix. \subsection{Comparison of the results when the costs are not considered} \label{sec4.2} In this section, the results of the MOBA model that did not consider cost information are presented in comparison to those of the BA model \cite{pietraszek2007use}. In the BA model, the cost ratio between CFN to CFP was set to one, which means that the CFN and CFP were the same. For this reason, the values of $p_{max}$ and $n_{max}$ in the MOBA model were set to the same values. When $p_{max} = n_{max}$, the overall reject rate $k_{max}$ in (\ref{eq4}) was equal to the reject rate of each class ($p_{max}$ or $n_{max}$), which ensures the comparability of the BA and MOBA models. The abstention parameters were set from 0.01 to 0.3 in steps of 0.02 since larger rejection rates are usually of no significance in practical applications \cite{simeone2012design}. For models that do not consider the costs, the classification results are typically evaluated based on the corresponding performance-rejection curves. Here, the accuracy (ACC), AUC, and G-mean (G) were used as the evaluation metrics. Note that the G-mean is the geometric mean of the sensitivity and specificity. That is, ACC-Rej, AUC-Rej, and G-Rej curves were obtained using each dataset. In this paper, only the trade-off curves of datasets \textit{pima}, \textit{haberman}, \textit{cmc}, and \textit{transfusion} are shown (Figures~\ref{f3}-\ref{f6}) to discuss the performance of the two compared models. Similar trade-off curves were obtained for the other datasets. \begin{figure*}[!htb] \centering \subfigure[ACC-Rej]{ \includegraphics[width=0.3\textwidth]{Fig3a} } \subfigure[AUC-Rej]{ \includegraphics[width=0.3\textwidth]{Fig3b} } \subfigure[G-Rej]{ \includegraphics[width=0.3\textwidth]{Fig3c} } \caption{Performance-rejection curves of BA and MOBA models using \textit{pima} dataset}\label{f3} \end{figure*} \begin{figure*}[!htb] \centering \subfigure[ACC-Rej]{ \includegraphics[width=0.3\textwidth]{Fig4a} } \subfigure[AUC-Rej]{ \includegraphics[width=0.3\textwidth]{Fig4b} } \subfigure[G-Rej]{ \includegraphics[width=0.3\textwidth]{Fig4c} } \caption{Performance-rejection curves of BA and MOBA models using \textit{haberman} dataset}\label{f4} \end{figure*} \begin{figure*}[!htb] \centering \subfigure[ACC-Rej]{ \includegraphics[width=0.3\textwidth]{Fig5a} } \subfigure[AUC-Rej]{ \includegraphics[width=0.3\textwidth]{Fig5b} } \subfigure[G-Rej]{ \includegraphics[width=0.3\textwidth]{Fig5c} } \caption{Performance-rejection curves of BA and MOBA models using \textit{cmc} dataset}\label{f5} \end{figure*} \begin{figure*}[!htb] \centering \subfigure[ACC-Rej]{ \includegraphics[width=0.3\textwidth]{Fig6a} } \subfigure[AUC-Rej]{ \includegraphics[width=0.3\textwidth]{Fig6b} } \subfigure[G-Rej]{ \includegraphics[width=0.3\textwidth]{Fig6c} } \caption{The performance-rejection curves of BA and MOBA models using \textit{transfusion} dataset}\label{f6} \end{figure*} Overall, the MOBA model provided a better trade-off between the performance (accuracy, AUC, and G-mean) and rejection than the BA model. In other words, for a fixed reject rate, the MOBA model provided higher values for the accuracy, AUC, and G-mean than those of the BA model. For the MOBA model, the values of the accuracy, AUC, and G-mean grew with increasing reject rates, while for the BA model, only the ACC-Rej curves exhibited an increasing trend and the AUC and G-mean values decreased in Figures~\ref{f4}-\ref{f6}. This is because the BA model minimized the error rate under an overall reject rate and did not consider the class-dependent performance. According to the definitions, the values of the AUC and G-mean were large only when both sensitivity and specificity were large. Thus, the MOBA model was shown to provide more balanced sensitivity and specificity values. The fluctuations observed in the trade-off curves may be attributed to the small sample size and randomness when constructing the models. \section{Conclusions} \label{sec5} In this paper, the MOBA model was proposed for cost-sensitive problems where the error costs are usually unknown and asymmetric. The MOBA model avoids setting cost information by optimizing essential metrics. A significant advantage of the MOBA model is its robustness towards different conditions, such as known or unknown cost information, evolving costs over time, and using different metrics to evaluate performance. Also, the MOBA model can accommodate unbalanced classification problems due to its ability to optimize the respective performance of two classes under class-dependent abstention constraints \cite{he2008learning,xiao2017class}. Experimental results in this study have shown that the MOBA model performed better than previous models. When the costs were known, the MOBA model obtained a greater number of lower-cost cases in the Wilcoxon rank sum test than did Tortorell's model. When costs were not considered, the MOBA model achieved better trade-offs between performance (accuracy, AUC, and G-mean) and abstention than the BA model. Abstaining classification has wide applications in safety-critical fields, such as medical diagnosis, fault detection, credit assessment, and so on. Rejection means it is difficult to make a definite decision given the current knowledge. More knowledge should be provided to reduce misclassification costs. The proposed MOBA model will show a promising utility owing to its advantages and better performance. In the future, we intend to extend the MOBA model to multi-class problems and apply the MOBA model to specific applications. \section*{Declarations of interest} None. \section*{References} \bibliographystyle{model1-num-names}

1. Introduction {#sec1} =============== Neonatal varicella is a rare condition, with the incidence of 2--6 per 100,000 live births per year \[[@B1]\]. Mostly, neonatal varicella is caused from maternal varicella, and the presentation of the disease is variable, ranging from mild symptoms to varicella pneumonia \[[@B2], [@B3]\], hepatitis \[[@B2], [@B4]\], or meningoencephalitis in disseminated disease \[[@B5], [@B6]\]. The worst outcome could be fatal \[[@B2], [@B3]\]. As maternal viremia happens, the transplacental infection occurs, which may start up to 48 hours before maternal rash develops. In neonates born to susceptible mothers who develop varicella symptoms from 4 to 5 days antepartum to 2 days postpartum, the mortality rate of the baby may reach 20--30% \[[@B1]--[@B3], [@B5]\]. Therefore, newborns whose mothers develop varicella 5 days prior to 2 days after birth should receive varicella zoster immunoglobulin (VZIG) in no time. Intravenous immunoglobulin (IVIG) can be used instead of VZIG if it is not available, which can be seen in many hospitals in Thailand. Moreover, the infant should be treated with intravenous acyclovir when lesions develop \[[@B2]\]. However, evidence supporting the use of acyclovir in neonatal varicella is limited to case reports and case series \[[@B7]\]. This case we are presenting is expected to be a supporting evidence of neonatal varicella treatment with IVIG and acyclovir. 2. Case Presentation {#sec2} ==================== A term male baby was born to a 27-year-old Thai female by caesarean section due to prolonged second stage of labor at 39 + 3 weeks in Trang Hospital, Thailand. All physical examinations of the baby were within normal limits. The baby\'s vital signs were as follows: respiratory rate 58 breaths/min, blood pressure 63/33 mmHg, heart rate 140 beats per minute, and temperature 36.8°C. One day after the delivery, the mother developed maculopapular rash over her trunk, suggestive of chicken pox. The rash was discovered two days after the delivery, and the decision was made to move the baby to the isolation room of newborn nursery ward. There was no rash on the baby\'s skin on the transfer date. There were 6 family members living together: the baby\'s mother, father, two other siblings, grandmother, and aunt. The baby\'s grandmother had herpes zoster for an unknown onset. Few weeks prior to the baby\'s birth, his aunt developed low-graded fever with vesicular skin eruptions, without seeking any medical care. In the family members\' opinion, they thought it was likely to be varicella. The skin rash disappeared prior to the baby\'s birth. This history was not given on admission. For treatment, IVIG 400 mg/kg/day was infused intravenously, started on the second day of life, the day when the baby was transferred to the isolation room. Other supportive treatments were also given. When the IVIG infusion was finished, intravenous acyclovir 25 mg/kg/day divided for every 8 hours was planned for seven days, started on the baby\'s second day of life. Expressed breast milk was collected in containers to restrict contact with mother. A day later, on the baby\'s third day of life, discrete vesicular rashes on erythematous background with discrete erythematous maculopapular rashes were found over the baby\'s trunk, arms, and legs, as shown in Figures [1(a)](#fig1){ref-type="fig"}--[1(c)](#fig1){ref-type="fig"}. The baby was clinically diagnosed with neonatal varicella. The rash completely disappeared on the baby\'s fifth day of life. The baby was discharged on his ninth day of life, when acyclovir infusion of seven days was completed. During hospital admission, he was afebrile and doing well. The follow-up appointment with a pediatrician was made at 2 weeks after hospital discharge. On the follow-up visit, the baby was healthy, well-fed, and did not show any signs of complications. 3. Discussion {#sec3} ============= Neonatal varicella is not a commonly encountered disease, yet it is an important disease we should be able to detect early because it can be fatal in newborns \[[@B2], [@B3]\]. Varicella-zoster virus is the causative agent of this disease. It can be spread to another person by direct contact or through inhalation of aerosols from vesicular fluid of skin lesions, either from acute varicella or from zoster. The infection can also spread from aerosolized respiratory secretions \[[@B8]\]. The diagnosis of neonatal varicella is made clinically. In this case, the infection was caused by infected family members who were always in close contact with the patient\'s mother. Nobody in the family knew that this virus could be transmitted transplacentally and result in neonatal varicella. Truthfully, if the information of the infected family was given at first, the mother would have been admitted to the isolation room because the risk of developing varicella during hospital admission was very high \[[@B9]\]. IVIG was given to the baby due to unavailability of VZIG in the hospital\'s medication list, which was also not available at some other hospitals in Thailand. From a previous study by Huang et al. \[[@B10]\], the combination of IVIG and intravenous acyclovir was administered to newborn infants whose mothers had developed varicella rash within 7 days before and 5 days after birth, and none of them developed varicella. In this case, the baby developed rash despite getting IVIG and acyclovir. After the baby was intravenously given IVIG and acyclovir, no complications occurred. The baby recovered fully and was discharged with no problems. In conclusion, the combination of IVIG and acyclovir might not effectively prevent neonatal varicella in this case, but the medication could prevent the baby from serious complications and shorten the clinical course. I would like to express my special thanks of gratitude to Dr. Paweena Booncharoen, whose kind supervision and valuable support have been the ones that help me completing my project. IVIG: : Intravenous immunoglobulin kg: : Kilogram mg: : Milligram VZIG: : Varicella zoster immunoglobulin. Ethical Approval ================ This study has been approved by the research ethics committee of Trang Hospital. It has been reviewed and approved by Committee on Human Rights related to researches involving human subjects, based on the Declaration of Helsinki. Consent ======= Patient consent has been obtained from the patient\'s father in Thai language. Disclosure ========== This study was presented as an oral presentation in AAD Annual Meeting 2019 and as an e-poster in World Congress of Dermatology 2019. Conflicts of Interest ===================== The author has no conflicts of interest to declare. {#fig1} [^1]: Academic Editor: Peter Olumese

\section*{acknowledgments} The author would like to thank Daniel Boyanovsky for useful discussions. This work was supported by Korea Science and Engineering Foundation under Grant No. 951-0207-056-2.

\section{Introduction} \label{intro} In the last two decades modern fabrication techniques made it possible to connect quantum dot systems with superconducting leads. Quantum dots were realized with carbon nanotubes \cite{Buitelaar2003,Vecino2004,Eichler2007,Grove-Rasmussen2009,Dirks2009, Herrmann2010, Pillet2010}, metallic particles \cite{Ralph}, semiconducting nanowires \cite{vanDam2006,Doh2008,Hofstetter2009,Herrmann2012arXiv}, single fullerene molecules \cite{Winkelmann2009}, self-assembled nanocrystals \cite{Katsaros2010} and graphene quantum dots \cite{Dirks2011}. The experiments show a gap in the Coulomb diamonds which is proportional to the superconducting gap, reflecting the BCS-density of states. In the sequential tunneling regime higher order quasiparticle tunneling processes are suppressed and current flows due to single quasiparticle tunneling. First transport theories were presented \cite{Whan1996}, using a master equation approach, where the rates were calculated on the basis of Fermi's golden-rule. Another method based on non-equilibrium Green's function was used by Yeyati \textit{ et al.} \cite{Levy_Yeyati1996} and Kang \cite{Kang1997} to describe resonant tunneling through an effective single level quantum dot in the limit of very strong Coulomb repulsion in the dot ($U \to \infty$ limit), where transport is governed by quasiparticle tunneling; the corresponding I-V curves show an intrinsic broadening of the BCS-like feature in the current in agreement with experimental observation \cite{Ralph}. For small Coulomb repulsion, higher order processes lead to Josephson current\cite{vanDam2006} and Andreev reflections \cite{Buitelaar2003,Vecino2004,Eichler2007,Doh2008,Grove-Rasmussen2009, Pillet2010,Dirks2011}, which appear as subgap features in the experiments. Both effects were studied intensely experimentally and theoretically \cite{Levy_Yeyati1996,Pala2007,Eichler2007,Governale2008 and were recently summarized in review articles of Refs. \onlinecite{DeFranceschi2010,Martin-Rodero2011}. Besides Andreev reflections also the Kondo effect \cite{Winkelmann2009} as well as Yu-Shiba-Rusinov bound states \cite{Grove-Rasmussen2009,Andersen2011,Franke2011} can lead to subgap features and are the subject of current research. If the temperature becomes comparable with the superconducting gap quasiparticles can get thermally excited across the gap, leading to additional subgap features \cite{Whan1996}. In the following we present a microscopic theory for transport through superconducting hybrid nanojunctions for finite superconducting gap $|\Delta| < \infty$ in the sequential tunneling limit. In particular, we trace out all degrees of freedom of the superconducting leads to obtain a generalized master equation for the reduced density matrix to lowest order in the tunneling Hamiltonian. {We differentiate from Ref.~\onlinecite{Whan1996} by going beyond the constant interaction implicitly used there, and from Refs. \onlinecite{Levy_Yeyati1996} and \onlinecite{Kang1997} since we also treat subgap features associated to many-body excitations of a quantum dot molecule (double quantum dot).} In contrast to Green's function techniques, see e.g. Ref. \onlinecite{Martin-Rodero2011}, this method enables one to treat the interactions on the system \textit{exactly}. Moreover, as shown on the example of a double quantum dot, our theory is easily scalable and allows an exact treatment of the Coulomb interaction and can treat any quantum dot set-up. Hence, we can describe lowest order quasiparticle transport of experimental relevant quantum dot systems (multiple quantum dots or multilevel quantum dots). We focus on transport involving thermally excited quasiparticles, and show that excited states of the quantum dot system can be observed in the current voltage spectroscopy in the Coulomb blockade region. Though transitions between two ground states are blocked due to the gap in the BCS-density of states, thermally excited quasiparticles can participate in transport through excited system states, giving a source of subgap features in superconducting hybrid systems. These subgap features are already present in lowest order of the perturbation theory, in contrast to Cooper pair transport which occurs only in fourth order in the tunneling coupling. Nevertheless, experiments suggest the existence of a regime in which quasiparticle transport dominates also in the subgap region \cite{DeFranceschi2010}. For a quantum dot coupled to a normal and a superconducting lead, a possible explanation for the subgap features observed in Ref. \onlinecite{Dirks2009} is given, where a carbon nanotube quantum dot is coupled to a normal and a superconducting contact. The paper is organized as follows: In Sect.~\ref{sect:model hamiltonian} we introduce the Hamiltonian in a system-bath model using a number conserving version of the Bogoliubov-Valatin transformation \cite{Josephson1962,Bardeen1962}. We describe the electrons of the superconducting leads as a combination of quasiparticle excitations of the BCS-ground state and Cooper pairs. For this purpose we introduce Cooper pair creation and annihilation operators. The explicit inclusion of these operators allows one to construct a theory which conserves the particle number in the tunneling process. In this way, for example, anomalous contributions to the tunneling rates due to Cooper pairing naturally vanish in second order. In Sect.~\ref{sect:transport theory and the gme}, the generalized master equation for the reduced density matrix is derived and used to calculate the current. In Sect.~\ref{sect:transport through multiple quantum dot devices} we apply the theory to the calculation of transport characteristics of two systems: the single level quantum dot (SD) and the double quantum dot (DD), the latter in two possible configurations cf. Fig.~\ref{fig:DD_configuration}. The SD is used to explain basic phenomena such as a gap opening in the Coulomb diamonds which is proportional to the superconducting gap, and transport involving thermally excited quasiparticles \cite{Whan1996}. On the other hand, the DD possess a richer many-body spectrum with several excited states. We visualize transitions through excited system states in the low bias regime using thermally excited quasiparticles. Due to the gap in the BCS-density of states, the ground state to ground state transition is not allowed in all cases, leading to transport through excited system states, appearing as peaks \textit{in} the Coulomb blockade region. The threshold for observing excited system states in the subgap region is that the energy difference between the excited state and its ground state must be smaller than $2|\Delta|$. We confirmed this threshold by means of the independently gated DD, where the detuning of the two sites changes the level spacing. Finally the N-QD-S system is investigated, where a quantum dot is coupled to a normal and a superconducting lead. In this case only the superconducting lead produces thermal lines in the Coulomb blockade region, giving a possible explanation for the subgap features in Ref. \onlinecite{Dirks2009}. \section{Model Hamiltonian}\label{sect:model hamiltonian} In the following we consider quantum dot systems weakly coupled to two superconducting leads. The total Hamiltonian is written in a system-bath model: \begin{equation} \hat H = \hat H_S + \hat H_B + \hat H_T, \end{equation} where $\hat H_S$ represents the Hamiltonian of the quantum dot system, $\hat H_B$ is the Hamiltonian of the superconducting leads, and $\hat H_T$ describes the tunneling between the system and the leads. Specifically, we focus on two systems, a single level quantum dot (SD) and a double quantum dot (DD). The SD has been the focus of many theoretical works before \cite{Whan1996,Levy_Yeyati1996,Kang1997,Pala2007,Governale2008}, and we use its simple Fock-space structure to demonstrate some generic effects resulting from the superconducting leads. \begin{figure} \includegraphics[width=0.75\columnwidth]{fig1} \caption{Sketch of the transport set-up of a double quantum dot (DD) coupled to superconducting leads. The DD is illustrated in the parallel (top panel) and serial (bottom panel) configuration. Tunneling events are depicted by arrows.} \label{fig:DD_configuration} \end{figure} We describe the SD by the single impurity Anderson model: \begin{equation}\label{eq:SD} \hat H_{SD} = \sum_{\sigma} \epsilon_d \dddag{\sigma}\dan{\sigma} + U \n{\uparrow}\n{\downarrow}, \end{equation} where $\n{\sigma} = \dddag{\sigma}\dan{\sigma}$ is the number operator of the electrons on the dot with spin $\sigma$. This model describes a quantum dot with on-site energy $\epsilon_d$ and Coulomb repulsion $U$ which can be occupied by at most two electrons. The highest occupied state is defined as $\ket{2} = \dddag{\uparrow} \dddag{\downarrow}\ket{0}$, the 1-particle states are defined as $\ket{1\sigma} = \dddag{\sigma}\ket{0}$, and $\ket{0}$ is the state with zero particles. For the DD we use a modified version of the Pariser-Parr-Pople Hamiltonian \cite{Pariser1953,Pople1953}: \begin{equation}\label{eq:DD hamiltonian} \begin{split} & \hat H_{DD} = \sum_{\substack{\alpha \in \{1,2 \} \\ \sigma \in \{ \uparrow,\downarrow\} }} \epsilon_{\alpha \sigma} \dddag{\alpha \sigma} \dan{\alpha \sigma} + \sum_{\sigma}\biggl( b \dddag{1 \sigma}\dan{2\sigma} + b^* \dddag{2\sigma} \dan{1\sigma} \biggr) \\ & + \sum_{\alpha} U_{\alpha} \biggl(\n{\alpha \uparrow} - \frac{1}{2} \biggr) \biggl(\n{\alpha \downarrow} - \frac{1}{2} \biggr) + V \bigl(\n{1} -1 \bigr)(\n{2} - 1). \end{split} \end{equation} Here, $\dddag{\alpha \sigma}$ are the creation operators for an electron on site $\alpha \in \{1,2\}$ with spin $\sigma$. They define the number operators $\n{\alpha \sigma} = \dddag{\alpha \sigma}\dan{\alpha \sigma} $. The operator $\n{\alpha} = \n{\alpha \uparrow } + \n{\alpha \downarrow}$ counts the number of electrons on site $\alpha$. In the general case we distinguish between the four on-site energies $\epsilon_{\alpha \sigma}$ and between the on-site Coulomb interactions $U_{\alpha}$. Electrons on different sites interact through the inter-dot Coulomb interaction $V$; $b$ describes the hopping between the two sites. In our set-up the on-site energies can be controlled by capacitively coupled gate electrodes. In the case of site-independent on-site energies and on-site Coulomb interaction the Hamiltonian can be diagonalized analytically \cite{Bulka2004,Hornberger2008}. The superconducting leads are described by the mean field form, $\hat H_B^{\text{MF}}$ of the pairing Hamiltonian, where we additionally inserted a unity represented by a product of Cooper pair annihilation and creation operators, \mbox{$\operatorname{\hat S}_\eta\operatorname{\hat S}^{\dagger}_\eta=1$}, which will be specified later in Sec.~\ref{sect:diagonalization_lead}. We find \begin{equation}\label{eq:MF} \begin{split} & \hat H_B^{\text{MF}} = \sum_{\eta k\sigma} \xi_{\eta k} \cdag{\eta k \sigma} \can{\eta k \sigma} +\sum_{\eta} \mu_{\eta} \hat N_{\eta}\\ & + \sum_{\eta k} \bigl ( \Delta_{\eta} \cdag{\eta k \uparrow} \cdag{\eta -k \downarrow} \operatorname{\hat S}_{\eta} +\Delta_{\eta}^* \operatorname{\hat S}^{\dagger}_{\eta} \can{\eta -k \downarrow} \can{\eta k \uparrow} \bigr)\\ & = \hat H_G +\sum_\eta \mu_\eta \hat N_\eta, \end{split} \end{equation} where $\xi_{\eta k} = \epsilon_{k} - \mu_{\eta}$ measures single particle energies $\epsilon_k$ with respect to the electrochemical potential $\mu_\eta$, and $\hat N_{\eta} = \sum_{k \sigma} \cdag{\eta k \sigma} \can{\eta k \sigma}$ counts the number of electrons in lead $\eta$. Finally, $\Delta_{\eta} = |\Delta_\eta|e^{i\phi_\eta} \equiv - \sum_l V_{lk} \braket{\operatorname{\hat S}^{\dagger}_\eta \can{\eta -k \downarrow} \can{\eta k \uparrow}}$ denotes the superconducting gap of lead $\eta$. Here $\braket{\bullet}$ denotes a thermal average calculated self-consistently using the mean field Hamiltonian of Eq.~(\ref{eq:MF}). The tunneling Hamiltonian, \begin{equation}\label{eq:tunneling Hamiltonian} \hat H_T = \sum_{\eta k \sigma \alpha} t_{\eta \alpha \sigma} \cdag{\eta k \sigma} \dan{\alpha \sigma} + t^*_{\eta \alpha \sigma} \dddag{\alpha \sigma} \can{\eta k \sigma}, \end{equation} describes the tunneling between the leads and the two sites of the DD, where the tunneling coefficients $t_{\eta \alpha \sigma}$ depend on the lead, site, and spin index. Depending on the choice of the tunneling coefficients the DD is described in parallel or in serial configuration, see Fig.~\ref{fig:DD_configuration}. For the single dot we skip the index $\alpha$ in Eq.~(\ref{eq:tunneling Hamiltonian}), as only one site is involved. \subsection{Diagonalization of the lead Hamiltonian}\label{sect:diagonalization_lead} The most famous way to diagonalize the mean field Hamiltonian, $\hat H_B^{\text{MF}}$, of Eq.~(\ref{eq:MF}) was first introduced by Bogoliubov \cite{Bogoliubov1958}. We are following Josephson and Bardeen \cite{Josephson1962,Bardeen1962} who modified the so called Bogoliubov transformation in a number conserving way. We adopt this idea and define the Bogoliubov transformation: \begin{equation}\label{eq:BT} \cdag{\eta k \sigma} = u_{\eta k } \gdag{\eta k \sigma} + \operatorname{sgn}{\sigma}\, v_{\eta k}^* \gan{\eta -k \bar \sigma} \operatorname{\hat S}^{\dagger}_{\eta}, \end{equation} where $\bar\sigma = -\sigma$. In Eq.~(\ref{eq:BT}) $\gdag{\eta k \sigma}$ creates a fermionic quasiparticle, often called bogoliubon, which is defined by \begin{equation} \{\gdag{\eta k \sigma}, \gan{\eta'k'\sigma'}\} = \delta_{\eta \eta'} \delta_{kk'} \delta_{\sigma \sigma'}, \end{equation} \begin{equation}\label{eq: QPV} \gan{\eta k \sigma} \ket{\text{GS}}_{\eta} = 0. \end{equation} Here $\ket{\text{GS}}_{\eta} $ denotes the ground state, or Cooper pair condensate of lead $\eta$ \cite{note_groundstate}. Bogoliubons are quasiparticle excitations of the Cooper pair condensate, meaning that the Cooper pair condensate is defined as the vacuum state of the bogoliubons, see Eq.~(\ref{eq: QPV}). The coefficients $u_{\eta k}$ and $v_{\eta k }$ are complex numbers and fulfill: \begin{equation}\label{eq:u2+v2} |u_{\eta k}|^2 +|v_{\eta k}|^2 =1. \end{equation} They read: \begin{equation}\label{eq:def_uk} u_{\eta_k} = \sqrt{ \frac{1}{2}\biggl( 1 + \frac{\xi_{\eta k}}{|E_{\eta k}|} \biggr) }, \end{equation} \begin{equation}\label{eq:def_vk} v_{\eta k} = e^{i\phi_{\eta}} \sqrt{ \frac{1}{2}\biggl( 1 - \frac{\xi_{\eta k}}{|E_{\eta k}|} \biggr) }, \end{equation} where $\phi_{\eta}$ is the phase of the superconducting gap $\Delta_{\eta}$. In the number conserving description, the Hamiltonian of Eq.~(\ref{eq:MF}) commutes with the particle number operator. Hence, it is required that the ground state must be an eigenstate of the particle number operator. We define the ground state of lead $\eta$ as \cite{Schrieffer,Ambegaokar} $\ket{\text{GS}}_{\eta} = \ket{0, {N}}_{\eta}$, where $\ket{0, {N}}_{\eta}$ represents a state with $N/2$ Cooper pairs and zero quasiparticle excitations. The Cooper pair annihilation operator $\operatorname{\hat S}_{\eta}$ annihilates a Cooper pair in lead $\eta$ and can formally be defined as\cite{Schrieffer}: \begin{equation}\label{eq:definition_s} \begin{split} \operatorname{\hat S}_{\eta} \ket{0,N}_{\eta} & = \ket{0,N-2}_\eta, \\ \operatorname{\hat S}_{\eta} \ket{{k\sigma}, N}_{\eta} &= \ket{{k\sigma},N-2}_\eta, \\ \gdag{k\sigma} \ket{0,N}_{\eta} &= \ket{{k\sigma},N}_\eta. \end{split} \end{equation} Eq.~(\ref{eq:definition_s}) implies that the Cooper pairs and the quasiparticles are decoupled: \begin{equation} \big[ \operatorname{\hat S}^{\dagger}_{\eta} , \gdag{k\sigma} \big] = 0, \quad \big[\operatorname{\hat S} , \gdag{k\sigma} \big] = 0, \end{equation} and the Cooper pair operators have the following properties, see App.~\ref{app:bcs GS CPO}: \begin{equation} \operatorname{\hat S}_{\eta} \operatorname{\hat S}^{\dagger}_{\eta}= 1, \quad \big[\operatorname{\hat S}_{\eta},\operatorname{\hat S}^{\dagger}_{\eta}] = \hat{\mathcal{P}}_{0,\eta}, \end{equation} where $\hat{\mathcal{P}}_0$ is the projector on states with zero Cooper pairs, and \begin{equation} \big[ \hat N , \operatorname{\hat S}^{\dagger} \big] = 2\operatorname{\hat S}^{\dagger}. \end{equation} Note that the transformation defined in Eq.~(\ref{eq:BT}) conserves the fermionic properties of the electron operators only if we restrict our Hilbert space to a subspace with more than zero Cooper pairs. In that subspace $\operatorname{\hat S}$ commutes with $\operatorname{\hat S}^{\dagger}$ and the Bogoliubov transformation is well defined. Applying the transformation of Eq.~(\ref{eq:BT}) on Eq.~(\ref{eq:MF}) we obtain that : \begin{equation}\label{eq:LH} \hat H_B - \sum_{\eta} \mu_{\eta} \hat N_{\eta} = \sum_{\eta k \sigma} E_{\eta k } \gdag{\eta k \sigma} \gan{\eta k \sigma} + E_G + T(\hat{\mathcal{P}}_0), \end{equation} where $T(\hat{\mathcal{P}}_0)$ are terms proportional to $\hat{\mathcal{P}}_0$. They vanish after truncating the Hilbert space and only diagonal contributions remain. In Eq.~(\ref{eq:LH}) $E_{\eta k} = \sqrt{\xi_{\eta k}^2 + |\Delta_{\eta}|^2}$ denotes the quasiparticle energy, and $E_G$ is a constant energy off-set, often referred to as the energy of the Cooper pair condensate. For later reference we note that the term $\sum_{\eta} \mu_{\eta} \hat N_{\eta}$ is not included in the diagonalization procedure and is still written in terms of electron operators. \section{Transport theory and the generalized master equation} \label{sect:transport theory and the gme} In this section we derive the generalized master equation in the presence of superconducting leads. Since the generalized master equation approach to transport through quantum dots has become rather standard in recent years (see e.g. the method article by Timm \textit{et al.} \cite{Timm2011} or the recent paper by Koller \textit{et al.} \cite{Koller2010}) we only go into details of the derivation of the master equation when the effect of the superconducting leads brings significant differences with respect to the normal conducting theory. The expectation value $\mathcal{O} = \braket{\hat{\mathcal{O}}} = \operatorname{Tr}\bigl(\hat{\mathcal{O} } \hat\rho\bigr) $ of any observable associated to an operator $\hat{\mathcal{O}}$ can be evaluated once the total density operator $\hat\rho$ is known, cf. Eq.~(\ref{eq:current def}) below. To this extent we start from the Liouville equation for the density operator in the interaction picture, see e.g. \cite{Blum1996}: \begin{equation}\label{eq:Liouville_I} i \hbar \frac{\partial}{\partial t} \ro{I}(t) = \bigl[ \hat H_{T,I}(t) , \ro{I}(t) \bigr]. \end{equation} Eq.~(\ref{eq:Liouville_I}) can be formally integrated and reinserted back into itself, \begin{equation} \begin{split} i \hbar \, \dot {\hat\rho}_{I}(t) & = \bigl[\hat H_{T,I}(t), \ro{I}(0) \bigr]\\ & - \frac{i}{\hbar} \int_0^t dt' \biggl[ \hat H_{T,I}(t), \bigl[ \hat H_{T,I}(t'), \ro{I}(t') \bigr] \biggr], \end{split} \end{equation} which is still exact and allows a perturbative treatment in the tunneling Hamiltonian $\hat H_T$. Prior to time $t= 0$ the bath and the system do not interact, meaning that the total density matrix is factorized into a system and a leads component: \begin{equation} \ro{I}(0) = \ro{S}(0) \ro{B}(0). \end{equation} The density matrix of the leads, $\ro{B}$, can be described by the equilibrium thermodynamic expression shown in Eq.~(\ref{eq:rho_B}). Further we assume that the leads have so many degrees of freedom that they stay in thermal equilibrium up to a correction of order $\hat H_T$. It is convenient to trace out the degrees of freedom of the leads and define the reduced density matrix: \begin{equation} \ro{red,I}(t) \equiv \operatorname{Tr_{B} } \ro{I}(t). \end{equation} In the Schrödinger picture, the master equation for the reduced density matrix reads: \begin{equation}\label{eq:red} \begin{split} & \dot{\hat \rho}_{red}(t) = \frac{i}{\hbar} \bigl[ \ro{red}(t), \hat H_S \bigr] - \biggl( \frac{i}{\hbar} \biggr)^2 \U{0}{t} \int_{0}^{t}dt' \times \\ &\times \operatorname{Tr_{B} } \biggl( \biggl[ \hat H_{T,I}(t), \biggl[\hat H_{T,I}(t'), \ro{red,I}(t') \ro{B} \biggr] \biggr] \biggr) \Udag{0}{t} , \end{split} \end{equation} where we neglect terms of order $\mathcal{O}(\hat H_T^3)$ and $\U{0}{t}= e^{-\frac{i}{\hbar} \hat H_S t}$ is the time evolution operator of the unperturbed system. \subsection{Superconducting leads} The features of the superconducting leads are revealed when using the Bogoliubov transformation (\ref{eq:BT}) to express the tunneling Hamiltonian. This yields additional terms compared to the normal conducting theory. \subsubsection{Thermodynamic properties of the leads} \label{sect:thermodynamic properties of the leads} The description of electrons in terms of bogoliubons and Cooper pairs makes it necessary to discuss the thermodynamic properties of the superconducting leads. In this section we drop for simplicity the lead index $\eta$, and consider only one lead. In order to calculate thermal expectation values we use the equilibrium density matrix of a superconductor: \begin{equation}\label{eq:rho_B} \ro{B} = \frac{e^{-\beta \hat H_G}}{Z_G}, \end{equation} where $\hat H_G = \hat H_B -\mu \hat N$, $\beta$ is the inverse thermal energy, and $Z_G$ is the partition function in the grand canonical ensemble. We find that the thermal expectation value of a pair of Bogoliubov quasiparticles is equal to the Fermi function: \begin{equation}\label{eq:fermi} \operatorname{Tr_{B} } \biggl( \gdag{k\sigma}\gan{k\sigma} \ro{B} \biggr) = \frac{1}{e^{ \beta E_k} + 1} =f^+(E_k), \end{equation} where the trace is over the many-body states \begin{equation}\label{eq:many-body excitation state} \ket{ \{n_{q\tau}\},N } = \prod_{q\tau}( \gdag{q\tau})^{n_{q\tau}}\ket{0,N}, \end{equation} with independent sums over the number of electrons $N$ in the Cooper pair condensate and the quasiparticle configuration $\{n_{q\tau}\}=\{ n_{q_1 \tau_1}, n_{q_2 \tau_2}, \, \dots \}$. \subsubsection{Time evolution of the quasiparticles} To proceed we have to specify the time evolution of the Bogoliubov and Cooper pair operators. We find: \begin{equation}\label{eq:time_gdag} \gdag{\eta k \sigma,I}(t) = e^{+\frac{i}{\hbar} (E_k + \mu_{\eta})t} \gdag{\eta k \sigma}, \end{equation} \begin{equation}\label{eq:time_sdag} \operatorname{\hat S}^{\dagger}_{\eta,I}(t) = e^{+\frac{i}{\hbar} 2 \mu_{\eta}t} \operatorname{\hat S}^{\dagger}_{\eta}, \end{equation} in agreement with the results of Josephson and Bardeen \cite{Josephson1962,Bardeen1962}. When calculating the time evolution it is important to remember that in the lead Hamiltonian the term $\mu_{\eta }\hat N_{\eta}$ is still written in terms of electron operators. Before we proceed, we like to emphasize the importance of the Cooper pair contribution for finite bias voltages. As already pointed out by Governale \textit{et al.} \cite{Governale2008}, in this case $\mu_\eta$ cannot be set to zero and the time evolution of the Cooper pair operators, Eq.~(\ref{eq:time_sdag}), plays an important role. Neglecting the Cooper pair contribution for finite bias voltages \cite{Kosov2012} violates the number conservation in the tunneling processes and can lead to coherences which would vanish in the number conserving case. \subsubsection{Difference to the normal conducting theory} To compute Eq.~(\ref{eq:red}) we rewrite the electron operators using the Bogoliubov transformation, Eq.~(\ref{eq:BT}), and insert the time evolution as in Eqs.~(\ref{eq:time_gdag}) and (\ref{eq:time_sdag}). This yields four different traces to be calculated. We find: \begin{equation}\label{eq:trace cdag c} \begin{split} & \operatorname{Tr_{B} }\biggl( \cdag{\eta k \sigma ,I }(t) \can{\eta' k' \sigma' ,I }(t') \ro{B} \biggr) = \\& \delta_{\eta \eta'}\delta_{k k'}\delta_{\sigma \sigma'} \biggl\{ |u_{\eta k}|^2 f^+(E_{\eta k}) e^{+\frac{i}{\hbar}( E_{\eta k } + \mu_{\eta})(t-t')} \\ & \phantom{\delta_{\eta \eta'}\delta_{k k'}\delta_{\sigma \sigma'} } + |v_{\eta k}|^2 f^-(E_{\eta k}) e^{-\frac{i}{\hbar}(E_{\eta k } - \mu_{\eta}) (t-t')} \biggr\}, \end{split} \end{equation} \begin{equation}\label{eq:trace can cdag} \begin{split} & \operatorname{Tr_{B} }\biggl( \can{\eta k \sigma ,I }(t) \cdag{\eta' k' \sigma' ,I }(t') \ro{B} \biggr) = \\& \delta_{\eta \eta'}\delta_{k k'}\delta_{\sigma \sigma'} \biggl\{ |u_{\eta k}|^2 f^-(E_{\eta k}) e^{-\frac{i}{\hbar}( E_{\eta k } + \mu_{\eta})(t-t')} \\ & \phantom{\delta_{\eta \eta'}\delta_{k k'}\delta_{\sigma \sigma'} } + |v_{\eta k}|^2 f^+(E_{\eta k}) e^{+\frac{i}{\hbar}(E_{\eta k } - \mu_{\eta}) (t-t')} \biggr\}, \end{split} \end{equation} \begin{equation}\label{eq:trace cdag cdag} \begin{split} & \operatorname{Tr_{B} }\biggl( \cdag{\eta k \sigma ,I }(t) \cdag{\eta' k' \sigma' ,I }(t') \ro{B} \biggr) = 0, \end{split} \end{equation} \begin{equation}\label{eq:trace can can} \begin{split} & \operatorname{Tr_{B} }\biggl( \can{\eta k \sigma ,I }(t) \can{\eta' k' \sigma' ,I }(t') \ro{B} \biggr) = 0, \end{split} \end{equation} where $f^-(E) = 1-f^+(E)$. Note that the trace in Eqs.~(\ref{eq:trace cdag cdag}) and (\ref{eq:trace can can}) are vanishing since the lead Hamiltonian, Eq.~(\ref{eq:LH}), conserves the particle number. \subsection{General Master Equation for the reduced density matrix} Collecting all the previous results and expressing Eq.~(\ref{eq:red}) in the basis of the system eigenstates, $\{ \ket{n} \}$, we obtain the Bloch-Redfield form of the general master equation (GME) for the reduced density matrix: \begin{equation}\label{eq:GME} \begin{split} \dot{\rho}_{nn'} &= -\frac{i}{\hbar} \bigl(E_n- E_{n'}\bigr) \rho_{nn'}(t) \\ &- \sum_{m m'}\biggl( R_{nn'mm'}^{N\rightarrow N+1} + R_{nn'mm'}^{N\rightarrow N-1} \biggr) \rho_{mm'}(t), \end{split} \end{equation} where $n$ is a collective quantum number of the many body states of the quantum dot system and $\rho_{nn'} \equiv \bra{n}\hat \rho_{red} \ket{n'} $. Here, the Redfield-tensors are defined as: \begin{equation}\label{eq:redfield tensor} \begin{split}& R_{nn'mm'}^{N\rightarrow N \pm 1} = \sum_{\eta} \biggl\{ \\& \delta_{m'n'} \sum_{ l } \bigl(\Gamma^+_{nllm} \bigr)^{N\rightarrow N\pm1}_{\eta} +\delta_{m n} \sum_{l} \bigl( \Gamma^-_{m'lln'} \bigr)^{N\rightarrow N\pm1}_{\eta} \\ & - \bigl(\Gamma^+_{m'n'nm}\bigr)^{N\rightarrow N\pm1}_{\eta} - \bigl( \Gamma^-_{m'n'nm}\bigr)^{N\rightarrow N\pm1}_{\eta} \biggr\}. \end{split} \end{equation} The rates $\Gamma$ in Eq.~(\ref{eq:redfield tensor}) originate from terms containing traces of the type of Eqs.~(\ref{eq:trace cdag c}) and (\ref{eq:trace can cdag}). Further, we distinguish between rates describing the increase and rates describing the decrease of the particle number on the system, emphasized with the superscript $N\rightarrow N\pm 1$. Their detailed form is presented in App.~\ref{app:rates}. The rates with the superscripts $\pm$ are connected by complex conjugation and reversing of the indices: \begin{equation}\label{eq:rel gamma+-} \bigl( \Gamma^{-}_{n m m'n'}\bigr)_{\eta}^{N\rightarrow N\pm1} = \biggl( \bigl( \Gamma^{+}_{n'm' m n}\bigr)_{\eta}^{N\rightarrow N\pm1} \biggr)^* . \end{equation} \subsection{Current} Having derived the GME for the reduced density matrix in Eq.~(\ref{eq:GME}), we can use it to calculate measurable quantities such as the current and the differential conductance. In this section we present an expression for the current derived from the second order GME of Eq.~(\ref{eq:GME}). To do this we introduce a current operator whose statistical average gives the total current: \begin{equation}\label{eq:current def} I_{\eta} = \operatorname{Tr} \bigl( \hat I_\eta\, \ro{tot} \bigr). \end{equation} In general, the current operator of lead $\eta$ is defined as the variation of the total particle number in lead $\eta$ with time: \begin{equation}\label{eq:current operator} \hat I_{\eta,I}(t) = - e \frac{d}{dt} \hat N_{\eta,I}(t) = \frac{ + i e}{\hbar} \biggl[ \hat N_{\eta,I}(t) , \hat H_{T,I}(t) \biggr]. \end{equation} Calculating the commutator of Eq.~(\ref{eq:current operator}), we see that the current operator has the same operatorial structure as the tunneling Hamiltonian: \begin{equation} \begin{split} \hat I_{\eta,I}(t) =\frac{ + i e}{\hbar} \sum_{ k \alpha } \biggl( t_{\eta \alpha \sigma} \cdag{\eta k \sigma,I}(t) \dan{\alpha \sigma,I}(t) \\ - t^*_{\eta \alpha \sigma} \dddag{\alpha \sigma,I}(t) \can{\eta k \sigma,I}(t) \biggr), \end{split} \end{equation} differing only in the prefactor and summation. Hence, by applying the same perturbation theory as before, we obtain for the current in lead $\eta$: \begin{equation}\label{eq:current} I_\eta(t) = e \sum_{n m l} \biggl( \bigl(\Gamma_{nllm}^{N\rightarrow N+1}\bigr) _\eta - \bigl(\Gamma_{nllm}^{N\rightarrow N-1}\bigr)_\eta \biggr) \, \rho_{mn}^N(t) . \end{equation} In Eq.~(\ref{eq:current}) we introduced the abbreviations \begin{equation} \begin{split} \bigl(\Gamma_{nm m'n'}^{N\rightarrow N\pm1}\bigr) _\eta &= \bigl(\Gamma_{nmm'n'}^+\bigr)^{N\rightarrow N\pm1}_\eta +\bigl(\Gamma_{nmm'n'}^-\bigr)^{N\rightarrow N\pm1}_\eta \\ &= 2\, \text{Re} \biggl( \bigl(\Gamma_{nmm'n'}^+\bigr)^{N\rightarrow N+1}_\eta \biggr), \end{split} \end{equation} exploiting Eq.~(\ref{eq:rel gamma+-}). This gives us rates which are real and read: \begin{equation}\label{eq:rate_N_N+1} \begin{split} \bigl(\Gamma_{nm m'n'}^{N\rightarrow N + 1}\bigr) _\eta & = \text{Re}\bigg( \tilde\Gamma_{nmm'n'}^\eta ~ D\big(E_{m'n'} - \mu_\eta + i \gamma \big) \\ & \phantom{= Re} \times f^+\big(E_{m'n'} - \mu_\eta +i\gamma \big) \bigg) , \end{split} \end{equation} \begin{equation}\label{eq:rate_N_N-1} \begin{split} \bigl(\Gamma_{nm m'n'}^{N\rightarrow N - 1}\bigr) _\eta &= \text{Re} \bigg( \tilde \Gamma_{m'n'n m}^\eta ~ D\big(E_{n'm'} - \mu_\eta + i\gamma \big)~ \\ & \phantom{= Re} \times f^-\big(E_{n'm'} - \mu_\eta+ i\gamma \big) \bigg), \end{split} \end{equation} where \begin{equation} \tilde\Gamma_{nmm'n'}^\eta= \frac{2\pi}{\hbar} \sum_{\sigma \alpha \alpha'} t_{\eta \alpha \sigma} t^*_{\eta \alpha' \sigma} \bra{n} \dan{\alpha \sigma} \ket{m}\bra{m'} \dddag{\alpha' \sigma} \ket{n'} . \end{equation} In Eqs.~(\ref{eq:rate_N_N+1}) and (\ref{eq:rate_N_N-1}) $E_{n'm'}= E_n'-E_m'$ denote differences between system eigenenergies and \begin{equation}\label{eq:dos} D(E)= \rho_N \text{Re} \bigg(\frac{|E|}{\sqrt{E^2-|\Delta|^2}} \bigg), \end{equation} is the BCS-density of states, with $\rho_N = \frac{V m k_F}{2 \pi^2 \hbar^2}$ labeling the density of states for normal leads which is assumed to be constant around the Fermi level; $V$ denotes the volume of the lead and $m$ is the electron mass. In order to renormalize the divergence of the density of states we introduced a finite lifetime $\hbar/\gamma$ of the quasiparticle states in the superconducting leads, leading to a Lorentzian broadening of the resonance condition, see App.~\ref{app:renormalization}. This assumption is also in agreement with the results of Levy Yeyati \textit{et al.} \cite{Levy_Yeyati1996}, where they showed that the broadening of the BCS-like features in the current is due to the coupling to the leads. Eq.~(\ref{eq:current}) is a general result and can be applied to any transport set-up where an arbitrary system with discrete levels is weakly coupled to superconducting or normal conducting leads. The normal conducting case is obtained by setting $|\Delta_{\eta}|= 0$ and $\gamma = 0$. The theory is valid in the so called weak coupling limit, which is defined by the following relations between fundamental energy scales of the system: $\Gamma \ll |\Delta| \ll U$ and $\Gamma \ll k_B T$, where $\Gamma$ is the level broadening due to hybridization with the leads, $U$ is the charging energy, and $|\Delta|$ is the superconducting gap. As proven for example in Ref. \onlinecite{Levy_Yeyati1996}, the inclusion of higher order terms only produces in this regime an effective broadening of the quasiparticle density of states without invalidating the sequential tunneling description. In this paper we are only interested in the stationary limit. Hence, we replace the density matrix in Eq.~(\ref{eq:current}) by its stationary solution which is determined from Eq.~(\ref{eq:GME}) by imposing $\dot \rho_{nn'}^N = 0$. \section{Transport through multiple quantum dot devices} \label{sect:transport through multiple quantum dot devices} In the preceding sections we developed a perturbative microscopic theory for the stationary current of quantum dot devices coupled to superconducting leads. In the following, we show the predictions of the theory for two models, the single level quantum dot (SD) and the double quantum dot (DD). In the transport set-up the bias and gate voltages influence the energy configuration of the leads and the system, respectively. Specifically, the bias voltage is modifying the electrochemical potential of the leads, which we choose to have a symmetric voltage drop. Therefore we define the chemical potentials of the left and right lead, respectively: \begin{equation} \mu_{L/R} = \mu_0 \pm e \frac{V_b}{2}, \end{equation} where $\mu_0$ is the equilibrium chemical potential. The gate voltages are modifying the on-site energies of the system: We replace $\epsilon_d \rightarrow \epsilon_d + e V_g$ in the SD- and $\epsilon_{\alpha} \rightarrow \epsilon_{\alpha} +e V_g^\alpha$ in the DD-Hamiltonian. Here $e = - |e|$ is the electron charge. In the following we neglect coherences in the GME, considering only diagonal contributions of the reduced density matrix $\rho_{nn}$ by setting $n=n'$ in Eq.~(\ref{eq:GME}). Hence, it suffices to use only two indices for the transition rates. Neglecting the coherences is a non trivial step in the derivation of the master equation for the system. Within the secular approximation, see Ref. \onlinecite{Blum1996}, justified in the weak coupling limit, only coherences between degenerate states can play a role. We can now distinguish three types of degeneracies in the many-body spectrum of a quantum dot molecule: spin degeneracy, orbital degeneracy, and degeneracy between states with different particle number. Spin degeneracies can be neglected in the presence of unpolarized or collinearly polarized leads \cite{Hornberger2008,Donarini2010}. Orbital degeneracies are system dependent and they are not present in the single and double quantum dot systems discussed in this paper. A detailed discussion of their effects can be found for example in Refs. \onlinecite{Schultz2009,Donarini2010}. A detailed analysis of Eq.~(\ref{eq:GME}) shows that only 'anomalous' terms originating from Eqs.~(\ref{eq:trace cdag cdag}) and (\ref{eq:trace can can}) could couple populations ($\rho_{N,N}$) with coherences ($\rho_{N-1,N+1}$). Since these terms are exactly vanishing in the number conserving description of the superconducting leads, coherences decouple from populations and vanish in the stationary limit due to the damping introduced by the "R" components. \begin{figure} \includegraphics[width = \columnwidth]{fig2} \caption{Panels (a) and (b): Density of states (continuous line) and Fermi function (dotted line) at $k_B T = 0.2\,$meV and $k_B T= 0.01\, $meV, respectively. Panels (c) and (d): Product of the density of states and the Fermi function for the temperatures used in Fig.~(a) and (b), respectively. } \label{fig:fermi_dos} \end{figure} \begin{figure} \includegraphics[width=\columnwidth]{fig3} \caption{(Color online) Illustration of the transition lines appearing in presence of superconducting leads. The green lines mark transitions at the \textbf Source and the \textbf Drain contacts, described by the inequalities of Eqs.~(\ref{eq:s+}), (\ref{eq:d+}), (\ref{eq:s-}), and (\ref{eq:d-}). The red lines mark transitions involving \textbf thermally excited quasiparticles, given by Eqs.~(\ref{eq:st+}), (\ref{eq:dt+}), (\ref{eq:st-}), and (\ref{eq:dt-}). The $E_g$-$N$ diagrams for the points (a)-(c) are sketched in Fig.~\ref{fig:single_level_E-N}. } \label{fig:trans_lines} \end{figure} In current voltage spectroscopy it is convenient to illustrate the conditions under which current is allowed to flow as lines in the stability diagrams. These so called transition lines are fixed by the energetic part of the transition rates at the source $\eta=S$ and the drain $\eta = D$ contact: \begin{equation}\label{eq:gamma_n_n+1} \big( \Gamma^{N\rightarrow N + 1}_{mn}\big)_\eta \propto f^+(\Delta E - \mu_\eta)D(\Delta E - \mu_\eta), \end{equation} \begin{equation}\label{eq:gamma_n+1_n} \big( \Gamma^{N +1 \rightarrow N}_{nm}\big)_\eta \propto f^-(\Delta E - \mu_\eta)D(\Delta E - \mu_\eta), \end{equation} neglecting the lifetime broadening $\gamma$ for simplicity, and with $\Delta E = E^{N+1}_m -E^N_n$ the energy difference of the two transport levels. Fig.~\ref{fig:fermi_dos} illustrates this product for two different temperatures: For high enough temperatures quasiparticles can be excited thermally across the gap giving a small peak in the transition rates \cite{Whan1996}. The peak positions define transition lines when plotted in a $V_g$-$V_b$ diagram. Notice that while the most pronounced peak survives also at zero temperature and defines a transport threshold, the second peak vanishes at low temperatures and essentially only processes at and close to the peak are relevant. For an $N\rightarrow N+1$ transition we denote transitions associated to the more pronounced peak as S+ and D+ when happening at the source or at the drain contact, respectively. Transitions involving thermally excited quasiparticles are called St+ and Dt+. In complete analogy, we classify transitions from $N+1 \rightarrow N$: We denote by S- and D- the more pronounced transitions at the source and at the drain, and by St- and Dt- their thermal counterparts. In total we find 8 different transition lines, as depicted in Fig.~\ref{fig:trans_lines}. In the following we derive transport conditions and provide equations for the transport lines. For convenience we introduce $\Delta E_g = \Delta E - \mu_0$. We start with the analysis of the $N\rightarrow N+1$ transitions, which are described by the rates in Eq.~(\ref{eq:gamma_n_n+1}). From the arguments we find that the rates do not vanish if \begin{align}\label{eq:s+} {\Delta E_g \leq - |\Delta| + \frac{e V_b}{2}}, & & \text{Source {S+}} \end{align} \begin{align}\label{eq:d+} { \Delta E_g \leq - |\Delta| - \frac{e V_b}{2}}. & & \text{Drain D+} \end{align} Another contribution comes from the thermally excited quasiparticles states, namely, if the argument of the Fermi function $f^+ (\Delta E - \mu_\eta)$ and of the density of states $D(\Delta E - \mu_\eta)$ is equal to $|\Delta|$. At this point the transition rates are peaked and contribute to the current: \begin{align}\label{eq:st+} {\Delta E_g = |\Delta| + \frac{e V_b}{2}}, & & \text{Source thermal St+} \end{align} \begin{align}\label{eq:dt+} { \Delta E_g = |\Delta| - \frac{e V_b}{2}}. & & \text{Drain thermal {Dt+}} \end{align} Since the thermally excited quasiparticles produce a peak rather than a step in the current voltage characteristic, the corresponding transport condition is formulated with an equality. Transitions from $N+1 \rightarrow N$ are described by the rate of Eq.~(\ref{eq:gamma_n+1_n}), leading in complete analogy to the previous case to the following transport conditions: \begin{align}\label{eq:s-} {- \Delta E_g \leq -|\Delta| - \frac{e V_b}{2}}, & & \text{Source {S-}} \end{align} \begin{align}\label{eq:d-} {- \Delta E_g \leq -|\Delta| + \frac{e V_b}{2}}, & &\text{Drain {D-}} \end{align} \begin{align}\label{eq:st-} {- \Delta E_g = |\Delta| - \frac{e V_b}{2}}, & & \text{Source thermal {St-}} \end{align} \begin{align}\label{eq:dt-} {- \Delta E_g = |\Delta| + \frac{e V_b}{2}}. & & \text{Drain thermal {Dt-}} \end{align} \subsubsection{Visualization of the transport conditions} \begin{figure} \includegraphics[width = \columnwidth]{fig4} \caption{(Color online) Visualization of the transport conditions of Eqs.~(\ref{eq:s+})-(\ref{eq:dt-}). We plotted the threshold of the transport inequalities as green lines (S$\pm$, D$\pm$); for the equalities coming from transitions involving thermally excited quasiparticles we used red lines (St$\pm$, Dt$\pm$ ). Choosing the reference level in the $N$ particle subspace, we found a scheme where transitions are energetically allowed to levels which lie in the shaded region below the green lines and to levels which align with the red lines. Dashed boxes mark the bias window $eV_b$. } \label{fig:scheme} \end{figure} To visualize the transport conditions of Eqs.~(\ref{eq:s+})-(\ref{eq:dt-}) we extend the scheme of Donarini \textit{et al.} of Ref. \onlinecite{Donarini2010} to superconducting leads. The scheme is depicted in Fig.~\ref{fig:scheme} and illustrates for which relative position of the systems eigenenergies $E_g^N = E_m^N - \mu_0 N$ transitions are energetically allowed. The bias window is marked with a dashed box. The green lines mark the borders of the inequalities, and the red lines the sharp equalities for the thermal transitions, meaning that transitions can occur to states lying below the green lines (shaded region), and to states which coincide with the red lines. In order to see a transition between two levels in the stability diagram a source and a drain transition must be allowed between the two levels (depicted as arrows in the $E_g$-$N$ diagrams of Fig.~\ref{fig:single_level_E-N}). We note that for a full analysis of the transport properties also the geometrical part of the rates must be taken into account and transport occurs only if $\tilde \Gamma \neq 0$. \subsection{ Single level quantum dot model} \begin{figure} \includegraphics[width = \columnwidth]{fig5} \caption{(Color online) \textbf{(a)-(d)}: $E_g$-$N$ diagrams for a single level quantum dot with $\Delta E_g > |\Delta|$ and at bias voltages as sketched in Fig.~\ref{fig:trans_lines}. For the simulations of Fig.~\ref{fig:SD_LT} $\Delta E_g > \Delta$ corresponds to a gate voltage $eV_g < -2.6 \,\text{meV}$. In (a) we cut the S+ line: the particle number on the system is increased by a tunneling event at the source contact and decreased at the drain. (b) Cut with the thermal line St+: the particle number of the system is increased by a tunneling event involving a thermally excited quasiparticle at the source contact and decreased by tunneling into empty states in the source and the drain contact, respectively. \textbf{(c)}: $E_g$-$N$ diagrams for a single level with $0<\Delta E_g < |\Delta|$. The two levels are only connected by two drain transitions, meaning that in this configuration the system is in thermal equilibrium with the drain contact. } \label{fig:single_level_E-N} \end{figure} The simplest example of a quantum dot system is the single level quantum dot presented in Eq.~(\ref{eq:SD}). Since only one level is involved, we can do most calculations analytically and understand the basic mechanism resulting from the superconducting leads. In Fig.~\ref{fig:SD_LT} the stationary current is shown as a function of bias and gate voltage for superconducting leads at $k_B T = 0.5 |\Delta| $. We observe the expected gap \cite{Grove-Rasmussen2009} between the Coulomb diamonds which is equal to $4 |\Delta|/e$. The gap can be explained using Fig. \ref{fig:trans_lines} and the corresponding Eqs.~(\ref{eq:s+})-(\ref{eq:dt-}). One dashed line marks the gate voltage where $\Delta E_g = 0$. Along this line the conditions under which current is allowed to flow read: $eV_b/2 > |\Delta|$ for the S+, D- lines, and $eV_b/2 < -|\Delta|$ for the S-, D+ lines, opening a bias window of $4|\Delta|/e$ where current is blocked for low temperatures $k_B T \ll |\Delta|$. For higher temperatures of $k_B T \approx 0.5 |\Delta| $ we observe small peaks in the Coulomb blockade region (green area) which are due to thermally excited quasiparticles; they correspond to the red lines in Fig.~\ref{fig:trans_lines}. In Fig.~\ref{fig:single_level_E-N} we show the energy particle number diagrams in the points (a)-(d), which lie on a vertical cut through Fig.~\ref{fig:trans_lines} at $\Delta E_g > |\Delta|$ which corresponds to a gate voltage $e V_g > 2.6 \, $meV in Fig.~\ref{fig:SD_LT}. In Fig.~\ref{fig:single_level_E-N} (a) we depicted the $E_g$-$N$ diagram for a cut with the S+ resonance line, where the particle number on the system is increased by a tunneling event at the source and decreased at the drain contact. For bias voltages smaller than the one at resonance (corresponding to larger $eV_b$ as $e$ is the negative charge of an electron) the S+, D- transitions remain open and current can flow. In Fig.~(\ref{fig:single_level_E-N}) (b) the $E_g$-$N$ diagram at the resonance line St+ is shown. In this case the bias voltage is not large enough to allow the transitions S+ of Eq.~(\ref{eq:s+}). For low temperatures no quasi particle is thermally excited meaning that only transitions from $1\rightarrow 0$ are energetically allowed (green arrows). For high enough temperatures, however, the particle number of the system can be increased by tunneling events involving thermally excited quasiparticles opening the St+ transition. By changing the sign of the bias voltage the role of the source and the drain is inverted, explaining the transition lines Dt+ and D+ (Fig.~\ref{fig:single_level_E-N}(c) and \ref{fig:single_level_E-N}(d)). Another interesting constellation of the energy level occurs in the region of $0< \Delta E_g < |\Delta|$ (Fig.~\ref{fig:single_level_E-N} (e)), where in the current-voltage characteristics the thermal lines are vanishing. Transitions can only occur at the drain contact, as the bias is not large enough to allow transitions at the source. Hence, the system is in thermal equilibrium with the drain contact and the occupation probabilities are related by the Boltzmann distribution: \begin{equation} \frac{\rho_0}{\rho_1} = e^{\beta(\Delta E_g + e V_b/2 )}, \end{equation} in the limit of $\gamma \rightarrow 0$. \begin{figure} \includegraphics[width = \columnwidth]{fig6} \caption{(Color online) (a) Current voltage characteristics of a SD coupled to superconducting leads. Parameters are $k_B T = 0.3 \,\text{meV}$ and $|\Delta| = 0.6\,$meV, $U = 4\,\text{meV} $, $\epsilon_d = -2 \,\text{meV}$, $e\Gamma = 0.001\,\text{meV}$. (b) Subgap features coming from thermally excited quasiparticles of the 0-1-particle transition, highlighted as a dashed box in (a). } \label{fig:SD_LT} \end{figure} \subsection{The double quantum dot} We have seen that the theory can reproduce well known results for the SD and we understood the properties of the thermal transitions in $E_g$-$N$ diagrams with only one non degenerated level per particle number. In the following we investigate a more advanced system, the double quantum dot, where the many body spectrum gives rise to more than one non degenerated level per particle number, so called excited system states. For normal conducting leads the excitations cannot be seen for low bias voltages, since transitions to the ground state are always possible, blocking transport through the excitations. In the last subsection we have seen that for superconducting leads the energy difference must be at least $|\Delta E_g| \geq eV_b/2 - |\Delta|$ to have non thermal source and drain transitions. Hence, we find situations where the transition to the ground state are energetically not allowed and transport occurs through excited system states. We start with equally gated dots with the same on-site energies and on-site Coulomb interactions, where it is possible to diagonalize the Hamiltonian analytically \cite{Bulka2004,Hornberger2008}. In the second part, the case of independently coupled dots is discussed, where the detuning of the two gate voltages influences the level spacing of the energy spectrum. Thus, excited states can be observed only in detuning ranges where the difference between the energy level of the excited state and its ground state is less than $2|\Delta|$. \subsubsection{Equally gated dots}\label{sect:equally gated dots} \begin{figure} \includegraphics[width = \columnwidth]{fig7} \caption{(Color online) (a) Current voltage characteristics of an equally gated DD in serial configuration at $k_B T = 0.2\,\text{meV}$, $|\Delta| = 0.4\,\text{meV}$, $U = 4 \, \text{meV}$, $V = 2 \,\text{meV}$, $b= -0.3 \,\text{meV}$, and $e\Gamma = 0.001 \,\text{meV}$. (b) I-V characteristics in the subgap region corresponding to the dashed box in (a). The distance between the 1-particle excited state and its ground state is equal to the coupling strength $2|b|$ of the two dots. Moreover, $2|b|<2|\Delta|$. The black arrow marks the transition line coming from transport through the 1-particle excited state. (c) I-V-characteristics in the subgap region, where we increased the coupling between the two dots ($b= -0.5 \,\text{meV}$), leading to a level spacing which is larger than $2|\Delta|$, hence transport through the excited system state is not allowed and the line disappears. } \label{fig:DD_eq_serial_current} \end{figure} \begin{figure} \includegraphics[width=\columnwidth]{fig8} \caption{(Color online) Sketch of the transition lines for the 0-1 particle transition of an equally gated DD. It shows two copies of Fig.~\ref{fig:trans_lines} where the labeling of the blue (orange lines) is the same as for the green (red) lines. The blue (orange) lines mark the transition lines corresponding to the 0- particle ground state to 1-particle first excited state transition. } \label{fig:0_1_excited_line} \end{figure} \begin{figure} \includegraphics[width = \columnwidth]{fig9} \caption{(Color online) $E_g$-$N$ diagram corresponding to the points of Fig.~\ref{fig:0_1_excited_line} where the dashed line cuts the transition lines for the case of an equally gated DD. In this case the distance between the 1-particle ground state to the 1-particle first excited state is equal to $2b< 2|\Delta|$, where $b$ is the tunneling strength between the two quantum dots. (A) Point on the thermal line St+ of the ground state to ground state transition. (B) Point on the thermal line St+ of the ground state to first excited state transition. (C) Point on the Dt- line of the ground state to ground state transition. (D) Point on the Dt- line of the ground state to first excited state transition; this line cannot be seen in the current voltage characteristics, as the ground state to ground state transitions are open. Hence, in the long time behavior the system will occupy the 1-particle ground state blocking the current through the excited state. } \label{fig:point_A_B_C_D} \end{figure} \begin{figure} \centering \includegraphics[width = 0.5 \columnwidth]{fig10} \caption{(Color online) $E_g$-$N$ diagram of point (B) in Fig.~\ref{fig:0_1_excited_line}, for a level spacing of the one particle energies larger than $2b > 2\Delta$. In contrast to Fig.~\ref{fig:point_A_B_C_D} the transition between the 0-particle ground state and the 1-particle excited state is open, blocking the current.} \label{fig:point_B_b>delta} \end{figure} For equally gated dots the on-site energies of the two sites are modulated with the same gate voltage. Hence, it is convenient to plot the current as a function of the bias and the gate voltage as for the SD. Fig.~\ref{fig:DD_eq_serial_current} shows the current of an equally gated DD in serial configuration. As for the SD we observe Coulomb blockade and the gap of $4|\Delta|/e$ between the tips of the diamonds. Transport carried by thermally excited quasiparticles is of particular interest, as it allows one to observe transitions through excited system states for low bias voltages, which are often diminished by the ground state transitions in the normal conducting case. In order to show some interesting phenomena resulting from the more complex spectrum, we concentrate on the 0- to 1-particle transition where three levels are involved. In the 1-particle spectrum, the difference between the ground state and the excited state is equal to $2|b|$, where $b<0$ is the tunneling strength between the two dots. Meaning that by tuning the coupling between the two dots it is possible to influence the level spacing. Fig.~\ref{fig:0_1_excited_line} shows a sketch of the transition lines expected for the $0-1$ transition for $|b|<|\Delta|$, where the red (green) lines show the ground state to ground state transitions, and the blue (orange) lines the ground state to first excited state transitions. For a better understanding of the transport properties we cut the transitions lines horizontally for a small bias voltage $eV_b/2 < |\Delta|$ in the Coulomb blockade region (points (A)-(D)), the corresponding $E_g$-$N$ diagrams are depicted in Fig.~\ref{fig:point_A_B_C_D}. In point (A) the difference between the ground states is equal to $\Delta E_g = eV_b/2 +|\Delta|$ opening the thermal transition St+ and current can flow. Following the dashed line to point (B), the 1 particle states are shifted down in energy until the St+ transition is allowed between the 0-particle ground state and the 1-particle excited state. Since $ |b| < \Delta$, the 1-particle ground state is energetically not accessible and current can flow through the excited state. We like to emphasize that the blocking of the ground state transition is only valid as long as the distance between the two 1-particle levels is smaller than $2|\Delta|$. For larger distances the ground state is energetically accessible, blocking the current through the excited state, c.f. Fig.~\ref{fig:point_B_b>delta}. In point (C) $eV_g$ is further decreased, the Dt- transition between the ground states is opening, and current can flow. Point (D) shows the typical energy configuration in which current through the excited state is blocked, even though the transition through the excited state is energetically allowed. The reason for that is the 1-particle ground state which can be populated, but transitions describing its depopulation are energetically not allowed, leading to a blocking of the current in the stationary limit. To demonstrate the important role of the level spacing we show the current voltage characteristics of an equally gated DD in the subgap region in Figs.~\ref{fig:DD_eq_serial_current}(b) and \ref{fig:DD_eq_serial_current}(c). In (b) the spacing of the 1-particle energy levels $|2b| < 2|\Delta|$, hence, the excited state can be observed in the current (arrow in Fig.~\ref{fig:DD_eq_serial_current}). In (c) we increase the tunneling strength between the two dots $2|b| > 2|\Delta|$ and the excited state line is vanishing, as explained in Fig.~\ref{fig:point_B_b>delta}. As in the case for $2|b|<2|\Delta|$ the excited level is in resonance with the St+ transition, however, due to the larger level spacing, the ground state transition opens and current is blocked. \subsubsection{Independently gated dots} \begin{figure} \includegraphics[width = 0.8 \columnwidth]{fig11} \caption{(Color online) (a)-(b) Current voltage characteristics of a DD coupled to normal conducting leads in serial (a) and in parallel (b) configuration. We fixed the bias voltage to $e V_b = 0.3\, $meV. (c)-(d) Current voltage characteristics of a DD coupled to superconducting leads in serial (c) and in parallel (d) configuration. We fixed the bias voltage to $e V_b = 0.3\,\text{meV}+ 2|\Delta|$ in order to obtain the same conditions as for the normal conducting case in (a)-(b). Parameters are: $T= 0.01\,\text{meV}$, $|\Delta| = 0.4 \,\text{meV}$, $e\Gamma = 0.001 \,\text{meV}$, $b= -0.2 \,\text{meV}$, $U = 4 \, \text{meV}$ and $V= 2 \,\text{meV}$. } \label{fig:DD_nc_sc} \end{figure} \begin{figure} \centering \includegraphics[width=\columnwidth]{fig12} \caption{(Color online) \textbf{(a)} $E_g$-$N$ diagram of the 0-1-particle transition for $eV_b/2 >|\Delta|$. In the 1-particle spectrum we plotted two situations which mark the borders of the current step. The dashed levels mark the left border (for small $\Sigma_g$) where the 1-particle levels lie above the 0-particle energy level. If the distance $\Delta E_g \leq eV_b/2 - |\Delta|$ current can flow through S+ and D- transitions. By lowering $e \Sigma_g$ the 1-particle energy levels move down in the $E_g$-$N$ diagram, while the transitions remain open. The solid lines mark the right border of the current steps, as for levels lying below the solid line the D- transition is closed and current is blocked. Thus, the width of the current steps in the current voltage characteristics is: $e \Delta \Sigma_g = eV_b -2 |\Delta|$. \textbf{(b)} $E_g$-$N$ diagram of the 0-1-particle transition involving thermal transitions. For the same arguments as in (a), the distance between two thermal lines in the current voltage characteristics is equal to $e \Delta \Sigma_g = eV_b + 2|\Delta|$.} \label{fig:IDD_E-N-0-1} \end{figure} \begin{figure} \centering \includegraphics[width = 0.35\columnwidth] {fig13} \caption{(Color online) $E_g$-$N$ diagram for the 0-1-particle transition. Transitions between the two 1-particle levels (dashed lines) and the 0-particle ground state are allowed through the thermal St+ transition. Increasing the gate voltage the levels move down in energy (solid lines) and the excited state transition can be observed when the excited level aligns with the St+ transition. Hence, the distance of two neighboring thermal transitions is equal to the level spacing. } \label{fig:IDD_E-N-0-1_level_spacing} \end{figure} In the last paragraph we considered a DD with both dots coupled to the same gate electrode. In most experiments, however, it is more convenient to couple the dots independently, which leads to a 'honeycomb' shaped current voltage characteristics \cite{VanderWiel2003}. For symmetric on-site energies and Coulomb repulsion it is possible to diagonalize the DD Hamiltonian of Eq.~(\ref{eq:DD hamiltonian}) analytically. Gating the dots independently destroys this symmetry, an analytical diagonalization is not possible, and one has to use numerical methods. We plot the current as a function of the detuning $\Delta_g= V_{g}^1 - V_{g}^2$, and the average of the two gate voltages $\Sigma_g = (V_{g}^1 + V_{g}^2)/2$. The current voltage characteristic for serial and parallel configuration is depicted for the normal conducting case in Fig.~\ref{fig:DD_nc_sc} (a)-(b) and for the superconducting case in Fig.~\ref{fig:DD_nc_sc} (c)-(d). Comparing both configurations, we observe for the serial one a decrease in the current for high detuning $\Delta_g$, while in the parallel configuration current can be observed over the entire voltage range. This difference is a consequence of the geometry of the set-up as the DD system remains unchanged. An increase of the detuning leads to a localization of the systems ground state at site 1 and transitions through site 2 are blocked. Since in serial configuration the right lead is only coupled to site 2, the localization of the wave function at site 1 leads to a decrease in the current. In parallel configuration, however, both sites are coupled to both leads and the ground state transition is always open. The left and right border of the current steps are given by the source and drain lines, respectively. They follow, in complete analogy to the simplest case, from energy conservation. In Fig.~\ref{fig:IDD_E-N-0-1} (a) we show the $E_g$-$N$ diagram for the 0 to 1-particle transition illustrating two limits: the ground states are (i) in resonance with the S+ transition (dashed line) and (ii) in resonance with the D- transition (solid line), describing the left and right borders of the current step in Fig.~\ref{fig:DD_nc_sc} (c-d). Starting at the S+ resonance, the energy levels of the 1-particle spectrum are moving down in energy by increasing the average gate voltage $\Sigma_g$. Both transitions (S+ and D-) remain open as long as the ground state lies in the blue (shaded) region. If the ground state lies below the solid line, the D- transition is closed and current is blocked. Hence, the width of the current steps in the current voltage characteristics in Fig.~\ref{fig:DD_nc_sc} (c-d) is equal to the size of the blue (shaded) region in Fig.~\ref{fig:IDD_E-N-0-1} (a), namely $e\Delta \Sigma_g = eV_b -2|\Delta|$. The same arguments hold for the distance of two corresponding thermal transitions, as illustrated in Fig.~\ref{fig:IDD_E-N-0-1} (b) the distance of two thermal lines is equal to $e\Delta \Sigma_g = eV_b + 2|\Delta|$. As we can see in Fig.~\ref{fig:DD_nc_sc} there exists a one to one correspondence of the transport conditions of the normal conducting to the superconducting case which leads to the same shape of the current voltage characteristics if $k_B T \ll |\Delta|$. Increasing the bias voltage by $2 |\Delta|$ compared to the normal conducting case $eV_b^{\text{\tiny{SC}}} =e V_b^{\text{\tiny{NC}}} + 2 |\Delta|$ leads to the same transport conditions. Although the shape of the current steps in Figs.~\ref{fig:DD_nc_sc} (a-b) and \ref{fig:DD_nc_sc} (c-d) look the same, they differ at the edges of the current steps, as in the superconducting case the sharp peaks of the quasiparticle density of states are reflected in the current. \subsubsection{Thermal effects}\label{sect:thermal effects DD} We have seen that the shape of the stability diagram can be explained using energy conservation, in complete analogy to the simplest case. In this section we discuss the case for small bias voltages $eV_b/2 < |\Delta| $, where current can flow due to thermally excited quasiparticles exclusively. As already observed above, thermally excited quasiparticles do not produce steps in the current voltage characteristics rather they appear as small peaks. This can be used to resolve transitions through excited system states whose energy difference to the ground state is less than $2 |\Delta|$. By detuning the gate voltages of the two sites of the DD we can change the level spacing of the systems eigenenergies; hence, the excited states are only observed in a certain detuning range. To analyze transitions through excited system states, c.f. Fig.~\ref{fig:DD_parallel_subgap}, we choose the parallel configuration to rule out the geometrical effect also leading to a decrease of the current for high detuning. If a line corresponding to an excited state disappears for higher detuning $\Delta_g$, we conclude that the energy difference to its ground state is larger than $2|\Delta|$. In Fig.~\ref{fig:DD_E_diff} we plotted the energy differences of the excited states with respect to their ground state for different values of the detuning $\Delta_g$, which are marked as red lines in Fig.~\ref{fig:DD_parallel_subgap}. Counting the number of levels lying under the red line in Fig.~\ref{fig:DD_E_diff} gives information about the number of visible excited lines. For instance, consider the case of $\Delta_g = 0$ in Fig.~\ref{fig:DD_E_diff}. Following the red line from small to high $\Sigma_g$ in Fig.~\ref{fig:DD_parallel_subgap}, we cross the $0$-$1$ particle transitions and observe three lines: two corresponding to the ground state, and one line in between corresponds to a transition through the 1-particle excited state. The distance between the leftmost ground state transition line and the excited line determines the level spacing of the one particle spectrum, see \ref{fig:IDD_E-N-0-1_level_spacing}. In the 2-particle spectrum the energy difference of one excited state lies under the red line. Hence we should see two lines coming from excited system states, namely the transition between the 1-particle ground state and the 2-particle excited state, and transitions between the 2-particle ground state and the 1-particle excited state. Along the horizontal cut at $\Delta_g = 2$ in Fig.~\ref{fig:DD_parallel_subgap}, excited states can only be observed for the 1-2 particle and the 2-3 particle transition. This is in agreement with Fig.~\ref{fig:DD_E_diff}, where only in the 2 particle subspace energy differences lie under the threshold of $2|\Delta|$. For higher detuning, e.g. $\Delta_g=4$, no excited states can be seen, as the detuning increases the level spacing, and all energy differences are larger than $2 |\Delta|$ Fig.~\ref{fig:DD_E_diff}. \begin{figure} \includegraphics[width= \columnwidth]{fig14} \caption{(Color online) Current voltage characteristics of a DD in parallel configuration for bias $V_b < 2|\Delta|/e$. Since the bias voltage is not high enough current can flow only due to thermally excited quasiparticles. The red lines correspond to Fig.~\ref{fig:DD_E_diff} where the energy differences of the excited states with respect to their ground state are plotted as a function of particle number. The number of visible excited states is proportional to the number of energy differences which are smaller than $2|\Delta|$ (red line in Fig.~\ref{fig:DD_E_diff}). Parameters are: $T= 0.01\,\text{meV}$, $eV_b = 0.3 \,\text{meV}$ $|\Delta| = 0.4 \,\text{meV}$, $e\Gamma = 0.001 \,\text{meV}$, $b= -0.2 \,\text{meV}$, $U = 4 \, \text{meV}$ and $V= 2 \,\text{meV}$. } \label{fig:DD_parallel_subgap} \end{figure} \begin{figure} \centering \includegraphics[width= 0.7 \columnwidth]{fig15} \caption{(Color online) Plot of the energy differences of the excited system states with respect to their ground state as a function of particle number. If the energy difference is smaller than $2|\Delta|$, transitions through these excited states can be seen in the current voltage characteristics. The threshold of $2|\Delta|$ is marked as a red horizontal line. We depicted the plots for three situations differing in the detuning $\Delta_g$. The three cases are marked as horizontal lines in Fig.~\ref{fig:DD_parallel_subgap}.} \label{fig:DD_E_diff} \end{figure} \subsection{The N-QD-S junction} \begin{figure} \includegraphics[width=\columnwidth]{fig16} \caption{(Color online) Sketch of the transition line of a QD coupled to a normal conducting (source) and a superconducting lead (drain). The difference to the S-QD-S system is that only the drain lines split due to the superconducting gap, the S+ and S- lines are described by the same equation. In this case a gap equal to $|\Delta|$ is opening, and the triangles are shifted apart. Thermal lines can be observed only for the drain.} \label{fig:trans_lines_n-qd-s} \end{figure} \begin{figure} \includegraphics[width=\columnwidth]{fig17} \caption{(Color online) Differential Conductance of a SD coupled to a normal conducting (source) and to a superconducting lead (drain) (N-QD-S system). The coupling to the lead is $e\Gamma = 0.01\,$meV. (a) Superconducting gap of $|\Delta| = 0.6\,$meV and temperature $k_B T = 0.1$meV. No thermal lines in the subgap region are visible. (b) The same temperature $k_B T = 0.1$meV, but for smaller gap $|\Delta| = 0.3$meV; quasiparticles get thermally excited across the gap leading to transport in the Coulomb blockade region. Parameters are $U= 4 \,\text{meV}$ and $\epsilon_d = -2 \,\text{meV}$. } \label{fig:n-qd-s_diff_cond} \end{figure} \begin{figure} \includegraphics[width=0.5\columnwidth]{fig18} \caption{(Color online) Visualization of the transport conditions for a N-QD-S system with $eV_b/2 < |\Delta|$, where the source is a normal and the drain a superconducting lead. They follow from Eqs.~(\ref{eq:s+})-(\ref{eq:dt-}) by setting $|\Delta| = 0$ in the equations corresponding to the source lead. } \label{fig:scheme_n-qd-s} \end{figure} \begin{figure} \includegraphics[width=\columnwidth]{fig19} \caption{$E_g$-$N$ diagrams corresponding to points (a) and (b) of Fig.~\ref{fig:trans_lines_n-qd-s}. (a) We see a positive current in the subgap region, which comes only due to the thermal smearing of the S+ transition. (b) The line connecting the S+ and the S- transition line in the Coulomb blockade region the system is in thermal equilibrium with the source contact. } \label{fig:E-N-diagram_n-qd-s} \end{figure} We close this paper by investigating a so called N-QD-S hybrid system, where a quantum dot system is coupled to a normal and to a superconducting lead, giving a possible explanation for the subgap features in Ref. \onlinecite{Dirks2009}. In the experiment of Ref. \onlinecite{Dirks2009} a carbon nanotube was contacted to two normal conducting leads and to a superconducting finger in between. The differential conductance between the superconducting finger and a normal lead is measured, realizing a N-QD-S hybrid system. It is possible to apply a bias voltage across the entire tube as well as between the superconductor and a normal conducting lead. The stability diagram in Fig.~2 (a) in Ref. \onlinecite{Dirks2009}, with no bias applied over the entire tube, reveals the typical Coulomb diamond pattern resulting from quasiparticle tunneling with no subgap features. By applying a bias voltage $V_{SD}$ over the entire tube, the gap in the stability diagram gets smaller with respect to the unbiased case and conductance lines can be seen in the Coulomb blockade region, c.f. Fig.~3 (a) of Ref. \onlinecite{Dirks2009}. The reduction of the gap in the stability diagram is proportional to the applied bias voltage of approximately $eV_{SD} \approx |\Delta|/2$, and is related to an effective reduction of the superconducting gap. For a smaller gap quasiparticles can get thermally excited across the gap leading to subgap transport in complete analogy to the S-QD-S case discussed above. We can model the N-QD-S system by setting $|\Delta_S| = 0$ for the normal conducting lead (source) in the master equation; the drain contact remains superconducting $|\Delta_D| = |\Delta|$. Hence, the transport conditions change slightly and can be summarized in the scheme of Fig.~\ref{fig:scheme_n-qd-s}. In Fig.~\ref{fig:trans_lines_n-qd-s} we schematically sketched the expected transition lines for a N-QD-S hybrid structure. In Fig.~\ref{fig:E-N-diagram_n-qd-s} we analyzed the two most important cases, marked as points (a) and (b) in Fig.~\ref{fig:trans_lines_n-qd-s}. Point (a) shows a paradoxical situation as the particle number of the system seems to be increased only at the drain contact, which would lead to a negative current at positive bias. However, if the two contacts have the same temperature, the thermal broadening of the S+ line gives a small contribution in the transition rates (dashed green arrow in Fig.~\ref{fig:E-N-diagram_n-qd-s} (a)) making the current positive. The situation in (b) shows again the system being in thermal equilibrium with the source contact. We can see that the lines with negative slope (drain lines) give a finite current in the Coulomb blockade region as observed in Fig.~3 (b) in the experiments. Thus, we claim that the subgap features observed in the experiments possibly are transitions involving thermally excited quasiparticles which are allowed due to the reduction of the superconducting gap. This argument is supported by the observation that for diamonds where the gap has the same size as before (edges of the stability diagram), no subgap lines can be observed. In Fig.~\ref{fig:n-qd-s_diff_cond} we show two $dI/dV-$ characteristic of a N-QD-S system corresponding to different superconducting gaps with the same temperature ($k_B T = 0.1$meV) in both cases. In (b) the superconducting gap ($|\Delta| = 0.3$meV) is only half of the gap in (a) ($k_B T = 0.6$meV). By reducing the gap, the temperature becomes large enough to excite quasiparticles across the gap, leading to conductance peaks in the Coulomb blockade region, as observed in the experiments However, a more complex modeling of the multi-terminal system is required to understand the experiments in all details. \section{Conclusion} In this work we developed a transport theory for nanostructures coupled to superconducting leads up to second order in the tunneling Hamiltonian. We used the Bogoliubov transformation to describe the electrons in the superconductors as Cooper pairs and Bogoliubov quasiparticle excitations, whereby we modified the Bogoliubov transformation in a number conserving way \cite{Josephson1962,Bardeen1962}, introducing Cooper pair creation and annihilation operators explicitly. We showed the predictions of the theory on two examples, the well known single level quantum dot, and the double quantum dot. The characteristic gap in the Coulomb diamonds, proportional to the superconducting gap, as well as negative differential conductance was observed in both cases. Further, we considered the double quantum dot in serial as well as in parallel configuration, see Fig.~\ref{fig:DD_configuration}, coupling the dots to the same as well as to two separate gate electrodes. We systematically analyzed the stability diagrams, extending the scheme of Ref. \cite{Donarini2010} for superconducting leads. We found that transport through excited system states occurs even for low bias voltages using thermally excited quasiparticles, leading to zero bias peaks in the conductance. Transitions through excited states can be observed if transitions through the ground state are energetically not allowed, namely if the distance between the energy levels of the excited state and the ground state is smaller than $2|\Delta|$. This effect can be seen in the the current voltage characteristics of an independently gated double quantum dot in parallel configuration without tuning parameters of the system, since the level spacing changes with the detuning $\Delta_g$ of the gate voltages. Hence the excited states can be seen only in certain detuning windows. Finally, we analysed the case where a quantum dot is coupled to a normal and a superconducting lead, giving a possible explanation for the subgap features of Ref. \onlinecite{Dirks2009} in terms of transport involving thermally excited quasiparticles. We conclude with the observation that thermally excited quasiparticles can lead to a finite current in the Coulomb blockade region. Besides the well known thermal transitions through the ground states, transitions through excited system states must be taken into account as they are an additional source of zero bias peaks in the conductance. For a better comparison with experiments the theory can be used to investigate more realistic systems such as carbon nanotube quantum double dots. Specifically, the current voltage spectroscopy in the low bias regime can be used to learn something about the spectrum of the set-up. Within our approach it is not possible to capture Josephson current and Andreev reflections as they are higher order processes. Yet, in the weak coupling regime lowest order quasiparticle transport gives not only the basic structure of the Coulomb diamonds but also the dominant subgap feature, i.e. thermally activated conductance peaks associated to quasiparticle transport. In order to observe the Josephson effect and Andreev reflections, the theory must be extended to higher order perturbation theory \cite{Governale2008,Martin-Rodero2011}. \begin{acknowledgments} We acknowledge financial support through DFG Program Nos. SFB631. \end{acknowledgments} \begin{appendix} \section{Properties of the Cooper pair operators} \label{app:bcs GS CPO} In the microscopic description of superconductive tunneling it is necessary to know the analytical form of the Cooper pair operators. However, a microscopic discussion of the Cooper pair operators and their influence on the transport properties of the hybrid superconductor-quantum dot junction is rather rare in the literature. In this section we show the connection between the Cooper pair operators and ground state of the particle number conserving lead Hamiltonian. Starting from the definition of Eq.~(\ref{eq:definition_s}), we can formally define the Cooper pair annihilation operator \cite{Schrieffer} as \begin{equation}\label{eq:explicit_def_s} \operatorname{\hat S}= \sum_{M = 0 }^{\infty} \sum_{\{n_{k\sigma}\}} \ket{\{n_{k\sigma}\},2M} \bra{\{n_{k\sigma}\},2M +2} , \end{equation} where $\{n_{k\sigma}\} = \{n_{k_1\sigma_1},n_{k_2\sigma_2},\dots\}$ is a set of quasiparticle occupation numbers. It follows that \begin{equation} \operatorname{\hat S}\operatorname{\hat S}^{\dagger} = 1, \end{equation} where we used \begin{equation} 1 = \sum_{M =0 }^{\infty}\sum_{\{n_{k\sigma}\}} \ketbra{\{n_{k\sigma}\},M}. \end{equation} In the full Hilbert space the Cooper pair creation and annihilation operators do not commute \begin{equation} \big[\operatorname{\hat S},\operatorname{\hat S}^{\dagger}\big] = \hat{\mathcal{P}}_0, \end{equation} where $\hat{\mathcal{P}_0}$ is the projector to states with zero Cooper pairs: \begin{equation} \hat{\mathcal{P}}_{0} = \sum_{\{n_{k\sigma}\}} \ket{\{n_{k\sigma}\},0} \bra{ \{n_{k\sigma}\},0}. \end{equation} Using that $\hat N \ket{\{n_{k\sigma}\}, M } = ( N_{\{n_{k\sigma}\}}^{\text{\tiny{QP}}} + M) \ket{\{n_{k\sigma}\}, M } $, with $N_{\{n_{k\sigma}\}}^{\text{\tiny{QP}}}$ being the number of quasiparticles in the string ${n_{k\sigma}}$, one obtains: \begin{equation} \begin{split} \big[ \hat N, \operatorname{\hat S} \big] &= -2\operatorname{\hat S},\\ \big[ \hat N, \operatorname{\hat S}^{\dagger} \big] &= 2\operatorname{\hat S}^{\dagger}. \end{split} \end{equation} \section{Rates} \label{app:rates} \subsection{Normal rates} In the stationary limit, ${\tau \rightarrow \infty}$, the normal rates read: \begin{widetext} \begin{equation}\label{eq:gamma+_n->n+1} \begin{split} & \bigl( \Gamma^{+}_{n m m'n'}\bigr)_{\eta}^{N\rightarrow N+1} = \lim_{\tau \rightarrow \infty} \biggl( \frac{1}{\hbar} \biggr)^2 \sum_{ k \sigma \alpha \alpha'} t_{\eta \alpha \sigma} t^*_{\eta \alpha' \sigma} \bra{n} \dan{\alpha \sigma} \ket{m}\bra{m'} \dddag{\alpha' \sigma} \ket{n'} \\ & \int_0^{\tau} dt_2 ~e^{\frac{i}{\hbar} E_{n'm'} t_2} \biggl[ |u_{\eta k}|^2 f^+(E_{\eta k}) e^{+\frac{i}{\hbar} (E_{\eta k} + \mu_{\eta})t_2} + |v_{\eta k}|^2 f^-(E_{\eta k}) e^{-\frac{i}{\hbar}(E_{\eta k} - \mu_{\eta})t_2} \biggr], \end{split} \end{equation} \begin{equation}\label{eq:gamma+_n->n-1} \begin{split} & \bigl( \Gamma^{+}_{n m m'n'}\bigr)_{\eta}^{N\rightarrow N-1} = \lim_{\tau \rightarrow \infty} \biggl ( \frac{1}{\hbar} \biggr)^2 \sum_{ k \sigma \alpha \alpha'} t_{\eta \alpha' \sigma} t^*_{\eta \alpha \sigma} \bra{n} \dddag{\alpha \sigma} \ket{m}\bra{m'} \dan{\alpha' \sigma} \ket{n'} \\ & \int_0^{\tau} dt_2~ e^{\frac{i}{\hbar} E_{n'm'} t_2} \biggl[ |u_{\eta k}|^2 f^-(E_{\eta k}) e^{-\frac{i}{\hbar} (E_{\eta k}+ \mu_{\eta})t_2} + |v_{\eta k}|^2 f^+(E_{\eta k}) e^{+\frac{i}{\hbar}(E_{\eta k} - \mu_{\eta})t_2} \biggr]. \end{split} \end{equation} \end{widetext} In the following we will show how to write the rates in Eqs.~(\ref{eq:gamma+_n->n+1}) and (\ref{eq:gamma+_n->n-1}) in terms of an integral over quasiparticle energies $E_{\eta k}$. Neglecting the lead index $\eta$, the energetic part of Eq.~(\ref{eq:gamma+_n->n+1}) is proportional to \begin{equation}\label{eq:energetic_part_Gamma} \begin{split} \bigl( \Gamma^{+}_{n m m'n'}\bigr)^{N\rightarrow N+1} \propto \sum_k\big(|u_k|^2 F_1(E_k) + |v_k|^2 F_2(E_k) \big) \end{split} \end{equation} where we defined \begin{equation} \begin{split} &F_1(E_k) = f^+(E_k) e^{\frac{i}{\hbar}(E_k + \omega)t_2} ,\\ &F_2(E_k) =f^-(E_k) e^{-\frac{i}{\hbar}(E_k - \omega)t_2}, \end{split} \end{equation} with $\omega = E_{n'm'} + \mu_{\eta}$. Recalling the definition of $u_k$ and $v_k$, c.f. Eqs.~(\ref{eq:def_uk}) and (\ref{eq:def_vk}), we see that \begin{equation}\label{eq:connection_u_v} \begin{split} &|u_k(-\xi_k)| = |v_k(\xi_k)|. \\ \end{split} \end{equation} Writing the sum as $\sum_k \to \int_{-\infty}^{\infty}d\xi_k \rho_N$, and exploiting Eqs.~(\ref{eq:connection_u_v}) and (\ref{eq:u2+v2}) we are able to to write Eq.~(\ref{eq:energetic_part_Gamma}) as: \begin{equation} \begin{split} &\int_0^\infty d\xi_k \big( F_1(E_k) + F_2(E_k) \big) . \end{split} \end{equation} Changing the integration variable from $\xi_k>0 \to E_k$ we obtain\begin{equation} \begin{split} \int_{|\Delta|}^{\infty}dE~D(E)\big( F_1(E) + F_2(E) \big), \end{split} \end{equation} where we defined the superconducting density of states as $D(E)=\rho_N\,\text{Re}\big(\frac{|E|}{\sqrt{E^2 + |\Delta|^2}})$. Due to the definition of the density of states with the real part, we can extend the integral to zero, and use that $F_2(-E) = F_1(E))$ to obtain \begin{equation} \begin{split} \int_{-\infty}^{\infty}dE\,D(E) F_1(E). \end{split} \end{equation} \subsection{Renormalization of the rates}\label{app:renormalization} \begin{figure} \includegraphics[width= 0.6 \columnwidth]{fig20} \caption{Contour in the complex plane used to integrate Eq.~(\ref{eq:lorentzian_integral}). } \label{fig:contour} \end{figure} In the lowest order approximation we find rates which are proportional to the BCS-density of states leading to divergences at the gap edges. We can can renormalize the rates by introducing a finite lifetime $(\gamma/\hbar)^{-1}$ in the exponents of Eq.~(\ref{eq:gamma+_n->n+1}) and Eq.~(\ref{eq:gamma+_n->n-1}). Since we are neglecting coherences the imaginary parts of the rates do not contribute to the dynamics of the system. For example consider the integral appearing in Eq.~(\ref{eq:gamma+_n->n+1}): \begin{equation}\label{eq:integral_renormalization1} \begin{split}& \text{Re}\bigg(\int_{-\infty}^{\infty}dE~\int_0^{\infty} dt_2 e^{\frac{i}{\hbar}(E + \omega + i \gamma)t_2} f^{+}(E)\,D(E) \bigg) \\& = \int_{-\infty}^{\infty}dE~ \frac{ \hbar\gamma}{(E+\omega)^2+\gamma^2}\, f^+(E)\, D(E), \end{split} \end{equation} where we introduced $\omega = E_{n'm'} + \mu_{\eta}$. Generalizing the integral for the cases ($N\rightarrow N \pm 1$) it reads \begin{equation}\label{eq:lorentzian_integral} \begin{split} \hbar \int_{-\infty}^{\infty}dE~ L(E,\omega)\,f^\pm(E)\, D(E) = \hbar \int_{-\infty}^{\infty}dE\, F(E), \end{split} \end{equation} where \begin{equation} L(E,\omega) = \frac{ \gamma}{(E+\omega)^2+\gamma^2} \end{equation} describes the Lorentzian and $F(E)= L(E,\omega) \,f^\pm(E)\, D(E)$. We can solve the integral of Eq.~(\ref{eq:lorentzian_integral}) using residue calculus hence. To this extend we analyze the singularities of the integrand and the area in which the integrand is analytic. The Lorentzian $L(E,\omega)$ has poles at \begin{equation} E = -\omega \mp i \gamma, \end{equation} with the corresponding residues: \begin{equation} Res_{E=-\omega \mp \gamma}\, L(E) = \frac{\pm i }{2}. \end{equation} The poles of the Fermi function $ f^\pm(E)$ are purely imaginary and equally distributed along the imaginary axis: \begin{equation} E= \frac{i \pi }{\beta} (2n + 1) \quad n \in \mathbb{Z}, \end{equation} with the residues \begin{equation} Res_{E= \frac{i \pi }{\beta} (2n + 1) }\, f^{\pm}(E) = \frac{\mp 1}{\beta}. \end{equation} The square roots in the BCS-density of states $D(E)$ have branch cuts along the real axis. In Fig.~\ref{fig:contour} we sketched the contour in the complex plane which is slightly shifted away from the real axis with $\epsilon =1/R$. In the limit $R\rightarrow \infty$ the integral along the semicircle vanishes and we are left with: \begin{equation}\label{eq:lorentzian_integral_2} \begin{split}& \lim_{R\rightarrow \infty} \int_{-R}^{R}dx F(x + i \epsilon) = 2\pi i \sum_{\alpha}Res_{z=\alpha} F(z). \end{split} \end{equation} In the limit $R\rightarrow \infty$ Eq.~(\ref{eq:lorentzian_integral_2}) is mapped back into the real integral of Eq.~(\ref{eq:lorentzian_integral}), and we find: \begin{equation} \begin{split}& \hbar \int_{-\infty}^{\infty}dE~ L(E)\,f^\pm(E)\, D(E) \\ =& \pi \hbar \text{Re}\bigg( f^+(-\omega + i\gamma) \, D(E-\omega + i\gamma) \bigg). \end{split} \end{equation} \end{appendix} \bibliographystyle{apsrev}

Background {#Sec1} ========== Viral hepatitis is the seventh leading cause of death globally \[[@CR1]\], and hepatitis C virus (HCV) is one of the leading causes of liver failure. The prevalence of HCV differs by geographical location. In Africa, a range in prevalence from 1.7 to 14.7% has been reported \[[@CR2]--[@CR5]\]. In Botswana, preliminary studies reported a low HCV seroprevalence of below 1% \[[@CR6]--[@CR8]\] although Botswana is endemic for both human immunodeficiency virus (HIV) and hepatitis B virus (HBV) with the prevalence of HIV/HBV co-infection ranging from 4 to 10.6% \[[@CR6], [@CR7]\]. Treatment of HCV has evolved rapidly with the introduction of direct acting antivirals (DAA) in the last decade. DAAs can achieve sustained virologic responses (SVR) of 95 to 100% \[[@CR9]\], have shortened treatment duration, and have overcome many challenges faced by administering interferon-based regimens. The World health organization (WHO) has recommended that DAA regimens be used for HCV treatment, and the choice of antiviral therapy and duration of treatment of the DAAs are dependent on HCV genotype. Therefore, studies of circulating HCV genotypes are important prior to widescale treatment rollout \[[@CR10]\]. In Botswana, there are no data on HCV genotypes. This study aimed to determine the HCV genotypes circulating in Botswana and to identify clinically relevant mutations within the HCV core region. Methods {#Sec2} ======= This is a retrospective cross-sectional pilot study utilizing 55 stored plasma samples collected between February 2015 and July 2016 from 55 liver disease patients at the liver clinic of Princess Marina Hospital, a referral hospital in Gaborone, Botswana. The study was conducted at Botswana Harvard AIDS Institute Partnership. HCV ribonucleic acid (RNA) was extracted using the Qiagen Viral RNA kit using 140 μl plasma samples according to the manufacturer's specifications (Qiagen, Hilden, Germany). The extracts were stored at − 80 °C prior to genotyping. Amplification targeted the partial core region with two primer sets -- outer core primers (5′ -- ACT GCC TGA TAG GGT GCT TGC -- 3′, nt 288 → 308 and 5′ -- ATG TAC CCC ATG AGG TCG GC -- 3′, nt 732 ← 751) and inner core primers (5′ -- AGG TCT CGT AGA CCG TGC A -- 3′, nt 321 → 339 and 5′ -- CAT GTG AGG GTA TCG ATG AC -- 3′, nt 705 ← 724) previously described, \[[@CR11], [@CR12]\] relative to the H77 reference \[[@CR13]\]. Nested PCR was carried out using Superscript (III) one-step RT-PCR with Platinum Taq DNA polymerase high fidelity kit (Invitrogen, USA) according to the manufacturer's instructions. PCR products were visualized on 1.5% agarose gel stained with ethidium bromide, and the positive amplicons were purified using the ZR DNA sequencing clean up kit (Zymo, Irvine, CA, USA) according to the manufacturer's specifications. Population sequencing was conducted using BigDye Terminator version 3.0 kit (Applied Biosystems; Foster City, CA, USA) sequencing chemistry on a 3130XL ABI genetic analyser (ABI PRISM 3130XL; Applied Biosystems). Sequence files were edited using Sequencher version 5.0 software (Gene Codes Corp., Ann Arbor, MI, USA). Phylogenetic analysis was used to evaluate HCV genotypes and included reference sequences from the Los Alamos HCV sequence database (<https://hcv.lanl.gov/content/index>). Sequences were aligned in Clustal X version 2.1. Additional phylogenetic inference was conducted using the Bayesian Markov chain Monte Carlo (MCMC) in the Bayesian Evolutionary Sampling Trees (BEAST) version 1.7.5 as described previously \[[@CR8], [@CR14]\]. Posterior probabilities \> 80% were deemed statistically significant. Sequences were exported to BioEdit software where nucleotide sequences were translated to amino acids. H77 (accession number AF009606 \[[@CR13]\]) served as the reference sequence, and mutations were visually analysed per amino acid position. Importantly, H77 is a genotype 1a reference; therefore, several genotype 1 references together with genotype 4 and 5 reference sequences were included in the analysis to distinguish between polymorphic regions that differ by genotype and true mutations. Sequences were submitted to National Center for Biotechnology Information (NCBI) GenBank under accession numbers MK392625 to MK392630. Statistical analysis was performed using R version 3.6.0 \[[@CR15]\]. Results {#Sec3} ======= Study participants were Batswana adults and 63.3% were female. The age ranged from 16 years to 74 years with a median of 44 years and an interquartile range (IQR) of 32--55 years. From the 55 participant samples we had access to, 6 (10.9%) of those were RNA positive as shown in Table [1](#Tab1){ref-type="table"}. The six participants had ages, ranging from 24 to 70 years with a median of 55.5 years as shown in Fig. [1](#Fig1){ref-type="fig"}a, and HCV was observed in older males and middle-aged women as shown in Fig. [1](#Fig1){ref-type="fig"}b. Table 1Characteristics of the HCV viraemic positive participantsSequence IDGenderGenotypeSubtypeAccession NumberLB_1Male55aMK392625LB_2Female55aMK392626LB_3Female55aMK392627LB_4Male55aMK392628LB_5Female44vMK392629LB_6Female44vMK392630 Fig. 1**a** Boxplot that shows the age distribution of HCV positive participants with a median of 55.5 years and interquartile range of 39.25--57.5 years. **b** The age distribution was further stratified by gender and observed in older men (\~ 65 years) and in middle-aged women Sequence analysis of 6 partial core regions revealed that four of the Botswana sequences belong to genotype 5a, as shown by the clustering with other genotype 5a strains from South Africa, Ethiopia, and Denmark. Two Botswana strains belong to genotype 4v as shown by the clustering pattern with other genotype 4v strains from Ethiopia, Cyprus, and the United Kingdom as shown in Fig. [2](#Fig2){ref-type="fig"}. Fig. 2Phylogenetic tree using Bayesian analysis of the 6 amplified Botswana HCV sequences and GenBank references. Reference sequences are labelled with the genotype + accession number + country. The genotype 5a sequences from Botswana are highlighted in red with the prefix LB, while the genotype 4v sequences from Botswana are highlighted in blue with the prefixes LB Further analysis was performed to identify core mutations within the viral sequences as shown in Table [2](#Tab2){ref-type="table"}. Table 2Core mutations observed in Botswana sequences and their association with disease progression and/or therapeutic responseMutationGenotypeMutation frequencyClinical relevanceReferenceG95E4v1Not characterized\[[@CR16]\]R9T5a4Not characterized\[[@CR17]\]K10Q5a4Increased risk of HCC\[[@CR18]\]T15I5a4Not characterized\[[@CR19], [@CR20]\]L36 V5a1Not characterized\[[@CR21]\]R70Q5a4Increased risk of HCC + poor response to IFN\[[@CR17], [@CR22], [@CR23]--[@CR25]\]*Abbreviations*: *HCC* Hepatocellular carcinoma, *IFN* Interferon Most mutations were observed in genotype 5 samples (R9T, K10Q, T15I, L36 V and R70Q). Only one mutation (G95E) was observed in genotype 4. Discussion {#Sec4} ========== This is the first report on the circulating HCV genotypes from Botswana. The finding of genotypes 5a and 4v agrees with previously reported data from southern Africa \[[@CR5], [@CR26], [@CR27]\]. Interestingly, genotype 4v has not been observed previously in southern Africa, although other genotypes, such as 4c, 4 g, 4 k, 4q and 4r, have been reported \[[@CR28], [@CR29]\]. Genotype 4v has been reported in central Africa and the Middle East. In two studies, genotype 4v was observed in Rwanda and Ethiopia \[[@CR30], [@CR31]\]. To achieve elimination of HCV by 2030, DAAs should be introduced and administered to patients as per genotype \[[@CR10]\]. According to the observed genotypes, Botswana will need to procure Ledipasvir, Daclatasvir and Sofosbuvir as the DAA regimens recommended to treat genotypes 4 and 5 for patients with and without liver cirrhosis \[[@CR10]\]. In this study, the HCV diversity in Botswana is low as compared to South Africa, since only two subtypes were observed. This observation could be due to small sample size and/or varying transmission dynamics. In the current study the core gene was amplified as the region of interest, despite the recommended NS5B region for HCV genotype classification \[[@CR32]\]. We selected a significant fragment of the core region to classify the HCV isolates, which also includes (but is not limited to) putative HCC core related mutations, since the core region has higher amplification rates compared to the NS5B region as previously reported \[[@CR12]\]. The core mutations in genotype 5a -- R9T, K10Q, T15I, L36 Vand R70Q -- account for most of the observed mutations. The K10Q mutation is associated with increased risk of HCC \[[@CR18]\]. This mutation was the nucleotide substitution A28C in the core gene bringing about an amino acid substitution from lysine to glutamine. Further analysis from 100 genotype 5 sequences downloaded from the GenBank showed that only 5% of these references contained the K10Q mutation. Interestingly, all four genotype 5 samples from the Botswana sequences had that mutation. Several studies in genotype 1 strains have indicated that the R70Q mutation increases the risk of HCC due to its oncogenic effect \[[@CR18], [@CR23]--[@CR25]\] and also confers resistance to IFN-based treatments \[[@CR22]\]. This mutation occurs due to nucleotide substitution G209A resulting in an amino acid change from arginine to glutamine \[[@CR18]\]. From the analysis with the GenBank genotype 5 sequences, it was interesting to note that 81% of the references contained the R70Q mutation compared to 16% with the wildtype amino acid. In a study conducted in South Africa, the R70Q mutation was observed in 90% of genotype 5a blood donors \[[@CR33]\]. Whether the impact of this mutation in genotype 1 strains is the same in genotype 5 remains unclear. Presence of this mutation in only genotype 5 isolates could either be due to transmission of a mutated strain or a naturally occurring drug resistant mutation to the SOC \[[@CR33]\]. However, the latter assumption cannot be confirmed since there are no data on duration of treatment. Despite the scarce data on impact of this mutation in genotype 5, all four genotype 5 individuals may have a poor response to the current SOC. Thus, DAA regimens may improve treatment of individuals infected with HCV strains harbouring these mutations. Only one mutation -- G95E -- was observed in one genotype 4v sample. However, this mutation has not been well characterized or associated with altering viral fitness or drug sensitivity. Most reported mutation analyses have been conducted on genotype 1 strains \[[@CR20]\]. Therefore, more longitudinal studies in mutation analysis based on other genotypes, such as 4 and 5, are warranted, as mutations differ by genotype. A major limitation of the study was the modest population size evaluated and the low amplification success rate; therefore, there was a limited number of positive samples available for analysis. Furthermore, as no HCV antigen enzyme-linked immunosorbent assay (ELISA) was conducted due to low sample volumes, the low amplification rate may reflect a low proportion of current HCV infections versus seropositive but resolved infections. The low amplification rate may also be due to multiple freeze-thaw cycles for sample use in other studies prior to the current analysis. The cross-sectional nature of this study made following up on the patients with HCC-risk factors impossible. Therefore, we could not observe the oncogenic effect of the reported mutations in genotype 5 patients. Conclusion {#Sec5} ========== In summary, genotypes 5a and 4v are the circulating genotypes in Botswana. The mutations observed in this study confer resistance to the SOC, and, as per WHO recommendations, there is a need to introduce DAAs as the new SOC for Botswana in order to achieve elimination of HCV by 2030. The DAAs for Botswana HCV patients should include Ledipasvir, Daclatasvir and Sofosbuvir. A longitudinal study with a larger representative population is warranted to develop more understanding of the HCV epidemiology in Botswana. DAA : Direct acting antivirals DNA : Deoxyribose nucleic acid ELISA : Enzyme-linked immunosorbent assay HBV : Hepatitis B virus HCC : Hepatocellular carcinoma HCV : Hepatitis C virus HIV : Human immune-deficiency virus IFN-α : Interferon -alpha IQR : Interquartile range MCMC : Markov chain Monte Carlo NCBI : National Center for Biotechnology Information RNA : Ribonucleic acid RT-PCR : Reverse transcriptase polymerase chain reaction SOC : Standard of care SVR : Sustained virologic response WHO : World health organisation **Publisher's Note** Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. The authors would like to acknowledge the study participants. We also acknowledge the University of Botswana, Princess Marina Hospital, National University of Science and Technology in Zimbabwe and the Botswana Harvard AIDS Institute Partnership for their support and contributions to the study. LB, MA, SS, SG, SM and WTC conceived and designed the experiments. LB, CCB, SS, ZMS, SM, TM, BBP, PM and WTC conducted the experiments. LB, WTC, JTB, MA, SM and SG analysed the results. EZ, RMM, ME, SG and JTB provided expert review of the manuscript. All authors contributed to the writing and reviewing of the manuscript. All authors reviewed and approved the final manuscript. MA, WTG, SM and SG were funded by the Sub-Saharan African Network for TB/HIV Research Excellence (SANTHE), a DELTAS Africa Initiative \[grant \# DEL-15-006\]. The DELTAS Africa Initiative is an independent funding scheme of the African Academy of Sciences (AAS)'s Alliance for Accelerating Excellence in Science in Africa (AESA) and supported by the New Partnership for Africa's Development Planning and Coordinating Agency (NEPAD Agency) with funding from the Wellcome Trust \[grant \# 107752/Z/15/Z\] and the UK government. The views expressed in this publication are those of the author(s) and not necessarily those of AAS, NEPAD Agency, Wellcome Trust or the UK government. SM was partially supported was supported by the Fogarty International Center and National Institute of Mental Health, of the National Institutes of Health under Award Number D43 TW010543. The opinions, findings and conclusions expressed in this article are those of the authors and do not necessarily reflect the views of the sponsors. The datasets used to support the results of this study are available from the corresponding author upon request. Sequences are available in the National Center for Biotechnology Information (NCBI) GenBank under accession numbers MK392625 to MK392630. The study was approved by the Health Research and Development Committee (HRDC) at the Ministry of Health and Wellness (HPDME 13/18/1 X (897). Permission to access raw data was granted by the ethics committee of the university of Botswana and Princess Marina hospital. All participants consented prior to participation in the study. Not applicable. SM is an editorial board member for BMC Infectious disease. The remaining authors declare no conflict of interest.

Background ========== Scientific background and explanation of rationale -------------------------------------------------- Inguinal hernias are the most common hernia; they account for 90% of all spontaneous hernias. Moreover, inguinal hernia repair is the most frequently performed procedure in general surgery. In Germany, more than 200,000 inguinal hernias are repaired annually \[[@B1],[@B2]\]. The standard method for repairing an inguinal hernia, originally described by Bassini in 1889, is to close the inguinal canal with sutures. Due to the high recurrence rate with this technique \[[@B3]\], new methods were established that used tension-free implantation of synthetic meshes. Furthermore, endoscopic/laparoscopic methods have been introduced, where the hernia canal is typically approached from the posterior side; this approach is directly opposite to the anterior approach used in the open surgery technique. Which of these approaches and methods yields the better results is strongly debated \[[@B4]\]. Among the various posterior techniques employing preperitoneal mesh implantation \[[@B2],[@B5]-[@B7]\], the two most widely accepted techniques are the transabdominal preperitoneal patch plastic (TAPP) repair and the total extraperitoneal (TEP) repair \[[@B8],[@B9]\]. There are many indications for both techniques, but the TAPP repair is particularly recommended for recurrent hernias (after an open preperitoneal patch plastic) and difficult hernias (sliding or incarcerated hernias) \[[@B2],[@B8],[@B10]\]. The TAPP repair has the advantage that it is easier to perform, can be better standardized, and offers the possibility to perform a diagnostic laparoscopy \[[@B2]\]. Thus, the type of hernia can be assessed immediately on both sides of the groin, and a bilateral repair can be performed without additional incisions. In general, the TAPP repair is easier to learn than the TEP repair \[[@B2],[@B10],[@B11]\]. Most randomized studies that compared laparoscopic with open repair found that laparoscopy was associated with less postoperative pain, earlier return to work, higher costs, a longer operating time, a longer learning period, and a higher recurrence and complication rate during the early learning phase \[[@B12]-[@B14]\]. In summary, open, mesh-based, tension-free repair remains the standard procedure; however, in the hands of a well-trained surgeon laparoscopic herniorrhaphy can produce excellent results comparable with those of open repair \[[@B14],[@B15]\]. Many factors determine whether patients experience postoperative pain, including the type of intervention, the presence of complications, patient age, and individual tolerance. In particular, the type of mesh fixation employed for the hernioplasty has a strong influence on postoperative pain \[[@B16],[@B17]\]. In a recent review, mesh fixation with glue was compared to mesh fixation with staples for an endoscopic inguinal hernia repair \[[@B18]\]. For the TAPP repair, two randomized controlled trials \[[@B19],[@B20]\], one non-randomized trial \[[@B21]\], and two case series \[[@B22],[@B23]\] were reviewed. The authors of the review found less postoperative pain and more rapid recovery after glue fixation than after staple fixation, without any significant difference in the recurrence rate. Nevertheless, the authors concluded that, due to the overall poor quality of the currently published data, future well-designed studies are required to demonstrate the superiority of fibrin glue over mechanical stapling for mesh fixation. Thus, well-designed, randomized studies that compare glue and staple fixations for TAPP repair remain warranted. Most previous studies applied only one mesh fixation method for patients with bilateral hernias. Only one study compared glue and staple fixations in single patients with bilateral hernias \[[@B24]\]; interestingly, that study was not included in the review by Schäfer et al. \[[@B18]\]. This study protocol describes a two-by-two trial design that will be used to compare the two fixation methods in two groups of randomized patients; one group with unilateral hernias and the other group with bilateral hernias. The latter group will allow a comparison of the two methods in a single patient. This design will eliminate the inter-individual bias that may influence the unilateral hernia group. This study focuses on early postoperative pain, particularly during the first two postoperative days. In many other studies, pain is first measured one month after surgery. Aim of the study ---------------- The aim of this trial is to investigate early and late postoperative pain after laparoscopic hernia repair with the TAPP technique, using either fibrin glue or staples for mesh fixation. Methods/Design ============== The study was planned according to the updated Consolidated Standards of Reporting Trials (CONSORT) statement. The study design is in compliance with the Declaration of Helsinki, the Guidelines of Good Clinical Practice issued by ICH, and the requirements of Swiss regulatory authorities \[[@B25]-[@B27]\]. Trial design ------------ This trial is a double blind, randomized, controlled trial with a two-by-two design, conducted at a single institution. The study aims to compare pain after mesh implantation, when fixed either with staples or glue, in a superiority analysis. The CONSORT diagram of the trial is depicted in Figure  [1](#F1){ref-type="fig"}. {#F1} Eligibility criteria and patient recruitment -------------------------------------------- Patients scheduled for elective laparoscopic primary inguinal hernia repair at the Department of Surgery, Kantonsspital St. Gallen, will be screened for enrollment in the trial. Informed consent will be obtained, at the latest, on the day before surgery, after ascertaining that the patient meets the eligibility criteria. Patients that match the following criteria are eligible for inclusion in the clinical trial: • age, 18 years or older • male gender • primary hernia repair (no re-operations for a recurrent hernia) • no chronic pain • no acute pain that requires analgesics, other than paracetamol or more than a single dose of non-steroidal anti-inflammatory drug treatment within 48 h before surgery • no current treatment with psychopharmaceutical drugs • good compliance can be expected • no mental incapacity • no known incompatibility (allergies) with Tisseel compounds • informed consent Under the following circumstances, patients will be excluded from the trial: • hernia is inoperable with TAPP, due to the size, previous abdominal surgery (assuming a damaged peritoneum), or expected adhesions • repair requires an intra-operative change to another hernia repair technique (e.g. open access methods) • the patient wishes to stop further trial participation When the patient refuses further participation after receiving surgery, the patient will receive a safety check-up during a one-month follow-up. Otherwise, the patient will not receive the study treatment and will not require a study-related safety follow-up. Study objectives and endpoints ------------------------------ The primary objective of this trial is to show that after TAPP repair for unilateral and bilateral inguinal hernias the acute postoperative pain caused by mesh fixation with tissue glue is less than with staple fixation. The primary efficacy endpoint of the TISTA trial is pain under physical stress at 24 h after surgery. Pain will be rated by the patient using a numeric rating scale (NRS). The NRS ranges from 0 to 10 in steps of one, with 0 representing no pain at all and 10 the worst imaginable pain. The NRS is comparable to the visual analogue scale (VAS) where the degree of pain is marked on a 10 cm long line. In practical terms, average NRS and VAS scores (in cm) can be considered equivalent. The patient is first asked to rate the pain at rest and then to rate it after physical stress. Physical stress is induced by a 90° bending of the hip joint on the operated side(s). The secondary endpoints of the trial and their definitions are summarized in Table  [1](#T1){ref-type="table"}. ###### Secondary endpoints and definitions **Secondary endpoints** **Definitions** ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------- Pain rated before surgery and at 4, 8, 12, 36, 48 h after surgery Measured on a numeric rating scale from 0-10 at rest and upon bending the hip joint on the operated side Operating time Time, in minutes, from the first skin incision to the application of dressing Length of hospital stay Time, in days (with one decimal precision), from start of surgery to hospital release Postoperative analgesic requirements Amount (g/day) and type (paracetamol, metamizole, morphine) of analgesic required after surgery, before hospital discharge Incidence of persistent pain (neuralgia) Neuralgia defined as:  • The presence of intermittent hyperesthesia, burning sensation, or jabbing pain in the ipsilateral, inguinal area nerves (genitofemoral nerve, lateral cutaneous femoral nerve, ilioinguinal and iliohypogastric nerve) \[[@B20],[@B22],[@B27]\]  • Evaluation will be performed 14 days, 1, 3 and 12 months after surgery  • The degree of pain will measured as described above Postoperative morbidity Categorized according to Dindo et al \[[@B28]\], and it includes:    • Wound infection, defined as infections treated without further surgery and identified by clinical examination without microbiological confirmation  • Hematoma or seroma formation, identified by clinical examination alone, before discharge from hospital, which does not require radiological confirmation  • Re-operation, defined as the need for re-operation during the initial hospital stay  • Bleeding  • Urinary retention, urinary tract infection  • Pulmonary infection  • Any serious deviation from the normal postoperative course Time to return to normal activities Time, in days, from hospital discharge to first working day Relapse Within 1 year of surgery Economic impact Calculated based on the following factors:  • Cost for glue, staples, and instruments to apply glue or staples  • If relevant: cost for excess operating time for the "slower" mesh fixation  • Length of hospital stay (cost/day)    • Medical leave of absence (converted into a monetary amount based on an average salary) Standardization of treatments ----------------------------- Patient intra- and peri-operative care will be standardized and maintained identical for each patient, except for the mesh fixation technique. The patients will receive combined neuraxial and general anesthesia. Transabdominal preperitoneal patch plastic repair (TAPP) -------------------------------------------------------- The abdomen will be accessed and the pneumoperitoneum will be achieved with standard laparoscopic techniques. The preperitoneal space will be exposed transabdominally with a sharp incision, followed by bluntly stripping the peritoneum that overlies the inguinal anatomy. An Ultrapro® mesh (Ethicon Inc., Johnson & Johnson, Somerville, NJ, USA) will be deployed and fixed in place with either staples or glue, and then the peritoneum will be returned to its anatomical position. Surgical technique ------------------ At an umbilical site, a Veress needle is inserted to induce a pneumoperitoneum, and then the needle is replaced with a 10 to 12 mm optical trocar. Next, two 5 mm trocars are positioned bilaterally on the umbilical line in the iliac fossa. An incision is made in the peritoneal wall, starting at the level of the superior margin of the internal inguinal ring and at the level of the epigastric vessels. The incision is extended medially, up to the residue of the umbilical artery, and then laterally, 3 to 4 cm past the inguinal ring; the total incision length is 7 to 8 cm. In the presence of a direct hernia, the hernial sac is directly isolated and reduced. In the case of an indirect or femoral hernia, the preperitoneal parapubic adipose tissue is carefully dissected medially to expose the horizontal pubic ramus and Cooper's ligament. Accurate dissection of the preperitoneal retrovesical tissue facilitates positioning the mesh. The internal inguinal ring is explored, and the hernial sac is isolated and reduced; this maneuver is performed to reveal the presence of perihernial lipomas, which can then be removed. Once the spermatic cord is freed from the peritoneal wall, a macroporous, partially absorbable Ultrapro mesh (Ethicon) is positioned. This type of mesh allows the surgeon to view the underlying structures, which facilitates positioning the mesh. The mesh is cut to 10 × 13 cm and placed in the preperitoneal space such that it is in medial contact with the paravesical area, it covers Cooper's ligament, it rests on the inguinal region, and it extends laterally over the epigastric vessels. Stapling -------- For this procedure, the mesh is fixed with an Endopath Multifeed Stapler with a 10 mm shaft and helical titanium staples (Protack, Covidien, Mansfield, MA, USA). The technique that we adopted involves positioning two metal clips at the level of Cooper's ligament and the pubic tubercle. Additionally 4 staples are placed laterally at the level of the deep inguinal ring medially and laterally to the inferior epigastric vessels with absorbable staples (Securestrap, Ethicon). The peritoneal flap is then closed with four additional absorbable staples. Tisseel™ tissue glue -------------------- For this procedure, the mesh is anchored with 2 mL of fibrin glue (Tisseel, Baxter, Deerfield, IL, USA). Tisseel is applied to both anterior and posterior sides of the mesh with a dedicated laparoscopic tool (Duplotip, Baxter) inserted into a 5 mm trocar. To obtain optimal bonding between the Tisseel on the mesh and the peritoneal wall, slight pressure is applied to the mesh with the Duplotip. The Tisseel is applied to the entire perimeter of the mesh, and in particular, at the level of the superior margin, the "triangle of disaster," in proximity of the prevesical fat, to assure good adhesion. The peritoneal flap is then closed with small, continuous, absorbable 2/0 sutures. Postoperative care ------------------ To reduce bias, the postoperative regime is standardized for the study. All patients will receive the same intra- and perioperative care, except for the mesh fixation technique. The postoperative care will also be uniformly administered, and changes will only be made when the patient's comfort or safety is compromised. All patients will be mobilized as soon as possible during the postoperative period. All patients will receive paracetamol (three 1 g doses per 24 h), administered either intravenously (early postoperative course) or orally (late postoperative course). Additional analgesia will be provided as required according to the following scheme: • First line reserve: metamizole, maximum 4 g/24 h • Second line reserve (if necessary): morphine, 1-2 mg, delivered intravenously The use of additional analgesics will be documented. Trial interventions ------------------- Patients will be randomized according to the following scheme: Primary arm 1: Patients with a unilateral inguinal hernia secondary arm 1: mesh fixation with tissue glue secondary arm 2: mesh fixation with staples Primary arm 2: Patients with bilateral inguinal hernias secondary arm 1: right side, tissue glue; left side, staples secondary arm 2: right side, staples; left side, tissue glue Methods for reducing bias ------------------------- The patients will be randomly assigned to the treatment, and the patients will be blinded to the assigned treatment. In primary arm 2 (patients with bilateral hernias), the bias of individual pain perception can be eliminated, because a single individual decides (blindly) which side is more painful. The treatment assignment will be based on a block randomized list (one for each primary arm) with variable block sizes (between 2 and 6 patients per block). After each patient signed the informed consent, his name will be entered into the study database. At the beginning of surgery, the patients scheduled for unilateral repair will be examined laparoscopically to determine whether the other side also requires repair. The final decision for uni- or bilateral repair can only be made after this intra-operative examination. Thus, randomization will occur during surgery by contacting the study nurse to receive the treatment allocation. In the absence of the study nurse, the data manager can provide the allocation information. The data manager generates the randomization list and maintains a secured copy of the assignment list. The pain scores are obtained by nurses or physicians that are not aware of the treatment assignment (observer blinding). In the case of unforeseen postoperative complications, the treating physician (surgeon) will know the treatment assignment and can be reached by phone. Furthermore, the data manager can unblind the assignment when necessary. Any unblinding will be recorded in the study documentation. Sample size calculation ----------------------- The sample size was calculated under the hypothesis that either glue fixation or staple fixation is superior for the primary endpoint (= pain after 24 h)(two-sided analysis). Lovisetto reported a mean VAS pain score of 1.9 for glue fixation and 2.6 for staple fixation one month after TAPP repair \[[@B19]\]. The observed difference of 0.7 is usually considered to be of no clinical relevance. We thus assumed a difference of 1 as a clinically relevant difference. Standard deviation was estimated to be equal 1, resulting in a standardized effect size of 1 which generally is considered clinically relevant. The sample size calculation was performed as two-sided analysis with a type I error of α = 0.05 and a power of (1-β) = 0.90. Based on these assumptions, we calculated that 23 patients per arm would be required for unilateral hernias (t-test for two independent samples with common variance). Assuming a drop-out rate of 20%, the total number of patients needed per arm is 29, resulting in a total of 58 patients with unilateral hernias. For the second arm with bilateral hernias we calculated that 19 patients would be required (t-test for paired samples), assuming an intra-individual correlation of 0.2 (derived from a preliminary, unpublished analysis of patients with bilateral inguinal hernias that had undergone a Lichtenstein procedure). With a drop-out rate of 20%, a total of 24 patients with bilateral hernias would be required, resulting in a total of 82 patients for the entire study. With an annual case load of approximately 100 in the previous years, we estimate that patient recruitment should be achieved within one or two years. Statistical analysis -------------------- The statistical analysis will be performed with SPSS Statistics for Windows (IBM Corp., Armornk, NY, USA). A two-sided p-value \<0.05 will be considered statistically significant. Descriptive statistics will be used to analyze the baseline characteristics. For the superiority analysis of pain intensity at 24 h after surgery, the Mann−Whitney U and Wilcoxon tests will be applied, as appropriate. To adjust for multiple testing, the Hochberg procedure will be applied \[[@B29]\]. Pain measurements performed at 4, 8, 12, 24, 36, 48 h after surgery by patients with unilateral and bilateral hernias will be analyzed in one mixed model to assess the mean ranks of the NRS scores with adjustments for time, treatment, and unilateral versus bilateral hernias. To compare other continuous secondary endpoints (e.g., operating time, length of hospital stay, postoperative administered analgesics, time to return to normal activity, and economic impact), the Mann−Whitney U and Wilcoxon tests will be applied, as appropriate. Chi-square statistics will be used to compare categorical secondary endpoints (e.g., incidence of chronic pain, rate of recurrence, and postoperative morbidity). Auxiliary, non-confirmatory analyses will be performed to assess the influences of baseline patient characteristics and treatment characteristics on the primary and secondary outcomes. Intermittent missing values will be replaced by linear interpolation (e.g. post-operative pain values within 48 h after surgery). No interim analysis is planned for this study. Data from patients withdrawn from the study will be disregarded, unless exclusion is based on a postoperative patient request, and the patient agrees to the use of existing documented data. The confirmatory analysis will be performed based on an intention-to-treat (ITT) design, and it will adhere to ITT principles. When a protocol violation occurs, a per protocol analysis (PP) will be performed for comparison. Ethical and legal considerations -------------------------------- This study will be conducted in agreement with either the Declaration of Helsinki (Tokyo, Venice, Hong Kong, Somerset West, and Edinburgh amendments) or the local laws and regulations of the country, whichever provides the greatest protection for the patient. This protocol, the patient information sheet, and the patient consent form have been reviewed and approved by the local Ethics Committee (EKSG 12/080) and by Swissmedic (TpP_I\_2012_003) prior to enrolling any patients in this trial. All patients will be informed of the aims of the study, the possible adverse events, the procedures and possible hazards to which they will be exposed, and the mechanism of treatment allocation. They will be informed about the strict confidentiality of their patient data and told that their medical records may be reviewed for trial purposes by authorized individuals other than their treating physician. The signed consent document will be maintained by the investigator. A copy of the signed consent document will be given to the patient or the patient's legally authorized representative. It will be emphasized that participation is voluntary, that the patient is allowed to refuse further participation in the study at any time, and that any refusal will not influence the patient's subsequent care. Documented informed consent must be obtained for all patients included in the study before they are registered or randomized in the study. Registration and randomization will be conducted in accordance with the national and local regulatory requirements. Maintaining patient confidentiality is the responsibility of the investigator. During the trial, the patients will be identified solely based on an individual identification code. The trial findings will be stored in accordance with local data protection law/ICH GCP Guidelines, and they will be handled in the strictest confidence. For the protection of these data, organizational procedures have been implemented to prevent distribution of the data to unauthorized individuals. The investigator will maintain a personal subject identification list (screening numbers with the corresponding subject names) to ensure the records can be identified when necessary. Discussion ========== This study was designed as a randomized trial with a two-by-two, parallel design. We aimed to compare patients with unilateral hernias that received mesh implants, fixed with either tissue glue or staples and to evaluate patients with bilateral hernias that received mesh implants fixed with tissue glue on one side and staples on the other side. In the latter group, the same patient will rate the pain associated with both methods; these ratings will not be influenced by inter-individual biases present in the unilateral hernia arm. To further reduce biases, both patients and observers (staff evaluating the pain) will be blinded to the assigned treatment. We expect that mesh implants fixed with tissue glue will be associated with less postoperative pain, shorter hospital stay, and a shorter sick leave after the operation compared to implants fixed with staples. In addition to enhancing patient comfort, the tissue glue fixation will also have an economic impact, by reducing health care costs. Because we can evaluate both inter-individual and intra-individual biases in pain perception, this study will provide important insights for future studies that aim to investigate pain. This study will therefore provide novel information that is currently lacking in the literature. Trial status ------------ The trial began on February 1^st^, 2013, and recruitment is ongoing. Abbreviations ============= CONSORT: Consolidated Standards of Reporting Trials; GCP: Good clinical practice; ICH: International Conference on Harmonisation; ITT: Intention to treat; NRS: Numeric rating scale (for pain measurement); PP: Per protocol; TAPP: Transabdominal preperitoneal patch plastic; TEP: Total extraperitoneal (repair); TISTA: This study's acronym (**TI**sseel versus **STA**ples); VAS: Visual analogue scale (for pain measurement). Competing interests =================== The authors declare that they have no competing interests. Authors' contributions ====================== SAM and IT developed the original study design. SAM, UB, and IT developed the research protocol. RW and IT performed the sample size calculations. SAM, BMS, CL, and IT are responsible for the clinical input. SAM, RW, IT, and UB drafted the manuscript. All authors have approved the final manuscript. Pre-publication history ======================= The pre-publication history for this paper can be accessed here: <http://www.biomedcentral.com/1471-2482/14/18/prepub> Acknowledgements ================ This trial is supported by a grant from the Clinical Trial Unit of the Kantonsspital St. Gallen.

\section{Conclusions} We demonstrate the first deep learning method for real-time and streaming target sound extraction. Future work includes the use of more constrained computing platforms, larger datasets with more classes, and multiple microphones. Our Waveformer architecture may be applicable to other acoustic applications like source separation and directional hearing, which deserves further exploration \section{Introduction} \label{sec:intro} Humans are exceptionally adept at attending their auditory focus to specific sounds even in a noisy environment~\cite{2020arXiv200605712O}. Recent works that aim to create a computational equivalent of this human capability formulate this problem as target sound extraction~\cite{2020arXiv200605712O,delcroix2021few,delcroix2022soundbeam}. The goal is to extract sound signals of interest from a mixture of various overlapping sounds, given clues which provide information about the target sound class such as embeddings of a one-hot label~\cite{2020arXiv200605712O}, audio clips~\cite{delcroix2022soundbeam,gfeller2021one}, and images~\cite{gao2019co,xu2019recursive}. Streaming target sound extraction could enable real-time intelligent acoustic applications for headphones, hearing aids and telephony by filtering out undesired sounds from the environment (e.g., traffic) and presenting only sounds of interest to the user (e.g., sirens). Recent works on target sound extraction have shown promising performance even for mixtures containing a large number of sound classes~\cite{2020arXiv200605712O}. However, none of these prior works demonstrate real-time streaming capabilities. In particular, the prior works for this task are based on non-streaming models and designed for offline processing, where the neural network has access to a large block ($\ge 1$~s) of audio samples~\cite{2020arXiv200605712O}. In contrast, real-time streaming applications impose significant {algorithmic and} computational constraints, requiring networks to operate on small blocks ($\le 10$~ms) with a limited number of lookahead samples for each block. All these factors can significantly degrade the performance~\cite{luo2019conv}. In this paper, we present the first deep learning method to perform target sound extraction in a streaming manner. Fig.~\ref{fig:main_fig} shows Waveformer, our encoder-decoder architecture where the encoder is a stack of dilated causal convolution (DCC) layers and the decoder is a transformer decoder \cite{https://doi.org/10.48550/arxiv.1706.03762}.\footnote{We call our network, Waveformer, since it uses an hybrid architecture with the causal convolution layers, common in WaveNet ~\cite{https://doi.org/10.48550/arxiv.1609.03499} based architectures, as the encoder and a transformer as the decoder.} {Our intuition is that much of the complexity in prior models comes with processing large receptive fields, especially at high sampling rates. For example, recent transformer based architectures proposed for speech separation \cite{subakan2022resource, luo2022tiny} implement chunk-based processing, where each chunk independently attends to all the chunks in the receptive field. Thus, to achieve a receptive field of length $R$, for each chunk, these models have an $\mathcal{O}(R)$ computational complexity. Instead, since DCC layers have a complexity of $\mathcal{O} (\log R)$ for achieving the same amount of receptive field (\xref{sec:encoder}), we use a stack of DCC layers as the encoder that processes the receptive field. We then use the decoder layer of the transformer architecture \cite{https://doi.org/10.48550/arxiv.1706.03762} as our model's decoder. We leverage the \textit{memory--target} paradigm of the transformer decoder, where the query-conditioned encoder's output is considered as the transformer decoder's `target' and the unconditioned encoder's output is considered as the transformer decoder's `memory'. Using self-attention on the `target' followed by cross-attention between the `target' and `memory,' the decoder generates the mask for extracting the specified target sound to produce the output signal.} To evaluate our network architecture, we implement a causal version of Conv-TasNet and a streaming version of ReSepformer~\cite{subakan2022resource} for the task of streaming target sound extraction. Evaluations show that our hybrid network architecture achieves state-of-the-art performance for this task. Further, the smallest and largest versions of our model have real-time factors (RTFs) of 0.66 and 0.94, respectively, on a consumer-grade CPU, demonstrating the real-time target sound extraction capability. This is shown to be faster than the two representative prior models while outperforming them in terms of the output signal quality. \section{Waveformer Architecture} We process individual audio chunks of duration $\tau$ seconds. For streaming, we need to operate at chunk level: an output chunk can depend on the current and past chunks. Thus, streaming models have an intrinsic latency equal to the duration of a single chunk. In real-time practical systems, it is desirable that this latency is on the order of 10 ms \cite{7952149}. Fig.~\ref{fig:main_fig} shows our proposed time-domain model architecture, which employs an encoder-decoder based mask generation network to generate an element-wise multiplicative mask in the latent space. Let $x_k \in R^{S}$ denote the current input audio chunk, where $S = \tau F_s$ is the number of audio samples included in the current chunk with a sampling rate of $F_s$. In the first step, a 1D-convolution layer with stride $L$ and kernel size $3L$ is applied to the input audio chunk $x_k$ to obtain the latent space representation, $y_k \in R^{E \times K}$, where $E$ is the latent space feature dimensions and $K = \frac{S}{L}$ is the feature sequence length in the latent space. {Setting the kernel size to $3L$ and stride to $L$ requires an overlap of $L$ samples with the previous and the future chunk, resulting in a lookahead of $L$ samples. In our experiments, we set $L = 32$ samples at 44.1 kHz. This results in a lookahead of around $0.73$~ms, which is negligible.} Given a one-hot or multi-hot query vector $q \in \{0, 1\}^{N_c}$, where $N_c$ is the total number of classes, streaming target sound extraction is achieved by computing feature masks $m_k \in R^{E \times K}$. With the mask generation network and element-wise multiplication denoted as $\mathcal{M}$ and $\odot$, respectively, the target sound signal, $\hat{x}_k \in R ^ S$, is computed as: \begin{align} y_k =& \text{Conv1d}(x_k), ~~~ m_k = \mathcal{M}(y_k\ |\ y_{k-1}, .., y_2, y_1, q) \notag\\ \hat{x}_k =& \text{ConvTranspose1d}(y_k \odot m_k).\notag \end{align} \subsection{Dilated causal convolution encoder}\label{sec:encoder} Our encoder is a stack of dilated causal convolution (DCC) layers \cite{https://doi.org/10.48550/arxiv.1609.03499}, and the decoder is a transformer network \cite{https://doi.org/10.48550/arxiv.1706.03762}. The motivation for such an architecture is that the encoder computes a \emph{contextful} representation of the input chunk, considering the previous chunks up to a certain receptive field, and the decoder conditions the encoder output with the query vector to estimate the target mask. While recent transformer models for speech separation \cite{subakan2021attention, subakan2022resource} have demonstrated performance gains over convolution based methods \cite{luo2019conv}, the latter have generally been more computationally efficient. We attribute this efficiency gap to the difference in the way existing transformer models process the receptive field compared to convolution based architectures. To achieve a receptive field of length $R$, given the chunk based processing in existing transformer architectures, each chunk individually attends to all previous chunks in the receptive field resulting in $\mathcal{O}(R)$ complexity. In contrast, convolution based models \cite{https://doi.org/10.48550/arxiv.1609.03499, luo2019conv} using a stack of $M$ DCC layers with kernel size $P$ and exponentially scaling dilation factors $\{2^0, 2^1, 2^2, ..., 2^{M-1}\}$ have a receptive field of $(P - 1) \cdot (2^M - 1)$. Its complexity is $\mathcal{O}(PM)$. With a small kernel size $P$, the computational complexity of the stacked DCC layers is $\mathcal{O}(P \cdot \log~(1 + \frac{R}{P - 1}))$ $\sim$ $\mathcal{O} (\log R)$ for it to have a receptive field of length $R$. We use 10 DCC layers with a kernel size of 3 and dilation factors $\{2^0, 2^1, 2^2, ..., 2^{10-1}\}$ in our encoder, resulting in a receptive field of $(3 - 1) \cdot (2 ^ {10} - 1) = 2046 $ samples in the latent space. With the initial input convolution stride $L$ set to 0.73~ms, our encoder's receptive field is $\approx 1.5 s$. Fig.~\ref{fig:main_fig} (b) shows an encoding of an input chunk of length 4. For chunk based streaming inference, the encoder maintains a context buffer for each DCC layer. This context is initially computed from the $1s$ receptive field, and then updated dynamically after encoding each subsequent chunk. For encoding a chunk, each DCC layer is fed with the output chunk of the previous layer, left padded with the context of length twice the layer's dilation. After encoding a chunk, the context is updated with the rightmost elements of the padded input for it to be used in encoding the next chunk. For each input chunk $y_k$, the DCC encoder computes an encoded representation, $e_k \in R^{E \times K}$. \begin{table} \label{single_target} \caption{Performance for the single-target extraction task. In our model, $E$ and $D$ correspond to encoder and decoder dimensionalities, respectively. RTF is the real-time factor for a consumer CPU.} \vspace{-.8em} \centering \begin{tabular}{lccc} \toprule Model & Model size & RTF & SI-SNRi \\ \midrule Conv-TasNet & 4.57M & 1.34 & 6.14 \\ ReSepformer & 13.24M & 1.60 & {7.26} \\ Ours ($E = 256;~D = 128$) & 1.10M & 0.66 & 9.02 \\ Ours ($E = 256;~D = 256$) & 1.69M & 0.75 & 9.40 \\ Ours ($E = 512;~D = 128$) & 3.29M & 0.88 & 9.26 \\ Ours ($E = 512;~D = 256$) & 3.88M & 0.94 & \textbf{9.43} \\ \bottomrule \end{tabular} \label{table:single_target} \vskip -0.15in \end{table} \subsection{Query-conditioned transformer decoder} \label{sec:decoder} To get the mask, the encoded representation computed above must be conditioned with the query, $q$. To this end, we first compute an embedding, $l \in R^{E \times 1}$, corresponding to $q$. This is achieved by using an embedding layer comprising three 512-dimensional feed-forward sub-layers with an $N_C$-dimensional input and an $E$-dimensional output. Our transformer decoder conditions the encoded chunk $e_k$ with the query embedding $l$ and derives the mask as follows. Fig.~\ref{fig:main_fig} (c) shows our decoder architecture. First, we perform multiplicative query integration~\cite{2020arXiv200605712O, delcroix2022soundbeam} to compute the conditioned representation: ${e_k}^\prime = e_k \odot l$. Since transformers are more computationally expensive with higher dimensionality, we first project the encoded representations, ${e_k}^\prime$ and $e_k$, to the decoder dimensions $D \le E$ with $1\times1$ convolution. This results in projected encoded representations, ${pe}_k, {{pe}_k}^\prime \in R^{D \times K}$. The decoded representations are then computed by passing ${{pe}_k, {pe}_k}^\prime$ to the transformer decoder layer's self-attention and cross-attention blocks, respectively, to obtain target mask ${pm}_k \in R^{D \times K}$ in the projected decoder space. It is then projected back to the encoder dimensions with another $1\times1$ convolution layer to obtain ${m_k}^\prime \in R^{E \times K}$. Since the bottleneck caused by the projection layers might affect the flow of gradients, as depicted in the diagram, we use a skip connection immediately after the multiplicative query integration to the output of the projection layer to compute the final mask: $m_k = {m_k}^\prime + {e_k}^\prime$. {Within the decoder, we use the chunk-based streaming attention scheme proposed in \cite{9413535}. As shown in Fig.~\ref{fig:main_fig}(c), for decoding the current chunk, $e_k$, the transformer decoder only attends to the samples in the current chunk, $e_k$, and one previous chunk, $e_{k-1}$. This ensures that the input length to the transformer decoder is fixed at $2K$ (current chunk + one previous chunk) and prevents the inference time from growing as the input audio length increases.} \section{Method} For an unnoticeable latency, real-time audio systems require an algorithmic latency less than to 10 ms \cite{7952149}. Targeting such systems, we look at the target sound extraction problem in the context of streaming. While a model's causality is usually defined at sample level, in a streaming application, causality is defined at chunk level: an output chunk can only depend on the current chunk and past chunks. As a result, streaming models have an intrinsic latency equal to the duration of a single chunk. In this work, we propose a deep neural network architecture for streaming target sound extraction that operates on individual audio chunks of arbitrary duration $\tau$, where $\tau$ can be set to the maximum algorithmic latency allowed by the system. \subsection{Streaming target sound extraction} \label{sec:stream} As shown in the Figure \ref{fig:top_level}, we propose a Tasnet \cite{https://doi.org/10.48550/arxiv.1711.00541} style time domain architecture that employs a mask generation network to generate an element-wise multiplicative mask in the latent space. Let $x_k \in R^{S}$ be the input audio chunk, where $S = \tau F_s$ is the chunk length in number of audio samples for a sampling rate $F_s$. In the first step, a 1d-convolution layer with stride $L$ is applied the time domain audio chunk $x_k$ to obtain the latent space representation of the input audio chunk, $y_k \in R^{E \times K}$, where $E$ is the latent space feature dimensions and $K = \frac{S}{L}$ is the chunk length in the latent space. Given a one-hot or multi-hot query vector $q^T \in \{0, 1\}^{N_c}$, where $N_c$ is the total number of classes, the problem of streaming target sound extraction can be formulating as computing the mask $m_k \in R^{E \times K}$. With the mask generation network denoted as $\mathcal{M}$, $\forall k \in \{1, 2, 3, ...\}$ estimated target audio chunk $\hat{x}_k \in R ^ S$ is computed in a streaming fashion as: \begin{gather*} y_k = Conv1d(x_k) \\ m_k = \mathcal{M}(y_k\ |\ y_{k-1}, .., y_2, y_1, q^T) \\ \hat{x}_k = ConvTranspose1d(y_k \odot m_k) \end{gather*} \subsection{Efficient processing of the receptive field using dilated causal convolutions} We employ an encoder-decoder architecture for the mask generation. The encoder is a stack of dilated causal convolution (DCC) layers \cite{https://doi.org/10.48550/arxiv.1609.03499} and the decoder is a transformer decoder \cite{https://doi.org/10.48550/arxiv.1706.03762}. The motivation for such an architecture for this problem is that, the encoder computes a \emph{contextful} representation of the input chunk, considering the previous chunks up to a certain receptive filed, and the decoder conditions the encoder output with the query vector to estimate the target mask. Recent works on transformer based models for speech separation \cite{subakan2021attention, subakan2022resource} have demonstrated the performance gains over convolution based methods \cite{luo2019conv}, but convolution based method have generally been more computationally efficient. We attribute this gap in efficiency to the difference in the way existing transformer based architectures process the receptive filed compared to convolution based architectures. To achieve a receptive field of length $R$, owing to chunk based processing in existing transformer architectures, each chunk individually attends to all previous chunks in the receptive field resulting in $\mathcal{O}(R)$ complexity. Convolution based architectures \cite{https://doi.org/10.48550/arxiv.1609.03499, luo2019conv} implementing a stack of $M$ DCC layers with a kernel size $P$, have a receptive field equal to $(P - 1)\ 2^{M - 1}$, however, the number of operations has the complexity $\mathcal{O}(PM)$. Writing in terms of the receptive field $R$, the computational complexity of the stacked DCC layers to achieve a receptive filed of length $R$ is $\mathcal{O}(1 + \log \frac{R}{P - 1})$ = $\mathcal{O} (\log R)$, as the kernel size $P$ is usually a small constant $\geq 1$. Hence, in our implementation, we use a stack of DCC layers as the encoder of our mask generation network. We use 10 layers and kernel size equal to 3 in our encoder, resulting a receptive field of $(3 - 1)\ .\ 2 ^ {(10 - 1)} = 1024$ samples in the latent space. With the initial input convolution stride $L$ set to be equivalent to 1 ms, receptive field of our encoder is $\approx1s$. In the figure \ref{fig:enc}, we show an example encoding of an input chunk of length 4. For chunk based streaming inference, the encoder maintains a context buffer for each DCC layer. This context is initially computed from the $\approx1s$ receptive field, and then updated dynamically after encoding each subsequent chunk. For encoding a chunk, each DCC layer is fed with the output chunk of the previous layer, left padded with the context of length twice the layer's dilation. After encoding a chunk, the context is updated with the rightmost elements of the padded input for it be used in encoding the next chunk. For each input chunk $y_k$ (output of the initial $Conv1d$ block), DCC encoder computes an encoded representation $e_k \in R^{E \times K}$. \subsection{Query conditioned decoding with linear attention} Encoder output only depends on the current input chunk and the encoder buffer computed from the receptive field. To obtain the final mask, this encoded representation has to be conditioned with the query vector $q^T$. We first compute an embedding, $l \in R^{E \times 1}$, corresponding to $q^T$ using an embedding layer. The embedding layer comprises of three feed-forward sub-layers, with the hidden dimension equal to 512, and the input \& output dimensions equal to $E$. The decoder conditions the encoded chunk $e_k$ with the query embedding $l$. Motivated by the success of the transformer based architectures for speech processing tasks \cite{https://doi.org/10.48550/arxiv.2010.13154, 9413535}, we employ a transformer decoder layer for estimating the target mask, shown in \ref{fig:dec}. Similar to \cite{2020arXiv200605712O, delcroix2022soundbeam}, we perform multiplicative query integration to compute conditioned representation: ${e_k}^\prime = e_k \odot l$. Since transformers are more computationally expensive at higher dimensionality, we first project the encoded representations, ${e_k}^\prime$ and $e_k$ to decoder dimensions $D \le E$, using 1x1 convolution layers. This results in projected encoded representations ${pe}_k, {{pe}_k}^\prime \in R^{D \times K}$. The decoded representations are then computed by passing ${{pe}_k, {pe}_k}^\prime$ to the transformer decoder layer's self-attention and cross-attention blocks, respectively, to obtain the target mask in the projected decoder space, ${pm}_k \in R^{D \times K}$. The target mask is then projected back to encoder dimensions using another 1x1 convolution layer to obtain the ${m_k}^\prime \in R^{E \times K}$. All projections include a ReLU non-linearity. Owing to this non-linearity, we observed that the projection layers affect the gradient flow to the embedding layer. To circumvent this, we add a skip connection immediately after the multiplicative query integration to the output of the projection layer to compute the final mask: $m_k = {m_k}^\prime + {e_k}^\prime$. The mask $m_k$ is then used to compute extract the target sound, as described in \ref{sec:stream}. \section{Related work} {\bf Universal sound separation.} The task here is to decompose a mixture of arbitrary sound types into their component sounds, regardless of the number of sounds in the mixture~\cite{kavalerov2019universal}. This becomes increasingly challenging as the number of possible sound types in the mixture increases. Several networks have been proposed for this task including convolutional long short-term memory networks~\cite{kavalerov2019universal}, time-dilated convolution networks~\cite{kavalerov2019universal} based on Conv-TasNet~\cite{luo2019conv}, and transformer networks~\cite{zadeh2019wildmix}. Prior work also proposed the use of embeddings learnt by a sound classifier trained on a large sound ontology~\cite{gemmeke2017audio} for conditioning a separation network. \vskip 0.05in\noindent {\bf Target sound extraction.} This aims at separating one or a limited number of sounds of interest from a mixture by conditioning a source-separation model with query information or clues about the target sound. This approach can circumvent the challenge of universal sound separation struggling to deal with a mixture of a large number of sounds. The clues may be provided as an embedding of an audio clip~\cite{delcroix2022soundbeam,gfeller2021one}, an image~\cite{gao2019co,xu2019recursive}, natural language text~\cite{kilgour2022text,liu2022separate}, onomatopoeic words~\cite{okamoto2022environmental}, or a one-hot sound label vector~\cite{2020arXiv200605712O}. The prior works~\cite{kong2020source,chen2022zero} have also evaluated the use of a sound event detector to detect the time when the target sound occurs in a mixture. Although, these works are often motivated for practical usage~\cite{liu2022separate,2020arXiv200605712O}, none of them use streaming models. In contrast, we design the first streaming network for target sound extraction using attention. \vskip 0.05in\noindent {\bf Speech-specific networks.} Prior work has also focused on speech enhancement~\cite{wisdom2019differentiable,eskimez2022personalized,chatterjee2022clearbuds}, speech separation~\cite{luo2019conv,luo2020dual,huang2014deep,hershey2016deep,isik2016single,yu2017permutation,wang2018voicefilter} and speech selection using clues provided to the network~\cite{tzinis2022heterogeneous,ephrat2018looking,directional}. For speech enhancement, neural networks have been proposed~\cite{eskimez2022personalized} to realize real-time operation. {Recently, efficient transformer based architectures have been proposed for speech recognition and separation tasks ~\cite{subakan2021attention,luo2022tiny,https://doi.org/10.48550/arxiv.2005.08100}. These method either use standard transformer blocks \cite{subakan2022resource} or convolution augmented transformer blocks \cite{luo2022tiny, https://doi.org/10.48550/arxiv.2005.08100}. In contrast, we use DCC layers to process the receptive field and a transformer decoder for generating the target mask for sound extraction. Recent work on ReSepformer~\cite{subakan2022resource} proposed a causal mode for their transformer method, which we use for our baseline comparisons.} \section{Related work} {\bf Universal sound separation} Universal sound separation~\cite{kavalerov2019universal} is the task of separating a mixture of arbitrary sound types into their component sounds, regardless of the number of sounds in the mixture. This task becomes increasingly challenging as the number of possible sound types in the mixture increases. Prior work in this area have made progress by constructing an open dataset with mixtures of sounds drawn from hundreds of different classes~\cite{wisdom2021s}. Several network architectures have been proposed for the task including convolutional long short-term memory networks~\cite{kavalerov2019universal}, time-dilated convolution networks~\cite{kavalerov2019universal} based on ConvTas-Net~\cite{luo2019conv}, and Transformer networks~\cite{zadeh2019wildmix}. Further work proposed the use of embeddings learned by a sound classifier trained on a large ontology of sounds~\cite{gemmeke2017audio} to condition a separation network with additional information. Unlike these works, our work is focused on the related problem of extracting target sound classes from a mixture. \noindent {\bf Target sound extraction.} Target source extraction refers to the task of separating one or more sounds of interest from a mixture by conditioning a source-separation model with query information or clues about the target like an embedding of an audio clip~\cite{delcroix2022soundbeam}, an image~\cite{gao2019co,xu2019recursive}, natural language text~\cite{kilgour2022text,liu2022separate}, onomatopoeic words~\cite{okamoto2022environmental}, or a one-hot sound label vector~\cite{2020arXiv200605712O}. Prior work~\cite{kong2020source,chen2022zero} has also evaluated the use of a sound event detector to detect the time when a target sound occurs in a mixture. Multiple modalities like natural language and sound samples have also been used in combination to exploit common information between them~\cite{kilgour2022text}. Although, these works are often motivated as being useful for real-time applications~\cite{liu2022separate,2020arXiv200605712O}, none of them are causal. In contrast, we design the first causal network for real-time target sound extraction of diverse sound events. \noindent {\bf Speech networks.} Prior work in the domain of speech has focused on speech enhancement~\cite{wisdom2019differentiable,eskimez2022personalized,chatterjee2022clearbuds}, which is the separation of relatively structured sounds of speech from non-speech interference, speech separation~\cite{luo2019conv,luo2020dual,huang2014deep,hershey2016deep,isik2016single,yu2017permutation,wang2018voicefilter} which is the separation of all voices in a mixture, and speech selection which is the extraction of a target speaker from a mixture given clues provided to the network~\cite{tzinis2022heterogeneous,ephrat2018looking}. For the task of speech enhancement, causal networks based on a complex convolution recurrent network and a deep convolution attention U-Net have been proposed~\cite{eskimez2022personalized} to achieve real-time performance. For speech separation, Transformer networks have recently been used to achieve state-of-the-art performance due to its ability to learn short and long-term dependencies~\cite{subakan2021attention,luo2022tiny}. However, the use of Transformer networks for the task of target sound extraction remains under-explored. In this work we present the first causal Transformer model for real-time target sound extraction. \section{Experiments and results} {\bf Dataset.} We use a synthetic sound mixture dataset created from the FSD Kaggle 2018 dataset~\cite{Fonseca2018_DCASE}. FSD Kaggle 2018 is a set of sound event and class label pairs, with 41 different sound classes, which are a subset of the Audioset ontology \cite{gemmeke2017audio}. Our synthetic dataset consists of 50k training samples, 5k validation samples and 10k test samples. Sound mixtures are created using the Scaper toolkit \cite{8170052}, each with 3-5 foreground sounds randomly sampled from the FSD Kaggle 2018 dataset, and a background sound randomly sampled from the TAU Urban Acoustic Scenes 2019 dataset~\cite{Mesaros2018_DCASE}. Foreground sound classes are randomly sampled without replacement so that each sample has 3-5 unique classes. We construct the sound mixtures by sampling 3-5s crops from each foreground sound and then pasting them on a 6s background sound. The SNRs of the foreground sounds are randomly chosen between 15 and 25 dB, relative to the background sound. Our training and validation data are sampled from the development splits of FSD Kaggle 2018 and TAU Urban Acoustic Scenes 2019, while our test samples are from the test splits. {From each mixture sample, 3 foreground sounds are randomly selected to be the target sounds. Only one of those 3 foreground sounds is used for the single-target extraction task, while others are used for the multi-target extraction task. The choices of the target foreground sounds are fixed after generating the validation and test sets, to ensure evaluations are reproducible. During training, however, the choices of the target foreground sounds in the training set are randomized. Since we mainly consider human listening applications for streaming target sound extraction, we run our experiments at a 44.1 kHz sampling rate to cover the full audible range.} \vskip 0.05in\noindent{\bf Evaluation setup.} Prior works \cite{2020arXiv200605712O, delcroix2022soundbeam} show that Conv-TasNet, originally proposed for speech separation, can also be used for target sound extraction. Further, ReSepformer proposes an efficient transformer architecture for speech separation that allows a streaming inference. Here, we compare the performance of our architecture with the causal or streaming implementations of Conv-TasNet and ReSepformer as described in the original papers~\cite{luo2019conv,subakan2022resource} for the target sound extraction task. For all the models, we set the stride of the initial convolution, $L$, to 32, which is about $0.73$ ms at 44.1 kHz. We train multiple configurations of our model with different encoder and decoder dimensions. We fix the number of DCC layers to 10, number of transformer layers to 1 and chunk length, $K$, to 13. This chunk length corresponds to 416 samples in the time domain, or a chunk duration of 9.43 ms. For Conv-TasNet, we follow the configuration used in \cite{2020arXiv200605712O} except for the number of repeats, which we set to 2. This ensures that the runtime of the Conv-TasNet baseline is not too large compared with that of our model's largest configuration. For the ReSepformer baseline, we set the model dimensionality to 512, number of blocks to 2, number of transformer layers to 2 and chunk size to 13 (9.43 ms). We perform label integration after the first transformer block, as we found that to be perform better than integrating it at the beginning. \vskip 0.05in\noindent{\bf Loss function and training hyper-parameters.} We use a linear combination of 90\% signal-to-noise-ratio (SNR) and 10\% scale-invariant-signal-to-noise-ratio (SI-SNR) \cite{https://doi.org/10.48550/arxiv.1811.02508} as the loss function for training. While a scale dependent loss, such as SNR loss, is essential for preserving the original amplitude of the desired sounds in the mixture, we observed that adding a fraction of the SI-SNR helps stabilize the training process especially for causal models. We set the initial learning rate to 5e-4 for our models and Conv-TasNet, and to 1.5e-4 for the ReSepformer. We use the \verb|ReduceLROnPlateau| learning rate scheduler to scale the learning rate by 0.1, if there is no improvement in the validation SI-SNR for more than 5 epochs. We use SI-SNR improvement (SI-SNRi) as the validation and test metric. We train the models for 100 epochs. We pick the weights after the epoch that resulted in the best validation SI-SNRi. \begin{figure}[t!] \centering \begin{subfigure}[l]{\linewidth} \includegraphics[width=\linewidth]{Figs/qual-mixed.png} \caption{Input mixture} \label{fig:qual_mixture} \end{subfigure} \centering \begin{subfigure}[l]{\linewidth} \includegraphics[width=\linewidth]{Figs/qual-gt-Bark.png} \caption{Ground-truth for label `Bark'} \label{fig:qual_gt_bark} \end{subfigure} \centering \begin{subfigure}[l]{\linewidth} \includegraphics[width=\linewidth]{Figs/qual-gt-Cough.png} \caption{Ground-truth for label `Cough'} \label{fig:qual_gt_cough} \end{subfigure} \centering \begin{subfigure}[l]{\linewidth} \includegraphics[width=\linewidth]{Figs/qual-output-single.png} \caption{Output of the single-target extraction of `Bark'} \label{fig:qual_output_single} \end{subfigure} \centering \begin{subfigure}[l]{\linewidth} \includegraphics[width=\linewidth]{Figs/qual-output-multi.png} \caption{Output of the multi-target extraction of `Bark' and `Cough'} \label{fig:qual_output_multi} \end{subfigure} \caption{{Visualization of time-domain waveforms of single-target and multi-target extraction (x-axis represents time). }} \label{fig:qual} \vskip -0.2in \end{figure} \vskip 0.05in\noindent{\bf Results.} We separately train the models for single-target and multi-target extraction tasks, and evaluate them on our testset. For multi-target evaluation, we train our model as well as baselines to make predictions with multi-hot query vectors, as opposed to one-hot queries used in the single-target evaluation. { During the multi-target training, 1-3 foreground sounds are randomly selected as target sounds. This training method using arbitrary number of target sources helps the model learn multi-target embeddings. The same model configurations are used for both the single-target and multi-target experiments. The Conv-TasNet and ReSepfromer baselines are also trained in the same way for the multi-target extraction task. We also evaluate the real-time factors (RTFs) of the models on an Intel Core i5 CPU using a single thread. RTF is computed by measuring the runtime consumed by the models to process a 416 sample audio chunk (9.43 ms at 44.1 kHz), and dividing that with the chunk duration, 9.43 ms. {For the RTF measurement, we include the padding for dilated convolution layers in our model's DCC encoder and Conv-TasNet's Temporal Convolution Network (TCN) blocks, accounting for the entire receptive field. In case of the ReSepformer, using a single chunk for RTF measurement excludes the overhead caused by causal attention masking in its inter-attention blocks. Consequently, the RTF value reported for ReSepformer is a lower bound of what is practically achievable.}} { Table~\ref{table:single_target} compares our models with different configurations with the baselines in terms of both the efficiency and the performance. We show that our approach results in 2.2-3.3 dB SI-SNRi improvement compared with the baselines, while being 1.5-2x more computationally efficient with 1.2-4x fewer parameters. Table~\ref{table:multi_target} compares the performance of our models with the baselines for the multiple target extraction task. It shows that our method outperforms the baselines by 1.2-1.4 dB for the 2-target case, and 1-1.2 dB for the 3-target case. As with prior work~\cite{2020arXiv200605712O}, the SI-SNR improvements are lower in the 3-target selection task since there is greater similarity between the input mixture and the target signal, compared to the single-target case, resulting in a larger input SI-SNR.} { In Fig.~\ref{fig:qual}, we qualitatively show an example single-target extraction and multi-target extraction from a 4-class input mixture, using our multi-target extraction model. Fig.~\ref{fig:qual_mixture} shows the input mixture waveform, and Figs.~\ref{fig:qual_gt_bark} and \ref{fig:qual_gt_cough} show the isolated ground-truth sounds. We provide the input mixture to our multi-target model, with a single target query followed by a two target query. Figs.~\ref{fig:qual_output_single} and \ref{fig:qual_output_multi} are the output waveforms obtained when the single target and the two targets are queried, respectively. The waveforms show that the model successfully recognizes the queried events and extracts the relevant sounds. It can also be observed that our model preserves the original amplitudes of the sounds in the input mixture well. } The use of higher dimensionality such as 512, in DCC layers is shown to perform better \cite{luo2019conv}. Using the 512 dimensions in the transformer decoder, however, can make it computationally intensive. So, it is better to run the DCC encoder at higher number of dimension than that of the transformer decoder. Since we use projection layers to convert between the encoder and decoder dimensions, to mitigate the effects the projection layers have on gradients, we include a skip connection in the decoder. Our ablation study shows that the skip connection improves the SI-SNRi from 9.06 to 9.43. \begin{table}[t!] \caption{SI-SNRi comparison in the multi-target extraction task.} \vspace{-.8em} \centering \begin{tabular}{l c c c c} \toprule & \multicolumn{3}{c}{\# selected classes} \\ Model & 1 & 2 & 3 & Mean \\ \midrule Conv-TasNet & 7.20 & 3.63 & 0.19 & 3.67 \\ ReSepformer & 7.42 & 3.56 & 0.33 & 3.77 \\ Ours ($E = 256;~D = 128$) & 9.06 & 4.78 & 1.51 & 5.11 \\ Ours ($E = 256;~D = 256$) & 9.12 & 4.76 & 1.31 & 5.06 \\ Ours ($E = 512;~D = 128$) & 9.39 & 4.92 & 1.39 & 5.23 \\ Ours ($E = 512;~D = 256$) & 9.29 & 4.92 & 1.35 & 5.19 \\ \bottomrule \end{tabular} \vskip -0.15in \label{table:multi_target} \end{table}

\section{Introduction} Developing valid bootstrap methods for time series has been a particularly challenging problem since the mid-1980s. Soon after Efron's seminal paper (cf. Efron (1979)) on the bootstrap for i.i.d.~observations the first attempts towards an extension to dependent data were made. While for univariate time series a variety of proposals exists that are asymptotically valid under certain assumptions on the dependence structure of the underlying process and for certain types of statistics, very little progress has been made for multivariate time series. This is due to the fact that in the multivariate context the distribution of most relevant statistics depends on the fourth-order moment structure of the underlying process, which many bootstrap methods developed for univariate time series are not able to imitate. Even under the assumption of a linear time series, that is, an $ {\mathbb R}^m $-valued, strictly stationary stochastic process $ ({\bf X}(t))_{t \in {\mathbb Z}} $ which fullfils \be {\bf X}(t) = \sum_{j=-\infty}^{\infty} {\bf B}_j \, {\bf e}(t-j), \ \ t\in {\mathbb Z}, \label{linearproc} \ee for certain $ m \times m $ real matrices $ {\bf B}_j $ and an i.i.d.~white noise process $ ({\bf e}(t))_{t \in {\mathbb Z}} $, the distribution of most statistics of interest depends on the fourth-order moment structure and many bootstrap methods fail. In contrast to this, for univariate linear time series -- that is for processes \eqref{linearproc} with dimension $ m=1 $ -- there are a number of scenarios in which the distribution of some statistic of interest depends only on first and second-order moments and established univariate bootstrap methods are successful. It should be emphasized at this point that whenever we use the term \emph{linear time series (or linear process)} in this work, we refer to a process as given by \eqref{linearproc} \emph{including} the i.i.d.~assumption on the innovations ${\bf e}(t)$, as it is done in many standard references for time series, cf. Brockwell and Davis (1991), among others. A subclass of (\ref{linearproc}) is that of the causal and invertible linear processes. In this case a sequence $ \{{\bf A}_j, j \in {\mathbb N}\}$ of $m\times m$ real matrices exists with $ \sum_{j=1}^\infty \|{\bf A}_j\|_F <\infty$, such that (\ref{linearproc}) also can be expressed as ${\bf X}(t) = \sum_{j=1}^{\infty} {\bf A}_j \, {\bf X}(t-j) + {\bf e}(t)$. $ \{{\bf X}(t),t \in {\mathbb Z}\}$ is then called a linear VAR($\infty$) process. Here $ \|\cdot\|_F$ refers to the Frobenius norm of a matrix. Notice that any process following expression (\ref{linearproc}) but with non-i.i.d.~white noise innovations, i.e.~where the $ {\bf e}_t$ are uncorrelated with zero mean but not independent, is nonlinear. Bootstrap methods for time series are usually formulated either in the time domain or in the frequency domain, with a few hybrid methods that combine both approaches. In the following paragraphs we will give a short overview of existing methods for the univariate and the multivariate setup, and discuss their respective limitations. Prominent examples in the time domain are the block bootstrap and variants thereof, the AR sieve bootstrap and the linear process bootstrap, among others. Block bootstrap methods are rough tools which are valid for a wide class of processes but typically their performance heavily depends on the block size. For univariate time series, the AR sieve bootstrap is known to be valid exclusively in those situations where the distribution of interest depends on the first and second-order characteristics of the process only, cf. Kreiss, Paparoditis and Politis (2011). There are some relevant statistics for which this is the case, like the sample mean for general stationary processes or sample autocorrelations for linear processes while sample autocovariances (even for linear processes) are already outside the range of its validity. The same remarks also can be made for the linear process bootstrap. To overcome some of the aforementioned limitations for univariate time series, Frangeskou and Paparoditis (2019) proposed a procedure that involves a wild bootstrap scheme to generate pseudo innovations which asymptotically correctly imitate the first, the second and the fourth-order moment structure of the true innovations. This extends the validity of the AR sieve bootstrap beyond the class of linear, causal and invertible processes. However, if one switches to multivariate processes the fourth-order moment structure shows up in the asymptotic distribution of almost any relevant statistic. This makes the bootstrap estimation problem much more involved. In particular, the AR sieve and the linear process bootstrap fail for multivariate noninvertible linear processes (\ref{linearproc}) and for such basic statistics as sample cross-correlations; see Jentsch and Kreiss (2010) and Meyer and Kreiss (2015) for more details. Thus, beyond causal and invertible linear processes, and apart from special cases like Gaussian time series, it seems that validity of the vector AR sieve bootstrap is essentially restricted to very elementary statistics like the sample mean. Even the aforementioned extension of the AR sieve via wild bootstrap-generated pseudo innovations is not available in the multivariate context. Consequently, for a wide class of stationary multivariate processes, including linear processes (\ref{linearproc}), and for many interesting statistics, the only available bootstrap method is essentially the time domain block bootstrap and variants thereof. Concerning frequency domain bootstrap methods for univariate time series, the situation is very similar. Interest is here focused on so-called integrated periodogram statistics. These are statistics which are obtained by integrating over all frequencies the periodogram multiplied with some function of interest. Many time domain statistics like autocorrelations or autocovariances have a frequency domain analogue, that is they can also be expressed as (functions of) integrated periodograms. Hurvich and Zeger (1987), Franke and H\"ardle (1992) and Dahlhaus and Janas (1996) developed for univariate time series a multiplicative bootstrap procedure for the periodogram. By construction, this scheme is capable of imitating the variance of periodogram ordinates at different frequencies but not their dependence structure across frequencies. Since for periodogram-based statistics, the fourth-order structure of the process is transmitted to their distribution through the weak dependence of the periodogram ordinates across different frequencies, the multiplicative periodogram bootstrap suffers from the same limitations as some of the time domain procedures discussed so far. More specifically, it is valid exclusively in those situations where the distribution of interest only depends on the spectral density, that is on the second-order structure of the process. Dahlhaus and Janas (1996) showed that a subclass of standardized integrated periodograms, the so-called \emph{ratio statistics}, falls into this category, but only under the assumption that the underlying time series is linear. Kreiss and Paparoditis (2012) proposed a modification of the multiplicative periodogram bootstrap which extends its validity for integrated periodograms to the entire class of univariate linear time series. Recently, Meyer et al. (2020) proposed a hybrid method which is valid for a very general class of weakly dependent univariate processes and which goes far beyond the linear process class. However, for multivariate time series the related inference problems are much more involved and no valid frequency domain bootstrap method exists so far. In principle, an analogue of the univariate multiplicative approach (Franke and H\"ardle (1992) and Dahlhaus and Janas (1996)), can be formulated in the multivariate setup by using the fact that -- for a wide class of stationary processes -- the periodogram matrices at any fixed set of frequencies are asymptotically independent and have a Wishart distribution; see for instance Brillinger (1981). However, applying such a bootstrap approach alone will fail in imitating the dependence across periodogram ordinates since it only can capture the second-order properties of the underlying multivariate process. Therefore, there seem to be no relevant statistics for multivariate time series for which such a procedure will be valid. This is true even for the linear process class (\ref{linearproc}), for which the distribution of statistics like auto or cross-correlations depends on the fourth-order moment structure of the process. In other words, for multivariate time series, any frequency domain bootstrap scheme that ignores the weak dependence structure of the periodogram matrix across frequencies, would be systematically failing to properly capture the distribution of a large class of relevant statistics. To summarize, a frequency domain bootstrap method that is capable of handling large classes of periodogram based statistics for general multivariate processes is not available so far. It is the purpose of the present paper to close this gap. Towards this end we introduce the concept of the \emph{multivariate frequency domain hybrid bootstrap} (MFHB). This procedure is composed of two ingredients. The first is the multivariate analogue of the previously mentioned multiplicative bootstrap approach proposed for univariate processes. This ingredient is used to imitate features of the distribution of interest that depend on the second-order structure of the process. The second ingredient of our MFHB uses an adaptation of the convolved subsampling idea (cf.~Tewes et al.~(2019)) to the multivariate frequency domain context. It is used to capture those features of the same distribution that the first ingredient is systematically missing out on, and which are crucially determined by the fourth-order structure of the process. Notice that the MFHB procedure we propose is structurally related to the hybrid procedure that was proposed in Meyer et al.~(2020) for univariate time series. However, the MFHB for multivariate time series has to take a number of obstacles into account that are not present in the univariate situation and that have to be solved in a novel way. In particular, the limiting distribution of multivariate integrated periodogram statistics -- as well as of functions thereof -- is a multivariate complex normal distribution where both covariance and relation matrix depend on the fourth-order moment structure of the underlying process. A valid frequency domain bootstrap procedure therefore has to imitate correctly not only the covariance but also the relation matrix of the corresponding distribution. Moreover, additional difficulties arise in the multivariate context from applying a frequency domain bootstrap to the class of smooth functions of integrated periodograms. Such an extension of the MFHB turns out to be important because it allows for applications to interesting classes of statistics, like for instance, to cross-correlations. In order to make such an extension possible, the second and the fourth-order characteristics of the covariance and of the relation matrix of the related smooth functions have to be separated and appropriately imitated by the two ingredients of the MFHB procedure. As we will see, in contrast to integrated periodograms, this separation can not be explicitly calculated and therefore the bootstrap scheme has to be appropriately modified. We show that the proposed MFHB procedure simultaneously establishes correct covariance and relation matrix estimation for both aforementioned classes of statistics. Moreover, the MFHB will be proven to be valid for integrated periodogram statistics as well as for smooth functions thereof and for a wide class of stochastic processes which includes many known linear and nonlinear multivariate time series models. The remainder of this paper is organized as follows. Section 2 states the assumptions we impose on the multivariate process and recalls some definitions and limiting results for multivariate integrated periodogram statistics. Section 3 presents the MFHB procedure for integrated periodograms and for functions thereof, discusses some of its features and establishes its asymptotic validity. Section 4 presents simulations that investigate the finite sample performance of the MFHB and compares it with that of the moving block bootstrap. The proofs of our main results are presented in the Appendix of this paper while the proofs of some technical results are deferred to the Supplementary Material. \section{Preliminaries and basic results} Consider an $ {\mathbb R}^m $-valued, weakly stationary stochastic process $ ({\bf X}(t))_{t \in {\mathbb Z}} $ with mean zero and component processes $ (X_{r}(t))_{t \in {\mathbb Z}} $, $ r=1,\ldots,m $. We denote the autocovariance matrix of the process by $ \boldsymbol\Gamma(h)=E({\bf X}(t+h){\bf X}(t)^\top) \in {\mathbb R}^{m \times m} $, for $ h \in {\mathbb Z} $, with entries $ \gamma_{rs}(h)={\rm Cov}(X_{r}(t+h),X_{s}(t)) $, $ r,s \in \{1,\ldots,m\} $, which fulfill $ \gamma_{rs}(h)= \gamma_{sr}(-h) $. Furthermore, the $k$-th order cumulant of $ ({\bf X}(t))_{t \in {\mathbb Z}} $ is denoted, for any $ r_1, \ldots, r_k \in \{1, \ldots, m\} $ and $ t, h_1, \ldots, h_{k-1} \in {\mathbb Z}$, by \begin{equation} \label{eq.cum} \textrm{cum}(X_{r_1}(t+h_1),X_{r_2}(t+h_2), \ldots, X_{r_{k-1}}(t+h_{k-1}), X_{r_{k}}(t)) \,. \end{equation} The following stationarity and weak dependence assumptions are imposed on the process under consideration. \begin{assumption} \label{assu1} $({\bf X}(t))_{t \in {\mathbb Z}}$ is eighth-order stationary, i.e. for all $ r_j \in \{1,2, \ldots, m\} $ all joint cumulants of the process up to order eight do not depend on the time point $ t $. We therefore write $c_{r_1r_2\ldots r_k}(h_1, h_2, \ldots, h_{k-1})$ for \eqref{eq.cum}. Furthermore, it holds for all $ r_1,\ldots,r_8 \in \{1, \ldots, m\} $ \begin{enumerate} \item[(i)] \ $ \sum_{h\in {\mathbb Z}} (1+|h|) |\gamma_{r_1r_2}(h) | <\infty$,\\ \item[(ii)] \ $ \sum_{h_1,h_2,h_3 \in {\mathbb Z}}(1+|h_1| + |h_2| +|h_3|) |c_{r_1r_2r_3r_4}(h_1,h_2,h_3)| <\infty \,$,\\ \item[(iii)]\ $ \sum_{h_1, \ldots,h_7 \in {\mathbb Z}} |c_{r_1\ldots r_8}(h_1, h_2, \ldots, h_7) | <\infty \,.$ \end{enumerate} \end{assumption} The above assumption is satisfied for a large class of stochastic processes which includes, for instance, the multivariate linear processes class (\ref{linearproc}) under certain moment assumptions on the innovations ${\bf e}_t$ and summability conditions on the matrices $ {\bf B}_j$; see Brillinger (1981). Notice that since $ \sum_{h\in {\mathbb Z}}|\gamma_{r_1 r_2}(h)| < \infty$, for all $ r_1,r_2 \in \{1,2,\ldots, m\}$, the process $({\bf X}(t))_{t \in {\mathbb Z}}$ possesses a spectral density matrix $ {\bf f}: [-\pi,\pi] \rightarrow {\mathbb C}^{m \times m} $ with entries \bee f_{rs}(\lambda) = \frac{1}{2\pi} \sum_{h \in {\mathbb Z}} \gamma_{rs}(h) \, e^{-ih\lambda}\,, \ \ \lambda\in [-\pi,\pi], \eee $r,s \in \{1,2, \ldots, m\}$, which is Hermitian, i.e., $ f_{rs}(\lambda)= f_{sr}(-\lambda)= \overline{f_{sr}(\lambda)} $ and $ \overline{\bf f}(\lambda) = {\bf f }(\lambda)$. Here, and throughout this work, $ \overline{\bf A} $ denotes the conjugate transpose of a complex-valued matrix $ {\bf A} $. Furthermore, ${\bf f}$ is bounded from above and is a continuous function of the frequency $\lambda$. In the following we additionally assume that the eigenvalues of the spectral density matrix ${\bf f}$ are all bounded away from zero over all frequencies. \begin{assumption} \label{assu2} A constant $ \delta >0$ exists such that the eigenvalues of the spectral density matrix, i.e., $ \sigma\big({\bf f}(\lambda)\big)$ satisfy \[ \min \big( \sigma({\bf f}(\lambda)\big) \big)> \delta,\] for all frequencies $ \lambda \in (-\pi, \pi ]$. \end{assumption} Given an $m$-dimensional vector time series $ {\bf X}(1), {\bf X}(2), \ldots, {\bf X}(n)$ stemming from $({\bf X}(t))_{t \in {\mathbb Z}}$, a common moment estimator of the cross covariances $ \gamma_{rs}(h)$, for $ -n< h < n$, is given by \bee \widehat \gamma_{rs}(h)= \begin{cases} \frac{1}{n} \sum_{t=1}^{n-h} X_{r}(t+h) X_{s}(t), & \textrm{ for } 0 \leq h \leq n-1\\ \frac{1}{n} \sum_{t=1}^{n-|h|} X_{r}(t) X_{s}(t+|h|), & \textrm{ for } -(n-1) \leq h \leq -1 \end{cases}\,. \eee In the following, the periodogram matrix $ {\bf I}: [-\pi,\pi] \rightarrow {\mathbb C}^{m \times m} $ with $ {\bf I}(\lambda)={\bf d}(\lambda) \, \overline{\bf d}(\lambda)$ is used as a basic statistic, where \[ {\bf d}(\lambda)=(d_{1}(\lambda),\ldots,d_{m}(\lambda))^\top ,\] and the $m$-dimensional vector of finite Fourier transforms $d_{r}$ is given by \bee d_{r}(\lambda)= \frac{1}{\sqrt{2\pi n}} \sum_{t=1}^{n} X_r(t) \, e^{-it\lambda}\,, \quad r=1,2, \ldots, m. \eee Notice that $ I_{rs}(\lambda)=d_r(\lambda) \, d_s(-\lambda) $ while $ I_{rr}(\lambda)I_{ss}(\lambda) =I_{rs}(\lambda)I_{sr}(\lambda) $. Furthermore, it holds \bee I_{rs}(\lambda) = \frac{1}{2\pi} \sum_{h=-(n-1)}^{n-1} \widehat\gamma_{rs}(h) \, e^{-ih\lambda}\,. \eee Let $ \lambda_{j,n} =2\pi j/n$, $ j \in {\mathcal G}(n)$, be the Fourier frequencies based on a sample size $n$, where \be \mathcal{G}(n) &:=& \{ j \in {\mathbb Z}: \; 1\leq |j| \leq [n/2] \}. \label{defGn} \ee Denote further for $ p,q,r,s \in \{1,2, \ldots, m\}$, by \bee f_{pqrs}(\lambda,\mu,\eta)= \frac{1}{(2\pi)^3} \sum_{h_1,h_2,h_3\in {\mathbb Z}} c_{pqrs}(h_1,h_2,h_3) \, e^{-i(h_1\lambda+ h_2\mu + h_3\eta)} \eee the fourth-order cumulant spectral density of $ ({\bf X}(t))_{t\in{\mathbb Z}}$. The following lemma, which describes the covariance structure of the elements of the periodogram matrix at the Fourier frequencies, is very useful for our subsequent analysis. \begin{lemma} \label{lemmacovper} Let $ ({\bf X}(t))_{t \in {\mathbb Z}} $ have finite fourth moments and be fourth-order stationary satisfying the summability conditions (i) and (ii) of Assumption \ref{assu1}. Then it holds for all Fourier frequencies $ \lambda_{j,n},\lambda_{k,n} \in [0,\pi] $, and for all $ r,s,v,w \in \{1,\ldots,m\} $ \bee {\rm Cov}(I_{rs}(\lambda_{j,n}),I_{vw}(\lambda_{k,n}))=S_1+S_2+S_3\,, \eee where \begin{align*} S_1 &= \frac{2\pi}{n} f_{rsvw}(\lambda_{j,n},-\lambda_{j,n},-\lambda_{k,n}) + \mathcal{O}\left(\frac{1}{n^2}\right)\,,\\ S_2 &= \begin{cases} f_{rv}(\lambda_{j,n}) \cdot f_{sw}(-\lambda_{j,n}) + \mathcal{O}\left(\frac{1}{n}\right)\,, & \lambda_{j,n}=\lambda_{k,n}\\ \mathcal{O}\left(\frac{1}{n^2}\right)\,, & \lambda_{j,n} \neq \lambda_{k,n} \end{cases}\,,\\ S_3 &= \begin{cases} f_{rw}(\lambda_{j,n}) \cdot f_{sv}(\lambda_{j,n}) + \mathcal{O}\left(\frac{1}{n}\right)\,, & \lambda_{j,n}=\lambda_{k,n} \in \{0,\pi\}\\ \mathcal{O}\left(\frac{1}{n^2}\right)\,, & \textrm{else} \end{cases}\,, \end{align*} and where all $ \mathcal{O}(\cdot) $ bounds are uniform over all Fourier frequencies. \end{lemma} As it is seen from the above result, the covariance between the elements of the periodogram matrix at different frequencies vanishes by the order $1/n$ and depends on the fourth-order cumulant spectral density $f_{rsvw}$. We will see that this rate is not fast enough so that the covariance between periodogram ordinates -- and consequently the fourth-order structure of the process -- show up in the limiting distribution of integrated periodogram statistics. This class of statistics, which we will consider in the following, is defined as functions of the periodogram matrix. In particular, we require: \begin{assumption} \label{assu3} For some $ J \in {\mathbb N} $, $ \varphi_j: [-\pi,\pi] \rightarrow {\mathbb C} $, $ j=1,2,\ldots,J $ are square-integrable functions which are bounded in absolute value. \end{assumption} Now, the vector of integrated periodogram statistics we consider is defined as \begin{align} {\bf M}_n= \left(M(\varphi_j,I_{r_js_j})=\int_{-\pi}^{\pi} \varphi_j(\lambda) I_{r_js_j}(\lambda) \, d\lambda\, , \,j =1,2, \ldots, J \right)^\top\,. \label{def_Mn} \end{align} Note that ${\bf M_n} $ is an estimator of the following vector of \emph{spectral means} \begin{align} {\bf M}= \left(M(\varphi_j,f_{r_js_j})=\int_{-\pi}^{\pi} \varphi_j(\lambda) f_{r_js_j}(\lambda) \, d\lambda\, , \,j =1,2, \ldots, J\right)^\top\,. \label{def_M} \end{align} Before proceeding with some limiting results regarding the behaviour of the estimators $ {\bf M}_n$, let us look at some examples. \begin{example} The sample cross-covariance $ \widehat\gamma_{rs}(h) $ at lag $ 0\leq h < n $ is an integrated periodogram statistic. This is due to the fact that choosing $ \varphi(\lambda)=e^{ih\lambda} $ it follows from straightforward calculations that $ \widehat{\gamma}_{rs}(h)=M(\varphi,I_{rs}) $ as well as $ \gamma_{rs}(h)=M(\varphi,f_{rs}) $. Notice that for $ -n < h <0$, $ \widehat\gamma_{rs}(h) = \widehat\gamma_{sr}(-h) $ is an estimator of $ \gamma_{rs}(h)$. \end{example} \begin{example} \label{examplesamplecorr} The sample cross-correlation $ \widehat\rho_{rs}(h)$ at lag $ 0\leq h < n $ is a function of integrated periodograms. To elaborate, let $ \varphi_1(\lambda)=e^{ih\lambda} $, $ \varphi_2(\lambda)=\varphi_3(\lambda)=1$ and consider the corresponding three-dimensional vector of spectral means. Then, $$ \widehat\rho_{rs}(h)=M(\varphi_1,I_{rs})/\sqrt{M(\varphi_2,I_{rr})M(\varphi_3,I_{ss})}$$ is an estimator of $ \rho_{rs}(h)=M(\varphi_1,f_{rs})/\sqrt{M(\varphi_2,f_{rr})M(\varphi_3,f_{ss})}$, the lag $h$ cross-correlation. Notice that $\rho_{rs}(h)$ and $ \widehat{\rho}_{rs}(h)$ are functions of the elements of the vectors $ {\bf M}=(M(\varphi_1,f_{rs}), M(\varphi_2,f_{rr}), M(\varphi_3,f_{ss}))^\top$ and $ {\bf M}_n=(M(\varphi_1,I_{rs}), M(\varphi_2,I_{rr}), M(\varphi_3,I_{ss}))^\top$, respectively. The limiting distribution of $ \widehat\rho_{rs}(h)$ will be derived in Example \ref{examplesamplecorrcontd}. \end{example} For practical calculations the integral in the expression for $ M(\varphi,I_{rs}) $ is commonly replaced by a Riemann sum using the Fourier frequencies. The corresponding approximation of $ M(\varphi,I_{rs}) $ is given by \bee M_{\mathcal{G}(n)}(\varphi,I_{rs})= \frac{2\pi}{n} \sum_{l \in \mathcal{G}(n)} \varphi(\lambda_{l,n}) I_{rs}(\lambda_{l,n}) \,. \eee Before discussing the asymptotic properties of ${\bf M}_n$, we evaluate on some properties of the complex normal distribution which are important for our subsequent discussion. An $ m $-dimensional complex random vector $ {\bf X}=(X_1, $ $ X_2,\ldots,X_m)^\top $ is called complex normal (or complex Gaussian) if and only if the $ 2m $-dimensional vector of real and imaginary parts \bee \big({\rm Re}({\bf X})^\top,{\rm Im}({\bf X})^\top \big)^\top := \big({\rm Re}(X_1),\ldots,{\rm Re}(X_m),{\rm Im}(X_1),\ldots,{\rm Im}(X_m) \big)^\top \eee has a $ 2m $-dimensional (real) normal distribution. The complex normal distribution is determined by three parameters: expectation $ \bm{\mu_X}=E({\bf X}) $, covariance matrix $ {\bf \Sigma_X} $, and relation matrix $ {\bf \Gamma_X} $, where \bee {\bf \Sigma_X}=E\big( [{\bf X}-\bm{\mu_X}] \, \overline{[{\bf X}-\bm{\mu_X}]} \big)\,, \quad {\bf \Gamma_X}=E\big( [{\bf X}-\bm{\mu_X}] \, [{\bf X}-\bm{\mu_X}]^\top \big)\,, \eee recalling that $ \overline {\bf A} $ denotes the conjugate transpose of any matrix or vector $ {\bf A} $. We therefore write \bee {\bf X} \sim \mathcal N_m^c(\bm{\mu_X},{\bf \Sigma_X},{\bf \Gamma_X})\,. \eee We are particularly interested in the case of centered vectors, i.e. $ \bm{\mu_X}={\bf 0} $. It is then easy to see that the covariance matrix of the real normal vector $ ({\rm Re}({\bf X})^\top,{\rm Im}({\bf X})^\top)^\top $ can be directly deduced from the parameters of the complex normal vector $ {\bf X} $ via \be {\bf X} \sim \mathcal N_m^c({\bf 0},{\bf \Sigma_X},{\bf \Gamma_X}) \quad \Leftrightarrow \quad \begin{pmatrix} {\rm Re}({\bf X})\\ {\rm Im}({\bf X}) \end{pmatrix} \sim \mathcal N_{2m}({\bf 0},{\bf G_X})\,, \label{equiv_complexreal} \ee where \begin{align*} {\bf G_X} & = E\left[\begin{pmatrix}{\rm Re}({\bf X})\\ {\rm Im}({\bf X})\end{pmatrix}\begin{pmatrix}{\rm Re}({\bf X})\\ {\rm Im}({\bf X})\end{pmatrix}^\top \right] \\ &= \left(\begin{array}{cc} \frac{1}{2}({\rm Re}({\bf \Sigma_X} )+{\rm Re}({\bf \Gamma_X}))& \frac{1}{2}(-{\rm Im}({\bf \Sigma_X}) +{\rm Im}({\bf \Gamma_X}))\\ & \\ \frac{1}{2}({\rm Im}({\bf \Sigma_X}) +{\rm Im}({\bf \Gamma_X})) & \frac{1}{2}({\rm Re}({\bf \Sigma_X}) -{\rm Re}({\bf \Gamma_X} )) \end{array} \right)\,. \end{align*} The above expression shows that $ {\bf G_X} $ is fully determined by $ {\bf \Sigma_X}$ and $ {\bf \Gamma_X}$ and vice versa. In particular, denote by ${\bf G}_{{\bf X};ij}$, $ i,j\in \{1,2\}$, the four matrices appearing in the $i$-th row and $j$-th column of the above block matrix $ {\bf G_X}$. Then $ {\bf \Sigma_X}= ({\bf G}_{{\bf X};11}+{\bf G}_{{\bf X};22}) + i\, ({\bf G}_{{\bf X};21} -{\bf G}_{{\bf X};12})$ while $ {\bf \Gamma_X} = ({\bf G}_{{\bf X};11}-{\bf G}_{{\bf X};22}) + i\, ({\bf G}_{{\bf X};21} +{\bf G}_{{\bf X};12})$.\\ The equivalence from \eqref{equiv_complexreal} also carries over to weak convergence: For a sequence of complex random vectors $ ({\bf X}_n)_{n \in {\mathbb N}} $ we have $ {\bf X}_n \stackrel{d}{\rightarrow} \mathcal N_m^c({\bf 0},{\bf \Sigma_X},{\bf \Gamma_X}) $ if and only if $ ({\rm Re}({\bf X}_n)^\top,{\rm Im}({\bf X}_n)^\top)^\top \stackrel{d}{\rightarrow} \mathcal N_{2m}({\bf 0},{\bf G_X}) $.\\ For $ \bm{\mu_X}={\bf 0} $ there are two particularly important special cases where the distribution is completely determined by the matrix $ {\bf \Sigma_X} $, the real normal case and the circularly symmetric case. The centered random vector $ {\bf X} $ is almost surely real-valued if and only if $ {\bf \Sigma_X}={\bf \Gamma_X} $, as a simple calculation shows. In this case one may switch to the usual notation for real-valued random vectors: \bee {\bf X} \sim \mathcal N_m^c({\bf 0},{\bf \Sigma_X},{\bf \Sigma_X}) \quad \Leftrightarrow \quad {\bf X} \sim \mathcal N_{m}({\bf 0},{\bf \Sigma_X})\,. \eee Another important special case is the circularly symmetric case. $ {\bf X} $ is called circularly symmetric if for all $ \varphi \in (-\pi,\pi] $ the distribution of $ e^{i\varphi} {\bf X} $ equals the distribution of $ {\bf X} $. This is the case if and only if $ \bm{\mu_X}={\bf 0} $ and $ {\bf \Gamma_X}={\bf 0} $. In this case the joint distribution of real and imaginary parts takes a very specific form, as can be seen from \eqref{equiv_complexreal}: \bee {\bf X} \sim \mathcal N_m^c({\bf 0},{\bf \Sigma_X},{\bf 0}) \Leftrightarrow \begin{pmatrix} {\rm Re}({\bf X})\\ {\rm Im}({\bf X}) \end{pmatrix} \sim \mathcal N_{2m} \left( \begin{pmatrix} {\bf 0}\\ {\bf 0} \end{pmatrix} \,, \frac{1}{2}\begin{pmatrix} {\rm Re}({\bf \Sigma_X}) & -{\rm Im}({\bf \Sigma_X})\\ {\rm Im}({\bf \Sigma_X}) & {\rm Re}({\bf \Sigma_X}) \end{pmatrix} \right)\,. \eee In particular, $ {\rm Re}({\bf X}) $ and $ {\rm Im}({\bf X}) $ then are identically distributed. \vspace*{0.2cm} Under Assumption \ref{assu1} and some additional weak dependence conditions on the process $ ({\bf X}(t))_{t\in{\mathbb Z}}$ it is known that, for any $ r,s \in \{1,\ldots,m\} $, $ M(\varphi,I_{rs}) $ is a consistent estimator for $ M(\varphi,f_{rs}) $ and that the following central limit theorem holds true for $ {\bf V}_n:=\sqrt{n}({\bf M}_n-{\bf M})$: \begin{align} {\bf V}_n=\sqrt{n} \Big( M(\varphi_j,I_{r_j s_j})-M(\varphi_j,f_{r_j s_j})\, , \, j=1,\ldots,J \Big)^\top \stackrel{d}{\longrightarrow} {\bf V}\,, \label{CLT} \end{align} where $ {\bf V} $ is a $ J $-dimensional complex normal random vector with mean zero, covariance matrix $ {\bf \Sigma} $, and relation matrix $ {\bf \Gamma} $. That is, \bee {\bf V}=(V_{r_j s_j}(\varphi_j), \, j=1,\ldots,J)^\top \sim \mathcal N_J^c({\bf 0},{\bf \Sigma},{\bf \Gamma}). \eee $ {\bf V} $ fulfils \be \overline{V_{r_j s_j}(\varphi_j(\cdot))} = V_{r_j s_j}(\overline{\varphi_j(- \,\cdot)})\,, \label{limit_prop} \ee and covariance and relation parameters are given as follows. The covariance matrix decomposes into $ {\bf \Sigma}={\bf \Sigma}_1+{\bf \Sigma}_2 $ where the $(j,k)$-th element of the matrix ${\bf \Sigma}$ is given by \be \Sigma_{j k}=\textrm{Cov}(V_{r_j s_j}(\varphi_j),V_{r_k s_k}(\varphi_k)) = \Sigma_{1;j k} + \Sigma_{2;j k} \label{covrelformula} \ee with \begin{align} \Sigma_{1;j k} &= 2\pi \int_{-\pi}^{\pi} \varphi_j(\lambda) \overline{\varphi_k(\lambda)} f_{r_j r_k}(\lambda) f_{s_j s_k}(-\lambda) \, d\lambda \label{limit_var1} \\ & \quad \quad + 2\pi \int_{-\pi}^{\pi} \varphi_j(\lambda) \overline{\varphi_k(-\lambda)} f_{r_j s_k}(\lambda) f_{s_j r_k}(-\lambda) \, d\lambda\, \,, \nonumber \end{align} and \begin{align} \Sigma_{2; jk} &= 2\pi \, \int_{-\pi}^{\pi} \int_{-\pi}^{\pi} \varphi_j(\lambda_1) \overline{\varphi_k(-\lambda_2)} f_{r_j s_j r_k s_k}(\lambda_1,-\lambda_1,\lambda_2) \, d\lambda_1 \, d\lambda_2 \,. \label{limit_var2} \end{align} The relation matrix decomposes into $ {\bf \Gamma}={\bf \Gamma}_1+{\bf \Gamma}_2 $ where the $(j,k)$-th element is given by \be \Gamma_{j k}=\textrm{Cov}(V_{r_j s_j}(\varphi_j),\overline{V_{r_k s_k}(\varphi_k)}) = \Gamma_{1;j k} + \Gamma_{2;j k} \label{covrelformula_rel} \ee with \begin{align} \Gamma_{1;j k} &= 2\pi \int_{-\pi}^{\pi} \varphi_j(\lambda) \varphi_k(-\lambda) f_{r_j r_k}(\lambda) f_{s_j s_k}(-\lambda) \, d\lambda \label{limit_rel1} \\ & \quad \quad + 2\pi \int_{-\pi}^{\pi} \varphi_j(\lambda) \varphi_k(\lambda) f_{r_j s_k}(\lambda) f_{s_j r_k}(-\lambda) \, d\lambda\, \,, \nonumber \end{align} and \begin{align} \Gamma_{2; jk} &= 2\pi \, \int_{-\pi}^{\pi} \int_{-\pi}^{\pi} \varphi_j(\lambda_1) \varphi_k(\lambda_2) f_{r_j s_j r_k s_k}(\lambda_1,-\lambda_1,\lambda_2) \, d\lambda_1 \, d\lambda_2 \,. \label{limit_rel2} \end{align} Observe that due to \eqref{limit_prop} it holds $ \Gamma_{j k} = \textrm{Cov}(V_{r_j s_j}(\varphi_j(\cdot)),V_{r_k s_k}(\overline{\varphi_k(- \,\cdot)})) $. Hence, $ {\bf \Gamma} $ can be obtained from $ {\bf \Sigma} $ by replacing $ \varphi_k(\cdot) $ with $ \overline{\varphi_k(- \,\cdot)} $, and we stated the explicit form merely for convenience reasons. We refer to Rosenblatt (1963), Brillinger (1981), Dahlhaus (1985) and Taniguchi and Kakizawa (2000). As it is seen from the above expressions the terms ${\bf \Sigma}_2$ and ${\bf \Gamma}_2$ appearing in the covariance and relation matrices of the limiting complex Gaussian distribution depend on the entire fourth-order moment structure of the process $({\bf X}(t))_{t\in{\mathbb Z}}$, and this dependence is due to the covariance of periodogram ordinates across frequencies; see Lemma~\ref{lemmacovper}. As argued in \eqref{equiv_complexreal}, the weak convergence $ {\bf V}_n \stackrel{d}{\rightarrow} {\bf V}$ stated in (\ref{CLT}) is equivalent to the statement that the $2J$-dimensional real random vector $ \big({\rm Re}({\bf V}_n)^\top , ({\rm Im}({\bf V}_n)^\top\big)^\top$ converges weakly to a $ 2J$-dimensional real normal distribution with mean zero and covariance matrix \begin{align} {\bf G} & = E\big[\left(\begin{array}{c}{\rm Re}({\bf V})\\ {\rm Im}({\bf V})\end{array}\right)\big({\rm Re}({\bf V})^\top, {\rm Im}({\bf V})^\top \big)\big] \nonumber\\ &= \left(\begin{array}{cc} \frac{1}{2}({\rm Re}({\bf \Sigma} )+{\rm Re}({\bf \Gamma}))& \frac{1}{2}(-{\rm Im}({\bf \Sigma}) +{\rm Im}({\bf \Gamma}))\\ & \\ \frac{1}{2}({\rm Im}({\bf \Sigma}) +{\rm Im}({\bf \Gamma})) & \frac{1}{2}({\rm Re}({\bf \Sigma}) -{\rm Re}({\bf \Gamma} )) \end{array} \right)\,. \label{defG} \end{align} Notice that the limiting distribution $ {\bf V}$ is not necessarily real-valued even if the functions $ \varphi_j, j=1,2, \ldots, J$ are real-valued. However, if the functions $\varphi_j$ satisfy $ \varphi_j(-\lambda)= \overline{\varphi_j(\lambda)}$ for $ j=1,2, \ldots, J$, then $ {\bf \Sigma}={\bf \Gamma}$ and $ {\bf V}$ is real valued, that is $ {\bf V}\sim {\mathcal N}_J({\bf 0}, {\bf \Sigma})$. \begin{example}[Ex.~\ref{examplesamplecorr} continued] \label{examplesamplecorrcontd} With the central limit theorem given for $ {\bf V}_n $ above, we can now state the limiting distribution of the sample cross-correlation. For the particular vectors $ {\bf M}_n $ and $ {\bf M} $ from Example \ref{examplesamplecorr} it can easily be seen that $ {\bf V}_n $ is real-valued and converges to a real Gaussian random vector $ {\bf V} $. Applying the delta method to this CLT it follows after straightforward but tedious computations that \bee \sqrt{n} \left( \widehat \rho_{rs}(h) - \rho_{rs}(h) \right) \stackrel{d}{\longrightarrow} \mathcal N(0,\tau^2)\,, \eee where \begin{align*} \tau^2 = \sum_{j \in {\mathbb Z}} &\left\{ \rho_{rr}(j)\rho_{ss}(j) + \rho_{rs}(j+h) \rho_{sr}(j-h) + \frac{c_{rsrs}(j,j-h,h)}{\gamma_{rr}(0) \gamma_{ss}(0)} \right. \\ &+ \frac{1}{2}\rho_{rs}(h)^2 \big[ \rho_{rr}(j)^2 + \rho_{ss}(j)^2 + 2\rho_{rs}(j)^2 \big] \\ &- 2\rho_{rs}(h) \big[ \rho_{rr}(j)\rho_{sr}(j-h) + \rho_{rs}(j)\rho_{ss}(j-h) \big] \\ &+ \frac{1}{4} \rho_{rs}(h)^2 \left[ \frac{c_{rrrr}(j,j,0)}{\gamma_{rr}(0)^2} + \frac{c_{ssss}(j,j,0)}{\gamma_{ss}(0)^2} + \frac{2 \, c_{rrss}(j,j,0)}{\gamma_{rr}(0) \gamma_{ss}(0)} \right] \\ &- \left.\frac{\rho_{rs}(h)}{\sqrt{\gamma_{rr}(0) \gamma_{ss}(0)}} \left[ \frac{c_{rsrr}(j,j-h,0)}{\gamma_{rr}(0)} + \frac{c_{rsss}(j,j-h,0)}{\gamma_{ss}(0)} \right] \right\} \,. \end{align*} \end{example} The above dependence of the variance of the limiting Gaussian distribution on the fourth-order structure of the process does not simplify even in the case of linear processes as this is the case for univariate linear processes. The following simple example illustrates this fact. \begin{example}[Example~\ref{examplesamplecorrcontd} continued] \label{examplefourthorder} Consider the following simple bivariate linear process, which is a vector moving average process of order $ 1 $: \be {\bf X}(t)=\begin{pmatrix} X_1(t)\\ X_2(t) \end{pmatrix} = \begin{pmatrix} \varepsilon_1(t)\\ \varepsilon_2(t) \end{pmatrix} + \begin{pmatrix} 1 & 1\\ 1 & -1 \end{pmatrix} \, \begin{pmatrix} \varepsilon_1(t-1)\\ \varepsilon_2(t-1) \end{pmatrix}\,, \label{exampleprocess} \ee where $ (\varepsilon_1(t))_{t \in {\mathbb Z}} $ and $ (\varepsilon_2(t))_{t \in {\mathbb Z}} $ are independent univariate i.i.d.~white noise processes, with $ E(\varepsilon_j(t))=0 $, $ {\rm E}(\varepsilon_j(t)^2)=1 $, and kurtosis ${\rm E}(\varepsilon_j(t)^4)=: \eta_j $, for $ j=1,2 $. We take a look at the cross-correlation at lag $ h=0 $, and derive the limiting variance of $ \sqrt{n} ( \widehat \rho_{12}(0) - \rho_{12}(0) ) $. For the process \eqref{exampleprocess} it holds $ \rho_{12}(0)=0 $, and the expression $ \tau^2 $ from Example \ref{examplesamplecorrcontd} simplifies considerably to \begin{align*} \tau^2 &= \rho_{11}(0)\rho_{22}(0) + 2\rho_{11}(1)\rho_{22}(1) + 2\rho_{12}(1)\rho_{21}(1) + \sum_{j \in {\mathbb Z}} \frac{c_{1212}(j,j,0)}{\gamma_{11}(0) \gamma_{22}(0)} \\ &= 1+ \frac{\eta_1-3}{9} + \frac{\eta_2-3}{9}\,. \end{align*} As it is seen -- and in contrast to what happens for univariate linear processes -- in the multivariate context the limiting variance depends on the fourth-order structure of the underlying white noise which can in general not be expressed in terms of second-order quantities of the process $ ({\bf X}(t)) $. \end{example} The considerations in this section made it clear that for a frequency domain bootstrap to be successful in the multivariate context, it has to appropriately imitate the second and the fourth order structure of the underlying stochastic process. This is what the MFHB procedure achieves. \section{The multivariate frequency domain hybrid bootstrap (MFHB)} We discuss the frequency domain procedure proposed in this paper in two parts. First we motivate and describe the MFHB procedure for integrated periodograms. Then we present a modification of the MFHB so that in can be successfully applied to functions of integrated periodograms. \subsection{The MFHB for integrated periodograms} \label{subsec_procedure} The MFHB procedure for integrated periodograms generates two sets of bootstrap pseudo random variables which will be denoted by the superscripts $ \ast $ and $ + $, respectively. The procedure is divided into three main steps which are denoted by Step I, Step II and Step III. In Step I independent pseudo periodogram matrices are generated which are denoted by $ {\bf I}^\ast$. This is done using the asymptotic complex Gaussian distribution of the vector of finite Fourier transforms and their asymptotic independence across frequencies. In Step II the idea of convolved bootstrap of subsamples, cf.~Tewes et al.~(2019), is adopted to develop an algorithm which generates a second set of pseudo periodogram matrices, denoted by $ {\bf I}^{+}$, which are independent of ${\bf I}^\ast$. The pseudo periodogram matrices $ {\bf I}^+$ correctly imitate the weak dependence of the ordinary periodogram matrices $ {\bf I}$ across frequencies within subsamples. Step III merges the integrated periodogram statistics based on the two bootstrapped periodograms $ {\bf I}^\ast$ and $ {\bf I}^+$ in an appropriate way. The merging ensures that replicates of the integrated periodograms based on $ {\bf I}^\ast$ imitate all features of the corresponding distribution up to those depending on the fourth-order structure of the process. The fourth-order features of this distribution are contributed by the corresponding statistics based on the pseudo periodograms $ {\bf I}^+$. Notice that the MFHB bootstrap approximations are designed in such a way that the covariance {\it and} the relation matrix of the limiting complex Gaussian distribution of the integrated periodograms is consistently estimated, see Remark \ref{rem32}. Furthermore, the bootstrap random matrices $ {\bf I}^\ast $ and $ {\bf I}^+ $ are defined on the same probability space, with probability measure $ P^\ast $, and are independent from each other. Consequently, all bootstrap expectations, variances and covariances are denoted in the following by $ {\rm E}^\ast $, $ {\rm Var}^\ast $, and $ {\rm Cov}^\ast $, respectively. The following algorithm implements the previously discussed ideas. \begin{enumerate} \item[] {\bf Step I.1} \ Let $ \widehat{\bf f}$ be an estimator of the spectral density matrix $ {\bf f}$ and denote by $\widehat{f}_{rs}(\lambda) $ the $ (r,s)$-th element of $ \widehat{\bf f}(\lambda)$. \item[] {\bf Step I.2} \ Generate, independently for $ j=1,2, \ldots,N:=[n/2]$, pseudo Fourier transforms \[ {\bf d}^\ast(\lambda_{j,n})=(d^\ast_{1}(\lambda_{j,n}), d^\ast_{2}(\lambda_{j,n}), \ldots, d_{m}^\ast(\lambda_{j,n}))^\top \sim {\mathcal N}_m^c({\bf 0}, \widehat{\bf f}(\lambda_{j,n}), {\bf 0})\] and calculate the pseudo periodogram matrices \bee {\bf I}^\ast(\lambda_{j,n}) = \begin{cases} {\bf d}^\ast(\lambda_{j,n})\overline{\bf d}^\ast(\lambda_{j,n}) & \mbox{for} \ j=1,2, \ldots, N\,,\\ {\bf I}^\ast(-\lambda_{j,n})^\top, & \mbox{for} \ j=-1, -2, \ldots, -N \end{cases} \eee with entries denoted by $ I^\ast_{rs}(\lambda_{j,n}) $, $ r,s\in\{1,2, \ldots, m\}$. \item[] {\bf Step I.3} \ Let \[ {\bf V}^\ast_n = \sqrt{n} \Big( M_{{\mathcal G}(n)}(\varphi_j,I^\ast_{r_j,s_j} )-M_{{\mathcal G}(n)}(\varphi_j,\widehat{f}_{r_j,s_j}) \, , \, j=1, 2, \ldots, J\Big)^\top,\] be the analogue of $ {\bf V}_n$ based on $ {\bf I}^\ast$. \item[] {\bf Step II.1} \ Select a positive integer $ b < n$ and consider the set of all periodogram matrices based on subsamples of length $b$. Denote by $ {\bf I}_t(\lambda_{j,b})$ the periodogram matrix of the subsample $ {\bf X}(t), {\bf X}(t+1), \ldots, {\bf X}(t+b-1)$ at frequency $\lambda_{j,b} = 2\pi j/b$, $\lambda_{j,b} \in {\mathcal G}(b)$. Let $ \widetilde{\bf f}(\lambda_{j,b})= (n-b+1)^{-1}\sum_{t=1}^{n-b+1} {\bf I}_t(\lambda_{j,b})$ be the average of the periodogram matrices of the subsamples at frequency $ \lambda_{j,b} $. \item[] {\bf Step II.2} \ Define the set of frequency domain residual matrices $$ {\bf U}_{t}(\lambda_{j,b}) = \widetilde{\bf f}^{-1/2}(\lambda_{j,b}){\bf I}_{t}(\lambda_{j,b}) \widetilde{\bf f}^{-1/2}(\lambda_{j,b}),$$ where $j=1,2, \ldots, [b/2]$, $t =1,2, \ldots, n-b+1$ and $ {\bf A}^{-1/2}$ denotes the square root of the inverse matrix ${\bf A}^{-1}$, i.e., ${\bf A}^{-1/2}{\bf A}^{-1/2}={\bf A}^{-1}$. \item[] {\bf Step II.3} \ Let $ k=[n/b]$ and generate i.i.d.~bootstrap random variables $ i_1, i_2, \ldots, i_k$ with a discrete uniform distribution on the set $ \{1,2, \ldots, n-b+1\}$. Let \[ {\bf I}^+(\lambda_{j,b}) = \frac{1}{k}\sum_{\ell=1}^k {\bf I}^+_{i_\ell}(\lambda_{j,b}),\] where $ {\bf I}^+_{i_\ell}(\lambda_{j,b}) = \widehat{\bf f}^{1/2}(\lambda_{j,b}){\bf U}_{i_\ell}(\lambda_{j,b}) \widehat{\bf f}^{1/2}(\lambda_{j,b}) $. We write $ {\bf I}^+_{\ell}(\lambda_{j,b}) $ for $ {\bf I}^+_{i_\ell}(\lambda_{j,b})$. The entries of the pseudo periodogram matrices $ {\bf I}^+(\lambda_{j,b}) $ are denoted $ I^+_{rs}(\lambda_{j,b}) $. \item[] {\bf Step II.4} \ Let \[ {\bf V}^+_n = \sqrt{kb} \Big(M_{{\mathcal G}(b)}(\varphi_j,I^+_{r_j,s_j} )-M_{{\mathcal G}(b)}(\varphi_j,\widehat{f}_{r_j,s_j}) \, , \, j=1, 2, \ldots, J\Big)^\top\] be the analogue of $ {\bf V}_n$ based on $ {\bf I}^+$. \item[] {\bf Step III.1} \ Let \[ {\bf G}_n^\ast = E^\ast\big[\left(\begin{array}{c}{\rm Re}({\bf V}^*_n)\\ {\rm Im}({\bf V}^*_n)\end{array}\right)\big({\rm Re}({\bf V}^*_n)^\top, {\rm Im}({\bf V}^*_n)^\top \big)\big], \] \[ {\bf G}_n^+= E^\ast\big[\left(\begin{array}{c}{\rm Re}({\bf V}^+_n)\\ {\rm Im}({\bf V}^+_n)\end{array}\right)\big({\rm Re}({\bf V}^+_n)^\top, {\rm Im}({\bf V}^+_n)^\top \big)\big].\] Furthermore, denote by $ {\bf \Sigma}_{1,n}^+ $ the $ J\times J$ matrix with $(j,k)$-th element given by \begin{align*} \sigma^+_{jk} & = \frac{4\pi^2}{b} \sum_{l=-[b/2]}^{[b/2]}\varphi_j(\lambda_{l,b}) \overline{\varphi_k(\lambda_{l,b})} S_{r_j s_js_kr_k}(\lambda_{l,b})\\ & \ \ \ \ + \frac{4\pi^2}{b} \sum_{l=-[b/2]}^{[b/2]}\varphi_j(\lambda_{l,b}) \overline{\varphi_k(-\lambda_{l,b})} S_{r_js_jr_ks_k}(\lambda_{l,b}). \end{align*} and $ {\bf \Gamma}_{1,n}^+ $ the $ J\times J$ matrix with $(j,k)$-th element given by \begin{align*} c^+_{jk} & = \frac{4\pi^2}{b} \sum_{l=-[b/2]}^{[b/2]}\varphi_j(\lambda_{l,b}) \varphi_k(-\lambda_{l,b}) S_{r_js_j s_k r_k}(\lambda_{l,b})\\ & \ \ \ \ + \frac{4\pi^2}{b} \sum_{l=-[b/2]}^{[b/2]}\varphi_j(\lambda_{l,b}) \varphi_k(\lambda_{l,b}) S_{r_js_jr_ks_k}(\lambda_{l,b}), \end{align*} where \[ S_{rsuw}(\lambda) = \frac{1}{n-b+1}\sum_{t=1}^{n-b+1}\big( \widetilde{I}_{t,rs}(\lambda)-\widehat{f}_{rs}(\lambda)\big) \big( \widetilde{I}_{t,uw}(\lambda) - \widehat{f}_{uw}(\lambda) \big),\] and $ \widetilde{I}_{t,rs}(\lambda)$ is the $(r,s)$-th element of the matrix $$ \widetilde{\bf I}_{t}(\lambda) = \widehat{\bf f}^{1/2}(\lambda){\bf U}_{t}(\lambda) \widehat{\bf f}^{1/2}(\lambda). $$ Finally, define the matrix \[ {\bf C}^+_n = \left(\begin{array}{cc} \frac{1}{2}({\rm Re}({\bf \Sigma}^+_{1,n} )+{\rm Re}({\bf \Gamma}^+_{1,n})) & \frac{1}{2}(-{\rm Im}({\bf \Sigma}^+_{1,n}) +{\rm Im}({\bf \Gamma}^+_{1,n}))\\ & \\ \frac{1}{2}({\rm Im}({\bf \Sigma}^+_{1,n}) +{\rm Im}({\bf \Gamma}^+_{1,n}))& \frac{1}{2}({\rm Re}({\bf \Sigma}^+_{1,n}) -{\rm Re}({\bf \Gamma}^+_{1,n})) \end{array} \right).\] \item[] {\bf Step III.2} Calculate \begin{equation*} {\bf G}_n^\circ = {\bf G}_n^\ast + \big({\bf G}_n^+ - {\bf C}_n^+\big) \end{equation*} and \begin{equation} \left(\begin{array}{c}{\rm Re}({\bf V}^\circ_n)\\ {\rm Im}({\bf V}^\circ_n)\end{array}\right) = \big({\bf G}_{n}^\circ \big)^{1/2}\big({\bf G}^\ast_n \big)^{-1/2} \left(\begin{array}{c}{\rm Re}({\bf V}^*_n)\\ {\rm Im}({\bf V}^*_n)\end{array}\right)\,. \label{eq.Co} \end{equation} \item[] {\bf Step III.3} Approximate the distribution of $ {\bf V}_n$ by that of \[ {\bf V}_n^\circ = {\rm Re}({\bf V}^\circ_n)+ i \,{\rm Im}({\bf V}^\circ_n)\,.\] \end{enumerate} \vspace*{0.3cm} The following series of remarks clarifies several aspects of the above bootstrap procedure. \begin{remark} \label{rem_gen_cn} In Step I.2, the problem of generating a complex normal random vector $ {\bf d}^\ast(\lambda_{j,n}) $ can in practice be reduced to that of generating a real normal random vector (which is usually a pre-implemented routine). Since $ {\mathcal N}_m^c({\bf 0}, \widehat{\bf f}(\lambda_{j,n}), {\bf 0}) $ is a circularly symmetric complex normal distribution, one may generate \bee \begin{pmatrix} \textrm{Re}({\bf d}^\ast(\lambda_{j,n}))\\ \textrm{Im}({\bf d}^\ast(\lambda_{j,n})) \end{pmatrix} \sim {\mathcal N}_{2m} \left( \begin{pmatrix} {\bf 0}\\ {\bf 0} \end{pmatrix} \,, \frac{1}{2}\begin{pmatrix} \textrm{Re}(\widehat{\bf f}(\lambda_{j,n})) & -\textrm{Im}(\widehat{\bf f}(\lambda_{j,n}))\\ \textrm{Im}(\widehat{\bf f}(\lambda_{j,n})) & \textrm{Re}(\widehat{\bf f}(\lambda_{j,n})) \end{pmatrix} \right)\,, \eee and then set $ {\bf d}^\ast(\lambda_{j,n})=\textrm{Re}({\bf d}^\ast(\lambda_{j,n})) + i \cdot \textrm{Im}({\bf d}^\ast(\lambda_{j,n})) $. \end{remark} \begin{remark} \label{rem32} {~} \begin{enumerate} \item[(i)] \ As already mentioned, the pseudo periodogram matrices ${\bf I}^\ast(\lambda_{j,n})$ in Step I.2 are generated using the fact that the $m$-dimensional vector of finite Fourier transforms converges towards a circular symmetric complex normal distribution. Notice that the $ {\bf I}^\ast(\lambda_{j,n})$ generated in this step are independent across frequencies. Therefore, the integrated periodogram statistic $ {\bf V}_n^\ast$ obtained in Step I.3 and based on these pseudo periodogram matrices, is only able to imitate the parts $ {\bf \Sigma_1}$ and $ {\bf \Gamma}_1$ of the limiting variance and relation matrices $ {\bf \Sigma}$ and $ {\bf \Gamma}$, respectively. This will be proven in Lemma \ref{le.BootLem2} (i). Recall that ${\bf \Sigma}_1$ and $ {\bf \Gamma}_1$ only depend on the spectral density matrix $ {\bf f}$ of the underlying process. \item[(ii)] \ In Step II.1 the periodogram matrices ${\bf I}_t(\lambda_{j,b})$ of the subsamples of length $b$ are used. In Step II.2 frequency domain residual matrices $ {\bf U}_{t}(\lambda_{j,b}) $ are defined which are i.i.d.~resampled in Step II.3 to obtain the pseudo periodogram matrix $ {\bf I}^+(\lambda_{j,b})$. Notice that the latter matrix is an average of $k$ independent matrices ${\bf I}_{\ell}^+(\lambda_{j,b}) $ calculated using $k$ randomly selected residual matrices $ {\bf U}_{i_\ell}(\lambda_{j,b})$ and after pre- and post-multiplying them with the square root of the estimated spectral density matrix $ \widehat{\bf f}(\lambda_{j,b})$. Observe that all quantities in Step II are calculated at the Fourier frequencies $ \lambda_{j,b}=2\pi j/b$, $ j\in {\mathcal G}(b)$ corresponding to the length $b$ of the subsamples. \item[(iii)] \ The integrated periodogram statistic ${\bf V}_n^+$ based on $ {\bf I}^+(\lambda_{j,b})$ is used to imitate the missing terms $ {\bf \Sigma}_2$ and $ {\bf \Gamma}_2$ which depend on the fourth order moment structure of the process. To do this appropriately, notice first that the pseudo statistic $ {\bf V}_n^+$ imitates asymptotically correct the entire covariance and relation matrices $ {\bf \Sigma}$ and $ {\bf \Gamma}$ as will be proven in Lemma \ref{le.BootLem2} (ii). Since we use $ {\bf V}_n^\ast$ generated in Step I to imitate the distribution of $ {\bf V}_n$ and the matrices $ {\bf \Sigma}_1$ and $ {\bf \Gamma}_1$, we have to subtract from the covariance and relation matrix of $ {\bf V}_n^+$ the corresponding parts $ {\bf \Sigma}_{1,n}^+$ and $ {\bf \Gamma}_{1,n}^+$ so that only the desired estimators of the components ${\bf \Sigma}_2 $ and ${\bf \Gamma}_2 $ are left. This is done in Step III.1 and Step III.2. In particular, in Step III.1 the elements $ \sigma_{jk}^+$ and $ c_{jk}^+$ of the matrices $ {\bf \Sigma}_{1,n}^+$ and $ {\bf \Gamma}_{1,n}^+$ are explicitly calculated and the corresponding matrix $ {\bf C}_n^+$ is obtained. The latter matrix is then subtracted from the matrix $ {\bf G}_n^+$, so that the obtained matrix $ {\bf G}_n^+-{\bf C}_n^+$ contains the elements of the covariance and relation matrix of $ {\bf V}_n^+$ which only depend on the fourth-order structure of the process, that is the parts $ {\bf \Sigma}_{2,n}^+ $ and $ {\bf \Gamma}_{2,n}^+$. These are used to estimate ${\bf \Sigma}_2 $ and $ {\bf \Gamma}_2$. Now, adding to this difference the matrix ${\bf G}_n^\ast$, leads to a new matrix $ {\bf G}_n^\circ$ which correctly imitates both parts of ${\bf \Sigma}$ and $ {\bf \Gamma}$, compare also Lemma \ref{le.BootLem2} (ii). We stress here the fact that in $ {\bf G}_n^\circ$, the parts corresponding to ${\bf \Sigma}_1 $ and ${\bf \Gamma}_1 $ are contributed by the bootstrap procedure based on the asymptotic Gaussianity of the finite Fourier transforms, that is by ${\bf \Sigma}_{1,n}^\ast $ and ${\bf \Gamma}_{1,n}^\ast $, while the parts $ {\bf \Sigma}_2$ and ${\bf \Gamma}_2$ by the convolved bootstrap procedure, that is by $ {\bf \Sigma}_{1,n}^+$ and ${\bf \Gamma}_{1,n}^+ $. \item[(iv)] \ In Step III.2 the matrix $ {\bf G}_n^\circ$ is used in equation (\ref{eq.Co}) to appropriately rescale the bootstrap vector $ {\bf V}_n^\ast$. The resulting bootstrap complex vector $ {\bf V}_n^\circ$ is finally used in Step III.3 to approximate the distribution of the complex vector ${\bf V}_n$. \end{enumerate} \end{remark} \begin{remark} In the definition of the frequency domain residual matrices ${\bf U}_{t}(\lambda_{j,b}) $ in Step II.2, the pre- and post-multiplication with the matrix $ \widetilde{\bf f}^{-1/2}(\lambda_{j,b})$ ensures that the resampled residual matrices $ {\bf U}_{i_\ell}(\lambda_{j,b})$ satisfy $$ {\rm E}^*({\bf U}_{i_\ell}(\lambda_{j,b})) = \widetilde{\bf f}^{-1/2}(\lambda_{j,b}) \frac{1}{n-b+1}\sum_{t=1}^{n-b+1}{\bf I}_{t}(\lambda_{j,b}) \widetilde{\bf f}^{-1/2}(\lambda_{j,b}) ={\bf Id}_m,$$ where $ {\bf Id}_m$ denotes the $m\times m $ unit matrix. Therefore, \[ {\rm E}^\ast( {\bf I}^+(\lambda_{j,b}) ) = {\rm E}^\ast( {\bf I}^+_{i_\ell}(\lambda_{j,b})) = \widehat{\bf f}^{1/2} (\lambda_{j,b})\cdot {\bf Id}_m\cdot\widehat{\bf f}^{1/2}(\lambda_{j,b}) = \widehat{\bf f}(\lambda_{j,b}) .\] Furthermore, by the consistency of $ \widetilde{\bf f}(\lambda)$ as an estimator of $ {\bf f}(\lambda)$, see the proof of Lemma 3.9, it holds true that for any fixed frequency $\lambda\in (0,\pi)$, $ {\bf U}_{t}(\lambda)=\widetilde{\bf f}^{-1/2}(\lambda) {\bf I}_t(\lambda) \widetilde{\bf f}^{-1/2}(\lambda) \stackrel{D}{\rightarrow} {\bf U}(\lambda)$, as $ b \rightarrow \infty$, where $ {\bf U}(\lambda)$ has the complex Wishart distribution of dimension $m$ and one degree of freedom, i.e., $ {\bf U}(\lambda)\sim W_m^C(1, {\bf Id}_m)$; see Brillinger (1981), Section 4.2. Recall that if $ {\bf U} (\lambda) \sim W_m^C(1,{\bf Id}_m)$ then $ {\rm E}({\bf U}(\lambda))={\bf Id}_m$ and ${\rm Cov}(U_{jk}(\lambda), U_{lm}(\lambda)) = \delta_{j,l}\delta_{k,m}$ for any two elements $ U_{jk}(\lambda)$ and $U_{lm} (\lambda)$ of the matrix ${\bf U}(\lambda)$. \end{remark} \begin{remark} The terms $ S_{rsuw}(\lambda)$ which appear in the expressions of $ \sigma_{jk}^+$ and $c_{jk}^+$ in Step III.1 are obtained by evaluating the covariance expressions $ {\rm Cov}^\ast(I_{i_1,r_js_j}(\lambda_{\ell,b}),I_{i_1,r_ks_k}(\lambda_{\ell,b}) )$ where $ I_{i_1,rs}(\lambda_{\ell,b}) $ denotes the $ (r,s) $-th element of $ {\bf I}_{i_1}(\lambda_{\ell,b}) $. To elaborate on the parameters they are estimating, consider the simplified form, \[ \widetilde{S}_{rsuw}(\lambda) = \frac{1}{n-b+1}\sum_{t=1}^{n-b+1}\big(I_{t,rs}(\lambda)- f_{rs}(\lambda)\big) \big( I_{t,uw}(\lambda) - f_{uw}(\lambda) \big),\] obtained after replacing estimated by true quantities and where $ I_{t,rs}(\lambda)$ denotes the $(r,s)$-th of the matrix $ {\bf I}_t(\lambda) $ defined in Step II.1. See also the proof of Lemma 3.9 (ii). It can then be shown that \begin{equation} \label{eq.S} \widetilde{S}_{rsuw}(\lambda_{j,b}) = f_{rw}(\lambda_{j,b})f_{su}(-\lambda_{j,b}) + o_P(1) = f_{rw}(\lambda_{j,b})f_{us}(\lambda_{j,b}) + o_P(1). \end{equation} To see this observe that \begin{align*} {\rm E}(\widetilde{S}_{rsuw}(\lambda_{j,b})) & = \frac{1}{n-b+1}\sum_{t=1}^{n-b+1}{\rm E}\big((I_{t,rs}(\lambda_{j,b})- f_{rs}(\lambda_{j,b}))( \overline{I_{t,wu}(\lambda_{j,b}) - f_{wu}(\lambda_{j,b})}) \big)\\ & = \frac{1}{n-b+1}\sum_{t=1}^{n-b+1}{\rm Cov}\big(I_{t,rs}(\lambda_{j,b}), I_{t,wu}(\lambda_{j,b}) \big) + \mathcal O(b^{-1}) \\ & = f_{rw}(\lambda_{j,b})f_{su}(-\lambda_{j,b}) +o(1), \end{align*} where the second equality follows because $ {\rm E}(I_{t,rs}(\lambda_{j,b} )) = f_{rs}(\lambda_{j,b}) + \mathcal O(b^{-1})$ with the $ \mathcal O(b^{-1})$ term independent of $t$, and the last equality follows using Lemma~\ref{lemmacovper}. By Lemma~\ref{le.BootLem1} below, it can further be shown that $ {\rm Var}(\widetilde{S}_{rsuw}(\lambda_{j,b}))$ $ \rightarrow 0$ (see also the proof of Lemma~\ref{le.BootLem2}), which justifies expression (\ref{eq.S}). \end{remark} \begin{remark} Notice that if the limiting distribution of $ {\bf V}_n $ is real-valued, i.e.~if $ {\bf V} \sim \mathcal{N}(0,{\bf \Sigma},{\bf \Sigma})$, then Step III.1 and Step III.2 simplify. In particular, in this case one can set the matrices $ {\bf G}_n^\ast$ and $ {\bf G}_n^+$ as real-valued and $ J \times J$ dimensional according to $ {\bf G}_n^\ast = {\rm E}^\ast({\bf V}_n^\ast ({\bf V}_n^\ast)^\top)$ and $ {\bf G}_n^+= {\rm E}^\ast({\bf V}_n^+ ({\bf V}_n^+)^\top)$. Consequently, the matrix $ {\bf C}_n^+ $ can then also be set as real-valued and $ J \times J$ dimensional according to $ {\bf C}^+_n={\bf \Sigma}^+_{1,n}$. \end{remark} \subsection{Smooth functions of integrated periodograms} \label{subsec_procedure_ratio} The MFHB bootstrap procedure proposed -- appropriately modified -- also can be applied to estimate the distribution of statistics which are functions of integrated periodograms such as, for instance, sample cross-correlations. To elaborate, suppose that \begin{equation} \label{eq.gfunction} g=(g_1, g_2, \ldots, g_L): {\mathbb C}^J \rightarrow {\mathbb C}^L \end{equation} is some (smooth) function and that the statistic of interest is given by $ \sqrt{n}({\bf R}_n - {\bf R}) $ where $ {\bf R}_n = g({\bf M}_n) $ and $ {\bf R} = g({\bf M}) $ for the vector of integrated periodograms $ {\bf M}_n $ and spectral means $ {\bf M} $ given in \eqref{def_Mn} and \eqref{def_M}. Sample cross-correlations can be expressed in this way as can be seen from Example~\ref{examplesamplecorr}.\\ Our bootstrap procedure from the previous section can be adapted to approximate the distribution of $ \sqrt{n}({\bf R}_n - {\bf R}) $. We first impose some smoothness assumption on the function $ g $. To do so, we can of course interpret $ g $ as a function defined on $ {\mathbb R}^{2J} $ via the identification \bee g({\bf z})=g\begin{pmatrix}{\rm Re}({\bf z})\\ {\rm Im}({\bf z}) \end{pmatrix} \quad \quad \forall \, {\bf z} \in {\mathbb C}^J\,. \eee Splitting up real and imaginary parts also for the values of $ g $ leads to the accompanying function $ \tilde g: {\mathbb R}^{2J} \rightarrow {\mathbb R}^{2L} $ given by \bee \tilde g({\bf x}):= \begin{pmatrix}{\rm Re}(g({\bf x}))\\ {\rm Im}(g({\bf x})) \end{pmatrix} \quad \quad \forall \, {\bf x} \in {\mathbb R}^{2J}\,. \eee \begin{assumption} \label{assu4} The function $ \tilde g: {\mathbb R}^{2J} \rightarrow {\mathbb R}^{2L} $ is continuously differentiable (in the real sense) in some neighbourhood around $ ( {\rm Re}({\bf M})^\top,{\rm Im}({\bf M})^\top)^\top $ with Jacobi matrix $ {\bf J}_{\tilde g}(( {\rm Re}({\bf M})^\top,{\rm Im}({\bf M})^\top)^\top) $. \end{assumption} Note that -- although considering complex-valued random variables and complex functions $ g $ -- we require only real differentiability of the accompanying function $ \tilde g $. This is a much less restrictive condition than assuming complex differentiability of $ g $. We define the random vectors \bee \widetilde {\bf R}_n:= \begin{pmatrix}{\rm Re}({\bf R}_n)\\ {\rm Im}({\bf R}_n) \end{pmatrix}\,, \quad \quad \widetilde {\bf R}:= \begin{pmatrix}{\rm Re}({\bf R})\\ {\rm Im}({\bf R}) \end{pmatrix} \,. \eee Now, applying the delta method to \eqref{CLT}, respectively \eqref{defG}, we get the limiting result \begin{align} \sqrt{n}(\widetilde {\bf R}_n - \widetilde {\bf R}) \stackrel{d}{\longrightarrow} {\bf J}_{\tilde g}(( {\rm Re}({\bf M})^\top,{\rm Im}({\bf M})^\top)^\top) \, \mathcal{N}_{2J}({\bf 0},{\bf G})\,. \label{CLT_smoothfun} \end{align} From the continuous mapping theorem, applied for $ h(x_1,x_2):=x_1+i\cdot x_2 $, $ x_1,x_2 \in {\mathbb R}^{L} $, it follows \begin{align} \sqrt{n}({\bf R}_n - {\bf R}) \stackrel{d}{\longrightarrow} {\bf W} \sim \mathcal{N}^c_{L}({\bf 0},{\bf \Sigma_R},{\bf \Gamma_R})\,, \label{CLT_smoothfun_g} \end{align} for suitable matrices $ {\bf \Sigma_R} $ and $ {\bf \Gamma_R} $ which can be obtained from \eqref{CLT_smoothfun} (the precise form of $ {\bf \Sigma_R} $ and $ {\bf \Gamma_R} $ is not needed for the ensuing bootstrap algorithm and its validity results).\\ We propose the following MFHB approximation of the distribution of $ \sqrt{n}({\bf R}_n-{\bf R})$. \ \begin{enumerate} \item[] {\bf Step} $\widetilde{\textbf{I}}$ \ Apply Steps I, II, III.1, and III.2 of the algorithm from Section \ref{subsec_procedure} to obtain the matrix $ {\bf G}_n^{\circ} $ and the pseudo periodogram matrices $ {\bf I}^*(\lambda_{j,n}) $, $ j \in \mathcal{G}(n) $. \item[] {\bf Step} $\widetilde{\textbf{II}}$ \ Let \begin{align*} {\bf M}^*_n &:= \Big( M_{{\mathcal G}(n)}(\varphi_j,I^\ast_{r_j,s_j} ), \, j=1, 2, \ldots, J\Big)^\top\,,\\ \widehat{\bf M}_n &:= \Big( M_{{\mathcal G}(n)}(\varphi_j,\widehat f_{r_j,s_j} ), \, j=1, 2, \ldots, J\Big)^\top\,, \end{align*} and \begin{align*} \begin{pmatrix}{\rm Re}({\bf W}^*_n)\\ {\rm Im}({\bf W}^*_n)\end{pmatrix} &:= \sqrt{n} \left[ \tilde g\begin{pmatrix}{\rm Re}({\bf M}^*_n)\\ {\rm Im}({\bf M}^*_n) \end{pmatrix} - \tilde g\begin{pmatrix}{\rm Re}(\widehat{\bf M}_n)\\ {\rm Im}(\widehat{\bf M}_n) \end{pmatrix} \right] \,. \end{align*} \item[] {\bf Step} $\widetilde{\textbf{III}}$ \ Let \begin{align*} \widetilde{\bf G}^*_n &:= E^*\left[\begin{pmatrix}{\rm Re}({\bf W}^*_n)\\ {\rm Im}({\bf W}^*_n)\end{pmatrix} \begin{pmatrix}{\rm Re}({\bf W}^*_n)\\ {\rm Im}({\bf W}^*_n)\end{pmatrix}^\top\right] - E^*\begin{pmatrix}{\rm Re}({\bf W}^*_n)\\ {\rm Im}({\bf W}^*_n)\end{pmatrix} \cdot E^*\begin{pmatrix}{\rm Re}({\bf W}^*_n)\\ {\rm Im}({\bf W}^*_n)\end{pmatrix}^\top \end{align*} and \begin{align*} \widetilde{\bf G}^\circ_n &:= {\bf J}_{\tilde g}\begin{pmatrix}{\rm Re}(\widehat{\bf M}_n)\\ {\rm Im}(\widehat{\bf M}_n)\end{pmatrix} \times {\bf G}^\circ_n \times {\bf J}_{\tilde g}\begin{pmatrix}{\rm Re}(\widehat{\bf M}_n)\\ {\rm Im}(\widehat{\bf M}_n)\end{pmatrix}^\top\,. \end{align*} \item[] {\bf Step} $\widetilde{\textbf{IV}}$ \ Calculate \bee \begin{pmatrix}{\rm Re}({\bf W}^\circ_n)\\ {\rm Im}({\bf W}^\circ_n)\end{pmatrix} := \big(\widetilde{\bf G}_{n}^\circ \big)^{1/2}\big(\widetilde{\bf G}^*_n \big)^{-1/2} \begin{pmatrix}{\rm Re}({\bf W}^*_n)\\ {\rm Im}({\bf W}^*_n)\end{pmatrix}\,. \eee \item[] {\bf Step} $\widetilde{\textbf{V}}$ \ Approximate the distribution of $\sqrt{n}({\bf R}_n-{\bf R})$ by that of \begin{align*} {\bf W}_n^\circ & := {\rm Re}({\bf W}_n^\circ)+ i \,{\rm Im}({\bf W}_n^\circ)\,. \end{align*} \end{enumerate} \vspace*{0.2cm} \begin{remark} {~} \begin{enumerate} \item[(i)] \ The algorithm above works on real and imaginary parts separately. This has the advantage that we do not have to impose complex differentiability assumptions on the function $ g $; it suffices to assume real differentiability of $ \tilde g $. \item[(ii)] \ In Step $ \widetilde{\textrm{II}} $ the bootstrap vector $ ( {\rm Re}({\bf W}_n^*)^\top,{\rm Im}({\bf W}_n^*)^\top)^\top $ imitates the structure of \bee \sqrt{n}(\widetilde {\bf R}_n - \widetilde {\bf R}) = \sqrt{n} \left[ \tilde g\begin{pmatrix}{\rm Re}({\bf M}_n)\\ {\rm Im}({\bf M}_n) \end{pmatrix} - \tilde g\begin{pmatrix}{\rm Re}({\bf M})\\ {\rm Im}({\bf M}) \end{pmatrix} \right]\,. \eee But as a comparison of \eqref{Th39_h3} in the proof of Theorem~\ref{th.2} with \eqref{CLT_smoothfun} shows, the limiting distribution of $ \sqrt{n}(\widetilde {\bf R}_n - \widetilde {\bf R}) $ is only partially captured by $ ( {\rm Re}({\bf W}_n^*)^\top,{\rm Im}({\bf W}_n^*)^\top)^\top $: While the factor determined by the Jacobi matrix is established properly, the matrix $ {\bf G} $ entering the covariance structure of the normal distribution is not captured. This is corrected in Steps $ \widetilde{\textrm{III}} $ and $ \widetilde{\textrm{IV}} $ where the bootstrap random vectors are first standardized via multiplication with $ \big(\widetilde{\bf G}^*_n \big)^{-1/2} $. Then the proper variance is established by $ \big(\widetilde{\bf G}^\circ_n \big)^{1/2} $ which simultaneously approximates both $ {\bf G} $ and the Jacobi matrix $ {\bf J}_{\tilde g}(( {\rm Re}({\bf M})^\top,{\rm Im}({\bf M})^\top)^\top) $. \end{enumerate} \end{remark} \begin{example} One very important statistic of interest that can be written as a smooth function of integrated periodograms is the sample cross-correlation. Examples \ref{examplesamplecorr} and \ref{examplesamplecorrcontd} show how the function $ g $ can be defined such that $ \sqrt{n}(\widehat \rho_{rs}(h)-\rho_{rs}(h)) $ takes the form $ \sqrt{n}({\bf R}_n - {\bf R}) $. The limiting variance of this expression in general takes the form indicated in \eqref{CLT_smoothfun_g}, and for this particular example it is given by the expression $ \tau^2 $ from Example \ref{examplesamplecorrcontd}. Since the function $ g $ fulfils Assumption \ref{assu4}, our bootstrap algorithm $ {\bf W}_n^\circ $ is asymptotically valid for this statistic (as is proven in Theorem \ref{th.2}), and successfully imitates the rather complicated expression $ \tau^2 $. \end{example} \subsection{Bootstrap validity} We need the following consistency assumption for the spectral density estimator $ \widehat{\bf f} $ used in Step I of the bootstrap procedure in Section \ref{subsec_procedure}. \begin{assumption} \label{assu5} The spectral density estimator $ \widehat{\bf f} $ is Hermitian, positive definite and satisfies \bee \sup_{\lambda\in[-\pi,\pi]}\|\widehat{\bf f}(\lambda)-{\bf f}(\lambda)\|_F \stackrel{P}{\rightarrow} 0\,. \eee \end{assumption} We refer to Wu and Zaffaroni (2018) for estimators of the spectral density matrix and for general classes of multivariate processes for which the above assumption is satisfied. We next establish the following two lemmas, the proofs of which are given in the Supplementary Material. \vspace*{0.2cm} \begin{lemma} \label{le.BootLem1} Let Assumption \ref{assu1} be fulfilled. Denote by $ I_{t;rs}(\lambda) $ the $(r,s)$-th element of the periodogram matrix $ {\bf I}_t(\lambda)$ based on the subsample $ {\bf X}(t), {\bf X}(t+1), \ldots, {\bf X}(t+b-1)$. Then, \begin{enumerate} \item[(i)] \ $ \sum_{\ell \in {\mathcal G}(b)} \big|\widetilde f_{rs}(\lambda_{\ell,b}) -E(I_{1;rs}(\lambda_{\ell,b})) \big| = {\mathcal O}_P(\sqrt{b^3/(n-b+1)})\,,$ \item[(ii)] \begin{align*} \sum_{\ell_1,\ell_2\in {\mathcal G}(b)} & \Big|\frac{\displaystyle 1}{\displaystyle n-b+1}\sum_{t=1}^{n-b+1} I_{t;r_js_j}(\lambda_{\ell_1,b})I_{t;r_ks_k}(\lambda_{\ell_2,b}) \\ & - E\big(I_{1;r_js_j}(\lambda_{\ell_1,b})I_{1;r_ks_k}(\lambda_{\ell_2,b}) \big) \Big| ={\mathcal O}_P(\sqrt{b^5/(n-b+1)})\,, \end{align*} and \begin{align*} \sum_{\ell\in {\mathcal G}(b)} & \Big|\frac{\displaystyle 1}{\displaystyle n-b+1}\sum_{t=1}^{n-b+1} I_{t;r_js_j}(\lambda_{\ell,b})I_{t;r_ks_k}(\lambda_{\ell,b}) \\ & - E\big(I_{1;r_js_j}(\lambda_{\ell,b})I_{1;r_ks_k}(\lambda_{\ell,b}) \big) \Big| ={\mathcal O}_P(\sqrt{b^3/(n-b+1)})\,. \end{align*} \end{enumerate} where the $ {\mathcal O}_P$ terms are uniformly in $r,s$, respectively $ r_j,s_j,r_k,s_k$. \end{lemma} \vspace*{0.2cm} In order to formulate our theoretical results dealing with the asymptotic properties of the bootstrap approximations proposed, we state the following assumption which summarizes our requirements regarding the behavior of the subsampling parameter $b$. \begin{assumption} \label{assu6} $ b=b(n) \rightarrow \infty$ as $ n \rightarrow \infty$ such that $ b^3/(n-b+1) \rightarrow 0$. \end{assumption} \vspace*{0.2cm} The following lemma investigates the asymptotic properties of the covariance and relation matrices of the random vectors $ {\bf V}^\ast_n$ and $ {\bf V}^+_n $ generated in the MFHB procedure and derives the limiting distribution of $ {\bf V}^\ast_n$. \begin{lemma} \label{le.BootLem2} If Assumptions \ref{assu1}, \ref{assu2}, \ref{assu3}, \ref{assu5} and \ref{assu6} are satisfied, then, \begin{enumerate} \item[(i)] \ $ {\rm Cov}^\ast({\bf V}_n^\ast) \stackrel{P}{\rightarrow} {\bf \Sigma}_1$ and $ E^\ast({\bf V}_n^\ast ({\bf V}_n^{\ast})^{ \top}) \stackrel{P}{\rightarrow} {\bf \Gamma}_1 $. \item[] \item[(ii)] \ $ {\bf G}_n^+ - {\bf C}_n^+ \stackrel{P}{\rightarrow} \left(\begin{array}{cc} \frac{1}{2}({\rm Re}({\bf \Sigma}_2 )+{\rm Re}({\bf \Gamma}_{2})) & \frac{1}{2}(-{\rm Im}({\bf \Sigma}_{2}) +{\rm Im}({\bf \Gamma}_{2}))\\ & \\ \frac{1}{2}({\rm Im}({\bf \Sigma}_{2}) +{\rm Im}({\bf \Gamma}_{2}))& \frac{1}{2}({\rm Re}({\bf \Sigma}_{2}) -{\rm Re}({\bf \Gamma}_{2})) \end{array} \right).$ \item[] \item[(iii)] \ $ {\bf V}^\ast_n \stackrel{d}{\rightarrow} {\mathcal N}_J^c({\bf 0}, {\bf \Sigma}_1, {\bf \Gamma}_1)$ in $ P $-probability. \end{enumerate} \end{lemma} \vspace*{0.2cm} Lemma~\ref{le.BootLem2} leads to the following result which establishes consistency of the MFHB procedure in estimating the distribution and the second-order moments of the random vector $ {\bf V}_n$. \begin{theorem} \label{th.1} If Assumptions \ref{assu1}, \ref{assu2}, \ref{assu3}, \ref{assu5} and \ref{assu6} are satisfied, then, \begin{enumerate} \item[(i)] \ $ {\rm Cov}^\ast({\bf V}_{n}^\circ) \stackrel{P}{\rightarrow} {\bf \Sigma}$, and $ E^\ast \big({\bf V}_{n}^\circ ( {\bf V}_{n}^\circ)^{\top}\big) \stackrel{P}{\rightarrow} {\bf \Gamma} $. \item[(ii)] \ $ {\bf V}_{n}^\circ \stackrel{d}{\rightarrow} {\bf V} $, in $ P $-probability. \end{enumerate} \end{theorem} The next result establishes validity of the MFHB procedure for smooth functions of integrated periodograms. \begin{theorem} \label{th.2} Let Assumptions \ref{assu1} to \ref{assu6} be satisfied. Then, as $n \rightarrow \infty$, \bee {\bf W}^\circ_n \stackrel{d}{\longrightarrow} {\bf W} \eee in $ P $-probability, where $ {\bf W}$ is the limiting distribution of $ \sqrt{n}({\bf R}_n-{\bf R}) $ given in \eqref{CLT_smoothfun_g}. \end{theorem} \section{Simulations} \subsection{Choice of the MFHB parameters} \label{SubSim1} The practical implementation of the MFHB procedure requires the choice of two parameters. The first is the spectral density estimator $ \widehat{{\bf f}}$ and the second the subsampling parameter $b$. Assumption 5 and Assumption 6 state our general requirements on these parameters focusing on the asymptotic properties they have to satisfy in order for the proposed bootstrap procedure to be consistent. Certainly, the choice of these parameters for a given sample size $n$ is an important venue of feature research. In the following we discuss some rather practical rules on how to choose these parameters.. Regarding the spectral density estimator $\widehat{{\bf f}}$, a variety of estimators exists which can be used in our procedure; see Brillinger (1981). As a simple approach, we use in the following kernel estimators obtained by locally averaging the periodogram matrix over frequencies close to the frequency of interest, i.e., $\widehat{{\bf f}}(\lambda) = (nh)^{-1}\sum_{j} K((\lambda-\lambda_{j,n})/h) {\bf I}_n(\lambda_{j,n})$. $K$ denotes the kernel function which determines the weights assigned to the ordinates of the periodogram matrix, while $h$ is the bandwidth that controls the number of periodogram ordinates effectively taken into account in order to obtain the kernel estimator $\widehat{{\bf f}}$. To select the parameter $h$ in practice, cross-validation type approaches have been proposed and investigated in the literature which also can be applied in our setting; see Robinson (1991) for details. For the subsampling parameter $b$, Assumption 6 solely states the required conditions on the rate at which this parameter has to increase to infinity with respect to the sample size $n$ in order to ensure consistency of the MFHB procedure. Our simulation experience with the choice of this parameter shows that the results obtained are not very sensitive with respect to the choice of $b$, provided that this parameter is not chosen too small. This motivates the suggestion of the following practical rule for selecting $b$. Select this parameter as the smallest integer which is larger or equal to $ 3\cdot n^{0.30}$. This rule satisfies the requirements of Assumption 6 and at the same time delivers a value of $b$ which is large enough for the MFHB procedure to perform well in practice. Notice that the numerical results presented in the next section are reported for different combinations of bandwidth and block size parameters, $ h $ and $b$. On the one hand this avoids a further increase of the computational burden caused by a cross-validation type choice of the bandwidth $h$. On the other hand it allows us to investigate the sensitivity of the bootstrap estimates with respect to different choices of the parameters involved. \subsection{Numerical results} In this section we investigate the finite sample performance of the MFHB procedure and compare it with that of the time domain moving block bootstrap (MBB). Note that, in view of our discussion in the Introduction, popular time domain bootstrap methods other than the MBB (and its variations) can not be considered as competitors for our MFHB since these procedures are asymptotically non-valid for a large class of statistics in the context of multivariate linear and non-linear time series. We consider time series of length $ n=100$ stemming from the following two bivariate processes considered in Tsay (2014). The first is a VAR(1) process driven by i.i.d. innovations, i.e., \[\hspace*{-2.5cm} \mbox{Model I:} \ \ \ \ {\bf X}(t) = {\bf \Phi}{\bf X}(t-1) + {\bf e}(t), \ \ {\bf \Phi}= \left(\begin{array}{rr} 0.8 & 0.4 \\ -0.3 & 0.6 \end{array} \right), \] where $ {\bf e}_t \sim {\mathcal N}_2({\bf 0}, {\bf S}_{\bf e})$ and $ {\bf S}_{\bf e}=(\sigma_{i,j})_{i,i=1,2}$, with $ \sigma_{1,1}=2.0$, $\sigma_{1,2}=0.5$ and $ \sigma_{2,2}=1$. The second model is a bivariate VARMA(2,1) process the innovations of which follow a bivariate GARCH-type process, that is, \[\hspace*{-1.5cm} \mbox{Model II:} \ \ \ \ {\bf X}(t) = {\bf \Phi}_1{\bf X}(t-1) + {\bf \Phi}_2{\bf X}(t-2) + {\bf \Theta} {\bf u}(t-1) + {\bf u}(t), \] with parameter matrices ${\bf \Phi}_1= \left(\begin{array}{rr} 0.816 & -0.623 \\ -1.116 & 1.074 \end{array} \right)$, \[ {\bf \Phi}_2 = \left(\begin{array}{rr} -0.643 & 0.592 \\ 0.615 & -0.133 \end{array} \right) \ \mbox{and} \ \ {\bf \Theta}= \left(\begin{array}{rr} 0 & -1.248 \\ -0.801 & 0 \end{array} \right).\] Furthermore, the innovations $ {\bf u}(t)$ are generated as $ {\bf u}(t)={\bf S}_{t}^{1/2}{\bf e}(t)$, where the $ {\bf e}(t)$'s are i.i.d.~${\mathcal N}_2({\bf 0}, {\bf Id}_2)$ distributed, and the volatility matrix $ {\bf S}_{t}$ evolves according to a BEKK(1,1) model, i.e.~$ {\bf S}_{t}={\bf A}_0{\bf A}^\top_0 + {\bf A}_1{\bf u}(t-1){\bf u}(t-1)^\top {\bf A}_1^\top + {\bf B}_1 {\bf S}_{t-1} {\bf B}_1^\top$, where \[{\bf A}_0= \left(\begin{array}{rr} 0.01 & 0 \\ 0 & 0.01 \end{array} \right), \ {\bf A}_1= \left(\begin{array}{rr} 0.15 & 0.20 \\ 0.06 & 0.40 \end{array} \right) \ \mbox{and} \ {\bf B}_1= \left(\begin{array}{rr} 0.9 & 0 \\ 0 & 0.9 \end{array} \right).\] Observe that for the VAR(1) model with Gaussian innovations the fourth-order cumulant spectral densities $f_{rsvw}$ equal zero, so the expressions of the covariance and relation matrices $ {\bf \Sigma} $ and $ {\bf \Gamma}$ simplify. This is not the case for the VARMA(2,1) process which is nonlinear due to the BEKK(1,1) generated innovations. The parameter matrices of this BEKK(1,1) model are very close to those of the same model fitted to the IBM stock and S\&P composite index in Tsay (2014), p. 418, Table 7.3. See also Francq and Zako\"ian (2016), Section 6, for a similar parametrization. We consider the problem of estimating the standard deviation of the cross-correlation estimates $ \widehat{\rho}(h)$ for the values $ h=-1, 0,+1$. Recall that the estimators considered also can be written as functions of integrated periodograms; see Example 2.3. We have generated 10,000 replications of both models in order to estimate the exact standard deviation of the sample cross-correlations considered. As already mentioned, the spectral density estimator $ \widehat{{\bf f}}$ used in the MFHB procedure is a kernel estimator obtained via smoothing the periodogram matrix with bandwidth $h$ and using the Bartlett-Priestley kernel; see Priestley (1981). Furthermore, and in order to see the sensitivity of the bootstrap methods compared with respect to the choice of the bootstrap parameters, the MFHB procedure has been applied using different choices of the bandwidth $h$ and of the subsampling parameter $b$. The same values of $b$ have also been used as block sizes in the MBB procedure. We also report results for $b=12$ which is the value of the subsampling parameter selected according to the rule proposed in Section \ref{SubSim1}. Table 1 presents the results for both models considered. The results reported in this table are based on $R=500$ repetitions where $ B=300$ bootstrap replications have been applied for each repetition. As this table shows, the MFHB procedure performs quite well for both models considered and outperforms the MBB. In particular, comparing the performance of both bootstrap procedures for the same subsampling parameter, respectively block size $b$, the MFHB mean square errors are in almost all cases considered, and independent of the choice of $h$, lower than those of the MBB procedure. Furthermore, the MFHB estimates seem to be less sensitive with respect to the choice of the subsampling parameter $b$ than the MBB procedure is with respect to the choice of the block size $b$. \begin{small} \begin{table}[ht] \label{table1} \hspace*{-1.2cm} \begin{tabular}{llllcllcllc} \hline & & & & & & & & & & \\ & &$\widehat{\rho}(-1)$ & & & $\widehat{\rho}(0)$& & &$\widehat{\rho}(+1) $& & \\ & & Mean& Std & MSE$\times$10& Mean & Std & MSE$\times$10 &Mean & Std & MSE$\times$10\\ \hline & & & & & & & & & & \\ & & & & \multicolumn{4}{c}{\normalsize{ MODEL I }} & & \\ & & \multicolumn{3}{l}{Est. Exact: 0.766} &\multicolumn{3}{l}{Est. Exact: 0.992} &\multicolumn{3}{l}{Est. Exact: 1.131} \\ {\normalsize MFHB} & & & & & & & & & & \\ h=0.10 & b=6 & 0.794 &0.123 & 0.158 &1.014 &0.129 & 0.176& 1.142& 0.167 & 0.283 \\ & b=8 &0.788 & 0.130 & 0.172 & 1.024 &0.140 &0.206 & 1.153& 0.188 & 0.358\\ & b=10 &0.787 & 0.129 &0.168 & 1.011 &0.146 &0.214 & 1.143& 0.192 & 0.368\\ & b=12 &0.788& 0.136 &0.188 & 1.000 &0.160 &0.256 & 1.130& 0.192 & 0.367\\ & b=16 &0.771 & 0.142 &0.203 & 0.958 &0.171 &0.300 & 1.104& 0.212 & 0.456\\ & & & & & & & & & & \\ h=0.12 & b=6 &0.824 & 0.124 &0.187 & 1.027 & 0.140 & 0.207 & 1.159& 0.172 & 0.308\\ & b=8 &0.828 & 0.122 &0.178 & 1.045 & 0.126 & 0.181 & 1.171& 0.170 & 0.300\\ & b=10 &0.821 & 0.117 & 0.162 & 1.030 & 0.129 & 0.181 & 1.152& 0.171 & 0.297\\ & b=12 &0.815& 0.127 &0.179 & 1.023 &0.147 &0.226 & 1.156& 0.170 & 0.295\\ & b=16 &0.799 & 0.145 &0.243 & 1.005 &0.167 &0.283 & 1.140& 0.191 & 0.365\\ & & & & & & & & & & \\ {\normalsize MBB} & & & & & & & & & & \\ & b=6 &0.906 & 0.112 & 0.323 & 1.016 &0.162& 0.273 & 1.092& 0.152 & 0.248\\ & b=8 & 0.849 & 0.122 & 0.201 & 0.977 &0.181 &0.332 & 1.055& 0.185 & 0.406\\ & b=10 & 0.832& 0.143 & 0.237 & 0.955 & 0.198 & 0.414 & 1.048& 0.199 & 0.475\\ & b=12 &0.795& 0.145 &0.217 & 0.933 &0.208 &0.477 & 1.020& 0.205 & 0.548\\ & b=16 &0.750 & 0.164 &0.271 & 0.875 &0.208 &0.559 & 0.966& 0.217 & 0.743\\ & & & & & & & & & & \\ & & & & \multicolumn{4}{c}{\normalsize{ MODEL II }} & & \\ & & \multicolumn{3}{l}{Est. Exact: 1.154} &\multicolumn{3}{l}{Est. Exact: 1.135} &\multicolumn{3}{l}{Est. Exact: 1.217} \\ {\normalsize MFHB} & & & & & & & & & & \\ h=0.10 & b=6 & 1.164 &0.232 & 0.570 &1.160 &0.166 & 0.300 &1.195& 0.196 & 0.424 \\ & b=8 &1.153 & 0.263 &0.620 & 1.168 &0.168 &0.301 & 1.196& 0.195 & 0.409\\ & b=10 &1.140 & 0.235 &0.633 & 1.148&0.181 &0.331 & 1.193& 0.183 & 0.353\\ & b=12 &1.113& 0.240 &0.785 & 1.115 &0.205 &0.428 & 1.193& 0.216 & 0.491\\ & b=16 &1.113 & 0.248 &0.742 & 1.084 &0.229 &0.536 & 1.144& 0.245 & 0.601\\ & & & & & & & & & & \\ h=0.12 & b=6 &1.167 & 0.224 & 0.544 & 1.195 &0.173 &0.337 & 1.222& 0.195 & 0.431\\ & b=8 &1.158 & 0.222 & 0.547 & 1.193 & 0.160 & 0.311 & 1.216& 0.185 & 0.399\\ & b=10 &1.184 & 0.230 & 0.538 & 1.200 &0.176 & 0.357 & 1.227& 0.186 & 0.387\\ & b=12 &1.149& 0.227 &0.578 & 1.159 &0.194 &0.394 & 1.221& 0.203 & 0.498\\ & b=16 &1.134 & 0.235 &0.661 & 1.111 &0.208 &0.437 & 1.168& 0.232 & 0.538\\ & & & & & & & & & & \\ {\normalsize MBB} & & & & & & & & & & \\ & b=6 &1.176& 0.208 & 0.454 & 1.229 & 0.255 &0.784 & 1.191& 0.221 & 0.524\\ & b=8 &1.157 & 0.245 &0.655 & 1.189 &0.273 & 0.782 & 1.187& 0.246 & 0.623\\ & b=10 &1.096 & 0.235 &0.730 & 1.131 &0.268 &0.718 & 1.162& 0.263 & 0.695\\ & b=12 &1.048& 0.238 &0.999 & 1.090 &0.271 &0.766 & 1.116& 0.269 & 0.735\\ & b=16 &1.024 & 0.245 &1.002 & 1.057 &0.289 &0.870 & 1.080& 0.290 & 0.888\\ & & & & & & & & & & \\ \hline \end{tabular} {\normalsize {\rm TABLE 1: \ Bootstrap estimates of the standard deviation of the sample cross-correlations $ \widehat{\rho}(h)$ for lags $ h \in \{-1,0,+1\}$ for time series of length $n=100$ stemming from Model I and Model II. MFHB refers to the estimates of the multivariate frequency domain hybrid bootstrap and MBB to those of the moving block bootstrap.}} \end{table} \end{small} \section{Appendix: Proofs} $\left.\right.$\\ {\sc Proof of Lemma~\ref{le.BootLem2} $(i)$:} The $ (j,k) $-th entry of the covariance matrix $ {\rm Cov}^\ast({\bf V}_{n}^\ast) $ is given by \bee {\rm Cov}^\ast(V_{n,j}^\ast,V_{n,k}^\ast) &=& {\rm Cov}^\ast \left( \frac{2\pi}{\sqrt n} \sum_{l_1 \in \mathcal G(n)} \varphi_j(\lambda_{l_1,n}) \big( I^\ast_{r_js_j}(\lambda_{l_1,n}) - \widehat f_{r_js_j}(\lambda_{l_1,n}) \big), \right. \\ & & \quad \quad \quad \left. \frac{2\pi}{\sqrt n} \sum_{l_2 \in \mathcal G(n)} \varphi_k(\lambda_{l_2,n}) \big( I^\ast_{r_ks_k}(\lambda_{l_2,n}) - \widehat f_{r_ks_k}(\lambda_{l_2,n}) \big) \right) \eee which equals \be \quad \quad \frac{4\pi^2}{n} \sum_{l_1,l_2 \in \mathcal G(n)} \varphi_j(\lambda_{l_1,n}) \overline{\varphi_k(\lambda_{l_2,n})} \, {\rm Cov}^\ast\big( I^\ast_{r_js_j}(\lambda_{l_1,n}), I^\ast_{r_ks_k}(\lambda_{l_2,n}) \big)\,. \label{proofh1} \ee The last expression can be decomposed, using $ \lambda_{-l,n}=-\lambda_{l,n} $ and $ I^\ast_{rs}(-\lambda_{l,n})=I^\ast_{sr}(\lambda_{l,n}) $, into \begin{align} \frac{4\pi^2}{n} \sum_{l_1,l_2=1}^{N} &\left\{ \varphi_j(\lambda_{l_1,n}) \overline{\varphi_k(\lambda_{l_2,n})} \, {\rm Cov}^\ast\big( I^\ast_{r_js_j}(\lambda_{l_1,n}), I^\ast_{r_ks_k}(\lambda_{l_2,n}) \big) \right. \label{proofh2}\\ &+ \varphi_j(-\lambda_{l_1,n}) \overline{\varphi_k(-\lambda_{l_2,n})} \, {\rm Cov}^\ast\big( I^\ast_{s_jr_j}(\lambda_{l_1,n}), I^\ast_{s_kr_k}(\lambda_{l_2,n}) \big) \nonumber\\ &+ \varphi_j(-\lambda_{l_1,n}) \overline{\varphi_k(\lambda_{l_2,n})} \, {\rm Cov}^\ast\big( I^\ast_{s_jr_j}(\lambda_{l_1,n}), I^\ast_{r_ks_k}(\lambda_{l_2,n}) \big) \nonumber\\ &\left.+ \varphi_j(\lambda_{l_1,n}) \overline{\varphi_k(-\lambda_{l_2,n})} \, {\rm Cov}^\ast\big( I^\ast_{r_js_j}(\lambda_{l_1,n}), I^\ast_{s_kr_k}(\lambda_{l_2,n}) \big) \right\}\,. \nonumber \end{align} For the first of the four similar summands in this expression we can calculate the following, where we used that $ {\bf I}^\ast(\lambda_{l_1,n}) $ and $ {\bf I}^\ast(\lambda_{l_2,n}) $ are independent for $ l_1\neq l_2 $ and $ 1 \leq l_1,l_2 \leq N $, and that $ {\bf I}^\ast(\lambda_{l_1,n}) $ has a complex Wishart $ W^C_m(1,{\bf \widehat f}(\lambda_{l_1,n}) )$ distribution (the covariance structure of which can be obtained from, e.g., Brillinger (1981), Section 4.2): \begin{align*} & \frac{4\pi^2}{n} \sum_{l_1,l_2=1}^{N} \varphi_j(\lambda_{l_1,n}) \overline{\varphi_k(\lambda_{l_2,n})} \, {\rm Cov}^\ast\big( I^\ast_{r_js_j}(\lambda_{l_1,n}), I^\ast_{r_ks_k}(\lambda_{l_2,n}) \big) \\ =\, & \frac{4\pi^2}{n} \sum_{l=1}^{N} \varphi_j(\lambda_{l,n}) \overline{\varphi_k(\lambda_{l,n})} \, {\rm Cov}^\ast\big( I^\ast_{r_js_j}(\lambda_{l,n}), I^\ast_{r_ks_k}(\lambda_{l,n}) \big) \\ =\, & \frac{4\pi^2}{n} \sum_{l=1}^{N} \varphi_j(\lambda_{l,n}) \overline{\varphi_k(\lambda_{l,n})} \, \widehat f_{r_jr_k}(\lambda_{l,n}) \, \overline{\widehat f_{s_js_k}(\lambda_{l,n})} \\ =\, & \frac{4\pi^2}{n} \sum_{l=1}^{N} \varphi_j(\lambda_{l,n}) \overline{\varphi_k(\lambda_{l,n})} \, f_{r_jr_k}(\lambda_{l,n}) \, f_{s_js_k}(-\lambda_{l,n}) \\ & + \frac{4\pi^2}{n} \sum_{l=1}^{N} \varphi_j(\lambda_{l,n}) \overline{\varphi_k(\lambda_{l,n})} \, \big(\widehat f_{r_jr_k}(\lambda_{l,n}) \, \widehat f_{s_js_k}(-\lambda_{l,n}) - f_{r_jr_k}(\lambda_{l,n}) \, f_{s_js_k}(-\lambda_{l,n}) \big)\,. \end{align*} The second summand on the last right-hand side vanishes asymptotically due to Assumption \ref{assu5} and $ N=\mathcal O(n) $. Since the first summand is a Riemann sum the last right-hand side converges in probability to \bee 2\pi \int_0^\pi \varphi_j(\lambda) \overline{\varphi_k(\lambda)} \, f_{r_jr_k}(\lambda) f_{s_js_k}(-\lambda) \, d\lambda\,. \eee With analogous calculations the other three summands in \eqref{proofh2} yield \bee & 2\pi \int_0^\pi \varphi_j(-\lambda) \overline{\varphi_k(-\lambda)} \, f_{s_js_k}(\lambda) f_{r_jr_k}(-\lambda) \, d\lambda \\ + & 2\pi \int_0^\pi \varphi_j(-\lambda) \overline{\varphi_k(\lambda)} \, f_{s_jr_k}(\lambda) f_{r_js_k}(-\lambda) \, d\lambda \\ + & 2\pi \int_0^\pi \varphi_j(\lambda) \overline{\varphi_k(-\lambda)} \, f_{r_js_k}(\lambda) f_{s_jr_k}(-\lambda) \, d\lambda\,. \eee Substituting $ \lambda $ with $ -\lambda $ in the first and second of these three terms shows that the limit in probability of \eqref{proofh2} is given by $ \Sigma_{1;jk} $.\\ As for the relation matrix, observe that the $ (j,k) $-th entry of $ E^\ast({\bf V}_{n}^\ast {\bf V}_{n}^{\ast \top}) $ is given by \bee {\rm Cov}^\ast(V_{n,j}^\ast,\overline{V_{n,k}^\ast}) = \frac{4\pi^2}{n} \sum_{l_1,l_2 \in \mathcal G(n)} \varphi_j(\lambda_{l_1,n}) \varphi_k(-\lambda_{l_2,n}) \, {\rm Cov}^\ast\big( I^\ast_{r_js_j}(\lambda_{l_1,n}), I^\ast_{r_ks_k}(\lambda_{l_2,n}) \big)\,, \eee which is \eqref{proofh1} if one replaces $ \varphi_k(\cdot) $ with $ \overline{\varphi_k(-\,\cdot)} $. Therefore, one can follow along the lines of the calculation for $ \Sigma_{1;jk} $ above to see that the $ (j,k) $-th entry of $ E^\ast({\bf V}_{n}^\ast {\bf V}_{n}^{\ast \top}) $ converges in probability to $ \Gamma_{1;jk} $. $ \hfill \square $ \ {\sc Proof of Lemma~\ref{le.BootLem2} $(ii)$:} \ We first establish the uniform (over the Fourier frequencies) consistency of $ \widetilde{{\bf f}}^{-1/2} $ and $ \widehat{{\bf f}}^{1/2}$ as estimators of $ {\bf f}^{-1/2}$ and $ {\bf f}^{1/2}$, respectively. We first show that \begin{equation}\label{eq.Lemma3.9.1} \max_{\ell \in {\mathcal G}(b)} \| \widetilde{{\bf f}}^{-1}(\lambda_{\ell,b}) - {\bf f}^{-1}(\lambda_{\ell,b})\|_F \stackrel{P}{\rightarrow} 0. \end{equation} For this notice first that \begin{equation}\label{eq.Lemma3.9.1b} \max_{\ell \in {\mathcal G}(b)} \| \widetilde{{\bf f}}(\lambda_{\ell,b}) - {\bf f}(\lambda_{\ell,b})\|_F \stackrel{P}{\rightarrow} 0. \end{equation} This follows since $ E(I_{1;rs}(\lambda_{\ell,b}) )= f_{rs}(\lambda_{\ell,b}) + \mathcal O(b^{-1})$ uniformly in $r,s$ and \begin{align*} \max_{\ell \in {\mathcal G}(b)} \| \widetilde{{\bf f}}(\lambda_{\ell,b}) - {\bf f}(\lambda_{\ell,b})\|_F & \leq \max_{\ell \in {\mathcal G}(b)} \| \widetilde{{\bf f}}(\lambda_{\ell,b}) - E\widetilde{{\bf f}}(\lambda_{\ell,b})\|_F\\ & \ \ \ \ \ \ + \max_{\ell \in {\mathcal G}(b)} \| E\widetilde{{\bf f}}(\lambda_{\ell,b}) - {\bf f}(\lambda_{\ell,b})\|_F\\ & = \max_{\ell \in {\mathcal G}(b)} \Big\{\sum_{r=1}^m\sum_{s=1}^m \big| \widetilde{f}_{rs}(\lambda_{\ell,b}) - E \widetilde{f}_{rs}(\lambda_{\ell,b}) \big|^2 \Big\}^{1/2} \\ & \ \ + \max_{\ell \in {\mathcal G}(b)} \Big\{\sum_{r=1}^m\sum_{s=1}^m \big| E(I_{1;rs}(\lambda_{\ell,b})) - f_{rs}(\lambda_{\ell,b}) \big|^2 \Big\}^{1/2}\\ & \leq \sum_{r=1}^m\sum_{s=1}^m \sum_{\ell\in{\mathcal G}(b)}\big| \widetilde{f}_{rs}(\lambda_{\ell,b}) - E \widetilde{f}_{rs}(\lambda_{\ell,b}) \big| + \mathcal O(b^{-1})\\ & = \mathcal O_P(\sqrt{b^3/(n-b+1)}) + \mathcal O(b^{-1}) \rightarrow 0, \end{align*} where the last convergence holds true because of Lemma 3.8 (i) and Assumption 6. Now to see (\ref{eq.Lemma3.9.1}) notice that $ \max_{\ell\in{\mathcal G}(b)}\|{\bf f}^{-1}(\lambda_{\ell,b})\|^2_F =\mathcal O(1)$ and that \begin{align} \label{eq.hilf1} \max_{\ell \in {\mathcal G}(b)} \| \widetilde{{\bf f}}^{-1} & (\lambda_{\ell,b}) - {\bf f}^{-1}(\lambda_{\ell,b})\|_F = \max_{\ell \in {\mathcal G}(b)} \| \widetilde{{\bf f}}^{-1}(\lambda_{\ell,b}) \big( {\bf f}(\lambda_{\ell,b}) - \widetilde{{\bf f}}(\lambda_{\ell,b})\big) {\bf f}^{-1}(\lambda_{\ell,b})\|_F \nonumber \\ & \leq \max_{\ell \in {\mathcal G}(b)} \| \widetilde{{\bf f}}^{-1}(\lambda_{\ell,b})\|_F \max_{\ell \in {\mathcal G}(b)} \| \widetilde{{\bf f}}(\lambda_{\ell,b}) - {\bf f}(\lambda_{\ell,b})\|_F \max_{\ell \in {\mathcal G}(b)} \| {\bf f}^{-1}(\lambda_{\ell,b})\|_F \nonumber \\ & \leq \big( \max_{\ell \in {\mathcal G}(b)} \| {\bf f}^{-1}(\lambda_{\ell,b})\|_F + \max_{\ell \in {\mathcal G}(b)} \| \widetilde{{\bf f}}^{-1}(\lambda_{\ell,b}) - {\bf f}^{-1}(\lambda_{\ell,b})\|_F\big) \nonumber \\ & \ \ \ \times \max_{\ell \in {\mathcal G}(b)} \| \widetilde{{\bf f}}(\lambda_{\ell,b}) - {\bf f}(\lambda_{\ell,b})\|_F \max_{\ell \in {\mathcal G}(b)} \| {\bf f}^{-1}(\lambda_{\ell,b})\|_F \end{align} By (\ref{eq.Lemma3.9.1b}) and for $n$ large enough such that $$ \max_{\ell \in {\mathcal G}(b)}\|\widetilde{{\bf f}}(\lambda_{\ell,b})-{\bf f}(\lambda_{\ell,b})\|_F \max_{\ell \in {\mathcal G}(b)}\|{\bf f}^{-1}(\lambda_{\ell,b})\|_F< 1,$$ expression (\ref{eq.hilf1}) leads to \begin{align*} \max_{\ell \in {\mathcal G}(b)} \| \widetilde{{\bf f}}^{-1}(\lambda_{\ell,b}) & - {\bf f}^{-1}(\lambda_{\ell,b})\|_F \\ & \leq \frac{\displaystyle \max_{\ell \in {\mathcal G}(b)}\|{\bf f}^{-1}(\lambda_{\ell,b})\|^2_F \max_{\ell \in {\mathcal G}(b)}\|\widetilde{{\bf f}}(\lambda_{\ell,b})-{\bf f}(\lambda_{\ell,b})\|_F }{\displaystyle 1- \max_{\ell \in {\mathcal G}(b)}\|{\bf f}^{-1}(\lambda_{\ell,b})\|_F \max_{\ell \in {\mathcal G}(b)}\|\widetilde{{\bf f}}(\lambda_{\ell,b})-{\bf f}(\lambda_{\ell,b})\|_F } \stackrel{P}{\rightarrow} 0. \end{align*} Assumption \ref{assu5} together with (\ref{eq.Lemma3.9.1}) implies that \begin{equation} \label{eq.Lemma3.9.2} \max_{\ell \in {\mathcal G}(b)} \| \widehat{{\bf f}}^{1/2}(\lambda_{\ell,b}) - {\bf f}^{1/2}(\lambda_{\ell,b})\|_F \stackrel{P}{\rightarrow} 0 \end{equation} and \begin{equation} \label{eq.Lemma3.9.3} \max_{\ell \in {\mathcal G}(b)} \| \widetilde{{\bf f}}^{-1/2}(\lambda_{\ell,b}) - {\bf f}^{-1/2}(\lambda_{\ell,b})\|_F \stackrel{P}{\rightarrow} 0. \end{equation} To see (\ref{eq.Lemma3.9.2}) notice that since $ \widehat{{\bf f}}$ is positive definite, we get by Assumption 2 and equation (1.3) in Schmitt (1992), that \[ \max_{\ell \in {\mathcal G}(b)} \| \widehat{{\bf f}}^{1/2}(\lambda_{\ell,b}) - {\bf f}^{1/2}(\lambda_{\ell,b})\|_F \leq \frac{1}{\sqrt{\delta}} \max_{\ell \in {\mathcal G}(b)}\| \widehat{{\bf f}}(\lambda_{\ell,b}) - {\bf f}(\lambda_{\ell,b})\|_F.\] For (\ref{eq.Lemma3.9.3}) we get by the same arguments as above and since $ {\bf A}^{-1/2} = ({\bf A}^{-1})^{1/2}$, the bound \[ \max_{\ell \in {\mathcal G}(b)} \| \widetilde{{\bf f}}^{-1/2}(\lambda_{\ell,b}) - {\bf f}^{-1/2}(\lambda_{\ell,b})\|_F \leq \sqrt{\sigma_{\max}} \max_{\ell \in {\mathcal G}(b)} \| \widetilde{{\bf f}}^{-1}(\lambda_{\ell,b}) - {\bf f}^{-1}(\lambda_{\ell,b})\|_F,\] where $ \sigma_{\max} = \max_{\lambda\in [0,\pi]} \max\sigma(f(\lambda))$. Equation (\ref{eq.Lemma3.9.3}) follows then by (\ref{eq.Lemma3.9.1}). Assertions (\ref{eq.Lemma3.9.2}) and (\ref{eq.Lemma3.9.3}) imply that \begin{equation} \label{eq.Root1} \max_{\ell \in {\mathcal G}(b)} \| \widehat{{\bf f}}^{1/2}(\lambda_{\ell,b}) \widetilde{{\bf f}}^{-1/2}(\lambda_{\ell,b}) - {\bf Id}_m\|_F \stackrel{P}{\rightarrow} 0 \end{equation} and \begin{equation} \label{eq.Root2} \max_{\ell \in {\mathcal G}(b)} \| \widetilde{{\bf f}}^{-1/2}(\lambda_{\ell,b}) \widehat{{\bf f}}^{1/2}(\lambda_{\ell,b}) - {\bf Id}_m\|_F \stackrel{P}{\rightarrow} 0. \end{equation} Using (\ref{eq.Root1}) and (\ref{eq.Root2}) it follows by straightforward calculations that \[ \textrm{Cov}^\ast\big(V_{n,j}^+,V_{n,k}^+\big) = \textrm{Cov}^\ast\big(\widetilde{V}_{n,j},\widetilde{V}_{n,k}\big) + o_P(1),\] where \[ \widetilde{V}_{n,j} = \sqrt{kb} \frac{2\pi}{b}\sum_{\ell \in {\mathcal G}(b)} \varphi_j(\lambda_{\ell,b}) \big(\widetilde{I}_{r_js_j}(\lambda_{\ell,b})- \widetilde{f}_{r_js_j}(\lambda_{\ell,b}) \big)\] and $ \widetilde{I}_{rs}(\lambda_{\ell,b})$ denotes the (r,s)-th element of the matrix $ k^{-1}\sum_{m=1}^k{\bf I}_{i_m}(\lambda_{\ell,b})$. Let $ I_{i_1,rs}(\lambda_{\ell,b})$ denote the $(r,s)$-th element of $ {\bf I}_{i_1}(\lambda_{\ell,b}) $. We then have \begin{align*} \textrm{Cov}^\ast(\widetilde{V}_{n,j}, & \widetilde{V}_{n,k}) = \frac{4\pi^2}{b}\sum_{\ell_1, \ell_2\in {\mathcal G}(b)} \varphi_j(\lambda_{\ell_1,b})\overline{\varphi_k(\lambda_{\ell_2,b})} \textrm{Cov}^\ast(I_{i_1,r_js_j}(\lambda_{\ell_1,b}), I_{i_1,r_ks_k}(\lambda_{\ell_2,b}))\\ & =\frac{4\pi^2}{b}\sum_{\ell\in {\mathcal G}(b)} \varphi_j(\lambda_{\ell,b})\overline{\varphi_k(\lambda_{\ell,b})} \textrm{Cov}^\ast(I_{i_1,r_js_j}(\lambda_{\ell,b}), I_{i_1,r_ks_k}(\lambda_{\ell,b}))\\ & \ \ + \frac{4\pi^2}{b}\sum_{\ell\in {\mathcal G}(b)} \varphi_j(\lambda_{\ell,b})\overline{\varphi_k(-\lambda_{\ell,b})} \textrm{Cov}^\ast(I_{i_1,r_js_j}(\lambda_{\ell,b}), I_{i_1,r_ks_k}(-\lambda_{\ell,b}))\\ & \ \ + \frac{4\pi^2}{b}\sum_{|\ell_1|\neq |\ell_2| \in {\mathcal G}(b)} \varphi_j(\lambda_{\ell_1,b})\overline{\varphi_k(\lambda_{\ell_2,b})}\textrm{Cov}^\ast(I_{i_1,r_js_j}(\lambda_{\ell_1,b}), I_{i_1,r_ks_k}(\lambda_{\ell_2,b}))\\ & \stackrel{P}{\rightarrow} \Sigma_{1;jk} + \Sigma_{2;jk}, \end{align*} using \begin{align*} \textrm{Cov}^\ast(& I_{i_1,r_js_j}(\lambda_{\ell_1,b}), I_{i_1,r_ks_k}(\lambda_{\ell_2,b})) \\ & =\frac{1}{n-b+1}\sum_{t=1}^{n-b+1}\big( I_{t,r_js_j}(\lambda_{\ell_1,b})-\widetilde{f}_{r_js_j}(\lambda_{\ell_1,b})\big) \overline{\big( I_{t,r_ks_k}(\lambda_{\ell_2,b})-\widetilde{f}_{r_ks_k}(\lambda_{\ell_2,b})\big)}, \end{align*} and Lemma~\ref{le.BootLem1}. By the same arguments it follows that \[ E^\ast\big(V_{n,j}^+V_{n,k}^+\big) = E^\ast(\widetilde{V}_{n,j}\widetilde{V}_{n,k}\big) + o_P(1),\] and that \[ E^\ast(\widetilde{V}_{n,j}\widetilde{V}_{n,k}\big) \stackrel{P}{\rightarrow} \Gamma_{1;jk} + \Gamma_{2;jk}.\] The above results show that $ {\bf G}_n^+ \stackrel{P}{\rightarrow} {\bf G}$. To conclude the proof we have to show that $ {\bf \Sigma}_{1,n}^+ \stackrel{P}{\rightarrow} {\bf \Sigma_1}$ and that $ {\bf \Gamma}^+_{1,n} \stackrel{P}{\rightarrow} {\bf \Gamma}_{1}$. Using (\ref{eq.Root1}) and (\ref{eq.Root2}) again, it follows that \begin{align*} \sigma^+_{jk} & = \frac{4\pi^2}{b} \sum_{\ell \in {\mathcal G}(b)}\varphi_j(\lambda_{\ell,b}) \overline{\varphi_k(\lambda_{\ell,b})} \mathring{S}_{r_j s_js_kr_k}(\lambda_{\ell,b})\\ & \ \ \ \ + \frac{4\pi^2}{b} \sum_{\ell \in {\mathcal G}(b)}\varphi_j(\lambda_{\ell,b}) \overline{\varphi_k(-\lambda_{\ell,b})} \mathring{S}_{r_js_jr_ks_k}(\lambda_{\ell,b})+ o_P(1). \end{align*} and \begin{align*} c^+_{jk} & = \frac{4\pi^2}{b} \sum_{\ell \in {\mathcal G}(b)}\varphi_j(\lambda_{\ell,b}) \varphi_k(-\lambda_{\ell,b}) \mathring{S}_{r_js_j s_k r_k}(\lambda_{\ell,b})\\ & \ \ \ \ + \frac{4\pi^2}{b} \sum_{\ell\in {\mathcal G}(b)} \varphi_j(\lambda_{\ell,b}) \varphi_k(\lambda_{\ell,b}) \mathring{S}_{r_js_jr_ks_k}(\lambda_{\ell,b}) + o_P(1), \end{align*} where \[ \mathring{S}_{rsuw}(\lambda) = \frac{1}{n-b+1}\sum_{t=1}^{n-b+1}\big( I_{t,rs}(\lambda)-\widetilde{f}_{rs}(\lambda)\big) \big( I_{t,uw}(\lambda) - \widetilde{f}_{uw}(\lambda) \big).\] From Lemma~\ref{le.BootLem1} we then conclude that $ \sigma^+_{jk}\stackrel{P}{\rightarrow} \Sigma_{1;jk} $ and that $ c^+_{jk}\stackrel{P}{\rightarrow} \Gamma_{1;jk} $. \hfil $\Box$ \ {\sc Proof of Lemma~\ref{le.BootLem2} $(iii)$:} We use the notation $ D^*_{rs}(\lambda)=I_{rs}^\ast(\lambda) - \widehat{f}_{rs}(\lambda)$. To show that the $J$-dimensional complex vector \[ {\bf V}^\ast_n=\Big(\frac{2\pi}{\sqrt{n}} \sum_{l\in{\mathcal G}(n)}\varphi_{j}(\lambda_{l,n})D_{r_js_j}^\ast(\lambda_{l,n}), j=1,2, \ldots, J\Big)^\top, \] converges weakly to the complex normal variable $ {\mathcal N}_J^c({\bf 0}, {\bf \Sigma}_1, {\bf \Gamma}_1)$, it suffices to show that the $2J$-dimensional real vector \[ {\bf V}^\ast_{n,T} = \frac{2\pi}{\sqrt{n}} \sum_{l\in{\mathcal G}(n)} \left(\begin{array}{c} {\rm Re}\{\varphi_1(\lambda_{l,n})D_{r_1s_1}^\ast(\lambda_{l,n})\}\\ \vdots \\ {\rm Re}\{\varphi_J(\lambda_{l,n})D_{r_Js_J}^\ast(\lambda_{l,n})\}\\ {\rm Im}\{\varphi_1(\lambda_{l,n})D_{r_1s_1}^\ast(\lambda_{l,n})\}\\ \vdots \\ {\rm Im}\{\varphi_J(\lambda_{l,n})D_{r_Js_J}^\ast(\lambda_{l,n})\}\\ \end{array} \right)\] converges to the $2J$ dimensional, real normal distribution $ {\mathcal N}_{2J}({\bf 0}, {\bf \Sigma}_V)$, where ${\bf \Sigma}_V$ denotes the limit of $E^\ast({\bf V}^\ast_{n,T} {\bf V}^{\ast \top}_{n,T})$, which in view of assertion (i) is well defined. Notice that the covariance matrix $ {\bf \Sigma}_1$ and the relation matrix $ {\bf \Gamma}_1$ together specify $ {\bf \Sigma}_V$ and vice versa. To simplify the presentation we give only the proof for the case where the functions $ \varphi_j$ are real-valued. The case of complex-valued $\varphi_j$ can be proved along the same lines but with a much more complicated notation. Write ${\bf V}_{n,T}^\ast = \sum_{l=1}^N {\bf Z}^\ast_l$, where the $ j $-th component of ${\bf Z}^\ast_l$ is given for $j=1,2, \ldots, J$, by $$ \frac{2\pi}{\sqrt{n}} \Big({\rm Re}\{\varphi_j(\lambda_{l,n})D_{r_js_j}^\ast(\lambda_{l,n})\} + {\rm Re}\{\varphi_j(-\lambda_{l,n})D_{r_js_j}^\ast(-\lambda_{l,n})\} \Big)$$ and for $ j=J+1, \ldots, 2J$, by $$ \frac{2\pi}{\sqrt{n}} \Big({\rm Im}\{\varphi_{j-J}(\lambda_{l,n})D_{r_{j-J}s_{j-J}}^\ast(\lambda_{l,n})\} + {\rm Im}\{\varphi_{j-J}(-\lambda_{l,n})D_{r_{j-J}s_{j-J}}^\ast(-\lambda_{l,n})\}\Big). $$ Observe that the random vectors $ {\bf Z}^\ast_l$ are independent. Thus in view of assertion (i), to establish the desired weak convergence, it suffices to show that Lyapunov's condition is satisfied, that is, that $ \sum_{l=1}^NE^\ast \|{\bf Z}^\ast_l\|^{2+\delta} \rightarrow 0$ for some $ \delta >0$. Choose $ \delta=2$. Using $ (a+b)^2\leq 2a^2 + 2b^2$ and $\varphi(\lambda) {\rm Re}\{D_{rs}^\ast(\lambda)\}^2 + \varphi(\lambda) {\rm Im}\{D_{rs}^\ast(\lambda)\}^2= \varphi(\lambda)|D^\ast_{rs}(\lambda)|^2$ we have \begin{align*} \sum_{l=1}^NE^\ast \|{\bf Z}^\ast_l\|^{4} & \leq \frac{(2\pi)^4}{n^2}\sum_{l=1}^NE^\ast\Big( 2\sum_{j=1}^{J}\varphi_{j}^2(\lambda_{l,n})|D^\ast_{r_js_j}(\lambda_{l,n})|^2 \\ & \ \ \ \ \ \ \ \ + 2\sum_{j=1}^{J}\varphi_{j}^2(-\lambda_{l,n})|D^\ast_{r_js_j}(-\lambda_{l,n})|^2 \Big)^2\\ & \leq \frac{4(2\pi)^4}{n^2} \max_{1\leq j \leq J} \sup_{\lambda\in[-\pi,\pi]} \varphi^2_j(\lambda) \sum_{l=1}^NE^\ast\Big(\sum_{j=1}^J|D^*_{r_js_j}(\lambda_{l,n})|^2 \\ & \ \ \ \ \ \ \ \ + \sum_{j=1}^J|D^*_{r_js_j}(-\lambda_{l,n})|^2\Big)^2\\ & \leq \frac{8(2\pi)^4}{n^2} \max_{1\leq j \leq J} \sup_{\lambda\in[-\pi,\pi]} \varphi^2_j(\lambda) \Big(\sum_{l=1}^N \sum_{j=1}^JE^\ast|D^*_{r_js_j}(\lambda_{l,n})|^4 \\ & \ \ \ \ \ \ \ \ +\sum_{l=1}^N \sum_{j=1}^J E^\ast|D^*_{r_js_j}(-\lambda_{l,n})|^4\Big). \end{align*} Now, recall the definition of $ D^*_{r_js_j}(\lambda)$ and verify by staightforward calculations that \begin{align*} E^\ast |I^*_{rs}(\lambda)-& \widehat{f}_{rs}(\lambda)|^4 = E^\ast[I^*_{rs}(\lambda)^2I^*_{sr}(\lambda)^2] + + \widehat{f}_{rs}(\lambda)^2\widehat{f}_{sr}(\lambda)^2 \\ & - 2\widehat{f}_{rs}(\lambda) E^\ast[I^*_{rs}(\lambda) I^*_{sr}(\lambda)^2] - 2\widehat{f}_{sr}(\lambda)E^\ast[ I^*_{rs}(\lambda)^2 I^*_{sr}(\lambda)] \\ & + \widehat{f}_{rs}(\lambda)^2 E^\ast[I^*_{sr}(\lambda)^2] + \widehat{f}_{sr}(\lambda)^2 E^\ast[I^*_{rs}(\lambda)^2] \\ & + 4 \widehat{f}_{rs}(\lambda) \widehat{f}_{sr}(\lambda) E^\ast[I^*_{rs}(\lambda) I^*_{sr}(\lambda)] \\ & - 2 \widehat{f}_{rs}(\lambda)\widehat{f}_{sr}(\lambda)^2 E^\ast[I^*_{rs}(\lambda)] - 2 \widehat{f}_{rs}(\lambda)^2 \widehat{f}_{sr}(\lambda)E^\ast[I^*_{sr}(\lambda)]\,. \end{align*} Now since $ {\bf I}^\ast(\lambda_{l,n}) $ has for every $ l=1,2, \ldots, N$ a complex Wishart distribution with parameters $1$ and $ \widehat{\bf f}(\lambda_{l,n}) $, for short, ${\bf I}^\ast(\lambda_{l,n}) \sim {\rm Wishart}\big(1,\widehat{\bf f}(\lambda_{l,n})\big)$, we get using Assumption \ref{assu5} and expressions for the moments of the complex Wishart distribution, see Withers and Nadarajah (2012), that $ E^\ast |I^*_{rs}(\lambda)- \widehat{f}_{rs}(\lambda)|^4 \stackrel{P}{\rightarrow} E | S_{rs}(\lambda) - f_{rs}(\lambda)|^4 ={\mathcal O}(1)$, for all $ r,s \in \{1,2, \ldots, m\}$, where the random variable $ {\bf S}(\lambda)=(S_{rs}(\lambda))_{r,s=1,2, \ldots, m}$ has the $ {\rm Wishart}(1,{\bf f}(\lambda))$ distribution. Therefore, \[ \sum_{l=1}^NE^\ast \|{\bf Z}^\ast_l\|^{4} \leq \frac{8(2\pi)^4}{n^2} \max_{1\leq j \leq J} \sup_{\lambda\in[-\pi,\pi]} \varphi^2_j(\lambda)\cdot N\cdot {\mathcal O}_P(1) = {\mathcal O}_P(n^{-1}) \rightarrow 0.\] $ \hfill \square $ \ {\sc Proof of Theorem~\ref{th.1}:} For assertion $(i)$, note that $ E^*({\bf V}_n^\circ)=E^*({\bf V}_n^*) = {\bf 0} $ which implies $ {\rm Cov}^*(({\rm Re}({\bf V}_n^*)^\top,{\rm Im}({\bf V}_n^*)^\top)^\top)= {\bf G}_n^* $. From the definition of $ {\bf C}_n^+ $ one can see that $ {\bf C}_n^+ $, and thus also $ {\bf G}_n^\circ $, are symmetric. It follows from \eqref{eq.Co} that \bee {\rm Cov}^*\left[ \begin{pmatrix} {\rm Re}({\bf V}_n^\circ)\\ {\rm Im}({\bf V}_n^\circ) \end{pmatrix} \right] = \big({\bf G}_{n}^\circ \big)^{1/2}\big({\bf G}^\ast_n \big)^{-1/2} {\bf G}^\ast_n \big({\bf G}^\ast_n \big)^{-1/2} \big({\bf G}_{n}^\circ \big)^{1/2} = {\bf G}_{n}^\circ \stackrel{P}{\rightarrow} {\bf G}\,, \eee due to Lemma~\ref{le.BootLem2}, \eqref{equiv_complexreal} and \eqref{defG}. Again using \eqref{equiv_complexreal} yields assertion $(i)$. Since $ ({\rm Re}({\bf V}_n^*)^\top,{\rm Im}({\bf V}_n^*)^\top)^\top $ is asymptotically normal due to Lemma~\ref{le.BootLem2} $ (iii) $, $ ({\rm Re}({\bf V}_n^\circ)^\top,{\rm Im}({\bf V}_n^\circ)^\top)^\top $ is asymptotically normal with covariance matrix $ {\bf G} $ which gives assertion $ (ii) $. $ \hfill \square $ \ {\sc Proof of Theorem~\ref{th.2}:} First note that $ \sqrt{n}({\bf M}^*_n -\widehat{\bf M}_n ) $ equals $ {\bf V}^*_n $ from Step I.3. Hence, Lemma~\ref{le.BootLem2} $ (iii) $ together with \eqref{equiv_complexreal} yields \be \sqrt{n}\left[\begin{pmatrix} {\rm Re}({\bf M}^*_n)\\ {\rm Im}({\bf M}^*_n) \end{pmatrix} -\begin{pmatrix} {\rm Re}(\widehat{\bf M}_n)\\ {\rm Im}(\widehat{\bf M}_n) \end{pmatrix} \right] \stackrel{d}{\longrightarrow} \mathcal N_{2J}({\bf 0},{\bf G_1}) \label{Th39_h1} \ee in $ P $-probability, where $ {\bf G_1} $ is defined analogous to $ {\bf G} $ in \eqref{defG} but with $ {\bf \Sigma} $ replaced by $ {\bf \Sigma_1} $ and $ {\bf \Gamma} $ replaced by $ {\bf \Gamma_1} $. We next apply the delta method to \eqref{Th39_h1}. Invoking the mean value theorem for all $ 2L $ component functions of $ \tilde g $ leads to \begin{align} \begin{pmatrix} {\rm Re}({\bf W}^*_n)\\ {\rm Im}({\bf W}^*_n) \end{pmatrix} = \big( \nabla \tilde g_i(\xi_i) \big)_{i=1,\ldots,2L} \times \sqrt{n}\left[\begin{pmatrix} {\rm Re}({\bf M}^*_n)\\ {\rm Im}({\bf M}^*_n) \end{pmatrix} -\begin{pmatrix} {\rm Re}(\widehat{\bf M}_n)\\ {\rm Im}(\widehat{\bf M}_n) \end{pmatrix} \right]\,, \label{Th39_h2} \end{align} where $ \nabla \tilde g_i(\cdot) $ denotes the gradient (as a row vector) of the $ i $-th component function of $ \tilde g $, and $ \xi_i \in {\mathbb R}^{2J} $ denotes a vector on the line segment between $ ({\rm Re}({\bf M}^*_n)^\top,{\rm Im}({\bf M}^*_n)^\top)^\top $ and $ ({\rm Re}(\widehat{\bf M}_n)^\top,{\rm Im}(\widehat{\bf M}_n)^\top)^\top $. We now have from \eqref{Th39_h1} that $ \| \xi_i- ({\rm Re}(\widehat{\bf M}_n)^\top,{\rm Im}(\widehat{\bf M}_n)^\top)^\top \| \rightarrow 0 $ in $ P^* $-probability for all $ i \in \{1,\ldots,2L\} $, and Assumption \ref{assu5} together with convergence of Riemann sums yields $ \| \widehat{\bf M}_n - {\bf M} \| \stackrel{P}{\rightarrow} 0 $. Therefore \eqref{Th39_h2} and \eqref{Th39_h1} imply \begin{align} \begin{pmatrix} {\rm Re}({\bf W}^*_n)\\ {\rm Im}({\bf W}^*_n) \end{pmatrix} \stackrel{d}{\longrightarrow} {\bf J}_{\tilde g}(( {\rm Re}({\bf M})^\top,{\rm Im}({\bf M})^\top)^\top) \, \mathcal{N}_{2J}({\bf 0},{\bf G_1}) \label{Th39_h3} \end{align} in $ P $-probability. We only need the asymptotic normality from this statement. Note that $ \widetilde {\bf G}_n^* $ is the covariance matrix of the last left-hand side and therefore \begin{align} \big( \widetilde {\bf G}_n^*\big)^{-1/2} \begin{pmatrix} {\rm Re}({\bf W}^*_n)\\ {\rm Im}({\bf W}^*_n) \end{pmatrix} \stackrel{d}{\longrightarrow} \mathcal{N}_{2J}({\bf 0},{\bf Id}_{2J}) \end{align} in probability. From the proof of Theorem~\ref{th.1} we have $ {\bf G}_{n}^\circ \stackrel{P}{\rightarrow} {\bf G} $. Also, the above considerations together with continuity of $ {\bf J}_{\tilde g} $ yield $ {\bf J}_{\tilde g}(({\rm Re}(\widehat{\bf M}_n)^\top,{\rm Im}(\widehat{\bf M}_n)^\top)^\top) \stackrel{P}{\rightarrow} {\bf J}_{\tilde g}(({\rm Re}({\bf M})^\top,{\rm Im}({\bf M})^\top)^\top) $. Thus, \begin{align*} \widetilde{\bf G}^\circ_n \stackrel{P}{\longrightarrow} {\bf J}_{\tilde g}\begin{pmatrix}{\rm Re}({\bf M})\\ {\rm Im}({\bf M})\end{pmatrix} \times {\bf G} \times {\bf J}_{\tilde g}\begin{pmatrix}{\rm Re}({\bf M})\\ {\rm Im}({\bf M})\end{pmatrix}^\top\,. \end{align*} Therefore we have \begin{align} \begin{pmatrix} {\rm Re}({\bf W}^\circ_n)\\ {\rm Im}({\bf W}^\circ_n) \end{pmatrix} \stackrel{d}{\longrightarrow} {\bf J}_{\tilde g}(( {\rm Re}({\bf M})^\top,{\rm Im}({\bf M})^\top)^\top) \, \mathcal{N}_{2J}({\bf 0},{\bf G}) \end{align} in probability. Applying the continuous mapping theorem with $ h(x_1,x_2):=x_1+i\cdot x_2 $, for $ x_1,x_2 \in {\mathbb R}^{L} $, it follows \begin{align} {\bf W}^\circ_n \stackrel{d}{\longrightarrow} {\bf W} \sim \mathcal{N}^c_{L}({\bf 0},{\bf \Sigma_R},{\bf \Gamma_R}) \end{align} in probability, for the matrices $ {\bf \Sigma_R},{\bf \Gamma_R} $ stated in \eqref{CLT_smoothfun_g} (the exact form of which does not need to be specified for the desired assertion to hold). $ \hfill \square $

\section{Introduction} The interest of the investigation of meson production in nucleon-nucleon and nucleon-antinucleon collisions near threshold has increased recently since different anomalies have been detected \ct{anom},\ct{eta},\ct{nbarn},\ct{cel}. These anomalies are characterized by unexpected enhancements in the production of the mesons which contain large admixtures of strange quarks in their wave function. The large OZI rule violations near threshold challenge most of the theoretical approaches to the dynamics of strangeness production in strong interactions (see discussion in \ct{ellis}). One of this anomalous behaviors is related to the surprisingly large cross section in proton-neutron collisions near threshold for the $\eta$ production \ct{eta},\ct{cel}. It has been found recently \ct{cel} that the ratio to the same process in p-p scattering in the same kinematical regime is about $7$. An explanation of this enhancement, based on the large contribution from $N^*(1535)$ ($S_{11}$) resonance has been suggested \ct{theory}. However the final result is very sensitive to the values of the meson-$S_{11}$ coupling constants, which at present are not well known. It is therefore difficult to establish the precise contribution from the $S_{11}$ to this cross section and confirm the validity of the mechanism. In here, we present a different explanation for the enhancement, which in our case is associated with a particular microscopic mechanism of QCD. In the vacuum of QCD there exist strong fluctuations of the gauge fields called instantons (see recent review \ct{shur}). The instantons induce flavor dependent quark-quark \ct{thooft} and quark-gluon interactions \ct{koch1} whose structure is related with the flavor and helicity properties of the quark zero-modes in the presence of the instanton fields. The most famous of them is t'Hooft's quark-quark interaction \ct{thooft} which has been successfully applied to describe the properties of the QCD vacuum and hadron spectroscopy \ct{shur}. Besides this interaction, which contains $2N_f$ quark legs, the instantons can also lead to different type of the vertices which include arbitrary numbers of quark and gluon legs. The gluons in these vertices produce easily strange quarks. Therefore these new type of interactions should lead to large violations of the OZI rule in the strong interaction. One example of them, which leads to a non-zero value of the chromomagnetic moment of the quark, has been discussed recently \ct{koch1}. In Section 2 we discuss the effective quark-multi-gluon Lagrangian induced by the instantons. In Section 3 the specific contribution of the mechanism to the cross section for the production of $\eta$ mesons in proton-neutron collisions is estimated. Finally in Section 4 the main conclusions of the present work will be drawn. \section{ Effective quark-gluon vertices induced by instantons} The effective Lagrangian induced by instantons has the following form~\cite{review} \footnote{ For anti-instantons one should interchange $R\leftrightarrow L$.} \begin{eqnarray} {\cal L}_{inst}&=&\int\prod_q(m_q\rho-2\pi^2\rho^3\bar q_R(1+\frac{i}{4} \tau^aU_{aa^\prime}\bar\eta_{a^\prime\mu\nu}\sigma_{\mu\nu})q_L) \nonumber\\ &\cdot &exp({-\frac{2\pi^2}{g}\rho^2U_{bb^{\prime}} \bar\eta_{b^\prime\gamma\delta} G^b_{\gamma\delta}}) \frac{d\rho}{\rho^5}d_0(\rho)d\hat{o}, \label{e1} \end{eqnarray} where $\rho$ is the instanton size, $\tau^a$ are the Pauli matrices associated with the generators of the $SU(2)_c$ subgroup of the $SU(3)_c$ color group, $d_0(\rho)$ is the density of the instantons, $d\hat{o}$ stands for integration over the instanton orientation in color space, $\int d\hat{o}=1$, $U$ is the orientation matrix of the instanton, $\bar\eta_{a\mu\nu}$ represent t'Hooft's symbols. This Lagrangian describes the effective interaction between gluons and quarks arising from QCD through the color dipole moment of the instanton. The effective quark--gluon interaction which generates the contribution to $\eta$ production can be obtained by expanding (\ref{e1}) in powers of the gluon strength and integrating over $d\hat{o}$. The effective parameter of this expansion is the ratio of the average instanton size $\rho_c$ to the confinement radius $R_{conf}$. This ratio, in realistic models for the QCD vacuum \ct{shur}, is rather small, $\rho_c^2/R_{conf}^2\approx 0.1$, and therefore we can restrict the expansion to the lowest order. The corresponding approximation contains contributions to the production of mesons with definite quantum numbers. To second order in $G_{\gamma\delta}^b$ we found out that effective Lagrangian, which is responsible for $\eta$ production, has the following form \begin{equation} \Delta L_{\eta}=-\int{\cal L}_{t'Hooft} \frac{\pi^3\rho^4d\rho}{8\alpha_s(\rho)} G_{\mu\nu}^a \widetilde{G}_{\mu\nu}^a, \label{e2} \end{equation} where ${\cal L}_{t'Hooft}$ is t'Hooft's interaction which for massless $u$ and $d$ quarks is ${\cal L}_{t'Hooft} = $ \begin{equation} \frac{16\pi^4d_0(\rho)\rho}{9} (\bar u_{R}u_{L}\bar d_{R}d_{L}+\frac{3}{8}( \bar u_{R}t^au_{L}\bar d_{R}t^ad_{L}-\frac{3}{4}\bar u_R\sigma_{\mu\nu}t^a \bar u_L\bar d_R\sigma_{\mu\nu}t^ad_L)), \label{e44} \end{equation} This quark-gluon vertex is presented in Fig.1. \begin{figure}[htb] \centering \epsfig{file=quarkg.eps,width=7.5cm} \vskip 0.5cm \caption{\it The effective quark-gluon vertex which gives rise to the contribution in $\eta$ meson production.} \end{figure} Within the simplest version of instanton liquid model for the QCD vacuum, the size of instantons is fixed \begin{equation} d_0(\rho)\sim\delta(\rho-\rho_c). \label{e3} \end{equation} $\rho_c\sim 1.6\div 2.0 \; GeV^{-1} $ is the average size of the instantons in the QCD vacuum. This value is obtained from hadron spectroscopy \ct{shur}. However this size is not very well defined, for example, recent lattice calculations \ct{lattice} give a value for it which is larger, $\rho_c\sim 2.5\; GeV^{-1}$. We will therefore in our calculation consider a large interval for $\rho_c\sim 1.6\div 2.5 \; GeV^{-1} $. By using the low energy theorem, based on the chiral anomaly, one can calculate the following matrix element \ct{voloshin}, \ct{shifman} \begin{equation} <0|\frac{3\alpha_s}{4\pi}G_{\mu\nu}^a \widetilde{G}_{\mu\nu}^a|\eta>\approx \sqrt{3}f_\pi m_\eta^2, \la{low} \end{equation} where we have neglected $m_u$ and $m_d$ with respect to $m_s$. Under the assumptions of ref. \ct{shifman} the quark-$\eta$ meson interaction from \re{e2} becomes \begin{equation} {\cal L}_{\eta}= \frac{2\pi^6\rho_c^6f_\pi m_\eta^2}{\sqrt{3} \alpha_s^2(\rho_c)} (\bar u_{R}u_{L}\bar d_{R}d_{L}+\frac{3}{8}( \bar u_{R}t^au_{L}\bar d_{R}t^ad_{L}-\frac{3}{4}\bar u_R\sigma_{\mu\nu}t^a \bar u_L\bar d_R\sigma_{\mu\nu}t^ad_L)) \phi_\eta. \la{et1} \end{equation} where $\phi_\eta$ is the $\eta$-field. It should be mentioned that this effective interaction can only be used if the frequency of gluon fields is small. This condition however holds for the production of the particles near threshold. \section{ Contribution of the instantons to the $\eta$ production cross section near threshold} A very unique feature of the interaction \re{et1} is its strong flavor dependence. Indeed it is non zero only if quarks have different flavors. Therefore, one can expect that the contribution of it to the $\eta$ meson production in proton-neutron collision is larger then in proton-proton and neutron-neutron ones, because in the former the flavor structure favors the probability of interaction. The contribution of the instanton interaction is determined by the matrix element of interaction \re{et1} \begin{equation} <NN|{\cal L}_\eta|NN\eta>\propto<NN|{\cal L}_{t'Hooft}|NN>. \la{mat} \end{equation} Due to the strong spin-flavor dependence of t'Hooft's interaction \re{e44}, the final result is very sensitive to the spin-flavor structure of the nucleon wave function. We will use a model for hadron structure which has proved to be consistent with the instanton picture and which will favor the proposed mechanism. It is well known that a significant violation of $SU(6)$ symmetry in the nucleon wave function is required to explain some properties of nucleons: $\Delta$-nucleon mass splitting, enhancement of $u$ quark distribution with respect to $d$ quark distribution at large Bjorken $x$ region in deep-inelastic scattering, behavior of nucleon form factors, etc... (see \ct{diquark}). It was shown \ct{koch2} that the instanton induced interaction can explain the nucleon-$\Delta$ mass difference, due to the formation of a quasi-bound scalar $ud$ diquark inside nucleon. At present, the instanton induced diquark configuration is widely under discussion, not only in connection with hadron spectroscopy, but also as a possible scenario for the formation of a color superconducting state in quark-gluon matter \ct{wilczek}. Therefore we feel encouraged to use this simplifying scenario. In the quark-diquark model, proton and neutron consist of a scalar $ud$ diquark and a u-quark or d-quark, respectively. It is evident that due to specific flavor properties of the instanton vertex only the interaction between quarks with different flavors, which are not included into diquark configuration, can lead to the $\eta$ meson production in nucleon-nucleon scattering (see Fig.2) \footnote{The diquark-quark interaction mediated by one instanton vanishes, as determined by t'Hooft's mechanism, since the two quarks that participate in the interaction have the same flavor. A diquark-quark interaction, through this mechanism, arises only if mediated by at least two {\it different} instantons, and is suppresed by the smallness of the density of instantons in the QCD vacuum.}. \begin{figure}[htb] \centering \epsfig{file=diquark.eps,width=7.5cm} \vskip 0.5cm \caption{\it The instanton contribution to the $\eta$ meson production in proton-neutron interaction} \end{figure} Therefore the instanton induced $\eta$ meson production in proton-proton interaction within pure quark-$ud$-diquark model should be very small in comparison with proton-neutron scattering \footnote{ If we neglect the production of the $\eta$ meson in the diquark-quark interaction and contributions from other mechanisms ($S_{11}$, ...) then the ratio of cross sections becomes $\sigma_{pp}/\sigma_{pn}\approx 0$.}. The experimental observation of a sizeable enhancement of the $\eta$ production in proton-neutron scattering, gives arguments in the favor of a large diquark component in the nucleon wave function, however the proposed dynamical mechanism favors the enhancement in other scenarios, e.g., bag model or naive quark model, as well \footnote{In the case of the pure $SU(6)$ nucleon wave function one can expect the ratio $\sigma_{pp}/\sigma_{pn}\approx 1/4$, based on the result of \ct{farrar} where the single instanton contribution to the ratio of amplitudes of elastic pp and pn scattering has been obtained, $\Phi(pp)/\Phi(pn)=-8/17$.}. Let us to estimate the cross section for $\eta$-meson production in proton-neutron collisions. We follow the notation: \begin{equation} p(p_1)+n(p_2)\rightarrow p(p_1^\prime)+n(p_2^\prime)+\eta(p_\eta). \label{reaction} \end{equation} The cross section is given by \begin{equation} d\sigma=\frac{dPS^3}{4\sqrt{(p_1.p_2)^2-m_p^4}} \sum_{spin}|M|^2, \la{cross} \end{equation} where \begin{equation} dPS^3=\frac{d^3p_1^\prime d^3p_2^\prime d^3p_\eta}{(2\pi)^92E_1^\prime 2E_2^\prime 2E_\eta}(2\pi)^4 \delta(p_1+p_2-p_1^\prime-p_2^\prime-p_\eta), \la{ps} \end{equation} is the 3-body phase space volume. $M$ is the matrix element of the interaction \re{mat} and $\sum_{spin}$ represents the spin summation over the final and spin averaging over the initial nucleon states. Only the colorless part of the instanton induced interaction can contribute to the cross section of the reaction \re{reaction}. Within the quark-diquark model one can factorize the matrix element and obtains using a nonrelativistic approximation for the quark wave functions of the nucleon, $M\sim <PN|\bar u_Ru_L\bar d_Rd_L|PN>\sim <P|\bar u_Ru_L|P><N|\bar d_Rd_L|N>$ \begin{equation} \approx\bar P_RP_L\bar N_RN_L. \label{mat2} \end{equation} The final result for the matrix element is \footnote{We took into account in \re{cross2} the additional factor two related to the contribution of anti-instantons.} \begin{equation} \sum_{spin}|M|^2=\frac{\pi^{12}\rho_c^{12}f_\pi^2m_\eta^4}{54\alpha_s^4 (\rho_c)} (8m_p^4-2m_p^2(t_1+t_2)+t_1t_2)F_N(t_1)^2F_N(t_2)^2, \label{cross2} \end{equation} where \begin{eqnarray} t_1&=&(p_1-p_1^\prime)^2\nonumber\\ t_2&=&(p_2-p_2^\prime)^2, \nonumber \end{eqnarray} and $F_N(t)$ is the strong form factor of the nucleon for which we will use the electromagnetic form factor \ct{dl} \begin{equation} F_N(t)=\frac{4m_p^2-2.79t}{(4m_p^2-t)(1-t/0.71)^2}. \la{ff1} \end{equation} In the numerical calculation we use the NLO approximation for the strong coupling constant \begin{equation} \alpha_s(\rho)=-\frac{2\pi}{\beta_1t}(1+\frac{2\beta_2log t}{\beta_1t}) , \label{e17} \end{equation} where \begin{equation} \beta_1=-\frac{33-2N_f}{6},\ \beta_2=-\frac{153-19N_f} {12}\nonumber \end{equation} and \begin{equation} t=log(\frac{1}{\rho_0^2\Lambda^2}+\delta). \label{e18} \end{equation} In Equation (\ref{e18}), the parameter, $\delta\approx1/\rho_0^2\Lambda^2$, provides a smooth interpolation of the value of $\alpha_s(\rho)$ from the perturbative $(\rho\rightarrow0)$ to the nonperturbative regime $(\rho\rightarrow\infty)$~\cite{a8}, and for $N_f=3$ we have used $\Lambda=230\; MeV$, and $\rho_0=1.6 \; GeV^{-1}$. The exact phase space integration of \re{cross} gives the total cross section of the reaction \re{reaction} as a function of the excess energy in c.m. system $Q=\sqrt{S}-m_n-m_p-m_\eta$, $S=(p_1+p_2)^2$, and is presented in Fig.3 for the different values of the instanton size $\rho_c$. This dependence is in qualitative agreement with experimental data of WASA/PROMICE Collaboration at CELSIUS \ct{cel}. It should be mentioned that the instanton induced cross section of the $\eta$ production in $pn$ scattering is very sensitive to the average size of instanton in the QCD vacuum and for $\rho_c=2.1 \; GeV^{-1}$ it describes the experimental data. \begin{figure}[htb] \centering \vskip -.5cm \centering \epsfig{file=etaf.eps,width=7.5cm} \caption{\it $Q$-dependence of $\eta$ production cross section for the different values of the average instanton size in comparison with experimental data of WASA/PROMICE Collaboration.} \end{figure} Before ending this section it is worth stressing the main difference between our approach and the hadronic approach. In the latter quantum numbers and energy considerations determine the intermediate meson, the $S_{11}$, whose coupling constant is at present unknown and whose properties, mass and width, have to be feed in from experiments. In our approach the calculation is basically parameter free, despite the fact that we have allowed some freedom into the size parameter reflecting the crudeness of the approximations involved. An independent determination of the $S_{11}$ coupling constant would clarify the relation between the nonperturbative microscoscopic mechanism presented here and the hadronic approach. The conclusions drawn thus far are only slightly affected by final state interactions \ct{fsi} due to their structure and magnitude compared with that of the of the enhancement. \section{Conclusion} We have shown that a novel mechanism for $\eta$ production in nucleon-nucleon scattering might be behind the anomalous enhancement found for proton-neutron collisions. The proposed mechanism is related with the presence of strong nonperturbative fluctuations of gluons fields in the QCD vacuum, called instantons. It has been shown that the instanton induced quark-gluon interaction, due specially to its singular flavor dependence, is able to explain the large $\eta$-meson production cross section near threshold in proton-neutron scattering. The calculation has been realized within a simplified scenario of baryon structure, the diquark description, which maximizes the flavor dependence of the interaction. However it is immediate that other models will not avoid the basic mechanism, whose origin is solely in the structure of the instanton interaction. The final result for the cross section of the $\eta$ production is very sensitive to the value of the instanton size. This observation opens the door to investigate the parameters of the QCD vacuum in the threshold production of $\eta$ (and other mesons) with high accuracy. Naturally our proposed mechanism appears in nature entangled with other ones. Their experimental separation is very important. A possible way is the determination of the energy and angular dependence of the production processes. Due to the point-like origin of the instanton-induced interaction, its dependences are totally different from those of other mechanisms. For example in $\eta$-meson production the competing mechanism is the production of an intermediate $S_{11}$-resonance and certainly its energy and angular dependence are very different. A very clean way to disentangle the various mechanisms is the measurement of the {\it spin} dependence of the cross section in the production processes by using polarized nucleon beams and targets. In this case the instantons give a large contribution to the double spin asymmetries due to very strong helicity dependence of t'Hooft's interaction \re{et1}. \section*{Acknowledgements} One of us (V.V.) thanks L. Alvarez-Ruso for tuition in the subject of final state interactions and for useful comments. This work was partially supported by DGICYT grant \# PB97-1227.

\section{\textbf{INTRODUCTION}} \label{section:intro} Hypothesis testing is one of the most relevant tasks in empirical work. Such tests include situations when the null and alternative hypothesis are assumed to belong to a parametric family of models. In a second type of tests, known as diagnostic or lack-of-fit tests, only the null hypothesis is assumed to belong to a parametric family leaving the alternative nonparametric. The latter type of testing has a distinguished and long literature starting with the work of Kolmogorov, see Stephens \cite{Stephens}, for testing the probability distribution function and in a time series context by Grenander and Rosenblatt \cite{GrenanderRos} for testing the white noise hypothesis. In a regression model context a new avenue of work started in Stute \cite{Stute97}, and Andrews \cite{Andrews97} with a more econometric emphasis, using partial sums empirical methodology, see also Stute et al. \cite{Stute_etal98b} or Koul and Stute \cite{KoulStute99}, among others. The methodology has attracted a lot of attention and it rivals tests based on a direct comparison between a parametric and a nonparametric fit to the regression model as first examined in H\"{a}rdle and Mammen \cite{HardleMammen} or Hong and White \cite{HongWhite}. One advantage of tests based on partial sums empirical methodology, when compared to the approach in \cite{HardleMammen}, is that the former does not require the choice of a bandwidth parameter for its implementation, see also Nikabadze and Stute \cite{NikabadzeStute} for some additional advantages. However, a possible drawback is that the asymptotic distribution depends, among other possible features, on the estimator of the parameters of the model under the null hypothesis in a nontrivial way, as it was shown in Durbin \cite{Durbin73}, and hence its implementation requires either bootstrap algorithms, see Stute et al. \cite{Stute_etal98a}, or the so-called Khmaladze's \cite{Khmaladze81} martingale transformation, see also earlier work and ideas in Brown at al. \cite{BrownDurbinEvans}. In this paper, though, we are interested in a third type of testing where neither the null hypothesis nor the alternative has a specific parametric form. This type of hypothesis testing can be denoted as testing for \emph{% qualitative }or \emph{shape} restrictions. Many shape properties (including monotonicity, convexity/concavity, strong convexity, log-convexity) are widespread in economics and other disciplines. It is also often of interest to analyse statistical and economic relationships that do not have a persistent shape pattern on the whole domain but rather switch the patterns in the domain (for example, \emph{U-shaped} or \emph{S-shaped} relations). To fix ideas, consider the nonparametric regression model% \begin{align} y_{i}& =m\left( x_{i}\right) +u_{i}\text{,} \label{1_1} \\ E[u_{i}|x_{i}]& =0, \notag \end{align}% where $x$ has bounded support $\mathcal{X=}:\left[ \underline{x},\overline{x}% \right] $ and $m\left( \cdot \right) $ is smooth. More specific conditions on the sequences $\left\{ u_{i}\right\} _{i\in \mathbb{Z}}$ and $\left\{ x_{i}\right\} _{i\in \mathbb{Z}}$ will be given in Condition $C1$ in Section % \ref{section:1step}. Our main aim is testing whether the regression function $m\left( x\right) $ possesses shape properties captured by a general null hypothesis \begin{equation} H_{0}:\text{ \ }m\in \mathcal{M}_{0}, \label{main_null} \end{equation}% where the class of interest $\mathcal{M}_{0}$ is a subset of smooth functions from $\mathcal{X}$ to $\mathbb{R}$. Our requirement on the class $% \mathcal{M}_{0}$ is given in Condition $C0$ below, which is formulated in a way to directly relate it to the methodology we employ later to approximate and estimate $m(\cdot )$. Before we introduce Condition $C0$, let us introduce some useful notations. Let $\mathcal{B}_{q,L}$ denote the set of all \emph{B-splines} of degree $q$ with knots that split $\mathcal{X}$ into $L^{\prime }$ equally spaced intervals.\footnote{% As it will be clear from our exposition later, the condition that these intervals are equally spaced is not important and is only imposed for the simplicity of the exposition. We only need that the system of knots has to become increasingly dense in $\mathcal{X}$. Also, in Section \ref{section:BPsplines} we shall motivate the choice of \emph{B-splines } in comparison to other nonparametric estimation techniques.} A generic element in this set is written as a linear combination $m_{\mathcal{B}}(x)\equiv \sum_{\ell =1}^{L}\beta _{\ell }p_{\ell ,L}\left( x\right) $, where $L=L^{\prime }+q$, and $\left\{ p_{\ell ,L}\left( \cdot \right) \right\} $ is the collection of the \emph{B-splines} base for the chosen system of knots (more details will be given in Section \ref{section:BPsplines}). Any element in $\mathcal{B}_{q,L}$ can be fully characterised by the vector $(\beta _{1},\ldots ,\beta _{L})\in \mathbb{R}% ^{L}$ and constraints on this vector can be mapped into constraints on the B-spline. For any set $S_{q,L}\subseteq \mathbb{R}^{L}$ we can, thus, define \begin{equation*} \mathcal{M}_{s_{q,L}}=:\left\{ m_{B}(\cdot )\in \mathcal{B}_{q,L}\;|\;(\beta _{1},\ldots ,\beta _{L})\in S_{q,L}\right\} . \end{equation*} \vskip 0.1in \noindent \emph{\textbf{Condition C0}. There is a set $S_{q,L} \subseteq \mathbb{R}^{L} $ for any $L'$ (and, hence, for any $L=L'+q$) that satisfies the following properties: \begin{itemize} \item[(a)] $S_{q,L}$ does not depend on data $\left\{ x_{i}\right\} _{i\in \mathbb{Z}}$ and, thus, is non-stochastic; \item[(b)] the boundary of $S_{q,L}$ consists of a finite number of smooth surfaces; \item[(c)] \begin{equation} \label{GT2} \mathcal{H}\left( \mathcal{M}_0,\mathcal{M}_{S_{q,L}}\right) \rightarrow 0\quad \text{as}\quad L\rightarrow \infty \text{,} \end{equation} where $\mathcal{H}$ is the Hausdorff distance in the supremum norm in the space of continuous functions from $\mathcal{X}$ to $\mathbb{R}$. \end{itemize}} \vskip 0.1in Condition $C0$ essentially states that the membership in $\mathcal{M}_0$ can be captured by restrictions on parameters $\{\beta_{\ell}\}_{\ell=1}^L$ which become necessary and sufficient restrictions as the system of knots becomes increasingly dense in $\mathcal{X}$. In addition, these restrictions on $\{\beta_{\ell}\}_{\ell=1}^L$ do not depend on the available data, which adds to the attractiveness of our approach for implementation purposes. Our idea will be to test the null hypothesis \begin{equation} H_{0}^{B}:\text{ \ }(\beta _{1},\ldots ,\beta _{L})\in S_{q,L} \label{main_nullB} \end{equation}% formulated in terms of the approximation for $m(\cdot )$. As may be expected, our methodology readily extends to situations when the domain $% \mathcal{X}$ is partitioned into several intervals with different null hypotheses in the spirit of (\ref{main_null}) formulated on different intervals in the partitioning. An illustration of this situation is given in Example \ref{example:deriv2} below. We shall now illustrate several examples of shape classes $\mathcal{M}_{0}$ that satisfy condition $C0$. In Section \ref{section:BPsplines}, after a more thorough discussion of B-splines we indicate the corresponding sets $% S_{q,L}$ for each of these examples. As we shall discuss in Section \ref{section:BPsplines}, the first two examples pertain to the case where in the constraints on the coefficients of the B-splines approximation are linear\footnote{Quasi-convexity in Example \ref{example:deriv2} may be an exception depending on how exactly one approaches the testing, as is clear from further discussion in Section \ref{section:BPsplines}.}, whereas in the last two examples these constraints are nonlinear. It is important to note that these examples are an illustration of the scope of our approach rather than a full list of properties that our methodology could test. \label{section:BPsplines} \begin{example}[Constraints, may hold simultaneously, on the derivatives of $% m\left( \cdot \right) $] \label{example:deriv1} \begin{equation} H_{0}:d_{r}\cdot m^{\left( r\right) }\left( x\right) \geq c_{r}\text{,}~\ \ \ \ r\in R\text{,} \label{eq:example1} \end{equation}% where $R$ is a finite subset of $\mathbb{N}^{+}$, $d_{r}\in \left\{ -1,1\right\} $ and $c_{r}$ are known constants. In this hypothesis, we specify conditions on inequalities for several derivatives simultaneously. If $c_{r}=0$, then we have familiar restrictions on the sign of the $r$-th derivative. Special cases in this example include testing for \textbf{monotonicity} ($r=1$ and $c_{1}=0$), testing for \textbf{% convexity/concavity} ($r=2$ and $c_{2}=0$), testing for \textbf{strong $% \lambda $-convexity} ($r=2$ and $c_{2}=\lambda >0$), testing for \textbf{% monotonicity and concavity simultaneously}, etc. \end{example} The properties of monotonicity, convexity/concavity or the ones in terms of higher-order derivatives are so commonplace in economics and other disciplines that the applications are too many to list. For instance, a demand function is expected to be a decreasing function of the price of a good, whereas a supply function is expected to be increasing. In single-object auctions, the equilibria analyzed commonly are those in which buyers play monotone strategy functions and mark-ups are often monotone functions of the bids (see e.g. \cite{Krishna}). In other economic relationships it is often of importance whether the marginal returns are increasing or decreasing, which naturally amounts to convexity and concavity, respectively. Note that in the context of isotone/monotone regressions the literature goes back to Brunk \cite{Brunk} and Wright \cite{Wright}, and in the context of convex regressions to Hildreth \cite{Hildreth}. A more detailed coverage of this literature is given in Section \ref{sec:methodology}. As mentioned above, our methodology can be used when different properties can be conjectured on different intervals of the partition $\mathcal{X}$. This is illustrated in the next example. \begin{example}[changing shape patterns on the domain; symmetry; quasi-convexity] \label{example:deriv2} In this case, we can take a partitioning of $\mathcal{% X}$ at points $\tau _{0}=\underline{x}<\tau _{1}<\ldots <\tau _{J-1}<\tau _{J}=\overline{x}$, and formulate the null hypothesis on the signs of various derivatives on the intervals in this partitioning. The null hypothesis in this case would be \begin{align} H_{0}& :d_{r_{j}}\cdot m^{\left( r_{j}\right) }\left( x\right) \geq 0,\quad x\in \left( \tau _{j},\tau _{j+1}\right] , \label{eq:example2} \\ & \text{ for some fixed }r_{j}\in \mathbb{N},\quad j=0,\ldots ,J-1\text{,} \notag \end{align}% $d_{r_{j}}\in \left\{ -1,1\right\} $.\footnote{% It goes without saying that we can have several constraints on each $[\tau _{j},\tau _{j+1}],\;j=0,\ldots ,J-1$.} The partitioning points $\tau _{j}$ either have to be known or have to be consistently estimated at a suitable rate before the testing procedure. This type of hypotheses include, among others, the important scenarios of $U$% \emph{-shape}, $S$\emph{-shape}. For example, \emph{U-shape} is the property of a function first decreasing and then increasing. So, we write the null hypothesis of \emph{U-shape} with the switch at $s_{0}$ as \begin{equation} H_{0}:\text{ \ }\left\{ \begin{array}{c} m\left( x^{1}\right) >m\left( x^{2}\right) \quad \text{ when }\quad x^{1}<x^{2}\leq s_{0} \\ m\left( x^{1}\right) >m\left( x^{2}\right) \quad \text{ when }\quad x^{1}>x^{2}\geq s_{0}.% \end{array}% \right. \label{monotone_3} \end{equation} The issues associated with testing properties outlined in this example relate to how functional pieces from different partition intervals are connected to each other. We can incorporate various scenarios -- e.g., we can require that different pieces are joint continuously, or smoothly at all or some partition points. It goes without saying that we can consider several different partitions and test simultaneously properties formulated on various intervals of these partitions. For example, one could consider a function defined on $[0,1]$ and conjecture that $m(\cdot)$ is convex on $(0,s_0$, concave on $(s_0,1]$, decreasing on $(0,0.3s_0]$, increasing on $(0.3s_0, 0.5s_0+0.5]$, and decreasing on $(0.5s_0, 1]$. Examples of \emph{U-shaped} relationships in economics and other disciplines can be found e.g. in \cite{CalabreseBaldwin}, \cite{Corrao_etal}, \cite{Goldin}, \cite{Groes_etal}. \cite{SuttonTrefler}, \cite{Weiman}, and in the discussions in \cite{Kostyshak}, \cite{LindMelhum} and \cite{Simonsohn}. Inverse \emph{U-shaped} relationships include the case of the so-called single-peaked preferences, which is an important class of preferences in psychology and economics. Examples of \emph{S-shaped} relationships between unexpected earnings and stock price in accounting can be found in \cite{DasLev} or \cite{FreemanTse}, among others. \emph{% S-shaped} growth curves of the adopted population in a large society is a generally accepted empirical feature of innovation diffusion (see discussions in \cite{Rogers} or \cite{Utterback}). Thus, testing for an \emph{S-shape} in this case would allow one to conclude whether technology evolves as one would expect. Another shape property in this class of shape-changing patterns are the properties of \textbf{% quasi-convexity} and \textbf{% quasi-concavity}. Formally, the class $\mathcal{M}_{0}$ in (\ref{main_null}) of quasi-convex functions is defined as $$ \mathcal{M}_{0}=\bigg\{ \phi (\cdot ):\;\forall \,x_{1},x_{2}\in \mathcal{X}% \; \; \forall \lambda \in \lbrack 0,1] \; \; \phi (\lambda x_{1}+(1-\lambda )x_{2})\leq \max {\big \{}\phi (x_{1}),\phi (x_{2}){\big \}} \bigg\}. $$ Since function $\phi $ is \textbf{quasi-concave} if and only if $-\phi $ is quasi-convex, we can easily write a description of the class of quasi-concave functions. A smooth function is quasi-convex if and only if it first decreases up to some point and then increases (giving a special case of monotonicity when such a switch point is located at one of the boundary points of the interval). Since this switch point may not be known a priori, it would have to be estimated. A further discussion of this is given later in Section \ref{section:BPsplines}. The concept of quasi-convexity/quasi-concavity is quite important in economics and other disciplines. For example, it is known that indirect utility functions are quasi-convex provided that the direct utility function is continuous. Thus strict quasi-concavity (and continuity) of the production function provides a sufficient condition for the differentiability of the cost function with respect to input prices. Finally, we mention another property related to this framework, which is the \textbf{symmetry} of a function around some point $s_0$. Without a loss of generality, we can take $s_0$ to be the center of the domain $[\underline{x}, \overline{x}]$\footnote{For a given $s_0$, one can only test for symmetry only on the interval $[s_0 - \min\{\overline{x}-s_0, s_0-\underline{x}\}, s_0+\min\{\overline{x}-s_0, s_0-\underline{x}\}]$, which is centered around $s_0$.} and write the class of symmetric functions as \begin{equation*} \mathcal{M}_{0}=\left\{ \phi (\cdot ):\;\forall \,x \in [\underline{x},s_0] \quad \phi (s_0-x)=\phi (s_0+x) \right\}. \end{equation*} \end{example} \begin{example}[$r$-convexity; $\protect\rho $-convexity] \label{example:rconvex} The null hypothesis of $\mathbf{r}$-\textbf{convexity}, $r \neq 0$, is described by the following $\mathcal{M}_0$; \begin{multline*} \mathcal{M}_0 = \bigg\{ \phi(\cdot): \; \phi(\cdot) \geq 0, \quad \forall \, x_1, x_2 \in \mathcal{X} \quad \forall \lambda \in [0,1] \\ \phi(\lambda x_1+(1-\lambda )x_2)\leq \log \left(\lambda e^{r\phi(x_1)}+(1-\lambda) e^{r\phi(x_2)}\right)^{\frac{1}{r}} \bigg\}, \end{multline*} and for $r=0$ this property is described by \begin{multline*} \mathcal{M}_0 = \bigg\{\phi(\cdot): \; \phi(\cdot) \geq 0, \quad \forall \, x_1, x_2 \in \mathcal{X} \quad \forall \lambda \in [0,1] \\ \phi(\lambda x_1+(1-\lambda )x_2)\leq \lambda \phi(x_1) + (1-\lambda) \phi(x_2) \bigg\}. \end{multline*} The definition of $\mathbf{r}$-\textbf{concave} functions would reverse the inequalities. If we consider functions that are twice continuously differentiable, then the class of $r$-convex functions can be described as (see e.g. Avriel \cite{Avriel72}) \begin{equation*} \mathcal{M}_0 = \bigg\{\phi(\cdot): \; \phi(\cdot) \geq 0, \quad \forall \, x \in \mathcal{X} \quad r \left(\phi^{\prime }(x) \right)^2 + \phi^{{\prime \prime }}(x) \geq 0 \bigg\}. \end{equation*} A concept closely related to $r$-convexity/concavity is what is known as $% \mathbf{\rho}$-convexity/concavity in the economics literature. Following \cite{CaplinNalebuff91a}, we can define the class of $\rho$-convex functions as follows: for $\rho \neq 0$, \begin{multline*} \mathcal{M}_0 = \bigg\{ \phi(\cdot): \; \phi(\cdot) > 0, \quad \forall \, x_1, x_2 \in \mathcal{X} \quad \forall \lambda \in [0,1] \\ \phi(\lambda x_1+(1-\lambda )x_2)\leq \left(\lambda \phi(x_1)^{\rho}+(1-\lambda) \phi(x_2)^{\rho}\right)^{\frac{1}{\rho}} \bigg\}% , \end{multline*} and for $\rho=0$, \begin{multline*} \mathcal{M}_0 = \bigg\{ \phi(\cdot): \; \phi(\cdot) > 0, \quad \forall \, x_1, x_2 \in \mathcal{X} \quad \forall \lambda \in [0,1] \\ \phi(\lambda x_1+(1-\lambda )x_2)\leq \phi(x_1)^{\lambda} \phi(x_2)^{1-\lambda}\bigg\}. \end{multline*} The definition of $\mathbf{\rho}$-\textbf{concave} functions would reverse the inequalities. It is obvious that for any $r=\rho$, a real-valued function $\phi$ is $r$-convex if and only if $e^{\phi}$ is $\rho$-convex. It is also clear that $\rho$-convexity gives the standard definition of convexity when $\rho=1$ and gives log convexity for $\rho=0$. In economics, $\rho $-concavity has been used to measure the curvature of demand functions, production functions, and distributions (i.e., density and distribution functions) of individual characteristics. It has also been employed in the areas of imperfect competition, auctions and mechanism design and public economics.\footnote{% See, for instance, \cite{AndersonRenault03}, \cite{CaplinNalebuff91b}, \cite{Cowan}, \cite{Ewerhart}, \cite{Moyes}, among others.} Log-convexity and log-concavity are of particular interest in economics. \end{example} Our final example illustrates further shape properties well developed in the mathematical literature. This example first requires defining a mean function. \begin{definition}[mean function; Niculescu \cite{Niculescu}] A function $N: \mathbb{R}^{+} \times \mathbb{R}^{+} \rightarrow \mathbb{R}^{+}$ is called a \textit{mean function} if \begin{itemize} \item $N(x_1, x_2) =N(x_2, x_1)$ \item $N(x, x) =x$ \item $x_1<N(x_1, x_2) < x_2$ whenever $x_1<x_2$ \item $N(ax_1, ax_2) =aN(x_1, x_2)$ for all $a>0$ \end{itemize} \end{definition} Examples of mean functions include the arithmetic mean (A), the geometric mean (G), the harmonic mean (H), the logarithmic mean and the identric mean. \begin{example}[$MN$-convexity] \label{example:meanconvex} For any two mean functions $M$ and $N$, the class of $\mathbf{MN}$-\textbf{% convex} is defined as \begin{equation*} \mathcal{M}_0 = \bigg\{ \phi(\cdot): \; \phi(\cdot) > 0, \quad \forall \, x_1, x_2 \in \mathcal{X} \quad \phi(M(x_1, x_2)) \leq N(\phi(x_1), \phi(x_2))% \bigg\}. \end{equation*} The definition of $\mathbf{MN}$-\textbf{concavity} would reverse the inequalities. When we have different combinations of arithmetic (A), geometric (G) and harmonic (H) means, we end up with the following special cases of ${MN}$% -convex functions (see e.g. \cite{Anderson_etal}): \begin{enumerate} \item $m$ is AG-convex if and only if $\frac{m^{\prime }(x)}{m(x)}$ is increasing \item $m$ is AH-convex if and only if $\frac{m^{\prime }(x)}{m^2(x)% }$ is increasing \item $m$ is GA-convex if and only if $x m^{\prime }(x)$ is increasing \item $m$ is GG-convex if and only if $\frac{x m^{\prime }(x)}{m(x)% }$ is increasing \item $m$ is GH-convex if and only if $\frac{x m^{\prime }(x)}{% m^2(x)}$ is increasing \item $m$ is HA-convex if and only if $x^2 m^{\prime }(x)$ is increasing \item $m$ is HG-convex if and only if $\frac{x^2 m^{\prime }(x)}{% m(x)}$ is increasing \item $m$ is HH-convex if and only if $\frac{x^{2}m^{\prime}(x)}{% m^{2}(x)}$ is increasing. \end{enumerate} \end{example} \vskip 0.1in We want to emphasize again that the examples given in this introduction are meant to illustrate the scope of applicability of our testing methodology rather than give an exhaustive list of potential application. \vskip 0.1in \subsection{\textbf{LITERATURE REVIEW ON TESTING FOR SHAPES}} $\left. {}\right. $ There is a range of the literature related to testing for monotonicity or convexity such as \cite{Bowman_etal}, Hall and Heckman \cite{HallHeckman}, Ghosal et al. \cite{GhosalSenVV}, Juditsky and Nemirovski \cite{JudNem}, Wang and Meyer \cite{WangMeyer}, Chetverikov \cite{Chetverikov}, Schlee \cite{Schlee}, Diack and Thomas-Agnan \cite{DiackTA}, Dumbgen and Spokoiny \cite{DumbgenSpok}, Abrevaya and Jiang \cite{AbrevayaJiang}, Baraud et al. \cite{Baraud_etal}, Dette et al. \cite{DetteHN}. Some of the work referenced above (such as \cite{Baraud_etal}, \cite{DumbgenSpok}, \cite{JudNem}) focuses on the regression function in the ideal Gaussian white noise model, while our framework does not require such assumptions. In the aforementioned literature some papers, such as \cite{Baraud_etal}, \cite{DiackTA} or \cite{HallHeckman}, allow the explanatory variable to take only deterministic values. \cite{AbrevayaJiang}, \cite{Chetverikov} and \cite{GhosalSenVV} treat the explanatory variable as random but either require its full stochastic independence with the unobserved regression error (\cite{Chetverikov}, \cite{GhosalSenVV}) or require the distribution of the error to be symmetric conditional on the explanatory variable, see \cite{AbrevayaJiang}, whereas we require a weaker mean independence of error condition. Some of the above-mentioned papers, such as \cite{Bowman_etal}, \cite{HallHeckman} or \cite{WangMeyer}, do not give any asymptotic theory. Many of the approaches are tailored to a specific type of shape and their extensions to more general shape properties do not appear straightforward, if at all possible. Moreover, these tests are often targeted to detecting specific deviations from the null hypothesis. The violations of the null hypothesis, however, can come from different sources. A parametric test for \emph{% U-shapes} was suggested in \cite{LindMelhum}, whereas \cite{Kostyshak} proposes a non-parametric test based on critical bandwidth giving sufficient conditions for its consistency. To the best of our knowledge, there is no formal approach in the literature to testing \emph{S-shapes} or general shape properties formulated on intervals in a partitioning. Also, to the best of our knowledge, this paper is the first to give a unified framework and a nonparametric methodology for testing (possibly simultaneously) a wide range of shape/qualitative constraints. Even for shape properties described in terms of the signs of the derivatives of the function, this methodology will, in particular, overcome some of the problems discussed above. But, of course, we want to make it clear that its applicability goes beyond such properties. To be more specific, we propose a test based on a transformation, introduced in Khmaladze \cite{Khmaladze81}, of the partial sums empirical process similar to that in Stute \cite{Stute97}. Some of the properties of the test is that it converges to a standard Brownian motion, so that critical values of standard functionals such as Kolmogorov-Smirnov, Cram\'{e}r -von-Mises or Anderson-Darling are readily obtained. As a consequence, our testing procedure has the same asymptotic distribution regardless of the shape constraints under consideration. Another feature of our testing procedure is its flexibility as it is able to test simultaneously for more than one shape constraint -- for instance, one could test monotonicity and log-convexity simultaneously. Finally, the test is easy to implement as it requires no more than \textquotedblleft recursive\textquotedblright\ least squares. The remainder of the paper is organized as follows. Section \ref{sec:methodology} introduces and motivates our nonparametric estimator of $% m\left( x\right) $ and it compares the methodology against rival nonparametric estimators. It also describes the form that the sets $S_{q,L}$ in Condition $C0$ take for the examples given in the introduction. In Section \ref{sec:Khmaladze}, we describe the test and examine its statistical properties. Because the Monte-Carlo experiment suggests that the asymptotic critical values are not a good approximation of the finite sample ones, Section \ref{sec:bootstrap} introduces a bootstrap algorithm. Section \ref{sec:simulations_and_data} presents a Monte-Carlo experiment and some empirical examples, whereas Section \ref{sec:conclusion} concludes with a summary and possible extensions of the methodology. The proofs are confined to the Appendix A which employs a series of lemmas given in Appendix B. \section{\textbf{METHODOLOGY AND REGULARITY CONDITIONS}} \label{sec:methodology} Before we present the procedure for testing the hypothesis given in (\ref{main_null}), we introduce and motivate the type of nonparametric estimator that we shall use to estimate the regression function $m\left( x\right) $ under $H_{0}$ and/or $H_{1}$. When testing for the null hypothesis of either monotonicity or convexity, several nonparameteric estimators have been considered in the literature. As mentioned in the introduction, early work on isotone/monotone regressions is Brunk \cite{Brunk} and Wright \cite{Wright}. Friedman and Tibshirani \cite{FT} combines local averaging and isotonic regression, Mukerjee \cite{Mukerjee} and Mammen \cite{Mammen91a} propose a two-step approach, which involves smoothing the data by using kernel regression estimators in the first step, and then deals with the isotonization of the estimator by projecting it into the space of all isotonic functions in the second step. Hall and Huang \cite{HallHuang} proposes an alternative method based on tilting estimation which preserves the optimal properties of the kernel regression estimator. Finally, a different approach to isotonization is based on rearrangement methods using a probability integral transformation, see Dette et al. \cite{DetteNP} or Chernozhukov et al. \cite{ChernozhukovFVG}. When the null hypothesis is that of convexity, Hildreth \cite{Hildreth} approach is based on estimating $m\left( \cdot \right) $ by least squares approach. Consistency, rate of convergence properties and pointwise convergence results for this estimator are established, respectively, in \cite{HansonPledger}, \cite{Mammen91b} and \cite{GroeneboomJW}. The global behaviour of such an estimator is examined in \cite{GuntuboyinaSen}. Birke and Dette \cite{BirkeDette} examine an estimator based on first obtaining unconstrained estimate of the derivative of the regression function which is isotonized and then integrated. Our main aim is to propose a testing methodology which is not only applicable for a wide range of shape properties but also able to perform valid statistical inferences. The narrowness of the above mentioned techniques would make it extremely difficult to make them work in our more general context. Besides, the implementation of these techniques is not always trivial and/or they often lack any asymptotic theory useful for the purpose of inference. A different approach might be based on series estimation using polynomial and, in particular, \emph{Bernstein} polynomials basis, see the survey by Chen \cite{ChenHandbook}, for results on sieve estimation in both nonparametric and semiparametric models. One motivation for the Bernstein polynomials is due to a well-known result that any continuous function can be approximated in the sup-norm by a Bernstein polynomial whose coefficients are the values of the function at uniform knot points. This property makes them an appealing candidate estimation method in our context. However, base Bernstein polynomials have an undesirable property of being highly correlated, making them difficult for our purposes -- in particular, to obtain a valid Khmaladze's transformation, which plays a key role in our results. This is discussed in more detail in Section \ref{section:BPsplines}. Due to many of the aforementioned reasons, in this paper we shall employ \emph{B-splines} and/or penalized \emph{B-splines} known as \emph{P-splines}. Our testing will rely on the ability to write our null hypothesis in terms of the coefficients of the \emph{B-spline} approximation, leading us to utilizing Condition $C0$ and testing (\ref{main_nullB}). Note that monotone regression splines (closely related to \emph{B-splines}) were introduced by Ramsay \cite{Ramsay} and later extended by Meyer \cite{Meyer} to estimate convex/concave function or functions that are, e.g., both monotone and convex. A distinct feature in our approach is, first, that we now take the use of \emph{B-splines} to a different level of generality and, second, that we derive the asymptotic theory as the number of knots goes to infinity, and, third, that we allow the number of constraints on the coefficients of \emph{B-splines} to increase with the number of knots. Both \cite{Ramsay} and \cite{Meyer} consider the number of knots, and hence the number of constraints, to be fixed. \subsection{\textbf{B-SPLINES (or P-SPLINES)}} \label{section:BPsplines} $\left. {}\right. $ \emph{B-splines} or \emph{P-splines} are constructed from polynomial pieces joined at some specific points denoted knots. Their computation is obtained recursively, see \cite{deBoor}, for any degree of the polynomial. It is well understood that the choice of the number of knots determines the trade-off between overfitting when there are too many knots, and underfitting when there are too few knots. The main difference between \emph{B-splines} and \emph{P-splines} is that the latter tend to employ a large number of knots but to avoid oversmoothing they incorporate a penalty function based on the second difference of the coefficients of adjacent \emph{B-splines}, in contrast to the second derivative employed in \cite{OSullivan86}, \cite{OSullivan88}, see \cite{EilersMarx} for details. The methodology and applications of constrained \emph{B-splines} and \emph{% P-splines} are discussed by many authors, too many to review here. For more detailed discussions, see, among others, the monographs \cite{deBoor} and \cite{Dierckx} for B-splines and \cite{EilersMarx}, \cite{BollaertsEM} for P-splines. Some literature on shape-preserving splines (for ordinary shapes such as monotonicity, convexity, etc.) includes, among others, \cite{LiNaikSwetits}, \cite{MammenTA}, \cite{Meyer} and \cite{Ramsay}. In general, the \emph{B-spline} basis of degree $q$ \begin{itemize} \item takes positive values on the domain spanned by $q+2$ adjacent knots, and is zero otherwise; \item consists of $q+1$ polynomial pieces each of degree $q$, and the polynomial pieces join at $q$ inner knots; \item at the joining points, the $q-1$th derivatives are continuous; \item except at the boundaries, it overlaps with $2q$ polynomials pieces of its neighbours; \item at a given $x$, only $q+1$ \emph{B-splines} are nonzero. \end{itemize} Assume that one is interested in approximating the regression function $m\left( x\right) $ in the interval $\left[\underline{x},\overline{x}\right]$. Then we split the interval $\left[\underline{x},\overline{x}\right]$ into $L^{\prime }$ equal length subintervals with $L^{\prime }+1$ knots\footnote{% Although it is possible to have nonequidistant subintervals, for simplicity we consider equally spaced knots. An alternative way to locate the knots is based on the quantiles of the $x$ distribution.}, where each subinterval will be covered with $q+1$ \emph{B-splines} of degree $q$. The total number of knots needed will be $L^{\prime }+2q+1$ (each boundary point $\underline{x}$, $\overline{x}$ is a knot of multiplicity $q+1$) and the number of \emph{B-splines} is $% L=L^{\prime }+q$. Thus, $m\left( x\right) $ is approximated by a linear combination of \emph{B-splines} of degree $q$ with coefficients $(\beta _{1},\ldots ,\beta _{L})$ as \begin{equation} m_{\mathcal{B}}\left( x;L\right) =\sum_{\ell =1}^{L}\beta _{\ell }p_{\ell ,L}\left( x;q\right) \text{,} \label{m_bers} \end{equation}% and where henceforth we shall denote the knots as $\{z^{k}\}$, $% k=1,\ldots,L^{\prime }+2q+1$, where $\underline{x}=z^{1}=\ldots =z^{q+1}$ and $% \overline{x}=z^{L^{\prime }+q+1}=\ldots =z^{L^{\prime }+2q+1}$. \emph{B-splines} share some properties which turns out to be very useful for our purpose. Indeed, \begin{eqnarray*} \left( a\right) ~\sum_{\ell =1}^{L}p_{\ell ,L}\left( x;q\right) &=&1\text{ \ \ for all }x\text{ \ and }q\text{.} \\ \left( b\right) \text{ }\frac{\partial }{\partial x}m_{\mathcal{B}}\left( x;L\right) &=&:m_{\mathcal{B}}^{\prime }\left( x;L\right) =q\sum_{\ell =1}^{L-1}\frac{\triangle \beta _{\ell +1}}{z^{l+1+q}-z^{l+1}}p_{\ell +1,L}\left( x;q-1\right) , \end{eqnarray*}% where $\triangle \beta _{\ell }=\beta _{\ell }-\beta _{\ell -1}$. In particular, $\left( a\right) $ indicates that \emph{B-splines}, as is the case with \emph{Bernstein} polynomials, are a partition of $1$. The property (b) states that the derivative of a \emph{B-spline} of degree $q$ becomes a \emph{B-spline} of degree $q-1$. Using this expression for the derivative and taking into account that the knot system effectively changes with the first and the last knots now removed (thus, the multiplicity of $\underline{x}$ and $\overline{x}$ becomes $q$ rather than $q+1$), one can derive an expression for the second derivative, and so on. We now describe the structure that the sets $S_{q,L}$ take in the Examples \ref{example:deriv1}-\ref{example:meanconvex} given in the previous section. One general idea, especially useful when dealing with shape constraints in Examples \ref{example:rconvex}-\ref{example:meanconvex}, for constructing $S_{q,L}$ is to enforce the shape property of interest at the chosen knots (for simplicity we can take equidistant ones). As $L\rightarrow \infty $, the knot system becomes dense in $\mathcal{X}$ and, thus, the shape property of interest becomes satisfied on an increasingly dense system of points in $\mathcal{X}$.\footnote{On a related note, \cite{Dierckx1980} shows that for cubic splines, if the sign of the second derivative is enforced at the knot points, then the respective convexity/concavity property of the approximation will hold on the whole domain. A similar property can be established for the sign of the first derivative and monotonicity.} \vskip 0.1in \noindent \textbf{Example \ref{example:deriv1} (Cont.).} For our general null hypothesis in (\ref{eq:example1}), the set $S_{q,L}$ is given by \begin{multline*} S_{q,L}=\bigg\{(\beta _{1},\ldots ,\beta _{L})\;|\;\forall \;r\in R,\quad \forall \;z^{k},k=1,\ldots ,L+q+1, \\ d_{r}\cdot m_{\mathcal{B}}\left( z^{k};L\right) \geq c_{r}\bigg\}. \end{multline*} When considering special cases of (\ref{eq:example1}), the set(s) $S_{q,L}$ takes much simpler and/or familiar structures. Indeed, consider $R=\{1\}$ and $c_1=0$, then the constraints become on the regression function to be (weakly) monotone and from the property $\left( b\right) $ of B-splines, it is easy to see that one can take \begin{equation*} S_{q,L} = \left\{(\beta_1, \ldots, \beta_L) \; | \; d_1 (\beta_{\ell+1}-\beta_{\ell})\geq 0, \quad \ell=1, \ldots, L-1 \right\}. \end{equation*} When $R=\{r\}$, $r>1$, $c_r=0$, then the conditions imposes on the set $S_{q,L}$ can be more elaborate than those in the monotonicity scenario. However, if the interval $\left[\underline{x},\overline{x}\right]$ is split into equidistance subintervals, the structure of $S_{q,L}$ can be described quite easily as \begin{equation} S_{q,L}=\left\{ (\beta _{1},\ldots ,\beta _{L})\;|\;d_{r}\sum_{k=0}^{r}(-1)^{r-k}\binom{r}{k}\beta _{\ell +k}\geq 0,\; \ell =q,\ldots ,L-q+1-r\right\} . \label{eq:sqlderiv} \end{equation}% A more refined form of $S_{q,L}$, which may be beneficial for small $L$, would also involve constraints that capture the behaviour of $% m_{\mathcal{B}}\left( x;L\right) $ around the boundary in a more precise way. These constraints are linear inequalities and involve relations among coefficients corresponding to the B-splines around the boundaries. The form of these inequalities would, however, be slightly different from the inequalities in (\ref{eq:sqlderiv}). Just to give an example, for $r=2$, $d_{r}=1$ and $c_r=0$ (thus, convexity testing), additional inequalities around the boundary would have the following form: \begin{align*} (q-1)\triangle \beta _{q+1}& \geq q\triangle \beta _{q},\;\; \ldots ,\;\;\triangle \beta _{3}\geq 2\triangle \beta _{2}, \\ (q-1)\triangle \beta _{L-q+1}& \leq q\triangle \beta _{L-q+2},\;\; \ldots ,\;\; \triangle \beta _{L-1}\leq 2\triangle \beta _{L}. \end{align*}% As $L\rightarrow \infty $, these additional constraints becomes less and less important as constraints (\ref{eq:sqlderiv}) essentially capture the whole convexity property. However, in a finite sample these constraints on the coefficients around the boundary can be important for the power of the test. \vskip 0.1in \noindent \textbf{Example \ref{example:deriv2} (Cont.).} In (\ref{eq:example2}), for each interval $[\tau _{j},\tau _{j+1}]$ in the partition, we have its own system of knots, so now for this interval we denote the number of unique knots as $% L_{j}^{\prime }+1$ and the number of base B-splines for that interval as $% L_{j}=L_{j}^{\prime }+q$. The approximating \emph{B-spline} on that interval is denoted as $m_{\mathcal{B};j}\left( x;L_{j}\right) =\sum_{\ell =1}^{L_{j}}\beta _{j;\ell }p_{j;\ell ,L_{j}}\left( x;q\right) $ and the set of the constraints on coefficients is denoted as $S_{j;q,L_{j}}$. It goes without saying that we assume that $\tau _{j}$ and $\tau _{j+1}$ are included into the knot system for the interval as boundary knots with the multiplicity $q+1$. Denote the complete system of knots on the interval $% (\tau _{j},\tau _{j+1}]$ as $\left\{ z_{j}^{k}\right\} _{k=1}^{L_{j}+q+1}$. Define \begin{multline*} S_{j; q,L_j} = \bigg\{(\beta_{j; 1}, \ldots, \beta_{j; L_j}) \; | \; \forall \; r \in R, \quad \forall \; z_j^k, \quad k=1, \ldots, L_j+ q+1, \\ d_{r_j} \cdot m_{\mathcal{B}; j}\left( z_j^k;L_j\right) \geq 0 \bigg\} \end{multline*} and consider this subset of $\mathbb{R}^{L_j}$ to be embedded into $\mathbb{R% }^{\sum_{j=0}^{J-1} L_j}$, where coefficients from other intervals are taken into account too. Then without any additional constraints on how pieces on different partition intervals are joined together, the constraint set in $% \mathbb{R}^{\sum_{j=0}^{J-1} L_j} $ can be written as $\bigcap_{j=0}^{J-1} S_{j; q,L_j}$. If one wants to impose restrictions that the curves on the adjacent subintervals are joined continuously, then the constraint set is $$S_{cont} = \left(\bigcap_{j=0}^{J-1} S_{j; q,L_j}\right) \bigcap \left(\bigcap_{j=0}^{J-2} \left\{ \beta_{j; L_j}= \beta_{j+1; 1}\right\} \right),$$ whereas under the restriction that the curves are smoothly joined, the constraint set becomes \begin{equation*} S_{smooth} = S_{cont} \bigcap \left( \bigcap_{j=0}^{J-2} \left\{ \beta_{j; L_j}-\beta_{j; L_j-1}= \beta_{j+1; 2}-\beta_{j+1; 1}\right\}\right). \end{equation*} To give a specific example, for the null hypothesis of the smooth \emph{% U-shape} function with switch at the known or estimated $s_{0}$, the constraint set is \begin{multline*} S_{smooth} = \bigg\{(\beta _{0;1},\ldots ,\beta _{0;L_{0}},\beta _{1;1},\ldots ,\beta _{1;L_{1}})\;|\;\beta _{0;\ell +1}-\beta _{0;\ell }\leq 0,\;\forall \ell \leq L_{0}-1,\\ \beta _{1;\ell +1}-\beta _{1;\ell }\geq 0, \; \forall \leq L_{1}-1,\; \; \beta _{0;L_{0}}=\beta _{1;1},\; \; \beta _{0;L_{0}}-\beta _{0;L_{0}-1}=\beta _{1;2}-\beta _{1;1}\bigg\}. \end{multline*} For testing quasi-convexity, the set of constraints is \begin{multline*} S_{q,L}=\bigg\{(\beta _{1},\ldots ,\beta _{L})\;|\;\forall \;z^{k_{1}}\leq z^{k_{2}}\leq z^{k_{3}},\quad k_{1},k_{2},k_{3}\in \{1,\ldots ,L+q+1\}, \\ m_{\mathcal{B}}\left( z^{k_{2}};L\right) \leq \max \left\{ m_{\mathcal{B}% }\left( z^{k_{1}};L\right) ,m_{\mathcal{B}}\left( z^{k_{3}};L\right) \right\} \bigg\}. \end{multline*}% By formulating the constraints in this way, we enforce quasi-convexity at the system of knots. Note, however, that quasi-convexity can also be tested by using the above methodology for testing U-shapes but would require estimating the switch point $s_0$. In this case we can make use of the relatively large literature on estimation of the mode or the maximum of a regression function by nonparametric methods (see \cite{Eddy80}, \cite{Eddy82} or \cite{Muller} for the case when the function is continuous at $s_0$, and \cite{DelgadoHidalgo} or \cite{Hidalgo} for the case when the function is discontinuous at $s_0$). Finally, for testing symmetry around the interval center $s_0$, given symmetric system of knots on $[\underline{x},s_0]$ and $[s_0, \overline{x}$ and, hence, the same number $L$ of B-splines of degree $q$ on these intervals, \begin{multline*} S_{q,L} = \bigg\{(\beta _{0;1},\ldots ,\beta _{0;L},\beta _{1;1},\ldots ,\beta _{1;L})\;|\;\beta_{0;\ell}= \beta_{1; L+1-l} \quad \forall l=1, \ldots, L\bigg\}. \end{multline*}% \vskip 0.1in \noindent \textbf{Example \ref{example:rconvex} (Cont.).} For the sake of brevity, we write the constraints for $r$-convexity only (it would be straightforward for a reader to write the set of constraints for $\rho$-convexity as these two notions are related): \begin{multline*} S_{q,L} = \bigg\{ (\beta_1, \ldots, \beta_L) \; | \; \forall \; z^{k} \quad k \in \{1, \ldots, L+ q+1\}, \\ r \left(m_{\mathcal{B}}'\left( z^{k};L\right)\right)^{2} + m_{\mathcal{B}}''\left( z^{k};L\right) \geq 0; \quad \beta_{\ell} > 0, \; \ell=1, \ldots, L \bigg\}. \end{multline*} \vskip 0.1in \noindent \textbf{Example \ref{example:meanconvex} (Cont.).} Convexity in mean for any two given mean functions $M$ and $N$ can be tested by using the set of constraints \begin{multline*} S_{q,L}=\bigg\{(\beta _{1},\ldots ,\beta _{L})\;|\;\forall \;z^{k_{1}},z^{k_{2}},\quad k_{1},k_{2}\in \{1,\ldots ,L+q+1\}, \\ m_{\mathcal{B}}\left( M\left( z^{k_{1}},z^{k_{2}}\right) ;L\right) \leq N\left( m_{\mathcal{B}}\left( z^{k_{1}};L\right) ,m_{\mathcal{B}}\left( z^{k_{2}};L\right) \right) , \\ \beta _{\ell }>0,\;\ell =1,\ldots ,L\bigg\}. \end{multline*}% This is a generic form of constraints for ${MN}$-convexity. Once we choose a specific form of mean functions $M$ and $N$, it is possible to write these constraints in a modified way. For illustration purposes, below we present alternative ways to write the constraints for GA-convex and HG-convex functions. For GA-convexity, we can consider \begin{multline*} S_{q,L} = \bigg\{ (\beta_1, \ldots, \beta_L) \; | \; \forall \; z^{k_1} < z^{k_2} , \quad k_1,k_2 \in \{1, \ldots, L+ q+1\}, \\ z^{k_1} m^{\prime }_{\mathcal{B}}\left(z^{k_1};L\right) \leq z^{k_2} m^{\prime }_{\mathcal{B}}\left(z^{k_2};L\right), \quad \beta_{\ell} > 0, \; \ell=1, \ldots, L\bigg\}. \end{multline*} For HG-convexity, we can consider \begin{multline*} S_{q,L} = \bigg\{ (\beta_1, \ldots, \beta_L) \; | \; \forall \; z^{k_1} < z^{k_2} , \quad k_1,k_2 \in \{1, \ldots, L+ q+1\}, \\ \frac{\left(z^{k_1}\right)^2 m^{\prime }_{\mathcal{B}}\left(z^{k_1};L\right)% }{m^2_{\mathcal{B}}\left(z^{k_1};L\right)} \leq \frac{\left(z^{k_2}\right)^2 m^{\prime }_{\mathcal{B}}\left(z^{k_2};L\right)}{m^2_{\mathcal{B}% }\left(z^{k_2};L\right)}, \text{ if }, \quad \beta_{\ell} > 0, \; \ell=1, \ldots, L \bigg\}. \end{multline*} \vskip 0.1in Note that in Examples \ref{example:deriv1}-\ref{example:deriv2} the constraints on the coefficients of \emph{% B-splines} happen to be linear\footnote{Quasi-convexity may be an exception depending on the exact approach.} and, therefore, especially easy to implement. We want to emphasize that it is a consequence of the type of the approximation basis that we employ. In other words, it is precisely the properties of \emph{B-splines} that make the testing of these hypotheses particularly easy. In Examples \ref{example:rconvex}-\ref{example:meanconvex} the constraints will in general be non-linear in such coefficients, even though in some special cases -- such as GA-convexity, HA-convexity, or AA-convexity (which of course results in the usual convexity) -- the constraints will still happen to be linear. \vskip 0.1in \noindent \textbf{Bernstein polynomials and other sieve bases.} We now discuss our motivation to not use other sieves bases in our testing approach. As already mentioned above, a strong candidate for approximating the regression function subject to qualitative properties is a \emph{Bernstein} polynomial given by \begin{equation*} m_{\mathcal{B}}\left( x;L\right) =\sum_{\ell =0}^{L}\beta _{\ell }\mathcal{B}% _{\ell ,L}\left( \frac{x-\underline{x}}{\overline{x}-\underline{x}}\right) \text{, \ \ \ }x\in \mathcal{X}\text{,} \end{equation*}% where $\mathcal{B}_{\ell ,L}\left( \tilde{x}\right) =:\left( \begin{array}{c} L \\ \ell% \end{array}% \right) \left( 1-\tilde{x}\right) ^{L-\ell }\tilde{x}^{\ell }$, $\tilde{x}% \in \lbrack 0,1]$, denotes the $\ell th$ base Bernstein polynomial. What makes it a strong candidate is the fact that we can take $\beta _{\ell }=m\left( \underline{x}+\ell /L(\overline{x}-\underline{x})\right) $, and hence, translate constrains on $m\left( x\right) $ into constraints on coefficients $\{\beta _{\ell }\}_{\ell =1}^{L}$. However, our main motivation not to use \emph{Bernstein} polynomials comes from the observation that, contrary to \emph{B-splines }or \emph{P-splines}, \emph{% Bernstein} polynomials are highly correlated. Indeed, using results in \cite{LeeParkYoo}, the eigenvalues of the matrix $\left\{ E\left( \mathcal{B}_{\ell _{1},L}\left( x\right) \mathcal{B}_{\ell _{2},L}\left( x\right) \right) =a_{\ell _{1},\ell _{2}}\right\} _{\ell _{1},\ell _{2}=0}^{L}$ are \begin{equation*} \lambda _{k}=\frac{1}{2L+1}\left( \begin{array}{c} 2L+1 \\ L-k% \end{array}% \right) /\left( \begin{array}{c} 2L \\ L% \end{array}% \right) \text{,} \end{equation*}% which implies that $\lambda _{L}\leq \left( 2L+1\right) ^{-1}L^{1/2}2^{1-2L}$ since $\left( \begin{array}{c} 2L \\ L% \end{array}% \right) \geq L^{-1/2}2^{2L-1}$. In addition, it is not difficult to see that $a_{\ell _{1},\ell _{2}}^{2}/a_{\ell _{1},\ell _{1}}a_{\ell _{2},\ell _{2}}\rightarrow e^{-1}$ if $\left\vert \ell _{1}-\ell _{2}\right\vert =o\left( L^{1/2}\right) $, which yields some adverse and important technical consequences for the test proposed in Section \ref{sec:Khmaladze}. Finally, other sieve bases could potentially be used but, the implementation of the test, even for case in Examples 1 and 2 would be laborious. One popular sieve basis is the power series $1,x,\ldots ,x^{L}$ for any $L$. However, the formulation of constraints becomes increasingly complicated when $L$ increases even if one restricts their attention to such cases as in Examples 1 and 2. For instance, for testing an increasing $m_{\mathcal{B}% }\left( x,L\right) =:\sum_{\ell =0}^{L}\beta _{\ell }x^{\ell }$, one may require all coefficients in the first derivative polynomial $\sum_{\ell =1}^{L} l \beta _{\ell }x^{\ell -1 } $ to be non-negative. These conditions are clearly sufficient but do not become necessary even as $L \rightarrow \infty$. This could lead to the test rejecting with probability one the null hypothesis even if it is true, with the main reason for this being the imposition of constraints that are not true or suitable. A better approach would be to use the one-to-one relationship between the Bernstein polynomials basis and the power basis $1,x,\ldots ,x^{L}$ for any $L$. A similar comment applies if we want to employ the Legendre polynomial base -- as it can be seen from the relationship between the Legendre and Bernstein polynomials given e.g. in Propositions 1 and 2 in \cite{Farouki}, and this, again, seems an unnecessary step.\footnote{% The relations between the Bernstein basis and some other polynomial bases have been addressed in the approximation literature. E.g., \cite{LiZhang} discusses not only the relation between the Bernstein basis and the Legendre basis but also the relation between the Bernstein basis and Chebyshev orthogonal bases.} We shall mention nonetheless that there is no reason to believe that the results or the methodology introduced in this paper cannot be implemented using a power series approximation or potentially some other approximation basis. However, this route seems more arduous than using \emph{B-splines}. Moreover, in the context of testing shape properties we believe it is more natural to employ local objects like \emph{B-splines} than the objects more a global type, like most of the polynomial bases. \subsection{\textbf{First step in the testing methodology and regularity conditions}} \label{section:1step} $\left. {}\right. $ In the remainder of the paper, we shall assume, without loss of generality, that $\mathcal{X}=\left[ 0,1\right]$. We now describe the first step in our methodology of testing $H_{0}$ in (\ref{main_null}) by giving estimators of $m(\cdot )$ under the alternative and the null hypothesis. To that end, write the \emph{B-splines} as a vector of functions \begin{equation} \boldsymbol{P}_{L}\left( x\right) =:\left( p_{1,L}\left( x;q\right) ,\ldots ,p_{L,L}\left( x;q\right) \right) ^{\prime }\text{; \ \ }\boldsymbol{P}_{k}=:% \boldsymbol{P}_{L}\left( x_{k}\right) \text{.} \label{p_appro} \end{equation}% Then, the standard series estimator of $m\left( x\right) $ is defined as the projection of $y$ onto the space spanned by $\boldsymbol{P}_{L}\left( x\right) $, that is% \begin{align} \widetilde{m}_{\mathcal{B}}\left( x_{i};L\right) & =\widetilde{b}^{\prime }% \boldsymbol{P}_{i}\text{, } \label{uncons} \\ \widetilde{b}& =\left( \widetilde{b}_{1},\ldots ,\widetilde{b}_{L}\right) ^{\prime }=\left( \frac{1}{n}\sum_{k=1}^{n}\boldsymbol{P}_{k}\boldsymbol{P}% _{k}^{\prime }\right) ^{+}\frac{1}{n}\sum_{k=1}^{n}\boldsymbol{P}_{k}y_{k}% \text{,} \notag \end{align}% where $B^{+}$ denotes the Moore-Penrose inverse of the matrix $B$. To obtain an estimator under the null hypothesis, we conduct a linear projection subject to suitable constraints. For general null hypothesis (\ref{main_null}), we consider estimation under the constraints in (\ref{main_nullB}), where $S_{q,L}$ satisfies condition $C0$: \begin{equation} \widehat{b}=\left( \widehat{b}_{1},\ldots ,\widehat{b}_{L}\right) =:\arg \min _{\substack{ b_{1},\ldots ,b_{L} \\ \text{s.t. }(b_{1},b_{2},\ldots ,b_{L})\in S_{q,L}}}\sum_{i=1}^{n}\left( y_{i}-\sum_{\ell =1}^{L}b_{\ell }p_{\ell ,L}\left( x_{i};q\right) \right) ^{2} \label{est_1} \end{equation}% This gives us the estimator of $m\left( \cdot \right) $ under $\left( \ref% {main_null}\right) $: \begin{equation} \widehat{m}_{\mathcal{B}}\left( x_{i};L\right) =\widehat{b}^{\prime }% \boldsymbol{P}_{i}\text{.} \label{m_estimate} \end{equation} When the estimation in (\ref{m_estimate}) involves only constraints linear in the parameters, such as in the case of the null hypothesis of an increasing function: \begin{equation} \label{conest_mon} \widehat{b}=\left( \widehat{b}_{1},\ldots ,\widehat{b}_{L}\right) =:\arg \min _{\substack{ b_{1},\ldots ,b_{L} \\ \text{s.t. } b_{1}\leq b_{2}\leq \ldots \leq b_{L}}}\sum_{i=1}^{n}\left( y_{i}-\sum_{\ell =1}^{L}b_{\ell }p_{\ell ,L}\left( x_{i};q\right) \right) ^{2}, \end{equation} then the constrained optimization problem becomes a quadratic programing problem with linear constraints. When the constraints are nonlinear, the constrained estimation may be slightly more challenging to implement but various global optimization techniques can be utilized or even some local optimization techniques as the range of plausible initial values of the coefficients of the B-spline approximation can be assessed from the data/model.\footnote{If the unconstrained least squares estimator is in the interior of $S_{q,L}$, then, of course, none of the constraints are binding and the constrained estimation is standard. The computational complications may only happen when some of the constraints are binding.} To get a better idea of what nonlinear constraints may look like, take e.g. Example \ref{example:meanconvex}. GA-convexity and HA-convexity there result in sets $S_{q,L}$ that have linear constraints only, whereas AG-, AH-, GG-, GH-, HG-, HH-convexities result in the quadratic constraints of the form \begin{equation} \beta ^{\prime }Q_{j}\beta \leq 0,\quad j=1,\ldots ,L+1-q. \label{ex:quadratic_1} \end{equation} Indeed, take for example AH-convexity. Under enough smoothness, this property can be written $\frac{d\,}{d\,x}\left( \frac{m^{\prime }\left( x\right) }{m^{2}\left( x\right) }\right) \geq 0$, which is equivalent to $m^{\prime \prime }\left( x\right) \geq \frac{2\left( m^{\prime }\left( x\right) \right) ^{2}}{m\left( x\right) }$. Taking into account that $m\left( x\right)$ is known to be positive, as in many economics examples, rewrite the last displayed inequality as \begin{equation*} m^{\prime \prime }\left( x\right) m\left( x\right) \geq 2\left( m^{\prime }\left( x\right) \right) ^{2}\text{.} \end{equation*}% Since all $m_{\mathcal{B}}(x;L)$, $m_{\mathcal{B}}^{\prime }(x;L)$, $m_{\mathcal{B}}^{\prime \prime }(x;L)$ are linear at in $\beta $, then for any $x$, this constraint is quadratic in $\beta $. When this inequality is enforced at $ L+1-q=L^{\prime }+1$ unique knots,\footnote{Recall an earlier discussion in Section \ref{section:BPsplines} of a general approach to constructing $S_{q,L}$ by enforcing the shape property of interest at the knots, which guarantees that the shape property becomes satisfied on an increasingly dense system of points in $\mathcal{X}$ as $L \rightarrow \infty$.} it gives us a system of quadratic inequalities (\ref{ex:quadratic_1}). It is worth noting that due to the nature of the B-splines, the matrices $Q_{j}$ are quite sparse. In fact, each $Q_{j}$ contains a non-zero $q\times q$ diagonal block while all the other elements in $Q_{j}$ are zero. This means that each inequality in (\ref{ex:quadratic_1}) will effectively contain only $q$ adjacent coefficients in $\beta $. The analogous conclusion will imply to other properties in Example \ref{example:meanconvex} which result in non-linear constraints.\footnote{% There is a rather rich literature on the quadratic optimization subject to quadratic constraints. A review can be found, e.g., in Park and Boyd (2017).} \vskip 0.1in We now introduce our regularity conditions. \begin{description} \item[\emph{Condition C1}] $\left\{ (x_{i},u_{i})^{\prime }\right\} _{i\in \mathbb{Z}}$ is a sequence of independent and identically distributed random vectors, where $x_{i}$ has support on $\mathcal{X}=:\left[ 0,1\right] $ and its probability density function, $f_{X}\left( x\right) $, is bounded away from zero. In addition, $E[u_{i}|x_{i}]=0$, $E[u_{i}^{2}|x_{i}]=\sigma _{u}^{2}$, and $u_{i}$ has finite $4th$ moments. \vskip 0.1in \item[\emph{Condition C2}] $m\left( x\right) $ is three times continuously differentiable on $\left[ 0,1\right] $. \vskip0.1in \item[\emph{Condition C3}] As $n\rightarrow \infty $, $L$ satisfies $% L^{2}/n+n/L^{4}=o\left( 1\right) $. \end{description} \vskip 0.1in Condition $C1$ can be weakened to allow for heteroscedasticity, e.g. $E\left[ u_{i}^{2}\mid x\right] =\sigma _{u}^{2}\left( x\right) $ as it was done in \cite{Stute97}. However, the latter condition complicates the technical arguments and for expositional simplicity we omit a detailed analysis of this case. However, in our empirical applications we present examples with heteroscedastic errors and illustrate how to deal with it in practice. Condition $C3$ bounds the rate at which $L$ increases to infinity with $n$. Condition $C2$ is a smoothness condition on the regression function $m\left( x\right) $. It guarantees that the approximation error or bias \begin{equation} m^{bias}\left( x\right) =:m_{\mathcal{B}}\left( x;L\right) -m\left( x\right) \label{m_bias} \end{equation}% is $O(L^{-3})$, see see Agarwal and Studden's \cite{AgarwalStudden} Theorems 3.1 and 4.1 or Zhou et al. \cite{Zhou}. It can be weakened to say that the second derivatives are H\"{o}lder continuous of degree $\eta >0$. In that case, $C3$ had to be modified to $L^{2}/n+n/L^{2+2\eta }=o\left( 1\right) $. In case of using \emph{P-splines} we refer to Claeskens \cite{Claeskens_etal} Theorem 2. \subsection{\textbf{THE TESTING METHODOLOGY}} $\left. {}\right. $ Having presented our estimators of $m\left( x\right) $ using or not the constraints induced by the null hypothesis, we now discuss the main part of the methodology for testing shape restrictions outlined in the introduction. We shall focus on the null hypothesis $\left( \ref{main_nullB}\right) $ in terms of the coefficients $\{ \beta _{\ell }\} _{\ell =1}^{L}$, with the alternative hypothesis being the negation of the null. Since the boundary of $S_{q,L}$ is taken to consist of a finite number of smooth surfaces, $% S_{q,L} $ can be described by a finite number of restrictions. In this way, our testing problem translates into the more familiar testing scenario when the null hypothesis is given as a set of constraints on the parameters of the model. However the main and key difference is that, in our scenario, the number of such constraints increases with the sample size. When testing for constraints among the parameters in a regression model, one possibility is via the (Quasi) Likelihood Ratio principle which compares the fits obtained by constrained and unconstrained estimates $\widehat{m}_{% \mathcal{B}}\left( x_{i};L\right) $ and $\widetilde{m}_{\mathcal{B}}\left( x_{i};L\right) $, respectively. That is, \begin{equation*} \mathcal{LR}_{n}\left( L\right) =\frac{1}{n}\sum_{i=1}^{n}\left( y_{i}-% \widehat{m}_{\mathcal{B}}\left( x_{i};L\right) \right) ^{2}-\frac{1}{n}% \sum_{i=1}^{n}\left( y_{i}-\widetilde{m}_{\mathcal{B}}\left( x_{i};L\right) \right) ^{2}\text{.} \end{equation*}% Although results in \cite{ChenChristensenJOE} might be used to obtain its asymptotic distribution, this route appears difficult to implement, since there is a high probability that both estimators $\widehat{m% }_{\mathcal{B}}\left( x_{i};L\right) $ and $\widetilde{m}_{\mathcal{B}% }\left( x_{i};L\right) $ coincide numerically. The latter is the case when the constraints are not binding. A second possibility is to employ the Wald principle, which involves checking if the constraints in $\left( \ref{main_nullB}\right) $ hold true for the unconstrained estimator $\widetilde{b}$ of the parameters $\beta =:\left\{ \beta _{\ell }\right\} _{\ell =1}^{L}$ given in $\left( \ref% {uncons}\right) $, or, in other words, if the data supports the set of constraints \begin{equation*} (\widetilde{b}_{1},\widetilde{b}_{2},\ldots ,\widetilde{b}_{L})\in S_{q,L}. \end{equation*}% Generally $S_{q,L}$ will involve some inequalities, and even when $L$ is fixed, the test would be difficult to implement or even to compute the critical values based on its asymptotic distribution. In our scenario, however, we would have two potential further technical complications. First, the number of constraints increases with the sample size $n$, which makes, from both a theoretical and practical point of view, this route very arduous, if at all feasible. Second, when some constraints describing $% S_{q,L}$ are binding, we are then dealing with estimation at the boundary which implies that the asymptotic distribution cannot be Gaussian. A third way to test for the null hypothesis $\left( \ref{main_nullB}\right) $ is to implement the Lagrange Multiplier,\emph{\ LM}, test -- that is to check if the residuals% \begin{equation} \widehat{u}_{i}=y_{i}-\widehat{m}_{\mathcal{B}}\left( x_{i};L\right) \text{,}% \quad i=1,\ldots ,n\text{.} \label{res_con} \end{equation}% and $x_{i}$ satisfy the orthogonality condition imposed by Condition $C1$. That is, we might base our test on whether or not the set of moment conditions \begin{equation} \mathcal{K}_{n}\left( x\right) =\frac{1}{n}\sum_{i=1}^{n}\mathcal{I}% _{i}\left( x\right) \widehat{u}_{i}\text{, \ \ \ \ }x\in \left[ 0,1\right] \label{T_n} \end{equation}% are significantly different from zero, where $\mathcal{I}$ is the indicator function and we abbreviate $\mathcal{I}\left( x_{i}<x\right) $ as $\mathcal{I}% _{i}\left( x\right) $. This approach was described and examined in \cite{Stute97} or \cite{Andrews97} with a more econometric emphasis. Tests based on $\mathcal{K}_{n}\left( x\right) $ are known as testing using partial sum empirical processes. Recall that in a standard regression model the \emph{LM} test is based on the first order conditions% \begin{equation*} \mathcal{LM}_{n}\left( L\right) =\frac{1}{n}\sum_{i=1}^{n}p_{\ell ,L}\left( x_{i};q\right) \widehat{u}_{i}\text{,} \end{equation*}% which has the interpretation of whether the residuals and regressors, $% p_{\ell ,L}\left( x_{i};q\right) $, satisfies the orthogonality moment condition induced by Condition $C1$. However to motivate the reasons to employ a transformation of $\mathcal{K}% _{n}\left( x\right) $, given in $\left( \ref{k_1}\right) $ or $\left( \ref% {k_2}\right) $ below, as the basis for our test statistic, it is worth examining the structure of $\mathcal{K}_{n}\left( x\right) $ given in $% \left( \ref{T_n}\right) $. For that purpose, we observe that% \begin{equation} \mathcal{K}_{n}\left( x\right) =\frac{1}{n}\sum_{i=1}^{n}\mathcal{I}% _{i}\left( x\right) u_{i}-\sum_{\ell =1}^{L}\left( \widehat{b}_{\ell }-\beta _{\ell }\right) \mathcal{P}_{n,\ell }\left( x;q\right) +\frac{1}{n}% \sum_{i=1}^{n}\mathcal{I}_{i}\left( x\right) m^{bias}\left( x_{i}\right) \text{,} \label{T_n1} \end{equation}% where $m^{bias}\left( x_{i}\right) $ was given in $\left( \ref{m_bias}% \right) $ and \begin{equation} \mathcal{P}_{n,\ell }\left( x;q\right) =:\frac{1}{n}\sum_{i=1}^{n}\mathcal{I}% _{i}\left( x\right) p_{\ell ,L}\left( x_{i};q\right) \text{.} \label{p_L} \end{equation}% Now, the third term on the right of $\left( \ref{T_n1}\right) $ is $O\left( L^{-3}\right) $ by Agarwal and Studden's \cite{AgarwalStudden} Theorems 3.1 and 4.1, and then Condition $C1$. On the other hand, standard arguments and Condition $C1$ imply that \begin{equation*} n^{1/2}\mathcal{U}_{n}\left( x\right) =\frac{1}{n^{1/2}}\sum_{i=1}^{n}% \mathcal{I}_{i}\left( x\right) u_{i}\overset{weakly}{\Rightarrow }\mathcal{U}% \left( x\right) =:\sigma _{u}\mathcal{B}\left( F_{X}\left( x\right) \right) \text{,} \end{equation*}% where $\mathcal{B}\left( z\right) $ denotes the standard Brownian motion and $F_{X}\left( x\right) $ the distribution function of $x_{i}$. Next, we discuss the contribution due to $\sum_{\ell =1}^{L}\left( \widehat{b% }_{\ell }-\beta _{\ell }\right) \mathcal{P}_{n,\ell }\left( x;q\right) $. In standard lack-of-fit testing problems with $L$ finite, when $\mathcal{K}% _{n}\left( x\right) =O_{p}\left( n^{-1/2}\right) $, its contribution is nonnegligible as first showed by \cite{Durbin73} and later by \cite{Stute97} in a regression model context. However the proof of Theorem \ref{M_n} suggests that $\sum_{\ell =1}^{L}\left( \widehat{b}% _{\ell }-\beta _{\ell }\right) \mathcal{P}_{n,\ell }\left( x;q\right) =O_{p}\left( \left( L/n\right) ^{1/2}\right) $, so that as $L$ increases with the sample size, it yields that the normalization factor for $\mathcal{K% }_{n}\left( x\right) $ is of order $n^{\alpha }$ for some $\alpha <1/2$. Thus the previous arguments suggest that under our conditions, we would have \begin{equation*} \left( \frac{n}{L}\right) ^{1/2}\mathcal{K}_{n}\left( x\right) =:-\left( \frac{n}{L}\right) ^{1/2}\sum_{\ell =1}^{L}\left( \widehat{b}_{\ell }-\beta _{\ell }\right) \mathcal{P}_{n,\ell }\left( x;q\right) \left( 1+o_{p}\left( 1\right) \right) \text{.} \end{equation*}% Results by Newey \cite{Newey97}, Lee and Robinson \cite{LeeRobinson} or Chen and Christensen \cite{ChenChristensenJOE} in a more general context, might suggest that the left side of the last displayed expression might converge to a Gaussian process when $\beta $ is in the interior of the set $% S_{q,L}$. However, when $\beta $ is at the boundary of $S_{q,L}$, then the asymptotic distribution is not Gaussian, and so to obtain the asymptotic distribution of $\left( n/L\right) ^{1/2}\sum_{\ell =1}^{L}\left( \widehat{b}% _{\ell }-\beta _{\ell }\right) \mathcal{P}_{n,\ell }\left( x;q\right) $ for inference purposes appears quite difficult, if at all possible. So, the purpose of the next section is to examine a transformation of $% \mathcal{K}_{n}\left( x\right) $ such that its statistical behaviour will be free from $\sum_{\ell =1}^{L}\left( \widehat{b}_{\ell }-\beta _{\ell }\right) \mathcal{P}_{n,\ell }\left( x;q\right) $. The consequence of the transformation would then be twofold. First, we would obtain that the transformation of $n^{1/2}\mathcal{K}_{n}\left( x\right) $ is $O_{p}\left( 1\right) $, which leads to better statistical properties of the test, and secondly and more importantly, the test will be pivotal in the sense that $% \sigma _{u}^{2}$ becomes the only unknown (although easy to estimate) of its asymptotic distribution. One consequence of our results is that the asymptotic distribution becomes independent of the null hypothesis under consideration. \section{\textbf{KHMALADZE'S TRANSFORMATION}} \label{sec:Khmaladze} This section examines a transformation of $\mathcal{K}_{n}\left( x\right) $ whose asymptotic distribution is free from the statistical behaviour of $% \left\{ \widehat{b}_{\ell }\right\} _{\ell =1}^{L}$. To that end, we propose a \textquotedblleft \emph{martingale}\textquotedblright\ transformation based on ideas by \cite{Khmaladze81}, see also \cite{BrownDurbinEvans} for earlier work. The (linear) transformation, denoted $% \mathcal{T}$, should satisfy that \begin{eqnarray} \left( \emph{i}\right) \text{ }n^{1/2}\left( \mathcal{TU}_{n}\right) \left( x\right) \overset{weakly}{\Rightarrow }\mathcal{U}\left( x\right) &=&:\sigma _{u}\mathcal{B}\left( F_{X}\left( x\right) \right) \notag \\ \left( \emph{ii}\right) \text{ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ }% n^{1/2}\left( \mathcal{TP}_{L}\right) \left( x\right) &=&0 \label{properties} \\ \left( \emph{iii}\right) \text{ \ \ \ \ \ \ \ \ \ \ \ \ }n^{1/2}\left( \mathcal{T}m^{bias}\right) \left( x\right) &=&o\left( 1\right) \text{,} \notag \end{eqnarray}% where% \begin{equation*} \mathcal{P}_{L}\left( x\right) =:E\left( \mathcal{P}_{n}\left( x;L\right) \right) =\int_{0}^{x}\boldsymbol{P}_{L}\left( z\right) f_{X}\left( z\right) dz\text{ \ \ }\left( \mathcal{P}_{n}\left( x;L\right) =:\left\{ \mathcal{P}% _{n,\ell }\left( x;q\right) \right\} _{\ell =1}^{L}\right) \text{,} \end{equation*}% with $\boldsymbol{P}_{L}\left( x\right) $ and $\mathcal{P}_{n,\ell }\left( x;q\right) $ given respectively in $\left( \ref{p_appro}\right) $ and $% \left( \ref{p_L}\right) $. \subsection{\textbf{ALL THE CONSTRAINTS ON $\protect\beta_{\ell}$ ARE LINEAR}% } \label{sec:Khmlinear} $\left. {}\right. $ We first present the form of Khmaladze's transformation when all the constraints characterizing $S_{q,L}$ are linear and, thus, the surfaces forming the boundary of $% S_{q,L}$ are hyperplanes. For any $x<1$, denote \begin{equation} A_{L}\left( x\right) =\int_{x}^{1}\left( \boldsymbol{P}_{L}\left( z\right) \boldsymbol{P}_{L}^{\prime }\left( z\right) \right) f_{X}\left( z\right) dz% \text{.} \label{L_x} \end{equation}% We then define the transformation $\mathcal{T}$ as% \begin{equation*} \left( \mathcal{TW}\right) \left( x\right) =\mathcal{W}\left( x\right) -\int_{0}^{x}\boldsymbol{P}_{L}^{\prime }\left( z\right) A_{L}^{+}\left( z\right) \left( \int_{z}^{1}\boldsymbol{P}_{L}\left( w\right) \mathcal{W}% \left( dw\right) \right) f_{X}\left( z\right) dz\text{, \ \ }x<1\text{.} \end{equation*}% It is easy to see that the transformation $\mathcal{T}$ satisfies condition $% \left( \emph{ii}\right) $ in $\left( \ref{properties}\right) $, so that the main concern will be to show that $\left( \emph{i}\right) $ and $\left( \emph{iii}\right) $ hold true. It is important to note that the constraints that proved to be binding in the estimation under the null hypothesis have to be incorporated when conducting the Khmaladze's transformation and, thus, $\boldsymbol{P}_{L}$ in $\left( \ref{p_appro}\right) $ has to be redefined. To give a specific example, consider the monotonicity testing and, thus, the constrained estimation as in (\ref{conest_mon}). Suppose the estimation results in one binding constraint $\widehat{b}_{\ell _{0}}=\widehat{b}_{\ell _{0}+1} $. Then, we redefine $P_k$ in $\left( \ref{p_appro}\right)$ as \begin{eqnarray*} \boldsymbol{P}_{k} &=&:\boldsymbol{P}_{L}\left( x_{k}\right), \quad \text{where} \\ \boldsymbol{P}_{L}\left( x\right) &=&:\left( p_{1,L}\left( x;q\right) ,\ldots ,p_{\ell _{0},L}\left( x;q\right) +p_{\ell _{0}+1,L}\left( x;q\right) ,p_{\ell _{0}+2,L}\left( x;q\right) ,...p_{L,L}\left( x;q\right) \right) ^{\prime }\text{,} \end{eqnarray*}% so that we effectively incorporate the binding constraint. The analogous methodology applies with several binding constraints. In what follows then, we shall make no distinction among various cases of binding constraints to simplify the exposition of the arguments. However, we shall emphasize that the use of the \textquotedblleft correct\textquotedblright $P_{L}\left( x\right) $ is crucial for the power of the test. Using $P_{L}\left( x\right) $ as given in $\left( \ref{p_appro}\right)$ without taking into account the binding the constraints will make the test to have only trivial power (see the discussion in Section \ref{sec:power}). The transformation $\mathcal{T}$ has only a theoretical value and as such, from an inferential point of view, we need to replace it by its sample analogue, which we shall denote by $\mathcal{T}_{n}$. To that end, for any $x\in \mathcal{X}$, define \begin{eqnarray} F_{n}\left( x\right) &=&\frac{1}{n}\sum_{k=1}^{n}\mathcal{I}_{k}\left( x\right) \label{notation} \\ \text{\ }C_{n}\left( x\right) &=&\frac{1}{n}\sum_{k=1}^{n}\boldsymbol{P}% _{k}u_{k}\mathcal{J}_{k}\left( x\right) ;~\ \ \ A_{n}\left( x\right) =\frac{1% }{n}\sum_{k=1}^{n}\boldsymbol{P}_{k}\boldsymbol{P}_{k}^{\prime }\mathcal{J}% _{k}\left( x\right) \notag \end{eqnarray}% where $\mathcal{I}\left( x\leq x_{k}\right) =:\mathcal{J}_{k}\left( x\right) =1-\mathcal{I}_{k}\left( x\right) $. In what follows, we shall abbreviate \begin{equation} C_{n,i}=:C_{n}\left( \widetilde{x}_{i}\right) ;\text{ \ }A_{n,i}=:A_{n}% \left( \widetilde{x}_{i}\right) \text{,} \label{A_n_i} \end{equation}% where $\widetilde{x}_{i}=x_{i}$ if $x_{i}+n^{-\varsigma }<z^{k\left( x_{i}\right) }$ and $=z^{k\left( x_{i}\right) }$ otherwise, with $z^{k\left( x\right) }$ denoting the closest knot $z^{k}$, $k=1,\ldots ,L$, bigger than $% x$ and $1/2<\varsigma <1$. The motivation to make this \textquotedblleft trimming\textquotedblright\ is because when $x_{i}$ is too close to $% z^{k\left( x_{i}\right) }$, the \emph{B-spline} is close but not equal to zero, which induces some technical complications in the proof of our main results. However, in small samples this\emph{\ }\textquotedblleft trimming\textquotedblright\ does not appear to be needed, becoming a purely technical argument. We define the sample analogue of $\left( \mathcal{TW}\right) \left( x\right) $ as \begin{equation} \left( \mathcal{T}_{n}\mathcal{W}\right) \left( x\right) =\mathcal{W}\left( x\right) -\frac{1}{n}\sum_{i=1}^{n}\boldsymbol{P}_{i}^{\prime }A_{n,i}^{+}\int_{\widetilde{x}_{i}}^{1}\boldsymbol{P}_{L}\left( w\right) \mathcal{W}\left( dw\right) \mathcal{I}_{i}\left( x\right) \text{.} \label{Khma_n} \end{equation} The transformation in $\left( \ref{Khma_n}\right) $ has a rather simple motivation. Suppose that we have ordered the observations according to $% x_{i} $, that is $x_{i-1}\leq x_{i}$, $i=2,...,n$, which would not affect the statistical behaviour of $\mathcal{K}_{n}\left( x\right) $. The latter follows by the well known argument that% \begin{equation} \sum_{i=1}^{n}g\left( x_{i}\right) =\sum_{i=1}^{n}g\left( x_{\left( i\right) }\right) \text{,} \label{g_i} \end{equation}% where $x_{\left( i\right) }$ is the $i-th$ order statistic of $\left\{ x_{i}\right\} _{i=1}^{n}$. So, we have that \begin{equation} \mathcal{K}_{n}\left( x_{j}\right) =\frac{1}{n}\sum_{i=1}^{n}\widehat{u}_{i}% \mathcal{I}_{i}\left( x_{j}\right) =:\frac{1}{n}\sum_{i=1}^{j}\widehat{u}_{i}% \text{.} \label{cusum} \end{equation}% Now, $\widehat{u}_{i}=u_{i}-\boldsymbol{P}_{i}^{\prime }A_{n}^{+}\left( 0\right) C_{n}\left( 0\right) $, where \begin{eqnarray*} C_{n}\left( 0\right) &=&\frac{1}{n}\sum_{k=1}^{n}\boldsymbol{P}_{k}u_{k}% \mathcal{I}_{k}\left( x_{i}\right) +\frac{1}{n}\sum_{k=1}^{n}\boldsymbol{P}% _{k}u_{k}\mathcal{J}_{k}\left( x_{i}\right) \\ &=&\frac{1}{n}\sum_{k=1}^{i-1}\boldsymbol{P}_{k}u_{k}+\frac{1}{n}% \sum_{k=i}^{n}\boldsymbol{P}_{k}u_{k}\text{.} \end{eqnarray*}% So, if instead of $C_{n}\left( 0\right) $ and $A_{n}\left( 0\right) $, we employed $C_{n,i}=n^{-1}\sum_{k=i}^{n}\boldsymbol{P}_{k}u_{k}$ and $% A_{n,i}=n^{-1}\sum_{k=i}^{n}\boldsymbol{P}_{k}\boldsymbol{P}_{k}^{\prime }$, we would replaced $\widehat{u}_{i}$ by \begin{equation} v_{i}=u_{i}-\boldsymbol{P}_{i}^{\prime }A_{n,i}^{+}C_{n,i} \label{u_rec} \end{equation}% in $\left( \ref{cusum}\right) $, so that it has a martingale difference structure as $E\left[ v_{i}\mid past\right] =0$, in comparison with $% \widehat{u}_{i}$ where $E\left[ \widehat{u}_{i}\mid past\right] \neq 0$. This is the idea behind the so-called (recursive) \emph{Cusum} statistic first examined in \cite{BrownDurbinEvans} and developed and examined in length in \cite{Khmaladze81}. Observe that $\left( % \ref{u_rec}\right) $ becomes the \textquotedblleft prediction error\textquotedblright of $u_{i}$ when we use the \textquotedblleft last\textquotedblright\ $j=i,\ldots,n$ observations. Thus, the preceding argument yields the Khmaladze's transformation \begin{equation} \left( \mathcal{T}_{n}\mathcal{K}_{n}\right) \left( x\right) =\mathcal{M}% _{n}\left( x\right) =:\frac{1}{n^{1/2}}\sum_{i=1}^{n}v_{i}\mathcal{I}% _{i}\left( x\right) \text{.} \label{k_1} \end{equation}% Observe that, using $\left( \ref{g_i}\right) $, we could write $\mathcal{M}% _{n}\left( x\right) $ as \begin{equation*} \mathcal{M}_{n}\left( x\right) =:\frac{1}{n^{1/2}}\sum_{i=1}^{n}v_{\left( i\right) }\mathcal{I}_{\left( i\right) }\left( x\right) \text{.} \end{equation*}% Now, Lemma \ref{bias} in Appendix B implies that the condition $\left( \emph{ii}\right) $ in $% \left( \ref{properties}\right) $ holds true when \begin{equation*} \mathcal{W}_{n}\left( x\right) =\frac{1}{n}\sum_{k=1}^{n}\boldsymbol{P}_{k}% \mathcal{I}_{k}\left( x\right) \text{,} \end{equation*}% so the technical problem is to show that (asymptotically) conditions $\left( \emph{i}\right) $ and $\left( \emph{iii}\right) $ in $\left( \ref{properties}% \right) $ also hold true. That is, to show that \begin{eqnarray*} \left( 1\right) &&\mathcal{M}_{n}\left( x\right) \overset{weakly}{% \Rightarrow }\mathcal{U}\left( x\right) \text{ \ \ }x\in \left[ 0,1\right] \\ \left( 2\right) &&\left( \mathcal{T}_{n}m^{bias}\right) \left( x\right) =o_{p}\left( n^{-1/2}\right) \text{.} \end{eqnarray*} Finally, it is worth mentioning that in $\left( \ref{Khma_n}\right) $ we might have employed $\mathcal{J}_{k}\left( x\right) =\mathcal{I}\left( x<x_{k}\right) $ instead of our definition $\mathcal{J}_{k}\left( x\right) =% \mathcal{I}\left( x\leq x_{k}\right) $. However because by definition of \emph{B-splines}, the matrix $A_{n,i}$, and hence $A_{L}\left( x_{i}\right) $% , might be singular, if we employed $\mathcal{J}_{k}\left( x\right) =% \mathcal{I}\left( x<x_{k}\right) $, then it would not be guaranteed that% \begin{equation*} \boldsymbol{P}_{i}^{\prime }-\boldsymbol{P}_{i}^{\prime }A_{n,i}^{+}A_{n,i}=0% \text{.} \end{equation*}% On the other hand, Theorem 12.3.4 in \cite{Harville} yields that the last displayed equation holds true when $\mathcal{J}_{k}\left( x\right) =\mathcal{I}\left( x\leq x_{k}\right) $. Now, $\left( 1\right) $ will be shown in the next theorem, whereas $\left( 2\right) $ is shown in Theorem \ref{M_nest}. \begin{theorem} \label{M_n}Under Conditions $C1-C3$, we have that \begin{equation*} \mathcal{M}_{n}\left( x\right) \overset{weakly}{\Rightarrow }\mathcal{U}% \left( x\right) ;\text{ \ \ }x\in \left[ 0,1\right] \text{.} \end{equation*} \end{theorem} Unfortunately, we do not observe $u_{i}$, so that to implement the transformation we replace $v_{i}$ by $\widehat{v}_{i}$, where $\widehat{v}% _{i}$ is defined as $v_{i}$ in $\left( \ref{u_rec}\right) $ but where we replace $u_{i}$ by $\widehat{u}_{i}$ as defined in $\left( \ref{res_con}% \right) $, yielding the statistic% \begin{equation} \widetilde{\mathcal{M}}_{n}\left( x\right) =:\frac{1}{n^{1/2}}\sum_{i=1}^{n}% \widehat{v}_{i}\mathcal{I}_{i}\left( x\right) \text{.} \label{k_2} \end{equation} \begin{theorem} \label{M_nest} Assuming that $H_{0}$ holds true, under Conditions $C1-C3$, we have that \begin{equation*} \widetilde{\mathcal{M}}_{n}\left( x\right) \overset{weakly}{\Rightarrow }% \mathcal{U}\left( x\right) \text{.} \end{equation*} \end{theorem} Denote the estimator of the variance of $u_{i}$, $\sigma _{u}^{2}$, by \begin{equation*} \widehat{\sigma }_{u}^{2}=\frac{1}{n}\sum_{i=1}^{n}\widehat{u}_{i}^{2}\text{.% } \end{equation*} \begin{proposition} \label{sigma_est}Under Conditions $C1-C3$, we have that $\widehat{\sigma }% _{u}^{2}\overset{P}{\rightarrow }\sigma _{u}^{2}$. \end{proposition} We then have the following corollary. \begin{corollary} Under $H_{0}$ and assuming Conditions $C1-C3$, for any continuous functional $g:\mathbb{R\rightarrow R}^{+}$, \begin{equation*} g\left( \widetilde{\mathcal{M}}_{n}\left( x\right) /\widehat{\sigma }% _{u}\right) \overset{d}{\rightarrow }g\left( \mathcal{U}\left( x\right) /\sigma _{u}\right) \text{.} \end{equation*} \end{corollary} \begin{proof} The proof is standard using Theorem $\ref{M_nest}$, Proposition \ref{sigma_est} and the continuous mapping theorem, so it is omitted. \end{proof} Denoting $\tilde{n}=:n-L-2$ and $\widetilde{\mathcal{M}}_{n}\left( x^{q}\right) =\widetilde{\mathcal{M}}_{n,q}$, where $x^{q}=q/n$, standard functionals are the Kolmogorov-Smirnov, Cram\'{e}r-von-Mises and Anderson-Darling tests defined respectively as{\ \begin{eqnarray} \mathcal{KS}_{n} &\mathcal{=}&\sup_{q=1,...,\tilde{n}}\left\vert \frac{% \widetilde{\mathcal{M}}_{n,q}}{\widehat{\sigma }_{u}}\right\vert \overset{d}{% \rightarrow }\sup_{x\in \left( 0,1\right) }\left\vert \mathcal{B}\left( F_{X}\left( x\right) \right) \right\vert \notag \\ \mathcal{C}v\mathcal{M}_{n} &\mathcal{=}&\frac{1}{\tilde{n}}\sum_{q=1}^{% \tilde{n}}\frac{\widetilde{\mathcal{M}}_{n,q}^{2}}{\widehat{\sigma }_{u}^{2}}% \overset{d}{\rightarrow }\int_{0}^{1}\mathcal{B}^{2}\left( F_{X}\left( x\right) \right) dx\text{,} \label{KSAD} \end{eqnarray}% }% \begin{equation*} \mathcal{AD}_{n}\mathcal{=}\frac{1}{\tilde{n}}\sum_{q=1}^{\tilde{n}}\frac{% \widetilde{\mathcal{M}}_{n,q}^{2}}{\widehat{\sigma }_{u}^{2}x^{q}\left( 1-x^{q}\right) }\overset{d}{\rightarrow }\int_{0}^{1}\frac{\mathcal{B}% ^{2}\left( F_{X}\left( x\right) \right) }{F_{X}\left( x\right) \left( 1-F_{X}\left( x\right) \right) }dx\text{.} \end{equation*} \vskip 0.1in \subsection{\textbf{NONLINEAR CONSTRAINTS ON $\protect\beta_{\ell}$}} $\left. {}\right. $ We turn now our attention to describing the Khmaladze's transformation in situations when some constraints describing $S_{q,L}$ may be non-linear, as it happens to be in our Examples 3 and 4. If the constrained estimate $\widehat{b}$ lies in the interior of $S_{q,L}$ (and, thus, coincides with the unconstrained estimate), then the transformation is conducted in the same way as in Section \ref{sec:Khmlinear}. However, the transformation will have a modified form if $% \widehat{b}$ lies on the boundary of $S_{q,L}$. We give its form and, in particular, discuss what objects plays the role of $\boldsymbol{P}_{L}\left( x\right)$ described in the previous section. For expositional simplicity suppose that $\widehat{b}$ belongs only to one of the smooth surfaces describing the boundary of $S_{q,L}$. Usually these surfaces will be defined by implicit functions but, applying the implicit function theorem, the explicit representation of this surface can be obtained either analytically or numerically (even if local, which would suffice since $\widehat{b}$ is consistent) with respect to one parameter expressed as a function of other parameters: suppose that for some $\ell _{0}$ we can express it as $% \beta_{\ell _{0}}=h(\beta _{1},\ldots ,\beta _{\ell _{0}-1},\beta _{\ell _{0}+1},\ldots ,\beta _{L})$. In cases such as AG-, AH-convexity and other similar properties in Example \ref{example:meanconvex}, for the reasons discussed in Section \ref{section:1step} this would be a restriction $\beta_{\ell _{0}}=h(\beta _{\ell _{0}-2},\beta _{\ell _{0}-1})$ obtained from an implicit function which is a polynomial of degree 2. Then for the purpose of conducting the Khmaladze's transformation, instead of approximating $m(\cdot )$ by the linear function $\sum_{k=0}^{\ell -1}\beta _{k}p_{k}\left( x_{i}\right) $, we consider the approximation given by \begin{equation} g\left( x_{i};\beta_{-\ell _{0}}\right) =:\sum_{k=0}^{\ell _{0}-1}\beta _{k}p_{k}\left( x_{i}\right) +h(\beta_{-\ell _{0}})p_{\ell _{0}}\left( x_{i}\right) +\sum_{k=\ell _{0}+1}^{L}\beta _{k}p_{k}\left( x_{i}\right) , \label{2} \end{equation}% where $\beta_{-\ell _{0}}=(\beta _{1},\ldots ,\beta _{\ell _{0}-1},\beta _{\ell _{0}+1},\ldots ,\beta _{L})$. Although this approximation function is nonlinear in parameters, it has a simple and useful structure, as it will become evident from our analysis below. To define the Khmaladze's transformation, denote the vector of first derivatives of $g\left( x;\beta_{-\ell _{0}}\right) $ with respect to the parameters as \begin{eqnarray*} \widetilde{P}\left( x;\beta_{-\ell _{0}}\right) &=&:\frac{\partial }{% \partial \beta_{-\ell _{0}}}g\left( x;\beta_{-\ell _{0}}\right) \\ &=&:\left\{ \widetilde{p}_{\ell }\left( x;\beta_{-\ell _{0}}\right) \right\} _{\ell =1}^{L}\text{,} \end{eqnarray*}% where \begin{equation*} \widetilde{p}_{\ell }\left( x;\beta_{-\ell _{0}}\right) =:p_{\ell }\left( x\right) +\frac{\partial h\left( \beta_{-\ell _{0}}\right) }{\partial \beta _{\ell }}p_{\ell _{0}}\left( x\right) ,\quad \ell \neq \ell _{0}. \end{equation*} Then, the Khmaladze's transformation of the test statistic $\mathcal{K}% _{n}\left( x\right) $ in (\ref{T_n}) has the following form: \begin{equation*} \mathcal{M}_{n}\left( x\right) =:\frac{1}{n^{1/2}}\sum_{i=1}^{n}v_{i}% \mathcal{I}_{i}\left( x\right) , \end{equation*}% where \begin{equation*} v_{i}=u_{i}-\widetilde{P}_{i}^{\prime }\left( \widehat{b}_{-\ell _{0}}\right) \mathcal{D}_{n}^{+}\left( \widehat{b}_{-\ell _{0}};i\right) \sum_{k=1}^{n}\widetilde{P}_{k}\left( \widehat{b}_{-\ell _{0}}\right) {u}_{k}\mathcal{J}_{k}\left( \widetilde{x}_{i}\right), \end{equation*} $\widetilde{P}_{i}\left( \beta_{-\ell _{0}}\right) =:\widetilde{P}% \left( x_{i};\beta_{-\ell _{0}}\right) $, and $\mathcal{D}_{n}\left(x; \beta _{-\ell _{0}}\right) =\sum_{k=1}^{n}\widetilde{P}_{k}\left( \beta _{-\ell _{0}}\right) \widetilde{P}_{k}^{\prime }\left( \beta_{-\ell _{0}}\right) \mathcal{J}_{k}\left(x\right) $, and $\mathcal{D}^{+}_{n}\left(x; \beta _{-\ell _{0}}\right)$ is the Moore-Penrose inverse of $\mathcal{D}_{n}\left(x; \beta _{-\ell _{0}}\right)$, and $\mathcal{D}_{n}\left( \beta _{-\ell _{0}};i\right) =\mathcal{D}_{n}\left(\widetilde{x}_{i}; \beta _{-\ell _{0}}\right) $ with $\widetilde{x}_{i}$ defined in the same way as in Section \ref{sec:Khmlinear}. Note that by employing $\widetilde{p}_{\ell }\left( x_i;\beta_{-\ell _{0}}\right)$ instead of $p_{\ell }\left( x_i\right)$, we have automatically incorporated in the transformation our binding restriction. In practice, instead of $v_{i}$ we use \begin{equation*} \widehat{v}_{i}=\widehat{u}_{i}-\widetilde{P}_{i}^{\prime }\left( \widehat{b}_{-\ell _{0}}\right) \mathcal{D}_{n}^{+}\left( \widehat{b} _{-\ell _{0}};i\right) \sum_{k=1}^{n}\widetilde{P}_{k}\left( \widehat{b} _{-\ell _{0}}\right) \widehat{u}_{k}\mathcal{J}_{k}\left( \widetilde{x}% _{i}\right) \end{equation*}% and consider a feasible version of the Khmaladze transformation: \begin{equation*} \widetilde{\mathcal{M}}_{n}\left( x\right) =:\frac{1}{n^{1/2}}% \sum_{i=1}^{n}\widehat{v}_{i}\mathcal{I}_{i}\left( x\right) , \end{equation*}% Just like in Section \ref{sec:Khmlinear}, we have the following result. \begin{theorem} \label{M_nest-nonlinear} Assuming that $H_{0}$ holds true, under Conditions $% C1-C3$, we have that \begin{equation*} \widetilde{\mathcal{M}}_{n}\left( x\right) \overset{weakly}{% \Rightarrow }\mathcal{U}\left( x\right) \text{.} \end{equation*} \end{theorem} This methodology can be generalized, of course, to the situation when the constrained estimate belongs to the intersection of several boundary surfaces. In this case we would express several parameter values as functions of other parameters and plug these functions into the linear approximation obtaining a nonlinear expression in the remaining parameters. We would define the derivative of the new approximation and adjust the Khmaladze transformation to account now for several nonlinear equality constraints. The rest of the methodology and the asymptotic result would remain the same as above. \subsection{\textbf{COMPUTATIONAL ISSUES}} $\left. {}\right. $ This section is devoted at how we can compute our statistic. In view of the CUSUM interpretation, we shall rely on the standard recursive residuals. We will illustrate the computational issues using notations from the case of all linear constraints but, of course, the methodology in the case of non-linear constraints will be the same. Note that since $f\left( x\right) $ is continuous the probability of a tie is zero, so that we can always consider the case $x_{i}<x_{i+1}$. Now with this view we have that \begin{equation*} \mathcal{M}_{n}\left( x\right) =:\frac{1}{n^{1/2}}\sum_{i=1}^{n}v_{i}% \mathcal{I}_{i}\left( x\right) ; \end{equation*}% can be written with $v_{i}$ replaced by $v_{i}=u_{i}-\boldsymbol{P}% _{i}^{\prime }A_{n}^{+}\left( x\right) C_{n}\left( x\right) $ and now \begin{equation*} \left( \frac{1}{n}\sum_{k=1}^{n}\boldsymbol{P}_{k}\boldsymbol{P}_{k}^{\prime }\mathcal{J}_{k}\left( x_{i}\right) \right) ^{+}\frac{1}{n}\sum_{k=1}^{n}% \boldsymbol{P}_{k}u_{k}\mathcal{J}_{k}\left( x_{i}\right) =:A_{n,i}^{+}C_{n,i}\text{.} \end{equation*} Then from a computational point of view is worth observing that \begin{equation*} A_{n,k}^{+}=A_{n,k+1}^{+}-\frac{A_{n,k+1}^{+}\boldsymbol{P}_{k}\boldsymbol{P}% _{k}^{\prime }A_{n,k+1}^{+}}{n+\boldsymbol{P}_{k}^{\prime }A_{n,k+1}^{+}% \boldsymbol{P}_{k}} \end{equation*}% and% \begin{equation*} A_{n.k}^{+}C_{n,k}=A_{n,k+1}^{+}C_{n.k+1}+A_{n,k}^{+}\boldsymbol{P}% _{k}\left( u_{k}-\boldsymbol{P}_{k}^{\prime }A_{n,k+1}^{+}C_{n,k+1}\right) \end{equation*}% see \cite{BrownDurbinEvans} for similar arguments. Alternatively, we could have considered the Cusum of backward recursive residuals, in which case we would have use the computational formulae,% \begin{equation*} \bar{A}_{n,k+1}^{+}=\bar{A}_{n,k}^{+}-\frac{\bar{A}_{n,k}^{+}\boldsymbol{P}% _{k+1}\boldsymbol{P}_{k+1}^{\prime }\bar{A}_{n,k}^{+}}{n+\boldsymbol{P}% _{k+1}^{\prime }\bar{A}_{n,k}^{+}\boldsymbol{P}_{k+1}} \end{equation*}% and% \begin{equation*} \overline{A}_{n,k+1}^{+}C_{n,k+1}=\overline{A}_{n,k}^{+}C_{n,k}+A_{n,k+1}^{+}% \boldsymbol{P}_{k+1}\left( u_{k+1}-\boldsymbol{P}_{k+1}^{\prime }\overline{A}% _{n,k}^{+}C_{n,k}\right) . \end{equation*} Of course in the previous formulas one would replace $u_{i}$ by $\widehat{u}% _{i}$ or $y_{i}$. \subsection{\textbf{POWER AND LOCAL ALTERNATIVES}} \label{sec:power} $\left. {}\right. $ Here we describe the local alternatives for which the tests based on $% \widetilde{\mathcal{M}}_{n}\left( x\right) $ have no trivial power. For that purpose, assume that the true model is such that \begin{equation*} H_{a}=m\left( x\right) +n^{-1/2}m_{1}\left( x\right) \text{,} \end{equation*}% where $m\left( x\right) \in \mathcal{M}_0$ and $m_{1}\left( x\right) $ is incompatible with $\mathcal{M}_0$ at least on some subset $\mathcal{X}% _{1}\subset \mathcal{X=}\left( 0,1\right) $ with the Lebesgue measure strictly greater than zero. Then, we have the result of Proposition \ref{Loc_alt}. \begin{proposition} \label{Loc_alt}Assuming Conditions $C1-C3$, under $H_{a}$ we have that \begin{equation*} \widetilde{\mathcal{M}}_{n}\left( x\right) +\mathcal{L}\left( x\right) \overset{weakly}{\Rightarrow }\mathcal{U}\left( x\right) \text{, \ \ }x\in % \left[ 0,1\right] \end{equation*}% where \begin{equation*} \mathcal{L}\left( x\right) =\int_{0}^{x}\left\{ m_{1}\left( v\right) -% \mathbf{P}_{L}^{\prime }\left( v\right) A_{L}^{+}\left( v\right) \int_{v}^{1}% \mathbf{P}_{L}\left( w\right) m_{1}\left( w\right) f_{X}\left( w\right) dw\right\} f_{X}\left( v\right) dv\text{.} \end{equation*} \end{proposition} One consequence of Proposition \ref{Loc_alt} is that not only tests based on $\widetilde{\mathcal{M}}_{n}\left( x\right) $ are consistent since $\mathcal{% L}\left( x\right) $ is a nonzero function, but that they also have a nontrivial power against local alternatives converging to the null at the \textquotedblleft parametric\textquotedblright\ rate $n^{-1/2}$. \paragraph*{\textbf{Role of constraints for the power of the test}} In our discussion of the Khmaladze's transformation we stated that in that transformation one has to enforce the binding constraints obtained in the constrained estimation under the null hypothesis. Here we present a simple example to illustrate what happens to the power of the test if such constraints are not enforced. To keep arguments simple, consider the case where $L=2$ when we use the B-splines to approximate the function and the null hypothesis, and take the null hypothesis to be that of an increasing regression function. That is, we employ the approximation $m_{\mathcal{B}}\left( x;2\right) = \beta _{1}p_{1,2}\left( x;q\right) + \beta _{2}p_{2 ,2}\left( x;q\right)$ and rewrite it in the following way: \begin{eqnarray*} m_{\mathcal{B}}\left( x;2\right) &=&\alpha _{1}\left( p_{1,2}\left( x;q\right) +p_{2,2}\left( x;q\right) \right) +\alpha _{2}p_{2,2}\left( x;q\right) \\ &=&\alpha _{1}p\left( x;q\right) +\alpha _{2}p_{2,2}\left( x;q\right) , \end{eqnarray*}% where $\alpha _{1}=\beta _{1}$ $\alpha _{2}=\beta _{2}-\beta _{1}$ and $% p\left( x;q\right) =:p_{1,2}\left( x;q\right) +p_{2,2}\left( x;q\right) $. Of course, because of the B-splines properties, we have $p\left( x;q\right)=1$ but for the sake of presenting a more general argument, we will keep the more general notation $p\left( x;q\right)$. The null hypothesis is then written as $\alpha _{2}\geq 0$. Suppose the null is not true and the true value is $\alpha _{2}<0$. When we estimate the model, we should expect that our estimator is of the form $\left( \widehat{% \alpha }_{1},0\right) $. We then define constrained residuals as \begin{equation*} \widehat{u}_{i}=y_{i}-\widehat{\alpha }_{1}p\left( x;q\right) . \end{equation*}% According to our methodology, at each step of the transformation we should only be projecting the residuals on $p\left( x;q\right)$. Let us analyze what happens if in the transformation we use as our $\mathbf{P% }_{k}$ that comes from the vector $\left( p\left( x_{k};q\right) ,{p}% _{2,2}\left( x_{k};q\right) \right) $ instead of using only $p\left( x_{k};q\right)$. In other words, we use the test statistic \begin{equation*} \widetilde{\mathcal{M}}_{n}\left( x\right) =:\frac{1}{n^{1/2}}\sum_{i=1}^{n}% \widehat{v}_{i}\mathcal{I}_{i}\left( x\right) \text{,} \end{equation*}% where $\widehat{u}_{i}=y_{i}-\widehat{\alpha }_{1}p\left( x;q\right) $ and the transformation is defined as follows: \begin{eqnarray*} \widehat{v}_{i} &=&\widehat{u}_{i}-\boldsymbol{P}_{i}^{\prime }A_{n,i}^{+}C_{n,i} \\ C_{n,i} &=&n^{-1}\sum_{k=i}^{n}\boldsymbol{P}_{k}\widehat{u}_{k};~\ \ \ A_{n,i}=n^{-1}\sum_{k=i}^{n}\boldsymbol{P}_{k}\boldsymbol{P}_{k}^{\prime }% \text{.} \end{eqnarray*} Rewrite the test statistic as \begin{equation*} \widetilde{\mathcal{M}}_{n}\left( x\right) =\frac{1}{n^{1/2}}% \sum_{i=1}^{n}v_{i}\mathcal{I}_{i}\left( x\right) +\frac{1}{n^{1/2}}% \sum_{i=1}^{n}\left( \widehat{v}_{i}-v_{i}\right) \mathcal{I}_{i}\left( x\right) \end{equation*}% and notice that the first term on the right-hand side converges to the Brownian motion. The second term is negligible under the null, as established earlier, and partly this is due to the result of Lemma \ref{bias} in Appendix B. It is rather obvious that the power of the test, as usual, comes from that term. For the test to have power we need that under the alternative the term $\frac{1}{n}\sum_{i=1}^{n}\left( \widehat{v}% _{i}-v_{i}\right) \mathcal{I}_{i}\left( x\right) $ converges somewhere different than zero on a set of a positive measure, so that once we multiply it by $n^{1/2}$, it diverges to $\pm \infty $. We have that% \begin{eqnarray} \widehat{u}_{i}-u_{i} &=&\widehat{m}_{\mathcal{B}}\left( x_{i};2\right) -m_{% \mathcal{B}}\left( x_{i};2\right) +\left( m_{\mathcal{B}}\left( x_{i};2\right) -m(x)\right) \notag \\ &=&\left( \widehat{\alpha }_{1}-\alpha _{1}\right) p\left( x_{i};q\right) -\alpha _{2}{p}_{2,2}\left( x_{i};q\right) +\left( m_{\mathcal{B}}\left( x_{i};2\right) -m(x)\right) . \label{eq:eqpower} \end{eqnarray} Lemma \ref{bias} implies that when the transformation uses $\mathbf{P}_{k}$ that comes from the vector $\left( p\left( x_{k};q\right) ,{p}_{2,2}\left( x_{k};q\right) \right) $, any linear combination $\boldsymbol{% \mathring{p}}\left( x\right) =:a_{1}p\left( x;q\right) +a_{2}{p}_{2,2}\left( x;q\right) $ satisfies $\left( \mathcal{T}_{n}\mathcal{B}_{n,2}\right) \left( x\right) =0$, where $\mathcal{B}_{n,2}\left( x\right) =n^{-1}\sum_{k=1}^{n}\boldsymbol{\mathring{p}}\left( x_{k}\right) \mathcal{I}% _{k}\left( x\right) $. This, taking into account representation (\ref{eq:eqpower}), implies that the part of $n^{-1}\sum_{i=1}^{n}\left( \widehat{% v}_{i}-v_{i}\right) \mathcal{I}_{i}\left( x\right) $ corresponding to $% \left( \widehat{\alpha }_{1}-\alpha _{1}\right) p\left( x_{i};q\right) -\alpha _{2}{p}_{2,2}\left( x_{i};q\right) $ in (\ref{eq:eqpower}) will be zero, and its part corresponding to the bias term $m_{\mathcal{B}}\left( x_{i};2\right) -m(x)$ in (\ref{eq:eqpower}) will be asymptotically negligible even when multiplied by $n^{1/2}$. Thus, the power of the test equals the trivial power because we failed to embed the binding constraint in the transformation. On the other hand, if in the transformation we use only the ``explanatory'' polynomial $p\left( x;q\right) $, we would now have that the contribution due to $\left( \widehat{\alpha }_{1}-\alpha _{1}\right) p\left( x_{i};q\right) -\alpha _{2}{p}_{2,2}\left( x_{i};q\right) $ in $% n^{-1}\sum_{i=1}^{n}\left( \widehat{v}_{i}-v_{i}\right) \mathcal{I}% _{i}\left( x\right) $ is exactly as that from $\alpha _{2}{p}_{2,2}\left( x_{i};q\right) $ and it is not equal to zero as ${p}_{2,2}\left( x_{i};q\right) $ is not in the space generated by $p\left( x_{i};q\right) $. In other words, % \begin{equation*} \frac{1}{n}\sum_{k=1}^{n}\left\{ {p}_{2,2}\left( x_{k};q\right) -\boldsymbol{% P}_{k}^{\prime }A_{n,k}^{+}\frac{1}{n}\sum_{j=1}^{n}\boldsymbol{P}_{j}{p}% _{2,2}\left( x_{j};q\right) \mathcal{J}_{j}\left( \widetilde{x}_{k}\right) \right\} \mathcal{I}_{k}\left( x\right) \neq 0 \end{equation*}% unless ${p}_{2,2}\left( x_{i};q\right) $ is proportional to ${p}_{2,2}\left( x_{i};q\right) $, which is not the case. \section{\textbf{BOOTSTRAP ALGORITHM}} \label{sec:bootstrap} One of our motivations to introduce a bootstrap algorithm for our test(s) is that although it is pivotal, our Monte Carlo experiment suggests that they suffer from small sample biases. When the asymptotic distribution does not provide a good approximation to the finite sample one, a standard approach to improve its performance is to employ bootstrap algorithms, as they provide small sample refinements. In fact, our Monte Carlo simulation does suggest that the bootstrap, to be described below, does indeed give a better finite sample approximation. The notation for the bootstrap is as usual and we shall implement the fast algorithm of WARP by \cite{GiacominiPW} in the Monte Carlo experiment. The bootstrap is based on the following \emph{3 STEPS}. \begin{description} \item[\emph{STEP 1}] Compute the unconstrained residuals% \begin{equation*} \widetilde{u}_{i}=y_{i}-\widetilde{m}_{\mathcal{B}}\left( x_{i};L\right) \text{, \ \ \ }i=1,...,n \end{equation*}% with $\widetilde{m}_{\mathcal{B}}\left( x_{i};L\right) $ as defined in $% \left( \ref{uncons}\right) $. \vskip 0.1in \item[\emph{STEP 2}] Obtain a random sample of size\emph{\ }$n$ from the empirical distribution of $\left\{ \widetilde{u}_{i}-\frac{1}{n}% \sum_{i=1}^{n}\widetilde{u}_{i}\right\} _{i=1}^{n}$. Denote such a sample as $\left\{ u_{i}^{\ast }\right\} _{i=1}^{n}$ and compute the bootstrap analogue of the regression model using $\widehat{m}_{\mathcal{B}}\left( x_{i};L\right) $, that is% \begin{equation} y_{i}^{\ast }=\widehat{m}_{\mathcal{B}}\left( x_{i};L\right) +u_{i}^{\ast }% \text{, \ }i=1,...,n\text{.} \label{reg_boot} \end{equation} \vskip0.1in \item[\emph{STEP 3}] Compute the bootstrap analogue of $\widetilde{\mathcal{M% }}_{n}\left( x\right) $ as \begin{equation*} \widetilde{\mathcal{M}}_{n}^{\ast }\left( x\right) =:\frac{1}{n^{1/2}}% \sum_{i=1}^{n}\widehat{v}_{i}^{\ast }\mathcal{I}_{i}\left( x\right) \end{equation*}% where% \begin{equation*} \widehat{v}_{i}^{\ast }=\widehat{u}_{i}^{\ast }-\boldsymbol{P}_{i}^{\prime }A_{n,i}^{+}C_{n,i}^{\ast };\text{ \ }C_{n,i}^{\ast }=:C_{n,i}^{\ast }\left( \widetilde{x}_{i}\right) =\frac{1}{n}\sum_{k=1}^{n}\boldsymbol{P}_{k}% \widehat{u}_{k}^{\ast }\mathcal{J}_{k}\left( \widetilde{x}_{i}\right) \end{equation*}% with $\widehat{u}_{i}^{\ast }=y_{i}^{\ast }-\boldsymbol{P}_{i}^{\prime }A_{n}^{+}\left( 0\right) C_{n}^{\ast }\left( 0\right) $, $i=1,...,n$. \end{description} \begin{theorem} \label{Mb_n}Under Conditions $C1-C3$, we have that for any continuous function $g:\mathbb{R\rightarrow R}^{+}$, (in probability), \begin{equation*} g\left( \widetilde{\mathcal{M}}_{n}^{\ast }\left( x\right) \right) \overset{d% }{\Rightarrow }g\left( \mathcal{U}\left( x\right) \right) \text{.} \end{equation*} \end{theorem} Finally, we can replace $\widehat{u}_{i}^{\ast }$ by $y_{i}^{\ast }$ in the computation of $\widetilde{\mathcal{M}}_{n}^{\ast }\left( x\right) $. That is, \begin{corollary} \label{cor:bootstrap} Under Conditions $C1-C3$, we have that \begin{equation*} \widetilde{\mathcal{M}}_{n}^{\ast }\left( x\right) -\widetilde{\widetilde{% \mathcal{M}}}_{n}^{\ast }\left( x\right) =0\text{,} \end{equation*}% where% \begin{equation*} \widetilde{\widetilde{\mathcal{M}}}_{n}^{\ast }\left( x\right) =:\frac{1}{% n^{1/2}}\sum_{i=1}^{n}\left( y_{i}^{\ast }-\boldsymbol{P}_{i}^{\prime }A_{n,i}^{+}\frac{1}{n}\sum_{k=1}^{n}\boldsymbol{P}_{k}y_{k}^{\ast }\mathcal{% J}_{k}\left( x_{i}\right) \right) \mathcal{I}_{i}\left( x\right) \text{.} \end{equation*} \end{corollary} The proof of Corollary \ref{cor:bootstrap} is immediate by Lemma \ref{bias} in the Appendix and therefore is omitted. \section{\textbf{MONTE CARLO EXPERIMENTS AND EMPIRICAL EXAMPLES}} \label{sec:simulations_and_data} \subsection{\textbf{MONTE CARLO EXPERIMENTS}} $\left. {}\right. $ In this section we present the results of several computational experiments. All the results in this section are given for cubic splines with different number of knots. We present the results for \emph{B-splines} as well as for \emph{P-splines} with penalties on the second differences of coefficients. The penalty parameter is chosen by cross-validation in the unconstrained estimation. In the tables \textquotedblleft \emph{KS}\textquotedblright\ refers to the Kolmogorov-Smirnov test statistic, \textquotedblleft \emph{CvM}% \textquotedblright\ refers to the \emph{Cram\'{e}r-von Mises} test statistic and \textquotedblleft \emph{AD}\textquotedblright\ to the Anderson-Darling integral test statistic. All three test statistics are based on a Brownian bridge. $L^{\prime }+1$ denotes the number of equidistant knots (including the boundary points) on the interval of interest. For example, when $% L^{\prime }=6$ and the interval is $[0,1]$, we consider knots $% 0,1/6,1/3,1/2,2/3,5/6,1$. In the implementation of P-splines in simulations, every simulation draw will give a different cross-validation parameter (we use ordinary cross validation described in Eilers and Marx, $1996$). In our simulation results for each $L^{\prime }$ we use a modal value of these cross-validation parameters. In all the scenarios below \begin{equation*} X\sim \mathcal{U}[0,1],\quad U\sim \mathcal{N}(0,\sigma ^{2}),\quad U\perp X. \end{equation*} In Scenarios 1, 3-5 the interval of interest if $[0,1]$ whereas in Scenario 2 of U-shape we consider individually intervals $[0,s_0]$ and $[s_0,1]$ with $s_0$ being the switch point. In the WARP bootstrap implementation, the demeaned residuals and $x$ are drawn independently. \vskip 0.1in \noindent \textbf{Scenario 1 (test for monotonicity)}. We take the following regression function: \begin{equation*} m\left( x\right) =x^{\frac{13}{4}}\text{,}\quad x\in \left[ 0,1\right] \text{% .} \end{equation*}% The results are summarized in Table \ref{MonteCarlo1}. \vskip 0.05in \begin{table}[tbp] \par \begin{center} \begin{tabular}{llllll} \hline\hline Setting & Method \hspace*{0.2in} & \multicolumn{2}{c}{B-splines} & \multicolumn{2}{c}{P-splines} \\ & & 10\% & 5\% & 10\% & 5\% \\ \hline $L^{\prime }=6$ & KS bootstrap & 0.113 & 0.054 & 0.0965 & 0.051 \\ $N=1000$ & CvM bootstrap & 0.1035 & 0.0475 & 0.093 & 0.0465 \\ $\sigma=0.25$ & AD bootstrap & 0.1065 & 0.054 & 0.1005 & 0.054 \\ \hline $L^{\prime }=9$ & KS bootstrap & 0.102 & 0.044 & 0.119 & 0.052 \\ $N=1000$ & CvM bootstrap & 0.101 & 0.048 & 0.11 & 0.0455 \\ $\sigma=0.25$ & AD bootstrap & 0.0985 & 0.042 & 0.1045 & 0.048 \\ \hline $L^{\prime }=14$ & KS bootstrap & 0.0945 & 0.043 & 0.105 & 0.0555\\ $N=1000$ & CvM bootstrap & 0.098 & 0.0425 & 0.0955 & 0.045\\ $\sigma=0.25$ & AD bootstrap & 0.093 & 0.043 & 0.096 & 0.049 \\ \hline $L^{\prime }=19$ & KS bootstrap & 0.089 & 0.0485 & 0.1055 & 0.059 \\ $N=1000$ & CvM bootstrap & 0.105 & 0.0555 & 0.1035 & 0.0505 \\ $\sigma=0.25$ & AD bootstrap & 0.1085 & 0.0545 & 0.106 & 0.0495 \\ \hline \end{tabular} \vspace*{0.05in} \end{center} \caption{{\protect\small Tests for monotonically increasing regression function in Scenario 1. Rejection rates in 2000 simulations. $L^{\prime }+1$ denotes the number of equidistant knots on $[0,1]$. $N$ denotes the number of observations in each simulation. $\protect\sigma$ is the standard deviation in the error distribution.}} \label{MonteCarlo1} \end{table} \vskip 0.1in \noindent \textbf{Scenario 2 (test for U-shape)}. The regression function is defined as \begin{align*} m(x) & =10 \left(\log(1+x)-0.33\right)^2. \end{align*} The graph of this function is U-shaped with the switch point at $% s_0=e^{0.33}-1$. In simulations $s_0$ is taken to be known. The results are summarized in Table \ref{MonteCarloU2}. We use two different \emph{B-splines} -- one on $[0,s_{0}]$ and the other on $[s_{0},1]$. We analyze the properties of the testing procedure in two approaches. In the first approach additional restrictions are imposed for the two B-splines to be joined continuously at $s_{0}$, and in the second approach these two B-splines are joined smoothly at $s_{0}$ (see details in Example 2). \begin{table}[tbp] \par \begin{center} \begin{tabular}{llllllllll} \hline\hline & & \multicolumn{4}{c}{Continuously joined} & \multicolumn{4}{c}{Smoothly joined} \\ \hline Setting & Method \hspace*{0.2in} & \multicolumn{2}{c}{B-splines} & \multicolumn{2}{c}{P-splines} & \multicolumn{2}{c}{B-splines} & \multicolumn{2}{c}{P-splines} \\ & & 10\% & 5\% & 10\% & 5\% & 10\% & 5\% & 10\% & 5\% \\ \hline $L^{\prime }=4$ & KS & 0.0935 & 0.048 & 0.113 & 0.051 & 0.098 & 0.062 & 0.1105 & 0.055 \\ $N=1000$ & CvM & 0.106 & 0.046 & 0.1 & 0.0515 & 0.101 & 0.0575 & 0.1005 & 0.05 \\ $\sigma=0.25$ & AD & 0.107 & 0.048 & 0.098 & 0.055 & 0.1015 & 0.0615 & 0.094 & 0.0505\\ \vspace*{0.02in} & & & & & & & & & \\ $L^{\prime }=6$ & KS & 0.1105 & 0.0555 & 0.112 & 0.051 & 0.107 & 0.0575 & 0.0935 & 0.0495\\ $N=1000$ & CvM & 0.101 & 0.0585 & 0.099 & 0.0525 & 0.1 & 0.0575 & 0.0955 & 0.048 \\ $\sigma=0.25$ & AD & 0.1015 & 0.0565 & 0.098 & 0.047 & 0.1055 & 0.055 & 0.0965 & 0.0485 \\ \hline \end{tabular}% \end{center} \caption{{\protect\small Tests for U-shape with the switch at $% s_0=e^{0.33}-1 $ in Scenario 2. Rejection rates in 2000 simulations. $% L^{\prime }+1$ denotes the number of equidistant knots on each subinterval $% [0,s_0]$ and $[s_0,1]$ . $N$ denotes the number of observations in each simulation. $\protect\sigma $ is the standard deviation in the error distribution.}} \label{MonteCarloU2} \end{table} \vskip 0.1in \noindent \textbf{Scenario 3 (analysis of power of the test)}. Take the regression function \begin{align*} m(x) & =(10x-0.5)^3 - \exp(-100(x-0.25)^2))\cdot \mathcal{I}(x<0.5) \\ & + (0.1(x-0.5)-\exp(-100(x-0.25)^2))\cdot \mathcal{I}(x>=0.5) \end{align*} and depicted in Figure \ref{fig:dip}. As expected, the power of the test depends on the variance of the error. The results are summarized in Table \ref{MonteCarlodip}. \begin{figure}[ht!] \includegraphics[width=0.4\textwidth]{dip.png} \caption{{\protect\small Plot of the regression function in Scenario 3.}} \label{fig:dip} \end{figure} \begin{table}[tbp] \par \begin{center} \begin{tabular}{llllll} \hline\hline Setting & Method \hspace*{0.2in} & \multicolumn{2}{c}{B-splines} & \multicolumn{2}{c}{P-splines} \\ & & 10\% & 5\% & 10\% & 5\% \\ \hline $L^{\prime }=6$ & KS bootstrap & 1 & 0.998 & 1 & 0.9885 \\ $N=1000$ & CvM bootstrap & 0.9155 & 0.7335 & 0.9975 & 0.9895 \\ $\sigma=0.5$ & AD bootstrap & 0.985 & 0.9375 & 0.9995 & 0.998 \\ \vspace*{0.05in} & & & & & \\ $L^{\prime }=9$ & KS bootstrap & 0.93 & 0.823 & 0.999 & 0.998 \\ $N=1000$ & CvM bootstrap & 0.862 & 0.766 & 0.998 & 0.9935 \\ $\sigma=0.5$ & AD bootstrap & 0.9195 & 0.8085 & 0.9985 & 0.9935 \\ \vspace*{0.05in} & & & & & \\ $L^{\prime }=12$ & KS bootstrap & 0.864 & 0.8175 & 0.996 & 0.9885 \\ $N=1000$ & CvM bootstrap & 0.8505 & 0.7895 & 0.989 & 0.9725\\ $\sigma=0.5$ & AD bootstrap & 0.8655 & 0.799 & 0.9885 & 0.974 \\ \vspace*{0.05in} & & & & & \\ $L^{\prime }=19$ & KS bootstrap & 0.639 & 0.5375 & 0.9815 & 0.9515 \\ $N=1000$ & CvM bootstrap & 0.5295 & 0.395 & 0.9435 & 0.8795 \\ $\sigma=0.5$ & AD bootstrap & 0.571 & 0.428 & 0.9445 & 0.892 \\ \hline $L^{\prime }=6$ & KS bootstrap & 1 & 1 & 1 & 1 \\ $N=1000$ & CvM bootstrap & 1 & 1 & 1 & 1 \\ $\sigma=0.25$ & AD bootstrap & 1 & 1 & 1 & 1 \\ \vspace*{0.05in} & & & & & \\ $L^{\prime }=9$ & KS bootstrap & 1 & 1 & 1 & 1 \\ $N=1000$ & CvM bootstrap & 1 & 1 & 1 & 1 \\ $\sigma=0.25$ & AD bootstrap & 1 & 1 & 1 & 1 \\ \vspace*{0.05in} & & & & & \\ $L^{\prime }=12$ & KS bootstrap & 0.9995 & 0.9995 & 1 & 1 \\ $N=1000$ & CvM bootstrap & 0.996 & 0.9945 & 1 & 1 \\ $\sigma=0.25$ & AD bootstrap & 1 & 0.996 & 1 & 1 \\ \vspace*{0.05in} & & & & & \\ $L^{\prime }=19$ & KS bootstrap & 0.996 & 0.986 & 1 & 1 \\ $N=1000$ & CvM bootstrap & 0.9155 & 0.836 & 1 & 1 \\ $\sigma=0.25$ & AD bootstrap & 0.955 & 0.891 & 1 & 1 \\ \hline \end{tabular} \vspace*{0.05in} \end{center} \caption{{\protect\small Tests for monotonicity in Scenario 3. Rejection rates in 2000 simulations. $L^{\prime }+1$ denotes the number of equidistant knots on $[0,1]$. $N$ denotes the number of observations in each simulation. $\protect\sigma$ is the standard deviation in the error distribution. }} \label{MonteCarlodip} \end{table} The power of monotonicity tests based on this regression function is considered in \cite{GhosalSenVV} and a similar regression function is considered in \cite{HallHeckman}. Note that \cite{GhosalSenVV} considers smaller sample sizes and also smaller standard deviation of noise with $\sigma=0.1$. \vskip 0.1in \noindent \textbf{Scenario 4 (analysis of power of the test)}. The regression function \begin{equation*} m(x) = x + 0.415\exp(-ax^2), \quad a>0. \end{equation*} and depicted in Figure \ref{fig:dipother}. The left-hand side graph in Figure \ref{fig:dipother} is for the case $a=50$ and the right-hand side graph in Figure \ref{fig:dipother}. In the latter case the non-monotonicity dip is smaller. These situations are considered to be challenging for monotonicity tests as these functions are somewhat close to the set of monotone functions (in any conventional metric). As expected, the power of the test depends on the value of parameter $a$ and also depends on the variance of the error. The results are summarized in Table \ref{MonteCarlodipother}. \begin{figure}[!h] \begin{center} \begin{minipage}{\linewidth} \begin{minipage}{0.5\linewidth} \includegraphics[width=0.95\linewidth]{dip2.png} \end{minipage} \begin{minipage}{0.5\linewidth} \includegraphics[width=0.95\linewidth]{dip3.png} \end{minipage} \end{minipage} \end{center} \caption{{\protect\small Plot of the regression function in Scenario 4. The left-hand side graph is for $a=50$ and the right-hand side graph is for $% a=20 $.}} \label{fig:dipother} \end{figure} \begin{table}[tbp] \par \begin{center} \begin{tabular}{llllllllll} \hline\hline & & \multicolumn{4}{c}{$a=50$} & \multicolumn{4}{c}{$a=20$} \\ \hline Setting & Method \hspace*{0.2in} & \multicolumn{2}{c}{B-splines} & \multicolumn{2}{c}{P-splines} & \multicolumn{2}{c}{B-splines} & \multicolumn{2}{c}{P-splines} \\ & & 10\% & 5\% & 10\% & 5\% & 10\% & 5\% & 10\% & 5\% \\ \hline $L^{\prime }=6$ & KS & 0.382 & 0.2425 & 0.575 & 0.423 & 0.25 & 0.139 & 0.338 & 0.216 \\ $N=1000$ & CvM & 0.3385 & 0.2245 & 0.5165 & 0.384 & 0.2465 & 0.1335 & 0.311 & 0.212 \\ AD $\sigma=0.5$ & AD & 0.434 & 0.228 & 0.6015 & 0.457 & 0.2475 & 0.1395 & 0.3355 & 0.221 \\ \vspace*{0.02in} & & & & & & & & & \\ $L^{\prime }=9$ & KS & 0.381 & 0.2545 & 0.572 & 0.4655 & 0.2 & 0.124 & 0.3405 & 0.221\\ $N=1000$ & CvM & 0.363 & 0.2355 & 0.536 & 0.3915 & 0.2055 & 0.132 & 0.3385 & 0.206 \\ $\sigma=0.5$ & AD & 0.4735 & 0.3165 & 0.6095 & 0.4555 & 0.228 & 0.145 & 0.348 & 0.206 \\ \vspace*{0.02in} & & & & & & & & & \\ $L^{\prime }=12$ & KS & 0.3975& 0.27 & 0.5995 & 0.4745 & 0.22 & 0.1375 & 0.343 & 0.2185 \\ $N=1000$ & CvM & 0.3905 & 0.2565 & 0.5545 & 0.4075 & 0.2335 & 0.152 & 0.3215 & 0.2065 \\ $\sigma=0.5$ & AD & 0.486 & 0.3505 & 0.614 & 0.5035 & 0.2525 & 0.1625 & 0.3415 & 0.2195 \\ \vspace*{0.02in} & & & & & & & & & \\ $L^{\prime }=19$ & KS & 0.4045 & 0.284 & 0.5905 & 0.476 & 0.2185 & 0.1235 & 0.3455 & 0.2265 \\ $N=1000$ & CvM & 0.418 & 0.308 & 0.5525 & 0.414 & 0.232 & 0.145 & 0.318 & 0.1995 \\ $\sigma=0.5$ & AD & 0.497 & 0.37 & 0.6125 & 0.4915 & 0.2485 & 0.155 & 0.341 & 0.2205 \\ \hline $L^{\prime }=6$ & KS & 0.9 & 0.8405 & 0.986 & 0.9625 & 0.5795 & 0.4605 & 0.756 & 0.6415 \\ $N=1000$ & CvM & 0.8295 & 0.7025 & 0.9615 & 0.913 & 0.5895 & 0.4195 & 0.741 & 0.626 \\ $\sigma=0.25$ & AD & 0.939 & 0.854 & 0.9835 & 0.9665 & 0.608 & 0.4505 & 0.7275 & 0.6295 \\ \vspace*{0.02in} & & & & & & & & & \\ $L^{\prime }=9$ & KS & 0.919 & 0.8385 & 0.986 & 0.9685 & 0.484 & 0.3355 & 0.6805 & 0.5645 \\ $N=1000$ & CvM & 0.8355 & 0.7275 & 0.966 & 0.911 & 0.46 & 0.347 & 0.6495 & 0.5065 \\ $\sigma=0.25$ & AD & 0.937 & 0.8695 & 0.99 & 0.9615 & 0.4995 & 0.374 & 0.6655 & 0.512 \\ \vspace*{0.02in} & & & & & & & & & \\ $L^{\prime }=12$ & KS & 0.9235 & 0.842 & 0.9865 & 0.9705 & 0.461 & 0.334 & 0.6785 & 0.5585 \\ $N=1000$ & CvM & 0.863 & 0.756 & 0.97 & 0.9325 & 0.436 & 0.318 & 0.6525 & 0.4965 \\ $\sigma=0.25$ & AD & 0.951 & 0.8895 & 0.9865 & 0.9705 & 0.492 & 0.3575 & 0.658 & 0.517 \\ \vspace*{0.02in} & & & & & & & & & \\ $L^{\prime }=19$ & KS & 0.925 & 0.8505 & 0.9865 & 0.974 & 0.4835 & 0.3505 & 0.698 & 0.585 \\ $N=1000$ & CvM & 0.8835 & 0.802 & 0.9745 & 0.9355 & 0.4595 & 0.345 & 0.6625 & 0.495 \\ $\sigma=0.25$ & AD & 0.941 & 0.8985 & 0.987 & 0.9695 & 0.486 & 0.3665 & 0.666 & 0.5105 \\ \hline $L^{\prime }=6$ & KS & 1 & 1 & 1 & 1 & 1 & 0.998 & 1 & 1 \\ $N=1000$ & CvM & 1 & 1 & 1 & 1 & 0.9985 & 0.9965 & 1 & 1 \\ $\sigma=0.1$ & AD & 1 & 1 & 1 & 1 & 0.9995 & 0.9975 & 1 & 1 \\ \vspace*{0.02in} & & & & & & & & & \\ $L^{\prime }=9$ & KS & 1 & 1 & 1 & 1 & 0.9935 & 0.9875 & 1 & 1 \\ $N=1000$ & CvM & 1 & 1 & 1 & 1 & 0.9915 & 0.9825 & 1 & 0.9955 \\ $\sigma=0.1$ & AD & 1 & 1 & 1 & 1 & 0.9915 & 0.9865 & 1 & 0.997 \\ \vspace*{0.02in} & & & & & & & & & \\ $L^{\prime }=12$ & KS & 1 & 1 & 1 & 1 & 0.968 & 0.9535 & 0.9995 & 0.998\\ $N=1000$ & CvM & 1 & 1 & 1 & 1 & 0.9539 & 0.932 & 0.9975 & 0.995 \\ $\sigma=0.1$ & AD & 1 & 1 & 1 & 1 & 0.969 & 0.951 & 0.998 & 0.997 \\ \vspace*{0.02in} & & & & & & & & & \\ $L^{\prime }=19$ & KS & 1 & 1 & 1 & 1 & 0.973 & 0.9515 & 0.9995 & 0.9995 \\ $N=1000$ & CvM & 1 & 1 & 1 & 1 & 0.9525 & 0.9285 & 0.9985 & 0.9955\\ $\sigma=0.1$ & AD & 1 & 1 & 1 & 1 & 0.966 & 0..948 & 0.999 & 0.997\\ \hline \end{tabular}% \end{center} \caption{{\protect\small Tests for monotonicity in Scenario 4. Rejection rates in 2000 simulations. $L^{\prime }+1$ denotes the number of equidistant knots on $[0,1]$ (including 0 and 1). $N$ denotes the number of observations in each simulation. $\protect\sigma $ is the standard deviation in the error distribution.}} \label{MonteCarlodipother} \end{table} The power of monotonicity tests based on this regression function is examined in \cite{GhosalSenVV} and a similar regression function was considered in \cite{Bowman_etal}. Note that \cite{GhosalSenVV} uses smaller sample sizes and also only $a=50$ and $\sigma=0.1$ to analyze power implications. \vskip 0.1in \noindent \textbf{Scenario 5 (test for log-convexity)}. We take the following regression function: \begin{equation*} m\left( x\right) =\exp(x^2), \quad x\in \left[ 0,1\right] \text{% .} \end{equation*}% The results are summarized in Table \ref{MonteCarlo:logconvex}. In this case, the results for P-splines are the same as for B-splines as the cross-validation criterion indicated 0 as the optimal penalty parameter in the overwhelming majority of simulations. \vskip 0.05in \begin{table}[tbp] \par \begin{center} \begin{tabular}{llllll} \hline\hline Setting & Method \hspace*{0.2in} & \multicolumn{2}{c}{B-splines} \\ & & 10\% & 5\% \\ \hline $L^{\prime }=6$ & KS bootstrap & 0.1045 & 0.0535 \\ $N=1000$ & CvM bootstrap & 0.1015 & 0.054 \\ $\sigma=0.25$ & AD bootstrap & 0.1015 & 0.0495 \\ \hline $L^{\prime }=9$ & KS bootstrap & 0.098 & 0.0435 \\ $N=1000$ & CvM bootstrap & 0.1 & 0.0445 \\ $\sigma=0.25$ & AD bootstrap & 0.1 & 0.048 \\ \hline $L^{\prime }=12$ & KS bootstrap & 0.1135 & 0.053 \\ $N=1000$ & CvM bootstrap & 0.0935 & 0.0465 \\ $\sigma=0.25$ & AD bootstrap & 0.095 & 0.0495 \\ \hline \end{tabular} \vspace*{0.05in} \end{center} \caption{{\protect\small Tests for log-convexity in Scenario 5. Rejection rates in 2000 simulations. $L^{\prime }+1$ denotes the number of equidistant knots on $[0,1]$. $N$ denotes the number of observations in each simulation. $\protect\sigma$ is the standard deviation in the error distribution.}} \label{MonteCarlo:logconvex} \end{table} \subsection{\textbf{APPLICATIONS}} $\left. {}\right.$ \noindent 1. \textbf{Hospital data} Here we use data on hospital finance for 332 hospitals in California in 2003.\footnote{% The dataset is from https://www.kellogg.northwestern.edu/faculty/dranove/\newline htm/dranove/coursepages/mgmt469.htm. The original dataset is for 333 hospitals but we had to remove the observation that had missing information about administrative expenses.} The data include many variables related to hospital finance and hospital utilization. We are interested in analyzing the effect of revenue derived from patients on administrative expenses. Figure \ref{fig:Hospital} is a scatter plot of the logarithm of patient revenue and the logarithm of administrative expenses with the fitted curve obtained using cubic \emph{B-splines} with $L^{\prime }+1=5$ uniform knots in the range of values of the log of patient revenue. The fitted cure is obtained under the monotonicity restriction. \begin{figure}[!h] \includegraphics[width=0.6\textwidth]{Hospital_fit.png} \caption{{\protect\small Hospital data. Plot of the log of administrative expenses and the log of patient revenue and the constrained (under monotonicity) fit by cubic B-splines with $L^{\prime }+1=5$ uniform knots in the domain of the log of patient revenue.}} \label{fig:Hospital} \end{figure} We conduct tests for the following hypotheses: (\textbf{a}) monotonicity; (% \textbf{b}) convexity; (\textbf{c}) monotonicity and convexity. In order to correct for heteroscedasticity of the errors, we estimate the scedastic function $\widehat{\sigma }^{2}(x)$ using residuals obtained in the unconstrained estimation using cubic \emph{B-splines} with the same set of knots. The scedastic function $\widehat{\sigma }^{2}(x)$ is estimated by regressing the logarithm of the squared unconstrained residuals on a linear combination of first-order \emph{B-splines} with $6$ knots in the domain of the log of patient revenue. We then consider the constrained residuals divided by $\widehat{\sigma }(x)$ when calculating \emph{KS}, \emph{CvM} and \emph{AD} test statistics and unconstrained residuals divided by $\widehat{\sigma }(x)$ when drawing bootstrap samples. After a bootstrap sample of residuals is drawn, we multiply each residual by the corresponding $\widehat{\sigma }(x)$ when generating a bootstrap sample of observations of the dependent variable. We implement the testing procedure by conducting the Khmaladze transformation both from the right end of the support (as is described theoretically in this paper) and from the left end of the support and obtained extremely similar results. More specifically, we only report results when the transformation is conducted from the right end of the support. In the case of $P$\emph{-splines}, we use the same \emph{B-spline} basis, take the second-order penalty and choose the penalization constant using the ordinary cross-validation criterion as in Eilers and Marx (1996). The penalty enters unconstrained optimization problems as well as constrained ones. Tables \ref{table:HospA}-\ref{table:HospC} present results of our testing. Namely, Table \ref{table:HospA} shows test statistics for the null hypothesis of the monotonically increasing regression function and also bootstrap critical values using both \emph{B-splines} and \emph{P-splines}. Table \ref{table:HospB} presents analogous results for the null hypothesis of convexity of the regression function. Table \ref{table:HospC} gives results for the joint null hypothesis of monotonicity and convexity. \begin{table}[tbp] \par \begin{center} \begin{tabular}{ccccccccc} \hline\hline \quad & \multicolumn{4}{c}{B-splines} & \multicolumn{4}{c}{P-splines} \\ \quad & Test statistic & \multicolumn{3}{c}{Bootstrap critical value} & Test statistic & \multicolumn{3}{c}{Bootstrap critical value} \\ Method & \quad & 10\% & 5\% & 1\% & \quad & 10\% & 5\% & 1\% \\ \hline KS & 1.0365 & 1.1801 & 1.3028 & 1.6393 & 0.7965 & 0.8502 & 0.9246 & 1.1044 \\ CvM & 0.2842 & 0.3428 & 0.45 & 0.7656 & 0.0614 & 0.1113 & 0.1356 & 0.2138 \\ AD & 1.5477 & 1.739 & 2.2068 & 3.5445 & 0.4691 & 0.664 & 0.8037 & 1.1031 \\ \hline \end{tabular} \end{center} \par \caption{{\protect\small Hospital data. Test statistics and bootstrap critical values under the null hypothesis of monotonicity of the regression function. Bootstrap critical values are from 400 bootstrap replications.}} \label{table:HospA} \end{table} \begin{table}[tbp] \par \begin{center} \begin{tabular}{ccccccccc} \hline\hline \quad & \multicolumn{4}{c}{B-splines} & \multicolumn{4}{c}{P-splines} \\ \quad & Test statistic & \multicolumn{3}{c}{Bootstrap critical value} & Test statistic & \multicolumn{3}{c}{Bootstrap critical value} \\ Method & \quad & 10\% & 5\% & 1\% & \quad & 10\% & 5\% & 1\% \\ \hline KS & 0.9064 & 1.3918 & 1.5949 & 1.9306 & 0.7678 & 1.3462 & 1.5503 & 1.9284 \\ CvM & 0.1188 & 0.5137 & 0.6783 & 1.1675 & 0.1302 & 0.4994 & 0.6404 & 1.1974 \\ AD & 0.7103 & 2.7328 & 3.6416 & 7.1427 & 0.7241 & 2.5773 & 3.3495 & 7.119 \\ \hline \end{tabular} \end{center} \par \caption{{\protect\small Hospital data. Test statistics and bootstrap critical values under the null hypothesis of convexity of the regression function. Bootstrap critical values are from 400 bootstrap replications.}} \label{table:HospB} \end{table} \begin{table}[tbp] \par \begin{center} \begin{tabular}{ccccccccc} \hline\hline \quad & \multicolumn{4}{c}{B-splines} & \multicolumn{4}{c}{P-splines} \\ \quad & Test statistic & \multicolumn{3}{c}{Bootstrap critical value} & Test statistic & \multicolumn{3}{c}{Bootstrap critical value} \\ Method & \quad & 10\% & 5\% & 1\% & \quad & 10\% & 5\% & 1\% \\ \hline KS & 1.218 & 1.4347 & 1.6063 & 1.9643 & 1.092 & 1.3568 & 1.6256 & 1.903 \\ CvM & 0.2497 & 0.5488 & 0.7375 & 1.4689 & 0.1895 & 0.4785 & 0.7751 & 1.221 \\ AD & 1.2513 & 3 & 3.8429 & 7.0125 & 0.9857 & 2.6499 & 3.8105 & 6.9091 \\ \hline \end{tabular} \end{center} \par \caption{{\protect\small Hospital data. Test statistics and bootstrap critical values under the null hypothesis of both monotonicity and convexity of the regression function. Bootstrap critical values are from 400 bootstrap replications.}} \label{table:HospC} \end{table} As we can see from Tables \ref{table:HospA}--\ref{table:HospC}, we do not reject any of the three hypotheses even at 10\% level. $\left. {}\right.$ \noindent 2. \textbf{Energy consumption in the Southern region of Russia.} The data are on daily energy consumption (in MWh) and average daily temperature (in Celsius) in the Southern region of Russia in the period from February 1, 2016 till January 31, 2018. The data have been downloaded from the official website of System Operator of the Unified Energy System of Russia.\footnote{% http://so-ups.ru/} We provide tests for U-shape with a switch at ${17.6}^{\circ }$ and convexity using the approaches outlined in section \ref{section:BPsplines} in the context of Examples 2 and 1, respectively. In order to correct for heteroscedasticity of the errors, we estimate the scedastic function $% \widehat{\sigma }^{2}(x)$ using residuals obtained in the unconstrained estimation using \emph{B-splines} (or \emph{P-splines}, respectively). The scedastic function is estimated using cubic \emph{B-splines} with 6 uniform knots and in the form of \begin{equation*} \sigma ^{2}(x)=\left( \sum_{k=1}^{8}c_{k}B_{k,8}\right) ^{2}. \end{equation*} Figure \ref{fig:SFO} gives scatter plots of the data together with fitted curves obtained under the U-shape constraint with the switch at $s_{0}={17.6}% ^{\circ }$. This constraint fit is obtained in accordance with the technique in section \ref{section:BPsplines}. Namely, we consider individual \emph{% B-spline} fits on intervals $[\underline{x},s_{0}]$ and $[s_{0},\overline{x}% ] $, where $\underline{x}$ and $\overline{x}$ are respectively lowest and highest values of the temperature in the sample. On each interval we use $% L^{\prime }+1=5$ uniform knots. The left-hand side figure only imposes the continuity of the fitted curve at the switch point, whereas the right-hand side figure imposes continuous differentiability. \begin{figure}[!h] \begin{center} \vspace{0in} \begin{minipage}{\linewidth} \begin{minipage}{0.5\linewidth} \includegraphics[width=0.95\linewidth]{SFOenergy_M5_Bsplines.png} \end{minipage} \begin{minipage}{0.5\linewidth} \includegraphics[width=0.95\linewidth]{SFOenergy_M5_Bsplines_cont1.png} \end{minipage} \end{minipage} \end{center} \caption{{\protect\small Energy consumption data. Plot of temperature and energy consumption and the constrained fit (under U-shape with the switch at ${17.6}^{\circ}$) using cubic B-spline with 5 uniform knots on each subinterval of temperature values. On the left-hand side the fitted curve is continuous at the switch point. On the right-hand side the fitted curve is continuously differentiable at the switch point.}} \label{fig:SFO} \end{figure} Tables \ref{table:SFOA}-\ref{table:SFOC} present results of our testing. Namely, Table \ref{table:SFOA} shows test statistics for the null hypothesis of \emph{U-shaped} regression function and also bootstrap critical values using both \emph{B-splines} and \emph{P-splines} in case when two \emph{% B-spline} curves are joined at the switch point in a continuous way. Table % \ref{table:SFOB} presents analogous results for the null hypothesis of \emph{% U-shaped} regression function when two \emph{B-spline} curves are joined at the switch point in a continuously differentiable way. Table \ref{table:SFOC} gives results for the null hypothesis of convexity. In all the cases Khmaladze's transformation is conducted from the right end of support. he bootstrap critical values obtained on the basis of $400$ bootstrap replications. As we can see, the null hypothesis of a \emph{U-shaped} relationship with the switch point at ${17.6}^{\circ }$ is not rejected at the $5\%$ level by any type of the test, whereas convexity is rejected. When testing convexity we use cubic splines with $L^{\prime }+1=7$ uniform knots on $[\underline{x},\overline{x}]$. \begin{table}[tbp] \par \begin{center} \begin{tabular}{ccccccccc} \hline\hline \quad & \multicolumn{4}{c}{B-splines} & \multicolumn{4}{c}{P-splines} \\ \quad & Test statistic & \multicolumn{3}{c}{Bootstrap critical value} & Test statistic & \multicolumn{3}{c}{Bootstrap critical value} \\ Method & \quad & 10\% & 5\% & 1\% & \quad & 10\% & 5\% & 1\% \\ \hline KS & 0.6204 & 1.0997 & 1.2139 & 1.4053 & 0.2482 & 0.5898 & 0.6199 & 0.7207 \\ CvM & 0.0698 & 0.3098 & 0.4048 & 0.5885 & 0.0038 & 0.0434 & 0.0498 & 0.0621 \\ AD & 0.5119 & 1.6541 & 2.1098 & 2.925 & 0.0647 & 0.397 & 0.4312 & 0.5324 \\ \hline \end{tabular} \end{center} \par \caption{{\protect\small Energy consumption data. Test statistics and bootstrap critical values under the null hypothesis of U-shaped regression function with the switch as ${17.6}^{\circ}$. Two B-spline curves are joined continuously at the switch point. Bootstrap critical values are from 400 bootstrap replications.}} \label{table:SFOA} \end{table} \begin{table}[tbp] \par \begin{center} \begin{tabular}{ccccccccc} \hline\hline \quad & \multicolumn{4}{c}{B-splines} & \multicolumn{4}{c}{P-splines} \\ \quad & Test statistic & \multicolumn{3}{c}{Bootstrap critical value} & Test statistic & \multicolumn{3}{c}{Bootstrap critical value} \\ Method & \quad & 10\% & 5\% & 1\% & \quad & 10\% & 5\% & 1\% \\ \hline KS & 0.8481 & 1.023 & 1.1749 & 1.407 & 0.505 & 0.6143 & 0.6524 & 0.7747 \\ CvM & 0.1472 & 0.2211 & 0.3247 & 0.5553 & 0.0469 & 0.0476 & 0.0542 & 0.0752 \\ AD & 0.9114 & 1.3676 & 1.8002 & 2.9199 & 0.2878 & 0.3713 & 0.4151 & 0.5547 \\ \hline \end{tabular} \end{center} \par \caption{{\protect\small Energy consumption data. Test statistics and bootstrap critical values under the null hypothesis of U-shaped regression function with the switch as ${17.6}^{\circ}$. Two B-spline curves are joined smoothly at the switch point. Bootstrap critical values are from 400 bootstrap replications.}} \label{table:SFOB} \end{table} \begin{table}[tbp] \par \begin{center} \begin{tabular}{ccccccccc} \hline\hline \quad & \multicolumn{4}{c}{B-splines} & \multicolumn{4}{c}{P-splines} \\ \quad & Test statistic & \multicolumn{3}{c}{Bootstrap critical value} & Test statistic & \multicolumn{3}{c}{Bootstrap critical value} \\ Method & \quad & 10\% & 5\% & 1\% & \quad & 10\% & 5\% & 1\% \\ \hline KS & 2.9812 & 1.1537 & 1.2652 & 1.5579 & 3.4626 & 0.5478 & 0.6891 & 1.0121 \\ CvM & 2.7713 & 0.3327 & 0.4389 & 0.71 & 3.2622 & 0.0402 & 0.0716 & 0.2076 \\ AD & 14.361 & 1.823 & 2.3809 & 3.856 & 17.23 & 0.3332 & 0.484 & 1.2158 \\ \hline \end{tabular} \end{center} \par \caption{{\protect\small Energy consumption data. Test statistics and bootstrap critical values under the null hypothesis of convexity of the regression function. Bootstrap critical values are from 400 bootstrap replications.}} \label{table:SFOC} \end{table} \section{\textbf{CONCLUSION}} \label{sec:conclusion} $\left. {}\right.$ This paper proposes a methodology for testing a wide range of shape properties of a regression function. The methodology relies on applying the Khmaladze's transformation to the partial sums empirical process in a nonparametric setting where \emph{B-splines} have been used to approximate the functional space under the null hypothesis. We establish that the proposed Khmaladze's transformation eliminates the effect of nonparametric estimation and results in asymptotically pivotal testing. To the best of our knowledge, this paper is the first implementation of the Khmaladze's transformation in a nonparametric setting. In our main examples we considered shape constraints that can be written as inequality constraints on the coefficients of the approximating regression splines. The generality of our procedure allows to test several shape properties simultaneously. The implementation is especially easy when the inequality constraints are linear, which is the case for shape properties expressed as linear inequality constraints on the derivatives. \section{\textbf{APPENDIX A}} For a $M\times M$ matrix $C=\left\{ c_{kj}\right\} _{k,j=1}^{M}$, $% \left\Vert C\right\Vert _{E}=\sum_{k,j=1}^{M}c_{kj}^{2}$ denotes the Euclidean norm and for a $M\times N$ matrix we define the spectral norm $% \left\Vert G\right\Vert $ as $\overline{\lambda }^{1/2}\left( G^{\prime }G\right) $, where $\overline{\lambda }\left( H\right) $ denotes the maximum eigenvalue of the matrix $H$. It is worth recalling the following inequality $\left\Vert CG\right\Vert _{E}\leq \left\Vert C\right\Vert _{E}\left\Vert G\right\Vert $. Finally $K$ denotes a generic finite and positive constant. We now introduce the following notation. We shall denote the fourth cumulant of a random variable $z$ by $\kappa _{4}\left( z\right) $ and for any $x_{i}$% , $i=1,...,n$, $\mathcal{I}\left( x^{1}<x_{i}<x^{2}\right) =:\mathcal{I}% _{i}\left( x^{1},x^{2}\right) $. Also we define \begin{eqnarray} A_{L,i} &=&A_{L}\left( \widetilde{x}_{i}\right) \text{, \ \ \ \ }\widetilde{% \boldsymbol{P}}_{i}=A_{L,i}^{+1/2}\boldsymbol{P}_{i}\text{; }\ i=1,...,n \label{q_1} \\ \overline{A}_{L}\left( x\right) &=&D_{L}\left( x\right) A_{L}\left( x\right) D_{L}\left( x\right) \text{; \ \ }\overline{A}_{L,i}=:\overline{A}_{L}\left( \widetilde{x}_{i}\right) \text{,} \notag \end{eqnarray}% where $A_{L}\left( x\right) $ was given in $\left( \ref{L_x}\right) $ and% \begin{equation} D_{L}\left( x\right) =diag\left( d_{1}\left( x\right) ,...,d_{L}\left( x\right) \right) \text{, \ }d_{\ell }\left( x\right) =\left\{ \begin{array}{c} 0\text{ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ if \ \ \ \ \ \ }x<z^{\ell -1} \\ L^{-q}\left( z^{\ell }-x\right) ^{-q-1/2}\ \ \ \ \text{if \ \ \ \ \ }z^{\ell -1}\leq x<z^{\ell } \\ L^{1/2}\text{ \ \ \ \ \ \ \ \ \ \ \ \ \ \ if \ \ \ \ \ \ \ }z^{\ell }\leq x% \text{.}% \end{array}% \right. \label{q_11} \end{equation} Observe that when $x=:z^{k}$, that is a knot, Condition $C1$ yields that% \begin{equation*} \overline{A}_{L}\left( z^{k}\right) =:L~A_{L}\left( z^{k}\right) =diag\left( \underline{0},B_{L}\left( z^{k}\right) \right) \text{,} \end{equation*}% where $\underline{0}$ is a square $k-1$ matrix of zeroes and the $L-k+1$ matrix $B_{L}\left( z^{k}\right) $ is positive definite, where the elements $% B_{L,\ell _{1},\ell _{2}}\left( z^{k}\right) $ of the matrix $B_{L}\left( z^{k}\right) $ are zero if $\left\vert \ell _{1}-\ell _{2}\right\vert >q$. The latter follows because there are only $q$ splines different than zero at a given value $x$. Finally, it is worth mentioning that for $x\in \left( z^{\ell -1},z^{\ell }\right) $, $d_{\ell }\left( x\right) =Kp_{\ell ,L}^{-1}\left( x;q\right) \left( z^{\ell }-x\right) ^{-1/2}$ and recalling that Harville's \cite{Harville} Section 20.2 yields that $\overline{A}% _{L}^{+}\left( z^{k}\right) =diag\left( \underline{0},B_{L}^{-1}\left( z^{k}\right) \right) $, where we abbreviate $p_{\ell ,L}\left( x;q\right) $ by $p_{\ell ,L}\left( x\right) $ in what follows. \subsection{\textbf{PROOF OF THEOREM \protect\ref{M_n}}} $\left. {}\right.$ We need to show $\left( \mathbf{a}\right) $ the finite dimensional distributions converge to a Gaussian random variable with covariance structure given by that of $\mathcal{U}\left( x\right) $ and $\left( \mathbf{% b}\right) $ the tightness of the sequence. We begin the proof of part $% \left( \mathbf{a}\right) $ showing the structure of the covariance structure of $\mathcal{M}_{n}\left( x\right) $. That is, for any $0\leq x^{1}\leq x^{2}\leq 1$, \begin{eqnarray} E\left( \mathcal{M}_{n}\left( x^{1}\right) \mathcal{M}_{n}\left( x^{2}\right) \mid X\right) &=&\frac{1}{n}\sum_{i,j=1}^{n}E\left( v_{i}v_{j}\mid X\right) \mathcal{I}_{i}\left( x^{1}\right) \mathcal{I}% _{j}\left( x^{2}\right) \label{pth1_1} \\ &&\overset{P}{\rightarrow }\sigma _{u}^{2}F_{X}\left( x^{1}\right) \text{.} \notag \end{eqnarray} Consider $i<j$ first and assume, without loss of generality, that $% x_{i}<x_{j}$. When $x_{i}+n^{-\varsigma }\leq z^{k\left( x_{i}\right) }$, that is $\widetilde{x}_{i}=x_{i}$, Condition $C1$ implies that% \begin{eqnarray} E\left( C_{n,j}u_{i}\mid X\right) &=&0;\text{ }E\left( C_{n,i}u_{j}\mid X\right) =\frac{\sigma _{u}^{2}}{n}\boldsymbol{P}_{j} \notag \\ E\left( C_{n,i}C_{n,j}^{\prime }\mid X\right) &=&\frac{\sigma _{u}^{2}}{n}% A_{n,j}\text{,} \label{th_1_a} \end{eqnarray}% and hence we obtain that% \begin{equation} E\left( v_{i}v_{j}\mid X\right) =\frac{\sigma _{u}^{2}}{n}\left\{ \boldsymbol{P}_{i}^{\prime }A_{n,i}^{+}A_{n,j}A_{n,j}^{+}\boldsymbol{P}_{j}-% \boldsymbol{P}_{i}^{\prime }A_{n,i}^{+}\boldsymbol{P}_{j}\right\} =0 \label{th_1_b} \end{equation}% because Harville's \cite{Harville} Theorem 12.3.4 yields that $% A_{n,j}A_{n,j}^{+}\boldsymbol{P}_{j}=\boldsymbol{P}_{j}$. On the other hand when $x_{i}+n^{-\varsigma }>z^{k\left( x_{i}\right) }$, that is $\widetilde{x% }_{i}\not=x_{i}$ and $x_{j}>z^{k\left( x_{i}\right) }$, $\left( \ref{th_1_a}% \right) $ holds true which implies $\left( \ref{th_1_b}\right) $. Finally when $z^{k\left( x_{i}\right) }-n^{-\varsigma }<x_{i},x_{j}\leq z^{k\left( x_{i}\right) }$, we obtain that \begin{eqnarray*} E\left( v_{i}v_{j}\mid X\right) &=&\frac{\sigma _{u}^{2}}{n}\boldsymbol{P}% _{i}^{\prime }A_{n}^{+}\left( z^{k\left( x_{i}\right) }\right) \boldsymbol{P}% _{j}\mathcal{I}\left( z^{k\left( x_{i}\right) }-\frac{1}{n^{\varsigma }}% <x_{i},x_{j}\leq z^{k\left( x_{i}\right) }\right) \\ &=&:\vartheta _{n}\left( i,j;X\right) \text{.} \end{eqnarray*}% Observe that in this case we have \begin{equation} E\left\Vert E\left( v_{i}v_{j}\mid X\right) \right\Vert =O\left( Ln^{-1-2\varsigma }\right) \label{e_v_i_vj} \end{equation}% because Lemmas \ref{Est_A_i} and \ref{Est_A_uni} imply that \begin{equation} \underline{\lambda }\left( \inf_{x\in \left[ 0,1\right] }A_{L}^{+1/2}\left( x\right) A_{n}\left( x\right) A_{L}^{+1/2}\left( x\right) \right) >\delta >0 \label{A_inv} \end{equation}% with probability approaching one, where \underline{$\lambda $}$\left( G\right) $ denotes the minimum eigenvalue of the matrix $G$, Lemma \ref% {Nor_P_i} implies that $E\left\Vert \widetilde{\boldsymbol{P}}% _{i}\right\Vert ^{2}=O\left( L\right) $ and \begin{equation} E\left( \mathcal{I}\left( z^{k\left( x_{i}\right) }-n^{-\varsigma }<x_{i}\leq z^{k\left( x_{i}\right) }\right) \right) =O\left( n^{-\varsigma }\right) \text{.} \label{e_ind} \end{equation} Next, when $i=j$, proceeding as above we obtain that \begin{equation*} E\left( v_{i}v_{j}\mid X\right) =\sigma _{u}^{2}\left( 1-\frac{\boldsymbol{P}% _{i}^{\prime }A_{n,i}^{+}\boldsymbol{P}_{i}}{n}\right) \text{.} \end{equation*}% Observe that, denoting $A_{n,i}^{\ast }=A_{n,i}-\boldsymbol{P}_{i}% \boldsymbol{P}_{i}^{\prime }/n$, we have that \begin{equation*} 1-\frac{\boldsymbol{P}_{i}^{\prime }A_{n,i}^{+}\boldsymbol{P}_{i}}{n}=1-2% \frac{\boldsymbol{P}_{i}^{\prime }A_{n,i}^{+}\boldsymbol{P}_{i}}{n}+\left( \frac{\boldsymbol{P}_{i}^{\prime }A_{n,i}^{+}\boldsymbol{P}_{i}}{n}\right) ^{2}+\frac{\boldsymbol{P}_{i}^{\prime }A_{n,i}^{+}A_{n,i}^{\ast }A_{n,i}^{+}% \boldsymbol{P}_{i}}{n}>0\text{.} \end{equation*} So, we conclude that the left side of $\left( \ref{pth1_1}\right) $ is, for any $0\leq x^{1}\leq x^{2}\leq 1$, \begin{equation} \frac{\sigma _{u}^{2}}{n}\sum_{i=1}^{n-1}\left( 1-\frac{\boldsymbol{P}% _{i}^{\prime }A_{n,i}^{+}\boldsymbol{P}_{i}}{n}\right) \mathcal{I}_{i}\left( x^{1}\right) +\frac{\sigma _{u}^{2}}{n}\sum_{i\neq j}^{n-1}\vartheta _{n}\left( i,j;X\right) \mathcal{I}_{i}\left( x^{1}\right) \text{.} \label{Theo1_1} \end{equation} The second term of $\left( \ref{Theo1_1}\right) $ is $o_{p}\left( 1\right) $ because $\left( \ref{e_v_i_vj}\right) $ implies that its first absolute moment is $O\left( L/n^{2\varsigma }\right) =o\left( 1\right) $ since $% \varsigma >1/2$ and Condition $C3$. The first term of $\left( \ref{Theo1_1}% \right) $ is also $o_{p}\left( 1\right) $ as we now show. Using the inequality $\mathcal{I}_{i}\left( x^{1}\right) \leq \sum_{k=1}^{k\left( x^{1}\right) -1}\mathcal{I}_{i}\left( z^{k};z^{k+1}\right) $, we obtain that \begin{eqnarray} &&E\left( \frac{1}{n^{2}}\sum_{i=1}^{n-1}\widetilde{\boldsymbol{P}}% _{i}^{\prime }\left( A_{L,i}^{+1/2}A_{n,i}A_{L,i}^{+1/2}\right) ^{+}% \widetilde{\boldsymbol{P}}_{i}\mathcal{I}_{i}\left( x^{1}\right) \right) \notag \\ &\leq &\sum_{k=1}^{k\left( x^{1}\right) -1}E\left( \frac{1}{n^{2}}% \sum_{i=1}^{n-1}\widetilde{\boldsymbol{P}}_{i}^{\prime }\left( A_{L,i}^{+1/2}A_{n,i}A_{L,i}^{+1/2}\right) ^{+}\widetilde{\boldsymbol{P}}_{i}% \mathcal{I}_{i}\left( z^{k};z^{k+1}\right) \right) \notag \\ &=&\frac{K}{n^{2}}\sum_{k=1}^{k\left( x^{1}\right) -1}\sum_{i=1}^{n}E\left\Vert \widetilde{\boldsymbol{P}}_{i}\mathcal{I}% _{i}\left( z^{k};z^{k+1}\right) \right\Vert ^{2}=O\left( \frac{L}{n}\right) \text{,} \label{eg} \end{eqnarray}% by $\left( \ref{A_inv}\right) $ and because $E\left\Vert \widetilde{% \boldsymbol{P}}_{i}\mathcal{I}_{i}\left( z^{k};z^{k+1}\right) \right\Vert ^{2}=O\left( 1\right) $ by Lemma \ref{Nor_P_i}, cf. $\left( \ref{Lemma2_4}% \right) $. So, we conclude that $\left( \ref{pth1_1}\right) $ holds true as it is well known that, under Conditions $C1$ and then $C3$, \begin{equation*} \sigma _{u}^{2}\frac{1}{n}\sum_{i=1}^{n}\mathcal{I}_{i}\left( x\right) \left( 1+\frac{L}{n}\right) \overset{P}{\rightarrow }\sigma _{u}^{2}F_{X}\left( x\right) \text{.} \end{equation*} To complete part $\left( \mathbf{a}\right) $, it suffices to show the asymptotic Gaussianity of $\mathcal{M}_{n}\left( x\right) $ for a fixed $x$ due to Cram\'{e}r-Wold device. First, we have already shown that% \begin{equation*} \sum_{i=1}^{n}E\left( \upsilon _{in}^{2}\left( x\right) \mid X\right) \overset{P}{\rightarrow }1\text{,} \end{equation*}% where by construction $\upsilon _{in}\left( x\right) =:\sigma _{u}^{-1}F_{X}^{-1/2}\left( x\right) \mathcal{I}_{i}\left( x\right) v_{i}/n^{1/2}$ is a martingale difference triangular array of r.v.'s. So, it remains to show the Lindeberg's condition for which a sufficient condition is \begin{equation} \sum_{i=1}^{n}E\left\vert \upsilon _{in}\left( x\right) \right\vert ^{4}=o\left( 1\right) \text{.} \label{v_in} \end{equation}% But the latter holds true because $\left( \ref{A_inv}\right) $ yields that \begin{equation*} E\left\Vert \boldsymbol{P}_{i}^{\prime }A_{n,i}^{+}C_{n,i}\mathcal{I}% _{i}\left( x\right) \right\Vert ^{4}\leq E\left\Vert \widetilde{\boldsymbol{P% }}_{i}\right\Vert ^{4}E\left\Vert \frac{1}{n}\sum_{j=1}^{n}\widetilde{% \boldsymbol{P}}_{j}u_{j}\mathcal{J}_{j}\left( \widetilde{x}_{i}\right) \right\Vert ^{4}=O\left( \frac{L^{4}}{n^{2}}\right) \text{,} \end{equation*}% since Lemma \ref{Nor_P_i} implies that $E\left\Vert \widetilde{\boldsymbol{P}% }_{i}\right\Vert ^{4}=O\left( L^{2}\right) $. So, $\left( \ref{v_in}\right) $ holds true by Condition $C3$, which concludes the proof of part $\left( \mathbf{a}\right) $. $\left( \mathbf{b}\right) $ Tightness follows by Lemmas \ref{Est_A_uni} and % \ref{tight} and using the same arguments as those for the proof of Billingsley's \cite{Billingsley} Theorem 22.1, see also Wu's \cite{Wu} Lemma 14, which completes the proof of the theorem. Observe that, for any $\delta >0$, $\delta \sum_{k\in \mathbb{Z}}\sup_{x\in \left( k\delta ,\left( k+1\right) \delta \right) }f_{X}\left( x\right) <K$ following Lemma 4 of Wu \cite{Wu}. \subsection{PROOF OF THEOREM \protect\ref{M_nest}} $\left. {}\right. $ We shall show that, uniformly in $x\in \left[ 0,1\right] $, \begin{equation} \widetilde{\mathcal{M}}_{n}\left( x\right) -\mathcal{M}_{n}\left( x\right) =o_{p}\left( 1\right) \text{.} \label{m_dif} \end{equation}% Because $\left( \ref{res_con}\right) $ yields that $\widehat{u}% _{i}-u_{i}=m^{bias}\left( x_{i}\right) -\boldsymbol{P}_{i}^{\prime }\left( \widehat{b}-\beta \right) $, we have that the left side of $\left( \ref% {m_dif}\right) $ is% \begin{equation} \frac{1}{n^{1/2}}\sum_{i=1}^{n}m^{bias}\left( x_{i}\right) \mathcal{I}% _{i}\left( x\right) -\frac{1}{n^{1/2}}\sum_{i=1}^{n}\left\{ \frac{% \boldsymbol{P}_{i}^{\prime }A_{n,i}^{+}}{n}\sum_{k=1}^{n}\boldsymbol{P}% _{k}m^{bias}\left( x_{k}\right) \mathcal{J}_{k}\left( \widetilde{x}% _{i}\right) \right\} \mathcal{I}_{i}\left( x\right) \label{m_dif2} \end{equation}% since the contribution due to $\boldsymbol{P}_{i}^{\prime }\left( \widehat{b}% -\beta \right) $ is zero by Lemma \ref{bias}. Now the first term of $\left( \ref{m_dif2}\right) $ is $o_{p}\left( 1\right) $ by Condition $C3$ and that Agarwal and Studden's \cite{AgarwalStudden} Theorems 3.1 and 4.1, see also Zhou et al. $\left( 1998\right) $, yields that $Em^{bias}\left( x_{i}\right) =O\left( L^{-3}\right) $, whereas together with Lemma \ref{Est_A_uni} and $\left( \ref{A_inv}\right) $, we obtain that the second term is, with probability approaching $1$, bounded by \begin{equation*} \frac{K}{L^{3}n^{1/2}}\sum_{i=1}^{n}\left\{ \left\Vert \widetilde{% \boldsymbol{P}}_{i}\right\Vert \frac{1}{n}\sum_{k=1}^{n}\left\Vert \widetilde{\boldsymbol{P}}_{k}\right\Vert \right\} =O_{p}\left( n^{1/2}/L^{2}\right) =o\left( 1\right) \end{equation*}% by Markov's inequality and Lemma \ref{Nor_P_i}\ and then Condition $C3$. This completes the proof of the theorem.\hfill $\blacksquare $ \subsection{PROOF OF PROPOSITION \protect\ref{sigma_est}} $\left. {}\right. $ First, using $\left( \ref{res_con}\right) $, we have that \begin{eqnarray*} \widehat{u}_{i}^{2} &=&u_{i}^{2}+m^{bias}\left( x_{i}\right) ^{2}+\left( \widehat{b}-\beta \right) ^{\prime }\boldsymbol{P}_{i}\boldsymbol{P}% _{i}^{\prime }\left( \widehat{b}-\beta \right) \\ &&+2u_{i}m^{bias}\left( x_{i}\right) +2u_{i}\boldsymbol{P}_{i}^{\prime }\left( \widehat{b}-\beta \right) +2m^{bias}\left( x_{i}\right) \boldsymbol{P% }_{i}^{\prime }\left( \widehat{b}-\beta \right) \text{.} \end{eqnarray*}% So, by Cauchy-Schwarz's inequality, it suffices to show that \begin{equation} \left( \widehat{b}-\beta \right) ^{\prime }\frac{1}{n}\sum_{i=1}^{n}% \boldsymbol{P}_{i}\boldsymbol{P}_{i}^{\prime }\left( \widehat{b}-\beta \right) \text{ \ and }\frac{1}{n}\sum_{i=1}^{n}m^{bias}\left( x_{i}\right) ^{2} \label{a1} \end{equation}% are$\ o_{p}\left( 1\right) $ because $n^{-1}\sum_{i=1}^{n}u_{i}^{2}-\sigma _{u}^{2}=o_{p}\left( 1\right) $. Now, the norm of the first expression in $% \left( \ref{a1}\right) $ is $O_{p}\left( L^{1/2}/n\right) =o_{p}\left( 1\right) $ because \begin{equation*} E\left( \left( A_{L}\left( 0\right) ^{-1/2}\frac{1}{n}\sum_{i=1}^{n}\mathbf{P% }_{i}u_{i}\right) \left( A_{L}\left( 0\right) ^{-1/2}\frac{1}{n}% \sum_{i=1}^{n}\mathbf{P}_{i}u_{i}\right) ^{\prime }\mid X\right) =\frac{% \sigma _{u}^{2}}{n}I_{L} \end{equation*}% and $\left\Vert I_{L}\right\Vert _{E}^{2}=L$. Note that by Lemma \ref% {Est_A_i}, $A_{n}\left( 0\right) $ is invertible with probability approaching $1$ since $A_{L}\left( 0\right) $ is. On the other hand, the second expression in $\left( \ref{a1}\right) $ is $o_{p}\left( 1\right) $ by Agarwal and Studden's \cite{AgarwalStudden} Theorems 3.1 and 4.1 and then Condition $C3$. This completes the proof of the proposition.\hfill $% \blacksquare $ \subsection{PROOF OF THEOREM \protect\ref{M_nest-nonlinear}} $\left. {} \right.$ In this proof, for the sake of notation simplicity, we will write $\beta$ and $\widehat{b}$ instead of $\beta_{-\ell_0}$ and $\widehat{b}% _{-\ell_0}$, respectively. Let \begin{equation*} G\left( x;\beta \right) =:\frac{\partial \widetilde{P}_{i}\left( x;\beta \right) }{\partial \beta ^{\prime }}\quad \text{and}\quad G_{i}\left( \beta \right) =:G_{i}\left( x_{i};\beta \right) . \end{equation*}% From the definition of $\widetilde{P}_{i}\left( x;\beta \right) $, it follows that $G\left( x;\beta \right) =H(\beta )p_{\ell _{0}}(x)$, where $% H(\beta )$ is the $(L-1)\times (L-1)$ Hessian of $h(\beta )$, whose norm is finite. Then, employing the form of our model and the approximation, by Taylor's approximation to $g(\cdot)$, we write \begin{eqnarray*} \widehat{u}_{i} &=&u_{i}-\left( g\left( x_{i};\widehat{b}\right) -g\left( x_{i};\beta \right) \right) +(m(x_{i})-g\left( x_{i};\beta \right) ) \\ &=&u_{i}-\left( \widehat{b}-\beta \right) ^{\prime }\widetilde{P}% _{i}\left( \beta \right) +\frac{1}{2}\left( \widehat{b}-\beta \right) ^{\prime }G_{i}\left( \widetilde{b}\right) \left( \widehat{b}% -\beta \right) +\left( m(x_{i})-g\left( x_{i};{\beta }\right) \right) \text{, } \end{eqnarray*} where $\widetilde{b}$ is an intermediate point between $\beta$ and $\widehat{b}$. So, it suffices to examine the contribution due to the following four terms: \begin{eqnarray*} \mathcal{A}_{1,n}\left( x;\widehat{b}\right) &=&:\frac{1}{n^{1/2}}% \sum_{i=1}^{n}\left\{ u_{i}-\widetilde{P}_{i}^{\prime }\left( \widehat{b }\right) \mathcal{D}_{n}^{+}\left( \widehat{b};i\right) \sum_{k=1}^{n}% \widetilde{P}_{k}\left( \widehat{b}\right) u_{k}\mathcal{J}_{k}\left( \widetilde{x}_{i}\right) \right\} \mathcal{I}_{i}(x) \\ \mathcal{A}_{2,n}\left( x;\widehat{b}\right) &=&:\frac{1}{n^{1/2}}% \sum_{i=1}^{n}\left\{ \widetilde{P}_{i}^{\prime }\left( \beta \right) -% \widetilde{P}_{i}^{\prime }\left( \widehat{b}\right) \mathcal{D}% _{n}^{+}\left( \widehat{b};i\right) \sum_{k=1}^{n}\widetilde{P}% _{k}\left( \widehat{b}\right) \widetilde{P}_{k}^{\prime }\left( \beta \right) \mathcal{J}_{k}\left( \widetilde{x}_{i}\right) \right\} \mathcal{I}% _{i}(x)\left( \widehat{b}-\beta \right) \\ \mathcal{A}_{3,n}\left( x;\widehat{b}\right) &=&:\left( \widehat{b}% -\beta \right) ^{\prime }\frac{1}{n^{1/2}}\sum_{i=1}^{n}\mathcal{I}% _{i}(x)G_{i}\left( \widetilde{b}\right) \left( \widehat{b}-\beta \right) \\ &&-\frac{1}{n^{1/2}}\sum_{i=1}^{n}\widetilde{P}_{i}^{\prime }\left( \widehat{b }\right) \mathcal{D}_{n}^{+}\left( \widehat{b};i\right) \frac{1}{n% }\left\{ \sum_{k=1}^{n}\widetilde{P}_{k}\left( \widehat{b}\right) \left( \widehat{b}-\beta \right) ^{\prime }G_{k}\left( \widetilde{b }\right) \left( \widehat{b}-\beta \right) \mathcal{J}_{k}\left( \widetilde{x}_{i}\right) \right\} \mathcal{I}_{i}(x)\text{,} \\ \mathcal{A}_{4,n}\left( \widehat{b}\right) &=&:\frac{1}{n^{1/2}}% \sum_{i=1}^{n}\left\{ m^{bias}\left( x_{i}\right) -\widetilde{P}_{i}^{\prime }\left( \widehat{b}\right) \mathcal{D}_{n}^{+}\left( \widehat{b}% ;i\right) \frac{1}{n}\sum_{k=1}^{n}\left( m^{bias}\left( x_{k}\right) \right) \mathcal{J}_{k}\left( \widetilde{x}_{i}\right) \right\} \mathcal{I}% _{i}(x), \end{eqnarray*}% with $m(x_{i})-g\left( x_{i};{\beta }\right) =:m^{bias}\left( x_{i}\right) $. Also introduce the following: \begin{eqnarray*} \mathcal{A}_{1,n}\left( x;{\beta }\right) &=&:\frac{1}{n^{1/2}}% \sum_{i=1}^{n}\left\{ u_{i}-\widetilde{P}_{i}^{\prime }\left( {\beta }% \right) \mathcal{D}_{n}^{+}\left( {\beta };i\right) \sum_{k=1}^{n}\widetilde{% P}_{k}\left( {\beta }\right) u_{k}\mathcal{J}_{k}\left( \widetilde{x}% _{i}\right) \right\} \mathcal{I}_{i}(x), \\ \mathcal{A}_{2,n}\left( x;\beta \right) &=&:\frac{1}{n^{1/2}}% \sum_{i=1}^{n}\left\{ \widetilde{P}_{i}^{\prime }\left( \beta \right) -% \widetilde{P}_{i}^{\prime }\left( {\beta }\right) \mathcal{D}_{n}^{+}\left( {% \beta };i\right) \sum_{k=1}^{n}\widetilde{P}_{k}\left( {\beta }\right) \widetilde{P}_{k}^{\prime }\left( \beta \right) \mathcal{J}_{k}\left( \widetilde{x}_{i}\right) \right\} \mathcal{I}_{i}(x)\left( \widehat{b}% -\beta \right) , \\ \mathcal{A}_{3,n}\left( x;{\beta }\right) &=&:\left( \widehat{b}-\beta \right) ^{\prime }\frac{1}{n^{1/2}}\sum_{i=1}^{n}\mathcal{I}% _{i}(x)G_{i}\left( \widetilde{b}\right) \left( \widehat{b}-\beta \right) \\ &&-\frac{1}{n^{1/2}}\sum_{i=1}^{n}\widetilde{P}_{i}^{\prime }\left( {\beta }% \right) \mathcal{D}_{n}^{+}\left( {\beta };i\right) \frac{1}{n}\left\{ \sum_{k=1}^{n}\widetilde{P}_{k}\left( {\beta }\right) \left( \widehat{b}% -\beta \right) ^{\prime }G_{k}\left( \widetilde{b}\right) \left( \widehat{b}-\beta \right) \mathcal{J}_{k}\left( \widetilde{x}% _{i}\right) \right\} \mathcal{I}_{i}(x). \end{eqnarray*}% We shall first examine the behaviour of $\mathcal{A}_{q,n}\left( x;\beta \right) $, $q=1,2,3$, and then we shall show that, uniformly in $x$, $% \mathcal{A}_{q,n}\left( x;\widehat{b}\right) -\mathcal{A}_{q,n}\left( x;\beta \right) =o_{p}\left( 1\right) $. Proceeding as with the proof of Theorem \ref{M_n}, we can establish \begin{equation*} \mathcal{A}_{1,n}\left( \beta \right) \overset{weakly}{\Rightarrow }\mathcal{% U}\left( x\right) . \end{equation*}% It is also obvious that $\mathcal{A}_{2,n}\left( x;\beta \right) =0$. Now we want to show that \begin{equation*} \sup_{x}\left\vert \mathcal{A}_{3,n}\left( x;\beta \right) \right\vert =o_{p}\left( 1\right) . \end{equation*}% To that end, we shall look at% \begin{eqnarray*} \mathcal{C}_{1,n}\left( x;\beta \right) &=&:\left( \widehat{b}-\beta \right) ^{\prime }\left\{ \frac{1}{n^{1/2}}\sum_{i=1}^{n}\mathcal{I}% _{i}(x)G_{i}\left( \widetilde{b}\right) \right\} \left( \widehat{b}% -\beta \right) \\ \mathcal{C}_{2,n}\left( x;\beta \right) &=&:\frac{1}{n^{1/2}}\sum_{i=1}^{n}% \widetilde{P}_{i}^{\prime }\left( \beta \right) \mathcal{D}_{n}^{+}\left( \beta ;i\right) \left\{ \left( \sum_{k=1}^{n}\widetilde{P}_{k}\left( \beta \right) \left( \widehat{b}-\beta \right) ^{\prime }G_{k}\left( \widetilde{b}\right) \left( \widehat{b}-\beta \right) \mathcal{J}% _{k}\left( \widetilde{x}_{i}\right) \right) \right\} \mathcal{I}_{i}(x)\text{% .} \end{eqnarray*}% We already know that $\widehat{b}-\beta =O_{p}\left( \left( L/n\right) ^{1/2}\right) $. So, by definition of $G\left( x;\beta \right) $, we obtain that \begin{eqnarray*} \left\Vert \mathcal{C}_{1,n}\left( x;\beta \right) \right\Vert &=&\left\Vert H\left( \widetilde{b}\right) \right\Vert \left\vert \frac{1}{n}% \sum_{i=1}^{n}Lp_{\ell _{0}}\left( x_{i}\right) \mathcal{I}% _{i}(x)\right\vert O_{p}\left( \frac{1}{n^{1/2}}\right) \\ &=&\left\Vert H\left( \widetilde{b}\right) \right\Vert \frac{1}{n}% \sum_{i=1}^{n}\left\vert Lp_{\ell _{0}}\left( x_{i}\right) \right\vert O_{p}\left( \frac{1}{n^{1/2}}\right) \\ &=&:O_{p}\left( n^{-1/2}\right) \text{,} \end{eqnarray*}% because $E\left\vert Lp_{\ell _{0}}\left( x_{i}\right) \right\vert \leq K$. So, because the last displayed expression yields that% \begin{equation*} \sup_{x}\left\vert \mathcal{C}_{1,n}\left( x;\beta \right) \right\vert =O_{p}\left( n^{-1/2}\right) \text{.} \end{equation*} Next, we examine $\mathcal{C}_{2,n}\left( x;\beta \right) $. Again, from the definition of $G\left( x;\beta \right) $ we have that \begin{equation*} \left( \widehat{b}-\beta \right) ^{\prime }G_{k}\left( \widetilde{b}\right) \left( \widehat{b}-\beta \right) =\left( \widehat{b}% -\beta \right) ^{\prime }H\left( \widetilde{b}\right) \left( \widehat{% \beta }-\beta \right) p_{\ell _{0}}\left( x_{k}\right) \text{.} \end{equation*}% From the fact that $\widehat{b}-\beta =O_{p}\left( \left( L/n\right) ^{1/2}\right) $ we can conclude that \begin{equation*} \mathcal{C}_{2,n}\left( \beta \right) =C\frac{L}{n^{3/2}}\sum_{i=1}^{n}% \widetilde{P}_{i}^{\prime }\left( \beta \right) \mathcal{D}_{n}^{+}\left( \beta ;i\right) \left\{ \sum_{k=1}^{n}\widetilde{P}_{k}\left( \beta \right) p_{\ell _{0}}\left( x_{k}\right) \iota _{L}^{\prime }H\left( \widetilde{% b }\right) \iota _{L}\mathcal{J}_{k}\left( \widetilde{x}_{i}\right) \right\} \mathcal{I}_{i}(x)\text{,} \end{equation*}% where $\iota _{L}=(1,\ldots ,1)^{\prime }$ and $C$ is some positive constant. We can establish that $\underline{\lambda }\left( \frac{L}{n}\mathcal{D}% _{n}\left( \beta ;i\right) \right) >0$ as this is what we have essentially in expression $\left( \ref{A_inv}\right) $ in the Appendix of the paper. Also note that \begin{eqnarray*} E\left( L^{1/2}\widetilde{P}_{i}^{\prime }\left( \beta \right) \right) &\simeq &L^{-1/2}\left( 1,...,1\right) ; \\ E\left( L^{1/2}\widetilde{P}_{k}\left( \beta \right) p_{\ell _{0}}\left( x_{k}\right) \right) &\simeq &L^{-1/2}\left( d_{1,h},...,d_{L,h}\right) ,^{\prime } \end{eqnarray*}% where $d_{\ell ,h}$, $\ell \neq \ell _{0}$, is a 0/1 variable which takes the values of 1 if and only if the function $h$ does depend on $\beta _{\ell }$. For instance, in the case of HA-convexity, which was given as an example, $h$ would depend only on parameters $\beta _{\ell _{0}-1}$ and $% \beta _{\ell _{0}-2}$ and, thus, we would only have $d_{\ell _{0}-1,h}=1$ and $d_{\ell _{0}-2,h}=1$ whereas the rest of such indicators would be zero. In this case we would have \begin{equation*} \mathcal{C}_{2,n}\left( x;\beta \right) =O_{p}\left( n^{-1/2}\right) \end{equation*}% once we observe that $L^{-1/2}\left( 1,...,1\right) L^{-1/2}\left( d_{1,h},...,d_{L,h}\right) ^{\prime }\simeq L^{-1}$. If $h$ depends on parameters whose number is growing at the rate $L^{1/2}$, then we would have \begin{equation*} \mathcal{C}_{2,n}\left( x;\beta \right) =O_{p}\left( \left( L/n\right) ^{1/2}\right) \end{equation*}% since now $L^{-1/2}\left( 1,...,1\right) L^{-1/2}\left( d_{1,h},...,d_{L,h}\right) ^{\prime }\simeq L^{-1/2}$ Of course, in the scenario when $h$ depends on all the other parameters, we have that $L^{-1/2}\left( 1,...,1\right) L^{-1/2}\left( d_{1,h},...,d_{L,h}\right) ^{\prime }\simeq 1,$which yields that \begin{equation*} \mathcal{C}_{2,n}\left( x;\beta \right) =O_{p}\left( \left( L^{2}/n\right) ^{1/2}\right) \text{,} \end{equation*}% which by Condition $C3$, it is $o_{p}\left( 1\right) $. Now we show that \begin{equation*} \mathcal{A}_{q,n}\left( x; \widehat{b}\right) -\mathcal{A}_{q,n}\left( x; \beta \right) =o_{p}\left( 1\right) \text{, \ \ \ }q=1,2,3\text{.} \end{equation*}% For that purpose, observe that by mean value theorem, we have that \begin{equation} \widetilde{P}_{i}\left( \widehat{b}\right) -\widetilde{P}_{i}\left( \beta \right) = H\left( \widetilde{b} \right) \left( \widehat{b}% -\beta \right) p_{\ell_0}\left( x_{i}\right) = O_{p}\left( \left( Ln\right) ^{-1//2}\right) \label{3} \end{equation}% because $\sup_{\beta }\left\Vert H\left( \beta \right) \right\Vert \leq K$, $% \widehat{b} -\beta =O_{p}\left( \left( L/n\right) ^{1/2}\right) $ and $% E\left\Vert p_{\ell_0 }\left( x_{k}\right) \right\Vert =O\left( L^{-1}\right) $. Next, we examine the behaviour of $\mathcal{D}_{n}^{+}\left( \widehat{b}% ;i\right) -\mathcal{D}_{n}^{+}\left( \beta ;i\right) $. By the mean value theorem and standard algebra, we have that \begin{eqnarray*} \mathcal{D}_{n}\left( \widehat{b};i\right) &=&\mathcal{D}_{n}\left( \beta ;i\right) +2H\left( \widetilde{b}\right) \left( \widehat{b}% -\beta \right) \sum_{k=1}^{n}\widetilde{P}_{k}\left( \beta \right) p_{\ell _{0}}\left( x_{k}\right) \mathcal{J}_{k}\left( \widetilde{x}_{i}\right) \\ &&+H\left( \widetilde{\beta }\right) \left( \widehat{b}-\beta \right) \left( \sum_{k=1}^{n}p_{\ell _{0}}^{2}\left( x_{k}\right) \mathcal{J}% _{k}\left( \widetilde{x}_{i}\right) \right) \left( \widehat{b}-\beta \right) ^{\prime }H^{\prime }\left( \widetilde{b}\right) \text{,} \end{eqnarray*} where $\widetilde{b}$ is an intermediate point between $\beta$ and $\widehat{b}$. The norm of the third term on the right is $O_{p}\left( 1\right) $ whereas the second term on the right is $O_{p}\left( \left( n/L\right) ^{1/2}\right) $. So, using the pseudo inverse of sum of matrices formula, it is easily seen that \begin{equation} \mathcal{D}_{n}^{+}\left( \widehat{b};i\right) -\mathcal{D}% _{n}^{+}\left( \beta ;i\right) =O_{p}\left( \left( n/L\right) ^{1/2}\right) \text{.} \label{5} \end{equation}% Now, we examine $\mathcal{A}_{q,n}\left( x;\widehat{b}\right) -\mathcal{% A}_{q,n}\left( x;\beta \right) $, $q=1,2,3$. Let's examine e.g. the behaviour $\mathcal{A}_{1,n}\left( x;\widehat{b}\right) -\mathcal{A}% _{1,n}\left( x;\beta \right) $. By definition, it is \begin{eqnarray*} \mathcal{A}_{1,n}\left( x;\widehat{b}\right) -\mathcal{A}_{1,n}\left( x;\beta \right) &=&\frac{1}{n^{1/2}}\sum_{i=1}^{n}\left\{ \widetilde{P}% _{i}^{\prime }\left( \widehat{b}\right) \mathcal{D}_{n}^{+}\left( \widehat{b};i\right) \sum_{k=1}^{n}\widetilde{P}_{k}\left( \widehat{b }\right) u_{k}\mathcal{J}_{k}\left( \widetilde{x}_{i}\right) \right. \\ &&\text{ \ \ \ \ \ \ \ \ \ \ \ \ \ \ }\left. -\widetilde{P}_{i}^{\prime }\left( \beta \right) \mathcal{D}_{n}^{+}\left( \beta ;i\right) \sum_{k=1}^{n}\widetilde{P}_{k}\left( \beta \right) u_{k}\mathcal{J}% _{k}\left( \widetilde{x}_{i}\right) \right\} \mathcal{I}_{i}(x)\text{.} \end{eqnarray*} It suffices to show that, uniformly in $x$, \begin{eqnarray*} \left( \mathbf{i}\right) \text{ \ \ }\frac{1}{n^{1/2}}\sum_{i=1}^{n}\left\{ \left( \widetilde{P}_{i}\left( \widehat{b}\right) -\widetilde{P}% _{i}\left( \beta \right) \right) ^{\prime }\mathcal{D}_{n}^{+}\left( \beta ;i\right) \sum_{k=1}^{n}\widetilde{P}_{k}\left( \beta \right) u_{k}\mathcal{J% }_{k}\left( \widetilde{x}_{i}\right) \right\} \mathcal{I}_{i}(x) &=&o_{p}\left( 1\right) \\ \left( \mathbf{ii}\right) \text{ \ \ }\frac{1}{n^{1/2}}\sum_{i=1}^{n}\left\{ \widetilde{P}_{i}^{\prime }\left( \beta \right) \mathcal{D}_{n}^{+}\left( \beta ;i\right) \sum_{k=1}^{n}\left( \widetilde{P}_{k}\left( \widehat{b}% \right) -\widetilde{P}_{k}\left( \beta \right) \right) u_{k}\mathcal{J}% _{k}\left( \widetilde{x}_{i}\right) \right\} \mathcal{I}_{i}(x) &=&o_{p}\left( 1\right) \\ \left( \mathbf{iii}\right) \text{ \ \ }\frac{1}{n^{1/2}}\sum_{i=1}^{n}\left% \{ \widetilde{P}_{i}^{\prime }\left( \beta \right) \left[ \mathcal{D}% _{n}^{+}\left( \widehat{b};i\right) -\mathcal{D}_{n}^{+}\left( \beta ;i\right) \right] \sum_{k=1}^{n}\widetilde{P}_{k}\left( \beta \right) u_{k}% \mathcal{J}_{k}\left( \widetilde{x}_{i}\right) \right\} \mathcal{I}_{i}(x) &=&o_{p}\left( 1\right) \text{.} \end{eqnarray*} The proof of (i)-(iii) follows quite easily based on $\left( \ref{3}\right) $ and $\left( \ref{5}\right) $, and that we have already shown that \begin{eqnarray*} \mathcal{D}_{n}^{+1/2}\left( \beta ;i\right) \sum_{k=i+1}^{n}\widetilde{P} _{k}\left( \beta \right) u_{k} &=&O_{p}\left( L^{1/2}\right) \text{.} \\ \underline{\lambda }\left( \frac{L}{n}\mathcal{D}_{n}\left( \beta ;i\right) \right) &>&0\text{ \ with probability }1\text{.} \end{eqnarray*} Analogously, we can easily establish that $\mathcal{A}_{q,n}\left( x; \widehat{b}\right) -\mathcal{A}_{q,n}\left( x; \beta\right) = o_p(1)$. Note that $\mathcal{A}_{4,n}\left( x;\widehat{b}\right) $ can be written as $\frac{1}{n^{1/2}}\sum_{i=1}^{n}m^{bias}(x_{i})\mathcal{I}_{i}(x)$ and in the same way as in the proof of Theorem \ref{M_nest} can be shown to be $o_{p}(1)$. $\blacksquare $ \subsection{\textbf{PROOF OF THEOREM \protect\ref{Mb_n}}} $\left. {}\right. $ Define% \begin{equation*} \mathcal{M}_{n}^{\ast }\left( x\right) =:\frac{1}{n^{1/2}}% \sum_{i=1}^{n}v_{i}^{\ast }\mathcal{I}_{i}\left( x\right) \text{,} \end{equation*}% where $v_{i}^{\ast }=u_{i}^{\ast }-\boldsymbol{P}_{i}^{\prime }A_{n,i}^{+}C_{n,i}^{\ast }$ and $C_{n,i}^{\ast }=:n^{-1}\sum_{k=1}^{n}% \boldsymbol{P}_{k}u_{k}^{\ast }\mathcal{J}_{k}\left( \widetilde{x}% _{i}\right) $. The proof is completed if we show that, (\emph{in probability}% ), \begin{equation*} \left( \mathbf{a}\right) \text{ }\mathcal{M}_{n}^{\ast }\left( x\right) \overset{weakly}{\Rightarrow }\mathcal{B}\left( F_{X}\left( x\right) \right) \end{equation*}% \begin{equation*} \left( \mathbf{b}\right) \text{ }\sup_{x\in \left( 0,1\right) }\left\vert \widetilde{\mathcal{M}}_{n}^{\ast }\left( x\right) -\mathcal{M}_{n}^{\ast }\left( x\right) \right\vert =o_{p^{\ast }}\left( 1\right) \text{.} \end{equation*} Part $\left( \mathbf{b}\right) $ holds true trivially using Lemma \ref{bias} because $\widehat{u}_{i}^{\ast }-\widehat{u}_{i}=\boldsymbol{P}_{i}^{\prime }\left( \widehat{b}^{\ast }-\widehat{b}\right) $. To show part $\left( \mathbf{a}\right) $ it suffices to show that $\left( \mathbf{a1}\right) $ the finite dimensional distributions converge to a Gaussian random variable with the appropriate covariance structure and $\left( \mathbf{a2}\right) $ tightness of the sequence. The proofs proceed similarly as those in Theorem % \ref{M_n}. Notice that in the proof of Theorem \ref{M_n}, we first conditioned on $X$ and then we examined its asymptotic unconditional limit. We begin with $\left( \mathbf{a1}\right) $. To that end, we first examine the structure of the second moments. Proceeding as in the proof of Theorem % \ref{M_n}, because $E^{\ast }u_{i}^{\ast 2}=:\widehat{\sigma }_{u}^{2}$, we have that% \begin{eqnarray*} E^{\ast }\left( \mathcal{M}_{n}^{\ast }\left( x^{1}\right) \mathcal{M}% _{n}^{\ast }\left( x^{2}\right) \right) &=&\frac{\widehat{\sigma }_{u}^{2}}{n% }\sum_{i=1}^{n}\left( 1-\frac{\boldsymbol{P}_{i}^{\prime }A_{n,i}^{+}% \boldsymbol{P}_{i}}{n}\right) \mathcal{I}_{i}\left( x^{1}\right) +o_{p}\left( 1\right) \\ &&\overset{P}{\rightarrow }\sigma _{u}^{2}F_{X}\left( x^{1}\right) \end{eqnarray*}% proceeding as in the proof of $\left( \ref{eg}\right) $ and by Proposition % \ref{sigma_est}. Next, we examine the weak convergence of $\mathcal{M}_{n}^{\ast }\left( x\right) $, which due to Cram\'{e}r-Wold device, it suffices to do so for a fixed $x$. First observe that we have shown that% \begin{equation*} \sum_{i=1}^{n}E^{\ast }\left( \upsilon _{in}^{\ast 2}\left( x\right) \mid X\right) \overset{P}{\rightarrow }1\text{,} \end{equation*}% where $\upsilon _{in}^{\ast }\left( x\right) =:\widehat{\sigma }% _{u}^{-1}F_{X}^{-1/2}\left( x\right) \mathcal{I}_{i}\left( x\right) v_{i}^{\ast }/n^{1/2}$ is a martingale difference triangular array of r.v.'s. So, to complete the proof, it suffices to show the Lindeberg's condition for which a sufficient condition is $\sum_{i=1}^{n}E^{\ast }\left\vert v_{i}^{\ast }\right\vert ^{4}=o_{p}\left( n^{2}\right) $, which follows proceeding as with the proof of Theorem \ref{M_n} and Proposition \ref% {sigma_est}. $\left( \mathbf{a2}\right) $ We now examine the tightness of% \begin{equation*} \frac{1}{n^{1/2}}\sum_{i=1}^{n}v_{i}^{\ast }\mathcal{I}_{i}\left( x\right) \text{.} \end{equation*}% The proof follows as that of Lemma \ref{tight} because again the proof there was done conditionally on $X$ and then we examined its asymptotic unconditional limit. Indeed, as we argued in the proof of Theorem \ref{M_n} part $\left( \mathbf{b}\right) $, it suffices to show that% \begin{equation} E^{\ast }\left( \frac{1}{n^{1/2}}\sum_{i=1}^{n}v_{i}^{\ast }\mathcal{I}% _{i}\left( x^{1};x^{2}\right) \right) ^{4}\overset{P}{\rightarrow }K\left\{ \frac{1}{n}\left( F_{X}\left( x^{2}\right) -F_{X}\left( x^{1}\right) \right) +\left( x^{2}-x^{1}\right) ^{2}\sup_{x\in \left( x^{1},x^{2}\right) }f_{X}^{2}\left( x\right) \right\}, \label{theo5_1} \end{equation} where $\mathcal{I}_{i}\left( x^{1},x^{2}\right) =\mathcal{I}_{i}\left( x^{2}\right) -\mathcal{I}_{i}\left( x^{1}\right) $. By Burkholder's inequality implies that the left side of $\left( \ref{theo5_1}\right) $ is bounded by% \begin{equation} \frac{K}{n^{2}}E\left( \sum_{i=1}^{n}\left( v_{i}^{\ast 2}-E\left( v_{i}^{\ast 2}\mid \mathcal{G}_{i}^{\ast }\right) \right) \mathcal{I}% _{i}\left( x^{1},x^{2}\right) \right) ^{2}+\frac{K}{n^{2}}\left( \sum_{i=1}^{n}E\left( v_{i}^{\ast 2}\mid \mathcal{G}_{i}^{\ast }\right) \mathcal{I}_{i}\left( x^{1},x^{2}\right) \right) ^{2}\text{,} \label{burk_2} \end{equation}% where $\mathcal{G}_{i}^{\ast }$ denotes the sigma algebra generated by $% \left\{ u_{i+1}^{\ast },...,u_{n}^{\ast }\right\} $ and $E\left( v_{i}^{\ast 2}\mid \mathcal{G}_{i}^{\ast }\right) =\widehat{\sigma }_{u}^{2}+\boldsymbol{% P}_{i}^{\prime }A_{n,i}^{+}C_{n,i}^{\ast }C_{n,i}^{\ast \prime }A_{n,i}^{+}% \boldsymbol{P}_{i}$ as it is easily seen. From here the proof proceeds as that of Lemma \ref{tight} after we observe that the only difference is that we have, say $\widehat{\sigma }_{u}^{2}$ instead of $\sigma _{u}^{2}$, and observing that $\widehat{\sigma }_{u}^{2}-\sigma _{u}^{2}=o_{p}\left( 1\right) $ by Proposition \ref{sigma_est} and that \begin{equation*} \kappa _{4}\left( u^{\ast }\right) =\frac{1}{n}\sum_{i=1}^{n}\widehat{u}% _{i}^{4}-3\left( \frac{1}{n}\sum_{i=1}^{n}\widehat{u}_{i}^{2}\right) ^{2}% \overset{P}{\rightarrow }Eu_{i}^{4}-3\left( Eu_{i}^{2}\right) ^{2}\text{,} \end{equation*}% so that the left side of $\left( \ref{burk_2}\right) $ is bounded by% \begin{equation*} K\left\{ \frac{1}{n}\left( F_{X}\left( x^{2}\right) -F_{X}\left( x^{1}\right) \right) +\left( x^{2}-x^{1}\right) ^{2}\sup_{z\in \left( x^{1},x^{2}\right) }f_{X}^{2}\left( z\right) \right\} \left( 1+o_{p}\left( 1\right) \right) \end{equation*}% and the conclusion follows by using the same arguments as those for the proof of Billingsley's \cite{Billingsley} Theorems 22.1, see also arguments in Wu's \cite{Wu} Lemma 14.\hfill $\blacksquare $ \section{\textbf{APPENDIX B}} In what follows we shall abbreviate $p_{\ell ,L}\left( x;q\right) $ by $% p_{\ell ,L}\left( x\right) $ for all $\ell =1,...,L$. \begin{lemma} \label{bias}Any linear combination of the \emph{B-splines}, $\boldsymbol{p}% _{L}\left( x\right) =:\sum_{\ell =1}^{L}a_{\ell }p_{\ell ,L}\left( x\right) $% , satisfies that $\left( \mathcal{T}_{n}\mathcal{B}_{n,L}\right) \left( x\right) =0$, where $\mathcal{B}_{n,L}\left( x\right) =n^{-1}\sum_{k=1}^{n}% \boldsymbol{p}_{L}\left( x_{k}\right) \mathcal{I}_{k}\left( x\right) $. \end{lemma} \begin{proof} The proof is immediate after we notice that $\int_{x_{i}}^{1}\boldsymbol{P}% _{L}\left( w\right) \mathcal{W}_{n}\left( dw\right) =A_{n,i}$. Indeed, $% \left( \ref{Khma_n}\right) $ implies that $\left( \mathcal{T}_{n}\mathcal{B}% _{n,L}\right) \left( x\right) $ is% \begin{eqnarray*} &&\frac{1}{n}\sum_{k=1}^{n}\left\{ \boldsymbol{p}_{L}\left( x_{k}\right) -% \boldsymbol{P}_{k}^{\prime }A_{n,k}^{+}\frac{1}{n}\sum_{j=1}^{n}\boldsymbol{P% }_{j}\boldsymbol{p}_{L}\left( x_{j}\right) \mathcal{J}_{j}\left( \widetilde{x% }_{k}\right) \right\} \mathcal{I}_{k}\left( x\right) \\ &=&\frac{1}{n}\sum_{k=1}^{n}\left( \boldsymbol{P}_{k}^{\prime }-\boldsymbol{P% }_{k}^{\prime }A_{n,k}^{+}\frac{1}{n}\sum_{j=1}^{n}\boldsymbol{P}_{j}% \boldsymbol{P}_{j}^{\prime }\mathcal{J}_{j}\left( \widetilde{x}_{k}\right) \right) a\mathcal{I}_{k}\left( x\right) \\ &=&\frac{1}{n}\sum_{k=1}^{n}\left( \boldsymbol{P}_{k}^{\prime }-\boldsymbol{P% }_{k}^{\prime }A_{n,k}^{+}A_{n,k}\right) a\mathcal{I}_{k}\left( x\right) \\ &=&0\text{,} \end{eqnarray*}% by Harville's \cite{Harville} Theorem 12.3.4, where $a=\left( a_{1},...,a_{L}\right) $. \end{proof} We now introduce some notation useful for the next lemmas. We shall denote $% \Lambda \left( r\right) =\left\{ x:z^{r-1}\leq x<z^{r}-n^{-\varsigma }\right\} $ and $\overline{\Lambda }\left( r\right) =\left\{ x:z^{r}-n^{-\varsigma }\leq x<z^{r}\right\} $. \begin{lemma} \label{Nor_P_i}Under Condition $C1$, we have that $E\left\Vert \widetilde{% \boldsymbol{P}}_{i}\right\Vert ^{s}=O\left( L^{s/2}+Ln^{\left( s/2-1\right) \varsigma }\right) $ for any $s\geq 2$, with $\widetilde{\boldsymbol{P}}_{i}$ given in $\left( \ref{q_1}\right) $. \end{lemma} \begin{proof} First, \begin{equation} E\left\Vert \widetilde{\boldsymbol{P}}_{i}\right\Vert ^{s}=\sum_{k=1}^{L}E\left\{ \left\Vert \widetilde{\boldsymbol{P}}% _{i}\right\Vert ^{s}\mathcal{I}_{i}\left( z^{k-1};z^{k}\right) \right\} \text{,} \label{lem_2_p} \end{equation}% where, for each $k=1,...,L$, we have that \begin{eqnarray} E\left\{ \left\Vert \widetilde{\boldsymbol{P}}_{i}\right\Vert ^{s}\mathcal{I}% _{i}\left( z^{k-1};z^{k}\right) \right\} &=&E\left\{ \left\Vert \overline{A}% _{L,i}^{+1/2}D_{L}\left( \widetilde{x}_{i}\right) \mathbf{P}_{i}\right\Vert ^{s}\mathcal{I}_{i}\left( z^{k-1};z^{k}\right) \right\} \notag \\ &=&KE\left\{ \left\Vert D_{L}\left( x_{i}\right) \mathbf{P}_{i}\right\Vert ^{s}\mathcal{I}\left( x_{i}\in \Lambda \left( k\right) \right) \right\} \label{Lemma2_3} \\ &&+KE\left\{ \left\Vert D_{L}\left( z^{k}\right) \mathbf{P}_{i}\right\Vert ^{s}\mathcal{I}\left( x_{i}\in \overline{\Lambda }\left( k\right) \right) \right\} \notag \end{eqnarray}% because $\overline{A}_{L,i}=:diag\left( 0,B_{L,i}\right) $ and \underline{$% \lambda $}$\left( B_{L,i}\right) >0$. The proof is now standard after observing that the first term on the right of $\left( \ref{Lemma2_3}\right) $ is bounded by \begin{eqnarray} &&E\left\Vert \frac{p_{k,L}^{s}\left( x_{i}\right) \mathcal{I}\left( x_{i}\in \Lambda \left( k\right) \right) }{L^{sq}\left( z^{k}-x_{i}\right) ^{s\left( q+1/2\right) }}\right\Vert +K\sum_{j=k+1}^{k+q}E\left\Vert L^{s/2}p_{j,L}^{s}\left( x_{i}\right) \mathcal{I}\left( x_{i}\in \Lambda \left( k\right) \right) \right\Vert \notag \\ &=&K\left( E\left\Vert x_{i}^{-s/2}\mathcal{I}_{i}\left( n^{-\varsigma };L^{-1}\right) \right\Vert +L^{s/2-1}\right) \label{Lemma2_4} \\ &=&O\left( n^{\left( s/2-1\right) \varsigma }+L^{s/2-1}\right) \text{.} \notag \end{eqnarray}% since $E\left( p_{j,L}^{s}\left( x_{i}\right) \right) =O\left( L^{-1}\right) $ and $p_{k,L}\left( x_{i}\right) \leq KL^{q}\left( z^{k}-x_{i}\right) ^{q}$% , whereas the second term on the right of $\left( \ref{Lemma2_3}\right) $ is bounded by% \begin{equation*} K\sum_{j=k+1}^{k+q}E\left\Vert L^{s/2}p_{j,L}^{s}\left( x_{i}\right) \mathcal{I}\left( x_{i}\in \overline{\Lambda }\left( k\right) \right) \right\Vert =O\left( L^{s/2-1}\right) \end{equation*}% proceeding as with the second term on the left of $\left( \ref{Lemma2_4}% \right) $. \end{proof} \begin{lemma} \label{Est_A_i}Under Conditions $C1$ and $C3$, we have that \begin{equation} \left\Vert A_{L,i}^{+1/2}A_{n,i}A_{L,i}^{+1/2}-I\right\Vert _{E}\overset{P}{% \rightarrow }0\text{.} \label{lem_51} \end{equation} \end{lemma} \begin{proof} We shall consider the scenario where $x_{i}\in \Lambda =:\left\{ \Lambda \left( r\right) ;\text{ }r=1,...,L\right\} $, the case when $x_{i}\in \overline{\Lambda }=:\left\{ \overline{\Lambda }\left( r\right) ;\text{ }% r=1,...,L\right\} $ is handled similarly, if not easier. Because when $% x_{i}\in \Lambda \left( r\right) $, $\widetilde{x}_{i}=x_{i}$, the matrix inside the norm in $\left( \ref{lem_51}\right) $ is $\overline{A}% _{L,i}^{+1/2}H_{n,i}\overline{A}_{L,i}^{+1/2}$, where \begin{equation} H_{n,i}=:D_{L}\left( x_{i}\right) \left( \frac{1}{n}\sum_{k=1;\neq i}^{n}% \boldsymbol{P}_{k}\boldsymbol{P}_{k}^{\prime }\mathcal{J}_{k}\left( x_{i}\right) -A_{L,i}\right) D_{L}\left( x_{i}\right) +\frac{1}{n}% D_{L}\left( x_{i}\right) \boldsymbol{P}_{i}\boldsymbol{P}_{i}^{\prime }D_{L}\left( x_{i}\right) \text{.} \label{h_n} \end{equation} The second term on the right of $\left( \ref{h_n}\right) $ is $o_{p}\left( 1\right) $ as we now show. Because \begin{equation} p_{\ell ,L}\left( x_{k}\right) p_{m,L}\left( x_{k}\right) =0\text{ \ \ if }% m\geq \ell +q\text{,} \label{p_p} \end{equation}% Cauchy-Schwarz and then Markov's inequalities implies that it suffices to show that \begin{equation} E\left\vert \frac{1}{n}\sum_{\ell =1}^{L}d_{\ell }^{2}\left( x_{i}\right) p_{\ell ,L}^{2}\left( x_{i}\right) \right\vert =o\left( 1\right) \text{.} \label{8_3_1} \end{equation}% But this is the case because, recall that $x_{i}\in \Lambda $, the left side is bounded by% \begin{equation*} \frac{K}{n}\sum_{r=1}^{L}E\left( \frac{1}{\left( z^{r}-x_{i}\right) }% +\sum_{\ell =r+1}^{r+q}Lp_{\ell ,L}^{2}\left( x_{i}\right) \right) \mathcal{I% }\left( x_{i}\in \Lambda \left( r\right) \right) =O\left( \frac{L\left( 1+\varsigma \log n\right) }{n}\right) =o\left( 1\right) \text{,} \end{equation*}% since $E\left( Lp_{\ell ,L}^{2}\left( x_{i}\right) \mathcal{I}\left( x_{i}\in \Lambda \left( r\right) \right) \right) =O\left( 1\right) $. It is worth observing that $\left( \ref{8_3_1}\right) $ also holds uniformly in $x$ since $p_{r,L}^{2}\left( x\right) <K$ and Condition $C3$. Next, the first term on the right of $\left( \ref{h_n}\right) $ is also $% o_{p}\left( 1\right) $. Indeed, because $\overline{A}_{L,i}=:diag\left( 0,B_{L,i}\right) $ and \underline{$\lambda $}$\left( B_{L,i}\right) >0$, it suffices to show that $\left\Vert H_{n,i}\right\Vert _{E}\overset{P}{% \rightarrow }0$ and more specifically, in view of $\left( \ref{p_p}\right) $% , to show that \begin{equation} \sum_{r=1}^{L}\left\{ \sum_{\ell =r}^{L}\frac{1}{n}\sum_{k=1;\neq i}^{n}d_{\ell }^{2}\left( x_{i}\right) \left( p_{\ell ,L}^{2}\left( x_{k}\right) \mathcal{J}_{k}\left( x_{i}\right) -a_{i,\ell \ell }\right) \right\} \mathcal{I}\left( x_{i}\in \Lambda \left( r\right) \right) \overset{% P}{\rightarrow }0\text{.} \label{H_i} \end{equation}% But $\left( \ref{H_i}\right) $ holds true because, conditionally on $x_{i}$, when $\ell =r$, we have that Condition $C1$ and $\mathcal{I}\left( x_{i}\in \Lambda \left( r_{1}\right) \right) \mathcal{I}\left( x_{i}\in \Lambda \left( r_{2}\right) \right) =0$ if $r_{1}\neq r_{2}$ implies that \begin{eqnarray*} &&E\left( \sum_{r=1}^{L}\frac{1}{n}\sum_{k=1;\neq i}^{n}\frac{\left( \overline{p}_{r,L}^{2}\left( x_{k}\right) \mathcal{J}_{k}\left( x_{i}\right) -E\left( \overline{p}_{r,L}^{2}\left( x_{k}\right) \mathcal{J}_{k}\left( x_{i}\right) \right) \right) }{\left( z^{r}-x_{i}\right) ^{2q+1}}\mathcal{I}% \left( x_{i}\in \Lambda \left( r\right) \right) \right) ^{2} \\ &=&\sum_{r=1}^{L}\frac{K}{n^{2}}\sum_{k=1;\neq i}^{n}\frac{E\overline{p}% _{r,L}^{4}\left( x_{k}\right) \mathcal{J}_{k}\left( x_{i}\right) -E^{2}\left( \overline{p}_{r,L}^{2}\left( x_{k}\right) \mathcal{J}_{k}\left( x_{i}\right) \right) }{\left( z^{r}-x_{i}\right) ^{4q+2}}\mathcal{I}\left( x_{i}\in \Lambda \left( r\right) \right) \\ &=&\frac{K}{n}\sum_{r=1}^{L}\left( z^{r}-x_{i}\right) ^{-1}\mathcal{I}\left( x_{i}\in \Lambda \left( r\right) \right) =O_{p}\left( \frac{L\log n}{n}% \right) \text{,} \end{eqnarray*}% because $\overline{p}_{r,L}\left( x_{k}\right) \mathcal{J}_{k}\left( x_{i}\right) \mathcal{I}\left( x_{i}\in \Lambda \left( r\right) \right) \leq K\left( z^{r}-x_{k}\right) ^{q}\mathcal{J}_{k}\left( x_{i}\right) \mathcal{I}% \left( x_{i}\in \Lambda \left( r\right) \right) $ and Markov's inequality, where \begin{equation} \overline{p}_{r,L}\left( x_{k}\right) =L^{-q}p_{r,L}\left( x_{k}\right) \label{p_bar} \end{equation}% and $E\left\{ \left( z^{r}-x_{i}\right) ^{-1}\mathcal{I}\left( x_{i}\in \Lambda \left( r\right) \right) \right\} =O\left( \log n\right) $. Next when $\ell >r$. Denoting $d_{\ell }^{2}\left( x_{i}\right) \left( p_{\ell ,L}^{2}\left( x_{k}\right) \mathcal{J}_{k}\left( x_{i}\right) -a_{i,\ell \ell }\right) =:\psi _{i,\ell }\left( x_{k}\right) $, the second conditional moments of the left side of $\left( \ref{H_i}\right) $ is% \begin{equation*} \frac{1}{n^{2}}\sum_{r=1}^{L}\left\{ \sum_{\ell _{1},\ell _{2}=r}^{L}\sum_{k=1;\neq i}^{n}E\left( \psi _{i,\ell _{1}}\left( x_{k}\right) \psi _{i,\ell _{2}}\left( x_{k}\right) \mid x_{i}\right) \right\} \mathcal{I}\left( x_{i}\in \Lambda \left( r\right) \right) =O\left( \frac{L^{2}}{n}\right) \end{equation*}% because $a_{i,\ell \ell }=O\left( L^{-1}\right) $, $d_{\ell }^{2}\left( x_{i}\right) =L$, $Ep_{\ell ,L}^{s}\left( x_{k}\right) =O\left( L^{-1}\right) $ for any $s\geq 1$ and $\left( \ref{p_p}\right) $ implies that \begin{equation*} \left\vert E\left( \psi _{i,\ell _{1}}\left( x_{k}\right) \psi _{i,\ell _{2}}\left( x_{k}\right) \mid x_{i}\right) \right\vert \leq K\mathcal{I}% \left( \left\vert \ell _{2}-\ell _{1}\right\vert \geq q\right) +KL\mathcal{I}% \left( \left\vert \ell _{2}-\ell _{1}\right\vert <q\right) \text{.} \end{equation*}% So, $\left( \ref{H_i}\right) $ holds true for these terms by Markov's inequality, Condition $C3$ and because $\sum_{r=1}^{L}\mathcal{I}\left( x_{i}\in \Lambda \left( r\right) \right) =1$, which concludes the proof of the lemma. \end{proof} \begin{lemma} \label{Est_A_uni}Under Conditions $C1$ and $C3$, we have that \begin{equation*} \sup_{x}\left\Vert A_{L}^{+1/2}\left( x\right) A_{n}\left( x\right) A_{L}^{+1/2}\left( x\right) -I\right\Vert _{E}\overset{P}{\rightarrow }0% \text{.} \end{equation*} \end{lemma} \begin{proof} Arguing as we did in the proof of Lemma \ref{Est_A_i}, it suffices to show that \begin{equation} \sup_{x}\left\Vert H_{n}\left( x\right) \right\Vert _{E}=:\sup_{1\leq r\leq L}\left\{ \sup_{x\in \Lambda \left( r\right) }+\sup_{x\in \overline{\Lambda }% \left( r\right) }\right\} \left\Vert H_{n}\left( x\right) \right\Vert _{E}% \overset{P}{\rightarrow }0\text{.} \label{Lemma_5_1} \end{equation}% We begin examining the contribution due to the sets $\Lambda \left( r\right) $. First, as we argued in Lemma \ref{Est_A_i}, we observe that $\left( \ref% {Lemma_5_1}\right) $ holds true if the diagonal elements of $H_{n}\left( x\right) $ converges uniformly to zero in probability. That is, it suffices to show that \begin{eqnarray} &&\sup_{1\leq r\leq L}\sup_{x\in \Lambda \left( r\right) }\sum_{\ell =r+q}^{L}\left\vert \frac{1}{n}\sum_{k=1}^{n}d_{\ell }^{2}\left( x\right) \left( p_{\ell ,L}^{2}\left( x_{k}\right) \mathcal{J}_{k}\left( x\right) -Ep_{\ell ,L}^{2}\left( x_{k}\right) \mathcal{J}_{k}\left( x\right) \right) \right\vert ^{2} \label{L_4} \\ &&+\left( \sup_{1\leq r\leq L}\sup_{x\in \Lambda \left( r\right) }\sum_{\ell =r}^{r+q-1}\left\vert \frac{1}{n}\sum_{k=1}^{n}d_{\ell }^{2}\left( x\right) \left( p_{\ell ,L}^{2}\left( x_{k}\right) \mathcal{J}_{k}\left( x\right) -Ep_{\ell r,L}^{2}\left( x_{k}\right) \mathcal{J}_{k}\left( x\right) \right) \right\vert \right) ^{2} \notag \\ &=&o_{p}\left( 1\right) \text{.} \notag \end{eqnarray}% Recall that due to the definition of \emph{B-splines}, for any $r$, $p_{\ell ,L}\left( x_{k}\right) \mathcal{J}_{k}\left( x\right) \mathcal{I}\left( x\in \Lambda \left( r\right) \right) =0$ if $\ell <r$. Because $x\in \Lambda \left( r\right) $ implies that $p_{\ell ,L}\left( x_{k}\right) \mathcal{J}_{k}\left( x\right) =p_{\ell ,L}\left( x_{k}\right) \mathcal{J}_{k}\left( z^{r}\right) $ and $d_{\ell }^{2}\left( x\right) =L$ for $\ell \geq r+q$, the first term on the left of $\left( \ref{L_4}\right) $ is \begin{eqnarray*} &&\sup_{1\leq r\leq L}\sum_{\ell =r+q}^{L}\left\vert \frac{L}{n}% \sum_{k=1}^{n}\left( p_{\ell ,L}^{2}\left( x_{k}\right) \mathcal{J}% _{k}\left( z^{r}\right) -Ep_{\ell ,L}^{2}\left( x_{k}\right) \mathcal{J}% _{k}\left( z^{r}\right) \right) \right\vert ^{2} \\ &\leq &\sum_{\ell =1}^{L}\left\vert \frac{L}{n}\sum_{k=1}^{n}\left( p_{\ell ,L}^{2}\left( x_{k}\right) \mathcal{J}_{k}\left( z^{\ell }\right) -Ep_{\ell ,L}^{2}\left( x_{k}\right) \mathcal{J}_{k}\left( z^{\ell }\right) \right) \right\vert ^{2} \\ &=&O_{p}\left( L^{2}/n\right) \end{eqnarray*}% by Condition $C1$ and that $Ep_{\ell ,L}^{4}\left( x_{k}\right) =O\left( L^{-1}\right) $, and then by Markov's inequality. So, to complete the proof we need to show that the second term on the left of $\left( \ref{L_4}\right) $ is also $o_{p}\left( 1\right) $. We shall look at the case when $\ell =r$, being the cases when $\ell >r$ similarly, if not easier, handled. To that end, we first notice that this term is bounded by \begin{equation} \left( \sup_{1\leq r\leq L}\sup_{x\in \Lambda \left( r\right) }\frac{1}{% \left( z^{r}-x\right) ^{1/2}}\left\vert \frac{1}{n}\sum_{k=1}^{n}\eta _{r}\left( x_{k};x\right) -E\eta _{r}\left( x_{k};x\right) \right\vert \right) ^{2}\text{,} \label{L_1} \end{equation}% where $\eta _{r}\left( x_{k};x\right) =\overline{p}_{r,L}^{2}\left( x_{k}\right) \mathcal{J}_{k}\left( x\right) /\left( z^{r}-x\right) ^{2q+1/2}$ with $\overline{p}_{r,L}\left( x_{k}\right) $ defined in $\left( \ref{p_bar}% \right) $. Consider points $z^{r}\left[ j\right] =z^{r}-\left( j-1\right) /n$, $% j=n^{1-\varsigma }+1,...,n/L$, so that $z^{r}\left[ n^{1-\varsigma }+1\right] =z^{r}-n^{-\varsigma }$ and denote $\eta _{r}\left( x_{k};x^{1},x^{2}\right) =\eta _{r}\left( x_{k};x^{2}\right) -\eta _{r}\left( x_{k};x^{1}\right) $. Now if $x\in \left( z^{r}\left[ j+1\right] ,z^{r}\left[ j\right] \right) $, we have that \begin{equation*} \left\vert \frac{1}{n}\sum_{i=1}^{n}\eta _{r}\left( x_{i};x,z^{r}\left[ j% \right] \right) \right\vert \leq \frac{1}{n}\sum_{i=1}^{n}\left\vert \eta _{r}\left( x_{i};z^{r}\left[ j+1\right] ,z^{r}\left[ j\right] \right) \right\vert \text{,} \end{equation*}% so that the triangle inequality yields that $\left( \ref{L_1}\right) $ is, except constants, bounded by% \begin{eqnarray*} &&\left( \sup_{1\leq r\leq L}\sup_{n^{1-\varsigma }+1\leq j\leq n/L}\frac{1}{% \left( z^{r}-z^{r}\left[ j\right] \right) ^{1/2}}\left\vert \frac{1}{n}% \sum_{i=1}^{n}\eta _{r}\left( x_{i};z^{r}\left[ j\right] ,z^{r}\left[ n^{1-\varsigma }+1\right] \right) -E\left( \cdot \right) \right\vert \right) ^{2} \\ &&+\left( \sup_{1\leq r\leq L}\sup_{n^{1-\varsigma }+1\leq j\leq n/L}\frac{1% }{\left( z^{r}-z^{r}\left[ j\right] \right) ^{1/2}}\left\vert \frac{1}{n}% \sum_{i=1}^{n}\left\vert \eta _{r}\left( x_{i};z^{r}\left[ j+1\right] ,z^{r}% \left[ j\right] \right) \right\vert -E\left( \cdot \right) \right\vert \right) ^{2} \end{eqnarray*}% \begin{equation} +\left( \sup_{1\leq r\leq L}\sup_{n^{1-\varsigma }+1\leq j\leq n/L}\frac{1}{% \left( z^{r}-z^{r}\left[ j\right] \right) ^{1/2}}E\left\vert \eta _{r}\left( x_{i};z^{r}\left[ j+1\right] ,z^{r}\left[ j\right] \right) \right\vert \right) ^{2}\text{.} \label{L_2} \end{equation} Because for any $q>0$, \begin{equation} E\left\vert \eta _{r}\left( x_{i};z^{r}\left[ j+1\right] ,z^{r}\left[ j% \right] \right) \right\vert ^{q}\leq K\int_{z^{r}\left[ j+1\right] }^{z^{r}% \left[ j\right] }\left( z^{r}-x\right) ^{-q/2}dx\leq K\left( z^{r}-z^{r}% \left[ j\right] \right) ^{-q/2}n^{-1}\text{,} \label{L_3} \end{equation}% the last term of $\left( \ref{L_2}\right) $ is bounded by $K\left( \sup_{1\leq r\leq L}\sup_{n^{1-\varsigma }+1\leq j\leq n/L}n^{\varsigma -1}\right) ^{2}=o\left( 1\right) $ because $\varsigma <1$. Next using $\left( \ref{L_3}\right) $ and the inequalities $\left( \sup_{x}\left\vert g\left( x\right) \right\vert \right) ^{q}=\sup_{x}\left\vert g\left( x\right) \right\vert ^{q}$ and $\sup_{\ell }\left\vert c_{\ell }\right\vert \leq \sum_{\ell }\left\vert c_{\ell }\right\vert $, the expectation of the second term of $\left( \ref{L_2}% \right) $ to the power $p/2$ is bounded by \begin{eqnarray*} &&\sum_{r=1}^{L}\sum_{j=n^{1-\varsigma }+1}^{n/L}\frac{1}{\left( z^{r}-z^{r}% \left[ j\right] \right) ^{p/2}}E\left\vert \frac{1}{n}\sum_{i=1}^{n}\left% \vert \eta _{r}\left( x_{i};z^{r}\left[ j+1\right] ,z^{r}\left[ j\right] \right) \right\vert -E\left( \cdot \right) \right\vert ^{p} \\ &=&K\sum_{r=1}^{L}\frac{1}{n^{p}}\sum_{j=n^{1-\varsigma }+1}^{n/L}\frac{1}{% \left( z^{r}-z^{r}\left[ j\right] \right) ^{p}} \\ &=&KL\sum_{j=n^{1-\varsigma }+1}^{n/L}j^{-p}=O\left( Ln^{\left( p-1\right) \left( \varsigma -1\right) }\right) =o\left( 1\right) \end{eqnarray*}% for any $\varsigma <1$ choosing $p$ large enough. Thus, the second term of $% \left( \ref{L_2}\right) $ is $o_{p}\left( 1\right) $. Finally, the first term of $\left( \ref{L_2}\right) $. First, as we argued with the second term, Condition $C1$ yields that its $p-th$ absolute moment is bounded by \begin{eqnarray*} &&\sum_{r=1}^{L}\sum_{j=n^{1-\varsigma }+1}^{n/L}\frac{1}{\left( z^{r}-z^{r}% \left[ j\right] \right) ^{p/2}}E\left\vert \frac{1}{n}\sum_{i=1}^{n}\eta _{r}\left( x_{i};z^{r}\left[ j\right] ,z^{r}\left[ n^{1-\varsigma }+1\right] \right) -E\left( \cdot \right) \right\vert ^{p} \\ &=&\frac{K}{n^{p/2}}\sum_{r=1}^{L}\sum_{j=n^{1-\varsigma }+1}^{n/L}\frac{% \log ^{p/2}n}{\left( z^{r}-z^{r}\left[ j\right] \right) ^{p/2}} \\ &=&O\left( \frac{L\log n}{n^{\left( p/2-1\right) \left( 1-\varsigma \right) }% }\right) \end{eqnarray*}% which is $o\left( 1\right) $ since we can always choose $p$ large enough such that $L=o\left( n^{p/2\left( 1-\varsigma \right) }\right) $ for any $% \varsigma <1$ and Condition $C3$. Notice that we can also bound the left side of $\left( \ref{L_3}\right) $ by $K\log n$ when $q=2$ there. To complete the proof of the lemma, we need to examine the contribution due to the sets $\overline{\Lambda }\left( r\right) $ into the left of $\left( % \ref{Lemma_5_1}\right) $. However, observing that when $x_{i}\in \overline{% \Lambda }\left( r\right) $, $\mathcal{J}_{k}\left( x_{i}\right) =\mathcal{J}% _{k}\left( z^{r}\right) $, the proof proceeds as that of the first term on the left of $\left( \ref{L_4}\right) $, and so it is omitted. \end{proof} \begin{lemma} \label{tight}Under Conditions $C1$ and $C3$, we have that for any $0\leq x^{1}<x^{2}\leq 1$, \begin{eqnarray} &&E\left( \frac{1}{n^{1/2}}\sum_{i=1}^{n}v_{i}\mathcal{I}_{i}\left( x^{1},x^{2}\right) \right) ^{4} \label{lem_1} \\ &=&K\left\{ \frac{1}{n}\left( F_{X}\left( x^{2}\right) -F_{X}\left( x^{1}\right) \right) +\left( x^{2}-x^{1}\right) ^{2}\sup_{x\in \left( x^{1},x^{2}\right) }f_{X}^{2}\left( x\right) \right\} \left( 1+o\left( 1\right) \right) \text{.} \notag \end{eqnarray} \end{lemma} \begin{proof} First, as we mentioned in Section 3, we notice that we can arrange the observations according to $x$ without modifying the properties of \begin{equation*} \frac{1}{n^{1/2}}\sum_{i=1}^{n}v_{i}\mathcal{I}_{i}\left( x^{1},x^{2}\right) \text{,} \end{equation*}% so that $v_{i}$ becomes a martingale difference sequence of r.v.'s. So, Burkholder's inequality implies that the left side of $\left( \ref{lem_1}% \right) $ is bounded by% \begin{equation} \frac{K}{n^{2}}E\left( \sum_{i=1}^{n}\left( v_{i}^{2}-E\left( v_{i}^{2}\mid \mathcal{G}_{i}\right) \right) \mathcal{I}_{i}\left( x^{1},x^{2}\right) \right) ^{2}+\frac{K}{n^{2}}\left( \sum_{i=1}^{n}E\left( v_{i}^{2}\mid \mathcal{G}_{i}\right) \mathcal{I}_{i}\left( x^{1},x^{2}\right) \right) ^{2}% \text{,} \label{lem_2} \end{equation}% where $\mathcal{G}_{i}$ denotes the sigma algebra generated by $\left\{ u_{i+1},...,u_{n}\right\} $ and $E\left( v_{i}^{2}\mid \mathcal{G}% _{i}\right) =\sigma _{u}^{2}+\boldsymbol{P}_{i}^{\prime }A_{n,i}^{+}C_{n,i}C_{n,i}^{\prime }A_{n,i}^{+}\boldsymbol{P}_{i}$ as it is easily seen. We shall first examine the second term of $\left( \ref{lem_2}% \right) $. By standard inequalities, that term is bounded by% \begin{equation} K\sigma _{u}^{4}\left( \frac{1}{n}\sum_{i=1}^{n}\mathcal{I}_{i}\left( x^{1},x^{2}\right) \right) ^{2}+\frac{K}{n^{2}}\left( \sum_{i=1}^{n}% \boldsymbol{P}_{i}^{\prime }A_{n,i}^{+}C_{n,i}C_{n,i}^{\prime }A_{n,i}^{+}% \boldsymbol{P}_{i}\mathcal{I}_{i}\left( x^{1},x^{2}\right) \right) ^{2}\text{% .} \label{lem_4} \end{equation}% Because Condition $C1$ yields \begin{equation*} \frac{1}{n}\sum_{i=1}^{n}\mathcal{I}_{i}\left( x^{1},x^{2}\right) \overset{P}% {\rightarrow }F_{X}\left( x^{2}\right) -F_{X}\left( x^{1}\right) \text{,} \end{equation*}% we have that the contribution of the first term of $\left( \ref{lem_4}% \right) $ into the left of $\left( \ref{lem_1}\right) $ is% \begin{equation*} \left( F_{X}\left( x^{2}\right) -F_{X}\left( x^{1}\right) \right) ^{2}\leq K\left( x^{2}-x^{1}\right) ^{2}\sup_{x\in \left( x^{1},x^{2}\right) }f_{X}^{2}\left( x\right) \text{.} \end{equation*}% Next, the second term of $\left( \ref{lem_4}\right) $ is also $K\left( x^{2}-x^{1}\right) ^{2}\sup_{x\in \left( x^{1},x^{2}\right) }f_{X}^{2}\left( x\right) $, as we now show. Indeed, Cauchy-Schwarz inequality and Lemma \ref% {Est_A_uni}, see also $\left( \ref{A_inv}\right) $, yield that this term is bounded by% \begin{equation} \frac{1}{n}\sum_{i=1}^{n}\widetilde{\boldsymbol{P}}_{i}^{\prime }\left( \widetilde{C}_{n,i}\widetilde{C}_{n,i}^{\prime }\right) ^{2}\widetilde{% \boldsymbol{P}}_{i}\mathcal{I}_{i}\left( x^{1},x^{2}\right) ~\frac{1}{n}% \sum_{i=1}^{n}\widetilde{\boldsymbol{P}}_{i}^{\prime }\widetilde{\boldsymbol{% P}}_{i}\mathcal{I}_{i}\left( x^{1},x^{2}\right) \text{,} \label{lem_5_3} \end{equation}% where $\widetilde{C}_{n,i}=A_{n,i}^{+1/2}n^{-1}\sum_{j=i+1}^{n}\boldsymbol{P}% _{j}u_{j}$. Now the second factor of $\left( \ref{lem_5_3}\right) $ converges to \begin{equation} E\left( \left\Vert \widetilde{\boldsymbol{P}}_{i}\right\Vert ^{2}\mathcal{I}% _{i}\left( x^{1},x^{2}\right) \right) =O\left( n^{\varsigma }+L\right) \left( F_{X}\left( x^{2}\right) -F_{X}\left( x^{1}\right) \right) \label{Lema_5_1} \end{equation}% because \begin{equation} \left\Vert \widetilde{\boldsymbol{P}}_{i}\right\Vert ^{2}=O\left( n^{\varsigma }+L\right) \text{.} \label{p_norm} \end{equation}% Next, because \begin{equation*} E\left( \widetilde{C}_{n,i}\widetilde{C}_{n,i}^{\prime }\mid X\right) =\frac{% \sigma _{u}^{2}}{n}diag\left( \text{\b{0}},1,...,1\right) \text{,} \end{equation*}% we have that the conditional expectation of the first factor of $\left( \ref% {lem_5_3}\right) $ is% \begin{equation} \frac{3}{n}\sum_{i=1}^{n}\frac{\left\Vert \widetilde{\boldsymbol{P}}% _{i}\right\Vert ^{2}}{n^{2}}\mathcal{I}_{i}\left( x^{1},x^{2}\right) +\frac{% \kappa _{u}\left( u\right) }{n}\sum_{i=1}^{n}\frac{\left\Vert \widetilde{% \boldsymbol{P}}_{i}\right\Vert ^{2}}{n^{4}}\sum_{j=i+1}^{n}\left\Vert D_{L}\left( x_{i}\right) \boldsymbol{P}_{j}\mathcal{J}_{k}\left( \widetilde{x% }_{i}\right) \right\Vert ^{4}\mathcal{I}_{i}\left( x^{1},x^{2}\right) \label{Lema_5_2} \end{equation} \begin{equation*} \leq O\left( n^{\varsigma }+L\right) \left( \frac{3}{n^{3}}\sum_{i=1}^{n}% \mathcal{I}_{i}\left( x^{1},x^{2}\right) +\frac{1}{n}\sum_{i=1}^{n}\frac{1}{% n^{4}}\sum_{j=i+1}^{n}\left\Vert D_{L}\left( x_{i}\right) \boldsymbol{P}_{j}% \mathcal{J}_{k}\left( \widetilde{x}_{i}\right) \right\Vert ^{4}\mathcal{I}% _{i}\left( x^{1},x^{2}\right) \right) \end{equation*}% because $\overline{\lambda }\left( D_{L}^{-1}\left( x_{i}\right) A_{n,i}^{+1/2}\right) <K$ by Lemma \ref{Est_A_uni} and $\left( \ref{p_norm}% \right) $. By standard arguments, the first term on the right of $\left( \ref% {Lema_5_2}\right) $ is \begin{equation*} 3\frac{n^{\varsigma }+L}{n^{2}}\left( F_{X}\left( x^{2}\right) -F_{X}\left( x^{1}\right) \right) \left( 1+o_{p}\left( 1\right) \right) \end{equation*}% Next we examine the second term of $\left( \ref{Lema_5_2}\right) $. First we have that \begin{eqnarray*} &&\left\Vert D_{L}\left( x_{i}\right) \boldsymbol{P}_{j}\mathcal{J}% _{k}\left( \widetilde{x}_{i}\right) \right\Vert ^{4}\mathcal{I}_{i}\left( x^{1},x^{2}\right) \\ &=&\sum_{k=1}^{L}\left\Vert D_{L}\left( x_{i}\right) \boldsymbol{P}_{j}% \mathcal{J}_{k}\left( \widetilde{x}_{i}\right) \right\Vert ^{4}\left\{ \mathcal{I}\left( x_{i}\in \Lambda \left( k\right) \right) +\mathcal{I}% \left( x_{i}\in \overline{\Lambda }\left( k\right) \right) \right\} \mathcal{% I}_{i}\left( x^{1},x^{2}\right) \text{.} \end{eqnarray*}% We shall examine the contribution due to the first term on the right, the second is similarly handle. Now, the expectation of a typical term of the last displayed expression is, conditionally on $x_{i}$, \begin{eqnarray*} &&\sum_{k=1}^{L}E\left\Vert D_{L}\left( x_{i}\right) \boldsymbol{P}_{j}% \mathcal{J}_{k}\left( x_{i}\right) \right\Vert ^{4}\mathcal{I}\left( x_{i}\in \Lambda \left( k\right) \right) \\ &\leq &K\sum_{k=1}^{L}\left( \frac{E\overline{p}_{k\left( x_{i}\right) ,L}^{4}\left( x_{j}\right) \mathcal{I}_{i}\left( z^{k\left( x_{i}\right) -1},z^{k\left( x_{i}\right) }\right) }{\left( z^{k\left( x_{i}\right) }-x_{i}\right) ^{4q+2}}\mathcal{J}_{k}\left( x_{i}\right) \right) \mathcal{I}% \left( x_{i}\in \Lambda \left( k\right) \right) \\ &&+L^{2}\sum_{k=1}^{L}E\left( p_{\ell ,L}^{4}\left( x_{j}\right) \mathcal{J}% _{k}\left( x_{i}\right) \right) \mathcal{I}\left( x_{i}\in \Lambda \left( k\right) \right) \\ &=&\left( K\left( z^{k\left( x_{i}\right) }-x_{i}\right) ^{-1}+L\right) \sum_{k=1}^{L}\mathcal{I}\left( x_{i}\in \Lambda \left( k\right) \right) \\ &=&K\left( n^{\varsigma }+L\right) \text{,} \end{eqnarray*}% because $Ep_{\ell ,L}^{4}\left( x_{j}\right) =O\left( L^{-1}\right) $. So the second term of $\left( \ref{Lema_5_2}\right) $ is \begin{equation*} O\left( \left( n^{\varsigma }+L\right) \frac{1}{n^{3}}\right) \frac{1}{n}% \sum_{j=1}^{n}\mathcal{I}_{i}\left( x^{1},x^{2}\right) \end{equation*}% which implies that by standard arguments that it is% \begin{equation*} O\left( \left( n^{\varsigma }+L\right) \frac{1}{n^{3}}\right) \left( F_{X}\left( x^{2}\right) -F_{X}\left( x^{1}\right) \right) \left( 1+o_{p}\left( 1\right) \right) \text{.} \end{equation*} Thus, gathering terms, we have that the first factor of $\left( \ref{lem_5_3}% \right) $, i.e. $\left( \ref{Lema_5_2}\right) $, is \begin{equation*} O\left( \left( \frac{n^{\varsigma }+L}{n}\right) ^{2}\frac{1}{n}\right) \left( F_{X}\left( x^{2}\right) -F_{X}\left( x^{1}\right) \right) \left( 1+o_{p}\left( 1\right) \right) \text{.} \end{equation*}% So, the last displayed expression together with $\left( \ref{Lema_5_1}% \right) $ implies that the second term of $\left( \ref{lem_4}\right) $ is \begin{equation*} O\left( \left( \frac{n^{\varsigma }+L}{n}\right) ^{3}\right) \left( F_{X}\left( x^{2}\right) -F_{X}\left( x^{1}\right) \right) ^{2}\leq K\left( x^{2}-x^{1}\right) ^{2}\sup_{x\in \left( x^{1},x^{2}\right) }f_{X}^{2}\left( x\right) \text{.} \end{equation*} To complete the proof we need to examine the first term of $\left( \ref% {lem_2}\right) $, whose first moment is \begin{eqnarray*} \frac{K}{n^{2}}\sum_{i=1}^{n}E\left( \left( v_{i}^{2}-E\left( v_{i}^{2}\mid \mathcal{G}_{i}\right) \right) \mathcal{I}_{i}\left( x^{1},x^{2}\right) \right) ^{2} &\leq &\frac{K}{n^{2}}\sum_{i=1}^{n}E\left( v_{i}^{4}\mathcal{I}% _{i}\left( x^{1},x^{2}\right) \right) \\ &\leq &\frac{K}{n}\left( F_{X}\left( x^{2}\right) -F_{X}\left( x^{1}\right) \right) \end{eqnarray*}% by standard arguments as $Ev_{i}^{4}<K$ proceeding as we did with the second term of $\left( \ref{lem_2}\right) $.\newpage \end{proof}

\section{Introduction}\label{sec:introduction} Supersymmetry (SUSY)~\cite{Golfand:1971iw,Wess:1973kz,Wess:1974tw} is one of the best motivated extensions to the Standard Model (SM). According to the latest fits~\cite{Buchmueller:2008qe,Buchmueller:2009fn,Bechtle:2009ty} to the precision observables a rather light scale is indicated and is in reach of the Large Hadron Collider (LHC). If SUSY happens to be realised in nature, the task will be to pin down the specific characteristics of the underlying model which will require many detailed studies. One of the interesting issues in this context is CP violation. While the observed amount of CP violation in the $K$ and $B$ sectors can be accommodated within the SM, another piece of evidence, the baryon asymmetry of the universe, requires a new source of CP violation~\cite{Cohen:1993nk,Gavela:1994dt,Rubakov:1996vz}. The Minimal Supersymmetric Standard Model (MSSM)~\cite{MSSM,Haber:1984rc,Martin:1997ns} has about one hundred free parameters and a large number of these may have non-zero CP-violating phases, see e.g.\ \cite{Ibrahim:2007fb}. Many of the phases are unphysical in the sense that they can be rotated away by a redefinition of the fields but not all can be systematically removed in this way. For example in the sectors of the charginos and neutralinos, the supersymmetric partners of Higgs bosons and $SU(2)\times U(1)$ gauge bosons, we can have up to three new phases. These include the phases of the $U(1)$ gaugino mass parameter $M_1$, $SU(2)$ gaugino mass parameter $M_2$ and the Higgsino mass parameter $\mu$. One of them can be rotated away and therefore we choose the convention where $M_2$ is kept real. Hence we have \begin{equation} M_1 = |M_1| e^{i \phi_1}\,, \qquad \mu = |\mu| e^{i \phi_\mu} \,. \end{equation} The other possible sources of complex parameters in the MSSM Lagrangian are trilinear couplings $A_f$ but they are not relevant to this study~\cite{Ellis:2008hq}. Certain combinations of the CP-violating phases are constrained by the experimental upper bounds on various electric dipole moments (EDMs), see e.g.~\cite{Ellis:2008zy}. Ignoring possible cancellations, the most severely constrained phase is that of $\mu$ which contributes to the EDMs at the one-loop level and has a $\mathrm{tan}\beta$ enhancement. In general for $\mathcal{O}(100)$~GeV masses, $|\phi_{\mu}|\lesssim0.01\pi$ and we therefore set $\phi_{\mu}=0$ throughout our study. The phase of $M_1$ has weaker constraints. Large phases can be accommodated in models with heavy first and second generation squarks and sleptons ($>$TeV) or if accidental cancellations occur~\cite{Kizukuri:1992nj,Ibrahim:1998je,Ibrahim:1999af,Brhlik:1998zn,Abel:2001vy,Arnowitt:2001pm}. Here we study the complete range of CP phases in order to see the general dependencies exhibited by our observables and the luminosity required to observe these within the LHC environment. Nevertheless, we want to stress that in the chosen scenario experimental bounds from EDMs can be evaded by arranging cancellations between various supersymmetric contributions for any value of $\phi_1$~\cite{Lee:2003nta,Lee:2007gn,Ellis:2008zy}. CP-odd observables are the unambiguous way of discovering hints of complex parameters in the model. One class of these observables include rate asymmetries of cross sections, branching ratios and angular distributions. The other possibility are the observables that can be defined using triple products of momenta and/or spins of particles, see~\cite{Kittel:2009fg,Hesselbach:2007dq,Kraml:2007pr} for a recent review. In any case studies of CP-violating phases will be challenging at the LHC~\cite{Langacker:2007ur,Ellis:2008hq,Deppisch:2009nj} and a precise determination of them is expected to be only possible at a future $e^+e^-$ linear collider. An interesting example of using triple product correlations is studying the 3-body decay of a neutralino $\tilde{\chi}_2^0$~\cite{AguilarSaavedra:2004hu,AguilarSaavedra:2003ur,Choi:2005gt,Bartl:2004jj}. Since the asymmetry is maximal in the rest frame of the decaying neutralino, the central idea of~\cite{AguilarSaavedra:2004hu,AguilarSaavedra:2003ur} was the full kinematic reconstruction of the production and decay process. Hence the CP-sensitive observables are calculated~\cite{AguilarSaavedra:2004hu,AguilarSaavedra:2003ur} in the rest frame of the decaying neutralino at an $e^+e^-$ collider. This can be done due to the clean experimental environment and the well-known initial state at such a machine. As this is not the case at the LHC, these observables are much more challenging due to the proton structure and further experimental uncertainties, hence making the observation of CP-violating effects very difficult. However, in the present analysis we show how a similar approach can be applicable at hadron colliders. By applying full event reconstruction, the observation of CP-odd asymmetries becomes feasible. In many scenarios the highest cross section at the LHC is predicted for the associated production of gluinos and squarks of the first two generations. If squarks can decay to a neutralino $\tilde{\chi}_2^0$ in the process \begin{equation} \tilde{q}_i \to \tilde{\chi}^0_2 + q \end{equation} one can expect an abundance of neutralinos. Large statistics will help to overcome the limitations of a similar analysis of the stop carried out in \cite{Ellis:2008hq}. The production of the neutralino $\tilde{\chi}^0_2$ is followed by the subsequent three-body decay \begin{equation} \tilde{\chi}^0_2 \to \tilde{\chi}^0_1 \ell^+\ell^-. \end{equation} The 3-body decay is discussed since sizeable contributions to CP violation are expected from interference diagrams and it offers the opportunity to use CP-sensitive observables based on triple product correlations. The process is sensitive to any phase that enters the neutralino sector ($\phi_{1}$ and $\phi_{\mu}$), but we set $\phi_{\mu}$ to zero as discussed above and concentrate on the effect of $\phi_{1}$. To extract information about the phase of $M_1$ we analyse the following triple product of momenta from the decay products of the $\tilde{q}_i$, \begin{equation} \mathcal{T}= \vec{p}_q \cdot (\vec{p}_{\ell^+} \times \vec{p}_{\ell^-}). \label{eq:TripProd} \end{equation} It originates from the covariant product present in the amplitude and can be expanded in terms of explicit momenta, \begin{equation} \epsilon_{\mu\nu\rho\sigma}p_a^\mu p_b^\nu p_c^\rho p_d^\sigma \longrightarrow E_a\;\overrightarrow{p_b}\cdot(\overrightarrow{p_c}\times\overrightarrow{p_d}) \pm \dots \label{eq:EpsExpan} \end{equation} When CP phases appear a non-zero asymmetry can be constructed from final state momenta using the triple product in Eq.~(\ref{eq:TripProd}) as explained below. The triple product $\mathcal{T}$ changes sign under the na\"{\i}ve\footnote{Na\"{\i}ve time reversal $\mathrm{T}_N$ reverses the momenta and spins of all particles without interchanging the initial and final states, see e.g.~\cite{Atwood:2000tu}. Recall that under the true time reversal also the initial and final states are interchanged.} time reversal $\mathrm{T}_N$ and consequently it is a $\mathrm{T}_N$-odd observable. Neglecting possible absorptive phases due to loops, Breit-Wigner propagators and final state interactions, that are expected to be small in the leptonic neutralino decay, $\mathrm{CPT}_N$ invariance is equivalent to CPT invariance~\cite{Atwood:2000tu} and hence a measurement of a $\mathrm{T}_N$-odd observable implies CP violation. In addition as these are CP-odd quantities any measurement of the phase made by this method not only reveals the magnitude of the phase but also the sign. It is important to note that the triple product correlations are a tree-level effect and therefore large asymmetries can be expected~\cite{Bartl:2004jj,Bartl:2002hi,Bartl:2003ck,Bartl:2006yv}. To be able to successfully observe CP violation we need to isolate squarks from antisquarks as the CP-conjugated process will have an asymmetry of the opposite sign. Normally this is done via charge identification but this will be impossible using the above decay chain with the light quark in the final state. However, for the first two generations, squark production typically dominates over anti-squark production due to the valence quarks present in the proton. Consequently our sample will have a majority of squarks (around 80\% in the analysed scenario) and the asymmetry will not be washed out. The proportion depends on the masses of squarks but this does not affect our analysis significantly. One of the major problems when trying to measure CP-sensitive observables with triple product correlations at the LHC are the large undetermined boosts given to produced particles. CP asymmetries have the largest value in the rest frame of the intermediate particle and the effect of the boost produces a severe dilution factor. To remove this dilution factor we have to boost all the particles in the triple product ($q, \ell^+, \ell^-$) into the rest frame of the intermediate particle in the decay ($\tilde{\chi}^0_2$ in our case). In SUSY decay chains this is complicated due to the presence of two LSPs and possible neutrinos that escape detection. Once the missing momenta are reconstructed it is then trivial to find the momentum of the particle that we are interested in and boost back into its rest frame. In principle this should return the CP asymmetry to the magnitude one would expect if the initial particle was produced at rest. So far in the literature the momentum reconstruction was used to determine the unknown masses of SUSY particles in the case when the decay chain proceeded via two-body decays~\cite{Nojiri:2003tu,Nojiri:2007pq,Cheng:2008mg,Cheng:2009fw}. In this analysis we change the approach and assume that the masses are already known with a reasonable precision from other measurements at the LHC. The new complications are three-body decays of supersymmetric particles (charginos and neutralinos) and possible missing neutrinos which means that we have less on-shell constraints. However, as we show in this paper, even in such a case momentum reconstruction is still possible after the inclusion of experimental uncertainties. Having all the momenta reconstructed one can perform a more detailed analysis of particle properties which could open a window for a more accurate determination of the underlying model at the LHC. Finally, in order to assess the feasibility of completing this study in the demanding experimental environment of the LHC we perform a simulation of our process using the Monte Carlo event generator \texttt{Herwig++}~\cite{Bahr:2008pv,Bahr:2008tx}. We apply simple selection cuts and smear final state particle momenta to account for finite detector resolution. Moreover, we make an analysis of the impact of the most important background processes relevant for our study. All of the analysis was done assuming a centre-of-mass energy of 14~TeV but it would not change markedly if we moved to 10~TeV except for an increase in the luminosity we require to see a statistically significant result. The paper is organised as follows. We begin by describing the process under consideration in Section~\ref{sec:formalism} and discuss the triple product correlations in more detail. In Section~\ref{sec:MomRec} we present the method we use to reconstruct the momenta of the invisible particles in our decay chain. Section~\ref{sec:results} discusses the potential for a measurement at the LHC both with and without the momentum reconstruction technique. Possible experimental factors that need to be considered and our final results are explained in Section~\ref{sec:experimental}. Finally, the Appendices list the Lagrangian, resulting couplings and give explicit matrix elements including the full spin correlations that are required in our study. \section{Formalism} \label{sec:formalism} \subsection{The process studied and the amplitude squared } \label{sec:process} One of the dominant SUSY channels at the LHC is associated squark-gluino production. At the tree level the production process Eq.~(\ref{eq-prod}) proceeds via the light quark exchange in the direct channel and squark/gluino exchange in the $t$ channel, as seen in Fig.~\ref{Fig:FeynProdA}, \begin{equation} gq \to \tilde{g} \tilde{q}_L. \label{eq-prod} \end{equation} We study the case where the squark subsequently decays via the following chain: \begin{equation} \tilde{q}_L \to \tilde{\chi}^0_j + q \to \tilde{\chi}^0_1 \ell^+\ell^- + q. \label{eq-decays} \end{equation} \begin{figure}[t!] \begin{center} \hspace{-0.4cm} \begin{minipage}[t]{5cm} \begin{center} {\setlength{\unitlength}{0.8cm} \begin{picture}(5,6) \put(0,2){\epsfig{file=us.eps,scale=0.5}} \put(5.3,1.9){{\small $\tilde{g}$}} \put(-0.1,1.9){{\small $q$}} \put(5.3,5.0){{\small $\tilde{q}_L$}} \put(-0.1,5.0){{\small $g$}} \put(2.6,3.8){{\small $q$}} \end{picture}} \end{center} \end{minipage} \hspace{0.cm} \begin{minipage}[t]{5cm} \begin{center} {\setlength{\unitlength}{0.8cm} \begin{picture}(5,6) \put(0,2){\epsfig{file=ult.eps,scale=0.5}} \put(5.3,1.9){{\small $\tilde{g}$}} \put(-0.1,1.9){{\small $q$}} \put(5.3,5.0){{\small $\tilde{q}_L$}} \put(-0.1,5.0){{\small $g$}} \put(2.9,3.4){{\small $\tilde{q}_L$}} \end{picture}} \end{center} \end{minipage} \begin{minipage}[t]{5cm} \begin{center} {\setlength{\unitlength}{0.8cm} \begin{picture}(5,6) \put(0,2){\epsfig{file=got.eps,scale=0.5}} \put(5.3,1.9){{\small $\tilde{q}_L$}} \put(-0.1,1.9){{\small $q$}} \put(5.3,5.0){{\small $\tilde{g}$}} \put(-0.1,5.0){{\small $g$}} \put(2.9,3.4){{\small $\tilde{g}$}} \end{picture}} \end{center} \end{minipage} \end{center} \vspace{-1.7cm} \caption{\label{Fig:FeynProdA}Feynman diagrams for the associated production process of squarks and gluinos at the LHC, $gq \to \tilde{g} \tilde{q}_L$.} \end{figure} \noindent The first step in the cascade decay chain is the two-body process $\tilde{q}_L \to q \tilde{\chi}^0_2$. Here, CP-violating couplings of the $\tilde{\chi}^0_2$ enter and are dominated by the phase $\phi_{1}$, see Appendix~\ref{sect:squarkdecay}. In addition, the spin vector of the $\tilde{\chi}^0_2$ has to be explicitly included in the amplitude since the full spin correlations have to be taken into account. We consider spectra where the second step in the cascade decay chain is the three-body decay of the neutralino, $\tilde{\chi}^0_2 \to \tilde{\chi}^0_1 \ell^+ \ell^-$ (cf.\ Appendix~\ref{sect:neutdecay}). The neutralino decay occurs via $Z^0$ exchange in the direct channel and via slepton $\tilde{\ell}_{L,R}$ exchanges in the $t$ and $u$ channels, as can be seen in Fig.~\ref{Fig:FeynDecayA}. It is sensitive to CP-violating supersymmetric couplings and its structure has been studied in detail in \cite{Choi:2005gt,Bartl:2004jj,MoortgatPick:1999di}, cf.\ Appendix~\ref{sec:lagrangian-couplings}. The phase $\phi_{1}$ affects the masses of $\tilde \chi_1^0$ and $\tilde{\chi}_2^0$, as well as its couplings and decay rates. \begin{figure}[t!] \begin{center} \hspace{-0.5cm} \begin{minipage}[t]{5cm} \begin{center} {\setlength{\unitlength}{0.8cm} \begin{picture}(5,8) \put(-5.2,-5.2){\special{psfile=prozq.ps hscale=65 vscale=60}} \put(4.7,1.7){{\small $\ell^{-}$}} \put(0.3,5.5){{\small $\tilde{\chi}^0_{2}$}} \put(4.7,5.5){{\small $\ell^{+}$}} \put(3.7,7.1){{\small $\tilde{\chi}^0_{1}$}} \put(2.2,3.8){{\small $Z^0$}} \end{picture}} \end{center} \end{minipage} \hspace{-0.5cm} \begin{minipage}[t]{5cm} \begin{center} {\setlength{\unitlength}{0.8cm} \begin{picture}(5,8) \put(-5.2,-6){\special{psfile=protq.ps hscale=65 vscale=65}} \put(4.7,5.5){{\small $\ell^{-}$}} \put(3.7,7.1){{\small $\ell^{+}$}} \put(4.7,1.5){{\small $\tilde{\chi}^0_{1}$}} \put(0.3,5.6){{\small $\tilde{\chi}^0_{2}$}} \put(2.1,3.9){{\small $\tilde{\ell}_{L,R}$}} \end{picture}} \end{center} \end{minipage} \begin{minipage}[t]{5cm} \begin{center} {\setlength{\unitlength}{0.8cm} \begin{picture}(5,8) \put(-5.2,-6){\special{psfile=prouq.ps hscale=65 vscale=65}} \put(3.7,7.1){{\small $\ell^{-}$}} \put(4.7,5.5){{\small $\ell^{+}$}} \put(0.3,5.6){{\small $\tilde{\chi}^0_{2}$}} \put(4.7,1.5){{\small $\tilde{\chi}^0_{1}$}} \put(2.1,3.9){{\small $\tilde{\ell}_{L,R}$}} \end{picture}} \end{center} \end{minipage} \end{center} \vspace{-1.7cm} \caption{\label{Fig:FeynDecayA}Feynman diagrams for the three-body leptonic neutralino decay $\tilde{\chi}^0_2\to \tilde{\chi}^0_1\ell^{+}\ell^{-}$.} \end{figure} Using the narrow width approximation and the formalism of~\cite{Haber:1994pe}, the squared amplitude $|T|^2$ of the full process can be factorised into the processes of production $pp\to \tilde{g} \tilde{q}_L$ and the subsequent decays $\tilde{q}_L \to q \tilde{\chi}^0_2$, $\tilde{\chi}^0_2\to\tilde{\chi}^0_1\ell^+\ell^-$ taking into account the full spin information of the decaying $\tilde{\chi}^0_2$. For the purpose of analysing neutralino decays we do not need to include decays of the gluino produced with a squark but they will be needed later for momentum reconstruction, see Sec.~\ref{sec:MomRec}. We apply the narrow-width approximation for the masses of the intermediate particles, $\tilde{q}_L$ and $\tilde{\chi}^0_2$, which is appropriate since the widths of the respective particles are in all cases much smaller than their masses and the mass differences between them are large enough, see Appendix~\ref{sec:kinematics}. The squared amplitude can then be expressed in the form \begin{eqnarray} |T|^2 &=& 4 |\Delta(\tilde{q}_L)|^2 |\Delta(\tilde{\chi}^0_2)|^2 P(\tilde{g}\tilde{q}_L) \Big{\{} P(\tilde{\chi}^0_2) D(\tilde{\chi}^0_2) + \sum^3_{a=1}\Sigma^a_P(\tilde{\chi}^0_2) \Sigma^a_D(\tilde{\chi}^0_2)\Big\}, \label{Tsquared} \end{eqnarray} where $a=1,2,3$ refers to the polarisation states of the neutralino $\tilde{\chi}^0_2$, which are described by the polarisation vectors $s^a(\tilde{\chi}^0_2)$, given in Appendix~\ref{sect:squarkdecay}. In addition, \begin{itemize} \item $\Delta(\tilde{q}_L)$ and $\Delta(\tilde{\chi}^0_2)$ are the propagators of the intermediate particles which lead to the factors $E_{\tilde{q}_L}/m_{\tilde{q}_L}\Gamma_{\tilde{q}_L}$ and $E_{\tilde{\chi}^0_2}/m_{\tilde{\chi}^0_2}\Gamma_{\tilde{\chi}^0_2}$ in the narrow-width approximation, \item $P(\tilde{g}\tilde{q}_L)$, $P(\tilde{\chi}^0_2)$ and $D(\tilde{\chi}^0_{2})$ are the terms in the production and decay that are independent of the polarisations of the decaying neutralino, whereas \item $\Sigma^a_P(\tilde{\chi}^0_{2})$ and $\Sigma^a_D(\tilde{\chi}^0_{2})$ are the terms containing the correlations between production and decay spin of the $\tilde{\chi}^0_2$. \end{itemize} According to our choice of the polarisation vectors $s^a(\tilde{\chi}^0_2)$, see Eq.~(\ref{eq-s1chi}) in Appendix~\ref{sect:squarkdecay}, $\Sigma^3_P/P(\tilde{\chi}^0_2)$ is the longitudinal polarisation, $\Sigma^1_P/P(\tilde{\chi}^0_2)$ is the transverse polarisation in the production plane, and $\Sigma^2_P/P(\tilde{\chi}^0_2)$ is the polarisation perpendicular to the reference plane of the neutralino $\tilde{\chi}^0_2$. \subsection{Structure of the T-odd asymmetry} \label{sec:Todd_Struc} The tools used in this paper to study CP-violating effects are T-odd observables that are based on triple products of the momenta of the involved particles. Here we study the visible decay products of the $\tilde{q}_L$ in the form of the following triple product that can be evaluated in different reference frames: \begin{equation} \label{eq:tpt} \mathcal{T}= \vec{p}_q \cdot (\vec{p}_{\ell^+} \times \vec{p}_{\ell^-}). \end{equation} The T-odd asymmetry is then defined as, \begin{eqnarray} \label{Asy} \mathcal{A}_{\mathcal{T}} = \frac{N_{\mathcal{T}_+}-N_{\mathcal{T}_-}}{N_{\mathcal{T}_+}+N_{\mathcal{T}_-}} &=& \frac{\int\mathrm{sign}\{ \mathcal{T}\} |T|^2d\,\mbox{lips}}{{\int}|T|^2d\,\mbox{lips}}\nonumber \\ & =& \frac{\int \mathrm{sign} \{ \mathcal{T}\} \Big( \sum^3_{a=1}\Sigma^a_P(\tilde{\chi}^0_2) \Sigma^a_D(\tilde{\chi}^0_2)\Big)d\,\mbox{lips}}{\int \Big(P(\tilde{\chi}^0_2) D(\tilde{\chi}^0_2)\Big)d\,\mbox{lips}}\:, \end{eqnarray} where $N_{\mathcal{T}_+}$ ($N_{\mathcal{T}_-}$) are the numbers of events for which $\mathcal{T}$ is positive (negative) and $d\,\mbox{lips}$ denotes Lorentz invariant phase space. The denominator in Eq.~(\ref{Asy}), ${\int}|T|^2d\,\mbox{lips}$, is equal to the total cross section, namely $\sigma(p p \to \tilde{q}_L \tilde{g} \to q \tilde{\chi}^0_1 \ell^+ \ell^- \tilde{g})$. In the corresponding numerator of Eq.~(\ref{Asy}), the triple-product correlations only enter via the spin-dependent terms. If the spin of the particles is neglected in the calculation the asymmetry will vanish. The asymmetry $\mathcal{A}_{\mathcal{T}}$ can be visualised as the difference between the number of events where the observed $\vec{p}_q$ (from the decaying $\tilde{q}_L$) lies above ($N_{\mathcal{T}_+}$) or below ($N_{\mathcal{T}_-}$) the plane spanned by $ ( \vec{p}_{\ell^+},\vec{p}_{\ell^-} ) $, normalised by the total number of these events. If no complex phases are present, the $\vec{p}_q$ will on average line up with the plane and no asymmetry will be seen. The asymmetry is not a Lorentz invariant quantity but it is maximal in the rest frame of the $\tilde{\chi}^0_2$. At the LHC the momentum of the $\tilde{\chi}^0_2$ will in general be undetermined. The asymmetry can still be formed in the laboratory frame but it will be significantly diluted compared to the $\tilde{\chi}^0_2$ rest frame due to the initial particle boost that is parametrised by the parton density functions (PDFs). The dilution is due to the fact that the relative orientation of the plane spanned by $\vec{p}_{\ell^+}, \vec{p}_{\ell^-}$ and the quark momentum can change when boosting from the neutralino rest frame to the laboratory frame. Hence, the quark that was initially above the plane can instead be observed to come from below the plane in the laboratory frame. A possible solution to bypass the effects of the dilution would be to reconstruct the momentum of the $\tilde{\chi}^0_1$ and use it to find the momentum of the $\tilde{\chi}^0_2$. We will discuss this method in Sec.~\ref{sec:MomRec}. In order to identify the T-odd contributions, we have to identify those terms in $|T|^2$, Eq.~(\ref{Tsquared}), which contain a triple product of the form shown in Eq.~(\ref{eq:tpt}). Triple products follow from expressions $i\epsilon_{\mu\nu\rho\sigma}a^\mu b^\nu c^\rho d^\sigma$, where $a$, $b$, $c$, $d$ are 4-momenta or spins of the particles involved, which are non-zero only when the momenta are linearly independent. The expressions $i\epsilon_{\mu\nu\rho\sigma}a^\mu b^\nu c^\rho d^\sigma$ are imaginary and when multiplied by the imaginary parts of the respective couplings they yield terms that contribute to the numerator of $\mathcal{A}_{\mathcal{T}}$, Eq.~(\ref{Asy}). In our process, T-odd terms with $\epsilon$-tensors are only contained in the spin-dependent contributions in the neutralino decay, $\Sigma^a_D(\tilde{\chi}^0_j)$. Examining the covariant product, $\epsilon_{\mu\nu\rho\sigma}a^\mu b^\nu c^\rho d^\sigma$, we can understand why the maximal asymmetry is seen in the rest frame of the $\tilde{\chi}^0_2$. If we expand the product in terms of the explicit momenta, Eq.~(\ref{eq:EpsExpan}), we see that in general we have triple products formed from all different combinations of the momenta of the four different particles in the covariant product: \begin{eqnarray} \epsilon_{\mu\nu\rho\sigma}p_a^\mu p_b^\nu p_c^\rho p_d^\sigma = & E_a\;\overrightarrow{p_b}\cdot(\overrightarrow{p_c}\times\overrightarrow{p_d})+E_c\;\overrightarrow{p_d}\cdot(\overrightarrow{p_a}\times\overrightarrow{p_b}) \label{eq:EpsExpanLab} \\ & -E_b\;\overrightarrow{p_c}\cdot(\overrightarrow{p_d}\times\overrightarrow{p_a})-E_d\;\overrightarrow{p_a}\cdot(\overrightarrow{p_b}\times\overrightarrow{p_c}). \nonumber \end{eqnarray} As we evaluate only one triple product we miss the other combinations and the asymmetry is not maximal. However, if we are in the rest frame of the $\tilde{\chi}^0_2$, the momentum components of the four vector vanish and we are now only left with the single triple product that we are interested in: \begin{equation} \epsilon_{\mu\nu\rho\sigma}p_a^\mu p_b^\nu p_c^\rho p_d^\sigma \longrightarrow m_a\;\overrightarrow{p_b}\cdot(\overrightarrow{p_c}\times\overrightarrow{p_d}). \label{eq:EpsExpanRest} \end{equation} Let us finally comment on the origin of the asymmetry, Eq.~(\ref{Asy}). Neutralinos in the process we study are polarised due to the nature of the electroweak coupling between the fermion, sfermion and neutralino, see also \cite{AguilarSaavedra:2004hu}. Therefore the momentum of the outgoing fermion is aligned with the spin of the neutralino $\tilde{\chi}^0_2$. The actual asymmetry arises solely in the neutralino $\tilde{\chi}^0_2$ decay as a result of the correlation between the neutralino spin and the momenta of outgoing leptons, consequently it requires us to take into account all spin correlations in the decay chain. \section{Momentum reconstruction} \label{sec:MomRec} \subsection{Dilution effects} \label{sec:DillEff} \begin{figure}[t!] \begin{picture}(16,7.5) \put(6.4,7.2){$\mathcal{T}=\overrightarrow{p}_q\cdot(\overrightarrow{p}_{\ell^+}\times\overrightarrow{p}_{\ell^-})$} \put(10.9,-0.2){$|\vec{p}_{\tilde{q}}|$ [GeV]} \put(2.9,7.2){\epsfig{file=MomPlot.eps,scale=0.8,angle=270}} \end{picture} \caption{\label{fig:MomPlot} The parton level asymmetry $\mathcal{A}_{\mathcal{T}}$, Eq.~(\protect\ref{Asy}), for the single decay chain given in Eq.~(\protect\ref{eq-decays}) as a function of the squark momentum, $|\vec{p}_{\tilde{q}}|$, in the laboratory frame. The scenario is given in Table~\protect\ref{tab:spectrum} with the phase $\phi_1=3\pi/2$. } \vspace{0.cm} \end{figure} The triple product that is constructed from momenta in the laboratory frame suffers from dilution factors at the LHC due to the frame being boosted with respect to the rest frame of the $\tilde{\chi}^0_2$, for a more detailed discussion see \cite{Ellis:2008hq}. It results in a considerable reduction in the maximum asymmetry observable when we introduce the PDFs which causes an undetermined boost to the system. To illustrate the dilution, Fig.~\ref{fig:MomPlot} shows how the parton level asymmetry $\mathcal{A}_{\mathcal{T}}$, Eq.~(\protect\ref{Asy}), is diluted in the laboratory frame when we produce the $\tilde{q}$ with varying initial momenta. As the $\tilde{\chi}^0_2$ is produced in the squark decay any boost is transferred. The plot was produced with an analytical calculation for the single decay chain given in Eq.~(\protect\ref{eq-decays}). The scenario displayed in Tab.~\ref{tab:spectrum} was used but with the phase set to $\phi_1=3\pi/2$. We see that the asymmetry is maximal, $\mathcal{A}_{\mathcal{T}}\sim14\%$, when the $\tilde{q}$ is at rest but drops to, $\mathcal{A}_{\mathcal{T}}\sim2.5\%$, when $|\vec{p}_{\tilde{q}}|\sim1$ TeV. The magnitude of the dilution clearly depends on the chosen scenario and in particular on the masses of the particles involved in the process. If it were possible to reconstruct the momentum of the $\tilde{\chi}^0_2$, we could however perform a Lorentz transformation of all the momenta and bypass the dilution factor, potentially recovering the full asymmetry. \subsection{Reconstruction procedure} \label{sec:RecProd} In the mSUGRA scenario that we have chosen to study in this paper (see Section~\ref{sec:scenario}) the full reconstruction of the event is made possible by considering the decay chains of both the particles produced in the hard collision. We include all the particles coming from both the $\tilde{q}_L$ and the $\tilde{g}$ because there are not enough kinematic constraints to perform reconstruction if only the $\tilde{q}$ chain is considered. Therefore we exclusively consider the production of $\tilde{q}_{L}$ and $\tilde{g}$ and their subsequent decays, see Fig.~\ref{fig:FullDiagram}. The $\tilde{q}_L$ decay chain will be the same as considered in Eq.~(\ref{eq-decays}) and the $\tilde{g}$ will decay as follows: \begin{equation}\label{eq:gluinodecay} \tilde{g} \to \tilde{t} + t \to \tilde{\chi}^+_1 + b + t \to \tilde{\chi}^0_1 \ell^+ \nu_{\ell} + b + t. \label{gludec} \end{equation} \begin{SCfigure}[50] \begin{picture}(7,8.15) \put(1,0.3){\epsfig{file=FullProc.eps,scale=0.5}} \put(6.9,.3){$\nu_{\ell}$} \put(6.6,5.3){$\ell^-_a$} \put(6.9,6.5){$\ell^+_a$} \put(6.7,7.7){$\tilde{\chi}^0_{1a}$} \put(0.9,5.7){$p$} \put(0.9,4){$p$} \put(3.5,4.2){$\tilde{g}$} \put(3.1,5.8){$\tilde{q}_L$} \put(4.9,5){$q$} \put(4.1,6.5){$\tilde{\chi}^0_2$} \put(4.9,4.2){$t$} \put(4.5,3.3){$\tilde{t}_1$} \put(5.2,2.4){$\tilde{\chi}^+_1$} \put(6.1,3.1){$b$} \put(7,2.6){$\tilde{\chi}^0_{1b}$} \put(7.2,1.4){$\ell^+_b$} \end{picture} \hspace{0.8cm} \caption{\label{fig:FullDiagram} The process studied for momentum reconstruction.} \end{SCfigure} \noindent In many scenarios the production of $\tilde{q}_L$ along with $\tilde{g}$ is the dominant source of the first and second generation squarks at the LHC and in the considered scenarios the branching ratios for the decay chain are favourable (cf.\ Sec~\ref{sec:results}). Assuming that all the masses in the decay chains are known, Sec.~\ref{sec:scenario}, the kinematics can be fully reconstructed using the set of invariant conditions and the measured missing transverse momentum. For our procedure we follow the methods for solving the kinematic equations very closely to those presented in \cite{Cheng:2007xv,Kawagoe:2004rz}. The novelty here is the inclusion of three-body decays of sparticles and allowing for additional missing momentum due to neutrinos in the final state. The difference lies also in using the mass constraints. For our purpose we assume that masses of sparticles are known and aim at the reconstruction of the momenta, whereas the previous studies used the above conditions to reconstruct masses. Rather than fully reconstructing the kinematics of both decay chains, an alternative idea may be to estimate the momentum of the $\tilde{\chi}^0_2$ and then boost into this approximate frame. A formula that estimates the momentum was presented in \cite{:1999fr} and is shown below, \begin{equation} \vec{p}_{\tilde{\chi}^0_2}^{\,\,\mathrm{approx}} \equiv \left(1+\frac{m_{\tilde{\chi}^0_1}}{M_{\ell\ell}}\right)\vec{p}_{\ell\ell}. \end{equation} This approach does not work in our study however as the approximation only becomes valid when $\vec{p}_{\tilde{\chi}^0_1}\to0$ in the rest frame of the $\tilde{\chi}^0_2$. At this kinematical endpoint the two leptons are back to back which causes the plane spanned by $\vec{p}_{l^+}$ and $\vec{p}_{l^-}$ to become badly defined. The triple product is therefore small and inaccurate leading to an asymmetry that is close to zero. Therefore the approach is not valid in our case. \vspace{0.5cm} In our process the following invariant equations can be formed. \begin{itemize} \item $\tilde{q}$ decay chain: \begin{eqnarray} m^2_{\tilde{\chi}^0_1} & = & (P_{\tilde{\chi}^0_{1a}})^2 \label{eq:chimass},\\ m^2_{\tilde{\chi}^0_2} & = & (P_{\tilde{\chi}^0_{1a}}+P_{\ell^+_a}+P_{\ell^-_a})^2,\\ m^2_{\tilde{q}} & = & (P_{\tilde{\chi}^0_2}+P_{q})^2. \end{eqnarray} \item $\tilde{g}$ decay chain: \begin{eqnarray} m^2_{\tilde{\chi}^+_1} & = & (P_{\tilde{\chi}^0_{1b}}+P_{\nu_{\ell}}+P_{\ell^+_b})^2 \label{eq:gME}\,,\\ m^2_{\tilde{t}} & = & (P_{\tilde{\chi}^+_1}+P_b)^2, \\ m^2_{\tilde{g}} & = & (P_{\tilde{t}}+P_{t})^2, \end{eqnarray} where $P$ denote 4-momenta of respective particles, and where necessary we label particles coming from squark and gluino decays with subscripts $a$ and $b$, respectively. \item We also have the missing transverse momentum constraint: \begin{eqnarray} \overrightarrow{p}^T_{miss} & = & \overrightarrow{p}^T_{\tilde{\chi}^0_{1a}} + \overrightarrow{p}^T_{\tilde{\chi}^0_{1b}} + \overrightarrow{p}^T_{\nu_{\ell}}\,. \label{eq:MET} \end{eqnarray} \item We combine the momenta of $\tilde{\chi}^0_{1b}$ and $\nu_{\ell}$ coming from the gluino together as it is impossible to resolve these two particles: \begin{equation} P_{\tilde{g}ME} = P_{\tilde{\chi}^0_{1b}}+P_{\nu_{\ell}}\,. \end{equation} \item An additional condition on the solutions is that the invariant $P^2_{\tilde{g}ME}$ has to be greater than the mass of the LSP, $m^2_{\tilde{\chi}^0_1}$: \begin{equation} P^2_{\tilde{g}ME}>m^2_{\tilde{\chi}^0_1}\,. \label{eq:pgmeinv} \end{equation} We apply this condition to each solution and discard as unphysical any that does not meet it. \end{itemize} After expressing the momenta of intermediate particles in terms of the final state particles \begin{eqnarray} P_{\tilde{\chi}^0_2} & = & P_{\tilde{\chi}^0_{1a}} + P_{l^+_a} + P_{l^-_a}\,,\\ P_{\tilde{\chi}^+_1} & = & P_{\tilde{g}ME} + P_{\ell^+_b} \,,\\ P_{\tilde{t}} & = & P_b + P_{\tilde{\chi}^0_{1b}}+P_{\ell^+_b}+P_{\nu_{\ell}} \,, \end{eqnarray} we now have a total of eight equations, Eq.~(\ref{eq:chimass}) - (\ref{eq:MET}), and eight unknowns: \begin{eqnarray} & \bigg(E_{\tilde{\chi}^0_{1a}},p_{\tilde{\chi}^0_{1a}}(x),p_{\tilde{\chi}^0_{1a}}(y),p_{\tilde{\chi}^0_{1a}}(z)\bigg)\,,\\ & \bigg(E_{\tilde{g}ME},p_{\tilde{g}ME}(x),p_{\tilde{g}ME}(y),p_{\tilde{g}ME}(z)\bigg)\:. \end{eqnarray} In principle the system can be solved to find $P_{\tilde{\chi}^0_1}$ and $P_{\tilde{g}ME}$. Equations Eq.~(\ref{eq:chimass}) and Eq.~(\ref{eq:gME}) are quadratic in $P_{\tilde{\chi}^0_1}$ and $P_{\tilde{g}ME}$ respectively, so we consider these last. Using on-shell conditions, quadratic terms in the remaining equations can be removed giving 6 linear equations, therefore we can simply use a matrix to give us solutions in terms of the energies $E_{\tilde{\chi}^0_1}$ and $E_{\tilde{g}ME}$. We first expand $\overrightarrow{p}_{\tilde{\chi}^0_1}$ and $\overrightarrow{p}_{\tilde{g}ME}$ in terms of other momenta contained in the respective decay chains: \begin{eqnarray} \overrightarrow{p}_{\tilde{\chi}^0_{1a}} & = & A\overrightarrow{p}_{\ell^+_a} + B\overrightarrow{p}_{\ell^-_a} + C\overrightarrow{p}_{q}\:, \label{eq:ABC} \\ \overrightarrow{p}_{\tilde{g}ME} & = & D\overrightarrow{p}_{\ell^+_b} + E\overrightarrow{p}_{b} + F\overrightarrow{p}_{t}\:.\label{eq:DEF} \end{eqnarray} We can now form the system of 6 linear equations for unknowns $A$-$F$ with $E_{\tilde{\chi}^0_{1a}}$ and $E_{\tilde{g}ME}$ as free parameters \begin{equation} \label{eq:MomMatEq} \mathcal{M}\left( \begin{array}{c} A \\ B \\ C \\ D \\ E \\ F \end{array} \right)= \left( \begin{array}{c} \frac{1}{2}(m^2_{\tilde{\chi}^0_1}-m^2_{\tilde{\chi}^0_2}) + P_{\ell^+_a}\cdot P_{\ell^-_a} + E_{\tilde{\chi}^0_{1a}}(E_{\ell^+_a}+E_{\ell^-_a}) \\ \frac{1}{2}(m^2_{\tilde{\chi}^0_2}-m^2_{\tilde{q}}) + P_{\ell^+_a}\cdot P_{q} + P_{\ell^-_a}\cdot P_{q} + E_{\tilde{\chi}^0_{1a}}E_{q}\\ \frac{1}{2}(m^2_{b}+m^2_{\tilde{\chi}^+_1}-m^2_{\tilde{t}}) + P_{\ell^+_b}\cdot P_{b} +E_{\tilde{g}ME}E_b \\ \frac{1}{2}(m^2_{\tilde{t}}+m^2_{t}-m^2_{\tilde{g}}) + P_{\ell^+_b}\cdot P_{t} + P_{b}\cdot P_{t} +E_{\tilde{g}ME}E_t \\ p^T_{miss}(x) \\ p^T_{miss}(y) \\ \end{array} \right)\:, \end{equation} where \begin{eqnarray} &&\hspace*{-1.5cm} \mathcal{M} = \left( \begin{array}{cccccc} (\overrightarrow{p}_{\ell^+_a}+\overrightarrow{p}_{\ell^-_a}) \overrightarrow{p}_{\ell^+_a} & (\overrightarrow{p}_{\ell^+_a}+\overrightarrow{p}_{\ell^-_a}) \overrightarrow{p}_{\ell^-_a} & (\overrightarrow{p}_{\ell^+_a}+\overrightarrow{p}_{\ell^-_a}) \overrightarrow{p}_{q} & 0 & 0 & 0 \\ \overrightarrow{p}_{\ell^+_a} \overrightarrow{p}_{q} & \overrightarrow{p}_{\ell^-_a} \overrightarrow{p}_{q} & \overrightarrow{p}_{q} \overrightarrow{p}_{q} & 0 & 0 & 0 \\ 0 & 0 & 0 & \overrightarrow{p}_b \overrightarrow{p}_{\ell^+_b} & \overrightarrow{p}_b \overrightarrow{p}_{b} & \overrightarrow{p}_b \overrightarrow{p}_t \\ 0 & 0 & 0 & \overrightarrow{p}_t \overrightarrow{p}_{\ell^+_b} & \overrightarrow{p}_b \overrightarrow{p}_{t} & \overrightarrow{p}_t \overrightarrow{p}_t \\ p_{\ell^+_a}(x) & p_{\ell^-_a}(x) & p_q(x) & p_{\ell^+_b}(x) & p_b(x) & p_t(x) \\ p_{\ell^+_a}(y) & p_{\ell^-_a}(y) & p_q(y) & p_{\ell^+_b}(y) & p_b(y) & p_t(y) \\ \end{array} \right)\nonumber \\ && \end{eqnarray} The matrix $\mathcal{M}$ can then be diagonalised to give solutions for each of the momentum components of $\overrightarrow{p}_{\tilde{\chi}^0_1}$ and $\overrightarrow{p}_{\tilde{g}ME}$ in terms of $E_{\tilde{\chi}^0_{1a}}$ and $E_{\tilde{g}ME}$. These solutions are substituted into the two quadratic equations, Eq.~(\ref{eq:chimass}) and Eq.~(\ref{eq:gME}), to produce two equations of the form: \begin{eqnarray} a_{11}E_{\tilde{\chi}^0_{1a}}^2+a_{12}E_{\tilde{\chi}^0_{1a}}E_{\tilde{g}ME} +a_{22}E_{\tilde{g}ME}^2 +a_1E_{\tilde{\chi}^0_{1a}} +a_2E_{\tilde{g}ME} +a \equiv F_A & = & 0\,, \label{eq:aEs}\\ b_{11}E_{\tilde{\chi}^0_{1a}}^2+b_{12}E_{\tilde{\chi}^0_{1a}}E_{\tilde{g}ME} +b_{22}E_{\tilde{g}ME}^2 +b_1E_{\tilde{\chi}^0_{1a}} +b_2E_{\tilde{g}ME} +b \equiv F_B & = & 0\,, \label{eq:bEs} \end{eqnarray} where the coefficients $a_{ij}$, $a_i$, $a$ and $b_{ij}$, $b_i$, $b$ are functions only of masses and measured momenta. We use, \begin{equation} F_A-\frac{a_{11}}{b_{11}}\times F_B =0 \end{equation} to produce the linear equation for $E_{\tilde{\chi}^0_{1a}}$, \begin{equation} E_{\tilde{\chi}^0_{1a}}=\frac{a_{11}b-a\,b_{11}-a_2b_{11}E_{\tilde{g}ME}+a_{11}b_2E_{\tilde{g}ME}- a_{22}b_{11}E^2_{\tilde{g}ME}+a_{11}b_{22}E^2_{\tilde{g}ME}}{-a_{11}b_1+a_1b_{11}+a_{12}b_{11}E_{\tilde{g}ME} -a_{11}b_{12}E_{\tilde{g}ME}}\,. \label{eq:EChi} \end{equation} This result can then be substituted into Eq.~(\ref{eq:aEs}) to obtain a quartic equation of the following form: \begin{equation} Q_4E_{\tilde{g}ME}^4+Q_3E_{\tilde{g}ME}^3+Q_2E_{\tilde{g}ME}^2+Q_1E_{\tilde{g}ME}+Q_0=0\,,\label{eq:quartic} \end{equation} where the various $Q$'s are functions of the $a$'s and $b$'s in Eqs.~(\ref{eq:aEs}) and (\ref{eq:bEs}). \subsection{Discussion of graphical solutions} Analysing the roots of the quartic equations, Eq.~(\ref{eq:quartic}), we select the solutions that are real and discard the solutions that contain imaginary parts. The selected roots are substituted into Eq.~(\ref{eq:EChi}) to find the corresponding solutions for $E_{\tilde{\chi}^0_{1a}}$. Using the values of $E_{\tilde{g}ME}$ and $E_{\tilde{\chi}^0_{1a}}$ together with the diagonalised matrix we can now calculate $A, B, C, D, E, F$, see Eqs.~(\ref{eq:ABC}),(\ref{eq:DEF}) and hence the components of $\overrightarrow{p}_{\tilde{g}ME}$ and $\overrightarrow{p}_{\tilde{\chi}^0_{1a}}$. In general, taking into account multiple roots, Eq.~(\ref{eq:quartic}) has 4 solutions. Thus we would have 4 real, 2 real and 1 complex pair or 2 complex pairs of roots. Only the real roots can yield physical solutions, therefore for each event we normally expect real solutions. As Eq.~(\ref{eq:quartic}) is derived from Eqs.~(\ref{eq:aEs}) and (\ref{eq:bEs}) they share the same set of solutions. Both Equations~(\ref{eq:aEs}) and (\ref{eq:bEs}) are polynomials of degree 2 in $E_{\tilde{\chi}^0_{1a}}$ and $E_{\tilde{g}ME}$ so they correspond to degree 2 curves in the $E_{\tilde{\chi}^0_{1a}}$ and $E_{\tilde{g}ME}$ plane (ellipses, hyperbolas or parabolas). The intersection points solve simultaneously both equations and at the same time the quartic Equation~(\ref{eq:quartic}). A graphical solution to this set of equations is shown in Fig.~\ref{fig:contours} for one sample event. In this example we have two real solutions, of which only one corresponds to the actual momenta in the event. \begin{figure}[t] \begin{picture}(16,8) \put(4.1,0.4){\epsfig{width=0.5\textwidth,file=contours_25-25_bw.eps}} \put(9.8,-0.){$E_{\tilde{g}ME}$ [GeV]} \put(2.2,7){$E_{\tilde{\chi}^0_{1a}}$ [GeV]} \end{picture} \caption{Ellipses in the $E_{\tilde{\chi}^0_{1a}}$ and $E_{\tilde{g}ME}$ plane corresponding to Equations~(\protect\ref{eq:aEs}) (solid line) and (\protect\ref{eq:bEs}) (dashed line) for one sample event. Out of the two real solutions the lower right one gives the correct momenta of the original event. \label{fig:contours}} \end{figure} In the realistic physical case one has to include uncertainties on measured momenta and masses. This will of course affect above equations and solutions. Smearing of momenta will typically result in shifting of the curves shown in Fig.~\ref{fig:contours}. Therefore it is possible that one can get two additional solutions or no real solutions at all. The consequences of experimental effects on our analysis will be discussed in Sec.~\ref{sec:expfactors}. \subsection{Practical approach} \label{sec:PracApproach} In general more than one solution to the above system will remain depending upon the resolvent. In some cases Eq.~(\ref{eq:pgmeinv}) reduces the number of solutions further. For the application of this study we therefore calculate the triple product, Eq.~(\ref{eq:tpt}), for each individual solution. If all solutions produce the same sign triple product we keep the event and use it in our asymmetry. If any of the solutions disagrees on the sign of the triple product we will discard the whole event and it will not contribute. This method has the disadvantage that we lose events and therefore statistical significance. However, we note that it does not introduce any bias, i.e.\ it can be checked that the asymmetry for accepted events remains the same. An alternative method we propose is to weight each solution with matrix element and phase space factors to select the solutions with the highest probability of being correct. A possible advantage of this method is that no events will be lost and this idea is followed up in \cite{newpaper}. An additional issue when trying to complete the momentum reconstruction procedure are combinatorial problems when trying to assign the measured momenta to the correct particle in the given event. If we take the leptons as an example, we know that opposite sign, same flavour leptons must come from the $\tilde{\chi}^0_2$. However, it is possible that a same flavour lepton could be produced from the $\tilde{\chi}^+_1$ in the opposite decay chain and be confused with those coming from the $\tilde{\chi}^0_2$. In order to assign the leptons correctly to the decay chains one can use the conditions for invariant masses. In the squark decay chain a useful observable is the invariant mass of two leptons and in the gluino decay chain one can use the invariant mass of the $b$-jet from the stop decay and the lepton from chargino decay. The end points of these distributions are given by \begin{eqnarray} M_{\ell^+_a \ell^-_a} &<& m_{\tilde{\chi}^0_2} - m_{\tilde{\chi}^0_1}\label{eq:lepinv}\,,\\ M_{b \ell_b^+} &<& \sqrt{\frac{ (m^2_{\tilde{t}_1} - m^2_{\tilde{\chi}^+_1} - m_b^2) (m^2_{\tilde{\chi}^0_2} - m^2_{\tilde{\chi}^0_1})}{m_{\tilde{\chi}^+_1} m_{\tilde{\chi}^0_2} } } \label{eq:blinv}\:. \end{eqnarray} It turns out that only a small fraction of the events, around 5\% in our scenario, satisfies both conditions simultaneously for the two possible assignments of the leptons in question. Moreover, if we run the momentum reconstruction algorithm, we find that physically acceptable solutions are only found in roughly 10\% of events where a wrong assignment has occurred. These conditions act as a strong discriminant, therefore we do not expect lepton combinatorics to be a relevant issue for this study. Another possible approach to this problem would be the subtraction of the opposite-sign opposite-flavour (OSOF) lepton pairs. In that case the quantities $N_{\mathcal{T}_+}$ and $N_{\mathcal{T}_-}$ in Eq.~(\ref{Asy}) would be defined as \begin{eqnarray} N_{\mathcal{T}_+} &=& [N_{e^+e^-} + N_{\mu^+\mu^-} - N_{e^+\mu^-} - N_{e^-\mu^+}]_{\mathcal{T}_+}\,, \label{eq:NTplus}\\ N_{\mathcal{T}_-} &=& [N_{e^+e^-} + N_{\mu^+\mu^-} - N_{e^+\mu^-} - N_{e^-\mu^+}]_{\mathcal{T}_-} \,. \label{eq:NTminus} \end{eqnarray} In Eqs.(\ref{eq:NTplus}) and (\ref{eq:NTminus}) one adds the number of events with a positive (negative) triple product using $e^+ e^-$ pairs (and the jet) to the number of events with a positive (negative) triple product using $\mu^+ \mu^-$ pairs and subtract the number of events with positive triple product using $e^+\mu^-$ and $e^-\mu^+$ pairs. For the combinatorial background one gets equal rates for same flavour and opposite flavour lepton pairs since the leptons are uncorrelated. Hence, this ``flavour subtraction'' procedure removes (up to statistical fluctuations) all combinatorial background resulting from the lepton pairing problem and also all background coming from $\tilde{\chi}_2^0 \to \tilde{\chi}_1^0 \tau^+ \tau^-$. \section{Numerical results} \label{sec:results} \subsection{Chosen scenario: spectrum and decay modes} \label{sec:scenario} In order to study the experimental prospects of measuring CP-violating effects at the LHC we use an MSSM scenario derived from mSUGRA parameters defined at the GUT scale, as shown in Tab.~\ref{tab:spectrum}. This scenario has been already used to analyse the properties of the neutralino sector in \cite{Choi:2005gt}. The values of the parameters at the electroweak scale have been derived using the RGE code \texttt{SPheno}~\cite{Porod:2003um}. Masses of the coloured particles are of order 500~GeV, apart from the light stop $\tilde{t}_1$, which has a mass of 171.0~GeV. The lightest supersymmetric particle (LSP) is the lightest neutralino with a mass of 78.1~GeV. The second neutralino and the light chargino have masses around 150~GeV, whereas the sleptons are around 200~GeV. Therefore we note that the two-body decay channels of $\tilde{\chi}^0_2$ and $\tilde{\chi}^\pm_1$ are closed and this gives a good opportunity to study CP-violation effects in their three-body decay modes. Details of the mass spectrum can be found in Tab.~\ref{tab:spectrum}. The values of the gaugino mass parameters reproduce the given spectrum in the case when all the CP phases are set to 0. In order to generate CP-violating effects in the following we will attribute a non-zero phase $\phi_1$ to the bino mass parameter, \begin{eqnarray} M_1 = |M_1| e^{i \phi_1}\, , \qquad 0 \leq \phi_1 < 2\pi \,, \end{eqnarray} while keeping the absolute value $|M_1|$ fixed as given in Tab.~\ref{tab:spectrum}. \begin{table}[t!]\renewcommand{\arraystretch}{1.3} \begin{center} \begin{tabular}{|c|c||c|c||c|c|} \hline Parameter & Value & Particle & Mass [GeV] & Particle & Mass [GeV] \\ \hline\hline $m_0$ & 150 GeV & $\tilde{g}$ & 496.5 & $\tilde{\chi}_1^0$ & 78.1 \\ $m_{1/2}$ & 200 GeV & $\tilde{d}_L$ & 484.1 & $\tilde{\chi}_2^0$ & 148.4 \\ $A_0$ & -650 GeV & $\tilde{d}_R$ & 466.4 & $\tilde{\chi}_1^\pm$ & 148.2 \\ $\tan\beta$ & 10 & $\tilde{u}_L$ & 477.9 & $\tilde{\chi}_2^\pm$ & 436.0 \\ \cline{5-6} $\mathrm{sign}\: \mu$ & $+$ & $\tilde{u}_R$ & 465.9 & $\tilde{e}_L$ & 207.5 \\ \cline{1-2} $|M_1|$ & 80.5 GeV & $\tilde{b}_1$ & 397.2 & $\tilde{e}_R$ & 173.1 \\ $M_2$ & 153.3 GeV& $\tilde{b}_2$ & 462.6 & $\tilde{\nu}_e$ & 192.0 \\ \cline{5-6} $M_3$ & 484.6 GeV& $\tilde{t}_1$ & 171.0 & $\tilde{\tau}_1$ & 149.4 \\ $\mu$ & 419.0 GeV& $\tilde{t}_2$ & 498.0 & $\tilde{\tau}_2$ & 212.5 \\ & & & & $\tilde{\nu}_\tau$ & 187.2 \\ \hline \end{tabular} \caption{mSUGRA input parameters at the GUT scale, MSSM parameters and particle masses in GeV from \texttt{SPheno 2.2.3}~\cite{Porod:2003um} for $\phi_1 = 0$ and with $m_t = 171$~GeV. \label{tab:spectrum}} \end{center} \end{table} Although we chose a specific scenario our method is applicable in a wider range of parameter points. For the decay chain that we concentrate on, we require the three-body decay of the $\tilde{\chi}^0_2$ which places the following constraints on the SUSY masses, \begin{eqnarray} m_{\tilde{\chi}^0_2}<m_{\tilde{\ell}}\label{eq:lepcons}\,,\\ m_{\tilde{\chi}^0_2}-m_{\tilde{\chi}^0_1}<m_Z.\label{eq:Zcons} \end{eqnarray} Equation~(\ref{eq:lepcons}) ensures that the decay, $\tilde{\chi}^0_2\to \tilde{\ell}\ell$ cannot occur whereas Eq.~(\ref{eq:Zcons}) forbids the decay $\tilde{\chi}^0_2\to Z\tilde{\chi}^0_1$. To complete momentum reconstruction from the gluino side, other decay modes can contribute with regard to the scenario presented here but their impact on the final result would be scenario dependent. For example if the decay $\tilde{t}\to t\tilde{\chi}^0_2$ became kinematically open the momentum reconstruction would actually become easier as there would be no neutrino in the final state and the system would become over-constrained. However, there would now be one additional top and one additional lepton in the final state that may cause new combinatorical difficulties. As already mentioned, the highest production rate is typically obtained for coloured final states containing squarks and gluinos at the LHC. In our scenario, where their masses are not very heavy, the total cross section for production of strongly interacting supersymmetric particles reaches 140~pb, see Tab.~\ref{tab:CrossSec}. For our purpose we will be interested in the inclusive production of left squarks and the associated production of left squark and gluino\footnote{Triple product asymmetry in right squarks decay chain has the opposite sign as compared to the left squarks (and the same sign as that of left anti-squarks), however in our case where the coupling of the right squark to the second lightest neutralino is suppressed due to small bino component of $\tilde{\chi}^0_2$, this contribution remains negligible.}. As was mentioned in Sec.~\ref{sec:introduction}, anti-squarks give a CP asymmetry with exactly the opposite sign to squarks. We note however that the inclusive cross section for left squark production is almost a factor of 4 larger than the cross section for left anti-squark production. This is a direct consequence of the fact that we have two protons in the initial state for which the abundance of quarks is significantly higher than of anti-quarks. Since we cannot distinguish experimentally squarks and anti-squarks of the first two generations this fraction of anti-squarks will cause some dilution in the asymmetry that will be however compensated for by a very high production rate. A similar situation occurs for associated squark gluino production for which the ratio of squarks to anti-squarks is 18.2~pb to 3.1~pb, see Tab.~\ref{tab:CrossSec}. Together with left squarks we also have the production of right squarks at a comparable rate. However, since the latter decay almost exclusively to the lightest neutralino in our scenario, as shown in Tab.~\ref{tab:brs}, they do not give rise to the CP-odd asymmetry. \begin{table}[t!] \renewcommand{\arraystretch}{1.3} \begin{center} \begin{tabular}{|c|c|c|} \hline\hline Produced Particles & \multicolumn{2}{c|}{Cross Section (pb)} \\ \hline \hline At least one coloured SUSY particle. & \multicolumn{2}{c|}{148} \\ \hline At least one $\tilde{g}$. & \multicolumn{2}{c|}{58.8} \\ \hline $\tilde{q}_L\tilde{q}_L+\tilde{q}_L\tilde{q}_L^*+\tilde{q}_L\tilde{q}_R+\tilde{q}_L\tilde{q}_R^*+\tilde{q}_L\tilde{g}$ & \multicolumn{2}{c|}{30.0} \\ \hline $\tilde{q}^*_L\tilde{q}_L+\tilde{q}^*_L\tilde{q}_L^*+\tilde{q}^*_L\tilde{q}_R+\tilde{q}^*_L\tilde{q}_R^*+\tilde{q}^*_L\tilde{g}$ & \multicolumn{2}{c|}{8.3} \\ \hline $\tilde{q}_L \tilde{g}$ & \multicolumn{2}{c|}{18.2} \\ \hline $\tilde{q}^*_L \tilde{g}$ & \multicolumn{2}{c|}{3.1} \\ \hline $\tilde{t}_1 \tilde{t}^*_1$ & \multicolumn{2}{c|}{66.3} \\ \hline \end{tabular}\\[0.5ex] \caption{Leading order cross sections at $\sqrt{s} = 14\mathrm{TeV}$ for direct production of various particles from \texttt{Herwig++ 2.3.2}~\cite{Bahr:2008tx,Bahr:2008pv} using \texttt{MRST 2004LO} PDF set~\cite{Martin:2007bv}. $\tilde{q}$ stands for squarks of the first and second generation. \label{tab:CrossSec}} \end{center} \end{table} Following the production process we include subsequent decays of squarks and gluinos. The dominant decay mode of the gluino is to the light stop and the top with a branching ratio of $BR=53.8\%$. The light stop then decays almost exclusively to $\tilde{\chi}^\pm_1 b$. Left squarks decay mainly to the light chargino and the second lightest neutralino with branching fractions of $65\%$ and $33\%$ respectively. Finally we consider the decays of chargino $\tilde{\chi}^\pm_1$ and neutralino $\tilde{\chi}^0_2$. Leptonic decays constitute in total $BR=61\%$ for the chargino and $BR=68\%$ for the neutralino decay modes. For decays into light leptons we have $BR=24.3\%$ and $BR=9\%$ for chargino and neutralino, respectively. The most interesting are the neutralino decays to electrons and muons that are used to construct the CP-sensitive triple product, Eq.~(\ref{eq:tpt}). The summary of the relevant branching ratios can be found in Tab.~\ref{tab:brs}. It may be noted that in the scenario presented, the branching ratio $\tilde{\chi}^0_2\to\tilde{\chi}^0_1\tau^+\tau^-$ is unusually high (59.3\%). This is due to the masses of the $\tilde{\chi}^0_2$ and $\tilde{\tau}$ being very close and hence the kinematic factors are favourable. The particular branching ratio is simply a coincidence in the scenario chosen however and is not a required feature for our study. \begin{table}\renewcommand{\arraystretch}{1.3} \begin{center} \begin{tabular}{|c|c||c|c||c|c|} \hline Decay & BR & Decay & BR & Decay & BR \\ \hline $\tilde{g} \to \tilde{t}_1 \bar{t} + \mathrm{c.c.}$ & 53.8\% & $\tilde{q}_L \to \tilde{\chi}_1^\pm q $ & 65\% & $\tilde{t}_1 \to \tilde{\chi}_1^+ b $ & 98.1\% \\ $\tilde{g} \to \tilde{b}_1 \bar{b} + \mathrm{c.c.}$ & 26.6\% & $\tilde{q}_L \to \tilde{\chi}_2^0 q $ & 33\% & $\tilde{t}_1 \to \tilde{\chi}_2^0 c $ & 1.6\% \\ \cline{5-6} $\tilde{g} \to \tilde{q}_R \bar{q} + \mathrm{c.c.}$ & 11.8\% & $\tilde{q}_L \to \tilde{\chi}_1^0 q $ & 1.5\% & $\tilde{\chi}_2^0 \to \tilde{\chi}_1^0 \tau^- \tau^+$ & 59.3\% \\ \cline{3-4} $\tilde{g} \to \tilde{b}_2 \bar{b} + \mathrm{c.c.}$ & 3.8\% & $\tilde{\chi}_1^+ \to \tilde{\chi}_1^0 \nu_\tau \tau^+$ & 37.2\% & $\tilde{\chi}_2^0 \to \tilde{\chi}_1^0 \nu \bar{\nu}$ & 23.6\% \\ $\tilde{g} \to \tilde{q}_L \bar{q} + \mathrm{c.c.}$ & 3.3\% & $\tilde{\chi}_1^+ \to \tilde{\chi}_1^0 q_u \bar{q}_d$ & 38.5\% & $\tilde{\chi}_2^0 \to \tilde{\chi}_1^0 q \bar{q}$ & 8.1\% \\ \cline{1-2} $\tilde{q}_R \to \tilde{\chi}^0_1 q $ & 98\% & $\tilde{\chi}_1^+ \to \tilde{\chi}_1^0 \nu_\mu \mu^+$ & 12.2\% & $\tilde{\chi}_2^0 \to \tilde{\chi}_1^0 e^+ e^- $ & 4.5\% \\ $\tilde{q}_R \to \tilde{\chi}^0_2 q $ & 1\% & $\tilde{\chi}_1^+ \to \tilde{\chi}_1^0 \nu_e e^+$ & 12.1\% & $\tilde{\chi}_2^0 \to \tilde{\chi}_1^0 \mu^+ \mu^- $ & 4.5\% \\ \hline \end{tabular} \caption{Branching ratios for the scenario defined in Tab.~\protect\ref{tab:spectrum} from \texttt{SPheno 2.2.3}~\protect\cite{Porod:2003um} for $\phi_1=0$. \label{tab:brs}} \end{center} \end{table} When we vary the phase of the $M_1$ parameter, the masses and couplings in the neutralino sector are affected. First we note that the changes in the neutralino masses are negligible and smaller than the possible experimental accuracy. It turns out however, that the phase $\phi_1$ has large impact on neutralino couplings and therefore the pattern of its decay modes. The most significant change is for the light chargino decays to the LSP and a fermion pair. With increasing phase the branching ratio for $\tilde{\chi}^+_1 \to \tilde{\chi}^0_1 \nu_\tau \tau^+$ rises and eventually reaches 70\% for $\phi_1 = \pi$. At the same time the branching ratios for decays to light leptons remain roughly at the level of 10\%. As we will explain later, a chargino decay to tau, followed by a leptonic tau decay can be used for momentum reconstruction in the same way as a chargino decay to an electron and a muon. On the other hand, the decay $\tilde{\chi}_2^0 \to \tilde{\chi}_1^0 \tau^+ \tau^-$ followed by a leptonic tau decay will be a background, since it gives incorrectly reconstructed momenta. Finally, we note that in the case of neutralino $\tilde{\chi}^0_2$ decays to light leptons, the respective branching ratios increase up to 5.5\% for $\phi_1 = \pi/2$. Since our procedure requires the sparticle masses to be known, we wish to comment on the possibility of the mass determination in our scenario. One of the standard approaches at the LHC is to study kinematic edges and endpoints of the invariant mass distributions, see e.g. \cite{Weiglein:2004hn}. In our case the masses of $\tilde{g}$, $\tilde{q}_L $, $\tilde{t}_1$, $\tilde{\chi}^0_{1,2}$ and $\tilde{\chi}^\pm_{1}$ are required. The possible invariant mass observables would include: \begin{itemize} \item quark and leptons in the decay chain of $\tilde{q}_L$ followed by the squark decay to chargino or neutralino $\tilde{\chi}^0_2$, \item lepton pair in the neutralino $\tilde{\chi}^0_2$ decay, cf.\ Eq.~(\ref{eq:lepinv}), \item top and bottom quarks in the gluino decay chain Eq.~(\ref{eq:gluinodecay}), \item bottom quark and lepton in the stop decay chain, cf.\ Eq.~(\ref{eq:blinv}), \item quark pairs in the decay of gluino to the 1st and 2nd generation squarks~\cite{Cho:2007qv}. \end{itemize} Fitting the above invariant masses to experimentally measured edges and endpoints should provide enough number of constraints to obtain required information. This method can be perhaps supplemented by a mass reconstruction method, see for example \cite{Nojiri:2003tu,Nojiri:2007pq,Cheng:2008mg,Cheng:2009fw,Kawagoe:2004rz,Kersting:2009ne,Kang:2009sk}. Therefore, in the following we assume that the masses are known. The impact of the mass uncertainties will be discussed in Section~\ref{sec:expfactors}. \subsection{CP asymmetry at the parton level}\label{sec:asy-parton} We start by discussing the dependence of the asymmetry on $\phi_1$, Eq.~(\ref{Asy}), at the partonic level studying only the decay chain presented in Eq.~(\ref{eq-decays}). In order to see a maximal effect, we place the $\tilde{q}_L$ at rest and calculate the triple product and the asymmetry in its distribution, Sec.~\ref{sec:Todd_Struc}. We see from Fig.~\ref{fig:PartonAsy} that the asymmetry in this scenario is roughly 14\% and this occurs when the phase $\phi_{1}$ is just above $\pi/2$ and just below $3\pi/2$. The asymmetry is produced by a complex interplay between different couplings in the $\tilde{\chi}^0_2$ decay, cf. Eqs.~(\ref{eq_dszz_to})-(\ref{eq_dselel_to}), and can vary significantly between scenarios. These couplings can all have different behaviour with respect to $\phi_1$ and in other scenarios the maximum asymmetry can be seen far from $\pi/2$. \begin{figure}[t!] \begin{picture}(16,7.5) \put(6.4,7.2){$\mathcal{T}=\overrightarrow{p}_q\cdot(\overrightarrow{p}_{\ell^+}\times\overrightarrow{p}_{\ell^-})$} \put(10.9,0.){$\phi_{1}/\pi$} \put(2.9,7.2){\epsfig{file=KrysRightPart.eps,scale=0.8,angle=270}} \end{picture} \caption{\label{fig:PartonAsy} The parton level asymmetry $\mathcal{A}_{\mathcal{T}}$, Eq.~(\protect\ref{Asy}), for the single decay chain given in Eq.~(\protect\ref{eq-decays}) in the rest frame of $\tilde{q}_L$ as a function of $\phi_1$. The scenario is given in Tab~\protect\ref{tab:spectrum}.} \vspace{0.cm} \end{figure} It should be noted from the previous plot that the asymmetry is obviously a CP-odd quantity that in addition to a measurement of the phase, also determines whether it lies above or below $\pi$, as seen in Fig.~\ref{fig:PartonAsy}. In comparison, using CP-even quantities, for example the mass, it is not possible to determine if the phase is above or below $\pi$. It must also be noted that these quantities have a weak dependence upon the phase and will be challenging to study. Moreover CP-even quantities alone do not give unambiguous proof of CP violation in the model, that can only be provided by CP-odd observables. \subsection{Influence of parton distribution functions on CP asymmetries} \label{sec:asymmetry-at-lhc} Experimentally the situation becomes significantly more complicated since in general, particles are not produced at rest but can be heavily boosted in the laboratory frame. Our observables can be significantly reduced in size by this effect as triple product correlations induced by spin effects are largest in the rest frame of the decaying particle. Essentially, a boosted frame can make the momentum vector of the quark appear to come from the opposite side of the plane formed by $\ell^+$ and $\ell^-$. As explained in Sec.~\ref{sec:Todd_Struc} this causes a severe dilution in the asymmetry that is measured at the LHC. There are two other dilution factors that have to be taken into account and give a further reduction in the observed asymmetry. The first one we consider is the contamination from anti-squarks $\tilde{q}_L^*$ that will be produced at the LHC along with squarks $\tilde{q}_L$. There is no way of identifying the charge of the decaying particle and consequently we need to include the effect of the $\tilde{q}_L^*$ in our analysis. Anti-squarks produce an asymmetry of the opposite sign to $\tilde{q}_L$ so if for example we had equal numbers of each of them, no overall asymmetry could be seen. However, the production cross section for $\tilde{q}_L^*$ is substantially lower than that of $\tilde{q}_L$ due to the valence quarks in the proton, see Tab.~\ref{tab:CrossSec} and at the LHC we would expect only roughly 20\% of the sample to be $\tilde{q}_L^*$. The other background contribution is that of neutralinos decaying to a tau pair, $\tilde{\chi}_2^0 \to \tilde{\chi}_1^0 \tau^+ \tau^-$ followed by leptonic tau decays to the pair of opposite-sign same-flavour leptons. Since the branching ratio for the above decay is large, even after inclusion of leptonic tau decays there is a significant number of events faking our signal, i.e. \begin{eqnarray}\label{eq:tautaubackg} BR(\tilde{\chi}^0_2 \to \tilde{\chi}^0_1 \tau^+ \tau^- \to \tilde{\chi}^0_1 \ell^+ \ell^-\nu_\ell\bar{\nu}_\ell \nu_\tau \bar{\nu}_\tau) \approx 2\times0.6\times0.175\times0.175 = 3.7\%. \label{eq:taudec} \end{eqnarray} compared with \begin{eqnarray} BR(\tilde{\chi}^0_2 \to \tilde{\chi}^0_1 \ell^+_{e,\mu} \ell^-_{e,\mu}) \approx 2\times0.045 = 9\%. \end{eqnarray} As the asymmetry calculated for such leptons is diluted, this background introduces a further dilution factor. However, this background can be easily removed using the flavour subtraction technique described above, Eqs.~(\ref{eq:NTplus}) and (\ref{eq:NTminus}). In addition we will see later that much of this background is removed after the application of simple selection cuts on the lepton energy and the invariant mass. We use the \texttt{MRST 2004LO} \cite{Martin:2007bv} PDF set in our analysis of the asymmetry and plot the integrated asymmetry $\mathcal{A_T}$ as a function of $\phi_{1}$ in Fig.~\ref{fig:PDFAsy}. We see that the inclusion of the PDFs and the $\tilde{q}_L^*$ sample reduce the asymmetry by about a factor of~8 in this scenario as compared to the result of Sec.~\ref{sec:asy-parton}. The maximum asymmetry is now $|\mathcal{A_T}|=1.7\%$. It must be noted that the dilution factor does depend on the scenario studied and changes in particular with the mass of the sparticle that is initially produced. For the calculation of the asymmetry we included the production channels shown in rows 3 and 4 of Table~\ref{tab:CrossSec} but we only take decays of individual $\tilde{q}_L$ and $\tilde{q}_L^*$ following the decay chain in Eq.~(\ref{eq-decays}). At this point correct identification was assumed for the final state particles and no hadronisation or detector effects were included but these assumptions will be relaxed in Sec.~\ref{sec:experimental}. The only backgrounds in the study are those discussed above. \begin{figure}[t!] \begin{picture}(16,8) \put(4.6,7.4){$\mathcal{T}=\overrightarrow{p}_q\cdot(\overrightarrow{p}_{\ell^+}\times\overrightarrow{p}_{\ell^-})$, $\,\,\sqrt{s} = 14\ \mathrm{TeV}$} \put(10.7,0.1){$\phi_{1}/\pi$} \put(6.2,5.95){\tiny{20 fb$^{-1}$}} \put(6.2,5.25){\tiny{50 fb$^{-1}$}} \put(6.2,4.6){\tiny{100 fb$^{-1}$}} \put(9.3,6.4){3$\sigma$ Observation} \put(2.7,7.4){\epsfig{file=LeftPDFCross.eps,scale=0.8,angle=270}} \end{picture} \caption{\label{fig:PDFAsy} The parton level asymmetry $\mathcal{A}_{\mathcal{T}}$, Eq.~(\protect\ref{Asy}), in the laboratory frame with PDFs included in the analysis using the scenario shown in Tab.~\protect\ref{tab:spectrum}. The coloured lines show the size of the asymmetry needed for a $3\sigma$ observation at the given luminosity, $\mathcal{L}$=(20 fb$^{-1}$, 50 fb$^{-1}$, 100 fb$^{-1}$), assuming squarks were produced via the channels shown in rows 3 and 4 of Tab.~\protect\ref{tab:CrossSec}. All produced $\tilde{q}$ and $\tilde{q}^*$ that follow the decay chain given in Eq.~(\protect\ref{eq-decays}) are taken into account.} \vspace{0.cm} \end{figure} Using the total production cross sections\footnote{Note that the total rate of squark production is actually larger than 30.0~pb quoted in Table~\ref{tab:CrossSec}, as some of the subprocesses give a pair of squarks e.g.$(\tilde{q}_L\tilde{q}_L)$ and both can contribute to our analysis. The total number of squarks in the sample at the given luminosity $\mathcal{L}$ and $\sqrt{s} = 14$~TeV is therefore $33.2\times \mathcal{L}$. In case of anti-squarks the number is $8.5\times \mathcal{L}$.} for $\tilde{q}_L$ and $\tilde{q}_L^*$, and respective branching ratios from Table~\ref{tab:brs} we can now estimate the integrated luminosity required to observe the asymmetry at the LHC with a certain significance. We assume that $N_{\mathcal{T}_+}\,(N_{\mathcal{T}_-})$, the numbers of events where $\mathcal{T}$ is positive (negative) as in Eq.~(\ref{Asy}), are binomially distributed, giving the following statistical error~\cite{Desch:2006xp}: \begin{equation} \label{eq:asymmerror} \Delta(\mathcal{A}_T)^{\rm stat}=2\sqrt{\epsilon(1-\epsilon)/N}\,, \end{equation} where $\epsilon=N_{\mathcal{T}_+}/(N_{\mathcal{T}_+}+N_{\mathcal{T}_-})=\frac 12 (1+\mathcal{A}_T)$, and $N = N_{\mathcal{T}_+}+N_{\mathcal{T}_-}$ is the total number of events. This can be rearranged to give the required number of events for a desired significance. The horizontal lines in Fig.~\ref{fig:PDFAsy} show an estimate of the amount of luminosity required for a $3\sigma$ observation of a non-zero asymmetry. In other words, an asymmetry can be seen at the 3$\sigma$ level where the asymmetry curve in Fig.~\ref{fig:PDFAsy} lies outside the luminosity band. \subsection{Impact of momentum reconstruction on the observable CP asymmetry} \label{sec:MomRecRes} In order to increase the statistical significance of our CP asymmetry we investigate the possibility of reconstructing the momenta of the invisible particles in our process. In principle a perfect reconstruction would return the magnitude of the asymmetry to that where the $\tilde{q}_L$ is at rest but in reality there are additional complications with this procedure, see Sec.~\ref{sec:MomRec}. The reconstruction is performed at the partonic level with PDF's included in the production process. Again our sample of events will contain $\tilde{q}^*_L$ which have an asymmetry of the opposite sign to that of $\tilde{q}_L$ as has already been discussed in Sec.~\ref{sec:asymmetry-at-lhc}. As we are looking exclusively at $\tilde{q}_L\tilde{g}$ and $\tilde{q}_L^*\tilde{g}$ production however when applying the momentum reconstruction, we actually have a smaller number of $\tilde{q}^*_L$ that dilute the asymmetry (15\%, see Table~\ref{tab:CrossSec}). After including this dilution factor we see our maximum asymmetry reduced from $\mathcal{A_T}\sim14\%$ to $\mathcal{A_T}\sim11\%$, see Figs.~\ref{fig:PartonAsy} and \ref{fig:MomRecAsyTau}, respectively. \begin{figure}[t!] \begin{picture}(16,8) \put(4.6,7.4){$\mathcal{T}=\overrightarrow{p}_q\cdot(\overrightarrow{p}_{\ell^+}\times\overrightarrow{p}_{\ell^-})$, $\,\,\sqrt{s} = 14\ \mathrm{TeV}$} \put(10.7,0.1){$\phi_{1}/\pi$} \put(6.2,5.2){\tiny{20 fb$^{-1}$}} \put(6.2,4.75){\tiny{50 fb$^{-1}$}} \put(6.2,4.3){\tiny{100 fb$^{-1}$}} \put(9.3,6.4){3$\sigma$ Observation} \put(2.7,7.4){\epsfig{file=MomRecCrossTau.eps,scale=0.8,angle=270}} \end{picture} \caption{\label{fig:MomRecAsyTau} The parton level asymmetry $\mathcal{A}_{\mathcal{T}}$, Eq.~(\protect\ref{Asy}), in the reconstructed $\tilde{q}_L$ rest frame. The coloured lines show the size of the asymmetry needed for a $3\sigma$ observation at the given luminosity, $\mathcal{L}$=(20 fb$^{-1}$, 50 fb$^{-1}$, 100 fb$^{-1}$) for the production channels $\tilde{q}_L\tilde{g}$, $\tilde{q}_L^*\tilde{g}$ Tab.~\protect\ref{tab:CrossSec}. The decay chains included are shown in Eq.~(\protect\ref{eq-decays}) and Eq~(\protect\ref{gludec}). As explained in the text correct jet and lepton assignment is assumed.} \vspace{0.cm} \end{figure} To calculate the luminosity we require to see a statistically significant effect at the LHC we include the production cross sections for channels $\tilde{q}_L\tilde{g}$, $\tilde{q}_L^*\tilde{g}$ (Tab.~\protect\ref{tab:CrossSec}) and the branching ratios from both the $\tilde{q}_L$ and $\tilde{g}$ decay chains Eq.~(\ref{eq-decays}) and Eq.~(\ref{gludec}). No hadronisation or detector effects were included in this section and correct identification was assumed for the quarks. For the leptons the correct assignment is made as is explained in Sec.~\ref{sec:PracApproach}. For the jets, a detailed experimental study would be required to look at this question as we need to include hard radiation, reconstruction and $b$-tagging efficiencies. However, a cursory examination suggests that the individual jets may be resolvable. For example, the jet coming from the $\tilde{q}_L$ decay is the hardest one for a high proportion of events. We would also require three other jets to have the invariant mass $m_t$. Finally we would then require at least one $b$-jet to be tagged. We would like to state here that the only backgrounds included in the study are those of $\tilde{q}_L^*\tilde{g}$ production and the decay chain Eq.~(\ref{eq:taudec}). After inclusion of all branching ratios the production rate for our process drops down to $200$~fb (after hadronic top decay), which results in approximately $20\,000$ events at the integrated luminosity $\mathcal{L} = 100\ \mathrm{fb}^{-1}$. One extra point to note is that we are able to use events where the $\tilde{\chi}^\pm_1$ in the opposite decay chain produces a $\tau^\pm$. Kinematically these events are similar to our normal signal events but we have one extra neutrino $\nu_{\tau}$ that will be invisible and simply contribute to our missing transverse momenta. An additional factor that reduces the number of reconstructed and accepted events are the multiple solutions, as discussed in Sec.~\ref{sec:MomRec}. We only use events when all solutions produce the same sign triple product and this allows us to use approximately $\sim60\%$ of the events. The horizontal lines in Fig.~\ref{fig:MomRecAsyTau} again show an estimate of the amount of luminosity required for a $3\sigma$ observation of a non-zero asymmetry. It can be seen that the luminosity lines are not flat but actually give a narrower band as $\phi_1\rightarrow\pi$. This effect is caused by the branching ratios, $\tilde{q}_L\to q \tilde{\chi}^0_2$, $\tilde{\chi}^0_2\to \tilde{\chi}^0_1 \ell^+ \ell^-$ and $\tilde{\chi}^+_1\to \tilde{\chi}^0_1 \overline{\nu}_{\ell} \ell^-$ altering with the change in phase and producing more events. We see that after 20~fb$^{-1}$ of well understood LHC data it may be possible to start seeing a statistically significant effect if large phases are present. It must be noted however that for this method to be successful we will require mass measurements of the individual particles involved in the decay chains. This will obviously require significant running time and may even need the help of a precision linear collider. We will address this issue in more detail in the next section. \section{Inclusion of experimental factors} \label{sec:experimental} To more realistically estimate if the study will be possible at the LHC some additional experimental factors need to be included in the analysis. For this purpose we simulate events using the Monte Carlo event generator \texttt{Herwig++ 2.3.2}, that has all the spin correlations included as required in our analysis. These events will be used in the following to perform momentum reconstruction and as a cross check for our analytic calculations. We start with the inclusion of the selection cuts that have to be used to resolve leptons and jets and these are listed below: \begin{eqnarray} E_{T}(j_1) &\geq& 100\ \mathrm{GeV}\,, \label{eq:CutJH} \\ E_{T}(j) &\geq& 25\ \mathrm{GeV}\,, \\ E_{T}(\ell_{e,\mu}) &\geq& 10\ \mathrm{GeV}\,, \label{eq:leppt} \\ M_{\ell^+\ell^-} &\geq& 20\ \mathrm{GeV}\,, \label{eq:lepinvmass} \\ |\eta| &\leq& 2.5\,. \label{eq:CutRap} \end{eqnarray} Here $E_{T}(j_1)$ is the transverse energy of the hardest jet, $E_{T}(j)$ is the transverse energy of all other jets, $E_{T}(\ell_{e,\mu})$ is the transverse energy of the leptons, $M_{\ell^+\ell^-}$ is the invariant mass of the opposite-sign same flavour lepton pair, and $\eta$ is the pseudorapidity of all the final state particles in the decay chain. Moreover we require at least two $b$-jets and that 1 $b$-jet plus 2 other jets (typically with the lowest $p_T$) should reconstruct the top quark. Since we need the top momentum to be reconstructed we only take into account tops that decay hadronically. One of the consequences of the application of the above cuts, especially Eq.~(\ref{eq:leppt}) and (\ref{eq:lepinvmass}) is the significant reduction in the background originating from $\tau$'s, Eq.~(\ref{eq:tautaubackg}). This is due to the rather low energy of the leptons coming from $\tau$ decays and the even lower invariant mass of the resulting lepton pair, which is peaking around 0. Already at this point, approximately 95\% of this background is removed. Another factor we include is the momenta of the resolved particles being smeared due to the intrinsic experimental precision. The accuracy for both jets and electrons follows the same function but with different coefficients~\cite{:2008zzm}: \begin{equation} \frac{\sigma_{E}}{E} = \sqrt {\frac{a^2_{j,e}}{E} + \frac{b^2_{j,e}}{E^2} + c^2_{j,e}}\ , \label{eq:JetAcc} \end{equation} where \begin{itemize} \item for jets $a_j = 0.6\ \mathrm{GeV}^{\frac{1}{2}}$, $b_j = 1.5\ \mathrm{GeV}$ and $c_j = 0.03$; \item for electrons the accuracy is better, with $a_e = 0.12\ \mathrm{GeV}^{\frac{1}{2}}$, $b_e = 0.2\ \mathrm{GeV}$ and $c_e = 0.01$. \end{itemize} The resolution for muons has a different functional dependence, \begin{equation} \frac{\sigma_{p_T}}{p_T} = \Bigg{\{} \begin{array}{cc} & 0.00008(p_T/\mathrm{GeV}-100)+0.03, \\ & 0.03, \end{array} \begin{array}{c} p_T>100 \,\mathrm{GeV}, \\ p_T<100 \,\mathrm{GeV}. \end{array} \end{equation} In addition we also have a finite resolution on the measurement of missing transverse energy, \begin{equation} \frac{\sigma^x_{MET}}{E_T}=\frac{\sigma^y_{MET}}{E_T}=\frac{0.57}{\sqrt{E_T/\mathrm{GeV}}}\,, \label{eq:METAcc} \end{equation} which are the errors on the $x$ and $y$ components of the MET vector and $E_T$ is the scalar sum of all visible transverse energy . The momenta smearing will only affect the observables when we perform momentum reconstruction as the LSP momenta will be reconstructed with limited precision. The triple product will not suffer however as the measurement only relies on the direction of the measured particles and not on the energy measurement. The direction can be found far more accurately and this error happens to be negligible for our observables. Finally, we also investigate the fact that the masses of the particles in the decay chains we are interested in will only be known with a certain precision at the LHC (we assume 10\% error), see Section~\ref{sec:scenario}. This error will again only affect the observables when we perform momentum reconstruction in order to boost into the rest frame of the $\tilde{\chi}^0_2$ and can cause the frame to be mis-measured, see Section~\ref{sec:expfactors} for more details. \subsection{Experimental factors without momentum reconstruction} Out of the experimental factors mentioned above, only the cuts affect the result for the triple product correlation measured in the laboratory frame. These cuts reduce the number of detectable events by $\approx50\%$ and consequently significantly increase the luminosity required to make a statistically significant measurement at the LHC, see Fig.~{\ref{fig:SqCuts}}. For example if large phases are present we may begin to see hints with integrated luminosity $\mathcal{L}=50\ \mathrm{fb}^{-1}$. With $\mathcal{L} = 300\ \mathrm{fb}^{-1}$ we could become sensitive to phases in the ranges $0.15\,\pi \lesssim \phi_{1} \lesssim 0.9\,\pi$ and $1.1\,\pi \lesssim \phi_{1} \lesssim 1.85\,\pi$ where the asymmetry $|\mathcal{A_T}|>0.7$\%. \begin{figure}[t!] \begin{picture}(16,8) \put(4.6,7.4){$\mathcal{T}=\overrightarrow{p}_q\cdot(\overrightarrow{p}_{\ell^+}\times\overrightarrow{p}_{\ell^-})$, $\,\,\sqrt{s} = 14\ \mathrm{TeV}$} \put(10.7,0.1){$\phi_{1}/\pi$} \put(6.2,5.35){\tiny{50 fb$^{-1}$}} \put(6.2,4.9){\tiny{100 fb$^{-1}$}} \put(6.2,4.4){\tiny{300 fb$^{-1}$}} \put(9.3,6.4){3$\sigma$ Observation} \put(2.7,7.4){\epsfig{file=LeftPDFCuts.eps,scale=0.8,angle=270}} \end{picture} \caption{\label{fig:SqCuts} The asymmetry $\mathcal{A}_{\mathcal{T}}$, Eq.~(\protect\ref{Asy}), in the laboratory frame for the decay chain Eq.~(\protect\ref{eq-decays}) after the cuts, Eqs.~(\protect\ref{eq:CutJH})-(\protect\ref{eq:CutRap}), have been applied. The coloured lines show the size of the asymmetry needed for a $3\sigma$ observation at the given luminosity, $\mathcal{L}$=(50 fb$^{-1}$, 100 fb$^{-1}$, 300 fb$^{-1}$), assuming squarks were produced via the channels shown in rows 3 and 4 of Tab.~\protect\ref{tab:CrossSec}. Momentum smearing for both the leptons and quarks was studied and found to have a negligible effect. All other relevant experimental details are mentioned in the text.} \vspace{0.cm} \end{figure} For the calculation of the asymmetry we included the production channels shown in rows 3 and 4 of Table~\ref{tab:CrossSec} taking into account the decays of individual $\tilde{q}_L$ and $\tilde{q}_L^*$ as listed in Eq~(\ref{eq-decays}). No explicit hadronisation was included but momentum smearing is expected to simulate some of this effect. We also need to include additional hard QCD radiation and other detector effects (for example fakes) in a full experimental study. Correct identification was assumed for the quark and two leptons. The backgrounds in the study are those from $\tilde{q}_L^*$ and Eq.~(\ref{eq:taudec}). It must be noted however, that significant pollution due to backgrounds will be expected for this signal from both the Standard Model and the MSSM. Further experimental cuts will certainly be required to improve the signal/background ratio and without a more detailed study it is hard to predict what effect this will have on the ability to complete this measurement at the LHC, especially as the asymmetries are rather small. However, a more detailed experimental analysis is beyond the scope of the present paper \cite{newpaper}. \subsection{Experimental factors with momentum reconstruction}\label{sec:expfactors} When we perform the momentum reconstruction we need to include the experimental precision on the momentum of the visible particles. This resolution is $\sim3\%$ for leptons and follows Eq.~(\ref{eq:JetAcc}) for jets. The corresponding effect on momentum reconstruction is a reduction in the number of events that have the same sign triple product. As stated in Sec.~\ref{sec:MomRecRes} we discard any events where we have solutions with opposite sign triple products. Discarding these events reduces the percentage we can use from $\sim60\%$ without momentum smearing down to $\sim30\%$ when we include it. The other difficulty momentum smearing creates is that all the reconstructed solutions can now have the wrong sign triple product as we no longer correctly reproduce the rest frame of the neutralino $\tilde{\chi}^0_2$. Inevitably this effect produces a decrease in the observed asymmetry from $\sim11\%$ to $\sim8\%$. We again include the cuts on all visible particles in our decay chain given by Eqs.~(\ref{eq:CutJH})-(\ref{eq:CutRap}). These cuts significantly reduce the number of visible events and remove $\sim80\%$ of the events compared with our initial na\"{\i}ve estimates. When we combine the cuts with the momentum reconstruction efficiency we are left with $\sim6\%$ of the initial events and this clearly increases the luminosity needed to make an observation statistically significant. After inclusion of these effects the number of events drops from $20000$ down to $1200$ at the integrated luminosity of $\mathcal{L} = 100\ \mathrm{fb}^{-1}$. This results in a $1\sigma$ absolute uncertainty of order $\sim 3\%$ on the asymmetry, according to Eq.~(\ref{eq:asymmerror}). Another possible experimental aspect we investigate is a 10\% uncertainty on the masses of the supersymmetric particles used in the momentum reconstruction. We found that this has a negligible effect on the momentum reconstruction as long as the mass differences between different particles in the decay chain are known better than $\mathcal{O}$(5 GeV). The assumption that the mass differences are known to a higher accuracy is reasonable as the main method of measuring masses in SUSY decay chains at the LHC will be via kinematic end points that are measured with high precision. \begin{figure}[t!] \begin{picture}(16,8) \put(4.6,7.4){$\mathcal{T}=\overrightarrow{p}_q\cdot(\overrightarrow{p}_{\ell^+}\times\overrightarrow{p}_{\ell^-})$, $\,\,\sqrt{s} = 14\ \mathrm{TeV}$} \put(10.7,0.1){$\phi_{1}/\pi$} \put(6.2,5.8){\tiny{50 fb$^{-1}$}} \put(6.2,5.2){\tiny{100 fb$^{-1}$}} \put(6.2,4.60){\tiny{300 fb$^{-1}$}} \put(9.3,6.4){3$\sigma$ Observation} \put(2.7,7.4){\epsfig{file=MomRecCrossCutsTau.eps,scale=0.8,angle=270}} \end{picture} \caption{\label{fig:MomRecCuts} The asymmetry $\mathcal{A}_{\mathcal{T}}$, Eq.~(\protect\ref{Asy}), in the reconstructed $\tilde{q}_L$ frame for $\tilde{q}_L\tilde{g}$, $\tilde{q}_L^*\tilde{g}$ production followed by the decay chains shown in Eq.~(\protect\ref{eq-decays}) and Eq~(\protect\ref{gludec}) with cuts Eqs.~(\protect\ref{eq:CutJH})-(\protect\ref{eq:CutRap}). The coloured lines show the size of the asymmetry needed for a $3\sigma$ observation at the given luminosity, $\mathcal{L}$=(50 fb$^{-1}$, 100 fb$^{-1}$, 300 fb$^{-1}$). The momenta of the final state particles have been smeared according to Eq.~(\protect\ref{eq:JetAcc})-(\protect\ref{eq:METAcc}) to replicate the LHC detector effects. All other relevant experimental details are mentioned in the text.} \vspace{0.cm} \end{figure} Figure~\ref{fig:MomRecCuts} shows the asymmetry and the luminosity required at the LHC to see a statistically significant result at the 3$\sigma$ level at the LHC once all the above factors have been taken into account. The production channels are $\tilde{q}_L\tilde{g}$, $\tilde{q}_L^*\tilde{g}$ and the branching ratios are included from both the $\tilde{q}_L$ and $\tilde{g}$ decay chains. No explicit hadronisation effects were included here and possible additional hard jets in the final state require further study. Also correct identification was assumed for all the leptons and quarks in both decay chains as explained in Sec.~\ref{sec:MomRecRes} and Sec.~\ref{sec:PracApproach}. Again the only backgrounds included in the study are those of $\tilde{q}_L^*\tilde{g}$ production and of taus that decay to visible leptons, Eq.~(\ref{eq:taudec}). We can see that using this method, with integrated luminosity $\mathcal{L} \sim 100\, \mathrm{fb}^{-1}$ we start to become sensitive if large phases happen to be present. After $\mathcal{L} \sim 300\, \mathrm{fb}^{-1}$ we expect to have sensitivity to phases in the ranges $0.2\,\pi \lesssim \phi_{1} \lesssim 0.85\,\pi$ and $1.15\,\pi \lesssim \phi_{1} \lesssim 1.8\,\pi$, and obviously more luminosity will improve this further. Note that a direct comparison between the methods with and without momentum reconstruction based on the above plots should not be performed. The backgrounds from both SM and MSSM will be more severe when we do not perform momentum reconstruction and clearly many new cuts will be required to isolate the signal, which is the consequence of being totally inclusive. For the method when we perform momentum reconstruction, we have a well defined final state that is difficult to be faked by Standard Model processes and is also uncommon for SUSY cascade decays. Moreover the multiple cuts on all the particles and the need for the missing momentum to be successfully reconstructed mean that many backgrounds will be rejected. \subsection{Ongoing work} \label{subsec:Ongoing} A new approach in the future study~\cite{newpaper} is to use matrix element weighting for each of the multiple solutions. This would significantly increase the number of accepted events and possibly increase the statistical significance of the result as we had observed that some of the ``wrong'' solutions obey rather unlikely kinematic configurations. However, the validity of this method needs further study. The other possible improvement is the application of the cuts on anti-squarks. In the associated production, Eq.~(\ref{eq-prod}), anti-squarks come always from the sea quarks, whereas squarks can originate both from the sea quarks and the valence quarks. Therefore one can expect some difference in their distributions. This would give a handle to cut out one of our most severe backgrounds. Preliminary studies showed that indeed this selection criteria can turn out to be useful depending on the actual scenario and the collected luminosity. It is clear that to fully understand if these measurements are possible at the LHC a detector simulation including backgrounds will be required. This study has to include the combinatorial issues that may be encountered with both the final state leptons and the jets. In addition, the impact of particle reconstruction inefficiencies, falsely identified particles (``fakes''), different detector acceptance for different particle species and other experimental considerations need to be considered. Such an analysis is in the line of future work. Further improvements in measuring the CP-violating phase $\phi_1$ would require the inclusion of more observables sensitive to the phase, like lepton opening angle or lepton angular distribution \cite{Choi:2005gt}. Nevertheless, in order to determine the phase in a precise and rather model-independent way, experimental input from the clean and unbiased measurements at a future linear collider \cite{Bartl:2004jj,Bartl:2005uh,Bartl:2006yv} is required in addition to the LHC. \section{Conclusions} In this paper we have investigated the problem of discovering CP-violating effects at the Large Hadron Collider. Our tool in this analysis was the triple product correlation of momenta of the quark and the lepton pair from the $\tilde{\chi}^0_2$ decay. This kind of T-odd observables can give us a handle to discover the presence of complex phases in the model. We took into account the full production and decay process of squark and gluino. Since triple products depend crucially on spin correlations, they have been included both in the analytical calculation and the event generation, that has been performed using \texttt{Herwig++ 2.3.2}. The process of special interest in our case was the squark decay into quark and neutralino $\tilde{\chi}_2^0$ followed by the three-body leptonic neutralino decay. For this decay in our scenario one can expect an asymmetry in the triple product distribution of up to 14\% when calculated in the rest frame of the decaying neutralino. The source of the CP violation in our case was the phase of the bino mass parameter $M_1$. Due to the hadronic experimental environment of the LHC, precise measurements will be a challenge both from experimental and theoretical point of view. Therefore we have considered two possible methods of measuring CP-violating effects in the neutralino decay. The first method calculates the triple product in the laboratory frame. It turns out however that the initial CP-odd asymmetry is diluted both by high boosts of produced particles and by the admixture of anti-squarks. This makes its observation far more difficult but can give us hints of CP violation after basic selection cuts in our example. As mentioned before the largest effects of CP violation can be expected in the rest frame of the decaying neutralino. We studied the use of momentum reconstruction of invisible LSPs and neutrinos to get access to the rest frame of the neutralino $\tilde{\chi}^0_2$. Using a set of invariant kinematic conditions we showed that it is possible to fully reconstruct the production and decay process on an event-by-event basis. The important result here is the possibility of reconstruction in the case of three-body decays and escaping neutrinos. Having fully reconstructed events we are able to boost particle momenta back to the rest frame of the neutralino $\tilde{\chi}^0_2$ and the asymmetry becomes as large as 10\%. Again, we include experimental event selection criteria as well as momentum smearing to account for finite detector resolution. We also consider the most severe background due to anti-squarks and we showed that the signal contamination from neutralino decays to leptonically decaying tau pairs does not pose a problem for our measurement. Applying momentum reconstruction and taking into account the experimental effects we have found that the CP-odd asymmetry drops down to 8\%. To avoid ambiguities and due to experimental constraints many events have to be discarded, nevertheless the left-over sample is enough to probe CP-violation effects. One should see a $3\sigma$ effect at $\mathcal{L} = 300\ \mathrm{fb}^{-1}$ and we expect sensitivity to the phase in the range $0.2\,\pi \lesssim \phi_{1} \lesssim 0.85\,\pi$. We emphasise that the asymmetry after momentum reconstruction is a much cleaner observable from a theoretical point of view, thanks to a well defined final state. Therefore, using the above technique there are good prospects for the observation or exclusion of CP-violating effects for a wide range of the phase $\phi_1$ after a few years of LHC running at the high luminosity. The full assessment of LHC's ability to resolve CP violation in the MSSM will definitely require a realistic simulation of detector effects, SM and SUSY backgrounds. That issue is beyond the scope of the present phenomenological analysis and therefore we leave this question for future study. The momentum reconstruction procedure turns out to be extremely useful method for studying new physics models at the LHC, as we have shown in this paper. Nevertheless, in order to reveal the underlying model, including CP-violating phases, a more precise machine, like an $e^+e^-$ linear collider, will be required. \section*{Acknowledgements} The authors wish to thank Philip Bechtle, Bj\"{o}rn Gosdzik, Bob McElrath and Filip Moortgat for valuable discussions. We also are grateful to David Grellscheid and Peter Richardson for their assistance in the use of \texttt{Herwig++}. KR is supported by the EU Network MRTN-CT-2006-035505 ``Tools and Precision Calculations for Physics Discoveries at Colliders'' (HEPTools). JT is supported by the UK Science and Technology Facilities Council (STFC). \newpage \section*{Appendices} \begin{appendix} \section{Interaction Lagrangian and neutralino mixing} \label{sec:lagrangian-couplings} Neglecting the mixing effects in the slepton sector, the interaction Lagrangian terms for $\tilde{\chi}^0_2$ decay are \begin{eqnarray} \label{eq:LagZll} && \mathcal{L}_{Z^0\ell^+\ell^-} = -\frac{g}{\cos\theta_W}Z_{\mu}\bar{\ell}\gamma^{\mu} [L_\ell P_L + R_\ell P_R]\ell\,, \\ && \mathcal{L}_{Z^0\tilde{\chi}^0_m\tilde{\chi}^0_n} = \frac{g}{2 \cos\theta_W} Z_{\mu}{\bar{\tilde{\chi}}}^0_m\gamma^{\mu} [{O}_{mn}^L P_L + {O}_{mn}^R P_R]{\tilde{\chi}}^0_n\,, \\ && \mathcal{L}_{\ell\tilde{\ell}\tilde{\chi}^0_k} = {g} f_{Lk}^{\ell}\: \bar{\ell}\: P_R\: \tilde{\chi}^0_k\: \tilde{\ell}_L + {g} f_{Rk}^{\ell}\: \bar{\ell}\: P_L\: \tilde{\chi}^0_k\: \tilde{\ell}_R + \textrm{h.c.}\,, \end{eqnarray} where $P_{L,R}=\frac{1}{2}(1\mp\gamma_5)$, $g = e/\sin\theta_W$ and $\theta_W$ is the weak mixing angle. The couplings are given by \begin{eqnarray} f_{Lk}^{\ell} &=& \frac{1}{\sqrt{2}} \, (\tan\theta_W N_{k1} + {N_{k2}})\,, \\ f_{Rk}^{\ell} &=& -\sqrt{2}\, \tan\theta_W N_{k1}^* \,, \\ {O}_{mn}^L &=& N_{m4} N_{n4}^* - N_{m3} N_{n3}^*\,,\\ {O}_{mn}^R &=& - {O}_{mn}^{L*}\,,\\ L_\ell &=& -\frac{1}{2} + \sin^2\theta_W\,,\\ R_\ell &=& \sin^2\theta_W\,, \end{eqnarray} where $N$ is the neutralino mixing matrix defined by Eq.~(\ref{eq:Nmatrix}). \vspace{0.5cm} The interaction Lagrangian for $\tilde{q}_L$ decay is \begin{equation} \mathcal{L}_{q\ti q \tilde{\chi}^0_k} = g f^{q}_{Lk}\: \bar{q}\: P_R\: \tilde{\chi}^0_k\: \ti q_L+ \mathrm{h.c.}, \end{equation} with \begin{eqnarray} f_{Lk}^{q} &=& \frac{1}{3 \sqrt{2}} \, (- T_{3q} \tan\theta_W N_{k1} + 3 {N_{k2}})\,, \label{eq:fLk} \end{eqnarray} where $T_{3q}$ is a weak isospin of a squark and we have neglected small Yukawa couplings for the first two generations of squarks and mixing between left and right squark states. \vspace{0.5cm} The complex symmetric neutralino mass matrix ${\cal M}_N$ is given in the basis $(\tilde{B}, \tilde{W}^3, \tilde{H}^0_1, \tilde{H}^0_2)$ by: \begin{equation} {\cal M}_N = \left( \begin{array}{cccc} M_1 & 0 & -m_Z c_\beta s_W & m_Z s_\beta s_W \\ 0 & M_2 & m_Z c_\beta c_W & -m_Z s_\beta c_W \\ -m_Z c_\beta s_W & m_Z c_\beta c_W & 0 & -\mu \\ m_Z s_\beta s_W & -m_Z s_\beta c_W & -\mu & 0 \\ \end{array} \right)\,, \end{equation} where the abbreviations $s_W=\sin\theta_W,\, c_W=\cos\theta_W,\, s_{\beta}=\sin\beta,\, c_{\beta}=\cos\beta$ have been used and $\tan\beta=v_2/v_1$ is the ratio of the vacuum expectation values of the Higgs fields. It is diagonalised by the unitary $4\times4$ matrix $N$ \begin{equation} {\rm diag}(m_{\tilde{\chi}_1^0},m_{\tilde{\chi}_2^0},m_{\tilde{\chi}_3^0},m_{\tilde{\chi}_4^0}) = N^*{\cal M}_N N^\dagger\,,\,\,\,\,\,\,\,\,\,\, (m_{\tilde{\chi}_1^0}<m_{\tilde{\chi}_2^0}<m_{\tilde{\chi}_3^0}<m_{\tilde{\chi}_4^0})\label{eq:Nmatrix} \end{equation} giving 4 non-negative Majorana neutralino mass eigenstates. The masses in Eq.~(\ref{eq:Nmatrix}) can be chosen real and positive by a suitable definition of the unitary matrix $N$. For convenience, in the following we use the notation: $m_i \equiv m_{\tilde{\chi}_i^0}$. \section{Amplitude squared including full spin correlations} \subsection{Neutralino production $\tilde{q}_L \to \tilde{\chi}^0_j q$\label{sect:squarkdecay}} The analytic expression for the production density matrix can be decomposed as \begin{equation} |M(\tilde{q}_L\to \tilde{\chi}^0_j q)|^2=P(\tilde{\chi}^0_j q)+\Sigma^a_P(\tilde{\chi}^0_j)\,, \label{eq_mq} \end{equation} whose spin-independent contribution reads \begin{equation} P(\tilde{\chi}^0_j q)=\frac{g^2}{2} |f_{Lj}^{q}|^2 (p_q p_{\tilde{\chi}^0_j})\,, \end{equation} where $p_q$ and $p_{\tilde{\chi}^0_j}$ denote the four-momenta of the quark $q$ and the neutralino $\tilde{\chi}^0_j$. The spin-dependent contributions is T-even and given by \begin{equation} \Sigma^a_P(\tilde{\chi}^0_j) = - \frac{g^2}{2} |f_{Lj}^{q}|^2 m_j (p_q s^a(\tilde{\chi}^0_j)) \,,\label{eq_prod-ea} \end{equation} where $s^a(\tilde{\chi}^0_j)$ denotes the spin-basis vector of the neutralino $\tilde{\chi}^0_j$. The three spin-basis four-vectors $s^1$, $s^2$ and $s^3$ form a right-handed system and provide, together with the momentum, an orthogonal basis system. They are chosen as: \begin{eqnarray} s^1(\tilde{\chi}^0_j)&=&\left(0,\frac{(\vec{p}_{\tilde{\chi}^0_j}\times \vec{p}_{\tilde{q}_L})\times \vec{p}_{\tilde{\chi}^0_j}}{|(\vec{p}_{\tilde{\chi}^0_j}\times \vec{p}_{\tilde{q}_L})\times \vec{p}_{\tilde{\chi}^0_j}|}\right)\,, \label{eq-s1chi}\\ s^2(\tilde{\chi}^0_j)&=&\left(0,\frac{\vec{p}_{\tilde{\chi}^0_j}\times \vec{p}_{\tilde{q}_L }} {|\vec{p}_{\tilde{\chi}^0_j}\times \vec{p}_{\tilde{q}_L}|}\right)\,, \label{eq-s2chi}\\ s^3(\tilde{\chi}^0_j)&=&\frac{1}{m_j}\left( |\vec{p}_{\tilde{\chi}^0_j}|, \frac{E_{\tilde{\chi}^0_j}}{|\vec{p}_{\tilde{\chi}^0_j}|} \vec{p}_{\tilde{\chi}^0_j}\right)\,. \label{eq-s3chi} \end{eqnarray} For a more detailed discussion see \cite{Ellis:2008hq}. \subsection{Neutralino three-body decay $\tilde{\chi}^0_j \to \tilde{\chi}^0_k \ell^+ \ell^-$ \label{sect:neutdecay}} Here we give the analytical expressions for the different contributions to the decay density matrix for the three-body leptonic neutralino decay, where we sum over the spins of the final-state particles~\cite{MoortgatPick:1999di}. The contributions independent of the polarisation of the neutralino $\tilde{\chi}^0_j$ \begin{equation} D(\tilde{\chi}^0_j)=D(Z Z)+ D(Z \tilde{\ell}_L)+ D(Z \tilde{\ell}_R)+ D(\tilde{\ell}_L \tilde{\ell}_L)+ D(\tilde{\ell}_R \tilde{\ell}_R) \label{eq_sumz}\,, \end{equation} are given by \begin{eqnarray} D(Z Z)&=& 8 \frac{g^4}{\cos^4\theta_W} |\Delta(Z)|^2 (L_{\ell}^2+R_{\ell}^2) \nonumber\\ & & \times \Big[ |O^{L}_{kj}|^2 (g_1+g_2) +(\mathrm{Re}\: O^{L}_{kj})^2 -(\mathrm{Im}\: O^{L}_{kj})^2) g_3 \Big]\,, \label{eq_dzz}\\ D(Z \tilde{\ell}_L)&=&4 \frac{g^4}{\cos^2\theta_W} L_{\ell}\, \mathrm{Re}\:\Big\{\Delta(Z) \Big[f^{\ell}_{L j} f^{\ell*}_{L k} \Delta_t^{*}(\tilde{\ell}_L) (2O^{L}_{kj} g_1 +O^{L*}_{kj} g_3) \nonumber\\ & &\phantom{4 \frac{g^4}{\cos^2\theta_W} L_{\ell} Re\Big\{\Delta(Z) \Big[} +f^{\ell*}_{L j} f^{\ell}_{L k} \Delta_u^{*}(\tilde{\ell}_L) (2O^{L*}_{kj} g_2 +O^{L}_{kj} g_3) \Big]\Big\}\,,\label{eq_dzel}\\ D(\tilde{\ell}_L \tilde{\ell}_L)&=& 2 g^4 \Big[ |f^{\ell}_{L j}|^2 |f^{\ell}_{L k}|^2 \big(|\Delta_t(\tilde{\ell}_L)|^2 g_1 +|\Delta_{u}(\tilde{\ell}_L)|^2 g_2\big)\nonumber\\ & &\phantom{2 g^4 \Big[} +\mathrm{Re}\:\big\{(f^{\ell*}_{L j})^2 (f^{\ell}_{L k})^2 \Delta_{t}(\tilde{\ell}_L) \Delta_u^{*}(\tilde{\ell}_L)\big\} g_3 \Big]\,, \label{eq_delel} \end{eqnarray} where $\Delta(Z)$ and $\Delta_{t,u}(\tilde{\ell}_{L})$ denote the propagators of the virtual particles in the direct channel and in both crossed channels (labelled $t,u$, cf.\ Fig.~\ref{Fig:FeynDecayA}). The quantities $D(Z\tilde{\ell}_R)$ and $D(\tilde{\ell}_R \tilde{\ell}_R)$ can be derived from Eqs.~(\ref{eq_dzel}) and (\ref{eq_delel}) by the substitutions \begin{equation} \label{eq_substdecayP} L_{\ell}\to R_{\ell}\,, \quad \Delta_{t,u}(\tilde{\ell}_L)\to \Delta_{t,u}(\tilde{\ell}_R)\,,\quad O^{L}_{kj}\to O^{R}_{kj}\,, \quad f_{L j,k}^{\ell}\to f_{R j,k}^{\ell}\,. \end{equation} The kinematic factors are \begin{eqnarray} g_1&=&(p_{\tilde{\chi}^0_k} p_{\ell^-})(p_{\tilde{\chi}^0_j} p_{\ell^+})\,, \label{eq_dkin1}\\ g_2&=& (p_{\tilde{\chi}^0_k} p_{\ell^+})(p_{\tilde{\chi}^0_j} p_{\ell^-})\,, \label{eq_dkin2}\\ g_3 &=& m_j m_k (p_{\ell^-} p_{\ell^+})\,.\label{eq_dkin3} \end{eqnarray} We can split the terms depending on the polarisation of the neutralino into T-even and T-odd contributions: \begin{equation} \Sigma_D^{a}(\tilde{\chi}^0_j)= \Sigma_D^{a,E}(\tilde{\chi}^0_j) +\Sigma_D^{a,O}(\tilde{\chi}^0_j)\,. \end{equation} The T-even contributions depending on the polarisation of the decaying neutralino $\tilde{\chi}^0_j$ \begin{equation} \Sigma_D^{a,E}(\tilde{\chi}^0_j)= \Sigma_D^{a,E}(ZZ) +\Sigma_D^{a,E}(Z \tilde{\ell}_L) +\Sigma_D^{a,E}(Z \tilde{\ell}_R) +\Sigma_D^{a,E}(\tilde{\ell}_L \tilde{\ell}_L) +\Sigma_D^{a,E}(\tilde{\ell}_R \tilde{\ell}_R)\, \label{eq_dssum-e} \end{equation} are \begin{eqnarray} \Sigma_D^{a,E}(ZZ)&=& 8 \frac{g^4}{\cos^4\theta_W} |\Delta(Z)|^2 (R^2_{\ell}-L_{\ell}^2) \nonumber\\ & &\times\Big[ -[(\mathrm{Re}\: O^{L}_{kj})^2 -(\mathrm{Im}\: O^{L}_{kj})^2]g^a_3+ |O^{L}_{kj}|^2(g^a_1-g^a_2) \Big]\,,\label{eq_dszz}\\ \Sigma_D^{a,E}(Z \tilde{\ell}_L)&=& \frac{4 g^4}{\cos^2\theta_W}L_{\ell}\, \mathrm{Re}\:\Big\{\Delta(Z) \Big[ f^{\ell}_{L j} f^{\ell*}_{L k} \Delta_t^{*}(\tilde{\ell}_L) \big(-2 O^{L}_{kj} g^a_1 +O^{L*}_{kj} g^a_3\big)\nonumber\\ & & \phantom{\frac{4 g^4}{\cos^2\theta_W}L_{\ell}\, Re\Big\{} + f^{\ell*}_{L j} f^{\ell}_{L k} \Delta_u^{*}(\tilde{\ell}_L) \big(2 O^{L*}_{kj} g^a_2 +O^{L}_{kj} g^a_3\big) \Big]\Big\}\,,\label{eq_dszel}\\ \Sigma_D^{a,E}(\tilde{\ell}_L \tilde{\ell}_L) & = & 2 g^4 \Big[ |f^{\ell}_{L j}|^2 |f^{\ell}_{L k}|^2 [|\Delta_{u}(\tilde{\ell}_L)|^2 g_2^a -|\Delta_{t}(\tilde{\ell}_L)|^2 g_1^a]\nonumber\\ & & \phantom{2 g^4 \Big[} + \mathrm{Re}\:\big\{ (f^{\ell*}_{L j})^2 (f^{\ell}_{L k})^2 \Delta_{t}(\tilde{\ell}_L) \Delta_u^{*}(\tilde{\ell}_L)g_3^a\big\}\Big]\,, \label{eq_dselel} \end{eqnarray} where the contributions $\Sigma^{a,E}_D(Z\tilde{\ell}_R)$ and $\Sigma^{a,E}_D(\tilde{\ell}_R \tilde{\ell}_R)$ are derived from Eqs.~(\ref{eq_dszel}) and (\ref{eq_dselel}) by applying the substitutions given by Eq.~(\ref{eq_substdecayP}) and in addition $g_{1,2,3}^a\to -g_{1,2,3}^a$. The kinematic factors are \begin{eqnarray} g^a_1&=& m_j (p_{\tilde{\chi}^0_k} p_{\ell^-}) (p_{\ell^+} s^a)\,, \label{eq422_3a}\\ g^a_2&=& m_j (p_{\tilde{\chi}^0_k} p_{\ell^+}) (p_{\ell^-} s^a)\,, \label{eq_dssub4}\\ g^a_3&=& m_k [(p_{\tilde{\chi}^0_j} p_{\ell^+}) (p_{\ell^-} s^a) -(p_{\tilde{\chi}^0_j} p_{\ell^-}) (p_{\ell^+} s^a)]\,. \label{eq_dssub5} \end{eqnarray} The T-odd contributions depending on the polarisation of the decaying neutralino $\tilde{\chi}^0_j$ \begin{equation} \Sigma_D^{a,O}(\tilde{\chi}^0_j)= \Sigma_D^{a,O}(ZZ) +\Sigma_D^{a,O}(Z \tilde{\ell}_L) +\Sigma_D^{a,O}(Z \tilde{\ell}_R) +\Sigma_D^{a,O}(\tilde{\ell}_L \tilde{\ell}_L) +\Sigma_D^{a,O}(\tilde{\ell}_R \tilde{\ell}_R)\, \label{eq_dssum-o} \end{equation} are \begin{eqnarray} \Sigma_D^{a, \mathrm{O}}(ZZ) &=& 8 \frac{g^4}{\cos^4\theta_W} |\Delta(Z)|^2 (L^2_{\ell} - R_{\ell}^2) \Big[ 2 \mathrm{Re}\:(O^{L}_{kj})\, \mathrm{Im}\:(O^{L}_{kj}) i g_4^a \Big]\,, \label{eq_dszz_to}\\ \Sigma_D^{a, \mathrm{O}}(Z \tilde{\ell}_L) & = & \frac{4 g^4}{\cos^2\theta_W}L_{\ell}\, \mathrm{Re}\:\Big\{\Delta(Z) \Big[-f^{\ell}_{L j} f^{\ell*}_{L k} O^{L*}_{kj} \Delta_t^{*}(\tilde{\ell}_L)\nonumber\\ & &\phantom{\frac{4 g^4}{\cos^2\Theta_W}L_{\ell}Re\Big\{\Delta(Z)\Big[} + f^{\ell*}_{L j} f^{\ell}_{L k} O^{L}_{kj} \Delta_u^{*}(\tilde{\ell}_L) \Big]g_4^a\Big\}\,,\label{eq_dszel_to}\\ \Sigma_D^{a, \mathrm{O}}(\tilde{\ell}_L \tilde{\ell}_L)&=& 2 g^4\, \mathrm{Re}\:\Big\{ (f^{\ell*}_{L j})^2 (f^{\ell}_{L k})^2 \Delta_{t}(\tilde{\ell}_L) \Delta_u^{*}(\tilde{\ell}_L) g_4^a\Big\}\,, \label{eq_dselel_to} \end{eqnarray} where the contributions $\Sigma^{a,O}_D(Z\tilde{\ell}_R)$ and $\Sigma^{a,O}_D(\tilde{\ell}_R \tilde{\ell}_R)$ are derived from Eqs.~(\ref{eq_dszel}) and (\ref{eq_dselel}) by applying the substitutions given by Eq.~(\ref{eq_substdecayP}). The kinematic factor is \begin{eqnarray} g^a_4 & = & i m_k \epsilon_{\mu \nu \rho \sigma} s^{a \mu} p_{\tilde{\chi}^0_j}^{\nu} p_{\ell^-}^{\rho} p_{\ell^+}^{\sigma}\,. \label{eq_dssub6} \end{eqnarray} \section{Kinematics} \label{sec:kinematics} \subsection{Phase space} The complete cross section for the process can be decomposed into the production cross section and the branching ratios of the subsequent decays: \begin{eqnarray} d\sigma_{\mathrm{tot}} &=& d{\sigma}(qg \to \ti q_L \tilde{g})\cdot\frac{E_{\ti q_L }}{m_{\ti q_L}\Gamma_{\ti q_L}}\, d\Gamma (\ti q_L \to q \tilde\chi^0_2)\,\cdot \, \frac{E_{\tilde{\chi}^0_2}}{m_{2} \Gamma_{\tilde{\chi}^0_2}} \,d\Gamma (\tilde\chi^0_2\to \tilde{\chi}^0_1 \ell^+ \ell^-)\, ,\label{dGamma} \end{eqnarray} where the factors $E/m \Gamma$ come from the use of the narrow-width approximation for the propagators of the $\ti q_L$ and $\tilde{\chi}^0_2$. This approximation is valid for $(\Gamma/m)^2 \ll 1$, which is satisfied for $\Gamma_{\ti q_L} \sim 4.7$~GeV and for $\Gamma_{\tilde{\chi}^0_2}\sim 3\times10^{-5}$~GeV where the width is small because only the three-body decay is kinematically possible. We have: \begin{eqnarray} d\Gamma (\ti q_L \to q \tilde\chi^0_2) &=&\frac2{E_{\ti q_L}}P(\tilde{\chi}^0_2q) \,d\Phi_{\ti q}\,, \\ d\Gamma (\tilde{\chi}^0_2 \to \tilde{\chi}^0_1 \ell^+ \ell^-) &=&\frac1{4 E_{\tilde{\chi}^0_2}}D(\tilde{\chi}^0_2)\,d\Phi_{\tilde{\chi}^0_2}\,, \end{eqnarray} where the phase-space factors in the laboratory system are given by: \begin{eqnarray} d\Phi_{\ti q} &=& \frac{1}{(2\pi)^2} \frac{E_q} {2||\vec{p}_{\tilde{q}}|\cos\theta_{\tilde{q}}-E_{\tilde{\chi}^0_2}-E_q|}~d\Omega_{\ti q}\,, \\ d\Phi_{\tilde{\chi}^0_2 } &=&\frac1{8(2\pi)^5} \frac{E_{\ell^+}}{||\vec{p}_{\tilde{\chi}^0_2}| \cos \theta_{\ell^+}-E_{\tilde{\chi}^0_1}-E_{\ell^+}-E_{\ell^-} \cos\alpha|}~E_{\ell^-} dE_{\ell^-}d\Omega_{\ell^+}d\Omega_{\ell^-}\,, \end{eqnarray} where $\alpha$ is the opening angle between the two final state leptons. \subsection{Integration limits} When evaluating the phase-space integral at the parton level, kinematical limits need to be determined on the integration variables and these are listed below. The three-body decay phase space of the $\tilde{\chi}^0_2$ has the following limits: \begin{eqnarray} E_{\ell^-}&<&\frac{m_2^2-m_1^2}{2(E_{\tilde{\chi}^0_2}-|\vec{p}_{\tilde{\chi}^0_2}|)}\,, \\ \cos\theta_{\ell^-}&<&\frac{2E_{\tilde{\chi}^0_2}E_{\ell^-}+m_1^2-m_2^2}{2E_{\ell^-}|\vec{p}_{\tilde{\chi}^0_2}|}\,. \end{eqnarray} For a more detailed discussion see \cite{Ellis:2008hq}. \end{appendix} \addcontentsline{toc}{section}{References}

Introduction ============ Cancer continues to dominate the health care landscape.[@b1-phmt-7-079] Although administration of pharmacological agents constitutes the major intervention in this disease space, other "nondrug" therapies constitute a significant component of the psychosocial--economic opportunity with respect to treatment. Much of the interventions with regard to cancer treatment, vis-a-vis both pharmacological and nonpharmacological modalities, focus on controlling the pain associated with this disease. Indeed, cancer pain accounts for considerable costs and may also contribute to noncompliance either as a result of treatment regimens or patients' quality of life.[@b2-phmt-7-079]--[@b4-phmt-7-079] Pediatric cancer contains another layer of complexity in that the patient often refers to the unit of the father, mother, and child, and as such, may affect compliance to therapy, the perception of pain, and quality of life. Other variables including the patient's other relatives and belief structure, provide additional considerations.[@b5-phmt-7-079]--[@b8-phmt-7-079] Various modalities have been used to address pediatric cancer pain. Although pharmacological options have been used to decrease pediatric pain, barriers to successful implementation include parents' reluctance to their child receiving pain medication, inadequate information/education pertaining to physician medication orders, subjective pain perception and severity scores, and fear of addiction or tolerance to pain medication among others.[@b8-phmt-7-079]--[@b10-phmt-7-079] Insurers are eager to adopt interventions for controlling cancer pain (pediatric and adult), since this would facilitate considerable cost saving to the health care industry as a means to increase patient compliance to therapy and disease management as well as to decrease other morbidities.[@b11-phmt-7-079]--[@b14-phmt-7-079] The application of martial arts in health maintenance has been well documented. Both physical (Karate, Tai Kwon Do, etc) and meditative (Tai Chi and Chi\[Qi\] Gung, etc) modalities within the scope of martial arts have been shown to increase flexibility, range of motion, mental focus, and decrease anxiety and stress.[@b15-phmt-7-079]--[@b17-phmt-7-079] Martial arts facilitate control over oneself and include different aspects of metered breathing, guided imagery, and physical as well as mental conditioning to overcome challenges. The positive effects of martial arts in disease states have also been described. Recent studies have shown that rheumatoid arthritis patients who practiced Tai Chi 50 minutes/week for 12 weeks had decreased pain and fatigue when compared with routine care.[@b18-phmt-7-079],[@b19-phmt-7-079] Similarly, patients suffering with osteoarthritis or fibro-myalgia, where there are less effective disease-modifying agents than for rheumatoid arthritis, can also benefit from Tai Chi.[@b20-phmt-7-079]--[@b22-phmt-7-079] In those studies, implementation of distinct Tai Chi regiments, either short form or long form (which pertains to the number of postures per form), has been correlated with positive findings pertaining to improved muscle strength, endurance, reduced stress, psychological well-being, flexibility, and muscle control. Even in pilot studies where there were no observed differences between Tai Chi and control groups over a 4--8-week regimen,[@b23-phmt-7-079] participants generally "preferred Tai Chi to their previous physiotherapeutic intervention, finding it more useful and more enjoyable". Furthermore, the implementation of Tai Chi to other diseases including Parkinson's[@b24-phmt-7-079] and heart diseases[@b25-phmt-7-079] has also been reported. Regarding malignancy, a recent meta-analysis demonstrated positive effects of Tai Chi and Qigong in adult cancer patients in improving their quality of life and decreasing markers of inflammation (C-reactive protein and cortisol).[@b26-phmt-7-079] Although there are reports of the application of martial arts in other disease states with demonstrations of beneficial effects, here we describe the effects of martial arts as an interventional therapeutic modality to reduce pediatric cancer pain. This program is further adaptable to hospital and clinic settings as well as noncancer pediatric disease space, thus providing widespread accessibility toward the control and mitigation of pediatric pain. Methods ======= Sixty-four children with childhood malignancies were enrolled in the martial arts therapy program, which encompassed both the meditative aspects of martial arts and the actual movements and techniques used in that discipline over the course of 12 months. The intervention was administered by specially trained martial arts instructors according to the standard operating procedures of the organization Kids Kicking Cancer ([www.kidskickingcancer.org](http://www.kidskickingcancer.org)) and patient consent was obtained prior to intervention. Each program consists of a 1-hour session, where patients are exposed to a mix of martial arts-based therapeutic interventions, either individually or in group setting. The session begins with identifying each child participant's (presession) pain rating as well as location of pain and cause where applicable (ie, "my pain was 7/10 after having chemotherapy, line/catheter/port placement, and/or injections", etc). Pain scores are recorded using an 11-point (0--10) verbal and/or continuous visual analog scale. The visual analog scale is a psychometric response scale commonly used in questionnaires to measure pain as an instrument for subjective characteristics or attitudes that cannot be directly or objectively measured. The participant begins with a "body scan" in which children lie down and begin to meditate, focusing upon the sensation of their breath. After a few moments (times differ based on the age and attention level of the group), the students are asked to imagine themselves perceiving their bodies beginning from the top of their head and proceed downward through their body relaxing the different muscle group on the voice direction of the "sensei" (martial arts instructor). The "body scan" also helps each participant focus on and become cognizant of each body part with respect to pain/fear/frustration and function and gently allows any unpleasant sensations to flow out of their body. The session continues with karate style martial arts movements modified for each patient according to his or her capabilities. After 30 minutes, participants rehydrate and discuss how they have used their breathing and meditation therapeutic techniques during that week to "bring in the light and blow out the darkness" of their procedures or medical situations. Each child defines his or her "darkness" (ie, "IV placement in the hand vein") and is invited to share with the class how they triumphed during the week using their "light" (via breath/meditation) to push out the darkness. This provides an individualized and collective patient empowerment as well as affirmation of successful pain management (participants are referred to as "powerful martial artists" and that "pain is a message that we do not have to listen to"). Each class incorporates a martial arts "Breath Brake®" -- a modality that incorporates three cycles of "breathing in (metered inhalation with slight lifting of the chest/body)" for 3 seconds, "holding (maintaining)" the breath for 3 seconds, and "breathing out (metered exhalation, where chest/body is slightly lowered and relaxed)" for 3 seconds, each time envisioning "breathing in the light and breathing out the darkness". The remaining 30 minutes include individualized or small group sessions, which focus on different aspects of martial arts movements and exercises, and further internalizing the concepts via physical touch points of parenthetically breaking the pain with punches and kicks as well as allowing the pain to pass through the individual by using the image of the breath to pass through and relax the area in pain. Because children on chemotherapy often have brittle bones, there are never any breaks or sparring involved in the program. However, the children do punch or kick a soft rubber target mitt when able. As the martial arts are focused on imagery, the children are invited to see their target as their disease, pain, anger, or any "darkness" the children choose. The children are asked to image themselves "destroying" their disease or whatever they define as the "darkness" and "see" it shatter with perceived reduction. As the martial arts are "ritually oriented", the children have a special bow and "karate yell" that they are asked to do several times during the class. At Kids Kicking Cancer, the children bow and call out the mantra "Power, Peace, Purpose"™. "Power" represents each child's ability to bring in a powerful martial arts energy. "Peace" represents how a child can blow out the pain, fear and sadness that he/she may face. "Purpose" represents each child's ability to project their triumph and "Teach the world" that pain, stress and adversity can be overcome. Because the pediatric patients face such darkness, their parents, caregivers, nurses, and doctors also become participants and also learn from them how to face down the stressors in their adult lives, further augmenting the positive influence of the program. The session concludes with another round of "power breathing" and post session pain scores. These studies were approved by the participating institutional review boards of Brooklyn Hospital Medical Center, Columbia Presbyterian Medical Center, and Maimonides Medical Center. Statistical analyses -------------------- Patient demographic and descriptive visit data were assessed using mean values, standard deviations, median values, modes, ratios, and proportions. The range of pain scores (overall ≥1 and within categories from 1 to 10) was crossed by total visits and tabulated by whether participant pain reduced at least 1 unit, stayed the same, or increased in intensity immediately after (post) participation session. From our perspective, a minimally clinically important reduction in pain is assigned to any lowering of pain level that children are able to accomplish. Furthermore, although any reduction in pain is clinically worthwhile in children, achieving greater reductions could be considered more important in pain relief than reductions in lower scores. The percentage of change in pain intensity after session was recorded to depict whether pain intensity reduced, stayed the same, or increased pre-/postvisit, where pain intensity reduced by at least 1 unit postintervention with ranges of pain intensity, minimum and maximum values, and inter-quartile ranges. Nonparametric Wilcoxon signed rank test for related (paired) samples was conducted for those which reported at least a pain intensity score of 1 presession with postsession scores. Mean percent change scores were also table reported by prepain score category of pain (on a scale of 1--10 of increasing pain) by visits. Median prepain and postpain intensity change scores were further compared between four age categories using a nonparametric Kruskal--Wallis procedure, followed by pairwise comparisons conducted using a nonparametric Mann--Whitney *U* test. Age groups were created by usinĝ25% cutoffs in the frequency age range. Nonparametric Mann--Whitney *U* test compared prepain and postpain difference scores by sex. Nonparametric Spearman's Rho correlation coefficient was conducted to examine if a possible relationship existed between prepain and postpain intensity scores and age. Statistical significance was considered achieved at *P*≤0.05. All statistical procedures were conducted using SPSS Version 22, IBM Corporation (Armonk, NY, USA). Results ======= Weekly pre- and postpain scale data were measured for 64 participants, 43 males (67.2%) and 21 females (32.8%; [Table 1](#t1-phmt-7-079){ref-type="table"}). Their ages ranged from 3 years to 19 years, with a median of 11 years, a mean of 11.2±3.9 years, and a mode of 12 years, and represented varied ethnicities (Caucasian 31%, African American 28%, Hispanic 33%, and Asian 5%). Pre-and postintervention data were obtained for 223 individual session visits. In all, 48.0% (107/223) of participants reported no pain (zero) prior to their presession, who were excluded from subsequent analysis. Data analyses therefore concentrated on those participants and visits where prepain intensity scores were recorded as ≥1. Sessions occurred weekly; however, participant pain scores were not measured every week, due to holidays and some absences. Mean number of patient participation pre- and post- (same day) measured visits was 1.8±1.6, with a range of visits from one to nine visits. Participants began their session visits ranging in pain scale score from ≥1 to as high as 10. Of 116 individual measured sessions where the participants began with a pain score of at least 1, pain intensity reduced ≥1 unit in 85.3% (99/116) of visits, increased in 6.9% (8/116), and remained the same in 7.8% (9/116; [Figure 1](#f1-phmt-7-079){ref-type="fig"}). [Table 2](#t2-phmt-7-079){ref-type="table"} represents the range of pain intensity scores for participants who had a pain score of 1 or greater prior to their visit (overall ≥1 and within categories from 1 to 10) crossed by total visits and tabulated by whether participant pain reduced at least 1 unit, stayed the same, or increased in intensity after (post) weekly participation session. In total, for all visits 85% of participants demonstrated at least one point reduction in pain, 7% demonstrated an increase, and 8% remained the same. Of these, the majority of individual visits started with pain scale scores of 3 or greater, with the largest pain intensity scores presession being in the range of 7--9 (48.2%; 56/116 visits). Interestingly, the majority (96.3%; 77/80) of individual sessions where participants began with a prepain intensity score of ≥5, reduced in pain intensity following the session; in these participants, only three remained the same and none increased in pain intensity postsession. [Table 3](#t3-phmt-7-079){ref-type="table"} represents the overall mean prepain and postpain scores of session visits for those who began the session with at least a pain score of 1 or higher. The overall mean pain score presession visit (n=116) was 5.95±2.64, which reduced to 3.03±2.45 postsession. The median value presession was 5.00 and postsession was 3.00. The interquartile value of pain intensity reduced from prevalues of 4.00 (25th percentiles) and 8.00 (75th percentiles) to postvalues of 1.00 and 5.00, respectively. The overall median pain score significantly reduced from 5.00 to 3.00 (40% reduction; ≤0.001). [Table 4](#t4-phmt-7-079){ref-type="table"} represents the mean percent change in pain intensity scores after (post) weekly session visit. Overall, where initial pain scores were ≥1 (116 visits), pain was reduced by a mean percent of 24.4%±1.4%. Where initial pain scores were ≥2 (102 visits), pain reduced by a mean percent of 51.7%±41.2%, and where beginning pain score was ≥3 (99 visits), pain reduced by a mean percent of 52.8%±36.0%. Prepain intensity scores that began at 4 or higher experienced mean percent reduction in postpain scores ranging from 43.5% to 67.5%. [Table 5](#t5-phmt-7-079){ref-type="table"} depicts the individual session changes that occurred for all 116 visits where the relationship of each presession to postsession pain score is reported in an individualized manner for each visit. For example, 24 participant sessions began with a prepain score of 7. Of these, four participants reported a reduction of pain to 0, one participant reported a reduction of pain to 2, five participants reported a reduction of pain to 4, seven participants reported a reduction of pain to 5, three participants reported a reduction of pain to 6, and two participants remained the same from their presession score of 7. No participants reported an increase in pain. The bolded values represent any prepain participant's score that increased after treatment. Only eight of 116 total sessions, representing 7% of the study participants, showed an increase from prepain scores, and nine participants remained the same as their preintervention pain scores. Thus, \>85% of those studied demonstrated a reduction from prepain scores after this intervention. When pain intensity scores were further compared between four age categories (3--6 years, 7--10 years, 11--14 years, and 15--19 years; [Table 6](#t6-phmt-7-079){ref-type="table"}), median pain scores reduced 1 unit through each successive older age group, beginning at a median of −1.00 in the lowest age group (3--6 years) and reducing to a median of −4.00 in the oldest age group (15--19 years), *P*≤0.001. Median prepain and postpain difference scores between males (−3.00) and females (−3.00) were identical (data not shown). In addition, all three ethnic groups experienced reductions in their mean prepain and postpain scores intervention; Caucasians reduced their pain scores by almost 3 units (2.99±3.01), African Americans by 2½ units (2.47±3.19), and Hispanics recorded slightly less reductions in pain scores (1.76±2.78); however, the differences were not statistically significant (\>0.05; data not shown). Discussion ========== The application of martial arts-related regimens in disease states is not well understood. Many of the studies that assess the effects of martial arts centralize around movement anomalies, which may be subject to mechanical or inflammatory processes in addition to its well-known effects to improve general psychological well-being.[@b15-phmt-7-079]--[@b20-phmt-7-079],[@b23-phmt-7-079],[@b24-phmt-7-079] Recent data on the effects of biological effects of martial arts have also been described. To this end, Ho et al[@b27-phmt-7-079] reported an increase in blood CD34+ progenitor cells in those engaged in the practice of Tai Chi suggesting its value as an antiaging therapy, while Chyu et al[@b28-phmt-7-079] reported an increase in serum insulin growth factor 1 levels, facilitating promotion and maintenance of muscle mass and neuronal function in overweight/obese premenopausal women who participated in a martial arts exercise intervention.[@b29-phmt-7-079],[@b30-phmt-7-079] The application of martial arts intervention in cancer has also been described. Recent data suggest that such modalities can be valuable as a means to improve physical and psychological quality-of-life parameters in their relation to treatment.[@b26-phmt-7-079] In those studies, the authors conduct a meta-analysis of the literature, which presents reports that martial arts intervention in cancer patients reported 1) less depressive symptoms,[@b31-phmt-7-079] 2) stability of insulin levels,[@b32-phmt-7-079] 3) improvement in self-esteem[@b33-phmt-7-079] and social/family well-being,[@b34-phmt-7-079] 4) strength and flexibility,[@b35-phmt-7-079] 5) cognitive function,[@b36-phmt-7-079]6) survival,[@b34-phmt-7-079] and 7) inflammatory/metabolic markers such as C-reactive protein and cortisol[@b36-phmt-7-079],[@b37-phmt-7-079] compared with control groups among a variety of adult malignancies. Although some reports did not demonstrate any statistically significant changes in select parameters,[@b38-phmt-7-079] there were demonstrable positive trends.[@b37-phmt-7-079] Our studies are the first to assess the effect of martial arts intervention in a pediatric population where the participants predominantly suffered from cancer-related pain. In evaluating pre- to post-test scores, it is important to acknowledge the possibility of "regression to the mean" where the study group will eventually trend to the population mean. However, the regression can occur in either direction. In our case, we have demonstrated that the intervention consistently reduced the perception of pain in an ethnically diverse pediatric population and that the major pain reduction was observed when the baseline preintervention scores 1) were \>3, 2) were most robust when pain scores were \>5, 3) positively reduced preintervention pain scores in \>85% of the participants, and 4) had better pain score reductions with increased age. Furthermore, we have shown an overall reduction in reported pain scores of ≥40% as a result of this program. It is interesting to note that this program is beneficial in the pediatric population among various age groups. Pain in the pediatric population is unique as it encompasses various stages of development and thus may translate pain registration and reporting differently as children mature. As many as 40% of children and adolescents complain of pain that occurs at least once weekly, and chronic pain affects at least 15%--20% of children. Furthermore, just as chronic pain is more prevalent in women than men, girls report more pain than boys.[@b39-phmt-7-079] Each year, 1.5 million children, irrespective of malignancy, undergo surgical procedures, and many receive inadequate pain relief; in 20% of cases, the pain can become chronic thus creating a lifelong medical issue.[@b40-phmt-7-079],[@b41-phmt-7-079] On a larger scale, failure to intervene early in children's pain may lead to impairment in functioning and disruption in families. Unaddressed pain heightens anxiety and fear, which, in turn, can increase perception of pain.[@b42-phmt-7-079] Perception of pain in pediatric medicine is complex and entails physiological, psychological, behavioral, and developmental factors. Although health care professionals often prefer practical methods, which reliably track the child's pain experience and pain control over time, researchers tend to focus on tools which are proven for reliability with different observers.[@b43-phmt-7-079] To this end, three main methods are currently used to measure pain intensity: self-report, behavioral, and physiological measures. Self-report measures are optimal and the most valid. Both verbal and nonverbal reports require a certain level of cognitive and language development for the child to understand and give reliable responses. Children's capability to describe pain increases with age and experience and changes throughout their developmental stages.[@b44-phmt-7-079] An additional variable often promulgated in the medical milieu is that children do not feel pain the way adults do[@b45-phmt-7-079] and that pain in adolescent patients is often underestimated and under treated[@b46-phmt-7-079] since adolescents tend to minimize or deny pain, especially in front of parents and friends, and some adolescents regress in behavior under stress.[@b45-phmt-7-079] In light of this, it is thus quite remarkable that this program is able to engender significant pain reduction in a wide range of age groups (3--19 years) and that older patients, in contrast to conventional thinking, actually demonstrated a more robust reduction to pain, thus providing applicability to a large proportion of the childhood population in pain. Although there have been reports of other alternative modalities used in childhood cancers including faith healing, massage, vitamins, chiropractic, and relaxation,[@b47-phmt-7-079] many of these interventions were instituted to cope with disease symptoms and side effects of medications and were often incorporated when families were dissatisfied with conventional medical care.[@b48-phmt-7-079] A unique strength of this program and the therapeutic intervention it provides the concept of "teaching the world" a psychosocial projection of one's own power to enable another in pain to surmount such, thus facilitating a group-think mentality, which has been reported as a successful value-added component in other disease states,[@b49-phmt-7-079] where those in pain help and empower each other to decrease the registration of pain. The distillation of this intervention "Power, Peace, Purpose"™ (the organization's tag line/brand) thus becomes a mantra for the patient to incorporate these concepts at any time, during and after the intervention, where pain in any definition needs to be addressed. Furthermore, the program has been incorporated into the medical landscape of numerous children's hospitals in the US, as well as other countries, and is thus well received in the allopathic domain as it pertains to pediatric cancer management. There are, however, limitations to our observations that would benefit from additional study. The degree to which intervention should be instituted and the frequency, which would be considered ideal, require further study. These studies were derived from sequential categorization of pain scores from participants in the program collected as they participated, without any a priori power analysis or detailed criteria for patients' involvement. A successful intervention would likely also be bolstered by the reinforcement of participating family members, treating clinicians, friends, and other children with cancer. Whether these collective components or individual isolated variables would contribute to a participant being a more likely "responder" to the intervention, as we have observed in \~85% of our participants, or whether the physical (hard) or meditative (soft) martial arts intervention approaches are more suitable for one participant over another is unknown. These comprise opportunities identified from our current studies in that such variables were not accounted for and may be valuable to incorporate into a streamlined intervention for more optimal pain reduction. In addition, the duration of intervention may be improved by initially increasing the frequency (per week or month) with subsequent maintenance as the patient becomes well versed with the modalities and can integrate them into care in a nonclass setting. Furthermore, the short-term duration of interventions in most of the participants in the current studies (less than four sessions in most) and the lack of consecutive visits may reduce the long-term effects of Qigong/Tai Chi/martial arts on cancer pain; thus, the optimal duration of such intervention has yet to be defined. In alignment with other alternatives, including Nia and Hypnosis, which are incorporated into diseases including malignancy,[@b50-phmt-7-079],[@b51-phmt-7-079] it would be ideal to maximize such martial arts-based therapeutic interventions since they are less costly, yet empowering, which synergize and can likely augment conventional allopathic therapy toward the benefit of the patient. Conclusion ========== Martial arts-based therapeutic interventions can be quite effective in mitigating pain in childhood cancer. Such interventions can be empowering and well received by the patient and his or her family, thus further facilitating its beneficial effects. Additional studies to determine duration and frequency of participation in addition to cost reduction and objective reporting (laboratory, inflammatory, and metabolic markers) can help streamline effective integration of the martial arts into pediatric oncology and other medical practices. The abstract of this paper was presented at the American Society of Hematology 55th Annual Meeting December 7--10, 2013, New Orleans, LA, as a poster presentation with interim findings. The poster's abstract was published in *Blood*, November 2013; 122: 5588. **Disclosure** The authors report no conflicts of interest in this work. {#f1-phmt-7-079} ###### Participant demographics Participants n=64 ------------------------ --------------- Sex  Male 43/64 (67.2%)  Female 21/64 (32.8%) Ethnicity  Caucasian 20/64 (31.3%)  African American 18/64 (28.1%)  Hispanic 21/64 (32.8%)  Asian 3/64 (4.7%) Age (years)  Mean ±SD age 11.2±3.9  Median age 11.0  Mode age 12.0  Age range 3--19 Session visits  Mean ±SD no of visits 1.8±1.6  Median no of visits 1.00  Range visits 1--9 ###### Pain reduced at least 1 unit on pain scale immediately after (post) weekly participation session, n (%) Prepain score before session Total no of visits Yes No Stayed the same ------------------------------ -------------------- ---------- -------- ----------------- 1 12 2 (17) 4 (33) 6 (50) 2 3 2 (67) 1 (33) 0 3 9 8 (89) 1 (11) 0 4 12 10 (83) 2 (17) 0 5 9 9 (100) 0 0 6 9 9 (100) 0 0 7 24 22 (92) 0 2 (8) 8 18 17 (94) 0 1 (6) 9 14 14 (100) 0 0 10 6 6 (100) 0 0 Total (\>1) 116 99 (85) 8 (7) 9 (8) ###### Participant visits where pain reduced at least 1 unit Pain before Pain after ------------------------------------------------- ------------- ------------ N (visits) 116 116 Mean 5.95 3.03 Median 5.00 3.00 Mode 7.00 0.00 Standard deviation 2.64 2.45 Difference between minimum and maximum in range 9 8  Minimum 1 0  Maximum 10 8 Percentiles 25 4.00 1.00 Percentiles 75 8.00 5.00 ###### Mean percent change in pain scale scores immediately after (post) weekly participation session Prepain score before session Total number of visits Mean ± SD ------------------------------ ------------------------ -------------- ≥1 116 −24.4%±1.4% ≥2 102 −51.7%±41.2% ≥3 99 −52.8%±36.0% 1 12 +2.1%±3.5% 2 3 −16.7%±1.4% 3 9 −44.4%±50.0% 4 12 −50.0%±56.4% 5 9 −67.5%±32.0% 6 9 −57.4%±32.4% 7 24 −43.5%±30.8% 8 18 −57.6%±28.8% 9 14 −57.1%±28.5% 10 6 −58.0%±30.3% **Notes:** (−) sign indicates the mean percent reduction in pain; (+) sign indicates the mean percent increase in pain. ###### Change from highest presession to lowest postsession pain scores Lowest postsession pain score -------------------------------------------------------------------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ----------- **Number of children with increased pain scores (bolded numbers)** **Highest presession pain score** **0** **1** **2** **3** **4** **5** **6** **7** **8** **Total** 1 2 6 **0** **0** **0** **0** **0** **1** **3** 12 2 2 0 0 **0** **0** **1** **0** **0** **0** 3 3 2 2 4 0 **0** **1** **0** **0** **0** 9 4 5 0 3 2 0 **1** **0** **1** **0** 12 5 3 1 2 2 1 0 **0** **0** **0** 9 6 2 1 1 2 1 2 0 **0** **0** 9 7 4 0 1 2 5 7 3 2 **0** 24 8 3 0 4 3 2 2 3 0 1 18 9 2 2 1 0 2 2 3 2 0 14 10 0 1 3 0 0 0 0 1 1 6 Total 25 13 19 11 11 16 9 7 5 116 ###### Reduction in pain scores by age groups Age group (years) Median Mean (standard deviation) 95% CI ------------------- -------- --------------------------- -------------- 3--6 −1.00 −1.83 (1.9) −3.9 to 0.2 7--10 −2.00 −2.84 (3.4) −4.1 to −1.6 11--14 −3.00 −2.94 (2.0) −3.6 to −2.3 15--19 −4.00 −4.10 (3.5) −5.4 to −2.8 Overall −2.00 −3.19 (3.0) −3.8 to −2.6

\section*{Hintergrund} \subsection*{Der Weg zum Mars} Der Mars ist wohl der Planet im Sonnensystem, der die meisten Fantasien weckt. Spätestens seitdem Giovanni Schiaparelli 1877 auf dem Mars angeblich Kanäle entdeckt hatte \autocite[Abb.~\ref{fig:schiaparelli}][]{meyer_meyers_1885,schiaparelli_decouvertes_1882,schiaparelli_observations_1897}, gilt er als der Himmelskörper, der in puncto außerirdisches Leben am meisten fasziniert. Besondere Berühmtheit erlangte das Buch „War of Worlds (Krieg der Welten)“ von H. G. Wells aus dem Jahre 1898. Ein darauf basierendes Hörspiel erzeugte während seiner Radioausstrahlung 1938 Irritationen unter der Bevölkerung der USA, die die Sendung als einen Bericht missverstand \autocite[][]{chilton_war_2016}. \begin{figure}[!ht] \centering \includegraphics[width=10cm]{Karte_Mars_Schiaparelli_MKL1888.jpg} \caption{Karte der Marsoberfläche von 1877 nach Beobachtungen von Giovanni Schiaparelli (Quelle: Meyers Konversations-Lexikon, 1888, \url{https://commons.wikimedia.org/wiki/File:Karte_Mars_Schiaparelli_MKL1888.png}).} \label{fig:schiaparelli} \end{figure} Jenseits dieser fantastischen Erzählungen versuchen Wissenschaftler bereits seit dem Beginn der Raumfahrt, den Mars genauer zu erkunden. Die erste erfolgreiche Mission zum Mars war die Sonde Mariner~4, die 1965 auf ihrem Vorbeiflug die ersten Aufnahmen der Oberfläche aus unmittelbarer Nähe zur Erde funkte\footnote{\url{https://nssdc.gsfc.nasa.gov/nmc/spacecraft/display.action?id=1964-077A}}. Als erster künstlicher Marssatellit schwenkte Mariner~9 1971 in den Orbit ein\footnote{\url{https://nssdc.gsfc.nasa.gov/nmc/spacecraft/display.action?id=1971-051A}} und erstellte die erste detaillierte Marskarte (Abb.~\ref{fig:mariner9}). \begin{figure}[!ht] \centering \includegraphics[width=10.0cm]{Mariner9_Map_1280.jpg} \caption{Karte der Marsoberfläche erstellt aus Fotografien der Marssonde Mariner~9 (Bild: USRA -- Universities Space and Research Association, Lunar and Planetary Catalog, \url{https://www.lpi.usra.edu/resources/mars_maps/mariner9/index.html}).} \label{fig:mariner9} \end{figure} Die ersten Landungen auf dem Mars gelangen 1976 mit \mbox{Viking 1\footnote{\url{https://nssdc.gsfc.nasa.gov/nmc/spacecraft/display.action?id=1975-075C}}} und Viking 2\footnote{\url{https://nssdc.gsfc.nasa.gov/nmc/spacecraft/display.action?id=1975-083C}}. Sie definierten bereits die grundlegenden Technologien für weiche Landungen auf dem Mars, die unter anderem aus Fallschirmen und Bremsraketen bestanden. Dass Flüge zum Mars generell eine große technologische Herausforderung sind, erkennt man daran, dass nur ca. 50\% aller Marsmissionen erfolgreich sind \autocite[][]{zundel_mars-mission_2016}. \subsection*{Landung auf Erde und Mars} Wenn Raumschiffe wie früher Apollo oder noch heute die Sojus-Kapseln auf der Erde landen, werden sie durch Fallschirme auf Geschwindigkeiten abgebremst, die eine nahezu weiche Landung ermöglichen. Der Aufprall der Apollo-Raumschiffe wurde durch eine Wasserung im Ozean abgemildert, während Düsen an den Sojus-Landemodulen kurz vor dem Aufsetzen zünden und die Restgeschwindigkeit auf akzeptable Werte reduzieren\footnote{\url{https://www.esa.int/kids/de/lernen/Leben_im_Weltraum/Astronauten/Achterbahnfahrt_zurueck_zur_Erde}}. \begin{figure}[!ht] \centering \includegraphics[width=7cm]{Apollo_15_descends_to_splashdown.jpg} \caption{Die Kapsel der Apollo 15-Mission wassert mit nur zwei intakten Fallschirmen (Bild: NASA)} \label{fig:apollo} \end{figure} Landungen auf dem Mars gestalten sich jedoch anders. Seit Viking~1 werden die Sonden zwar durch Fallschirme abgebremst. Auf den letzten Kilometern wird die weiche Landung jedoch meistens durch Düsen bewerkstelligt. Auf diese Weise konnten das mobile Labor Curiosity und die geologische Messeinheit InSight erfolgreich auf dem Mars aufsetzen. \begin{figure}[!ht] \centering \includegraphics[width=7cm]{Expedition57_SoyuzMS-09_Landing.jpg} \caption{Der Fallschirm nach der Landung des Sojus MS-09-Raumschiffs, mit dem Alexander Gerst am 20. Dezember 2018 auf die Erde zurückkehrte (Bild: NASA, \url{https://www.flickr.com/photos/nasahqphoto/45477952845}, \href{https://creativecommons.org/licenses/by-nc-nd/2.0/legalcode}{CC BY-NC-ND 2.0}).} \label{fig:sojus} \end{figure} Auch der gemeinsam von ESA\footnote{European Space Agency, Europa} und Roskosmos\footnote{Russland} entwickelte ExoMars-Rover wird 2021 mit einer Landeeinheit durch Bremsraketen auf dem Mars landen (siehe Abb.~\ref{fig:sequenz}). Bei zukünftigen astronautischen Missionen wird es nicht anders sein. \begin{figure}[!ht] \centering \includegraphics[width=11cm]{593484main_pia14839_Curiosity.jpg} \caption{Künstlerische Darstellung der letzten Sequenz der Landung des Mars-Roboters „Curiosity“. Das Bremsmanöver wird nicht mit Fallschirmen sondern mit Triebwerken durchgeführt (Bild: NASA/JPL-Caltech).} \label{fig:curiosity} \end{figure} Doch warum ist das so? Die einfache Antwort ist: Die Atmosphäre ist zu dünn. Fallschirme alleine können die Sinkgeschwindigkeit nicht ausreichend reduzieren. Somit ist der Luftwiderstand zu gering, um genügend Reibung zu erzeugen. Die vom Fallschirm erzeugte Bremskraft reicht nicht aus, um die von der Gravitation ausgeübte Beschleunigung ausreichend zu kompensieren. Die Details werden im nächsten Abschnitt erläutert. \begin{figure}[!ht] \centering \includegraphics[width=11cm]{ExoMars_Rover_LandingSequence.jpg} \caption{Schematische Darstellung der Landesquenz des ExoMars-Rovers. Mehrere große Fallschirme bremsen die Landeeinheit ab. Jedoch reicht die Reduktion der Geschwindigkeit für eine weiche Landung nicht aus. Die Sonde würde zerstört. Daher werden in der letzten Phase Triebwerke eingesetzt, so dass die Landeeinheit schließlich sanft aufsetzt (Bild: ESA).} \label{fig:sequenz} \end{figure} \subsection*{Physik des Fallschirms} Der Fall mit Fallschirm ist eine Modifikation des freien Falls mit einer zusätzlichen zeitabhängigen Randbedingung durch ein bremsendes Medium. Dadurch entsteht eine Kraft, die der Beschleunigung entgegen wirkt. Im Vakuum existiert ein solches Medium jedoch nicht. Dann entspricht dies dem idealisierten freien Fall. Um die nachfolgenden Betrachtungen zu erleichtern, nehmen wir an, dass der Fall stets senkrecht zur Oberfläche, d. h. parallel zur Gravitationswirkung verläuft. \subsubsection*{Freier Fall ohne Luftwiderstand} Gemäß des Newtonschen Gravitationsgesetzes übt die Gravitationskraft $F_g$ auf einen Körper der Masse $m$ eine Beschleunigung $a$ aus. Daher kann man schreiben: \begin{align*} F_g &= m\cdot a \\[5pt] \Leftrightarrow G\cdot \frac{M\cdot m}{r^2} &= m\cdot a\\[5pt] \Leftrightarrow a &=G\cdot \frac{M}{r^2} \equiv g \end{align*} \medskip Dieser Term stellt eine allgemeine Bewegungsgleichung für eine gleichmäßig beschleunigte Bewegung dar. Hierbei ist $r$ der Abstand zwischen den Massenmittelpunkten. Daraus folgen per Integration die bekannten Gleichungen für die Geschwindigkeit und den Ort der beschleunigten Masse zum Zeitpunkt $t$. \begin{equation} v(t) = v_0 + G\cdot \frac{M}{r^2}\cdot t = v_0 + g\cdot t \end{equation} \begin{equation} s(t) = s_0 + v_0\cdot t + \frac{1}{2}\cdot G\cdot \frac{M}{r^2}\cdot t^2 = s_0 + v_0\cdot t + \frac{1}{2}\cdot g\cdot t^2 \end{equation} \medskip Dabei stellen $s_0$ den Ort und $v_0$ die Geschwindigkeit zu Beginn der Beschleunigung dar. \subsubsection*{Freier Fall mit Luftwiderstand} \label{s:FFmL} Die Kraft des Luftwiderstands\footnote{Herleitung unter: \url{https://www.leifiphysik.de/mechanik/reibung-und-fortbewegung/luftreibung}} wirkt in umgekehrter Richtung der Gravitationswirkung und beträgt: \begin{equation} F_w = c_w \cdot \frac{1}{2} \cdot v^2 \cdot \varrho \cdot A \end{equation} \medskip Diese Kraft\footnote{Bei laminarer Strömung folgt der Luftwiderstand aus dem Stokesschen Gesetz.} wird bestimmt durch Größen, die das Medium und das Objekt im Luftstrom charakterisieren. Auf der einen Seite ist die Kraft proportional zur Dichte des Mediums, auf der anderen Seite zur Querschnittsfläche, auf die die Kraft wirkt, sowie zum Quadrat der Strömungsgeschwindigkeit. Der Strömungswiderstandskoeffizient bzw. Widerstandsbeiwert $c_w$ ist eine dimensionslose Proportionalitätskonstante, die den Einfluss der Form des Objekts auf die Strömung und damit auf den Luftwiderstand beschreibt. Umgangssprachlich beziffert sie die Windschlüpfrigkeit. Sie kann durch kontrollierte Experimente z.~B. im Windkanal experimentell ermittelt werden. \medskip Die Dichte hängt zudem vom atmosphärischen Druck $p$ und der Temperatur $T$ ab. $R_S$ ist die spezifische Gaskonstante. \begin{equation} \varrho = \frac{p}{R_S \cdot T} \end{equation} \medskip Die Newtonsche Kraftgleichung lautet demnach: \begin{align} m\cdot a &= F_g - F_w \\[5pt] \Leftrightarrow m\cdot a &= m\cdot g - c_w \cdot \frac{1}{2} \cdot v^2 \cdot \varrho \cdot A \end{align} \medskip Aus der Erfahrung ist bekannt, dass der Fallschirmspringer nach dem Öffnen des Fallschirms eine konstante Endgeschwindigkeit $v_E$ erreicht -- im Gegensatz zur Situation des freien Falls im Vakuum. Wenn $v_E=\mathsf{const.}$, folgt $a=0$. Damit erhalten wir die Bestimmungsgleichung der Endgeschwindigkeit eines Fallschirmsprungs. \begin{align} m\cdot a = 0 &= m\cdot g - c_w \cdot \frac{1}{2} \cdot v^2 \cdot \varrho \cdot A\\[5pt] \Leftrightarrow m\cdot g &= c_w \cdot \frac{1}{2} \cdot v^2 \cdot \varrho \cdot A\\[5pt] \Leftrightarrow v_E &= \sqrt{\frac{2\cdot m\cdot g}{c_w\cdot \varrho\cdot A}}\label{e:vE} \end{align} \medskip Anders als beim freien Fall im Vakuum ist die Geschwindigkeit nicht unabhängig von der Masse des fallenden Gegenstands. \medskip Aus der Bewegungsgleichung folgen wiederum die Geschwindigkeits- und Ortsfunktionen. Setzen wir als Anfangsbedingungen $v_0=0$ und $s_0=0$, wird die Lösung übersichtlicher, und es folgt\footnote{Das zu lösende Integral enthält eine Funktion der Form $\frac{1}{1-x^2}$, deren Stammfunktion artanh ist.}: \begin{equation} v(t) = v_E \cdot \tanh\left(\frac{g\cdot t}{v_E}\right) \end{equation} \begin{equation} s(t) = \frac{v_E^2}{g}\cdot\ln\left(\cosh\left(\frac{g\cdot t}{v_e}\right)\right) \end{equation} \medskip Auch hier erkennt man: \begin{equation} \lim_{t\rightarrow\infty} v(t) = v_E \end{equation} \subsection*{Fallschirme für Marsmissionen} \label{s:fallschirme} Um Sonden in der Marsatmosphäre abzubremsen, werden typischerweise mehrere verschiedene Fallschirme benutzt. So setzt man auch für die Landung des Rovers der ExoMars-Mission Schirme für die unterschiedlichen Geschwindigkeitsbereiche ein \autocite[][siehe auch Abb.~\ref{fig:sequenz}]{esa_exomars_2018}. \medskip Beim Eintritt in die Atmosphäre bremst das Landemodul zunächst durch die Reibung. In der nächsten Phase werden spezielle Überschall-Fallschirme entfaltet. Sie haben bei diesen Geschwindigkeiten Widerstandsbeiwerte von etwa $c_w=0,4$ \autocite[][]{takayanagi_development_2016}. \medskip In der letzten Phase bremsen besonders große Fallschirme die Landemodule auf geringe Geschwindigkeiten ab. Im Fall des ExoMars-Landemoduls kommt ein Fallschirm mit einem Durchmesser von \unit[35]{m} zum Einsatz \autocite[][]{esa_first_2018}. Ein typischer Widerstandsbeiwert für solche Fallschirme ist $c_w=1,28$ \autocite[][S.~47]{hillje_entry_1967}. \clearpage \section*{Vorbereitung} Zur Einstimmung sollten sich die Schülerinnen und Schüler die nachfolgenden Videos ansehen. Sie zeigen Animationen von Landungen zweier Marssonden. \bigskip Curiosity has landed (Dauer: 2:30)\\ \url{https://youtu.be/N9hXqzkH7YA} \bigskip InSight: Landing on Mars (Dauer: 3:18)\\ \url{https://youtu.be/C0lwFLPiZEE} \bigskip Falls Sie die Eigenschaften des Mars nicht selber einführen, lassen Sie die Schülerinnen und Schüler die grundlegenden Charakteristika (z.~B.~Größe, Masse, Entfernung, Atmosphäre) selbst recherchieren. Hierzu bieten sich Artikel aus Wikipedia an. Nach dieser Einleitung lassen Sie die Schülerinnen und Schüler folgende Fragen beantworten. Ermuntern Sie sie zu Diskussionen. \begin{itemize} \item Warum hilft ein Fallschirm, einen Absprung aus einem Flugzeug unbeschadet zu überleben? \item Wie entsteht die Bremswirkung eines Fallschirms? \item Warum funktioniert ein Fallschirm auf dem Mond nicht? \item Was muss man beim Einsatz eines Fallschirms auf dem Mars beachten, damit er gut funktioniert? \end{itemize} \section*{Aufgaben} Die Schülerinnen und Schüler berechnen die Endgeschwindigkeiten für eine Landung auf dem Mars bzw. auf der Erde. Aus dem Vergleich wird ersichtlich, dass diese auf dem Mars für eine weiche Landung zu hoch ist. Bei den Berechnungen werden folgende vereinfachende Annahmen gemacht. \begin{itemize} \item Senkrechter Fall \item Auftrieb vernachlässigt \item Der Fallschirm hat keine Masse. \item Konstante Gravitationsbeschleunigung während des Falls \item Homogene Dichte der Atmosphäre während des Falls \item Konstante atmosphärische Temperatur \item Reine CO$_2$-Atmosphäre auf dem Mars \item Ideales Gas \end{itemize} \newpage \begin{table}[!ht] \centering \renewcommand{\arraystretch}{1.2} \caption{Wichtige Größen zur Berechnung der Aufgaben.} \label{tab:values} \begin{tabular}{lcc} \hline \multicolumn{1}{c}{Größe} & Formelzeichen & Wert \\ \hline Masse der ExoMars\footnote-Landeeinheit & $m$ & \unit[1140]{kg}\\ Schwerebeschleunigung Mars & $g_M$ & \unit[3,71]{$\frac{\mathsf{m}}{\mathsf{s}^2}$}\\ Schwerebeschleunigung Erde & $g_E$ & \unit[9,81]{$\frac{\mathsf{m}}{\mathsf{s}^2}$}\\ Widerstandsbeiwert des Fallschirms & $c_w$ & 1,28 \\ Querschnittsfläche des Fallschirms & $A$ & \unit[960]{$\mathsf{m}^2$}\\ Atmosphärischer Druck Mars & $p_M$ & \unit[636]{Pa}\\ Atmosphärischer Druck Erde & $p_E$ & \unit[101325]{Pa}\\ Dichte Luft bei Normaldruck & $\varrho_L$ & \unit[1,3]{$\frac{\mathsf{kg}}{\mathsf{m}^3}$}\\ Mittlere Temperatur Mars & $T_M$ & \unit[280]{K} \\ Spezielle Gaskonstante CO$_2$ & $R_{S,\mathsf{CO}_2}$ & \unit[188,9]{$\frac{\mathsf{J}}{\mathsf{kg}\cdot\mathsf{K}}$}\\ \hline \end{tabular} \end{table} \footnotetext{ExoMars ist ein Marserkundungsprogramm von ESA (Europa) und Roskosmos (Russland).} \subsection*{Endgeschwindigkeit eines Fallschirmsprungs} Beim freien Fall im Vakuum beschleunigt ein Objekt gleichmäßig. Bei einem Fall mit einem Fallschirm innerhalb einer Atmosphäre strebt die Fallgeschwindigkeit jedoch einem konstanten Wert zu. \medskip Leite aus dem Prinzip \begin{displaymath} F = m\cdot a \end{displaymath} \medskip die Gleichung für die Berechnung der Endgeschwindigkeit eines Falls am Fallschirm $v_E$ her. \begin{displaymath} v_E = \sqrt{\frac{2\cdot m\cdot g}{c_w\cdot \varrho\cdot A}} \end{displaymath} \medskip Die wirkende Kraft $F$ setzt sich aus der Gravitationskraft \begin{displaymath} F_g=m\cdot g \end{displaymath} \medskip und der Kraft des Luftwiderstands \begin{displaymath} F_w=c_w\cdot \frac{1}{2}\cdot v^2\cdot\varrho\cdot A \end{displaymath} \medskip zusammen. Beide Kräfte wirken in entgegengesetzte Richtungen. \subsection*{Landung am Fallschirm auf dem Mars} Nutze die Daten aus Tab.~\ref{tab:values}, um die Endgeschwindigkeit des ExoMars-Rovers bei der Landung zu berechnen, die lediglich durch das Bremsen mit einem Fallschirm der Querschnittsfläche $A$ erzielt wird. Drücke das Ergebnis in den Einheiten m/s und km/h aus. \medskip Nimm zur Berechnung der Dichte der Marsatmosphäre an, dass diese ausschließlich aus Kohlenstoffdioxid (CO$_2$) besteht und die in der Tabelle angegebenen Eigenschaften hat. Der Zusammenhang zwischen Dichte und atmosphärischem Druck folgt aus der allgemeinen Gasgleichung für ideale Gase. \begin{displaymath} \varrho = \frac{p}{R_S \cdot T} \end{displaymath} \subsection*{Landung am Fallschirm auf der Erde} Ermittle die Endgeschwindigkeit für ein Bremsmanöver mit Fallschirm auf der Erde. \subsection*{Diskussion} Wie in Abb.~\ref{fig:sequenz} gezeigt und auf Seite~\pageref{s:fallschirme} beschrieben, werden Sonden, die auf dem Mars landen, nacheinander mit verschiedenen Fallschirmen abgebremst. Erkläre, warum es für die Berechnung der Endgeschwindigkeit genügt, lediglich den letzten Fallschirm in dieser Sequenz zu betrachten. \medskip Erläutere den Unterschied der beiden berechneten Fälle. Beziehe dich hierbei auf die Eigenschaften der beiden Planeten und ihren Atmosphären. \medskip Was ist für die unterschiedlichen Geschwindigkeiten ausschlaggebend? \medskip Würde der ExoMars-Rover die Landung auf diese Weise überstehen? Überlege, was mit einem Auto geschähe, das mit derselben Geschwindigkeit gegen eine Wand fährt. \medskip Auf welche Weise verringert man die Sinkgeschwindigkeit, um eine weiche Landung auf dem Mars zu gewährleisten? \subsection*{Dimension eines idealen Fallschirms} Auf welche Querschnittsfläche bzw. welchen Durchmesser müsste man den Fallschirm auf dem Mars verändern, damit die Endgeschwindigkeit so groß wie auf der Erde ist? Ist das realistisch? \clearpage \section*{Lösungen} \subsection*{Endgeschwindigkeit eines Fallschirmsprungs} Die Herleitung an Stelle der Gl.~\ref{e:vE} auf Seite~\pageref{s:FFmL} erläutert. \subsection*{Landung am Fallschirm auf dem Mars} Man kann entweder die Funktion der Dichte $\varrho(p,T)$ in die Bestimmungsgleichung von $v_E$ einsetzen oder die Dichte auf dem Mars ausrechnen. \begin{displaymath} \varrho_M = \frac{p_M}{R_S \cdot T_M}=\frac{\unit[636]{Pa}}{\unit[188,9]{\frac{\mathsf{J}}{\mathsf{kg}\cdot\mathsf{K}}}\cdot \unit[280]{K}}=\unit[0,012]{\frac{\mathsf{kg}}{\mathsf{m}^3}} \end{displaymath} \begin{align*} v_{E,M} &= \sqrt{\frac{2\cdot m\cdot g_M}{c_w\cdot \varrho_M\cdot A}}=\sqrt{\frac{2\cdot m\cdot g_M\cdot R_{S,\mathsf{CO}_2}\cdot T_M}{c_w\cdot p_M\cdot A}}\\[5pt] &= \sqrt{\frac{2\cdot \unit[1140]{kg}\cdot\unit[3,71]{\frac{\mathsf{m}}{\mathsf{s}^2}}\cdot\unit[188,9]{\frac{\mathsf{J}}{\mathsf{kg}\cdot\mathsf{K}}}\cdot\unit[280]{K}}{1,28\cdot \unit[636]{Pa}\cdot\unit[960]{m^2}}}\\[5pt] &= \unit[23,93]{\frac{m}{s}} = \unit[86,1]{\frac{km}{h}} \end{align*} \subsection*{Landung am Fallschirm auf der Erde} \begin{align*} v_{E,E} &= \sqrt{\frac{2\cdot m\cdot g_E}{c_w\cdot \varrho_L\cdot A}}= \sqrt{\frac{2\cdot \unit[1140]{kg}\cdot\unit[9,81]{\frac{\mathsf{m}}{\mathsf{s}^2}}}{1,28\cdot\unit[1,3]{\frac{\mathsf{kg}}{\mathsf{m}^3}} \cdot\unit[960]{m^2}}}\\[5pt] &= \unit[3,74]{\frac{m}{s}} = \unit[13,5]{\frac{km}{h}} \end{align*} \subsection*{Diskussion} Die Endgeschwindigkeit hängt nicht von der Anfangsgeschwindigkeit ab, vorausgesetzt der Bremsvorgang dauert lange genug. Somit sind die Geschwindigkeiten, die durch die vorgeschalteten Fallschirme erzielt werden, für die Berechnung irrelevant. \medskip Die Endgeschwindigkeit ist auf dem Mars mehr als 6 mal so hoch wie auf der Erde. Obwohl die gravitative Anziehung auf dem Mars erheblich geringer als auf der Erde ist, fällt eine Sonde wie der ExoMars-Rover dort schneller zu Boden. \medskip Letztendlich ist die Atmosphäre auf dem Mars zu dünn, als dass sie einen ausreichend großen Widerstand auf den Fallschirm ausüben könnte. Die Erdatmosphäre ist bedeutend dichter. \medskip Der ExoMars-Rover würde den Aufprall mit einer Geschwindigkeit von \unit[86]{km/h} sicher nicht überstehen. \medskip Aus diesem Grund benötigt man auf dem Mars Triebwerke, um die Geschwindigkeit bei der Landung ausreichend zu verringern. \subsection*{Dimension eines idealen Fallschirms} \begin{displaymath} v_E = \sqrt{\frac{2\cdot m\cdot g}{c_w\cdot \varrho\cdot A}} \Leftrightarrow A = \frac{2\cdot m\cdot g}{c_w\cdot \varrho\cdot v_E^2} \end{displaymath} \medskip Um nun die Querschnittsfläche des Fallschirms zu berechnen, setzen wir die Eigenschaften des Mars ein und nehmen als Endgeschwindigkeit $v_{E,E}$. \begin{align*} A_{id} &= \frac{2\cdot m\cdot g_M}{c_w\cdot \varrho_M}\cdot\frac{1}{v_{E.E}^2} = \frac{2\cdot m\cdot g_M}{c_w\cdot \varrho_M}\cdot\frac{c_w\cdot \varrho_L\cdot A}{2\cdot m\cdot g_E} = \frac{\varrho_L\cdot g_M}{\varrho_M\cdot g_E}\cdot A\\[5pt] &=\frac{\unit[1,3]{\frac{kg}{m^3}} \cdot \unit[3,71]{\frac{m}{s^2}}}{\unit[0,012]{\frac{kg}{m^3}} \cdot \unit[9,81]{\frac{m}{s^2}}}\cdot A = 40,89\cdot A = \unit[39251]{m^2} \end{align*} \medskip Der Durchmesser eines runden Fallschirms mit dieser Querschnittsfläche beträgt: \begin{displaymath} d=2\cdot\sqrt{\frac{A}{\pi}}=\unit[224]{m} \end{displaymath} \medskip Das ist völlig unrealistisch. Nicht nur kann man solch große Fallschirme nicht bauen und testen. Auch besitzen sie eine große Masse, die das Gesamtgewicht des Landemoduls enorm anwachsen ließe. \clearpage \section*{Danksagung} Der Autor bedankt sich bei den Lehrern Matthias Penselin, Florian Seitz und Martin Wetz für ihre wertvollen Hinweise, Kommentare und Änderungsvorschläge, die in die Erstellung dieses Materials eingeflossen sind. Weiterer Dank gilt Herrn Dr. Volker Kratzenberg-Annies für seine gewissenhafte Durchsicht. \printbibliography \clearpage \ \vfill \medskip Diese Unterrichtsmaterialien sind im Rahmen des Projekts {\em Raum für Bildung} am Haus der Astronomie in Heidelberg entstanden. Weitere Materialien des Projekts finden Sie unter: \begin{center} \href{http://www.haus-der-astronomie.de/raum-fuer-bildung}{http://www.haus-der-astronomie.de/raum-fuer-bildung} und \href{http://www.dlr.de/next}{http://www.dlr.de/next} \end{center} Das Projekt findet in Kooperation mit dem Deutschen Zentrum für Luft- und Raumfahrt statt und wird von der Joachim Herz Stiftung gefördert.\\ \begin{center} \includegraphics[height=1.5cm]{HdA_500x260.png} \hspace*{4em} \includegraphics[height=1.5cm]{DLR_Logo_svg.png} \hspace*{4em} \includegraphics[height=1.5cm]{Joachim_Herz_Logo_sRGB.png} \end{center} \end{document} \section*{Zusammenfassung} In ferner Zukunft werden sicher Menschen zum Mars fliegen. Schon heute führen Sonden und Fahrzeuge auf der Marsoberfläche Messungen durch. Landungen können auf dem Mars jedoch meistens nicht alleine mit Fallschirmen durchgeführt werden. Stattdessen verwendet man Landemodule, die mit Triebwerken die Geschwindigkeit auf akzeptable Werte verringern. Im Vergleich zur Erde übt der Mars zwar eine geringere Anziehungskraft auf die Sonden aus. Jedoch ist seine Atmosphäre sehr viel dünner als die der Erde. Mit simplen Rechnungen und einigen vereinfachenden Annahmen berechnen Schülerinnen und Schüler die Geschwindigkeit einer Marssonde mit Fallschirm. Dadurch erkennen sie, dass aufgrund der dünnen Marsatmosphäre die Endgeschwindigkeit keine weiche Landung zulässt. \section*{Lernziele} Die Schülerinnen und Schüler \begin{itemize} \item ermitteln eine Bestimmungsgleichung für die Endgeschwindigkeit eines Falls mit Fallschirm, \item berechnen die Endgeschwindigkeiten von Fallschirmabwürfen auf dem Mars und auf der Erde, \item erläutern die Gründe dafür, dass Fallschirme alleine keine weiche Landung zulassen, \item berechnen die Größe eines Fallschirms, der eine weiche Marslandung ermöglichen würde. \end{itemize} \section*{Materialien} \begin{itemize} \item Arbeitsblätter (erhältlich unter \href{http://www.haus-der-astronomie.de/raum-fuer-bildung}{http://www.haus-der-astronomie.de/raum-fuer-bildung}) \item Stift \item Taschenrechner \item Computer/Tablet/Smartphone mit Internetzugang (optional) \end{itemize} \section*{Stichworte} Mars, Erkundung, Reibung, Luftwiderstand, Luftdruck, Dichte, Strömung \section*{Dauer} 90 Minuten \end{titlepage}

1. Introduction =============== Missed abortion (MA), also known as silent miscarriage, usually occurs when an embryo or fetus dies, but the body does not recognize the pregnancy loss and continues to release hormones.^\[[@R1]\]^ MA is very common complication, which affects ∼15% of all clinically recognized pregnancies.^\[[@R2]\]^ The loss of a pregnancy is often distressing for women and their partners, with adverse effects on their social and psychological wellbeing.^\[[@R3]\]^ In China, the incidence of MA has been rapidly increasing (13. 4%) presenting itself as an important public health problem.^\[[@R4]\]^ The pathogenesis of MA has not yet been fully understood, and it is commonly believed that the combination of multiple factors may lead toward this outcome. The main known cause of MA involves genetic factors, such as chromosomal abnormalities, that is, the number of chromosomes or their structural abnormalities.^\[[@R5]\]^ The other risk factors for MA include an abnormal intrauterine environment, endocrine disorders, immune dysfunction, excessive smoking, severe systemic infections, nutritional deficiencies, environmental chemicals, and trauma.^\[[@R6]--[@R8]\]^ Over the years, environmental pollution has become another important risk factor for MA. It has been shown that the ambient air pollutants may influence the progress of the pregnancy and fetus growth, and may lead to poor birth outcome. Furthermore, exposure to harmful chemical and physical factors have shown strong association with cessation of embryonic development, which may consequently lead to MA. In this study, we investigated the association of tobacco, heavy metals, and phthalates exposure with MA. During the early pregnancy, harmful substances abundantly present in the tobacco smoke, such as nicotine and cotinine, reduce progesterone secretion, and may consequently lead to spontaneous abortion. Phthalates are a plastic plasticizer widely used in plastic industry for medicinal products, food packaging, personal care products, and polyvinyl chloride.^\[[@R9],[@R10]\]^ As environmental endocrine disruptors, phthalates have shown to have estrogen activity, which can induce the reproductive and developmental toxicity.^\[[@R11]\]^ Urine or blood levels are usually used to assess the chemical exposure. However, a lot of chemicals have a short half-life, therefore the level of these chemicals in urine or blood may not reflect the actual whole-body exposure. For this reason, hair samples might provide a better option. Hair is a biological sample that is easy to collect, transport, and storage. Additionally, hair has certain enrichment of metal elements, so it is widely used for trace elements analysis.^\[[@R12]\]^ Moreover, compared to blood or urine samples, hair has substantially longer detection window (month to year) which may enable retrospective investigation of chronic and past consumption.^\[[@R13]\]^ This memory effect the hair has due to accumulation of chemicals provides possibility to examine exposure to variety of pesticides (including different chemical categories such as organochlorines, organophosphates, pyrethroids, and other), polybrominated diphenyl ethers, polycyclic aromatic hydrocarbons, polychlorinated biphenyls, and phthalates.^\[[@R12],[@R14]\]^ Consequently, this study uses the hair as biological sample to assess the exposure to tobacco, heavy metals, and phthalates. In the following paper, we measured the levels of tobacco and heavy metals exposure by examining nicotine and cotinine, and plumbum, cadmium and manganese, respectively, in hair samples. We measured the exposure to heavy metals by detecting plumbum (Pb), cadmium (Cd), manganese(Mn) in hair samples. In addition, we analyzed the exposure to phthalates by testing phthalic acid esters, including dimethyl phthalate (DMP), diethyl phthalate (DEP), dibutyl phthalate (DBP), butyl benzyl phthalate (BBP), and di-2-ethyl hexyl phthalate (DEHP) in the same hair samples. 2. Methods ========== 2.1. Study population --------------------- Pregnant women, \<20 weeks' gestation, age between 19 and 46 years, were recruited from the Department of Obstetrics and Gynecology of Guangxi Medical University Affiliated Hospital and the People Hospital of Guangxi from March to May 2012. All study subjects received a detailed description of the study protocol and signed the informed consent approved by the institutional review board of Guangxi Medical University. Trans-vaginal ultrasound examination was used to obtain real-time visual images of the *developing embryo* or fetus, empty gestational sac or an embryo/fetus without cardiac activity was used as diagnosis criteria for miss abortion.^\[[@R15]\]^ The pregnant woman who was diagnosed as miss abortion was recruited as MA group. The pregnant women having ectopic pregnancy, hydatidiform mole, twin pregnancy, gynecological inflammation, and reproductive tract abnormalities were excluded from the study. Finally, 42 patients with MAs and 57 controls completed the questionnaire and provided hair samples. The case group and control groups were matched by ethnologic and gestation length. Each subject completed a questionnaire survey which included subject\'s general information (age, weight, gestational age, ethnologic data, etc.), previous drugs consumption, harmful factors exposure (heavy metals, pesticide, chemical agent, etc.) and partner\'s general situation (age, smoking, drinking, work environment, etc.). Additionally, the questionnaire had question about frequency and quantity for tobacco, heavy metals, and plastic products exposure during one year before pregnancy and during early stages of pregnancy. Pregnant women hair samples were cut from occipital clingy scalp using clear stainless steel scissors. Since hair grows at a rate of ∼13 mm per month,^\[[@R16]\]^ assay of a 3.9 cm length of hair close to scalp covered the early gestational period. The hair samples were put into clean bag and tagged. 2.2. Tobacco, heavy metal, and phthalate exposure assessment ------------------------------------------------------------ The concentration of nicotine, cotinine in the hair samples was analyzed using gas chromatography (6890N, Agilent, America), according to the following method: hair samples were cleaned by detergent (pH = 7.0), tap water, and deionized water, and then left to dry out. Samples were then cut into 1 to 2 mm long using a clear stainless steel scissors. Hair samples (∼0.48--0.52 g) were then placed in a glass tube, by adding 5.00 mL NaOH (2M) and 20 μL quinolone internal standard (0.109 mg/mL). The samples were ultrasonicated for 30 minutes and incubated for 4 hours at 40 ± 1°C for digestion. After incubation, 5.00 mL methanol--methylene chloride mixed solvent (1:3, V: V) was added into the tube. The process to transfer nicotine and cotinine from the digestion liquid to the organic phase was done by vortexing the sample for 2 minutes and then transferring 2.50 mL organic extract solution into a glass centrifuge tube. Around 20μL glacial acetic acid was added. The mix solution was dried by nitrogen. The dry sample was added 0.20 mL methanol and blander vortex mixed for 1 minutes, finally centrifuged with 2000 rpt/min for 5 minutes. Detection of phthalates was done as follows: 0.1 g of hairs samples were transferred in a glass tube containing 1.0 mL 2 mol/L NaOH solution. The tube was then placed in 37°C constant temperature bath box for 2.5 hours to digest adequately. The digested liquid was added in 1 mL methylene chloride extraction solvent. After blending, the mixed liquid was extracted using ultrasonic method for 30 minutes, and then centrifuged. Finally, phthalates were analyzed by gas chromatography. For hair heavy metal analysis, ∼0.5 g hair samples were added to a sample cup containing 5 mL of HNO~3~--H~2~O~2~ (3:1, V: V). The mixture was pretreated at 100°C for 30 minutes using multipurpose pretreatment device (XT-9800, Shanghai Xintuo analytic instrument science and technology Ltd.). After cooling, the mixture was then digested for additional 8 minutes on 0.1 MPa using an optical fiber pressure control closed microwave digestion system (MK-III, Shanghai Xinke Microwave Digestion Sample Testing Technology Research Institute). Consequently, the sample was volatilized acid at 160°C in the pretreatment of heating apparatus to nearly dry, then transferred to 25 mL volumetric flask and constant volume with deionized water. The concentration of Mn in the digestion liquid was determined by flame method using atomic absorption spectrophotometer (AA-6800F, Shimadzu, Japan). 5 mL of the digestion liquid was added 0.2 mL of 0.1 g/mL (NH~4~)~2~SO~4~--KH~2~PO~4~ matrix modifier, then with the deionized water to capacity 10 mL. The concentration of Pb and Cd was determined by graphite furnace method using atomic absorption spectrophotometer (AA6300C, Shimadzu, Japan). 2.3. Statistical analysis ------------------------- The differences between groups were analyzed with Student\'s *t*-test for continuous variables and the chi-square test for categorical variables. The odds ratio (OR) associated with cigarette smoking, heavy metals, and phthalates exposure were calculated by comparing the control group with the case group. All variables identified on the basis of biologic plausibility and preliminary analyses as possible confounders were included in logistic-regression models, which were then simplified with the use of a backward-elimination technique. All final models included the same set of statistically significant covariates and each final model also included questionnaire survey and hair analysis of tobacco exposure, heavy exposure, and phthalates exposure. Logistic regression analysis was used to compare MA group and the control group for all potential factors, including the background information, environmental, and behavioral factors. Odds ratio (OR) and 95% confidence interval (CI) were used to estimate the associations between the underlying factors and the MA. Statistical significance was defined as *P* \< .05. All statistical analyses were performed with SPSS 17.0 (SPSS, Chicago, IL). 3. Results ========== 3.1. Participant characteristics -------------------------------- Participant\'s baseline characteristics and exposure factors data were obtained using the questionnaire survey (Table [1](#T1){ref-type="table"}). Briefly, the results showed that the average age of women who suffered from MA (case group) was slightly higher compared to the women with normal pregnancy (control group) (*P* \< .05). In addition, controls were more likely to have premarital health check-up compared to the case group (*P* \< .05). Furthermore, the exposure to passive smoking and heavy metals during pregnancy was significantly higher in the case group compared to a control group (*P* \< .05); while no difference in levels of education, race, active smoking, infection, plasticizer contact during pregnancy, etc. were observed between the 2 groups (*P* \> .05). ###### Characteristics of the study subjects.  3.2. Comparison of concentration in the case and control groups --------------------------------------------------------------- Different levels of nicotine, cotinine, Mn, Ca, Pb, DMP, DEP, DBP, BBP, and DEHP from the participant\'s hair samples are shown in Table [2](#T2){ref-type="table"}. In order to find the liminal value of the risk factors, we used several cutoff values to compare the risk factor levels. The obtained results showed higher levels of nicotine (*P = .*037) and cotinine (*P = .*018) in the case group compared to the control group. These data re-confirmed results *obtained* elsewhere, nicotine and cotinine are risk factors for MA.^\[[@R19]\]^ A total of 0.5 μg/g levels of nicotine and the 1 μg/g levels of cotinine found in control group were higher compared to case group. However, exposure to high levels of nicotine (\>0.5 μg/g) and cotinine (\>1 μg/g), the proportions of case group (23.81%, 28.57%) were significantly higher compared to the control group (7.02%, 12.28%). Additional, no significant changes in the Mn level were observed between the groups (*P* \> .05); while the levels of Pb (*P* = .038) and Cd (*P* = .010) were significantly higher in the case group. ###### Distribution of smoking chemicals, heaven mental and phthalates in the hairs samples of 2 groups.  Furthermore, we were unable to detected DMP and DEP from the hair samples. The reason for this may be since these 2 kinds of phthalic acid ester in hair were under limit of detection (LOD). However, we successfully detected DBP, BBP, and DEHP from the 2 groups; no significant differences in DBP and BBP were observed (*P \> .*05), while the significantly higher concentration of DEHP was observed in the case group (*P = .*001). At the 0 to 4 mg/L level of DEHP, the proportion of control group (91.23%) was higher than those of the cases (59.52%). But at the 4 to 6 mg/L level of DEHP, the proportion of case group (19.05%) was significantly higher than those of the controls (3.51%). 3.3. Potential factor exposure and risk of MA --------------------------------------------- Table [3](#T3){ref-type="table"} shows the logistic regression model that was used to estimate the potential factors associated with MA. In the initial analysis, statistically significant factors, that is, age, Pb, Cd, nicotine, cotinine, and DEHP were entered in the model. Age significantly increased the risk of MA (*P = .*012, OR 1.172, 95% CI 1.036--1.327). Pb exposure was significantly associated with an increased MA risk (*P = .*070, OR 8.26, 95% CI 1.779--38.342). Cd exposure increased the risk of MA (*P = .*040, OR 8.931, 95% CI 2.003--39.811). Nicotine exposure was associated with an increased MA risk (*P = .*090, OR 8.347, 95% CI 1.707--40.418). Cotinine exposure was also a risk factor for MA (*P = .*029, OR 4.376, 95% CI 1.159--16.513). Compared to low DEHP content in the hair, high DEHP content increased the possibility of MA (*P = .*020, OR 1.863, 95% CI 1.103--3.146). ###### Model of logistic regression of potential factors on missed abortion.  4. Discussion ============= According to the results of the present study, the statistically higher average age was observed in the case group (28.76) compared to the control group (26.7), suggesting the age is a risk factor for MA. This data reconfirmed findings from previous studies,^\[[@R17]--[@R19]\]^ which demonstrated that age was strongly associated with MA occurrence. This might be because in older pregnancy ovaries, uteri are reduced, and the egg quality drops, causing chromosomal changes.^\[[@R18],[@R19]\]^ Studies have shown that,^\[[@R17]--[@R20]\]^ before getting pregnant, partner that smokes presents a risk factor for pregnant women abortion. Pregnant women exposure to tobacco smoke also increases the risk of placental abruption, placenta previa, fetal growth restriction, and other pregnancy complications.^\[[@R21]\]^ Active smokers inhale first hand smoke, while passive smokers inhale side stream smoke which due to the incomplete combustion of tobacco, contains harmful material more dangerous than the mainstream smoke.^\[[@R22]\]^ Tobacco smog effect in pregnant women is mainly due to harmful substances (such as nicotine, cotinine, carbon monoxide, and tar) in the smoke which cause pathological changes leading to the placenta calcification, which in turn includes shorter microvilli, reduced placental blood vessels, and increased collagen content of fuzzy matrix.^\[[@R23]\]^ It can furthermore increase the placenta and uterus contraction, reducing the flow of blood leading to miscarriage and stillbirth.^\[[@R13]\]^ Nicotine and cotinine from the mother\'s body may enter the placental circulation having direct toxic effect on the fetal cardiovascular system.^\[[@R24]\]^ In this study, the levels of nicotine OR = 3.31, 95% CI: 0.01387--0.54547, and cotinine (OR = 2.16, 95% CI: 0.51264--0.58607), in the hair of the case group were higher compared to normal pregnancy group, and the differences were statistically significant (*P* \< 0.05). This implies that pregnant women early exposure to tobacco smoke environment increases the risk of MA. Exposure to environmental pollution and occupational, heavy metal elements can translocate through the placenta from pregnant woman\'s body to the fetus where they accumulate.^\[[@R25]\]^ Erikson et al^\[[@R26]\]^ found that excessive exposure to Mn during pregnancy, in addition to inducing obvious neurotoxicity in pregnant women, it can also cause long-term growth retardation in offspring. Contrary, if pregnant women should experience serious lack of Mn, this might lead to fetal congenital malformations, stillbirth, or habitual abortion.^\[[@R27]\]^ Our research results showed no statistically significant difference between Mn levels in the hair of control group and case group, implying the occurrence of MA was not necessarily associated to hair Mn content. However, the levels of Pb, Cd in the hair of the case group were higher compared to controls. The toxicity of Pb on fetal embody works through reduced membrane Na^+^, K^+^ atpase activity. Pb can inhibit ovarian corpus luteum cells synthetic progesterone key enzyme activity of steroid hormones; and can interfere with ovarian hormone stimulation on the lining of the uterus. Furthermore, Pb can cause endometrial atrophy, and make disintegrating uterine decidua cells release a large number of phosphatase. Transforming arachidonic acid to prostaglandins causes lining of the uterus to contract, which may be the main reason behind miscarriage and stillbirth.^\[[@R28]\]^ According to the European Economic Community (EEC) and the organization for economic co-operation and development (OECD) teratogenic content classification, Cd belongs to potential teratogenic. Cd contributes in developmental toxicity in animals.^\[[@R29]\]^ Cd suppresses the embryonic cells through the placenta and DNA replication and protein biosynthesis, especially on the nucleoli rRNA synthesis or significantly inhibited transcription. Existing research has shown that the accumulation of Cd can provoke morphological changes of the placenta, such as increased maternal fetal capillaries gap, with smaller diameter. Cd may in turn affect the blood supply of the placenta and fetal blood flow.^\[[@R27]\]^ Pb, Cd may also inhibit the biosynthesis of embryonic cells and damage the structure of the cell membrane, in turn causing pathological changes, and miscarriage. We detected 5 phthalates (DMP, DEP, DBP, BBP, and DEHP) in the hair of pregnant women. Besides the concentration of DMP and DEP that were below LOD, we could detect DBP, BBP and DEHP of all participants. This means that a lot of pregnant women had no knowledge about phthalate exposure and were probably not taking any precautions to avoid these chemicals. According to our research, the levels of DBP and BBP were high but without statistical significance in the 2 groups. However, the level of DEHP in the case group was significantly higher compared to the controls. In lower level of DEHP, there were no differences between the 2 groups. But in the higher level, there were significant differences between the 2 groups. The obtained results were similar with the does--response relationships between urinary phthalate concentrations and risk of clinical pregnancy loss.^\[[@R30]\]^ Our results showed that value OR for DEHP was 1.83, suggest that the exposure to high level of DEHP may be a risk factor for MA. Adverse female reproductive and developmental outcomes have been observed in several animal studies where the exposure to specific phthalates \[DEP, DEHP, di-*n*-butyl phthalate (DnBP), and di-iso-butyl phthalate (DiBP)\] were linked with decreased embryo survival, increased incidence of resorptions, reduced number and size of litters, and increased abortion rates in rats.^\[[@R31],[@R32]\]^ Schmidt reported that chronic occupational exposure to high levels of phthalates has been linked with decreased rates of pregnancy and higher rates of miscarriages in female factory workers.^\[[@R33]\]^ The present study has some limitations. Firstly, we did not measure the levels of human chorionic gonadotropin (hCG) in pregnant women, since once MA occurs the levels of hCG must decline. If we wanted to find the relation between MA and the levels of hCG, we should detect the levels of hCG before the embryo die, which is impossible in a case-control design. Secondly, due to a limited number in this group, the estimated OR was associated with a wide CI. If the sample size was larger, the results would be more convincing. Finally, this study did not test for chemicals in the blood or urine to confirm the results of the experiments. In summary, our results suggest that high gestational age, passive smoking, heavy metals and the phthalate exposure are risk factors for MA. Premarital health examination is a strong protective factor against MA. Abbreviations: BBP = butyl benzyl phthalate, Cd = cadmium, CI = confidence interval, DBP = dibutyl phthalate, DEHP = di-2-ethyl hexyl phthalate, DEP = diethyl phthalate, DiBP = di-iso-butyl phthalate, DMP = dimethyl phthalate, DnBP = di-n-butyl phthalate, EEC = The European Economic Community, hCG = human chorionic gonadotropin, LOD = limit of detection, MA = missed abortion, Mn = manganese, OECD = The Organization for Economic Co-operation and Development, OR = odds ratio, Pb = plumbum. This study was supported by Natural Science Foundation of Guangxi (2013JJAA40161), National Natural Science Foundation of China (81360440) and Guangxi Colleges and Universities Key Laboratory of Prevention and Control of Highly Prevalent Diseases grants. The authors have no conflicts of interest to disclose.

\section{INTRODUCTION} Long period comet C/2014 B1 (Schwartz) (hereafter ``B1'') follows a hyperbolic orbit with semimajor axis $a$ = -2160 AU, eccentricity $e$ = 1.00442 and inclination $i$ = 28.4\degr~(orbital elements are from JPL's HORIZONS on-line database). Although B1 is technically not gravitationally bound to the Sun, the eccentricity excess above $e$ = 1 is so small that an interstellar origin is unlikely. Instead, along with numerous other long-period comets having slightly hyperbolic orbits, B1 is of probable Oort cloud origin. Perihelion occurred at $q$ = 9.557 AU on UT 2017 September 10. Apart from its atypically large perihelion distance, the special feature of B1 is its peculiar, indeed unique, optical morphology. Specifically, its shape resembles a discus or bi-convex lens viewed edgewise. Moreover, the major axis of symmetry of this elongated object lies nearly (but not exactly) perpendicular to the projected orbit and shows no relation to the anti-solar direction. In other comets, solar gravity and solar radiation pressure are the dominant forces shaping the large-scale morphology of the dust coma. Small dust particles are accelerated by the momentum of solar photons and pushed into a radiation pressure swept tail that is aligned approximately in the anti-solar direction. Large particles are less influenced by radiation pressure and are also ejected more slowly from the nucleus; they instead populate a narrow region along the orbital plane, appearing as a band or ``trail'' in the plane of the sky. The morphology of B1, and the lack of a clear relation to the anti-solar and projected orbit directions together present a puzzle regarding the nature of the dust and the mechanism of its ejection. In this paper we present observations taken to characterize B1 and to attempt to understand its unique and distinctive appearance. \section{OBSERVATIONS} Observations were acquired using the Wisconsin-Indiana-Yale-NOAO (WIYN) 0.9 m telescope, the Nordic Optical Telescope (NOT) 2.5 m, and the Keck 10 m telescope. A journal of observations, including basic instrumental parameters and the observing geometry for each date of observation, is given in Table (\ref{geometry}). In the Table, we express dates as DOY (Day of Year) where DOY = 1 on UT 2014 January 1. For reference, B1 was discovered on DOY = 28 (Schwartz and Sato 2014) and perihelion occurred on DOY = 1349. We corrected for spatial sensitivity variations across the charge-coupled detectors employed at each telescope, using flat fields computed from the median of a large number of uncorrelated night sky images (at Keck) or obtained from an illuminated patch on the inside of the telescope dome (at the other telescopes). Photometric calibration of the red filter data was obtained from observations of Sun-like stars from the catalog by Landolt (1992) and from field objects measured in the Sloan Digital Sky Survey Data Release 14 (Abolfathi et al.~2018). Magnitudes obtained using Sloan filters were transformed to the Johnson-Cousins BVR system using the relations determined by Jester et al.~(2005). In both cases, the standard star data are accurate to $\sim \pm$0.01 magnitudes in BVR and the photometric stability of the atmosphere was typically $\pm$0.01 to $\pm$0.02 magnitudes. The accuracy of the comet photometry (Table \ref{photometry}) is further limited by uncertainty in the sky background owing to the presence of field stars and galaxies. We present a composite of the data from each night listed in Table (\ref{geometry}) in Figure (\ref{composite}). These images are from telescopes with widely different collecting areas, have different total integration times and resulting sensitivities, and were taken in different sky conditions. While the extent of the measurable coma varies from image to image, the overall stability of the morphology is noteworthy, particularly with respect to the varying projected anti-solar vector, marked in yellow in each panel. A high quality image from the NOT telescope is shown in Figure (\ref{image}). \subsection{Photometry} We took measurements using an aperture with an angular radius scaled inversely with geocentric distance ($\Delta$, see Table \ref{geometry}), so as to always measure the light scattered from within a fixed linear (not angular) distance from the nucleus (which we assumed to be located at the opto-center of the comet). In this way, the measured apparent brightness of a fixed cross-section coma should vary in accordance with the inverse square law without need for an additional geometric correction related to the surface brightness distribution in the coma. Table (\ref{photometry}) lists apparent magnitudes obtained within a circular aperture of projected radius 4$\times$10$^4$ km at the comet, denoted $m_R(40)$. If there were no change in the coma of B1 then $m_R(40)$ would vary according to the inverse-square law, here expressed as \begin{equation} m_R(40) = H_R(40) + 2.5\log_{10}\left(r_H^2 \Delta^2\right) - 2.5\log_{10}(\Phi(\alpha)) \label{abs} \end{equation} \noindent where $r_H$ and $\Delta$ are the heliocentric and geocentric distances expressed in AU and $\Phi(\alpha) \le 1$ is the phase function at phase angle $\alpha$. Equation (\ref{abs}) defines the absolute red magnitude, $H_R(40)$, as the magnitude the comet would have if it could be observed from $r_H = \Delta = 1$ AU and $\alpha =$ 0\degr. The phase functions of comets are reasonably well-approximated by simple functions of $\alpha$, provided $\alpha \lesssim$ 50\degr. We write $-2.5\log_{10}(\Phi(\alpha))$ = $k \alpha$, where $k$ is a constant. The phase functions of cometary nuclei fall in the range 0.02 $\le k \le$ 0.1 (Kokotanekova et al.~2017). The phase functions measured in active near-Sun comets tend to be smaller; we take $k$ = 0.4 magnitudes degree$^{-1}$ (Meech and Jewitt 1987). Over the modest range of phase angles taken by B1 ($\alpha \le$ 3.8\degr, Table \ref{geometry}), even a $\pm$50\% uncertainty in the coefficient would have only a modest effect ($\pm$0.08 magnitudes) on $H_R(40)$. Absolute magnitudes computed from Equation (\ref{abs}) are listed in Table (\ref{photometry}) and plotted in Figure (\ref{absmagplot}). The error bars in the Figure are computed from $\sigma_T = (\sigma_m^2 + 0.08^2)^{1/2}$, where $\sigma_m$ is the measurement error listed in Table (\ref{photometry}) and 0.08 is our estimate of the phase function uncertainty. B1 shows an intrinsic brightening of $\sim$1 magnitude between the epoch of discovery in 2014 and perihelion in 2017. A least-squares fit to the data from Table (\ref{photometry}), plotted in Figure (\ref{absmagplot}) as a red line, shows that the cross-section varies as $C_e \propto r_H^{-4.9\pm0.4}$. This is steeper than the $r_H^{-2}$ dependence expected of a super-volatile sublimator. However, the gradient of this fit is heavily dependent on the one 2014 measurement and so we do not imbue it with further signifiance. The absolute magnitude is related to the effective scattering cross-section, $C_e$, by \begin{equation} C_e = \frac{1.5\times 10^6}{p_R} 10^{-0.4 H_R(40)} \label{area} \end{equation} \noindent where $p_R$ is the red geometric albedo. Comet dust albedos are, in general, not well measured, with evidence for a wide range of values amongst comets and even temporal variability (declining from 0.12 to 0.04) within a single object (Ishiguro et al.~2010). We here assume a nominal albedo $p_R$ = 0.1, while acknowledging that the true value may be larger or smaller by a factor of several, especially if the grains are icy (as, indeed, seems likely at 10 AU). Derived values of $C_e$ from Equation (\ref{area}) are listed in Table (\ref{photometry}). Both the large cross-sections listed there and the extended morphology of B1 indicate an origin by scattering from the coma, with a minimal (and poorly constrained) contribution from the central nucleus. The optical colors of B1 (B-V = 0.85$\pm$0.03, V-R = 0.58$\pm$0.03) are slightly redder than the average colors of 25 long-period comets (B-V = 0.78$\pm$0.02, V-R = 0.47$\pm$0.02, see Jewitt 2015), and redder than the Sun (for which B-V = 0.64$\pm$0.02, V-R = 0.35$\pm$0.01) but less red than the ``ultrared'' matter (B-V = 1.06$\pm$0.02, V-R = 0.66$\pm$0.02) that is a distinctive feature of many Kuiper belt objects (Jewitt 2015). These colors are consistent with scattering from dust; the resonance fluorescence bands from abundant gas phase radicals (e.g.~CN, C$_2$, C$_3$) are heavily concentrated towards short wavelengths and would produce colors bluer than observed, especially in B-V. Indeed, cometary gas rarely dominates the optical spectrum except near $r_H \sim$ 1 AU, and has not been detected in a comet as far from the Sun as is B1. \subsection{Structure of the Coma} While the comet intrinsically brightened by about one magnitude, the morphology of B1 in our data changed little from that in Figure (\ref{image}) in four years of observation. In order to quantify this, we separately measured the position angles of the central axes of the coma extending to the north and the south of the nucleus, labeled $\theta_N$ and $\theta_S$, respectively, as a function of time. To do this, we used perpendicular cross-sections at different distances from the nucleus. The position angles were determined from linear least-squares fits to the position of peak surface brightness, weighted by the uncertainty on the position of the peak. Table (\ref{photometry}) lists the resulting $\theta_N$ and $\theta_S$ together with their uncertainties. For the latter, we found that the formal errors from the least-squares fits are small compared to systematic errors caused by structure in the sky background, especially at large projected distances from the nucleus where the surface brightness of the coma is very low. (In addition, our representation of the long axis of the coma as linear is itself only an approximation, and some of the images hint at curvature which we do not here further consider). The listed errors are our best estimates of the true uncertainties in $\theta_N$ and $\theta_S$, including these systematic, background effects. Figure (\ref{angle_plot}) compares the measured angles, as a function of time, with the position angles of the projected antisolar ($\theta_{-\odot}$) and orbital velocity ($\theta_{-V})$ vectors. Neither $\theta_N$ nor $\theta_S$ shows any relation to $\theta_{-\odot}$ or $\theta_{-V}$, showing that the shape of the coma is a property of the comet, not a result of external influences. The long axes of the coma measured to the north and the south of the nucleus appear nearly, but not quite, aligned. We define the angle between the north and south extensions as $\theta_{N-S} = |\theta_N - \theta_S |$, finding a weighted mean value $\theta_{N-S}$ = 176\degr.1$\pm$0\degr.8 (Table \ref{photometry}), which is close to but significantly different from 180\degr. The only hint of an angle relationship is between $\theta_{N-S}$ and the Earth elevation, $\delta_{E}$, shown in Figure (\ref{tilt}) (data from Tables \ref{geometry} and \ref{photometry}). Such a relationship would be expected if $\theta_{N-S}$ is affected by projection of the three-dimensional dust distribution into the plane of the sky. However, a weighted least-squares fit (plotted as a straight line in Figure \ref{tilt}) gives $\theta_{N-S} = 176.5\pm0.8 - (0.93\pm0.37)\delta_{E}$, showing that the dependence is significant only at the 2.5$\sigma$ level. We therefore do not interpret it further. We also measured the full width at half maximum of the coma, as a function of $\ell$, the projected angular distance from the opto-center measured along the long axis of the coma. To do this, we first rotated the image to bring the long axis to vertical, and then made horizontal cuts across the coma to determine the angular Full Width at Half Maximum (FWHM). Figure (\ref{fwhm}) shows the resulting FWHM, called $\theta_{1/2}$ (orange bars), as a function of $\ell$. As remarked earlier and apparent in Figure (\ref{image}), the optical appearance of B1 is discus-like, with a width that decreases with increasing distance from the nucleus. Figure (\ref{fwhm}) shows the FWHM in fact increases with distance from the nucleus, giving a double fan type structure to the coma. Least-squares fits to the data of Figure (\ref{fwhm}) having the form $\theta_{1/2} = a \ell + b$, give $a$ = 0.81$\pm$0.02 to the north and $a = -0.86\pm0.02$ to the south. The opening angles of the fans are calculated from tan$^{-1} a$ = 39\degr$\pm$1\degr~in the north and 41\degr$\pm$1\degr~in the south. The close similarity between these angles reflects the symmetry apparent in Figure (\ref{image}). \subsection{Radial Surface Brightness} We computed the surface brightness profile, $\Sigma(\phi)$ [magnitudes per arcsecond$^{-2}$], where $\phi$ is the angular distance from the optocenter. We present the data from UT 2018 March 11 because these have the least background contamination from field objects; profiles from other dates are similar but of lower quality. The surface brightness within a nested set of concentric, circular apertures is plotted in Figure (\ref{profile}). The background was determined from the median of the pixels in a concentric annulus having inner and outer radii 54\arcsec~and 64\arcsec, respectively. At small angles from the optocenter, $\phi \lesssim 2\arcsec$, the surface brightness is affected by convolution with the point spread function of the image. Conversely, at large angles, $\phi \gtrsim$ 10\arcsec, the accuracy of the surface brightness determination is increasingly limited by the uncertainty in the brightness of the background sky. We estimated the sky uncertainty by experimenting with different sky apertures; its value is dominated by imperfections in the flat fielding of the data and by residual trails of field objects. The surface brightness profile is usefully characterized by its gradient, $m = d \ln \Sigma(\phi)/d \ln \phi$. A coma produced in steady-state should, in the absence of radiation pressure, have $m = -1$ while, in the presence of radiation pressure, the gradient should asymptotically steepen towards $m = -1.5$ as a result of the steady acceleration of the particles (Jewitt and Meech 1987). While most simply derived for the case of a spherically symmetric, steady-state coma, the $m$ = -1 gradient is a consequence only of the equation of continuity, and applies equally well to anisotropic comae provided they remain optically thin. Therefore, it is reasonable to use concentric apertures and to interpret the surface brightness gradient of B1 in the above terms even though its coma is clearly anisotropic. The measured gradient in our best data (from UT 2018 March 11) B1 is $m = -1.10\pm0.01$ in the range $2\arcsec \le \phi \le 10\arcsec$, more similar to the steady-state value than to the radiation pressure limit. However, we also find that $m$ is itself an apparent function of $\phi$ ($m$ becomes more negative as $\phi$ increases, see Figure \ref{profile}). If real, this steepening coma is probably due either to fading of the grains (e.g.~due to ice loss or to their disaggregation into tiny sub-components) or to radiation pressure acceleration at large angles. However, the outer portion of the surface brightness profile where the gradient changes is precisely where the sky subtraction uncertainties become dominant (see error bars in Figure \ref{profile}) and so we cannot reach any definitive interpretation of this region. Gradients measured in other data from Table (\ref{geometry}) are consistent with the value measured on March 11, but are less accurate as a result of background uncertainties. \section{DISCUSSION} The following special features of B1 are deserving of explanation; \noindent 1) a persistent bi-convex lens shaped appearance (Figure \ref{image}) with an underlying cone-like dust distribution (Figure \ref{fwhm}). \noindent 2) the weak evidence for the effect of radiation pressure on the dust (Figures \ref{image} and \ref{profile}) \noindent 3) the long-term stability of the position angle of the coma (Figure \ref{angle_plot}) \noindent 4) the fact that the position angles of the coma to the north and the south of the nucleus differ by nearly (but not exactly) 180\degr~(Table \ref{photometry}) Before resorting to a multi-parameter numerical model, we first consider what more can be learned about B1 based on the above observations and order of magnitude considerations. \subsection{Particle Size} We can immediately exclude the possibility that small particles dominate the scattering cross-section in the coma of B1 because these should be deflected into a classical, radiation pressure swept tail that is not seen. The outer isophotes (contoured in the right-hand panel of Figure \ref{image}) do show a large-scale ($\gtrsim$ 30\arcsec) east-west asymmetry that might be related to radiation pressure, corresponding to a distance $L \sim 2.2\times 10^8$ m at the comet, in the plane of the sky. If this ``tail'' makes an angle $\alpha$ to the line of sight, the physical length is $L/\tan(\alpha)$. We take $\alpha$ = 4\degr~(equal to the phase angle, see Table \ref{geometry}) to find $L \sim 3\times 10^9$ m, as our best estimate of the scale on which radiation pressure acts. Following convention, we define $\beta$ as the ratio of the acceleration due to radiation pressure to the acceleration due to solar gravity. For a sphere of radius $a$ and density $\rho$ this ratio is \begin{equation} \beta = \frac{3Q_{pr} L_{\odot}}{16\pi G M_{\odot}c \rho a} \label{beta} \end{equation} \noindent where $L_{\odot} = 4\times 10^{26}$ W and $M_{\odot} = 2\times 10^{30}$ kg are the luminosity and mass of the Sun, $G = 6.67\times 10^{-11}$ N kg$^{-2}$ m$^2$ is the gravitational constant and $c = 3\times 10^8$ m s$^{-1}$ is the speed of light. Dimensionless quantity $Q_{pr}$ is the radiation pressure efficiency, which is of order unity for particles with $a \gtrsim \lambda$, where $\lambda \sim$ 0.5 $\mu$m is the wavelength of light, and $Q_{pr} \ll 1$ otherwise (Bohren and Huffman 1983). Substituting $\rho$ = 500 kg m$^{-3}$, Equation (\ref{beta}) gives $\beta \sim a_{\mu m}^{-1}$, where $a_{\mu m}$ is the particle radius expressed in microns. The distance over which a particle is accelerated in time, $t$, is just $L = g_{\odot}(1) t^2/(a_{\mu m}r_H^2)$, where $g_{\odot}(1)$ = 0.006 m s$^{-2}$ is the gravitational acceleration at 1 AU and $r_H $ is expressed in AU. We take $r_H $ = 10 AU as representative for the period of observations discussed here. For example, consider a micron-sized particle ($a_{\mu m}$ = 1) released from the nucleus with zero initial velocity. B1 has been active for $t >$ 4 years (1.2$\times 10^8$ s, c.f.~Table \ref{photometry}), during which time the distance travelled relative to the nucleus would be $L \sim 4\times 10^{11}$ m ($\sim$ 2.5 AU), fully two orders of magnitude larger than the scale of the coma in Figure (\ref{image}). Even given the crude nature of this calculation, it is evident that the absence of a tail in B1 requires that the optically dominant particles have sizes $a \gg$ 1 $\mu$m. If we instead keep $t = 1.2\times 10^8$ s and set $L = 3\times 10^{9}$ m, representing the approximate scale for radiation pressure deflection as estimated above, we solve to find $\beta \lesssim$ 0.01 ($a \gtrsim 100~\mu$m) for the effective size of the particles. Smaller particles ($\beta \gtrsim$ 0.01) would have been accelerated by radiation pressure over distances larger than the size of the observed coma, forming a tail that is not observed. This argument is clearly approximate, but it serves to make the point that the stability of the morphology over four years and the absence of a radiation-swept tail imply that the effective particle size in B1 is very large. The prevalence of large particles in other distant comets has been noted (e.g.~Jewitt et al.~2017) and interpreted as a reflection of the effects of inter-particle cohesion, which binds small particles to the nucleus (Jewitt et al.~2018). \subsection{Production Rate} The dust production rate can be estimated to order of magnitude from the time dependence of the scattering cross-section in Table (\ref{photometry}). We observe that between UT 2014 February 26 and 2017 March 22, an interval of $\Delta t = 9.6\times 10^7$ s, the cross-section increased by $\Delta C_e = 1.8\times 10^4$ km$^2$. For an optically thin collection of spheres, the rates of change of the mass and cross-section are related by \begin{equation} \frac{dM}{dt} = \frac{4}{3} \rho \overline{a} \frac{\Delta C_e}{\Delta t} \label{dmbdt} \end{equation} \noindent where $\rho$ = 500 kg m$^{-3}$ is the assumed grain density (e.g.~see Figure 2 of Fulle et al.~2018) and $\overline{a}$ is the average particle radius. With $\overline{a} \gtrsim$ 100 $\mu$m, we find from Equation (\ref{dmbdt}) that $dM/dt \gtrsim$ 10 kg s$^{-1}$. Strictly, $dM/dt$ represents the difference between mass added to the 40,000 km radius photometry aperture at the nucleus and lost from the aperture by outflow at its outer edge. It therefore sets a strong lower limit to the mass loss rate from the nucleus. B1 is too distant and too cold for water ice to sublimate. However, its heliocentric distance and temperature are such that exposed amorphous water ice, if present, should crystallize (Guilbert-Lepoutre 2012). Crystallization, accompanied by the release of trapped gases and the expulsion of dust through gas drag forces could account for, or perhaps trigger, the observed activity. B1 is also warm enough for a very small area of exposed CO or CO$_2$ ice to drive the mass loss. By solving the energy balance equation (e.g.~as in Jewitt et al.~2017, 2018) for sublimating ice at the sub-solar point on a nucleus at $r_H$ = 10 AU, we find maximum CO and CO$_2$ sublimation rates of $f_s(CO) = 4\times 10^{-5}$ kg m$^{-2}$ s$^{-1}$ and $f_s(CO_2) = 2\times 10^{-5}$ kg m$^{-2}$ s$^{-1}$, respectively. To supply 10 kg s$^{-1}$ would require an ice patch of 0.3 km$^2$ for CO and 0.6 km$^2$ for CO$_2$. These are very modest surface areas, as may be seen, for example, by comparison with the Centaur comet 29P/Schwassman-Wachmann 1. At $r_H$ = 6 AU, 29P loses CO at $\sim$2000 kg s$^{-1}$ (Senay and Jewitt 1994), requiring a sublimating surface of 16 km$^2$, calculated in the same way. If the dust to gas mass ratio, $f_{dg}$, is different from unity then these sublimating areas should be multiplied by $f_{dg}^{-1}$. The ratio was $f_{dg}$ = 4 in 67P/Churyumov-Gerasimenko (Rotundi et al.~2015), $f_{dg} >$ 5 in long-period comet C/1995 O1 (Hale-Bopp) (Jewitt and Matthews 1999) and values as high as $f_{dg}$ = 30 have been reported (Reach et al.~2000). \subsection{Disk Ejection Model} \label{disky} The images suggest that the coma of B1 is a disk-shaped figure viewed edge-wise. If we identify the long axis of the coma with the projected rotational equator, then the projected pole of the nucleus must lie near the plane of the sky at position angle $\sim$80\degr~to 85\degr~in Figure (\ref{image}). The expanding north and south arms of the coma (Figure \ref{fwhm}) then represent the equatorial plane of the coma, where more dust is emitted than at larger latitudes. In this scenario, the near north-south symmetry of the coma is a product of rotation. Physically, we imagine that such a disk could result only if the equatorial rotational velocity of the nucleus is a significant fraction of the gravitational escape velocity and of the dust launch speeds. Particles would escape preferentially from low latitudes with the assistance of centripetal acceleration, while those emanating at higher latitudes would fall back along suborbital trajectories without contributing to the coma. Depending on the ratio of the rotation period to the night-side cooling time, the dust ejection could be continuous in azimuth or restricted to the day-side of the nucleus. We envision that the angle $|\theta_N - \theta_S|$ = 176\degr.1$\pm$0\degr.4 is not quite 180\degr~because the Earth is not exactly in the projected equator of B1. To explore this geometry and to test some of the inferences made above based on order-of-magnitude considerations, we used the Monte Carlo dust dynamics model of Ishiguro et al.~(2007). The model follows the motions of dust particles under the action of solar gravity and radiation pressure, after their ejection from the nucleus according to specified speed and direction parameters. With many free or under-constrained parameters the model cannot, in general, offer unique solutions for the particle and ejection parameters. The Ishiguro model is nevertheless valuable in allowing an exploration of parameter regimes that are consistent with the imaging data. We explored a range of parameters to try to match the morphology of B1 on UT 2018 March 11. As expected, plausible solutions from the generated models all required the optical dominance of large particles ejected from the nucleus with small speeds. The models also required emission from the nucleus over a narrow range of latitudes and a pole direction inclined to the line of sight by $\sim$90\degr. In all these regards, the Monte Carlo models support the inferences made above on the basis of simple physical arguments. A specific example is shown in Figure (\ref{bestmodel}). The parameter assumptions used to generate this model include a power-law distribution of particle sizes, with differential index $q$ = 3.5, dust emission continuous over four years, a velocity vs.~radiation pressure parameter relation, $V = V_1 \beta^{u1}$, with $V_1$ = 100 m s$^{-1}$ and $u1$ = 1/2, and a range of particle sizes from $\beta_{min}$ = 10$^{-4}$ to $\beta_{max} = 10^{-2}$. Experiments show that, while $\beta_{min}$ is poorly defined, $\beta_{max}$ (a measure of the smallest particles) cannot be substantially increased without destroying the similarity between the model and the data through the formation of a prominent tail (Figure \ref{betamax}). Thus, we conclude that the particle radii fall in the range from $a_{min}$ = 0.1 mm to $a_{max} \sim$ 10 mm and that their speeds of ejection, by the above relation, lie in the range $1 \lesssim V \le $ 10 m s$^{-1}$. The best-fit rotation vector points at right ascension 90\degr~and declination -3\degr~(or its opposite). The position angle of the rotation vector, projected into the plane of the sky, is 84\degr, in substantial agreement with the visual estimate of 80\degr~to 85\degr, mentioned above. In the model shown, dust is ejected from a band extending $\pm$10\degr~from the equator. Isotropic and hemispheric emission models failed to reproduce the lens-like appearance of B1 and can be rejected. In short, the Monte Carlo models substantially support inferences made above on the basis of visual examination of B1. The peculiar and stable discus shape of the coma can be matched by equatorial emission of large ($a >$ 0.1 mm), slow ($V <$ 10 m s$^{-1}$) particles from a nucleus having a spin pole near the plane of the sky. \subsection{Opposing Jets Model} An alternate possibility is that the north and south arms of the coma reflect ejection from diametrically opposite active area (``jet'') sources on the nucleus, one for each of the two arms. We consider this possibility less attractive than the disk ejection model for several reasons. First, the assumption of two comparably active, diametrically opposite jets continuously active for $\ge$ 4 years, while possible, is completely ad-hoc. Second, this conjecture is energetically problematic because the source regions would necessarily be located about $\pm$90\degr~from the subsolar point, near the sunrise and sunset terminators, where the illumination needed to drive the mass loss is weakest. Third, jets in comets are typically curved and/or temporally modulated by nucleus rotation (e.g.~Larson et al.~1987), whereas the coma in B1 is not. \subsection{Lorentz Force Model} The fact that the elongation of B1 is unrelated to the anti-solar and projected orbit directions raises the possibility that other forces (in particular, the Lorentz force), might be at play. The Lorentz force has been suggested as an agent in shaping sub-structures (the ``striae'') in the tails of some near-Sun comets (Ip et al.~1985) and in the coma of distant comet C/1995 O1 (Hale-Bopp) (Kramer at al.~2014) but has not otherwise been suspected. Dust particles exposed to sunlight are charged to a positive potential $W \sim$ 5 V by the loss of photoelectrons (Kimura and Mann~1998, Pavlu et al.~2008). On a grain of radius $a$, this potential corresponds to a charge $q = 4\pi \epsilon_0 W a$, where $\varepsilon_0 = 8.85\times10^{-12}$ F m$^{-1}$ is the permittivity of free space. The charge on the grain renders the particle susceptible to the Lorentz force, $\vec{F_L}$, given by $\vec{F_L} = q(\vec{v}\times \vec{B})$, where $\vec{B}$ is the magnetic flux density in the solar wind and $\vec{v}$ is the relative velocity of the wind passing the comet. Since $\vec{v}$ is radial, the relevant component of the magnetic field is the azimuthal component, and $\vec{F_L}$ is perpendicular to the solar system mid-plane. In-situ measurements over a range of heliocentric distances show that $\vec{B}$, while highly variable on the $\sim$1 month solar rotation timescale, varies inversely with $r_H$. We write $B = B_1 r_H^{-1}$ where $B_1 = 600$ (T m) is a constant (determined from Figure 11 of Balogh and Erdos 2013) and $r_H$ is in meters. For example, the flux density at 10 AU is $B = 0.4$ nT, with factor-of-two fluctuations occurring on the solar rotation timescale. The solar wind speed is also variable, depending on activity on the Sun, but is well represented by $v$ = 500 km s$^{-1}$, roughly independent of $r_H$ (Balogh and Erdos 2013). Combining these relations we define the ratio of the Lorentz force, $|\vec{F_L}|$, to the gravitational force as the ``magnetic $\beta$'', $\beta_m$; \begin{equation} \beta_m = \frac{3 \varepsilon_0 B_1 W r_H v }{G M_{\odot} \rho a^2} \label{beta_m} \end{equation} \noindent where $r_H$ is expressed in m, and the other quantities are as defined above. Finally, the ratio of the Lorentz acceleration to the radiation pressure acceleration from Equations (\ref{beta}) and (\ref{beta_m}) is \begin{equation} \frac{\beta_m}{\beta} = \frac{16\pi \epsilon_0 B_1 W v c }{Q_{pr} L_{\odot}} \left[ \frac{r_H}{a}\right] \end{equation} \noindent Substituting, and now expressing $r_H$ in AU and $a$ in microns, we obtain \begin{equation} \frac{\beta_m}{\beta} = 0.07 \left[\frac{r_H}{a_{\mu m}}\right]. \label{betaratio} \end{equation} \noindent Equations (\ref{beta}) and (\ref{beta_m}) are plotted in Figure (\ref{betaplot}), for three values of the heliocentric distance, $r_H$ = 1, 10 and 30 AU. The Figure and Equation (\ref{betaratio}) show that, for example, at $r_H$ = 1 AU, $\beta = \beta_m$ for $a = 0.07~\mu$m. Such tiny particles ($a/\lambda \sim 0.1$) are inefficient optical scatterers and, therefore, magnetic effects are not observed in comets near Earth even though tiny particles are abundant (McDonnell et al.~1986). At the perihelion distance of B1, namely $r_H \sim$ 10 AU, we find from Equation (\ref{betaratio}) that $\beta = \beta_m$ when $a = 0.7~\mu$m. Such micron-sized particles are efficient scatterers of optical photons and would be detected if they were abundant in B1. As described above, small particles are depleted from the coma of B1 as a result of cohesion and the effective particle radius is instead a relatively large $a \gtrsim 100~\mu$m. Substituting $r_H$ = 10 AU, $a_{\mu m} \gtrsim$ 100 into Equation (\ref{betaratio}) we find $\beta_m/\beta \lesssim 0.007$ showing that the effects of the Lorentz force can be safely ignored in the context of the lens-like morphology of this comet. The images of B1 also provide other evidence against the role of Lorentz forces. This is because, with $\vec{v}$ radial and $\vec{B}$ acting in the azimuthal direction, the direction of $\vec{F_L}$ is necessarily close to the normal to the ecliptic, at all times. Between 2014 and 2018, the ecliptic longitude of B1 changed by $\sim$60\degr, and the position angle of the ecliptic normal varied from 0\degr~to 20\degr. In contrast, the measured position angles of the north and south arms of the coma are stable to within the measurement uncertainties (mostly $\pm$1\degr~or $\pm$2\degr) in the interval 2014 - 2018, and certainly do not show any variation as large as 20\degr~(Table \ref{photometry}). In addition, we note that the optical (BVR) colors of B1 are redder than sunlight, typical of all reliably measured comets (Jewitt 2015), and so provide no evidence for the optically small (blue) particles that would be affected by Lorentz forces. \subsection{Concluding Remarks} The special morphology of B1 is an artifact of several effects. First, the dust particles must be ejected slowly, with a speed, $V$, comparable to the gravitational escape velocity from the nucleus, $V_e$. We write $V_e \sim 0.5 r_n$ (m s$^{-1}$), where $r_n$ is the nucleus radius expressed in kilometers and where we have assumed a nucleus density $\rho$ = 500 kg m$^{-3}$ (c.f.~Jorda et al.~2016, Kokotanekova et al.~2017). We set $V = V_e$. Then, the speed range $1 \lesssim V \le $ 10 m s$^{-1}$ corresponds to nucleus radii 2 $\le r_n \le 20$ km. While the radius of the nucleus of B1 has not been measured, values in this range are typical of well-measured cometary nuclei (Lamy et al.~2004). Second, the equatorial rotational velocity of the nucleus must be a significant fraction of $V_e$, in order to provide modulation of the ejection by latitude. For a sphere of density $\rho$, the critical period is $P = (3\pi/(G \rho))^{1/2}$ which, with $\rho$ = 500 kg m$^{-3}$ gives $P = 1.2\times 10^4$ s (3.3 hr). Somewhat longer critical rotation periods are possible for aspherical nuclei in rotation about a short axis. Third, we suppose that the Earth is located close to the projected rotational equator of the nucleus of B1. Our proposed model of B1 can thus be tested for consistency by determination of the nucleus size and by measurement of its rotation vector. Such measurements will require the use of high resolution imaging with the Hubble Space Telescope; our proposal to obtain such images was recently rejected. The time-integrated mass loss from B1 is not well constrained. To take an extreme case, we note that the rate at perihelion is $dM/dt \gtrsim$ 10 kg s$^{-1}$ and suppose that this rate has been sustained at all distances $r_H \le$ 30 AU. On its current orbit, B1 spends 30 years (10$^9$ s) with $r_H \le$ 30 AU, corresponding to a total mass loss of $\sim$10$^{10}$ kg. This compares with the mass of a 2 km to 20 km radius nucleus $M \sim$ 10$^{13}$ kg to 10$^{16}$ kg. A fractional mass loss of $\Delta M/M \sim$ 10$^{-3}$ to 10$^{-6}$ is unlikely to generate a measurable non-gravitational acceleration, or to materially change the spin rate of the nucleus. Moreover, while B1 presumably made previous journeys through the planetary region, it seems unlikely that prior mass loss has rivalled a significant fraction of the nucleus mass, or that outgassing torques have changed the spin of the nucleus. Instead, we imagine that the nucleus spin is primordial, accumulated as the nucleus grew, or in some collisional event occurring in the ancient protoplanetary disk. Whatever the origin of the spin, mass loss has been limited to those brief, near-perihelion periods when gas drag forces exceed gravitational and cohesive forces binding material to the nucleus. Finally, we ask why has the discus morphology not been noticed in other comets? The probable reason is that most observations in the published literature refer to comets much closer to the Sun than B1, usually with $r_H \le$ 5 AU and frequently with $r_H$ as small as 1 to 2 AU. In these comets, the equilibrium temperatures are higher and the sublimation fluxes (which depend exponentially on temperature) are orders of magnitude larger than in B1. As a result, the cohesion bottleneck (in which small particles are bound to the nucleus at large $r_H$), is broken, and small particles flood the coma (Gundlach et al.~2015, Jewitt et al.~2018). Because they are small, the ejected particles in near-Sun comets attain terminal velocities $V \gg V_e$, erasing the role of centripetal assistance. While the discus morphology is not expected (and has not been reported) in comets nearer the Sun, we predict that discus comets will become more common as future studies increasingly probe comets far from the Sun, particularly those discovered pre-perihelion and beyond the cohesion bottleneck (Jewitt et al.~2018). \clearpage \section{SUMMARY} We present a study of the high-perihelion, long-period comet C/2014 B1 (Schwartz) in the heliocentric distance range 11.9 AU to 9.6 AU. \begin{enumerate} \item The comet exhibits a unique and morphologically stable discus-shaped coma, whose orientation remains fixed even as the projected anti-solar and negative velocity vectors vary in response to the changing observing geometry. This shows that the coma morphology is an intrinsic property of the comet, not a product of external influences. \item The lack of a prominent dust tail implies a large mean particle radius, 0.1 $\le a \le$ 10 mm, too large to be significantly affected by radiation pressure. The depletion of smaller particles is tentatively attributed to inter-particle cohesion, whose effects dominate at small sizes. Even larger particles may be present, but are rare. \item The absolute brightness of the coma increases by $\sim$1 magnitude, proving that mass-loss is on-going at 10 AU. We infer a dust mass production rate, $dM/dt \gtrsim$ 10 kg s$^{-1}$, that can be sustained by equilibrium sublimation from exposed supervolatile ices covering as little as 0.3 km$^2$ (for CO) to 0.6 km$^2$ (for CO$_2$) of the surface. Water ice is too cold to sublimate at 10 AU, but crystallization of amorphous water ice might play a role in liberating trapped supervolatiles. \item The discus-shaped coma is consistent with preferential equatorial ejection of dust from a nucleus whose rotational pole lies close to the plane of the sky, and from which the ejection speed (1-10 m/s) is comparable to the escape speed. Centripetal assistance near the equator can then help launch large dust particles that cannot escape from high latitudes. A Monte Carlo simulation based on this scenario successfully matches the morphology. The implied nucleus radius is 2 - 20 km. \item Based on our model, we predict that the discus shape should be most common in cometary comae that are dominated by large, slow particles, since only these would give rise to a latitude modulation of the dust ejection. The effect is absent in near-Sun comets, because the strong gas flux launches smaller particles far above the escape speed. \end{enumerate} \acknowledgments Based in part on observations made with the Nordic Optical Telescope, operated by the Nordic Optical Telescope Scientific Association at the Observatorio del Roque de los Muchachos, La Palma, Spain, of the Instituto de Astrofisica de Canarias. Some of the data were obtained at the W. M. Keck Observatory, which is operated as a scientific partnership among the California Institute of Technology, the University of California and the National Aeronautics and Space Administration. The Observatory was made possible by the generous financial support of the W. M. Keck Foundation. Also based in part on data from Kitt Peak National Observatory, National Optical Astronomy Observatory, which is operated by the Association of Universities for Research in Astronomy (AURA) under a cooperative agreement with the National Science Foundation. We thank Jessica Agarwal, Jing Li, Man-To Hui, Quanzhi Ye and the anonymous referee for comments on the manuscript.

Influenza in pregnancy, particularly pandemic influenza, has been shown to cause more severe illness and hospitalization in pregnant women than in the nonpregnant population \[[@CIT0001], [@CIT0002]\]. Influenza vaccination is recommended by the World Health Organization (WHO) for pregnant women, although vaccination is not standard policy in many low and middle income countries \[[@CIT0003]\]. Reasons for this vary but include logistical and financial constraints, as well as a lack of evidence for influenza burden in some countries. Recent randomized trials of influenza vaccine in pregnancy in low income settings have demonstrated a protective effect against maternal and infant influenza \[[@CIT0004]\] as well as increased birthweight in South Asia \[[@CIT0007], [@CIT0008]\]. Maternal vaccination in pregnancy provides the potential for protecting the infant early in life, but evidence on timing of influenza vaccination in pregnancy on efficacy in infants is lacking \[[@CIT0009]\]. Maternal influenza vaccination earlier in pregnancy may provide protection in infants born preterm and potentially more time for protection of the mother against illness during pregnancy, thereby perhaps improving birth outcomes. However, vaccination in the third trimester may lead to higher levels of antibody transfer from the mother to the infant and hence improve protection against illness early in life. One study examining maternal and cord blood antibody response to the H1N1 vaccine at delivery found a lower level of antibodies for those vaccinated earlier in pregnancy \[[@CIT0010]\] but another study found comparable antibody levels \[[@CIT0011]\]. Neither examined protection from influenza in the infants. The South African maternal influenza vaccine trial found higher levels of transplacental transfer associated with a longer interval between vaccination and delivery in HIV uninfected women \[[@CIT0012]\]. In low resource settings, the number of antenatal care (ANC) visits is increasing, but early ANC attendance is still low. Therefore it would be useful for governments considering implementing a maternal influenza vaccination program to know whether the timing of vaccine delivery in pregnancy is important to increase efficacy in infants. We conducted an individually randomized population-based trial of maternal influenza vaccination in rural southern Nepal from April 2011 through September 2013 and estimated the efficacy of the vaccine among infants from birth through 6 months of age \[[@CIT0007]\]. We used the data from this trial to examine whether vaccine efficacy varied by gestational age at vaccination. METHODS {#s6} ======= Details of the trial methods are presented elsewhere \[[@CIT0007], [@CIT0013], [@CIT0014]\]. Briefly, a census of women of childbearing age (15--40 years) was conducted in 9 Village Development Committees (VDC) of rural Sarlahi district in the low lying plains of Nepal, bordering Bihar, India. Women who identified themselves as pregnant and of gestational age between 17 and 34 weeks were offered enrollment in the trial. Women not pregnant at that time were visited every 5 weeks and asked if they had had a period since the last visit. If yes, the date of last menstrual period (LMP) was recorded. If not, they were offered a pregnancy test and if positive were consented into the trial. Hence prevalent and incident pregnancies were enrolled in the trial. Women were enrolled in 2 distinct annual cohorts. Cohort 1 was enrolled from April 25, 2011, through April 24, 2012, and cohort 2 from April 25, 2012, through April 24, 2013. Pregnant women between 17 and 34 weeks gestation were individually randomized to receive the currently recommended trivalent inactivated influenza vaccine as early as possible in pregnancy within this gestational window. Randomization was conducted in blocks of 8, stratified by VDC and gestational age (17--25 weeks and 26--34 weeks) within VDC, based on date of LMP obtained from the 5 weekly surveillance. This resulted in balanced numbers of women in each stratum of gestational age, but with most women enrolled and vaccinated early in pregnancy (except for the prevalent pregnancies). To better address the question of whether timing of maternal vaccination altered efficacy, the number of women vaccinated later in pregnancy was increased in the second cohort (which were all incident pregnancies identified prior to 17 weeks gestation) by randomizing them to vaccine or placebo as in cohort 1 but randomly assigning the timing of vaccine receipt to be spread evenly over 17--34 weeks gestation. The assignment to vaccine or saline placebo continued to follow the randomization strategy used in the first cohort (blocked by 8, stratified by VDC and 17--25 vs 26--34 weeks gestation). The primary outcomes of the original trial were maternal influenza-like illness (ILI), laboratory confirmed influenza in infants, and low birth weight (LBW). There were multiple secondary outcomes including preterm birth, maternal laboratory confirmed influenza and others \[[@CIT0007], [@CIT0014]\]. To address whether vaccine timing in pregnancy alters efficacy, this secondary analysis examined the following outcomes: the rise in geometric mean of the hemagglutination-inhibition antibody (HAI) titers to influenza antigens obtained from serum samples taken prior to vaccination and within the first week postpartum, mean influenza HAI titers in cord blood, maternal to infant antibody transfer ratio, maternal seroconversion and seroprotection, laboratory confirmed influenza in women during pregnancy and through 180 days postpartum, laboratory confirmed influenza in infants from birth through 180 days of age, and prevalence of low birth weight (\<2500 g taken as soon after birth as possible but no later than 72 hours after birth). Two different vaccine formulations were used in the trial. These were the recommended vaccines for seasonal use in northern and southern hemispheres during those time periods. Between April 25, 2011, and October 15, 2012, a vaccine containing A/H3N2 Perth, A/H1N1 California, and B Brisbane (Victoria) was used, and from October 2012 to September 2013 the vaccine contained A/H3N2 Victoria, A/H1N1 California, and B Wisconsin (Yamagata). The first vaccine formulation covered all of the 1st annual cohort and part of the 2nd annual cohort trial periods. The study proposed to collect blood samples from a subset of participants to confirm seroprotection and seroconversion. Ultimately, a convenience blood sample of 208 women-infant pairs was collected prior to vaccination, at 1 week and 3 months postpartum, and cord blood was collected from the infant after delivery. HAI assays were used to test for antibodies to influenza antigens contained in the vaccines by a CLIA certified laboratory at Cincinnati Children's Hospital. Laboratory staff were blinded to the vaccine assignments and timing of vaccine. Maternal samples were collected prior to vaccination, at home visits postpartum, and cord blood was collected at facilities or home (if the delivery occurred there) and transported on ice to the field laboratory where they were stored in liquid nitrogen and later shipped to Cincinnati in dry shippers. Detailed laboratory methods are provided in the supplement to the primary results \[[@CIT0007]\]. Enrolled women were visited weekly from the time of vaccination through 6 months postpartum to assess morbidity (self-reported fever, persistent cough, sore throat, nasal congestion, and myalgia) in the past week. If positive for fever and at least any one of the other symptoms in the past week, a mid-nasal swab was collected, stored in PrimeStore® Molecular Transport Medium (Longhorn Diagnostics LLC, Bethesda, MD) and later analyzed using real-time polymerase chain reaction (RT-PCR) for the presence of influenza. Maternal laboratory confirmed influenza was defined as maternal report of fever and at least one of the following: cough, sore throat, nasal congestion, or myalgia concurrent with a PCR-positive nasal swab. Infants were visited weekly from birth, and mothers were asked whether the infant had fever, cough, wheeze, difficulty breathing, or ear discharge on each day in the past week. If the mother reported the infant had any of these symptoms on any day, a mid-nasal swab was collected and later analyzed using RT-PCR for the presence of influenza. Episodes of infant influenza were defined as any one of fever, cough, wheeze, difficulty breathing, or ear discharge on at least 1 day concurrent with a PCR-positive nasal swab. Episodes were considered distinct if separated by at least 7 symptom-free days. Other outcomes examined were influenza like illness in pregnancy and through 6 months postpartum in women, small-for-gestational-age (SGA), defined as having a birthweight less than the 10th percentile of the international Intergrowth-21 birthweight standard/reference at a specific gestational age \[[@CIT0015]\], preterm defined as \< 37 completed weeks gestation at birth, and mean difference between treatment groups in birthweight in grams and gestational age in weeks. HAI titers were logarithmically transformed and geometric mean titers reported. Seroconversion was defined as a pre-vaccination HAI titer of 1:10 or less and a post-delivery maternal titer of 1:40 or larger, or a pre-vaccination titer greater than 1:10 and an increase in titer of 4 or more post-vaccination. Seroprotection was defined as HAI antibody titer ≥ 1:40. Incidence rates for influenza were calculated by dividing the number of cases of illness by the number of days at risk by treatment group and compared using incidence risk ratios and 95% confidence intervals from mixed effects binomial regression models with log link to account for correlation between repeated morbidity measures. Days with symptoms and the subsequent 7 days were excluded from days at risk. The differences in mean birth weights and gestational ages between treatment groups were calculated using linear regression. Prevalence of LBW, preterm birth, and SGA by treatment group were compared using risk ratios and 95% confidence intervals from Poisson regression models with a log link and robust variance. Verbal informed consent was obtained from all women who participated in the trial. The study was approved by the Institutional Review Boards (IRB) of the Nepal Health Research Council, the Institute of Medicine at Tribhuvan University, the Johns Hopkins Bloomberg School of Public Health (JHBSPH), and Cincinnati Children's Hospital. IRBs at Seattle Children's Hospital, the University of Washington, and George Washington University granted oversight to the IRB at JHBSPH. The trial was registered with ClinicalTrials.gov (NCT01034254). Funding for the study was provided by the Bill and Melinda Gates Foundation (grant 50274). RESULTS {#s7} ======= A total of 3693 women were consented and vaccinated between April 2011 and September 2013, resulting in 3646 live births (2090 women, 2063 live births in cohort 1, 1603 women, 1583 live births in cohort 2). There was no difference between the placebo and vaccinated groups of women on a variety of demographic, socioeconomic, morbidity, and reproductive history characteristics at the time of enrollment, stratified by timing of vaccination in pregnancy (17--25 and 26--34 weeks) ([Supplemental Table 1](#sup1){ref-type="supplementary-material"}). The incidence rate ratio (IRR) was not statistically significantly different from one for laboratory confirmed influenza between infants of women vaccinated earlier (IRR: 0.73 (95% confidence interval \[CI\]: 0.51, 1.05) or later in pregnancy (IRR: 0.63 (95% CI: 0.37, 1.08) in the cohorts combined, nor in each cohort ([Table 1](#T1){ref-type="table"}). The point estimates for the infant incidence rate ratios were more protective in the second cohort, which is consistent with greater protection in this cohort due to a better match between the vaccine and circulating virus, but there was no difference in protection by timing of vaccination in either cohort. The reduction in low birth weight in the combined cohorts was 17% among infants of mothers vaccinated between 17 and 25 weeks, although it was reduced by 10% in those vaccinated between 26 and 34 weeks, but this difference was not statistically significant ([Table 1](#T1){ref-type="table"}). ###### Infant Influenza Incidence and Prevalence of Low Birth Weight by Cohort and Treatment Group Cohort 1 Cohort 2 Combined ------------------------------- ------------------- ------------------- ------------------- ------------------- ------------------- ------------------- **Infant Influenza**  Vaccine cases (person-years) 44 (255.6) 16 (79.6) 8 (129.9) 6 (117.2) 52 (385.5) 22 (196.8)  Placebo cases (person-years) 50 (266.8) 19 (74.7) 21 (138.3) 15 (117.9) 71 (386.1) 34 (192.5)  Vaccine incidence^a^ 172.2 201.0 61.6 51.2 134.9 111.8  Placebo incidence^a^ 201.8 254.4 151.8 127.3 183.9 176.6  IRR (95% CI) 0.85 (0.56, 1.29) 0.79 (0.41, 1.53) 0.41 (0.18, 0.91) 0.40 (0.16, 1.02) 0.73 (0.51, 1.05) 0.63 (0.37, 1.08) **Low Birthweight**  Vaccine n/N (%) 142/606 (23.4) 51/201 (25.4) 63/296 (21.3) 59/277 (21.3) 205/902 (22.7) 110/478 (23.0)  Placebo n/N (%) 164/584 (28.1) 55/195 (28.2) 79/305 (25.9) 66/276 (23.9) 243/889 (27.3) 121/471 (25.7)  RR (95% CI) 0.83 (0.69, 1.01) 0.90 (0.65, 1.25) 0.82 (0.61, 1.10) 0.89 (0.65, 1.21) 0.83 (0.71, 0.98) 0.90 (0.72, 1.12) Abbreviations: CI, confidence interval; IRR, incidence rate ratio; RR, relative risk. ^a^Incidence: cases/1000 person-years. The geometric mean of the influenza HAI titers in cord blood was higher in infants of mothers vaccinated later than earlier in pregnancy for the first vaccine formulation used (245.0 (95% CI: 6.0, 962.2) versus 137.3 (95% CI: 83.2, 226.4) for A/H1N1/California, 43.1 (95% CI: 22.4, 82.8) versus 210.9 (95% CI: 77.5, 574.0) for A/H3N2/Perth and 94.3 (95% CI: 56.4, 157.7) versus 160.0 (95% CI: 93.9, 272.6) for B/Brisbane (V)), although these differences were not statistically significant ([Table 2](#T2){ref-type="table"}). With the second vaccine formulation used in the trial, geometric mean HAI titers were higher in infants whose mothers were vaccinated later in pregnancy, except for the A/H1N1/California antigen (198.3 (95% CI: 138.7, 283.4) versus 284.0 (95% CI: 195.5, 412.5)), but none of these differences were statistically significant. There were no statistically significant differences between early and later vaccination in pregnancy in the rise in geometric mean HAI titers from prevaccination to within 1 week postpartum in women, nor in the cord:maternal antibody transfer ratios ([Table 2](#T2){ref-type="table"}). Maternal seroconversion and seroprotection rates did not differ by timing of vaccination for the 2 vaccine formulae or specific antigens, but the sample sizes were small ([Table 2](#T2){ref-type="table"}). These comparisons included women in the placebo (N = 32) and vaccine (N = 20) groups who contracted influenza, which would have increased their antibody titers through natural infection. ###### Geometric Mean Hemagglutination-Inhibition Antibodies (HAI) to Influenza Antigens in Cord Blood, Rise in Geometric Mean HAI Antibodies From Pre-vaccination to Within 1 Week Postpartum in Women, Infant-Mother Transfer Ratio, Maternal Seroprotection and Seroconversion by Vaccine Type and Timing of Vaccination Vaccine 1 A/H1N1/California A/H3N2/Perth B/Brisbane (V) ----------------------------- ------------------------- ------------------------- ------------------------- ------------------------- ------------------------- -------------------- **Cord blood**  17--25 weeks 137.3 (83.2, 226.4) 19.8 (11.5, 33.9) 43.1 (22.4, 82.8) 14.0 (8.8, 22.2) 94.3 (56.4, 157.7) 24.6 (15.8, 38.4)  26--34 weeks 254.0 (67.0, 962.2) 18.3 (6.4, 52.3) 210.9 (77.5, 574.0) 14.1 (4.8, 41.4) 160.0 (93.9, 272.6) 18.3 (8.4, 40.2) **Maternal increase**  17--25 weeks 7.8 (4.3, 14.3) 0.9 (0.8, 1.2) 2.0 (1.4, 3.0) 1.0 (0.7, 1.4) 6.1 (4.2, 8.7) 1.7 (1.3, 2.3)  26--34 weeks 10.9 (4.1, 28.7) 1.3 (0.7, 2.4) 3.7 (1.0, 13.3) 1.4 (1.0, 1.9) 4.7 (1.7, 13.0) 1.8 (1.1, 3.0) **Infant-mother ratio**  17--25 weeks 1.0 (0.8, 1.4) 1.0 (0.8, 1.3) 0.9 (0.7, 1.0) 0.6 (0.5, 0.8) 0.9 (0.7, 1.1) 0.5 (0.4, 0.8)  26--34 weeks 1.0 (0.4, 2.3) 0.9 (0.6, 1.3) 0.7 (0.6, 0.9) 0.6 (0.5, 0.9) 1.1 (0.6, 1.9) 0.3 (0.2, 0.4) **Seroprotection** **n/N (%**) **n/N (%**) **n/N (%**) **n/N (%**) **n/N (%**) **n/N (%**) Pre-vaccine maternal  17--25 weeks 5/20 (25.0) 9/24 (37.5) 8/20 (40.0) 8/24 (33.3) 4/20 (20.0) 8/24 (33.3)  26--34 weeks 3/9 (33.3) 3/8 (37.5) 6/9 (66.7) 3/8 (37.5) 4/9 (46.4) 4/8 (50.0) **Post-vaccine maternal**  17--25 weeks 18/20 (90.0) 7/24 (29.2) 13/20 (65.0) 8/24 (33.3) 20/20 (100.0) 16/24 (66.7)  26--34 weeks 8/9 (88.9) 4/8 (50.0) 8/9 (88.9) 3/8 (37.5) 9/9 (100.0) 7/8 (87.5) **Maternal seroconversion**  17--25 weeks 14/20 (70.0) 1/24 (4.2) 4/20 (20.0) 0/24 (0.0) 17/20 (85.0) 4/24 (16.7)  26--34 weeks 6/9 (66.7) 1/8 (12.5) 3/9 (33.3) 0/8 (0.0) 6/9 (66.7) 1/8 (12.5) **Vaccine 2** A/H1N1/California A/H3N2/Victoria B/Wisconsin (Y) Vaccine Placebo Vaccine Placebo Vaccine Placebo Geometric Mean (95% CI) Geometric Mean (95% CI) Geometric Mean (95% CI) Geometric Mean (95% CI) Geometric Mean (95% CI) Geometric Mean (95% CI) **Cord blood**  17--25 weeks 284.0 (195.5, 412.5) 26.2 (13.3, 51.6) 284.6 (183.7, 440.9) 55.3 (30.4, 100.5) 274.5 (198.2, 380.2) 80.0 (53.5, 119.6)  26--34 weeks 198.3 (138.7, 283.4) 27.9 (18.7, 41.6) 319.3 (229.9, 443,5) 38.4 (26.0, 56.7) 372.6 (277.9, 499.5) 66.9 (51.7, 86.6) **Maternal increase**  17--25 weeks 5.4 (3.3, 8.9) 0.8 (0.6, 1.1) 6.9 (4.7, 10.3) 1.2 (0.6, 2.1) 2.5 (2.1, 3.1) 1.0 (0.7, 1.3)  26--34 weeks 6.3 (4.4, 9.0) 1.0 (0.9, 1.1) 10.4 (7.4, 14.8) 1.4 (1.0, 2.0) 2.5 (1.9, 3.3) 1.2 (1.0, 1.4) **Infant-mother ratio**  17--25 weeks 1.6 (1.1, 2.4) 1.2 (0.8, 1.8) 1.2 (0.9, 1.5) 0.9 (0.6, 1.6) 0.8 (0.6, 1.0) 0.5 (0.3, 0.8)  26--34 weeks 1.2 (0.9, 1.5) 1.1 (1.0, 1.3) 1.3 (1.1, 1.5) 0.8 (0.6, 1.0) 0.8 (0.6, 1.0) 0.4 (0.3, 0.5) **Seroprotection** n/N (%) n/N (%) n/N (%) n/N (%) n/N (%) n/N (%) **Pre-vaccine maternal**  17--25 weeks 15/29 (51.7) 11/27 (40.7) 16/29 (55.2) 16/27 (59.3) 28/29 (96.6) 27/27 (100.0)  26--34 weeks 22/49 (44.9) 20/42 (47.6) 23/49 (46.9) 23/42 (54.8) 48/49 (98.0) 41/42 (97.6) **Post-vaccine maternal**  17--25 weeks 27/29 (93.1) 8/27 (29.6) 29/29 (100.0) 18/27 (66.7) 29/29 (100.0) 27/27 (100.0)  26--34 weeks 46/49 (93.9) 20/42 (47.6) 48/49 (98.0) 27/42 (64.3) 49/49 (100.0) 42/42 (100.0) **Maternal seroconversion**  17--25 weeks 16/29 (55.2) 0/27 (0.0) 22/29 (75.9) 4/27 (14.8) 11/29 (37.9) 2/27 (7.4)  26--34 weeks 35/49 (71.4) 1/42 (2.4) 40/49 (81.6) 6/42 (14.3) 22/49 (44.9) 2/42 (4.8) Abbreviation: CI, confidence interval. Corresponding to the point estimates of the reduction in LBW being greater in infants of women who were vaccinated earlier in pregnancy, the mean difference in birth weight between the infants of vaccinated and placebo mothers was greater for those vaccinated earlier in pregnancy ([Table 3](#T3){ref-type="table"}). However, none of these differences between early and late vaccination were statistically significant. Similarly, there were no statistically significant differences in a variety of other secondary outcomes (ILI in women and infants, maternal laboratory confirmed influenza, SGA, preterm, stillbirths, infant mortality, and congenital defects) ([Table 3](#T3){ref-type="table"}). ###### Incidence Rate Ratios, Relative Risks, and Mean Differences Between Vaccine and Placebo for Other Outcomes of the Maternal Influenza Vaccination by Gestational Age at Vaccination Cohort 1 Cohort 2 Combined --------------------------------- --------------------- -------------------- -------------------- --------------------- --------------------- -------------------- **IRR (95% CI**) **IRR (95% CI**) **IRR (95% CI**) **IRR (95% CI**) **IRR (95% CI**) **IRR (95% CI**) ILI pregnancy 0.85 (0.60, 1.21) 0.58 (0.22, 1.53) 0.57 (0.33, 1.00) 0.65 (0.30, 1.42) 0.76 (0.56, 1.02) 0.62 (0.33, 1.15) ILI postpartum 0.96 (0.64, 1.43) 1.13 (0.64, 2.02) 0.82 (0.44, 1.55) 0.53 (0.25, 1.14) 0.92 (0.66, 1.30) 0.86 (0.54, 1.37) All ILI 0.89 (0.67, 1.18) 0.95 (0.58, 1.57) 0.69 (0.44, 1.08) 0.58 (0.33, 1.03) 0.83 (0.65, 1.05) 0.77 (0.52, 1.12) Flu pregnancy 0.37 (0.13, 1.03) 0.94 (0.13, 6.68) 1.31 (0.41, 4.18) 1.57 (0.38, 6.52) 0.62 (0.30, 1.31) 1.32 (0.42, 4.14) Flu postpartum 0.52 (0.20, 1.35) 1.46 (0.25, 8.71) NA 0.20 (0.02, 1.67) 0.62 (0.25, 1.54) 0.56 (0.16, 1.90) All flu 0.44 (0.22, 0.88) 1.21 (0.33, 4.47) 1.50 (0.49, 4.58) 0.73 (025, 2.07) 0.62 (0.35, 1.10) 0.89 (0.39, 2.00) ILI infants 1.04 (0.86, 1.27) 1.01 (0.75, 1.37) 0.73 (0.55, 0.96) 1.04 (0.79, 1.36) 0.92 (0.78, 1.08) 1.03 (0.84, 1.26) **RR (95% CI**) **RR (95% CI**) **RR (95% CI**) **RR (95% CI**) **RR (95% CI**) **RR (95% CI**) Preterm 0.97 (0.77, 1.24) 0.84 (0.51, 1.38) 0.94 (0.67, 1.34) 0.79 (0.52, 1.20) 0.97 (0.79, 1.18) 0.81 (0.59, 1.11) SGA 0.86 (0.73, 1.00) 0.97 (0.77, 1.24) 0.95 (0.75, 1.20) 1.05 (0.83, 1.32) 0.88 (0.78, 1.01) 1.02 (0.86, 1.20) Stillbirths 1.35 (0.62, 2.92) 0.72 (0.23, 2.23) 0.92 (0.34, 2.51) 1.16 (0.36, 3.78) 1.18 (0.63, 2.21) 0.90 (0.40, 2.03) Infant deaths 0.93 (0.47, 1.83) 1.00 (0.36, 2.82) 2.22 (1.02, 4.84) 0.96 (0.51, 1.82) 1.36 (0.82, 2.24) 0.98 (0.57, 1.69) Congenital defects 0.86 (0.31, 2.36) 2.94 (0.60, 14.42) 0.86 (0.23, 3.17) 0.97 (0.20, 4.77) 0.86 (0.39, 1.90) 1.75 (0.59, 5.20) **Mean (95% CI**) **Mean (95% CI**) **Mean (95% CI**) **Mean (95% CI**) **Mean (95% CI**) **Mean (95% CI**) Birthweight difference (g) 34.2 (−17.5, 85.8) 10.9 (−77.5, 99.2) 77.1 (5.9, 148.2) 45.6 (−30.5, 121.7) 47.7 (5.9, 89.6) 30.5 (−27.4, 88.4) Gestational age difference (wk) −0.07 (−0.34, 0.21) 0.15 (−0.26, 0.55) 0.04 (−0.29, 0.37) 0.29 (−0.05, 0.63) −0.03 (−0.25, 0.18) 0.23 (−0.03, 0.49) Abbreviations: CI, confidence interval; g, grams; ILI, influenza-like illness; IRR, incidence rate ratios; NA, not applicable; RR, relative risks; SGA, small-for-gestational-age; wk, week. DISCUSSION {#s8} ========== The primary outcomes of the original trial, maternal ILI, infant laboratory-confirmed influenza, and LBW were significantly reduced by maternal vaccination across the 2 cohorts combined. In this subanalysis of the maternal influenza vaccination trial, we did not find any statistical evidence that the impact of the vaccine on maternal and infant antibody responses or health outcomes differed by the timing of vaccination during pregnancy. There was some suggestion from point estimates that antibody levels in cord blood were higher for most antigens in the vaccine among infants whose mothers were vaccinated later in pregnancy, but the numbers of samples tested for antibodies was small, limiting the power to detect meaningful differences. Additionally, there was a somewhat higher birth weight among infants whose mothers were vaccinated earlier in pregnancy, but none of these differences between early and late vaccination were statistically significant. Many studies have compared first to second and third trimester safety outcomes for influenza vaccination, but few studies have compared efficacy outcomes between second and third trimester influenza vaccination. An observational study in Australia found a 33% reduction in hospitalizations for respiratory illness among newborns of women vaccinated in the third trimester of pregnancy but no effect if vaccinated earlier in pregnancy \[[@CIT0016]\]. An observational study in Nicaragua found reductions in preterm birth if vaccinated in the second and third trimesters of 13% and 34%, respectively, and a 20% and 36% reduction in LBW in infants of women vaccinated in the second and third trimester, respectively \[[@CIT0017]\]. A retrospective analysis of 7 Vaccine Safety Datalink sites in the United States did not find any association between vaccine exposure and preterm and SGA overall, nor by trimester \[[@CIT0018]\]. Sperling et al. examined the immunogenicity of influenza vaccine in pregnancy and found seroconversion rates to be similar in the second trimester (13 to \<28 weeks) compared to early third trimester (28 to \<34 weeks) \[[@CIT0019]\]. We were unable to identify any randomized trials that have examined efficacy of infant health outcomes by timing of influenza vaccination, although we identified 2 randomized trials that examined antibody response to pandemic H1N1 vaccination in cord blood by gestational age at vaccination \[[@CIT0010], [@CIT0011]\]. One found lower antibody response at birth for infants of mothers vaccinated earlier in pregnancy \[[@CIT0010]\], but the other trial did not \[[@CIT0011]\]. Recent prospective observational studies of tetanus-diphtheria-acellular pertussis (Tdap) immunization in pregnancy have evaluated the effect of second versus third trimester vaccination on cord blood antibodies and infant seropositivity rates \[[@CIT0020], [@CIT0021]\]. These studies found that early second trimester Tdap immunization increases neonatal antibodies, suggesting that for Tdap, a broader immunization window would maximize infant protection from pertussis in both full-term and preterm infants. Physiogically, transplacental antibody transfer occurs starting in the second trimester but increases significantly in the third trimester. Therefore, it would be expected that antibody titers may be higher in the cord blood of infants whose mothers were vaccinated later in pregnancy. However, these infants have had the benefit of higher maternal antibody levels and maternal protection from illness for less time during pregnancy, perhaps explaining the slightly higher impact of second trimester influenza vaccination on mean birth weights and greater reduction in LBW relative to the vaccine's impact in the third trimester. Strengths of this study include a large individually randomized population-based controlled trial design with vaccinations occurring over 2 annual cohorts. In addition, the design included randomization stratified on timing of vaccination during pregnancy, allowing this analysis to retain the advantages of the randomized design within these strata. The availability of antibody measurements to the vaccine antigens was also helpful in further explaining the health outcome findings. Limitations of the study include self-report of morbidity, although influenza in mothers and infants was confirmed through analysis of nasal swabs by RT-PCR. However, maternal influenza was only confirmed for febrile illness, whereas the infant influenza was confirmed on the basis of any ILI symptom. Although there was some indication from point estimates of differences in the impact of timing of vaccination on cord blood antibodies and birth weight, these difference were not statistically significant. A larger sample size might have been able to strengthen the evidence for these differences, especially with respect to outcomes HAI antibody titers. However, the actual differences are not of that much public health importance. Another limitation is the lack of gold standard gestational age measurements using ultrasound. However, the 5-weekly surveillance for pregnancy and confirmation by pregnancy test reduced the recall period of date of LMP. This approach to gestational age estimation was recently validated and found quite accurate against early ultrasound measurements in a similar population in Bangladesh \[[@CIT0022]\]. The public health importance of these findings is that vaccination in either the second or third trimester of pregnancy is equally likely to provide protection from influenza. In many low income settings where women often present for antenatal care infrequently or late influenza immunization administered anytime in the 2nd or 3rd trimester of pregnancy could provide protection. This makes the logistics of implementing a maternal influenza immunization strategy more feasible in resource-limited settings. Supplementary Data {#s9} ================== Supplementary materials are available at *Clinical Infectious Diseases* online. Consisting of data provided by the authors to benefit the reader, the posted materials are not copyedited and are the sole responsibility of the authors, so questions or comments should be addressed to the corresponding author. ###### Click here for additional data file. ***Authors' Contributions***. J. K., M. C. S., J. M. T., J. A. E., and J. K. contributed to the design of the study. M. C. S. conceived of the study and secured funding for the project. S. K. K., S. C. L., L. S., J. M. T., L. C. M., and J. K. supervised the conduct of the study in the field. J. K. drafted the manuscript, served as the study statistician, and J. K. and L. C. M. designed/implemented the data analysis. H. Y. C. assisted in study implementation and cord blood collection. N. K. provided data analysis support. All authors read and approved the final manuscript. ***Disclaimer.*** The funding agency had no role in the design, conduct, analysis, and interpretation of the findings, or in the decision to publish. ***Financial support.*** This work was supported by a grant from the Bill and Melinda Gates Foundation (grant OPP50274). ***Potential conflicts of interest.*** H. Y. C. has received research support from GSK, Sanofi Pasteur, and Novavax. J. E. has been a consultant for Pfizer, a member of a Data Safety Monitoring Board for GlaxoSmithKline (GSK) influenza antiviral studies, and her institution has received research support for clinical studies from GSK, Gilead, Chimerix, and Roche. All authors have submitted the ICMJE Form for Disclosure of Potential Conflicts of Interest. Conflicts that the editors consider relevant to the content of the manuscript have been disclosed.

\section{Introduction} With an isotropic equivalent energy release on the order of $10^{52} - 10^{54}$ ergs, Gamma Ray Bursts (GRBs) are among the most energetic events in the Universe besides the Big Bang. GRBs can be separated into two classes: short and long bursts. Long bursts, with durations longer than 2 seconds \citep[e.g.][]{chryssa93} are associated with the collapse of a very massive star and the formation of a black hole \citep{woosley93}. Short bursts are thought to be the result of a neutron star (NS) - NS or NS - black hole merger \citep[e.g][]{eichler89, paczynski91}. The leading theoretical model for GRBs and their afterglows is the fireball model \citep[see ][ and references within]{meszaros97, sari98, zhang04} in which the GRB is produced by internal shocks in a relativistic fireball, while the afterglow is produced in external shocks which are created when the fireball encounters the ambient external medium. The multi-wavelength mission {\it Swift} \citep{gehrels04} was launched on 2004-November-20 in order to hunt for GRBs. It is in low-Earth orbit at an altitude of 600 kms. The \swift\ observatory is equipped with three telescopes: a) the Burst Alert Telescope \citep[BAT, ][]{barthelmy05}, b) the X-Ray Telescope \citep[XRT, ][]{burrows05}, and c) the UV-Optical Telescope \citep[UVOT, ][]{roming05}. The BAT is a coded mask experiment and operates in the 15-350 keV energy range. The XRT detector is a copy of the MOS CCDs used on-board XMM \citep{tur01}. It operates between 0.3-10 keV in three observing modes, Photon Counting (PC) which is equivalent to the full-frame mode on XMM, Windowed Timing (WT), and Low-Rate Photo-diode mode (LrPD) which is only used for extremely bright sources \citep{hill04}. The UVOT is a sister instrument to XMM's Optical Monitor \citep[OM, ][]{mason01} and includes a similar set of filters to the OM \citep{mason01, roming05}. As reported by \citet{nousek05}, \swift\ has observed a general behavior of GRB afterglow X-ray light curves: a fast decay with slope $\alpha_1$ in the first hundred seconds is followed by a much shallower decay slope $\alpha_2$. This continues over a span of several thousands of seconds after the burst and is followed by a steeper decay slope $\alpha_3$ \citep[see also ][]{bing05}. The mean decay slopes of the 27 afterglows discussed in \citet{nousek05} are: $\alpha_1$=3.38\plm1.27, $\alpha_2$=0.76\plm0.34, and $\alpha_3$=1.33\plm0.32. In this paper we report the \swift\ observations of GRB\,050603. The paper is organized as follows: In \S\ref{observe} we describe the observations and the data reduction and analysis, in \S\ref{results} we present the results which are then discussed in \S\ref{discuss}. Throughout the paper decay and energy spectral indices $\alpha$ and $\beta$ are defined as $F_{\nu}(t,\nu)\propto (t-t_0)^{-\alpha}\nu^{-\beta}$ with $t_0$ the trigger time of the burst. Luminosities are calculated assuming a $\Lambda$CDM cosmology with $\Omega_{\rm M}$=0.27, $\Omega_{\Lambda}$=0.73 and a Hubble constant of $H_0$=71 km s$^{-1}$ Mpc$^{-1}$ using the luminosity distances given by \citet{hogg99}. All errors are 1$\sigma$ unless stated otherwise. \section{\label{observe} Observations and data reduction} GRB\,050603 was detected by the BAT on 2005-June-03 at 06:29:05 UT \citep{retter05} with the trigger ID 131560. Due to engineering tests of the \swift\ satellite, the narrow-field instruments, UVOT and XRT, were unable to observe the GRB afterglow until about 9.5 and 11 hours after the burst, respectively. The UVOT began observing at 15:42:59 UT \citep{brown05a}. The observations were made in the V filter with exposure time ratios of 1:8:1 per observing window so that the entire observation is not ruined if there is high background from the earth limb at the beginning or end of the observation. Only the middle exposures were used, to eliminate the problems of high background. The NASA \swift\ Data Center (SDC)-processed sky images were astrometrically aligned with respect to an image of the field from the Digitized Sky Survey. Aperture photometry was performed using a 6$''$ source aperture and a background annulus with inner and outer radii of 14$''$ and 30$''$ respectively. Once the afterglow in individual observations had fallen below 3$\sigma$ above background they were co-added to bring the S/N ratio above a 3$\sigma$ detection. The GRB\,050603 observations by the UVOT also revealed an infrequent problem in which the images do not contain events from the whole exposure time indicated in the header. Extension 4 of Sequence 00131560001 reported an exposure time of 1500s. However, a comparison of the count rate measured by the instrument and the counts in the image indicated that the image contained only 110 seconds of data. Thus only 7.3\% of the data were actually recorded. This resulted in an erroneous magnitude being reported in Brown et al. (2005a), as pointed out by Berger (2005). We have corrected this by comparing the count rate as measured by the UVOT with the counts in the actual image to calculate the exposure time. When the header and calculated exposure times differed by more than 5\%, the exposure keyword was changed to the calculated value. The software bug causing this problem was fixed on 2005-September-14 and data taken before that date will be checked and corrected in the archive. The XRT started to take data at 17:19:27 UT \citep{racusin05}. GRB050603 was observed over a period of about two weeks for a total of about 220 ks. The detailed observation log is listed in Table~\ref{obs_log}. All observations were performed in PC mode. The XRT data were reduced by the {\it xrtpipeline} software version 0.9.9. Source photons were selected by {\it XSELECT} in a circular region with a radius of r=47$''$ and the background photons in a circular region close by with a radius r=96$''$. For the spectral data events with grades 0-12 were selected with {\it XSELECT}. The spectral data were re-binned by {\it grppha} 3.0.0 having at least a minimum of 20 photons per bin. The spectra were analyzed by {\it XSPEC} version 11.3.2. The auxiliary response files were created by {\it xrtmkarf} and the standard response matrix swxpc0to12\_20010101v007.rmf was used. Background-subtracted X-ray light curves in the 0.3-10.0 keV energy range were constructed using the ESO Munich Image Data Analysis Software {\it MIDAS} (version 04Sep). The binning was dynamically performed. At the beginning of the observations the binning was set to 50 photons per bin while at later times it was reduced to 10 photons per bin as listed in Table\,\ref{obs_log}. Also at later times the source extraction radius was reduced to 10 pixels (corresponding to r=23.6$''$) in order to avoid confusion by the background. The X-ray light curve was fitted by power law and broken power law models in {\it XSPEC}. The count rates were converted into an unabsorbed flux by an energy conversion factor (ECF) using a power law fit with the absorption column density fixed to the Galactic value \citep[1.2$\times 10^{20}$ cm$^{-2}$, ][]{dic90}. Only one ECF with ECF=3.76$\times 10^{-11}$ ergs s$^{-1}$ cm$^{-2}$ count$^{-1}$ was applied for the whole light curve. As shown below in \S\ref{xrayspec}, there is no obvious change of the spectral parameters between early and later observations. \section{\label{results} Results} \subsection{Positions} All positions given for GRB\,050603 are listed in Table~\ref{position}. Figure~\ref{050603_error} displays the UVOT V-filter image of the GRB\,050603 field with the BAT and XRT error circles superimposed as given in Table~\ref{position}. The XRT position given in Table~\ref{position} is corrected for the XRT boresight offset \citep{moretti05} and differs slightly from the positions given by \citet{grupe05} and \citet{racusin05}. This new XRT position is in excellent agreement with the UVOT, optical, and radio positions \citep{brown05a, berger05a, cameron05}. \subsection{BAT} The BAT mask-weighted light curve (Figure~\ref{050603_bat_lc}) shows three fast-rise-exponential-decay (FRED) like spikes with peaks at 2.7 and 0.85 s before the trigger and 0.15 s afterwards. Each spike had a width of 0.6 s FWHM. The left panel of Figure~\ref{050603_bat_lc} displays the light curve of the whole 15-350 keV energy band. The right panel shows the light curve split into four energy bands: 15--25~keV, 25--50~keV, 50--100~keV, and 100--350~keV. The BAT light curve shows a harder spectrum during the spikes compared to the BAT observation after the spikes (Figure~\ref{050603_bat_lc_color}). The time-averaged spectrum between $T_0$-3s and $T_0$+18s in the 15-150 keV band shown in Figure~\ref{050603_bat_spec} is well fitted by a single power law with an energy spectral slope $\beta_{\gamma}=0.17^{+0.07}_{-0.08}$ ($\chi^2/\nu$=51/57). The spectrum is background subtracted. It has the standard 80-channel BAT energy binning. GRB\,050603 had $T_{90}$=12\plm2s which classifies it as a long burst. The fluence in the observed 15--150~keV band was (7.6\plm0.3)$\times10^{-6}$ ergs cm$^{-2}$. Based on the redshift z=2.812 \citep{berger05b}, the k-corrected rest-frame 100--500~keV and 20~keV--2~MeV total isotropic equivalent energies $E_{\rm \gamma, iso}$=3.2$\times10^{53}$ ergs and $E_{\rm \gamma, iso}$=1.26$\times10^{54}$ ergs, respectively, are some of the largest measured among all GRBs detected by \swift\ \citep{nousek05}. The total energies are comparable to other high-energetic pre-\swift\ bursts such as GRB\,990123 \citep{briggs99, corsi05}, GRB\,000131 \citep{andersen00}, or GRB\,010222 \citep{zand01}. \subsection{XRT data} \subsubsection{\label{xrt_lc} XRT light curve} Figure~\ref{050603_lc} displays the 0.3-10 keV flux X-ray light curve of GRB\,050603. The initial count rate at the beginning of the XRT observation was 0.06 counts s$^{-1}$ which converts to a 0.3-10.0 keV unabsorbed flux $F_{\rm X}=3\times10^{-12}$ ergs s$^{-1}$ cm$^{-2}$. The 2.0-10.0 keV unabsorbed flux was $F_{\rm X}=2\times10^{-12}$ ergs s$^{-1}$ cm$^{-2}$. The count rate was low enough that the data are not affected by pileup. The decay slope, derived from the 0.3-10.0 keV flux light curve shown in Figure~\ref{050603_lc}, is unusually steep with $\alpha=1.76^{+0.15}_{-0.07}$. This makes GRB\,050603 a relatively rapidly fading afterglow in its late phase \citep{nousek05}. The data are consistent with one simple power law throughout the observation with a $\chi^2/\nu$=14.3/17. Just before the detection of GRB\,050603 the XRT MOS CCD was hit by a micro-meteorite which caused severe damage to columns 294 and 320. These columns (and several adjacent ones) have been disabled on-board and are not useable. If the PSF of a source overlaps these bad columns, the measured flux can be incorrect. Since the source position on the detector changes with each orbit, this can lead to errors in the light curve. We have verified that the GRB was not positioned on the bad columns during this observation. \subsubsection{\label{xrayspec} X-ray spectral analysis} Table~\ref{spectral} lists the results of the spectral analysis of the XRT data. A simple power law model with Galactic absorption fits the spectra well for the segments 001 and 002 (Table\,\ref{obs_log}) data, resulting in X-ray energy spectral slopes $\beta_{\rm X}$=0.80\plm0.17 and 0.62\plm0.13 for segments 001 and 002, respectively. The simultaneous fit to the segment 001 and 002 data is shown in Figure~\ref{050603_spec}. The energy ranges of the two spectra are different due to the S/N and teh binning of the channels using {\it grppha}. This fits results in an X-ray spectra slope $\beta_{\rm X}$=0.71\plm0.10. In all cases, no additional intrinsic absorption is required. Leaving the absorption column density as a free parameter results in an absorption column density which is significantly below the Galactic value. As shown in Figure~\ref{050603_spec} there are some apparent residuals around 0.5 and 2.0 keV. These features are due to systematic errors in the still ongoing calibration of the auxiliary response file \citep[see the XRT calibration document XRT-OAB-CAL-ARF-v3\footnote{ The calibration document XRT-OAB-CAL-ARF-v3 can be found under: http://swift.gsfc.nasa.gov/docs/heasarc/caldb/swift/docs/xrt/index.html} and][]{romano05}. \subsection{UVOT} The V magnitudes for GRB\,050603 are listed in Table 4 and the V-band light curve is shown in Figure~\ref{050603_lc_xrt_uvot}. The first measurement, V=18.2 at 9.5 hours after the burst, is brighter than any other optical \swift\ afterglow at the same time, many of which are not optically detected at all (see e.g. Roming et al. 2005b), and is among the brightest of all optical afterglows with the exception of GRB\,030329 \citep[see e.g. ][]{berger05c}. The X-ray and optical decay slopes are similar, with the UVOT points decaying with a power law slope of 1.8\plm0.2 with $\chi^2/\nu$=18/7. There is no indication of a break near 12 hours as suggested by Berger \& Becker (2005), though the errors and wide temporal sampling do not allow a definitive statement. The steepness of the decay indicates that the break may have already occurred before our observations began, which would put a putative break at less than T+9 hours post-burst. The large value of $\chi^2/\nu$ indicates that the deviations from the simple power law may be real, though the largest deviations are only 2 $\sigma$. Li et al. (2006), in an independent analysis of the UVOT data, also concluded that this afterglow exhibited real fluctuations from a powerlaw decay. The power law decay reported by Li et al. (2006) is 1.86\plm0.06, which is consistent with our results and also in agreement with that seen in the X-rays. \subsection{Other wavelengths} GRB\,050603 was the target of several ground-based observations at radio and optical wavelengths. \citet{cameron05} reported the Very Large Array (VLA) position at 8.46 GHz as listed in Table~\ref{position}. The flux density was 262\plm41 $\mu$Jy at 8.4 hours after the burst. The afterglow was also observed by SCUBA at 450 and 850 $\mu$m at 11.2 and 13.3 hours after the burst, respectively. \citet{barnard05} reported of a 1.2$\sigma$ detection with a flux density of 2.408\plm1.973 mJy at 450 $\mu$m, but no detection at 850$\mu$m. \citet{berger05a} reported an $R$-band detection of the afterglow with the 2.5m du Pont telescope at Las Campanas Observatory 3.4 hours after the burst with $R$=16.5 mag. The optical position is given in Table~\ref{position}. \citet{berger05b} measured the redshift of the afterglow as z=2.821 based on a single emission line which was interpreted as Ly$\alpha$. This observation was performed with the Magellan/Baade telescope 2.13 days after the burst. GRB\,050603 was also detected by the IBIS instrument on board INTEGRAL in the 40-300 keV energy range \citep{gotz05}. However, because the burst was outside the field of view it could not be localized. The duration of the burst seen by IBIS agreed with the results from the BAT. \citet{golenetskii05} reported the detection of GRB050603 by Konus-Wind in the observed 20keV-3MeV range. Its fluence in this energy band was $(3.41\pm0.06) \times 10^{-5}$ ergs cm$^{-2}$ with a duration of 6~s and an observed $E_{\rm peak}$=349\plm28 keV. \section{\label{discuss}Discussion} GRB\,050603 was a particularly bright and energetic burst. The 15-150 keV fluence of 7.6$\times~10^{-6}$ ergs cm$^{-2}$ places it among the top 10\% of \swift\ bursts. The isotropic energy of 1.26$\times10^{54}$ ergs makes it one of the most energetic bursts ever detected. Combined with the rest-frame $E_{\rm peak}$=1.3 MeV, GRB\,050603 is consistent with the Amati relation \citep{amati02}. With an optical magnitude of V=18.2 9.5 hours after the burst it was also the brightest optical afterglow seen so far by \swift\ at this late time after the burst. The only burst that had a comparable magnitude (V=18.9) at 10 hours after the burst was GRB\,050525 \citep{blustin06}. All other bursts were far below this magnitude at this time after the burst. In X-rays, however, the afterglow of GRB\,050603 was not the brightest one among other \swift\ bursts observed at similar times after the trigger. The X-ray afterglow of GRB\,050603 decayed with a rather steep slope of $\alpha$=1.76\plm0.07, compared with the mean decay slope of $\alpha=1.34 \pm 0.32$ for 27 bursts listed in \citet{nousek05}. This is intermediate between the slopes of $\sim 1.3 \pm 0.1$ obtained from Chandra grating observations, XMM observations, and Beppo-SAX afterglows discussed in \citet{gendre06}, but is somewhat shallower than the mean decay slope of $\alpha_{\rm X}$=2.0\plm0.3 from their survey of Chandra imaging afterglow observations. It is also intermediate between the expectations for the ``normal'' afterglow slope following the end of energy injection by the central engine, and the steeper slope expected following the end of energy injection by the central engine. We note, however, that the decay slope in the V-band filter ($\alpha=1.8\pm0.2$) is consistent with the X-ray decay slope, consistent with the typical signature of an afterglow after the jet break \citep{sari99}. This suggests the possibility that GRB\,050603 had an early jet break before 2.9 hours (rest-frame) after the burst. If we assume that the jet break happened before the start of the UVOT and XRT observations, we can use the dependence of jet angle on break time \citep{sari99, frail01} to place a limit on the jet opening angle of $\Theta_j < 1\degdot3$, where we have used $E_{\rm iso}=1.26\times 10^{54}$ ergs and we assumed the density of the circum-burst matter is 0.1 cm$^{-3}$. We note that \citet{bloom03} argue for a larger typical circum-burst density of 10 cm$^{-3}$, which would imply a slightly larger jet angle limit of 2\degdot3. In order to check the jet interpretation, we compare the observed $\alpha_{\rm X}$=$1.76^{+0.15}_{-0.07}$ and $\beta_{\rm X}$=0.71\plm0.10 to the prediction of the jet model. For the familiar case of $p>2$ (where $p$ is the electron spectral index), one requires that $\alpha_{\rm X} - 2\beta_{\rm X} = 0$ ($\nu > \nu_c$) or $\alpha_{\rm X} - 2\beta_{\rm X} = 1$ ($\nu_m < \nu <\nu_c$), where $\nu_m$ is the typical synchrotron frequency (injection frequency) and $\nu_c$ is the cooling frequency \citep{rhoads99}. In this case, we have $\alpha_{\rm X} - 2\beta_{\rm X} = 0.34^{+0.18}_{-0.12}$. Both cases are inconsistent with the data at the $>2.8\sigma$ level. We therefore consider the case of $1 < p <2$ \citep{dai01}. For $\nu > \nu_c$, this case has $2\alpha_{\rm X}-\beta_{\rm X}=3$, while the data have $2\alpha_{\rm X}-\beta_{\rm X}=2.81^{+0.32}_{-0.12}$, in good agreement with theory. In this regime the electron index is $p=2\beta_{\rm X} = 1.4 \pm 0.2$. This case also has $\alpha_{\rm O}=\alpha_{\rm X}$ as long as the self-absorption frequency is below the optical band, in agreement with the observations. In this case, we expect $\beta_{\rm O} = \beta_{\rm X}$ if the optical band is above the cooling frequency, or $\beta_{\rm O}=0.2 \pm 0.1$ if the optical band is between the injection frequency and the cooling frequency; however, with only V-band data we cannot constrain $\beta_{\rm O}$ directly. The observed spectral index between the X-ray and optical bands is $\beta_{\rm OX}$\footnote{Here we define the optical to X-ray spectral slope $\beta_{\rm OX}= log (5460\AA \times f_{\rm 5460\AA}) - log (1{\rm keV} \times f_{\rm 1KeV}$)}=0.04. These spectral slopes suggest that there is a break in the spectrum between the optical and X-rays suggesting that the optical band is between the injection and cooling frequencies. We therefore conclude that the jet interpretation fits these data provided there is a flat electron energy spectrum $p=1.4$ \citep{dai01}. We note however, that this interpretation is at odds with the expectations from \citet{liang05} who find the relation $E_{\gamma,iso,52}=0.85\times(\frac{E_{\rm peak}}{\rm 100~keV})^{1.94}\times t^{-1.34}$, which predicts a jet break in the optical light curve at 1 day after the burst in the rest-frame, or 3.8 days in the observed frame. \acknowledgments We would like to thank the anonymous referee for valuable comments and suggestions that improved the paper. We would also like to thank Vicki Barnard and Brian Cameron for their information on the SCUBA and VLA observations. We made use of data obtained through the High Energy Astrophysics Science Archive Research Center Online Service, provided by the NASA/Goddard Space Flight Center. This work has been supported by NASA contract NAS5-00136, SAO grant GO5-6076 BASIC and NASA grant NNG05GF43G.

\section*{Acknowledgment} We thank our shepherd Lorenzo Cavallaro and the anonymous reviewers for their constructive and valuable feedback. This work is sponsored in part by NSF grants CCF-18-45893, CCF-18-22965, CCF-16-19123, CNS-18-42456, CNS-18-01426, CNS-16-18771, CNS-16-17670, CNS-15-64055, and CNS-15-63843; ONR grants N00014-17-1-2010, N00014-16-1-2263, and N00014-17-1-2788; an NSF CAREER award; an ARL Young Investigator (YIP) award; a Google Faculty Fellowship; a JP Morgan Faculty Research Award; a DiDi Faculty Research Award; a Google Cloud grant; a Capital One Research Grant; and an Amazon Web Services grant. Any opinions, findings, conclusions, or recommendations expressed herein are those of the authors, and do not necessarily reflect those of the US Government, ONR, ARL, NSF, Captital One, Google, JP Morgan, DiDi, or Amazon. \subsection{Detailed Model Architecture} \label{subsec:detailed_arch} \vspace{.1cm}\noindent\textbf{Multi-head self-attention.} Given the embeddings of all input token (see Section~\ref{subsec:pretrain_method}) in $l$-th layer, $(E_{l,1},...,E_{l,n})$, the self-attention layer updates each embedding with the following steps. It first maps $E_{l,i}$ to three embeddings, \emph{i.e., } query embedding $q_i$, key embedding $k_i$, and value embedding $v_i$: \begin{equation*} \label{eq:kqv} \centering q_i=f_{q}(W_{q};E_{l,i}); \ k_i=f_{k}(W_{k};E_{l,i}); \ v_i=f_{v}(W_{v};E_{l,i}) \end{equation*} Here $f_{q}$, $f_{k}$, and $f_{v}$ are affine transformation functions (\emph{i.e., } a fully-connected layer) parameterized by $W_{q}$, $W_{k}$, and $W_{v}$, respectively. It then computes attention $a_{ij}$ between $E_{l,i}$ and all other embeddings $E_{l,\{j|j\neq i\}}$ by taking the dot product between the $E_{l,i}$'s query embedding $q_i$ and $E_{l,j}$'s key embedding $k_j$: $a_{ij}=q_i\cdot k_j$. Intuitively, attention $a$ is a square matrix, where each cell $a_{ij}$ indicates how much attention $E_{l,i}$ should pay to $E_{l,j}$ when updating itself. It then divides every row of $a$ by $\sqrt{d_{emb}}$ (the dimension of the embedding vectors) and scale it by softmax to ensure them sum up to 1: \begin{equation*} \label{eq:scale} \centering a'_{ij} = \frac{\exp(a_{ij})}{\sum^n_{j=1}\exp(a_{ij})} \end{equation*} The scaled attention $a'_{ij}$ will be multiplied with value embedding $v_j$ and summed up: \begin{equation*} \label{eq:single_head} \centering E^h_{k+1,i} = \sum^n_{j=1} a'_{ij} v_j \end{equation*} Here $h$ in $E^h_{k+1,i}$ denote the updated embedding belong to attention head $h$. Assume we have total $H$ attention heads, the updated embeddings will finally go through an 2-layer feedforward network $f_{out}$ parameterized by $W_{out}$ with skip connections~\cite{he2016deep} to update embeddings from all heads: \begin{equation} \label{eq:multi_head} \centering E_{l+1,i} = f_{out}(concat(E^0_{l+1,i},...,E^H_{l+1,i});W_{out}) \end{equation} \vspace{.1cm}\noindent\textbf{How to use the embedding.} The learned embeddings $E_{l,i}$ after the last self-attention layer encodes the execution semantics of each instruction and the overall function. Consider predicting masked input in pretraining. Let layer $l$ be the $g_p$'s last self-attention layer. The model will stack a 2-layer multi-layer perceptron (MLP) for predicting the masked codes and values: \begin{equation*} MLP(E_{l,i}) = softmax(tanh(E_{l,i}\cdot W_1)\cdot W_2), i\in\mathbb{MP} \end{equation*} $W_1\in \mathbb{R}^{d_{emb}\times d_{emb}}$ and $W_2\in \mathbb{R}^{d_{emb}\times |V|}$ where $|V|$ is the vocabulary size of the code token or bytes. As shown in Equation~\ref{eq:pretrain}, each embedding will have 9 stacked MLPs (\emph{i.e., } one for predicting code and the rest for predicting bytes). \subsection{\textsc{Trex}\xspace Hyperparameters} \label{subsec:hyperparm} \vspace{.1cm}\noindent\textbf{Network architecture.} We use 12 self-attention layers with each having 8 self-attention heads. The embedding dimension is $d_{emb}=d_{func}=768$, which is also the embedding dimension used in bi-LSTM. We set 3072 as the hidden layer size of MLP in the self-attention layer. We adopt GeLU~\cite{hendrycks2016gaussian}, known for addressing the problem of vanishing gradient, as the activation function for \textsc{Trex}\xspace's self-attention module. We use the hyperbolic tangent (tanh) as the activation function in finetuning MLP. We set the dropout rate 0.1 for pretraining do not use dropout in finetuning. \vspace{.1cm}\noindent\textbf{Pretraining.} We fix the largest input length to 512 and choose the effective batch size for both pretraining and finetuning as 512. As 512 batch size with each sample of length 512 is too large to fit in our GPU memory (11 GB), we aggregate the gradient every 64 batches and then updates the weight parameter. This setting results in the actual batch size as 8 ($64\times8=512$). We pick learning rate $5\times10^{-4}$ for pretraining. Instead of starting with the chosen learning rate at first epoch, we follow the common practice of using small warmup learning rate at first epoch. We use $10^{-7}$ as the initial warmup learning rate and gradually increases it until reaching the actual learning rate ($5\times10^{-4}$) after first epoch. We use use Adam optimizer, with $\beta_1=0.9$, $\beta_2=0.98$, $\epsilon=10^{-6}$, and weight decay $10^{-2}$. \subsection{More Experiments} \label{subsec:more_exp} \vspace{.1cm}\noindent\textbf{Cross-project generalizability.} While we strictly separate the functions for pretraining, finetuning, and testing, the functions of finetuning and testing can come from the same project (but strictly different functions). Therefore, in this section, we further separate the functions in finetuning and testing by extracting them from \emph{different projects}. For example, we can finetune the model on Coreutils while testing on OpenSSL. Specifically, we select 3 projects, \emph{i.e., } Binutils, Coreutils, and OpenSSL, which have the largest number of functions, and evaluate how \textsc{Trex}\xspace performs. We allow the functions to come from different architectures, optimizations, and obfuscations, and follow the same setup as described in Section~\ref{sec:impl}. \begin{table}[!t] \footnotesize \setlength{\tabcolsep}{7pt} \centering \renewcommand{\arraystretch}{1.1} \setlength\aboverulesep{0.4pt} \setlength\belowrulesep{0.4pt} \caption{\textsc{Trex}\xspace results (in AUC score) on function pairs with training and testing functions extracted from different projects. } \label{tab:cross-dataset} \begin{tabular}{r|c|c|c} \toprule[1.1pt] \backslashbox{Train}{Test} & \textbf{Coreutils} & \textbf{Binutils} & \textbf{OpenSSL} \\ \midrule[.9pt] \rowcolor{lightblue} \textbf{Coreutils} & 0.947 & 0.945 & 0.940 \\ \hline \textbf{Binutils} & 0.945 & 0.945 & 0.944 \\ \hline \rowcolor{lightblue} \textbf{OpenSSL} & 0.936 & 0.939 & 0.956 \\ \bottomrule[1.1pt] \end{tabular} \end{table} Table~\ref{tab:cross-dataset} shows that \textsc{Trex}\xspace's AUC score does not drop dramatically ($<2$\%) when the functions are coming from different projects when compared to coming from same projects (the diagonal). This observation indicates \textsc{Trex}\xspace generalizes to unseen function pairs in an entirely different dataset. Note that the functions in each function pair can come from arbitrary architectures, optimizations, and obfuscations (last column in Table~\ref{tab:result}). As we have shown in Section~\ref{subsec:rq2}, the numbers achieved by \textsc{Trex}\xspace in Table~\ref{tab:cross-dataset} even outperforms the existing baselines when their functions can only come from different architectures and within only the same projects. \vspace{.1cm}\noindent\textbf{Effectiveness of bi-LSTM encoding.} As described in Section~\ref{subsec:input_repr}, we treat the numeric values as an 8-byte sequence and use bi-LSTM to combine them into a single representation (embedding). The structure of bi-LSTM is known to capture the potential dependencies between different bytes (with different significance) in the byte-sequence. Indeed, besides bi-LSTM, there are other possible differentiable modules such as multi-layer perceptron (MLP) or simple summation can also combine the input bytes. Therefore, we study the performance of \textsc{Trex}\xspace in predicting masked tokens when we vary the modules for byte-sequence combination. Specifically, we follow the same setting described above by selecting a random 10,000 function binaries as the testing set and evaluate the testing PPL achieved by pretraining \textsc{Trex}\xspace. Figure~\ref{fig:pretrain-loss} shows the validation PPL in 10 epochs of pretraining \textsc{Trex}\xspace based on (1) bi-LSTM, (2) 2-layer (with 1024 hidden size) MLP, and (3) simple summation (SUM), to combine the byte-sequence. We can observe that bi-LSTM is obviously better than other two, achieving the lowest PPL. This indicates that bi-LSTM helps \textsc{Trex}\xspace the most in terms of learning approximate execution semantics. \vspace{.1cm}\noindent\textbf{Vulnerability search performance.} We quantify the accuracy of \textsc{Trex}\xspace in searching vulnerable functions in the firmware images and compare it to that of SAFE. As SAFE does not work for MIPS, we study how it performs on NETGEAR R7000 model, the only model that runs on ARM architecture from Table~\ref{tab:cve_case}. Specifically, we compile OpenSSL to ARM and x64 with \texttt{O3}, and feed both our compiled and firmware's function binaries to \textsc{Trex}\xspace and SAFE to compute embeddings. Based on the function embeddings, we search the compiled OpenSSL functions in the NETGEAR R7000's embedded OpenSSL libraries, and test their top-1/3/5/10 errors. For example, the top-10 error measures when the query function does not appear in the top-10 most similar functions in the firmware. \begin{figure}[!t] \centering \subfloat[Same architecture]{ \includegraphics[width=0.47\linewidth]{./figs/case/vuln_error_same.pdf} \label{subfig:same}} \subfloat[Cross architecture]{ \includegraphics[width=0.47\linewidth]{./figs/case/vuln_error_cross.pdf} \label{subfig:cross}} \caption{Top-1/3/5/10 error of \textsc{Trex}\xspace and SAFE in searching functions in firmware. (a) The query functions and the firmware are from same architecture (ARM). (b) The query functions are from x64 but the firmware are from ARM.} \label{fig:vuln_error} \end{figure} Figure~\ref{fig:vuln_error} shows that \textsc{Trex}\xspace consistently outperforms SAFE, achieving 24.3\% lower error rate on average. Moreover, as shown in Figure~\ref{subfig:cross}, when the query functions are from x64 (the firmware binaries are from ARM), \textsc{Trex}\xspace outperforms SAFE by a greater margin, achieving 25.3\% lower error rate in cross-architecture search. \begin{table*}[!t] \footnotesize \setlength{\tabcolsep}{1.5pt} \centering \renewcommand{\arraystretch}{1} \setlength\aboverulesep{0.4pt} \setlength\belowrulesep{0.4pt} \caption{Details of the real-world 180 firmware images we collected from DD-WRT. } \label{tab:firmware} \begin{tabular}{l|l|l|l?{1.1pt}l|l|l|l?{1.1pt}l|l|l|l} \toprule[1.1pt] \textbf{Vendor} & \textbf{Model} & \textbf{ISA} & \textbf{Year} & \textbf{Vendor} & \textbf{Model} & \textbf{ISA} & \textbf{Year} & \textbf{Vendor} & \textbf{Model} & \textbf{ISA} & \textbf{Year} \\ \midrule[.9pt] \rowcolor{lightblue} 8devices &Carambola 2 &MIPS &2020 &8devices &Carambola 2 8MB &MIPS &2018 &8devices &Lima &MIPS &2020 \\ \hline Actiontec &MI424WR &ARM &2019 &Alfa &AIP-W502U &MIPS &2020 &Alfa &SOLO48 &MIPS &2020 \\ \hline \rowcolor{lightblue} Belkin &F5D8235-4 &MIPS &2020 &Buffalo &BHR-4GRV &MIPS &2020 &Buffalo &WBMR-HP-G300H &MIPS &2020 \\ \hline Buffalo &WHR-1166D &MIPS &2020 &Buffalo &WHR-300HP2 &MIPS &2018 &Buffalo &WHR-600D &MIPS &2018 \\ \hline \rowcolor{lightblue} Buffalo &WXR-1900DHP &ARM &2020 &Buffalo &WZR-1166DHP &ARM &2020 &Buffalo &WZR-600DHP2 &ARM &2020 \\ \hline Buffalo &WZR-900DHP &ARM &2020 &Buffalo &WZR-HP-G450H &MIPS &2020 &Comfast &CF-E325N &MIPS &2020 \\ \hline \rowcolor{lightblue} Comfast &CF-E355AC &MIPS &2020 &Comfast &CF-E380AC &MIPS &2020 &Comfast &CF-WR650AC &MIPS &2020 \\ \hline Compex &WP546 &MIPS &2020 &Compex &WPE72 &MIPS &2020 &D-Link &DAP-2230 &MIPS &2020 \\ \hline \rowcolor{lightblue} D-Link &DAP-2330 &MIPS &2020 &D-Link &DAP-2660 &MIPS &2020 &D-Link &DAP-3320 &MIPS &2020 \\ \hline D-Link &DAP-3662 &MIPS &2020 &D-Link &DHP-1565 A1 &MIPS &2020 &D-Link &DIR-825 C1 &MIPS &2020 \\ \hline \rowcolor{lightblue} D-Link &DIR-825 Rev.B &MIPS &2020 &D-Link &DIR-835 A1 &MIPS &2020 &D-Link &DIR-859 &MIPS &2020 \\ \hline D-Link &DIR-860L A1 &ARM &2020 &D-Link &DIR-860L B1 &MIPS &2020 &D-Link &DIR-862 &MIPS &2020 \\ \hline \rowcolor{lightblue} D-Link &DIR-866 &MIPS &2020 &D-Link &DIR-868l Rev.A &ARM &2020 &D-Link &DIR-868l Rev.B &ARM &2020 \\ \hline D-Link &DIR-868l Rev.C &ARM &2020 &D-Link &DIR-869 &MIPS &2020 &D-Link &DIR-878 A1 &MIPS &2020 \\ \hline \rowcolor{lightblue} D-Link &DIR-880l &ARM &2020 &D-Link &DIR-882 A1 &MIPS &2020 &D-Link &DIR-885l &ARM &2020 \\ \hline D-Link &DIR-890l &ARM &2020 &D-Link &DIR632A &MIPS &2020 &GL.iNet &AR150 &MIPS &2020 \\ \hline \rowcolor{lightblue} Gateworks &GW2382 &ARM &2018 &Gateworks &GW2388 16M &ARM &2018 &Gateworks &GW2391 &ARM &2018 \\ \hline Gateworks &Laguna GW2382 &ARM &2020 &Gateworks &Laguna GW2388 32M &ARM &2020 &Gateworks &Laguna GW2391 &ARM &2020 \\ \hline \rowcolor{lightblue} Gigaset &SX763 &MIPS &2020 &JJPlus &JA76PF &MIPS &2020 &Linksys &E1700 &MIPS &2020 \\ \hline Linksys &E2100L &MIPS &2020 &Linksys &EA6350 &ARM &2020 &Linksys &EA6400 &ARM &2020 \\ \hline \rowcolor{lightblue} Linksys &EA6500 v2 &ARM &2020 &Linksys &EA6700 &ARM &2020 &Linksys &EA6900 &ARM &2020 \\ \hline Linksys &EA8500 &ARM &2020 &Linksys &RE7000 &MIPS &2020 &Linksys &WRT610N v1.0 &MIPS &2020 \\ \hline \rowcolor{lightblue} NETGEAR &AC1450 &ARM &2020 &NETGEAR &EX6200 &ARM &2018 &NETGEAR &R6250 &MIPS &2020 \\ \hline NETGEAR &R6300v2 &ARM &2020 &NETGEAR &R6400 &ARM &2020 &NETGEAR &R6700 &ARM &2020 \\ \hline \rowcolor{lightblue} NETGEAR &R6700v3 &ARM &2020 &NETGEAR &R7000 &ARM &2020 &NETGEAR &R7000P &ARM &2020 \\ \hline NETGEAR &R7500v1 &ARM &2020 &NETGEAR &R7500v2 &ARM &2020 &NETGEAR &R7800 &ARM &2020 \\ \hline \rowcolor{lightblue} NETGEAR &R8500 &ARM &2020 &NETGEAR &R8900 &ARM &2020 &NETGEAR &R9000 &ARM &2020 \\ \hline NETGEAR &WG302v2 &ARM &2020 &NETGEAR &WNDR3700v4 &MIPS &2020 &NETGEAR &WNDR4300 &MIPS &2020 \\ \hline \rowcolor{lightblue} NETGEAR &WNDR4500 &MIPS &2020 &NETGEAR &WNDR4500v2 &MIPS &2020 &NETGEAR &XR450 &ARM &2020 \\ \hline NETGEAR &XR500 &ARM &2020 &NETGEAR &XR700 &ARM &2020 &Pronghorn &SBC &ARM &2020 \\ \hline \rowcolor{lightblue} Senao &ECB3500 &MIPS &2020 &Senao &ECB9750 &MIPS &2019 &Senao &EOC1650 &MIPS &2020 \\ \hline Senao &EOC5510 &MIPS &2020 &Senao &EOC5610 &MIPS &2020 &Senao &EOC5611 &MIPS &2020 \\ \hline \rowcolor{lightblue} Senao &NOP8670 &ARM &2020 &TP-Link &Archer A7 V5 &MIPS &2020 &TP-Link &Archer C1900 &ARM &2020 \\ \hline TP-Link &Archer C25 V1 &MIPS &2020 &TP-Link &Archer C5 V1 &MIPS &2020 &TP-Link &Archer C7 V1 &MIPS &2020 \\ \hline \rowcolor{lightblue} TP-Link &Archer C7 V2 &MIPS &2020 &TP-Link &Archer C7 V3 &MIPS &2020 &TP-Link &Archer C7 V4 &MIPS &2020 \\ \hline TP-Link &Archer C8 V1 &ARM &2020 &TP-Link &Archer C9 V1 &ARM &2020 &TP-Link &Archer C9 V2 &ARM &2020 \\ \hline \rowcolor{lightblue} TP-Link &Archer C9 V3 &ARM &2020 &TP-Link &Deco-M4 &MIPS &2020 &TP-Link &TL-WDR3600 V1 &MIPS &2020 \\ \hline TP-Link &TL-WDR4300 V1 &MIPS &2020 &TP-Link &TL-WDR4310 V1 &MIPS &2020 &TP-Link &TL-WDR4900 V2 &MIPS &2020 \\ \hline \rowcolor{lightblue} TP-Link &TL-WR1043N V5 &MIPS &2020 &TP-Link &TL-WR1043ND &MIPS &2020 &TP-Link &TL-WR1043ND V2 &MIPS &2020 \\ \hline TP-Link &TL-WR1043ND V4 &MIPS &2020 &TP-Link &TL-WR2543ND &MIPS &2020 &TP-Link &TL-WR710N V1 &MIPS &2020 \\ \hline \rowcolor{lightblue} TP-Link &TL-WR710N V2.1.0 &MIPS &2020 &TP-Link &TL-WR810N V1 &MIPS &2020 &TP-Link &TL-WR810N V2 &MIPS &2020 \\ \hline TP-Link &TL-WR842N V1 &MIPS &2020 &TP-Link &TL-WR842N V2 &MIPS &2020 &TRENDnet &TEW-811DRU &ARM &2018 \\ \hline \rowcolor{lightblue} TRENDnet &TEW-812DRU V2 &ARM &2018 &TRENDnet &TEW-818DRU &ARM &2020 &TRENDnet &TEW-828DRU &ARM &2020 \\ \hline Ubiquiti &AirGrid M2 &MIPS &2020 &Ubiquiti &AirGrid M5 &MIPS &2020 &Ubiquiti &AirGrid-M5-XW &MIPS &2020 \\ \hline \rowcolor{lightblue} Ubiquiti &AirRouter &MIPS &2020 &Ubiquiti &AirRouter-HP &MIPS &2020 &Ubiquiti &AirWire &MIPS &2020 \\ \hline Ubiquiti &BulletM2 HP &MIPS &2020 &Ubiquiti &BulletM5 HP &MIPS &2020 &Ubiquiti &LS SR71A &MIPS &2020 \\ \hline \rowcolor{lightblue} Ubiquiti &NanoBeam AC &MIPS &2020 &Ubiquiti &NanoBeam M2 XW &MIPS &2020 &Ubiquiti &NanoBeam M5 XW &MIPS &2020 \\ \hline Ubiquiti &NanoBridge M2 &MIPS &2020 &Ubiquiti &NanoBridge M2 XW &MIPS &2020 &Ubiquiti &NanoBridge M3 &MIPS &2020 \\ \hline \rowcolor{lightblue} Ubiquiti &NanoBridge M365 &MIPS &2020 &Ubiquiti &NanoBridge M5 XW &MIPS &2020 &Ubiquiti &NanoBridge M900 &MIPS &2020 \\ \hline Ubiquiti &NanoStation M2 &MIPS &2020 &Ubiquiti &NanoStation M3 &MIPS &2020 &Ubiquiti &NanoStation M365 &MIPS &2020 \\ \hline \rowcolor{lightblue} Ubiquiti &Pico M5 &MIPS &2020 &Ubiquiti &Power AP N &MIPS &2020 &Ubiquiti &PowerBeam M5 M400 XW &MIPS &2020 \\ \hline Ubiquiti &PowerBridge M10 &MIPS &2020 &Ubiquiti &PowerBridge M5 &MIPS &2020 &Ubiquiti &Rocket M2 Titanium XW &MIPS &2020 \\ \hline \rowcolor{lightblue} Ubiquiti &Rocket M2 XW &MIPS &2020 &Ubiquiti &Rocket M5 Titanium XW &MIPS &2020 &Ubiquiti &Rocket M5 X3 XW &MIPS &2020 \\ \hline Ubiquiti &Rocket M5 XW &MIPS &2020 &Ubiquiti &RocketM2 &MIPS &2020 &Ubiquiti &RocketM3 &MIPS &2020 \\ \hline \rowcolor{lightblue} Ubiquiti &RocketM365 &MIPS &2020 &Ubiquiti &RocketM5 &MIPS &2020 &Ubiquiti &RocketM900 &MIPS &2020 \\ \hline Ubiquiti &RouterStation &MIPS &2020 &Ubiquiti &RouterStation Pro &MIPS &2020 &Ubiquiti &sunMax &MIPS &2020 \\ \hline \rowcolor{lightblue} Ubiquiti &UAP-AC-MESH &MIPS &2020 &Ubiquiti &UAP-AC-PRO &MIPS &2020 &Ubiquiti &UAP-LR &MIPS &2020 \\ \hline Ubiquiti &UAP-LR-v2 &MIPS &2020 &Ubiquiti &UAP-v2 &MIPS &2020 &Ubiquiti &locoM2 &MIPS &2020 \\ \hline \rowcolor{lightblue} Ubiquiti &locoM2 XW &MIPS &2020 &Ubiquiti &locoM5 &MIPS &2020 &Ubiquiti &locoM5 XW &MIPS &2020 \\ \hline Ubiquiti &locoM900 &MIPS &2020 &WiliGear &WBD-500 &MIPS &2020 &YunCore &XD3200 &MIPS &2020 \\ \bottomrule[1.1pt] \end{tabular} \end{table*} \subsection{Probing Learned Execution Semantics} \label{subsec:probe_case} As discussed in Section~\ref{sec:overview}, training the model to predict masked micro-trace codes and values compels the model to learn execution semantics by concrete examples. It thus automates the extraction of the code's dynamic features without manual effort. In this section, we study concrete code examples showing the potential hint in the code that the model likely leverages to predict the masked part. \begin{figure}[!t] \centering \includegraphics[width=\linewidth]{./figs/case/arithmetic.pdf} \caption{The partial micro-trace code and value sequence from function \texttt{BinarySource\_get\_rsa\_ssh1\_pub} in PuTTy-0.74 compiled to x64 with \texttt{O0}. We mask the register \colorbox{lightgray}{\texttt{rax}} at line 3 and the opcode \colorbox{lightgray}{\texttt{add}} at line 4. We also mask their corresponding micro-trace values (where opcode has only dummy values (8 \texttt{\#\#}s) as described in Section~\ref{subsec:input_repr}). We highlight the contextual hint in \colorbox{myblue}{blue} that \textsc{Trex}\xspace leverages to predict the masked \colorbox{lightgray}{\texttt{rax}} and \colorbox{lightred}{red} that \textsc{Trex}\xspace uses to predict the masked \colorbox{lightgray}{\texttt{add}}.} \label{fig:arithmetic} \end{figure} \vspace{.1cm}\noindent\textbf{Predicting arithmetic instructions.} Consider the example in Figure~\ref{fig:arithmetic}, which shows the function's micro-trace (\emph{i.e., } the dynamic value and assembly code). For ease of exposition, the format of micro-trace values (\emph{e.g., } the value of opcode and constants) are not exactly the same as the actual format we handle, as described in Section~\ref{subsec:input_repr}. We mask the register \texttt{rax} and opcode \texttt{add} and their corresponding values in a sequence of arithmetic instructions. To correctly predict the masked value \texttt{0x4c0469a} of \texttt{rax}, the model should understand the approximate execution semantics of \texttt{sub} in line 2, which subtracts \texttt{rax} with \texttt{rcx}. It may also observe the value of \texttt{rax} at line 4, which exactly equals the result of adding \texttt{rax} with 7 at line 3. Therefore, we see the model predicts \texttt{rax} with 96\% confidence and \texttt{rcx} with only 3\%. It also predicts the value of \texttt{rax} correctly (right table). To correctly predict \texttt{add} at line 4, the model should observe that the \texttt{rax} at line 5 has the same exact value with the result of \texttt{rax+rcx} at line 4, so it predicts the opcode at line 4 to be \texttt{add}. As shown in Figure~\ref{fig:arithmetic}, our pretrained model predicts \texttt{add} with 98\% confidence and \texttt{mov} with only 1\%, which implies that the model approximately understands the execution semantics of \texttt{add} and \texttt{mov} so it predicts that \texttt{add} is much more likely. \vspace{.1cm}\noindent\textbf{Predicting stack operations.} Consider the example in Figure~\ref{fig:stack}. We mask the register and constant of the instruction in function epilogue -- it increments the stack pointer \texttt{rsp} by \texttt{0x20} to deallocate the local variable stored on stack. To correctly predict the masked \texttt{rsp} and its value \texttt{0x8b4a3f}, the model should observe the \texttt{rsp} is decremented (due to opcode \texttt{sub}) by \texttt{0x20} from \texttt{0x8b4a3f} at line 3, which is part of the function prologue. Therefore, the model should understand the execution semantics of \texttt{sub} and basic syntax of function prologue and epilogue. To predict the masked constant \texttt{0x20}, the model should notice that line 3 decrements the stack pointer by \texttt{0x20}, which is the size of local variables. As the model predicts \texttt{0x20} with 99\% confidence, this implies the model likely learns patterns of function prologue and epilogue, including the context such as \texttt{push ebp}, \texttt{pop ebp}, etc. This observation indicates that our pretraining task assists in learning common function idiom (function calling convention) beyond execution semantics of individual instructions. Learning such knowledge can potentially help other tasks beyond function similarity, such as identifying function boundaries~\cite{pei2021xda}. \begin{figure}[!t] \centering \includegraphics[width=.95\linewidth]{./figs/case/stack.pdf} \caption{The partial code and micro-trace from function \texttt{closeUnixFile} in SQLite-3.34.0 compiled to x64 with \texttt{O0}. We mask the register \colorbox{lightgray}{\texttt{rsp}} and constant \colorbox{lightgray}{\texttt{0x20}} at line 4. We also mask their corresponding micro-trace values. We highlight the contextual hint in \colorbox{lightred}{red} that \textsc{Trex}\xspace leverages to predict the masked \colorbox{lightgray}{\texttt{rsp}} and \colorbox{lightgray}{\texttt{0x20}}.} \label{fig:stack} \end{figure} \end{appendix} \section{Case Studies} \label{sec:case} In this section, we study how \textsc{Trex}\xspace can help discover new vulnerabilities in a large volume of latest firmware images. \begin{table}[!t] \footnotesize \setlength{\tabcolsep}{4pt} \centering \renewcommand{\arraystretch}{1} \setlength\aboverulesep{0.4pt} \setlength\belowrulesep{0.4pt} \caption{Vulnerabilities we have confirmed (\cmark) in firmware images (latest version) from 4 well-known vendors and products.} \label{tab:cve_case} \begin{tabular}{r?{1.1pt}c|c|c|c} \toprule[1.1pt] \multicolumn{1}{c|}{\multirow{2}{*}{\textbf{CVE}}} & \multicolumn{1}{c|}{\multirow{2}{*}{\textbf{\begin{tabular}[c]{@{}c@{}}Ubiquiti\\ sunMax\end{tabular}}}} & \multicolumn{1}{c|}{\multirow{2}{*}{\textbf{\begin{tabular}[c]{@{}c@{}}TP-Link\\ Deco-M4\end{tabular}}}} & \multicolumn{1}{c|}{\multirow{2}{*}{\textbf{\begin{tabular}[c]{@{}c@{}}NETGEAR\\ R7000\end{tabular}}}} & \multicolumn{1}{c}{\multirow{2}{*}{\textbf{\begin{tabular}[c]{@{}c@{}}Linksys\\ RE7000\end{tabular}}}} \\ \multicolumn{1}{c|}{} & \multicolumn{1}{c|}{} & \multicolumn{1}{c|}{} & \multicolumn{1}{c|}{} & \multicolumn{1}{c}{} \\ \midrule[.9pt] CVE-2019-1563 & \cmark & \cmark & \cmark & \xmark \\ \hline \rowcolor{lightblue} CVE-2017-16544 & \xmark & \cmark & \xmark & \xmark \\ \hline CVE-2016-6303 & \cmark & \cmark & \cmark & \cmark \\ \hline \rowcolor{lightblue} CVE-2016-6302 & \cmark & \cmark & \cmark & \cmark \\ \hline CVE-2016-2842 & \cmark & \cmark & \cmark & \cmark \\ \hline \rowcolor{lightblue} CVE-2016-2182 & \cmark & \cmark & \cmark & \cmark \\ \hline CVE-2016-2180 & \cmark & \cmark & \cmark & \cmark \\ \hline \rowcolor{lightblue} CVE-2016-2178 & \cmark & \cmark & \cmark & \cmark \\ \hline CVE-2016-2176 & \cmark & \cmark & \xmark & \cmark \\ \hline \rowcolor{lightblue} CVE-2016-2109 & \cmark & \cmark & \xmark & \cmark \\ \hline CVE-2016-2106 & \cmark & \cmark & \xmark & \cmark \\ \hline \rowcolor{lightblue} CVE-2016-2105 & \cmark & \cmark & \xmark & \cmark \\ \hline CVE-2016-0799 & \cmark & \cmark & \xmark & \cmark \\ \hline \rowcolor{lightblue} CVE-2016-0798 & \cmark & \cmark & \xmark & \cmark \\ \hline CVE-2016-0797 & \cmark & \cmark & \xmark & \cmark \\ \hline \rowcolor{lightblue} CVE-2016-0705 & \cmark & \cmark & \xmark & \cmark \\ \bottomrule[1.1pt] \end{tabular} \end{table} Firmware images often include third-party libraries (\emph{i.e., } BusyBox, OpenSSL). However, these libraries are frequently patched, but the manufacturers often fall behind to update them accordingly~\cite{OWASP}. Therefore, we study whether our tool can uncover function binaries in firmware images similar to known vulnerable functions. We find existing state-of-the-art binary similarity tools (\emph{i.e., } SAFE~\cite{massarelli2019safe}, Asm2Vec~\cite{ding2019asm2vec}, Gemini~\cite{xu2017neural}) all perform their case studies on the firmware images and vulnerabilities that have already been studied before them~\cite{costin2014large, feng2016scalable, david2016statistical, chen2016towards}. Therefore, we decide to collect our own dataset with more updated firmware images and the latest vulnerabilities, instead of reusing the existing benchmarks. This facilitates finding new (1-day) vulnerabilities in most recent firmware images that are not disclosed before. We crawl firmware images (in their latest version) in 180 products including WLAN routers, smart cameras, and solar panels, from well-known manufacturers' official releases and third-party providers such as DD-WRT~\cite{ddwrt}, as shown in Table~\ref{tab:firmware}. We collect firmware images with only the latest version, which are much more updated than those studied in the state-of-the-art (dated before 2015, which are all patched before their studies). For each function in the firmware images, we construct function embedding and build a firmware image database using Open Distro for Elasticsearch~\cite{elastic}, which supports vector-based indexing with efficient search support based on NMSLIB~\cite{boytsov2013engineering}. We extract the firmware images using binwalk~\cite{binwalk}. In total, we collect 180 number of firmware images from 22 vendors. These firmware images are developed by the original vendors, or from third-party providers such as DD-WRT~\cite{ddwrt}. Most of them are for WLAN routers, while some are deployed in other embedded systems, such as solar panels and smart cameras. Among the firmware images, 127 of them are compiled for MIPS 32-bit and 53 are compiled for ARM 32-bit. 169 of them are 2020 models, and the rest are 2019 and 2018 models. \begin{table}[!t] \footnotesize \setlength{\tabcolsep}{7pt} \centering \renewcommand{\arraystretch}{1} \setlength\aboverulesep{0.4pt} \setlength\belowrulesep{0.4pt} \caption{We study 16 vulnerabilities from OpenSSL and BusyBox, two widely-used libraries in firmware images.} \label{tab:cve} \begin{tabular}{r|l|l} \toprule[1.1pt] \textbf{CVE} & \textbf{Library} & \textbf{Description} \\ \midrule[.9pt] CVE-2019-1563 & OpenSSL & Decrypt encrypted message \\ \hline \rowcolor{lightblue} CVE-2017-16544 & BusyBox & Allow executing arbitrary code \\ \hline CVE-2016-6303 & OpenSSL & Integer overflow \\ \hline \rowcolor{lightblue} CVE-2016-6302 & OpenSSL & Allows denial-of-service \\ \hline CVE-2016-2842 & OpenSSL & Allows denial-of-service \\ \hline \rowcolor{lightblue} CVE-2016-2182 & OpenSSL & Allows denial-of-service \\ \hline CVE-2016-2180 & OpenSSL & Out-of-bounds read \\ \hline \rowcolor{lightblue} CVE-2016-2178 & OpenSSL & Leak DSA private key \\ \hline CVE-2016-2176 & OpenSSL & Buffer over-read \\ \hline \rowcolor{lightblue} CVE-2016-2109 & OpenSSL & Allows denial-of-service \\ \hline CVE-2016-2106 & OpenSSL & Integer overflow \\ \hline \rowcolor{lightblue} CVE-2016-2105 & OpenSSL & Integer overflow \\ \hline CVE-2016-0799 & OpenSSL & Out-of-bounds read \\ \hline \rowcolor{lightblue} CVE-2016-0798 & OpenSSL & Allows denial-of-service \\ \hline CVE-2016-0797 & OpenSSL & NULL pointer dereference \\ \hline \rowcolor{lightblue} CVE-2016-0705 & OpenSSL & Memory corruption \\ \bottomrule[1.1pt] \end{tabular} \end{table} Table~\ref{tab:cve_case} shows 16 vulnerabilities (CVEs) we use to search in the firmware images. We focus on the CVEs of OpenSSL and BusyBox, as they are widely included in the firmware. For each CVE, we compile the corresponding vulnerable functions in the specified library version and computes the vulnerable function embeddings via \textsc{Trex}\xspace. As the firmware images are stripped so that we do not know with which optimizations they are compiled, we compile the vulnerable functions to both MIPS and ARM with \texttt{-O3} and rely on \textsc{Trex}\xspace's capability in cross-architecture and optimization function matching to match functions that are potentially compiled in different architectures and with different optimizations. We then obtain the firmware functions that rank top-10 similar to the vulnerable function and manually verify if they are vulnerable. We leverage \texttt{strings} command to identify the OpenSSL and BusyBox versions indicative of the corresponding vulnerabilities. Note that such information can be stripped for other libraries so it is not a reliable approach in general. As shown in Table~\ref{tab:cve_case}, we have confirmed all 16 CVEs in 4 firmware models developed by well-known vendors, \emph{i.e., } Ubiquiti, TP-Link, NETGEAR, and Linksys. These cases demonstrate the practicality of \textsc{Trex}\xspace, which helps discover real-world vulnerabilities in large-scale firmware databases. Table~\ref{tab:cve} shows the details of the 16 vulnerabilities that \textsc{Trex}\xspace uncover in the 4 firmware images shown in Table~\ref{tab:cve_case}. The description of ``allow denial-of-service'' usually refers to the segmentation fault that crashes the program. The cause of such error can be diverse, which is not due to the other typical causes shown in the table (\emph{e.g., } integer overflow, buffer over-read, etc.). \section{Conclusion} \label{sec:conclusion} We introduced \textsc{Trex}\xspace to match semantically similar functions based on the function execution semantics. Our key insight is to first pretrain the ML model to explicitly learn approximate execution semantics based on the functions' mico-traces and then transfer the learned knowledge to match semantically similar functions. Our evaluation showed that the learned approximate execution semantics drastically improves the accuracy of matching semantically similar functions -- \textsc{Trex}\xspace excels in matching functions across different architectures, optimizations, and obfuscations. We plan to explore in our future work how the learned execution semantics of the code can further boost the performance of broader (binary) program analysis tasks such as decompilation. \section{Discussion} \label{sec:discussion} \vspace{.1cm} \noindent\textbf{Definition of pretraining and finetuning.} A typical difference between pretraining and finetuning in NLP is pretraining is often self-/un-supervised while finetuning is supervised. Besides, pretraining is often believed responsible for heavy-lifting, with huge amount of training corpus and parameter updates. While in finetuning, the training process often does not significantly update the ``knowledge'' of the pretrained model. Instead, it is often light-weight with small extra layers and number of training epochs. In general, pretraining learns task-agnostic general knowledge while finetuning learns for task-specific downstream tasks. In the setting of this project, such distinction becomes not so obvious, as our finetuning task (function similarity) does not necessarily requires any labeling effort. Particularly, we can cheaply collect the same source code and compile into different architectures, optimizations, and obfuscations. Therefore, we believe our finetuning task can also be leveraged to learn general semantic representation of the code. By encouraging the model to produce similar embeddings for same functions with different transformations, we teach the model invariances of underlying code semantics. We plan to explore this direction in the future work. \section{Evaluation} \label{sec:eval} Our evaluation aims to answer the following questions. \begin{itemize} \item RQ1: How accurate is \textsc{Trex}\xspace in matching semantically similar function binaries across different architectures, optimizations, and obfuscations? \item RQ2: How does \textsc{Trex}\xspace compare to the state-of-the-art? \item RQ3: How fast is \textsc{Trex}\xspace compared to other tools? \item RQ4: How much does pretraining on micro-traces help improve the accuracy of matching functions? \end{itemize} \begin{table}[!t] \footnotesize \setlength{\tabcolsep}{5.5pt} \centering \renewcommand{\arraystretch}{1} \setlength\aboverulesep{0.4pt} \setlength\belowrulesep{0.4pt} \caption{\textsc{Trex}\xspace results (in AUC score) on function pairs across architectures, optimizations, and obfuscations. } \label{tab:result} \begin{tabular}{r?{1.1pt}ccccc} \toprule[1.1pt] \multirow{2}{*}{} & \multicolumn{5}{c}{\textbf{Cross-}} \\ \cline{2-6} & \multicolumn{1}{c|}{\textbf{ARCH}} & \multicolumn{1}{c|}{\textbf{OPT}} & \multicolumn{1}{c|}{\textbf{OBF}} & \multicolumn{1}{c|}{\textbf{\begin{tabular}[c]{@{}l@{}}ARCH+\\ OPT\end{tabular}}} & \multicolumn{1}{c}{\textbf{\begin{tabular}[c]{@{}l@{}}ARCH+\\ OPT+\\ OBF\end{tabular}}} \\ \midrule[.9pt] \rowcolor{lightblue} \textbf{Binutils} & \multicolumn{1}{c|}{0.993} & \multicolumn{1}{c|}{0.993} & \multicolumn{1}{c|}{0.991} & \multicolumn{1}{c|}{0.959} & 0.947 \\ \hline \textbf{Coreutils} & \multicolumn{1}{c|}{0.991} & \multicolumn{1}{c|}{0.992} & \multicolumn{1}{c|}{0.991} & \multicolumn{1}{c|}{0.956} & 0.945 \\ \hline \rowcolor{lightblue} \textbf{Curl} & \multicolumn{1}{c|}{0.993} & \multicolumn{1}{c|}{0.993} & \multicolumn{1}{c|}{0.991} & \multicolumn{1}{c|}{0.958} & 0.956 \\ \hline \textbf{Diffutils} & \multicolumn{1}{c|}{0.992} & \multicolumn{1}{c|}{0.992} & \multicolumn{1}{c|}{0.990} & \multicolumn{1}{c|}{0.970} & 0.961 \\ \hline \rowcolor{lightblue} \textbf{Findutils} & \multicolumn{1}{c|}{0.990} & \multicolumn{1}{c|}{0.992} & \multicolumn{1}{c|}{0.990} & \multicolumn{1}{c|}{0.965} & 0.963 \\ \hline \textbf{GMP} & \multicolumn{1}{c|}{0.992} & \multicolumn{1}{c|}{0.993} & \multicolumn{1}{c|}{0.990} & \multicolumn{1}{c|}{0.968} & 0.966 \\ \hline \rowcolor{lightblue} \textbf{ImageMagick} & \multicolumn{1}{c|}{0.993} & \multicolumn{1}{c|}{0.993} & \multicolumn{1}{c|}{0.989} & \multicolumn{1}{c|}{0.960} & 0.951 \\ \hline \textbf{Libmicrohttpd} & \multicolumn{1}{c|}{0.994} & \multicolumn{1}{c|}{0.994} & \multicolumn{1}{c|}{0.991} & \multicolumn{1}{c|}{0.972} & 0.969 \\ \hline \rowcolor{lightblue} \textbf{LibTomCrypt} & \multicolumn{1}{c|}{0.992} & \multicolumn{1}{c|}{0.994} & \multicolumn{1}{c|}{0.991} & \multicolumn{1}{c|}{0.981} & 0.970 \\ \hline \textbf{OpenSSL} & \multicolumn{1}{c|}{0.992} & \multicolumn{1}{c|}{0.992} & \multicolumn{1}{c|}{0.989} & \multicolumn{1}{c|}{0.964} & 0.956 \\ \hline \rowcolor{lightblue} \textbf{PuTTy} & \multicolumn{1}{c|}{0.992} & \multicolumn{1}{c|}{0.995} & \multicolumn{1}{c|}{0.990} & \multicolumn{1}{c|}{0.961} & 0.952 \\ \hline \textbf{SQLite} & \multicolumn{1}{c|}{0.991} & \multicolumn{1}{c|}{0.994} & \multicolumn{1}{c|}{0.993} & \multicolumn{1}{c|}{0.980} & 0.959 \\ \hline \rowcolor{lightblue} \textbf{Zlib} & \multicolumn{1}{c|}{0.990} & \multicolumn{1}{c|}{0.991} & \multicolumn{1}{c|}{0.990} & \multicolumn{1}{c|}{0.979} & 0.965 \\ \midrule[.9pt] \textbf{Average} & \multicolumn{1}{c|}{0.992} & \multicolumn{1}{c|}{0.993} & \multicolumn{1}{c|}{0.990} & \multicolumn{1}{c|}{0.967} & 0.958 \\ \bottomrule[1.1pt] \end{tabular} \end{table} \subsection{RQ1: Accuracy} \label{subsec:rq1} We evaluate how accurate \textsc{Trex}\xspace is in matching similar functions across different architectures, optimizations, and obfuscations. As shown in Table~\ref{tab:result}, we prepare function pairs for each project (first column) with 5 types of partitions. (1) ARCH: the function pairs have \emph{different architectures} but same optimizations without obfuscations (2nd column). (2) OPT: the function pairs have \emph{different optimizations} but same architectures without obfuscations (3rd column). (3) OBF: the function pairs have \emph{different obfuscations} with same architectures (x64) and no optimization (4th column). (4) ARCH+OPT: the function pairs have \emph{both different architectures and optimizations} without obfuscations (5th column). (5) ARCH+OPT+OBF: the function pairs can come from arbitrary architectures, optimizations, and obfuscations (6th column). Table~\ref{tab:result} reports the mean testing AUC scores of \textsc{Trex}\xspace on each project with 3 runs. On average, \textsc{Trex}\xspace achieves $>0.958$ (and up to 0.995) AUC scores, even in the most challenging setting where the functions can come from different architectures, optimizations, and obfuscations at the same time. We note that \textsc{Trex}\xspace performs the best on cross-optimization matching. This is intuitive as the syntax of two functions from different optimizations are not changed significantly (\emph{e.g., } the name of opcode, operands remain the same). Nevertheless, we find the AUC scores for matching functions from different architectures is only 0.001 lower, which indicates the model is robust to entirely different syntax between two architectures. On matching functions with different obfuscations, \textsc{Trex}\xspace's results are 0.026, on average, lower than that of cross-optimizations, which indicates the obfuscation changes the code more drastically. Section~\ref{subsec:rq2} will show the specific results of \textsc{Trex}\xspace on each obfuscations. \subsection{RQ2: Baseline Comparison} \label{subsec:rq2} \vspace{.1cm}\noindent\textbf{Cross-architecture search.} As described in Section~\ref{sec:impl}, we first compare \textsc{Trex}\xspace with SAFE and Gemini on OpenSSL-1.0.1f and OpenSSL-1.0.1u with their reported numbers (as they only evaluate on these two projects). We then run SAFE's released model on our dataset and compare to \textsc{Trex}\xspace. We use our testing setup (see Section~\ref{sec:impl}) to evaluate SAFE's trained model, where 90\% of the total functions of those in Table~\ref{tab:dataset} are used to construct the testing set. These testing sets are much larger than that in SAFE, where they only evaluate on 20\% of the OpenSSL functions. Note that the dataset used in SAFE are all compiled by \texttt{GCC-5.4} at the time when it is publicized (November 2018), while ours are compiled by \texttt{GCC-7.5} (April 2020). We find these factors (the more diverse dataset and different compilers) can all lead to the possible dataset distribution shift, which often results in the decaying performance of ML models when applied in the security applications~\cite{jordaney2017transcend}. To study the distribution shift, we measure the KL-divergence~\cite{kullback1951information} between SAFE's dataset (OpenSSL compiled by \texttt{GCC-5.4}) and our dataset. Specifically, we use the probability distribution of the raw bytes of the compiled projects' binaries, and compute their KL-divergence between SAFE and ours. As OpenSSL is also a subset of our complete dataset, we compute the KL-divergence between our compiled OpenSSL and that of SAFE as the baseline. We find the KL-divergence is 0.02 between SAFE's dataset and ours, while it decreases to 0.0066 between our compiled OpenSSL and that of SAFE. This indicates that the underlying distribution of our test set shifts from that of SAFE's. Moreover, the KL-divergence of 0.0066 between the same dataset (OpenSSL) but only compiled by different GCC versions implies that the compiler version has much smaller effect to the distribution shift than the different software projects. \begin{figure}[!t] \centering \includegraphics[width=0.8\linewidth]{./figs/compare/roc-drawio-safe-gemini.pdf} \caption{ROC curves of matching functions in OpenSSL across different architectures. \textsc{Trex}\xspace outperforms the reported results of SAFE and Gemini and the results of running SAFE's trained model on our testing set.} \label{fig:cross-arch-roc} \end{figure} As shown in Figure~\ref{fig:cross-arch-roc}, our ROC curve is higher than those reported in SAFE and Gemini. While SAFE's reported AUC score, 0.992, is close to ours, when we run their trained model on our testing set, its AUC score drops to 0.976 -- possibly because our testing set is much larger than theirs. This observation demonstrates the generalizability of \textsc{Trex}\xspace -- when pretrained to approximately learn execution semantics explicitly, it can quickly generalize to match unseen (semantically similar) functions with only a minimal training set. \begin{figure}[!t] \centering \includegraphics[scale=.4]{./figs/compare/cross-arch-auc-legend.pdf} \vspace{.1cm} \includegraphics[width=\linewidth]{./figs/compare/cross-arch-auc.pdf} \caption{Comparison between \textsc{Trex}\xspace and SAFE on matching functions in each project compiled to different architectures (see Table~\ref{tab:dataset}).} \label{fig:cross-arch-auc-safe} \end{figure} Figure~\ref{fig:cross-arch-auc-safe} shows that \textsc{Trex}\xspace consistently outperforms SAFE on all projects, \emph{i.e., } by 7.3\% on average. As the SAFE's model is only trained on OpenSSL, we also follow the same setting by training \textsc{Trex}\xspace on only OpenSSL, similar to the cross-project setting described in Section~\ref{subsec:rq1}. \vspace{.1cm}\noindent\textbf{Cross-optimization search.} We compare \textsc{Trex}\xspace with Asm2Vec and BLEX on matching functions compiled by different optimizations. As both Asm2vec and Blex run on single architecture, we restrict the comparison on x64. Besides, since Asm2Vec uses Precision@1 and Blex uses accuracy as the metric (discussed in Section~\ref{sec:impl}), we compare with each tool separately using their metrics and on their evaluated dataset. \begin{table}[!t] \footnotesize \setlength{\tabcolsep}{8pt} \centering \renewcommand{\arraystretch}{1} \setlength\aboverulesep{0.4pt} \setlength\belowrulesep{0.4pt} \caption{Comparison between \textsc{Trex}\xspace and Asm2Vec (in Precision@1) on function pairs across optimizations. } \label{tab:result-asm2vec} \begin{tabular}{r?{1.1pt}cccc} \toprule[1.1pt] \multicolumn{1}{c?{1.1pt}}{} & \multicolumn{4}{c}{Cross Compiler Optimization} \\ \cline{2-5} \multicolumn{1}{l?{1.1pt}}{} & \multicolumn{2}{c?{1.1pt}}{\texttt{O2} and \texttt{O3}} & \multicolumn{2}{c}{\texttt{O0} and \texttt{O3}} \\ \cline{2-5} \multicolumn{1}{c?{1.1pt}}{\textbf{}} & \multicolumn{1}{c|}{\textsc{Trex}\xspace} & \multicolumn{1}{c?{1.1pt}}{Asm2Vec} & \multicolumn{1}{c|}{\textsc{Trex}\xspace} & Asm2Vec \\ \midrule[.9pt] \rowcolor{lightblue} Coreutils & \multicolumn{1}{c|}{\textbf{0.955}} & \multicolumn{1}{c?{1.1pt}}{0.929} & \multicolumn{1}{c|}{\textbf{0.913}} & 0.781 \\ \hline Curl & \multicolumn{1}{c|}{\textbf{0.961}} & \multicolumn{1}{c?{1.1pt}}{0.951} & \multicolumn{1}{c|}{\textbf{0.894}} & 0.850 \\ \hline \rowcolor{lightblue} GMP & \multicolumn{1}{c|}{\textbf{0.974}} & \multicolumn{1}{c?{1.1pt}}{0.973} & \multicolumn{1}{c|}{\textbf{0.886}} & 0.763 \\ \hline ImageMagick & \multicolumn{1}{c|}{\textbf{0.971}} & \multicolumn{1}{c?{1.1pt}}{0.971} & \multicolumn{1}{c|}{\textbf{0.891}} & 0.837 \\ \hline \rowcolor{lightblue} LibTomCrypt & \multicolumn{1}{c|}{\textbf{0.991}} & \multicolumn{1}{c?{1.1pt}}{0.991} & \multicolumn{1}{c|}{\textbf{0.923}} & 0.921 \\ \hline OpenSSL & \multicolumn{1}{c|}{\textbf{0.982}} & \multicolumn{1}{c?{1.1pt}}{0.931} & \multicolumn{1}{c|}{\textbf{0.914}} & 0.792 \\ \hline \rowcolor{lightblue} PuTTy & \multicolumn{1}{c|}{\textbf{0.956}} & \multicolumn{1}{c?{1.1pt}}{0.891} & \multicolumn{1}{c|}{\textbf{0.926}} & 0.788 \\ \hline SQLite & \multicolumn{1}{c|}{\textbf{0.931}} & \multicolumn{1}{c?{1.1pt}}{0.926} & \multicolumn{1}{c|}{\textbf{0.911}} & 0.776 \\ \hline \rowcolor{lightblue} Zlib & \multicolumn{1}{c|}{\textbf{0.890}} & \multicolumn{1}{c?{1.1pt}}{0.885} & \multicolumn{1}{c|}{\textbf{0.902}} & 0.722 \\ \midrule[.9pt] Average & \multicolumn{1}{c|}{\textbf{0.957}} & \multicolumn{1}{c?{1.1pt}}{0.939} & \multicolumn{1}{c|}{\textbf{0.907}} & 0.803 \\ \bottomrule[1.1pt] \end{tabular} \end{table} Table~\ref{tab:result-asm2vec} shows \textsc{Trex}\xspace outperforms Asm2Vec in Precision@1 (by 7.2\% on average) on functions compiled by different optimizations (\emph{i.e., } between \texttt{O2} and \texttt{O3} and between \texttt{O0} and \texttt{O3}). As the syntactic difference introduced by optimizations between \texttt{O0} and \texttt{O3} is more significant than that between \texttt{O2} and \texttt{O3}, both tools have certain level of decrease in AUC scores (5\% drop for \textsc{Trex}\xspace and 14\% for Asm2Vec), but \textsc{Trex}\xspace's AUC score drops much less than that of Asm2Vec. To compare to Blex, we evaluate \textsc{Trex}\xspace on Coreutils between optimizations \texttt{O0} and \texttt{O3}, where they report to achieve better performance than BinDiff~\cite{bindiff}. As Blex show the matched functions of each individual utility in Coreutils in a barchart without including the concrete numbers of matched functions, we estimate their matched functions using their reported average percentage (75\%) on all utilities. \begin{figure}[!t] \centering \includegraphics[width=.95\linewidth]{./figs/compare/cross-opt-blex.pdf} \caption{Cross-optimization function matching between \texttt{O0} and \texttt{O3} on Coreutils by \textsc{Trex}\xspace and Blex. We sort the 109 utility binaries in Coreutils by their number of functions, and aggregate the matched functions every 10 utilities.} \label{fig:cross-opt-blex} \end{figure} Figure~\ref{fig:cross-opt-blex} shows that \textsc{Trex}\xspace consistently outperforms Blex in number of matched functions in all utility programs of Coreutils. Note that Blex also executes the function and uses the dynamic features to match binaries. The observation here thus implies that the learned execution semantics from \textsc{Trex}\xspace is more effective than the hand-coded features in Blex for matching similar binaries. \vspace{.1cm}\noindent\textbf{Cross-obfuscation search.} We compare \textsc{Trex}\xspace to Asm2Vec on matching obfuscated function binaries with different obfuscation methods. Notably, Asm2Vec is evaluated on obfuscations including bogus control flow (\texttt{bcf}), control flow flattening (\texttt{cff}), and instruction substitution (\texttt{sub}), which are subset of our evaluated obfuscations (Table~\ref{tab:dataset}). As Asm2Vec only evaluates on 4 projects, \emph{i.e., } GMP, ImageMagic, LibTomCrypt, and OpenSSL, we focus on these 4 projects and \textsc{Trex}\xspace's results for other projects are included in Table~\ref{tab:result}. \begin{table}[!t] \footnotesize \setlength{\tabcolsep}{2.5pt} \centering \renewcommand{\arraystretch}{1} \setlength\aboverulesep{0.4pt} \setlength\belowrulesep{0.4pt} \caption{Comparison between \textsc{Trex}\xspace and Asm2Vec (in Precision@1) on function pairs across differet obfuscations. } \label{tab:cross-obf-asm2vec} \begin{tabular}{r|r?{1.1pt}l|l|l|l?{1.1pt}l} \toprule[1.1pt] & & GMP & LibTomCrypt & ImageMagic & OpenSSL & Average \\ \midrule[.9pt] \rowcolor{lightblue} & \textsc{Trex}\xspace & \textbf{0.926} & \textbf{0.938} & \textbf{0.934} & \textbf{0.898} & \textbf{0.924} \\ \cline{2-7} \rowcolor{lightblue} \multirow{-2}{*}{\texttt{bcf}} & Asm2Vec & 0.802 & 0.920 & 0.933 & 0.883 & 0.885 \\ \midrule[.9pt] \multirow{2}{*}{\texttt{ccf}} & \textsc{Trex}\xspace & \textbf{0.943} & \textbf{0.931} & \textbf{0.936} & \textbf{0.940} & \textbf{0.930} \\ \cline{2-7} & Asm2Vec & 0.772 & 0.920 & 0.890 & 0.795 & 0.844 \\ \midrule[.9pt] \rowcolor{lightblue} & \textsc{Trex}\xspace & \textbf{0.949} & \textbf{0.962} & \textbf{0.981} & \textbf{0.980} & \textbf{0.968} \\ \cline{2-7} \rowcolor{lightblue} \multirow{-2}{*}{\texttt{sub}} & Asm2Vec & 0.940 & 0.960 & 0.981 & 0.961 & 0.961 \\ \midrule[.9pt] \multirow{2}{*}{All} & \textsc{Trex}\xspace & \textbf{0.911} & \textbf{0.938} & \textbf{0.960} & \textbf{0.912} & \textbf{0.930} \\ \cline{2-7} & Asm2Vec & 0.854 & 0.880 & 0.830 & 0.690 & 0.814 \\ \bottomrule[1.1pt] \end{tabular} \end{table} Table~\ref{tab:cross-obf-asm2vec} shows \textsc{Trex}\xspace achieves better Precision@1 score (by 14.3\% on average) throughout all different obfuscations. Importantly, the last two rows show when multiple obfuscations are combined, \textsc{Trex}\xspace performance is not dropping as significant as Asm2Vec. It also shows \textsc{Trex}\xspace remains robust under varying obfuscations with different difficulties. For example, instruction substitution simply replaces very limited instructions (\emph{i.e., } arithmetic operations as shown in Section~\ref{sec:overview}) while control flow flattening dramatically changes the function code. Asm2Vec has 12.2\% decreased score when the obfuscation is changed from \texttt{sub} to \texttt{ccf}, while \textsc{Trex}\xspace only decreases by 4\%. \subsection{RQ3: Execution Time} \label{subsec:rq3} We evaluate the runtime performance of generating function embeddings for computing similarity. We compare \textsc{Trex}\xspace with SAFE and Gemini on generating functions in 4 projects on x64 compiled by \texttt{O3}, \emph{i.e., } Binutils, Putty, Findutils, and Diffutils, which have disparate total number of functions (see Table~\ref{tab:dataset}. This tests how \textsc{Trex}\xspace scales to different number of functions compared to other baselines. Since the offline training (\emph{i.e., } pretraining \textsc{Trex}\xspace) of all the learning-based tools is a one-time cost, it can be amortized in the function matching process so we do not explicitly measure the training time. Moreover, the output of all tools are function embeddings, which can be indexed and efficiently searched using locality sensitive hashing (LSH)~\cite{gionis1999similarity, rajaraman2011mining}. Therefore, we do not compare the matching time of function embeddings as it simply depends on the performance of underlying LSH implementation. Particularly, we compare the runtime of two procedures in matching functions. (1) Function parsing, which transforms the function binaries into the format that the model needs. (2) Embedding generation, which takes the parsed function binary as input and computes function embedding. We test the embedding generation using our GPU (see Section~\ref{sec:impl}). \begin{figure}[!t] \centering \includegraphics[scale=.4]{./figs/runtime/runtime_bar_legend.pdf} \subfloat[Function parsing]{ \includegraphics[width=0.47\linewidth]{./figs/runtime/runtime_bar_parse.pdf} \label{subfig:runtime_bar_parse}} \subfloat[Embedding generation]{ \includegraphics[width=0.47\linewidth]{./figs/runtime/runtime_bar_emb.pdf} \label{subfig:runtime_bar_emb}} \caption{Runtime performance (lower is better) of \textsc{Trex}\xspace, SAFE, and Gemini on (a) function parsing and (b) embedding generation. The time is log-scaled.} \label{fig:runtime} \end{figure} Figure~\ref{fig:runtime} shows that \textsc{Trex}\xspace is more efficient than the other tools in both function parsing and embedding generation for functions from 4 different projects with different number of functions (Table~\ref{tab:dataset}). Gemini requires manually constructing control flow graph and extracting inter-/intra-basic-block feature engineering. It thus incurs the largest overhead. For generating function embeddings, our underlying network architectures leverage Transformer self-attention layers, which is more amenable to parallezation with GPU than the recurrent (used by SAFE) and graph neural network (used by Gemini)~\cite{vaswani2017attention}. As a result, \textsc{Trex}\xspace runs up to 8$\times$ faster than SAFE and Gemini. \subsection{RQ4: Ablation Study} \label{subsec:rq4} In this section, we aim to quantify how much each key component in \textsc{Trex}\xspace's design helps the end results. We first study how much does pretraining, argued to assist learning approximate execution semantics, help match function binaries. We then study how does pretraining \emph{without micro-traces} affect the end results. We also test how much does incorporating the micro-traces in the pretraining tasks improve the accuracy. \vspace{.1cm}\noindent\textbf{Pretraining effectiveness.} We compare the testing between AUC scores achieved by \textsc{Trex}\xspace (1) with pretraining (except the target project that will be finetuned), (2) with 66\% of pretraining functions in (1), (3) with 33\% of pretraining functions in (1), and (4) without pretraining (the function embedding is computed by randomly-initialized model weights that are not pretrained). The function pairs can come from arbitrary architectures, optimizations, and obfuscations. \begin{figure}[!t] \centering \includegraphics[scale=.4]{./figs/ablation/ablation-pretrain-effective-legend.pdf} \includegraphics[width=.95\linewidth]{./figs/ablation/ablation-pretrain-effective.pdf} \caption{Comparison of testing AUC scores between models pretrained with different fraction of the pretraining set.} \label{fig:ablation-pretrain} \end{figure} Figure~\ref{fig:ablation-pretrain} shows that the model's AUC score drops significantly (on average 15.7\%) when the model is not pretrained. Interestingly, we observe that the finetuned models achieve similar AUC scores, \emph{i.e., } with only 1\% decrease when pretrained with 33\% of the functions compared to pretraining with all functions. This indicates that we can potentially decrease our pretraining data size to achieve roughly the same performance. However, since our pretraining task does not require collect any label, we can still collect unlimited binary code found in the wild to enrich the semantics that can be observed. \vspace{.1cm}\noindent\textbf{Pretraining w/o micro-traces.} We try to understand whether including the micro-traces in pretraining can really help the model to learn better execution semantics than learning from only static assembly code, which in turn results in better function matching accuracy. Specifically, we pretrain the model on the data that contains only dummy value sequence (see Section~\ref{sec:method}), and follow the same experiment setting as described above. Besides replacing the input value sequence as dummy value, we accordingly remove the prediction of dynamic values in the pretraining objective (Equation~\ref{eq:pretrain}). We also compare to SAFE in this case. Figure~\ref{fig:ablation-pretrain-finetune-trace} shows that the AUC scores decrease by 7.2\% when the model is pretrained without micro-trace. The AUC score of pretrained \textsc{Trex}\xspace without micro-traces is even 0.035 lower than that of SAFE. However, the model still performs reasonably well, achieving 0.88 AUC scores even when the functions can come from arbitrary architectures, optimizations, and obfuscations. Moreover, we observe that pretraining without micro-traces has less performance drop than the model simply not pretrained (7.2\% vs. 15.7\%). This demonstrates that even pretraining with only static assembly code is indeed helpful to improve matching functions. One possible interpretation is that similar functions are statically similar in syntax, while understanding their inherently similar execution semantics just further increases the similarity score. \begin{figure}[!t] \centering \includegraphics[scale=.3]{./figs/ablation/ablation-pretrain-finetune-trace-legend.pdf} \includegraphics[width=\linewidth]{./figs/ablation/ablation-pretrain-finetune-trace.pdf} \caption{Comparison of testing AUC scores between models pretrained \emph{with} micro-trace and pretrained \emph{without} micro-trace.} \label{fig:ablation-pretrain-finetune-trace} \end{figure} \vspace{.1cm}\noindent\textbf{Pretraining accuracy.} We study the testing perplexity (PPL defined in Section~\ref{sec:impl}) of pretraining \textsc{Trex}\xspace to directly validate whether it indeed helps \textsc{Trex}\xspace to learn the approximate execution semantics. The rationale is the good performance, \emph{i.e., } the low PPL, on unseen testing function binaries indicates that \textsc{Trex}\xspace highly likely learns to generalize based on its learned approximate execution semantics. \begin{figure}[!t] \centering \includegraphics[width=0.85\linewidth]{./figs/pretrain/input-combine.pdf} \caption{Testing PPL of pretraining \textsc{Trex}\xspace in 10 epochs. We compare different designs of combining byte-sequence in the micro-trace (see Section~\ref{subsec:input_repr}).} \label{fig:pretrain-loss} \end{figure} The green line in Figure~\ref{fig:pretrain-loss} shows the validation PPL in 10 epochs of pretraining \textsc{Trex}\xspace. The testing set is constructed by sampling 10,000 random functions from the projects used in pretraining (as described in Section~\ref{sec:impl}). We observe that the PPL of \textsc{Trex}\xspace drops to close to 2.1, which is far below that of random guessing (\emph{e.g., } random guessing PPL is $2^{-\log(1/256)}=256$), indicating that \textsc{Trex}\xspace has around 0.48 confidence on average on the masked tokens being the correct value. Note that a random guessing has only 1/256=0.004 confidence on the masked tokens being the correct value. \section{Implementation and Experimental Setup} \label{sec:impl} We implement \textsc{Trex}\xspace using fairseq, a sequence modeling toolkit~\cite{ott2019fairseq}, based on PyTorch 1.6.0 with CUDA 10.2 and CUDNN 7.6.5. We run all experiments on a Linux server running Ubuntu 18.04, with an Intel Xeon 6230 at 2.10GHz with 80 virtual cores including hyperthreading, 385GB RAM, and 8 Nvidia RTX 2080-Ti GPUs. \vspace{.1cm}\noindent\textbf{Datasets.} To train and evaluate \textsc{Trex}\xspace, we collect 13 popular open-source software projects. These projects include Binutils-2.34, Coreutils-8.32, Curl-7.71.1, Diffutils-3.7, Findutils-4.7.0, GMP-6.2.0, ImageMagick-7.0.10, Libmicrohttpd-0.9.71, LibTomCrypt-1.18.2, OpenSSL-1.0.1f and OpenSSL-1.0.1u, PuTTy-0.74, SQLite-3.34.0, and Zlib-1.2.11. We compile these projects into 4 architectures \emph{i.e., } x86, x64, ARM (32-bit), and MIPS (32-bit), with 4 optimization levels (OPT), \emph{i.e., } \texttt{O0}, \texttt{O1}, \texttt{O2}, and \texttt{O3}, using \texttt{GCC-7.5}. Specifically, we compile the software projects based on its makefile, by specifying \texttt{CFLAGS} (to set optimization flag), \texttt{CC} (to set cross-compiler), and \texttt{--host} (to set the cross-compilation target architecture). We always compile to dynamic shared objects, but resort to static linking when we encounter build errors. We are able to compile all projects with these treatments. We also obfuscate all projects using 5 types of obfuscations (OBF) by Hikari~\cite{hikari} on x64 -- an obfuscator based on \texttt{clang-8}. The obfuscations include bogus control flow (\texttt{bcf}), control flow flattening (\texttt{cff}), register-based indirect branching (\texttt{ibr}), basic block splitting (\texttt{spl}), and instruction substitution (\texttt{sub}). As we encounter several errors in cross-compilation using Hikari (based on Clang)~\cite{hikari}, and the baseline system (\emph{i.e., } Asm2Vec~\cite{ding2019asm2vec}) to which we compare only evaluates on x64, we restrict the obfuscated binaries for x64 only. As a result, we have 1,472,066 functions, as shown in Table~\ref{tab:dataset}. \begin{table*}[!t] \footnotesize \setlength{\tabcolsep}{2.2pt} \centering \renewcommand{\arraystretch}{1} \setlength\aboverulesep{0.2pt} \setlength\belowrulesep{0.2pt} \caption{Number of functions for each project across 4 architectures with 4 optimization levels and 5 obfuscations. } \label{tab:dataset} \rowcolors{1}{}{lightblue} \begin{tabular}{c|c|l|l|l|l|l|l|l|l|l|l|l|l|l?{1.1pt}r} \toprule[1.1pt] \multirow{2}{*}{ARCH} & \multirow{2}{*}{\begin{tabular}[c]{@{}l@{}}OPT\\ OBF\end{tabular}} & \multicolumn{14}{c}{\bf \# Functions} \\ \cline{3-16} & & \multicolumn{1}{c|}{Binutils} & \multicolumn{1}{c|}{Coreutils} & \multicolumn{1}{c|}{Curl} & \multicolumn{1}{c|}{Diffutils} & \multicolumn{1}{c|}{Findutils} & \multicolumn{1}{c|}{GMP} & \multicolumn{1}{c|}{ImageMagick} & \multicolumn{1}{c|}{Libmicrohttpd} & \multicolumn{1}{c|}{LibTomCrypt} & \multicolumn{1}{c|}{OpenSSL} & \multicolumn{1}{c|}{PuTTy} & \multicolumn{1}{c|}{SQLite} & \multicolumn{1}{c?{1.1pt}}{Zlib} & \multicolumn{1}{c}{\textbf{Total}} \\ \midrule[.9pt] & \texttt{O0} & 25,492 & 19,992 & 1,067 & 944 & 1,529 & 766 & 2,938 & 200 & 779 & 11,918 & 7,087 & 2,283 & 157 & 75,152 \\ \cline{2-16} & \texttt{O1} & 20,043 & 14,918 & 771 & 694 & 1,128 & 704 & 2,341 & 176 & 745 & 10,991 & 5,765 & 1,614 & 143 & 60,033 \\ \cline{2-16} & \texttt{O2} & 19,493 & 14,778 & 765 & 693 & 1,108 & 701 & 2,358 & 171 & 745 & 11,001 & 5,756 & 1,473 & 138 & 59,180 \\ \cline{2-16} & \texttt{O3} & 17,814 & 13,931 & 697 & 627 & 983 & 680 & 2,294 & 160 & 726 & 10,633 & 5,458 & 1,278 & 125 & 55,406 \\ \cline{2-16} \multirow{-5}{*}{ARM} & \multicolumn{14}{c?{1.1pt}}{Total \# Functions of ARM} & \textbf{249,771} \\ \midrule[.5pt] \multirow{5}{*}{MIPS} & \texttt{O0} & 28,460 & 18,843 & 1,042 & 906 & 1,463 & 734 & 2,840 & 200 & 779 & 11,866 & 7,003 & 2,199 & 153 & 76,488 \\ \cline{2-16} & \texttt{O1} & 22,530 & 13,771 & 746 & 653 & 1,059 & 670 & 2,243 & 176 & 745 & 10,940 & 5,685 & 1,530 & 139 & 60,887 \\ \cline{2-16} & \texttt{O2} & 22,004 & 13,647 & 741 & 653 & 1,039 & 667 & 2,260 & 171 & 743 & 10,952 & 5,677 & 1,392 & 135 & 60,081 \\ \cline{2-16} & \texttt{O3} & 20,289 & 12,720 & 673 & 584 & 917 & 646 & 2,198 & 161 & 724 & 10,581 & 5,376 & 1,197 & 121 & 56,187 \\ \cline{2-16} & \multicolumn{14}{c?{1.1pt}}{Total \# Functions of MIPS} & \textbf{253,643} \\ \midrule[.5pt] & \texttt{O0} & 37,783 & 24,383 & 1,335 & 1,189 & 1,884 & 809 & 3,876 & 326 & 818 & 12,552 & 7,548 & 2,923 & 204 & 95,630 \\ \cline{2-16} & \texttt{O1} & 32,263 & 20,079 & 1,013 & 967 & 1,516 & 741 & 3,482 & 280 & 782 & 11,578 & 6,171 & 2,248 & 196 & 81,316 \\ \cline{2-16} & \texttt{O2} & 32,797 & 21,082 & 1,054 & 1,006 & 1,524 & 728 & 3,560 & 265 & 784 & 11,721 & 6,171 & 2,113 & 183 & 82,988 \\ \cline{2-16} & \texttt{O3} & 34,055 & 22,482 & 1,020 & 1,052 & 1,445 & 707 & 3,597 & 284 & 760 & 11,771 & 5,892 & 1,930 & 197 & 85,192 \\ \cline{2-16} \multirow{-5}{*}{x86} & \multicolumn{14}{c?{1.1pt}}{Total \# Functions of x86} & \textbf{358,261} \\ \midrule[.5pt] \multirow{5}{*}{x64} & \texttt{O0} & 26,757 & 17,238 & 1,034 & 845 & 1,386 & 751 & 2,970 & 200 & 782 & 12,047 & 7,061 & 2,190 & 151 & 73,412 \\ \cline{2-16} & \texttt{O1} & 21,447 & 12,532 & 739 & 600 & 1,000 & 691 & 2,358 & 176 & 745 & 11,120 & 5,728 & 1,523 & 137 & 58,796 \\ \cline{2-16} & \texttt{O2} & 20,992 & 12,206 & 734 & 596 & 976 & 689 & 2,374 & 171 & 742 & 11,136 & 5,703 & 1,380 & 132 & 57,831 \\ \cline{2-16} & \texttt{O3} & 19,491 & 11,488 & 662 & 536 & 857 & 667 & 2,308 & 160 & 725 & 10,768 & 5,390 & 1,183 & 119 & 54,354 \\ \cline{2-16} & \multicolumn{14}{c?{1.1pt}}{Total \# Functions of x64} & \textbf{244,393} \\ \midrule[1.1pt] & \texttt{bcf} & 27,734 & 17,093 & 998 & 840 & 1,388 & 746 & 2,833 & 200 & 782 & 10,768 & 7,069 & 2,183 & 151 & 72,785 \\ \cline{2-16} & \texttt{cff} & 27,734 & 17,093 & 998 & 840 & 1,388 & 746 & 2,833 & 200 & 782 & 10,903 & 7,069 & 2,183 & 151 & 72,920 \\ \cline{2-16} & \texttt{ibr} & 27,734 & 17,105 & 998 & 842 & 1,392 & 746 & 2,833 & 204 & 782 & 12,045 & 7,069 & 2,183 & 151 & 74,084 \\ \cline{2-16} & \texttt{spl} & 27,734 & 17,093 & 998 & 840 & 1,388 & 746 & 2,833 & 200 & 782 & 10,772 & 7,069 & 2,183 & 151 & 72,789 \\ \cline{2-16} & \texttt{sub} & 27,734 & 17,093 & 998 & 840 & 1,388 & 746 & 2,833 & 200 & 782 & 11,403 & 7,069 & 2,183 & 151 & 73,420 \\ \cline{2-16} \multirow{-5}{*}{x64} & \multicolumn{14}{c?{1.1pt}}{Total \# Obfuscated Functions} & \textbf{365,998} \\ \midrule[1.1pt] \multicolumn{15}{c?{1.1pt}}{Total \# Functions} & \textbf{1,472,066} \\ \bottomrule[1.1pt] \end{tabular} \end{table*} \vspace{.1cm}\noindent\textbf{Micro-tracing.} We implement micro-tracing by Unicorn~\cite{quynh2015unicorn}, a cross-platform CPU emulator based on QEMU~\cite{bellard2005qemu}. We micro-execute each function 3 times with different initialized registers and memory, generating 3 micro-traces (including both static code and dynamic values) for pretraining (masked LM). We leverage multi-processing to parallelize micro-executing each function and set 30 seconds as the timeout for each run in case any instruction gets stuck (\emph{i.e., } infinite loops). For each function (with 3 micro-traces), we append one additional dummy trace, which consists of only dummy values (\texttt{\#\#}). This setting encourages the model to leverage its learned execution semantics (from other traces with concrete dynamic values) to predict the masked code with \emph{only} code context, which helps the finetuning task as we only feed the functions' static code as input (discussed in Section~\ref{sec:intro}). \vspace{.1cm}\noindent\textbf{Baselines.} For comparing cross-architecture performance, we consider 2 state-of-the-art systems. The first one is SAFE~\cite{massarelli2019safe} that achieves state-of-the-art function matching accuracy. As SAFE's trained model is publicly available, we compare \textsc{Trex}\xspace with both SAFE's reported results on their dataset, \emph{i.e., } OpenSSL-1.0.1f and OpenSSL-1.0.1u, and running their trained models on all of our collected binaries. The second baseline is Gemini~\cite{xu2017neural}, which is shown outperformed by SAFE. As Gemini does not release their trained models, we compare our results to their reported numbers directly on their evaluated dataset, \emph{i.e., } OpenSSL-1.0.1f and OpenSSL-1.0.1u. For cross-optimization/obfuscation comparison, we consider Asm2Vec~\cite{ding2019asm2vec} and Blex~\cite{egele2014blanket} as the baselines. Asm2Vec achieves the state-of-the-art cross-optimization/obfuscation results, based on learned embeddings from static assembly code. Blex, on the other hand, leverages functions' dynamic behavior to match function binaries. As we only find a third-party implementation of Asm2Vec that achieves extremely low Precision@1 (the metric used in Asm2Vec) from our testing (\emph{e.g., } 0.02 vs. their reported 0.814), and we have contacted the authors and do not get replies, we directly compare to their reported numbers. Blex is not publicly available either, so we also compare to their reported numbers directly. \vspace{.1cm}\noindent\textbf{Metrics.} As the cosine similarity between two function embeddings can be an arbitrary real value between -1 and 1, there has to be a mechanism to threshold the similarity score to determine whether the function pairs are similar or dissimilar. The chosen thresholds largely determine the model's predictive performance~\cite{pagani2018beyond}. To avoid the potential bias introduced by a specific threshold value, we consider the receiver operating characteristic (ROC) curve, which measures the model's false positives/true positives under different thresholds. Notably, we use the area under curve (AUC) of the ROC curve to quantify the accuracy of \textsc{Trex}\xspace to facilitate benchmarking -- the higher the AUC score, the better the model's accuracy. Certain baselines do not use AUC score to evaluate their system. For example, Asm2Vec uses Precision at Position 1 (Precision@1), and Blex uses the number of matched functions as the metric. Therefore, we also include these metrics to evaluate \textsc{Trex}\xspace when needed. Specifically, given a sequence of query functions and the target functions to be matched, Precision@1 measures the percentage of matched query functions in the target functions. Here ``match'' means the query function should have the \emph{top-1 highest} similarity score with the ground truth function among the target functions. To evaluate pretraining performance, we use the standard metric for evaluating the language model -- perplexity (PPL). The lower the PPL the better the pretrained model in predicting masked code. The lowest PPL is 1. \vspace{.1cm}\noindent\textbf{Pretraining setup.} To strictly separate the functions in pretraining, finetuning, and testing, we pretrain \textsc{Trex}\xspace on all functions in the dataset, \emph{except the functions in the project going to be finetuned and evaluated for similarity.} For example, the function matching results of Binutils (Table~\ref{tab:result}) are finetuned on the model pretrained on all other projects without Binutils. Note that pretraining \emph{is agnostic to any labels} (\emph{e.g., } ground-truth indicating similar functions). Therefore, we can always pretrain on large-scale codebases, which can potentially include the functions for finetuning (this is the common practice in transfer learning~\cite{devlin2018bert}). It is thus worth noting that our setup of separating functions for pretraining and finetuning makes our testing significantly more challenging. We keep the pretrained model weights that achieve the best validation PPL for finetuning. The validation set for pretraining consists of 10,000 random functions selected from Table~\ref{tab:dataset}. \vspace{.1cm}\noindent\textbf{Finetuning setup.} We choose 50,000 random function pairs for each project and randomly select \emph{only 10\%} for training, and the remaining is used as the testing set. We keep the training and testing functions \emph{strictly non-duplicated} by ensuring the functions that appear in training function pairs not appear in the testing. As opposed to the typical train-test split (\emph{e.g., } 80\% training and 20\% testing~\cite{massarelli2019safe}), our setting requires the model to generalize from few training samples to a large number of unseen testing samples, which alleviates the possibility of overfitting. Moreover, we keep the ratio between similar and dissimilar function pairs in the finetuning set as roughly 1:5. This setting follows the practice of contrastive learning~\cite{saunshi2019theoretical,chen2020simple}, respecting the actual distribution of similar/dissimilar functions as the number of dissimilar functions is often larger than that of similar functions in practice. \vspace{.1cm}\noindent\textbf{Hyperparameters.} We pretrain and finetune the models for 10 epochs and 30 epochs, respectively. We choose $\alpha=0.125$ in Equation~\ref{eq:pretrain} such that the cross-entropy loss of code prediction and value prediction have the same weight. We pick $\xi=0.1$ in Equation~\ref{eq:cosembloss} to make the model slightly inclined to treat functions as dissimilar because functions in practice are mostly dissimilar. We fix the largest input length to be 512 and split the functions longer than this length into subsequences for pretraining. We average the subsequences' embeddings during finetuning if the function is split to more than one subsequences. In this paper, we keep most of the hyperparameters fixed throughout the evaluation if not mentioned explicitly (complete hyperparameters are defined in Appendix~\ref{subsec:hyperparm}). While we can always search for better hyperparameters, there is no principled method to date~\cite{bergstra2012random}. We thus leave as future work a more thorough study of \textsc{Trex}\xspace's hyperparameters. \section{Introduction} \label{sec:intro} Semantic function similarity, which quantifies the behavioral similarity between two functions, is a fundamental program analysis capability with a broad spectrum of real-world security usages, such as vulnerability detection~\cite{brumley2008automatic}, exploit generation~\cite{avgerinos2014automatic}, tracing malware lineage~\cite{jang2013towards, bayer2009scalable}, and forensics~\cite{luo2014semantics}. For example, OWASP lists ``using components with known vulnerabilities'' as one of the top-10 application security risks in 2020~\cite{OWASP}. Therefore, identifying similar vulnerable functions in massive software projects can save significant manual effort. When matching semantically similar functions for security-critical applications (\emph{e.g., } vulnerability discovery), we often have to deal with software at \emph{binary level}, such as commercial off-the-shelf products (\emph{i.e., } firmware images) and legacy programs. However, this task is challenging as the functions' high-level information (\emph{i.e., } data structure definitions) are removed during the compilation process. The problem aggravates when the functions are compiled to run on different instruction set architectures with various compiler optimizations or even simple obfuscations. There has been a growing interest in using Machine Learning (ML) to tackle these challenges~\cite{xu2017neural, massarelli2019safe, ding2019asm2vec}. Notably, ML models learn function representations (\emph{i.e., } embeddings) from instructions in function binaries and match the functions based on the learned representation. These approaches have obtained the state-of-the-art results~\cite{xu2017neural, massarelli2019safe, ding2019asm2vec}, outperforming the traditional signature-based approaches~\cite{bindiff} that rely on the hand-coded features (\emph{e.g., } number of basic blocks) extracted by domain experts. The ML-based approaches are also amenable to large-scale function matching as computing similarity score as embedding distance is exceptionally efficient (\emph{e.g., } taking around 0.1 seconds searching over one million functions)~\cite{feng2016scalable}. \vspace{0.01cm}\noindent\textbf{Execution semantics.} Despite the impressive progress, it remains challenging for these approaches to match semantically similar functions with disparate syntax and structure~\cite{mckee2019software}. An inherent cause is that the code semantics is characterized by \emph{its execution effects}. However, all existing learning-based approaches are \emph{agnostic to program execution semantics}, taking the static code as the only input. Such a setting can easily lead a model into learning simple pattern matching, limiting their accuracy when such patterns are absent or changed~\cite{payer2014similarity, aghakhani2020malware}. Take a pair of x86 instructions as an example. \texttt{mov eax,2;lea ecx,[eax+4]} are semantically equivalent to \texttt{mov eax,2;lea ecx,[eax+eax*2]}. To determine their similarity statically, the most salient pattern is their common substring (both sequences share the tokens \texttt{mov}, \texttt{eax}, \texttt{lea}, \texttt{ecx}), which does not encode the key reason of equivalence. Without grasping the execution semantics, an ML model can easily learn such spurious patterns without understanding the inherent cause of the equivalence: \texttt{[eax+eax*2]} computes the same exact address as \texttt{[eax+4]} when \texttt{eax} is 2. \vspace{0.01cm}\noindent\textbf{Limitations of existing dynamic approaches.} Conceptually, dynamic analysis captures the perfect execution semantics. However, traditional dynamic analysis is inappropriate here as it is hard to find a concrete program input exercising a certain function of interest. While symbolic execution can help, it is known for poor scalability to large programs. Therefore, existing dynamic approaches~\cite{egele2014blanket} \emph{relax} the function initial states with random values that are often under-constrained -- they may not be feasible from executing the program regularly. For example, a function \texttt{bar()} can only be executed when called by function \texttt{foo()}. When directly executing \texttt{bar()}, the constraints of \texttt{bar()}'s initial states imposed in \texttt{foo()} is ignored. While these approaches can directly execute individual functions and compare their execution behavior, such a relaxation introduces \emph{noisy and incomplete traces}. For example, for functions taking only values from specified range, \emph{e.g., } $\sqrt{x}$ and $arcsin(x)$, the randomized function inputs (\emph{e.g., } negative numbers for $\sqrt{x}$ and numbers outside $[-1,1]$ for $arcsin(x)$) often trigger the similar behavior such as the exception handling. Therefore, existing dynamic approaches suffer from false positives~\cite{ding2019asm2vec}. \vspace{0.01cm}\noindent\textbf{Our approach.} This paper presents \textsc{Trex}\xspace (\emph{TRansfer-learning EXecution semantics}) that matches semantically similar functions based on their static instructions but augmented with execution semantics. Instead of using noisy and incomplete dynamic trace to directly determine similarity, \textsc{Trex}\xspace first pretrains the model to observe and understand how each instruction executes in a function. It then \emph{transfers} and \emph{composes} the learned knowledge from pretrained model to match semantically similar functions. Our extensive experiments suggest that the learned knowledge of executions in pretraining significantly boosts the performance of matching semantically similar function binaries -- \textsc{Trex}\xspace excels in matching functions from different architectures, optimizations, and obfuscations. \begin{figure*}[!t] \centering \includegraphics[width=\linewidth]{./figs/general.pdf} \caption{ The workflow of \textsc{Trex}\xspace. We first pretrain the model on the function code and its micro-traces (\emph{i.e., } dynamic register values) using the masked LM task. We then finetune the pretrained model on the function pairs (\emph{e.g., } same source function compiled to different architectures with different optimizations/obfuscations) by stacking new neural network layers for function similarity tasks. Finetuning updates both the pretrained model and the stacked layers. During inference, the finetuned model computes the function embedding, whose distance encodes the function similarity.} \label{fig:general} \end{figure*} Our key insight is that the function's behavior can be \emph{composed} by each instruction's execution semantics, which can be learned by observing and understanding how each instruction executes in the dynamic traces. For example, given instructions \texttt{mov eax,2;sub eax,1;add eax,3} and their dynamic traces, we can train the model to predict the value of \texttt{eax} after executing these instructions. This training objective is designed to force the model to understand the execution semantics of \texttt{mov}, \texttt{sub}, and \texttt{add}, \emph{i.e., } how they manipulate the operands and updates the results, and learn to \emph{compose their execution effects} to understand their overall behavior. We can then \emph{transfer} the model's learned knowledge to compose instructions' execution semantics to match semantically similar functions. As a result, our approach is not constrained by noisy and incomplete traces -- it matches functions based \emph{only} on their \emph{complete} static instructions but \emph{augmented} with rich knowledge of execution semantics. Such a design also saves significant runtime overhead as we do not need to execute any function on-the-fly when matching them~\cite{kapravelos2013revolver}. Given a sequence of assembly instructions and their micro-traces (\emph{i.e., } register values), we pretrain the model with the masked language modeling (masked LM) task. Notably, it masks random parts in the sequence and asks the model to predict masked parts based on their context. This design forces the model to learn how a function executes to correctly infer the missing values. It thus \emph{automates} learning execution semantics without any manual feature engineering. It also enables learning more \emph{comprehensive} execution semantics of diverse instructions, as we can mask various parts of the function, imposing the model to learn different aspects of execution semantics to make correct prediction. Moreover, this task is fully \emph{unsupervised}. Therefore, \textsc{Trex}\xspace can be trained and further improved using arbitrary binaries found in the wild. To encode both the static and dynamic information as the model inputs, we design a hierarchical Transformer~\cite{vaswani2017attention} that supports modeling concrete micro-trace values. By contrast, existing approaches often treat the numerical values as a dummy token~\cite{massarelli2019safe, ding2019asm2vec} to avoid prohibitively large vocabulary size, which cannot effectively learn the rich dependencies between concrete values that likely encode key function semantics. Moreover, our architecture's self-attention layer is designed to model long-range dependencies in a sequence~\cite{vaswani2017attention} efficiently. Therefore, \textsc{Trex}\xspace can support roughly 170$\times$ longer sequence and runs 8$\times$ faster than existing neural architectures, essential to learning embeddings of long function execution traces. We evaluate \textsc{Trex}\xspace on 1,472,066 functions collected from 13 popular open-source software projects across 4 architectures (x86, x64, ARM, and MIPS) and compiled with 4 optimizations (\texttt{O0}-\texttt{O3}), and 5 obfuscation strategies~\cite{hikari}. \textsc{Trex}\xspace outperforms the state-of-the-art systems by 7.8\%, 7.2\%, and 14.3\% in matching functions across different architectures, optimizations, and obfuscations, respectively. Our ablation studies show that the pretraining task significantly improves the accuracy of matching semantically similar functions (by 15.7\%). We also apply \textsc{Trex}\xspace in searching vulnerable functions in 180 real-world firmware images developed by well-known vendors and deployed in diverse embedded systems, including WLAN routers, smart cameras, and solar panels. Our case study shows that \textsc{Trex}\xspace assists in finding 16 real-world CVEs in these firmware images, which have not been disclosed in any previous studies. We make the following contributions. \begin{itemize} \item We propose a new approach to matching semantically similar functions: we first train the model to learn program execution semantics and then transfer the learned knowledge to match semantically similar functions. \item We implement micro-execution to collect dynamic traces on different architectures. We then develop a novel neural architecture -- hierarchical Transformer -- that can efficiently learn execution semantics from micro-traces. \item We implement \textsc{Trex}\xspace and evaluate it on 1,472,066 functions from 13 popular software projects and libraries. \textsc{Trex}\xspace outperforms the state-of-the-art tools by 7.8\%, 7\%, and 14.3\%, in cross-architecture, optimization, and obfuscation function matching, respectively, while running up to 8$\times$ faster. Moreover, \textsc{Trex}\xspace helps uncover 16 vulnerabilities in 180 real-world firmware images with latest version that are not disclosed by previous studies. We release \textsc{Trex}\xspace at \url{anonymous.4open.science/r/644f4ceb-ab22-4f04-a91c-0565e091ca31}. \end{itemize} \section{Introduction} \label{sec:intro} Semantic function similarity, which quantifies the behavioral similarity between two functions, is a fundamental program analysis capability with a broad spectrum of real-world security usages, such as vulnerability detection~\cite{brumley2008automatic}, exploit generation~\cite{avgerinos2014automatic}, tracing malware lineage~\cite{jang2013towards, bayer2009scalable}, and forensics~\cite{luo2014semantics}. For example, OWASP lists ``using components with known vulnerabilities'' as one of the top-10 application security risks in 2020~\cite{OWASP}. Therefore, identifying similar vulnerable functions in massive software projects can save significant manual effort. When matching semantically similar functions for security-critical applications (\emph{e.g., } vulnerability discovery), we often have to deal with software at \emph{binary level}, such as commercial off-the-shelf products (\emph{i.e., } firmware images) and legacy programs. However, this task is challenging, as the functions' high-level information (\emph{e.g., } data structure definitions) are removed during the compilation process. Establishing semantic similarity gets even harder when the functions are compiled to run on different instruction set architectures with various compiler optimizations or obfuscated with simple transformations. Recently, Machine Learning (ML) based approaches have shown promise in tackling these challenges~\cite{xu2017neural, massarelli2019safe, ding2019asm2vec} by learning robust features that can identify similar function binaries across different architectures, compiler optimizations, or even some types of obfuscation. Specifically, ML models learn function representations (\emph{i.e., } embeddings) from function binaries and use the distance between the embeddings of two functions to compute their similarity. The smaller the distance, the more similar the functions are to each other. Such approaches have achieved state-of-the-art results~\cite{xu2017neural, massarelli2019safe, ding2019asm2vec}, outperforming the traditional signature-based methods~\cite{bindiff} using hand-crafted features (\emph{e.g., } number of basic blocks). Such embedding distance-based strategy is particularly appealing for large-scale function matching---taking only around 0.1 seconds searching over one million functions~\cite{feng2016scalable}. \vspace{.1cm}\noindent\textbf{Execution semantics.} Despite the impressive progress, it remains challenging for these approaches to match semantically similar functions with disparate syntax and structure~\cite{mckee2019software}. An inherent cause is that the code semantics is characterized by \emph{its execution effects}. However, all existing learning-based approaches are \emph{agnostic to program execution semantics}, training only on the static code. Such a setting can easily lead a model into matching simple patterns, limiting their accuracy when such spurious patterns are absent or changed~\cite{payer2014similarity, aghakhani2020malware}. For instance, consider the following pair of x86 instructions: \texttt{mov eax,2;lea ecx,[eax+4]} are semantically equivalent to \texttt{mov eax,2;lea ecx,[eax+eax*2]}. An ML model focusing on syntactic features might pick common substrings (both sequences share the tokens \texttt{mov}, \texttt{eax}, \texttt{lea}, \texttt{ecx}) to establish their similarity, which does not encode the key reason of the semantic equivalence. Without grasping the approximate execution semantics, an ML model can easily learn such spurious patterns without understanding the inherent cause of the equivalence: \texttt{[eax+eax*2]} computes the same exact address as \texttt{[eax+4]} when \texttt{eax} is 2. \vspace{.1cm}\noindent\textbf{Limitations of existing dynamic approaches.} Existing dynamic approaches try to avoid the issues described above by directly comparing the dynamic behaviors of functions to determine similarity. As finding program inputs reaching the target functions is an extremely challenging and time-consuming task, the prior works perform under-constrained dynamic execution by initializing the function input states (\emph{e.g., } registers, memory) with random values and executing the target functions directly~\cite{egele2014blanket}. Unfortunately, using such under-constrained execution traces directly to compute function similarities often result in many false positives~\cite{ding2019asm2vec}. For example, providing random inputs to two different functions with strict input checks might always trigger similar shallow exception handling codes and might look spuriously similar. \vspace{.1cm}\noindent\textbf{Our approach.} This paper presents \textsc{Trex}\xspace (\emph{TRansfer-learning EXecution semantics}) that trains ML models to learn the approximate execution semantics from under-constrained dynamic traces. Unlike prior works, which use such traces to directly measure similarity, \textsc{Trex}\xspace pretrains the model on diverse traces to learn each instruction's execution effect in its context. \textsc{Trex}\xspace then finetunes the model by \emph{transferring} the learned knowledge from pretraining to match semantically similar functions (see Figure~\ref{fig:general}). Our extensive experiments suggest that the approximately learned knowledge of execution semantics in pretraining significantly boosts the accuracy of matching semantically similar function binaries -- \textsc{Trex}\xspace excels in matching functions from different architectures, optimizations, and obfuscations. \begin{figure*}[!t] \centering \includegraphics[width=\linewidth]{./figs/general.pdf} \caption{ The workflow of \textsc{Trex}\xspace. We first pretrain the model on the functions' micro-traces, consisting of both instructions and dynamic values, using the masked LM task. We then finetune the pretrained model on the semantically similar function (only static instruction) pairs by stacking new neural network layers for function similarity tasks. Finetuning updates both the pretrained model and the stacked layers. During inference, the finetuned model computes the function embedding, whose distance encodes the function similarity.} \label{fig:general} \end{figure*} Our key observation is that while under-constrained dynamic execution traces tend to contain many infeasible states, they still encode precise execution effects of many individual instructions. Thus, we can train an ML model to observe and learn the effect of different instructions present across a large number of under-constrained dynamic traces collected from diverse functions. Once the model has gained an approximate understanding of execution semantics of various instructions, we can train it to match semantically similar functions by leveraging its learned knowledge. As a result, during inference, we do not need to execute any functions on-the-fly while matching them~\cite{kapravelos2013revolver}, which saves significant runtime overhead. Moreover, our trained model does not need the under-constrained dynamic traces to match functions, it only uses the function instructions, but they are \emph{augmented} with rich knowledge of execution semantics. In this paper, we extend micro-execution~\cite{godefroid2014micro}, a form of under-constrained dynamic execution, to generate \emph{micro-traces} of a function across multiple instruction set architectures. A micro-trace consists of a sequence of aligned instructions and their corresponding program state values. We pretrain the model on a large number of micro-traces gathered from diverse functions as part of training data using the masked language modeling (masked LM) task. Notably, masked LM masks random parts in the sequence and asks the model to predict masked parts based on their context. This design forces the model to learn approximately how a function executes to correctly infer the missing values, which automates learning execution semantics without manual feature engineering. Masked LM is also fully \emph{self-supervised}~\cite{devlin2018bert} -- \textsc{Trex}\xspace can thus be trained and further improved with arbitrary functions found in the wild. To this end, we design a hierarchical Transformer~\cite{vaswani2017attention} that supports learning approximate execution semantics. Specifically, our architecture models micro-trace values explicitly. By contrast, existing approaches often treat the numerical values as a dummy token~\cite{massarelli2019safe, ding2019asm2vec} to avoid prohibitively large vocabulary size, which cannot effectively learn the rich dependencies between concrete values that likely encode key function semantics. Moreover, our architecture's self-attention layer is designed to model long-range dependencies in a sequence~\cite{vaswani2017attention} efficiently. Therefore, \textsc{Trex}\xspace can support roughly 170$\times$ longer sequence and runs 8$\times$ faster than existing neural architectures, essential to learning embeddings of long function execution traces. We evaluate \textsc{Trex}\xspace on 1,472,066 functions collected from 13 popular open-source software projects across 4 architectures (x86, x64, ARM, and MIPS) and compiled with 4 optimizations (\texttt{O0}-\texttt{O3}), and 5 obfuscation strategies~\cite{hikari}. \textsc{Trex}\xspace outperforms the state-of-the-art systems by 7.8\%, 7.2\%, and 14.3\% in matching functions across different architectures, optimizations, and obfuscations, respectively. Our ablation studies show that the pretraining task significantly improves the accuracy of matching semantically similar functions (by 15.7\%). We also apply \textsc{Trex}\xspace in searching vulnerable functions in 180 real-world firmware images developed by well-known vendors and deployed in diverse embedded systems, including WLAN routers, smart cameras, and solar panels. Our case study shows that \textsc{Trex}\xspace helps find 16 CVEs in these firmware images, which have not been disclosed in previous studies. We make the following contributions. \begin{itemize} \item We propose a new approach to matching semantically similar functions: we first train the model to learn approximate program execution semantics from micro-traces, a form of under-constrained dynamic traces, and then transfer the learned knowledge to identify semantically similar functions. \item We extend micro-execution to support different architectures to collect micro-traces for training. We then develop a novel neural architecture -- hierarchical Transformer -- to learn approximate execution semantics from micro-traces. \item We implement \textsc{Trex}\xspace and evaluate it on 1,472,066 functions from 13 popular software projects and libraries. \textsc{Trex}\xspace outperforms the state-of-the-art tools by 7.8\%, 7\%, and 14.3\%, in cross-architecture, optimization, and obfuscation function matching, respectively, while running up to 8$\times$ faster. Moreover, \textsc{Trex}\xspace helps uncover 16 vulnerabilities in 180 real-world firmware images with the latest version that are not disclosed by previous studies. We release the code and dataset of \textsc{Trex}\xspace at \url{https://github.com/CUMLSec/trex}. \end{itemize} \section{Methodology} \label{sec:method} This section describes \textsc{Trex}\xspace's design specifics, including our micro-tracing semantics, our learning architecture's details, and pretraining and finetuning workflow. \subsection{Micro-tracing Semantics} \label{subsec:microx-semantics} We implement micro-execution by Godefroid~\cite{godefroid2014micro} to handle x64, ARM, and MIPS, where the original paper only describes x86 as the use case. In the following, we briefly explain how we micro-execute an individual function binary, highlighting the key algorithms in handling different types of instructions. \vspace{.1cm}\noindent\textbf{IR Language.} To abstract away the complexity of different architectures' assembly syntax, we introduce a low-level intermediate representation (IR) that models function assembly code. We only include a subset of the language specifics to illustrate the implementation algorithm. Figure~\ref{fig:microx-syntax} shows the grammar of the IR. Note that the IR here only serves to facilitate the discussion of our micro-tracing implementation. In our implementation, we use real assembly instructions and tokenize them as model’s input (Section~\ref{subsec:input_repr}). Notably, we denote memory reads and writes by \texttt{load($e$)} and \texttt{store($e_{v},e_{a}$)} (\emph{i.e., } store the value expression $e_v$ to address expression $e_a$), which generalize from both the load-store architecture (\emph{i.e., } ARM, MIPS) and register-memory architecture (\emph{i.e., } x86). Both operations can take as input $e$ -- an expression that can be an explicit hexadecimal number (denoting the address or a constant), a register, or a result of an operation on two registers. We use \texttt{jmp} to denote the general jump instruction, which can be both direct or indirect jump (\emph{i.e., } the expression $e_a$ can be a constant $c$ or a register $r$). The jump instruction can also be unconditional or conditional. Therefore, the first parameter in \texttt{jmp} is the conditional expression $e_c$ and unconditional jump will set $e_c$ to \texttt{true}. We represent function invocations and returns by \texttt{call} and \texttt{ret}, where \texttt{call} is parameterized by an expression, which can be an address (direct call) or a register (indirect call). \begin{figure}[!t] \centering \includegraphics[width=\linewidth]{./figs/microx-semantics/language.pdf} \caption{Low-level IR for representing assembly code. The IR abstracts away the actual assembly syntax that are disparate across different architectures.} \label{fig:microx-syntax} \end{figure} \vspace{.1cm}\noindent\textbf{Micro-tracing algorithm.} Algorithm~\ref{alg:microx} outlines the basic steps of micro-tracing a given function $f$. First, it initializes the memory to load the code and the corresponding stack. It then initializes all registers except the special-purpose register, such as the stack pointer or the program counter. Then it starts linearly executing instructions of $f$. We map the memory address \emph{on-demand} if the instruction access the memory (\emph{i.e., } read/write). If the instruction reads from memory, we further initialize a random value in the specific memory addresses. For call/jump instructions, we first examine the target address and skip the invalid jump/call, known as ``forced execution''~\cite{peng2014x}. By skipping unreachable jumps and calls, it can keep executing the function till the end of the function and exposes more behaviors, \emph{e.g., } skipping potential input check exceptions. Since the \texttt{nop} instructions can serve as padding between instructions within a function, we simply skip \texttt{nop}. We terminate the micro-tracing when it finishes executing all instructions, reaches \texttt{ret}, or times out. Figure~\ref{fig:arithmetic} and~\ref{fig:stack} demonstrate sample micro-traces of real-world functions. \subsection{Input Representation} \label{subsec:input_repr} Formally, given a function $f$ (\emph{i.e., } assembly code) and its micro-trace $t$ (by micro-executing $f$), we prepare the model input $x$, consisting of 5 types of token sequence with the same size $n$. Figure~\ref{fig:arch} shows the model input example and how they are masked and processed by the hierarchical Transformer to predict the corresponding output as a pretraining task. \input{microx} \vspace{.1cm}\noindent\textbf{Micro-trace code sequence.} The first sequence $x_f$ is the assembly code sequence: $x_f=\{mov, eax, +, ...\}^n$, generated by tokenizing the assembly instructions in the micro-trace. We treat all symbols appear in the assembly instructions as tokens. Such a tokenization aims to preserve the critical hint of the syntax and semantics of the assembly instructions. For example, we consider even punctuation to be one of the tokens, \emph{e.g., } ``,'', ``['', ``]'', as ``,'' implies the token before and after it as destination and source of \texttt{mov} (in Intel syntax), respectively, and ``['' and ``]'' denote taking the address of the operands reside in between them. We take special treatment of numerical values appear in the assembly code. Treating numerical values as regular text tokens can incur prohibitively large vocabulary size, \emph{e.g., } $2^{32}$ number of possibilities on 32-bit architectures. To avoid this problem, we move all numeric values to the micro-trace value sequence (that will be learned by an additional neural network as detailed in the following) and replace them with a special token \texttt{num} (\emph{e.g., } last token of input in Figure~\ref{fig:arch}). With all these preprocessing steps, the vocabulary size of $x_f$ across all architectures is 3,300. \vspace{.1cm}\noindent\textbf{Micro-trace value sequence.} The second sequence $x_t$ is the micro-trace value sequence, where each token in $x_t$ is the dynamic value from micro-tracing the corresponding code. As discussed in Section~\ref{sec:overview}, we keep \emph{explicit} values (instead of a dummy value used by existing approaches) in $x_t$. Notably, we use the dynamic value for each token (\emph{e.g., } register) in an instruction before it is executed. For example, in \texttt{mov eax,0x8; mov eax,0x3}, the dynamic value of the second \texttt{eax} is \texttt{0x8}, as we take the value of \texttt{eax} before \texttt{mov eax,0x3} is executed. For code token without dynamic value, \emph{e.g., } \texttt{mov}, we use dummy values (see below). \vspace{.1cm}\noindent\textbf{Position sequences.} The position of each code and value token is critical for inferring binary semantics. Unlike natural language, where swapping two words can roughly preserve the same semantic meaning, swapping two operands can significantly change the instructions. To encode the inductive bias of position into our model, we introduce \emph{instruction position sequence} $x_c$ and \emph{opcode/operand position sequence} $x_o$ to represent the relative positions between the instructions and within each instruction. As shown in Figure~\ref{fig:arch}, $x_c$ is a sequence of integers encoding the position of each instruction. All opcodes/operands within a single instruction share the same value. $x_o$ is a sequence of integers encoding the position of each opcode and operands within a single instruction. \vspace{.1cm}\noindent\textbf{Architecture sequence.} Finally, we feed the model with an extra sequence $x_a$, describing the input binary's instruction set architecture. The vocabulary of $x_a$ consists of 4 architectures: $x_a=\{$x86, x64, ARM, MIPS$\}^n$. This setting helps the model to distinguish between the syntax of different architecture. \begin{figure}[!t] \centering \includegraphics[width=\linewidth]{./figs/arch-pretrain.pdf} \caption{Model architecture with input-output example. The masked input is marked in gray. In pretraining, the loss function consists of the cross-entropy losses of both (1) code prediction and (2) value prediction, where the value prediction consists of predicting each of the 8 bytes of the micro-trace value.} \label{fig:arch} \end{figure} \vspace{.1cm}\noindent\textbf{Encoding numeric values.} As mentioned above, treating concrete values as independent tokens can lead to prohibitively large vocabulary size. We design a \emph{hierarchical input encoding scheme} to address this challenge. Specifically, let $x_{t_i}$ denote the $i$-th value in $x_t$. We represent $x_{t_i}$ as an (padded) 8-byte fixed-length byte sequence $x_{t_i}=$\{\texttt{0x00}, ..., \texttt{0xff}\}$^8$ ordered in Big-Endian. We then feed $x_{t_i}$ to a 2-layer bidirectional LSTM (bi-LSTM) and take its last hidden cell's embedding as the value representation $t_i=bi$-$LSTM(x_{t_i})$. Here $t_i$ denotes the output of applying the embedding to $x_{t_i}$. To make the micro-trace code tokens without dynamic values (\emph{e.g., } opcode) align with the byte sequence, we use a dummy sequence (\texttt{\#\#}) with the same length. Figure~\ref{fig:arch} (right-hand side) illustrates how bi-LSTM takes the byte sequence and computes the embedding. Such a design significantly reduces the vocabulary size as now we have only 256 possible byte values to encode. Moreover, bi-LSTM encodes the potential dependencies between high and low bytes within a single value. This setting thus supports learning better relationships between different dynamic values (\emph{e.g., } memory region and offset) as opposed to treating all values as dummy tokens~\cite{shi2019learning}. \subsection{Pretraining with Micro-traces} \label{subsec:pretrain_method} This section describes the pretraining task on the function binaries and their micro-traces, focusing on how the model masks its input $x$ (5 sequences) and predicts the masked part to learn the approximate execution semantics. \vspace{.1cm}\noindent\textbf{Learned embeddings.} We embed each token in the 5 sequences with the same embedding dimension $d_{emb}$ so that they can be summed as a single sequence as input to the Transformer. Specifically, let $E_f(x_f)$, $E_t(x_t)$, $E_c(x_c)$, $E_o(x_o)$, $E_a(x_a)$ denote applying the embedding to the tokens in each sequence, respectively. We have $E_{i}$, the embedding of $x_i$: \begin{equation*} E_{i} = E_f(x_{f_i})+E_t(x_{t_i})+E_c(x_{c_i})+E_o(x_{o_i})+E_a(x_{a_i}) \label{eq:pos} \end{equation*} Here $x_{f_i}$ denote the $i$-th token in $x_f$, where other sequence (\emph{i.e., } $x_t$, $x_c$, $x_o$, $x_a$) follow the similar denotation. Note that $E_t(x_{t_i})$ leverages the bi-LSTM to encode the byte sequences (see Section~\ref{subsec:input_repr}), while the others simply multiplies the token (\emph{i.e., } one-hot encoded~\cite{harris2010digital}) with an embedding matrix. Figure~\ref{fig:arch} (right-hand side) illustrates the two embedding strategies. \vspace{.1cm}\noindent\textbf{Masked LM.} We pretrain the model using the masked LM objective. Formally, for each sequence $(x_1,x_2,...,x_n)$ in a given training set, we randomly mask out a selected percentage of tokens in each sequence. Specifically, we mask the code token and value token in $x_f$ and $x_t$, respectively, and replace them with a special token \texttt{<MASK>} (marked gray in Figure~\ref{fig:arch}). As masked LM trains on micro-traces without requiring additional labels, it is fully unsupervised. Let $m(E_i)$ denote the embedding of the masked $x_i$ and $\mathbb{MP}$ a set of positions on which the masks are applied. The model $g_p$ (to be pretrained) takes as input a sequence of embeddings with random tokens masked: $(E_1,..., m(E_i),...,E_n), i\in \mathbb{MP}$, and predicts the code and the values of the masked tokens: $\{\hat{x}_{f_i}, \hat{v}_i|i\in \mathbb{MP}\}=g_p(E_1,..., m(E_i),...,E_n)$. Let $g_p$ be parameterized by $\theta$, the objective of training $g_p$ is thus to search for $\theta$ that minimizes the cross-entropy losses between (1) the predicted masked code tokens and the actual code tokens, and (2) predicted masked values (8 bytes) and the actual values. For ease of exposition, we omit summation over all samples in the training set. \begin{equation} \label{eq:pretrain} \centering \argmin_{\theta} \sum\limits_{i=1}^{|\mathbb{MP}|} (-x_{f_i}\log(\hat{x}_{f_i})+ \alpha\sum\limits_{j=1}^{8}-x_{t_{ij}}\log(\hat{x}_{t_{ij}})) \end{equation} $\hat{x}_{t_{ij}}$ denotes the predicted $j$-th byte of $x_{t_i}$ (the $i$-th token in $x_t$). $\alpha$ is a hyperparameter that weighs the cross-entropy losses between predicting code tokens and predicting values. \vspace{.1cm}\noindent\textbf{Masking strategy.} For each chosen token to mask, we randomly choose from the masking window size from $\{1,3,5\}$, which determines how many consecutive neighboring tokens of the chosen tokens are also masked~\cite{joshi2020spanbert}. For example, if we select $x_5$ (the 5-th token) in Figure~\ref{fig:arch} to mask and the masking window size is 3, $x_4$ and $x_6$ will be masked too. We then adjust accordingly to ensure the overall masked tokens still account for the initially-chosen masking percentage. \vspace{.1cm}\noindent\textbf{Contextualized embeddings.} We employ the self-attention layers~\cite{vaswani2017attention} to endow contextual information to each embedding $E_i$ of the input token. Notably, let $E_{l}=(E_{l,1},...,E_{l,n})$ denote the embeddings produced by the $l$-th self-attention layer. We denote the embeddings before the model ($E_i$ as defined in Section~\ref{subsec:input_repr}) as $E_{0,i}$. Each token's embedding at each layer will attend to all other embeddings, aggregate them, and update its embedding in the next layer. The embeddings after each self-attention layer are known as \emph{contextualized embeddings}, which encodes the context-sensitive meaning of each token (\emph{e.g., } \texttt{eax} in \texttt{mov eax,ebx} has different embedding with that in \texttt{jmp eax}). This is in contrast with static embeddings, \emph{e.g., } word2vec commonly used in previous works~\cite{duandeepbindiff, ding2019asm2vec}, where a code token is assigned to a fixed embedding regardless of the changed context. The learned embeddings $E_{l,i}$ after the last self-attention layer encodes the approximate execution semantics of each instruction and the overall function. In pretraining, $E_{l,i}$ is used to predict the masked code. While in finetuning, it will be leveraged to match similar functions (Section~\ref{subsec:finetuning_method}). \subsection{Finetuning for Function Similarity} \label{subsec:finetuning_method} Given a function pair, we feed each function's \emph{static code} (instead of micro-trace as discussed in Section~\ref{sec:intro}) to the pretrained model $g_p$ and obtain the pair of embedding sequences produced by the last self-attention layer of $g_p$: $E^{(1)}_{k}=(E^{(1)}_{k,1},...,E^{(1)}_{k,n})$ and $E^{(2)}_{k}=(E^{(2)}_{k,1},...,E^{(2)}_{k,n})$ where $E^{(1)}_{k}$ corresponds to the first function and $E^{(2)}_{k}$ corresponds to the second. Let $y=\{-1,1\}$ be the ground-truth indicating the similarity (1 -- similar, -1 -- dissimilar) between two functions. We stack a 2-layer Multi-layer Perceptrons $g_t$, taking as input the average of embeddings for each function, and producing a function embedding: \begin{equation*} g_t(E_{k})=tanh((\sum\limits_{i=1}^n E_{k,i})/n)\cdot W_1)\cdot W_2 \end{equation*} Here $W_1\in \mathbb{R}^{d_{emb}\times d_{emb}}$ and $W_2\in \mathbb{R}^{d_{emb}\times d_{func}}$ transforms the average of last self-attention layers embeddings $E_k$ with dimension $d_{emb}$ into the function embedding with the dimension $d_{func}$. $d_{func}$ is often chosen smaller than $d_{emb}$ to support efficient large-scale function searching~\cite{xu2017neural}. Now let $g_t$ be parameterized by $\theta$, the finetuning objective is to minimize the cosine embedding loss ($l_{ce}$) between the ground-truth and the cosine distance between two function embeddings: \begin{equation*} \argmin_{\theta}\ l_{ce}(g_t(E^{(1)}_{k}),g_t(E^{(2)}_{k}), y) \end{equation*} where \begin{equation} \label{eq:cosembloss} l_{ce}(x_1,x_2,y) = \left\{\begin{matrix} 1-cos(x_1,x_2) & y=1\\ max(0,cos(x_1,x_2)-\xi) & y=-1 \end{matrix}\right. \end{equation} $\xi$ is the margin usually chosen between 0 and 0.5~\cite{paszke2019pytorch}. As both $g_p$ and $g_t$ are neural nets, optimizing Equation~\ref{eq:pretrain} and Equation~\ref{eq:cosembloss} can be guided by gradient descent via backpropagation. After finetuning $g_t$, the 2-layer multilayer perceptrons, and $g_p$, the pre-trained model, we compute the function embedding $f_{emb}=g_t(g_p(f))$ and the similarity between two functions is measured by the cosine similarity between two function embedding vectors: $cos(f^{(1)}_{emb},f^{(2)}_{emb})$. \section{Overview} \label{sec:overview} In this section, we use the real-world functions as motivating examples to describe the challenges of matching semantically similar functions. We then overview our approach, focusing on how our pretraining task (masked LM) addresses the challenges. \subsection{Challenging Cases} \begin{figure*}[!t] \centering \subfloat[Cross-architecture]{ \includegraphics[width=0.3\linewidth]{./figs/challenges/x86_arm.pdf} \label{subfig:x86_arm}} \subfloat[Cross-optimization]{ \includegraphics[width=0.33\linewidth]{./figs/challenges/O0_O3.pdf} \label{subfig:O0_O3}} \subfloat[Cross-obfuscation]{ \includegraphics[width=0.33\linewidth]{./figs/challenges/subobf.pdf} \label{subfig:subobf}} \caption{Challenging cases of matching semantically similar functions across different instruction architectures, optimizations, and obfuscations. (\textbf{Left}) the function \texttt{priv\_encode\_gost} is from \texttt{libcrypto.a} in \texttt{openssl-1.0.1f}. The upper function is compiled to x86 while the lower is compiled to ARM. (\textbf{Middle}) the function \texttt{<wd\_comparator>} is from \texttt{basenc} in \texttt{coreutils-8.32}. The upper and lower function is compiled by GCC-7.5 with \texttt{-O0} and \texttt{-O3}, respectively. (\textbf{Right}) the function \texttt{<CMS\_add0\_cert>} is from \texttt{libcrypto.a} in \texttt{openssl-1.0.1u}. The upper function is compiled using \texttt{clang} with default options. The lower function is compiled by turning on the instruction substitution using Hikari~\cite{hikari}, \emph{e.g., } \texttt{-mllvm -enable-subobf}.} \label{fig:challenge} \end{figure*} We use three semantically equivalent but syntactically different function pairs to demonstrate some challenges of learning from only static code. Figure~\ref{fig:challenge} shows the (partial) assembly code of each function. \vspace{.1cm}\noindent\textbf{Cross-architecture example.} Consider the functions in Figure~\ref{subfig:x86_arm}. Two functions have the same execution semantics as both functions take the lower 12-bit of a register and compare it to \texttt{0x80}. Detecting this similarity requires understanding the approximate execution semantics of \texttt{and} in x86 and \texttt{lsl}/\texttt{lsr} in ARM. Moreover, it also requires understanding how the values (\emph{i.e., } \texttt{0xfff} and \texttt{0x14}) in the code are manipulated. However, all existing ML-based approaches~\cite{massarelli2019safe} only learn on static code without observing each instruction's real execution effect. Furthermore, to mitigate the potentially prohibitive vocabulary size (\emph{i.e., } all possible memory addresses), existing approaches replace all register values and memory addresses with an abstract dummy symbol~\cite{massarelli2019safe,duandeepbindiff}. They thus cannot access the specific byte values to determine inherent similarity. \vspace{.1cm}\noindent\textbf{Cross-optimization example.} Now consider two functions in Figure~\ref{subfig:O0_O3}. They are semantically equivalent as \texttt{[ebp+8]} and \texttt{[esp+4]} access the same memory location, \emph{i.e., } the function's first argument pushed on the stack by the caller. To detect such similarity, the model should understand \texttt{push} decreases the stack pointer \texttt{esp} by 4. The model should also notice that \texttt{mov} at line 2 assigns the decremented \texttt{esp} to \texttt{ebp} such that \texttt{ebp+8} in the upper function equals \texttt{esp+4} in the lower function. However, such dynamic information is not reflected in the static code. \vspace{.1cm}\noindent\textbf{Cross-obfuscation example.} Figure~\ref{subfig:subobf} demonstrates a simple obfuscation by instruction substitution, which essentially replaces \texttt{eax+1} with \texttt{eax-(-1)}. Detecting the equivalence requires understanding approximately how arithmetic operations such as \texttt{xor}, \texttt{sub}, and \texttt{add}, executes. However, static information is not enough to expose such knowledge. \subsection{Pretraining Masked LM on Micro-traces} \label{subsec:overview_pretrain} This section describes how the pretraining task, masked LM, on functions' micro-traces encourages the model to learn execution semantics. Although it remains an open research question to explicitly prove certain knowledge is encoded by such language modeling task~\cite{saunshi2019theoretical}, we focus on describing the intuition behind the masked LM -- why predicting masked codes and values in micro-traces can help address the challenging cases in Figure~\ref{fig:challenge}. \vspace{.1cm}\noindent\textbf{Masked LM.} Recall the operation of masked LM: given a function's micro-trace (\emph{i.e., } values and instructions), we mask some random parts and train the model to predict the masked parts using those not masked. Note that pretraining with masked LM does not need any manual labeling effort, as it only predicts the masked part in the input micro-traces without any additional labeling effort. Therefore, \textsc{Trex}\xspace can be trained and further improved with a substantial number of functions found in the wild. The benefit of this is that a certain instruction not micro-executed in one function is highly likely to appear in at least one of the other functions’ micro-traces, supporting \textsc{Trex}\xspace to approximate diverse instructions' execution semantics. \vspace{.1cm}\noindent\textbf{Masking register.} Consider the functions in Figure~\ref{subfig:subobf}, where they essentially increment the value at stack location \texttt{[rbp-0x2c]} by 1. The upper function directly loads the value to \texttt{eax}, increments by 1, and stores the value in \texttt{eax} back to stack. The lower function, by contrast, takes a convoluted way by first letting \texttt{ecx} to hold the value -1, and decrements \texttt{eax} by \texttt{ecx}, and stores the value in \texttt{eax} back to stack. We mask the \texttt{eax} at line 3 in the upper function. We find that our pretrained model can correctly predict its name and dynamic value. This implies the model understands the semantics of \texttt{add} and can deduce the value of \texttt{eax} in line 3 after observing the value of \texttt{eax} in line 2 (before the addition takes the effect). We also find the model can recover the values of masked \texttt{ecx} in line 4 and \texttt{eax} in line 5, implying the model understands the execution effect of \texttt{xor} and \texttt{sub}. The understanding of such semantics can significantly improve the robustness in matching similar functions -- when finetuned to match similar functions, the model is more likely to learn to attribute the similarity to their \emph{similar execution effects}, instead of their syntactic similarity. \vspace{.1cm}\noindent\textbf{Masking opcode.} Besides masking the register and its value, we can also mask the opcode of an instruction. Predicting the opcode requires the model to understand the execution effect of each opcode. Consider Figure~\ref{subfig:O0_O3}, where we mask \texttt{mov} in line 2 of upper function. We find our pretrained model predicts \texttt{mov} with the largest probability (larger than the other potential candidates such as \texttt{add}, \texttt{inc}, etc.). To correctly predict the opcode, the model should have learned several key aspects of the function semantics. First, according to its context, \emph{i.e., } the value of \texttt{ebp} at line 3 and \texttt{esp} at line 2, it learns \texttt{mov} is most probable as it assigns the value of \texttt{esp} to \texttt{ebp}. Other opcodes are less likely as their execution effect conflicts with the observed resulting register values. This also implicitly implies the model learns the approximate execution semantics of \texttt{mov}. Second, the model also learns the common calling convention and basic syntax of x86 instructions, \emph{e.g., } only a subset of opcodes accept two operands (\texttt{ebp,esp}). It can thus exclude many syntactically impossible opcodes such as \texttt{push}, \texttt{jmp}, etc. The model can thus infer \texttt{ebp} (line 3 of upper function) equals to \texttt{esp}. The model may have also learned \texttt{push} decrements stack pointer \texttt{esp} by 4 bytes, from other masked samples. Therefore, when the pretrained model is finetuned to match the two functions, the model is more likely to learn that the semantic equivalence is due to that \texttt{[ebp+8]} in the upper function and \texttt{[esp+4]} in the lower function refer to the same address, instead of their similar syntax. \vspace{.1cm}\noindent\textbf{Other masking strategies.} Note that we are not constrained by the number or the type of items (\emph{i.e., } register, opcode, etc.) in the instructions to mask, \emph{i.e., } we can mask complete instructions or even a consecutive sequence of instructions, and we can mask dynamic values of random instructions’ input-output. Moreover, the masking operation dynamically selects random subsets of code blocks and program states at each training iteration and on different training samples. As a result, it enables the model to learn the diverse and composite effect of the instruction sequence, essential to detecting similarity between functions with various instructions. In this paper, we adopt a completely randomized strategy to choose what part of the micro-trace to mask with a fixed masking percentage (see Section~\ref{subsec:pretrain_method} for details). However, we envision a quite interesting future work to study a better (but still cheap) strategy to dynamically choose where and how much to mask. \section{.5pt}{5pt plus 2pt minus 2pt}{5pt plus 2pt minus 2pt} \begin{document} \title{\textsc{Trex}\xspace: Learning Execution Semantics from\\ Micro-Traces for Binary Similarity} \makeatletter \newcommand{\linebreakand}{% \end{@IEEEauthorhalign} \hfill\mbox{}\par \mbox{}\hfill\begin{@IEEEauthorhalign} } \makeatother \author{\IEEEauthorblockN{Kexin Pei} \IEEEauthorblockA{Columbia University\\ kpei@cs.columbia.edu} \and \IEEEauthorblockN{Zhou Xuan} \IEEEauthorblockA{University of California, Riverside\\ zxuan004@ucr.edu} \linebreakand \IEEEauthorblockN{Junfeng Yang} \IEEEauthorblockA{Columbia University\\ junfeng@cs.columbia.edu} \and \IEEEauthorblockN{Suman Jana} \IEEEauthorblockA{Columbia University\\ suman@cs.columbia.edu} \and \IEEEauthorblockN{Baishakhi Ray} \IEEEauthorblockA{Columbia University\\ rayb@cs.columbia.edu}} \date{} \maketitle \input{abst} \input{intro} \input{overview} \input{threat} \input{methodology} \input{impl} \input{eval} \input{case} \input{related} \input{conclusion} \input{ack} \bibliographystyle{plain} \section{Related Work} \label{sec:related} \subsection{Binary Similarity} \vspace{.1cm}\noindent\textbf{Traditional approaches.} Existing static approaches often extract hand-crafted features by domain experts to match similar functions. The features often encode the functions' syntactic characteristics. For example, BinDiff~\cite{bindiff} extracts the number of basic blocks and the number of function calls to determine the similarity. Other works~\cite{myles2005k, khoo2013rendezvous, crussell2013andarwin, farhadi2014binclone} introduce more carefully-selected static features such as n-gram of instruction sequences. Another popular approach is to compute the structural distance between functions to determine the similarity~\cite{bindiff, bourquin2013binslayer, david2014tracelet, pewny2014leveraging, eschweiler2016discovre, david2017similarity, huang2017binsequence}. For example, BinDiff~\cite{bindiff} performs graph matching between functions' call graphs. TEDEM~\cite{pewny2014leveraging} matches the basic block expression trees. BinSequence~\cite{huang2017binsequence} and Tracelet~\cite{david2014tracelet} uses the edit distance between functions' instruction sequence. As discussed in Section~\ref{sec:intro}, both static features and structures are susceptible to obfuscations and optimizations and incur high overhead. \textsc{Trex}\xspace automates learning approximate execution semantics without any manual effort, and the execution semantics is more robust to match semantically similar functions. In addition to the static approaches, dynamic approaches such as iLine~\cite{jang2013towards}, BinHunt~\cite{ming2012ibinhunt}, iBinHunt~\cite{ming2012ibinhunt}, Revolver~\cite{kapravelos2013revolver}, Blex~\cite{egele2014blanket}, Rieck \textit{et al.~}\cite{rieck2011automatic}, Multi-MH~\cite{pewny2015cross}, BinGo~\cite{chandramohan2016bingo}, ESH~\cite{david2016statistical}, BinSim~\cite{ming2017binsim}, CACompare~\cite{hu2017binary}, Vulseeker-pro~\cite{gao2018vulseeker}, and Tinbergen~\cite{mckee2019software} construct hand-coded dynamic features, such as values written to stack/heap~\cite{egele2014blanket} or system calls~\cite{ming2017binsim} by executing the function to match similar functions. These approaches can detect semantically similar (but syntactically different) functions by observing their similar execution behavior. However, as discussed in Section~\ref{sec:intro}, these approaches can be expensive and can suffer from false positives due to the under-constrained dynamic execution traces~\cite{jiang2009automatic, ding2019asm2vec}. By contrast, we only use these traces to learn approximate execution semantics of individual instructions and transfer the learned knowledge to match similar functions without directly comparing their dynamic traces. Therefore, we are much more efficient and less susceptible to the imprecision introduced by these under-constrained dynamic traces. \vspace{.1cm}\noindent\textbf{Learning-based approaches.} Most recent learning-based works such as Genius~\cite{feng2016scalable}, Gemini~\cite{xu2017neural}, Asm2Vec~\cite{ding2019asm2vec}, SAFE~\cite{massarelli2019safe}, DeepBinDiff~\cite{duandeepbindiff} learn a function representation that is supposed to encode the function syntax and semantics in low dimensional vectors, known as function embeddings. The embeddings are constructed by learning a neural network that takes the functions' structures (control flow graph)~\cite{feng2016scalable,xu2017neural,duandeepbindiff} or instruction sequences~\cite{ding2019asm2vec, massarelli2019safe} as input and train the model to align the function embedding distances to the similarity scores. All existing approaches are based only on static code, which lacks the knowledge of function execution semantics. Moreover, the learning architectures adopted in these approaches require constructing expensive graph features (attributed CFG~\cite{feng2016scalable, xu2017neural}) or limited in modeling long-range dependencies (based on Word2Vec~\cite{ding2019asm2vec, duandeepbindiff}). By contrast, \textsc{Trex}\xspace learns approximate execution semantics to match functions. Its underlying architecture is amenable to learning long-range dependencies in sequences without heavyweight feature engineering or graph construction. \subsection{Learning Program Representations} There has been a growing interest in learning neural program representation as embeddings from ``Big Code''~\cite{allamanis2018survey}. The learned embedding of the code encodes the program's key properties (\emph{i.e., } semantics, syntax), which can be leveraged to perform many applications beyond function similarity, such as program repair~\cite{parihar2017automatic, wang2017semantics}, recovering symbol names and types~\cite{chua2017neural, patrick2020probabilistic}, code completion~\cite{raychev2014code}, decompilation~\cite{fu2019coda, katz2018using}, prefetching~\cite{shi2019learning}, and many others that we refer to Allamanis \textit{et al.~}\cite{allamanis2018survey} for a more thorough list. Among these works, the most closest work to us is XDA~\cite{pei2021xda}, which also leverages the transfer learning to learn general program representations for recovering function and instruction boundaries. However, XDA only learns from static code at the raw byte level, which lacks the understanding of execution semantics. The core technique proposed in this paper -- learning approximate execution semantics from micro-traces -- is by no means limited to only function similarity task but can be applied to any of the above tasks. Indeed, we plan to explore how the learned semantics in our model can transfer to other (binary) program analysis tasks in our future work. \section{Threat Model} \label{sec:threat} We assume no access to the debug symbols or source while comparing binaries. Indeed, there exist many approaches to reconstruct functions from stripped binaries~\cite{bao2014byteweight, shin2015recognizing, andriesse2017compiler, di2017rev, pei2021xda}. Moreover, we assume the binary can be readily disassembled, \emph{i.e., } it is not packed nor transformed by virtualization-based obfuscator~\cite{vmprotect, ugarte2015sok}. \vspace{.1cm}\noindent\textbf{Semantic similarity.} We consider two semantically similar functions as having the same input-output behavior (\emph{i.e., } given the same input, two functions produce the same output). Similar to previous works~\cite{xu2017neural, massarelli2019safe, ding2019asm2vec}, we treat functions compiled from the same source as similar, regardless of architectures, compilers, optimizations, and obfuscation transforms.

Background ========== The successful mating of ewe lambs within their first year of age reduces the effective cost of replacement ewes, increases the lifetime productivity of ewes in the flock, reduces the generation interval \[[@B1]\] and reduces the culling risks \[[@B2]\]. Moreover, an earlier first lambing brings about a higher productivity of the ewe throughout its life \[[@B3]\]. In some breeds such as the Rasa Aragonesa, even though females reach an adequate live weight at 12--13 months, the mean age at first lambing is 17--18 months \[[@B3]\], leading to extended unproductive periods. Therefore, it is challenging to find endocrine markers that could help to select the most precocious ewes, particularly if they could be used very early in life. Sheep, as other mammals, are born with a limited and highly variable number of ovarian germ cells, which decreases drastically with age \[[@B4]\]. At birth, the oocytes are mostly contained in resting primordial and early growing follicles that constitute the so-called follicular reserve. Genetic factors and maternal nutrition during gestation both seem to contribute to this inherently high variation in the follicular reserve in sheep \[[@B4]-[@B8]\]. At 12--14 weeks after birth, an important population of gonadotropin-sensitive follicles is already present in the ovary \[[@B9],[@B10]\]. Moreover, recent results suggest that the number of growing antral follicles present in the ovaries of young adult cattle may be related to their fertility \[[@B11]\]. Therefore, the follicular populations present in the ovaries of young ruminants seem to have a great impact on their adult reproductive life. Presently, anti-Müllerian hormone (AMH) is the best endocrine marker of the ovarian follicular reserve of growing follicles, and it is used as a predictor of the ovarian response to gonadotropins in the human \[[@B12],[@B13]\], mouse \[[@B14]\] and bovine species \[[@B15]\]. In the cow, AMH concentration in plasma is highly correlated with the number of antral follicles of 3 to 7 mm in diameter, which are the main targets of superovulatory treatments \[[@B15]\]. Also, AMH measurement can help to predict the aptitude of individuals to produce high or low numbers of embryos after superovulation, a repeatable and possibly heritable trait in the cow \[[@B16]\]. Recently, similar results were reported in mares, indicating that AMH could be a predictor of the number of follicles available in a follicle aspiration program \[[@B17]\]. Although a direct relationship between AMH and fertility has not been reported to date in domestic animals, some evidence supports this idea. On one hand, in young adult cattle, a high degree of correlation among AMH, antral follicle count (AFC) and the number of healthy follicles and oocytes in ovaries has been demonstrated; and on the other hand, lower numbers of antral growing follicles (AFC) have been related with suboptimal fertility \[[@B11]\]. Up to now, there are no data on plasma AMH concentration in sheep, though it would be of great interest for the future selection of animals for embryo production, as well as of young animals with a high fertility in the adult age. Therefore, the aims of the present study were, first, to determine if AMH can be detected in the plasma of prepubertal Rasa Aragonesa ewes and to evaluate its relationship with the ovarian response to a gonadotropin treatment, and second, to relate AMH with fertility at first and second mating opportunities. With this aim, the plasma AMH concentration of ewe lambs aged around 3.6 months was assessed as well as the ovarian response to an equine chorionic gonadotropin (eCG) challenge, and their fertility was established after mating at 10 and 14 months. Results ======= Plasma AMH concentration and its relationship with eCG-induced ovulation in prepubertal ewe lambs ------------------------------------------------------------------------------------------------- AMH was detected in the plasma of 71 out of 76 ewe lambs (94%). The concentrations were highly variable between animals, ranging from 0 (*n* = 5) to 590 pg/mL. The age and weight of the ewe lambs with nondetectable AMH concentrations ranged from 88 to 126 days and from 23.5 to 28 kg. AMH was detected in the youngest lamb of this study (68 days) as well as in the lightest lamb (19 kg), reaching 362.5 and 53.7 pg/mL, respectively. AMH concentration in plasma was not correlated with age (ρ = −0.19; NS) or live weight (ρ = −0.08; NS). Plasma AMH concentration was correlated with the occurrence of ovulation in response to eCG (ρ = 0.42; P \< 0.0001) but not with the number of ovulations when only considering ovulating ewes (ρ = 0.08; NS). Neither the occurrence of ovulation nor the number of ovulations of ovulating ewes were correlated with age (ρ = 0.13 and −0.27, respectively; both NS) or live weight (ρ = 0.04 and 0.02; both NS). As expected, age and live weight were highly correlated (ρ = 0.38; P \< 0.001). Non-ovulating ewe lambs had lower plasma AMH concentrations than ovulating ewe lambs with 1, 2 or ≥ 3 ovulations (all P \< 0.004; Table [1](#T1){ref-type="table"}). However, no significant differences in AMH were found between the different groups of ovulating ewe lambs. No significant differences in live weight or age were found between any of the groups. ###### **Plasma anti-Müllerian hormone (AMH) concentration (pg/mL), age (d) and live weight (kg) at the moment of ovarian stimulation with 600 IU eCG, for each ovulation group (0, 1, 2 and ≥3 ovulations**) **of prepubertal ewe lambs (means ± SEM)** **No. of ovulations** **No. of ewes** **Age** **Live weight** **AMH** ----------------------- ----------------- ------------- ----------------- ------------------ 0 19 108 ± 2 ^a^ 25.0 ± 0.4 ^a^ 43.0 ± 14.6 ^a^ 1 17 118 ± 4 ^a^ 25.5 ± 0.6 ^a^ 126.9 ± 34.0 ^b^ 2 21 104 ± 5 ^a^ 24.7 ± 0.4 ^a^ 99.8 ± 18.3 ^b^ ≥3 19 108 ± 4 ^a^ 24.7 ± 0.8 ^a^ 163.1 ± 41.0 ^b^ Within columns, means with different superscripts differ significantly: a,b: P \< 0.004. Relationship between plasma AMH concentration of prepubertal ewe lambs and their fertility at the first and second service periods ---------------------------------------------------------------------------------------------------------------------------------- Ultrasound pregnancy diagnosis was coincident with fertility results both in the first (28.9%, n = 22 pregnant ewes) and in the second mating (77.8%, n = 42 pregnant ewes), at 10 and 14 months, respectively, revealing the absence of abortions and of mistakes in the lambing records. The overall fertility after the two consecutive mating opportunities was 84.2%. The fertility of adult ewes at first, at second and after both consecutive service periods was correlated with their AMH concentration in the prepubertal phase (ρ = 0.34, P \< 0.01; ρ = 0.33, P \< 0.05; ρ = 0.36, P \< 0.01, respectively). At the first mating, neither age nor live weight was significantly different between pregnant and non-pregnant ewes (315 ± 5 vs. 311 ± 3 days and 33.4 ± 0.6 vs. 32.6 ± 0.5 kg, respectively; both NS). Prepubertal plasma AMH concentration was higher in ewes that became pregnant at the first mating than in those which failed to conceive at first but became pregnant at the second mating (P \< 0.05) or in non-pregnant ewes after two service periods (P \< 0.01). It was also higher in ewes which became pregnant at the second mating than in non-pregnant ewes (P \< 0.05; Figure [1](#F1){ref-type="fig"}). {#F1} The performance of AMH concentration at prepubertal age for predicting the fertility of adult ewes at first mating is shown in Figure [2](#F2){ref-type="fig"}. The receiver-operating characteristic (ROC) curve represents the relationship between the sensitivity of the diagnostic test based on AMH concentration at prepubertal age to predict the occurrence of pregnancy in adults (true-positive percentage) and 100 minus the specificity of the test to predict the failure of pregnancy (false-positive percentage) at first mating. {#F2} The predicting performance of plasma AMH concentration, assessed by the AUC (area under the ROC curve), was 0.698 ± 0.063, significantly different from 0.5 (the AUC value corresponding to a non-informative or "random test") (P \< 0.01). The AUC values estimated by the leave-one-out validation and 10-fold cross-validation methods were somewhat lower (0.660 ± 0.066 and 0.650 ± 0.013, respectively), although also different from 0.5 (P \< 0.02 and P \< 0.001, respectively). The cutoff point of plasma AMH concentration that could be used for a screening test of the future ability of ewes to get pregnant at first mating was found to be 97 pg/mL. This AMH concentration corresponded to a sensitivity value to predict the occurrence of pregnancy of 68.2% (15/22) and a specificity to predict the failure of pregnancy of 72.2% (39/54). The positive and negative predictive values (PPV and NPV) were 50.0% and 84.4%, respectively (Table [2](#T2){ref-type="table"}). The fertility at first mating of ewes with AMH concentrations equal or higher than 97 pg/mL before puberty was 34.8 percentage points higher than that of ewes with lower AMH concentrations (P \< 0.001). ###### Number and percentages (between brackets) of lambing and non-lambing ewes at first mating predicted through their plasma AMH concentration at prepubertal age using a cutoff value of 97 pg/mL **Selection criterion** **AMH group** **Lambing** **Non-lambing** **No. of ewes** ------------------------- --------------- --------------- ----------------- ----------------- AMH ≥ 97 pg/mL 15 (50.0) ^a^ 15 (50.0) 30   \< 97 pg/mL 7 (15.2) ^b^ 39 (84.8) 46 No selection \-\-- 22 (28.9) 54 (71.1) 76 Within columns, percentages with different superscripts differ: a,b: P \< 0.001. The fertility at second mating or after both consecutive service periods was also higher in ewes with AMH concentration equal or higher than 97 pg/mL before puberty (93.3 vs. 71.8%; P \< 0.03 and 96.7 vs. 76.1%; P \< 0.004, respectively). In the present work, the benefit of selecting replacement ewes by AMH (calculated as the difference between PPV, 50% and overall flock fertility at first mating, 28.9%), was 21.1 percentage points. As PPV (as well as NPV) depends on the prevalence of the character in the population \[[@B18]\] (herein, the overall flock fertility), PPV for assumed flock fertility values at first mating from 5 to 95% has been depicted in Figure [3](#F3){ref-type="fig"}. It is shown that the benefit obtained selecting replacement ewes by prepubertal AMH would be higher for intermediate than for either high or low flock fertility values. {#F3} Discussion ========== This study shows for the first time that: (1) AMH is present in the plasma of prepubertal ewe lambs; (2) plasma AMH concentration is related to the occurrence of ovulation in ewe lambs after administration of an ovarian stimulation treatment, possibly reflecting that a population of gonadotropin-responsive follicles is already present at this age in most ewes; and (3) plasma AMH concentration before puberty could be used as a predictor of the fertility of adult ewes at first mating. This test would allow farmers to perform a precocious selection of replacement ewe lambs with the highest expected fertility at first mating. In the conditions of the present study, plasma AMH concentrations did not depend on age and weight, being detectable in most ewe lambs. Nevertheless, longitudinal studies would be necessary to accurately determine the age and weight at which AMH becomes detectable as well as the precise relationships of AMH with age and live weight. In ewe lambs at around 3.6 months of age, plasma AMH concentration was about 3--4-fold higher in ovulating than in non-ovulating ewe lambs after eCG treatment. This difference could be due to the presence of a different numbers of antral gonadotropin-responsive follicles between animals at this age. In this way, the mean age of the ewe lambs of this study (15.6 weeks) would be coincident with a period where a peak in the total number of follicles ≥ 3 mm has been observed in other breeds of sheep, varying between 14 weeks \[[@B13]\] and 16 weeks \[[@B19]\]. Therefore, the relationship found between the ovulatory response to eCG and plasma AMH concentrations could be explained by the fact that they likely both reflect the presence of antral gonadotropin-responsive follicles in the ovaries of ewe lambs. In adult cows, a high correlation of plasma AMH concentrations with the number of corpora lutea after superovulation was found \[[@B15]\]. However, in the present study a relationship between AMH and the number of ovulations in ovulating ewe lambs was not found. This discrepancy could be attributed to physiological differences between prepubertal and adult females and/or to differences in the ovarian stimulation treatments. In this way, the ovarian response to a single eCG treatment without previous progestogen priming might be different to the response to a superovulation protocol consisting in multiple FSH decreasing doses with previous progestogen synchronization. In the present work, neither plasma AMH concentration before puberty nor fertility at first mating was related with age or live weight. Moreover, all ewes received the same feeding plane during the whole experiment. Therefore, the observed differences in plasma AMH concentrations amongst ewes of similar age and similar live weight may be due to inherent variations in the follicular population, possibly influencing fertility at first mating. Genetic factors or maternal nutrition during pregnancy could therefore be responsible for this asynchronous follicular development; as in sheep, alterations in the maternal diet have been demonstrated to affect fetal ovarian development \[[@B5],[@B7]\]. Nevertheless, further studies are needed to elucidate the causes of this asynchronous development of the ovary between the same aged and weighted ewes, as well as to determine whether it could affect ovarian function and reproductive performance during the adult life. Apart from age, live weight and nutritional status, it is widely known that successful pregnancy is affected by many other factors, such as season, endocrine status or uterine conditions \[[@B20]\]. In the present work, healthy ewe lambs were reared under the same conditions to minimize the impact of such factors, although possible unknown individual disorders can not be discarded. The results of the present study showed that ewes with higher prepubertal plasma AMH concentrations displayed a higher probability of becoming pregnant at first mating than those with lower plasma AMH concentrations. Moreover, those ewes which failed to conceive after two consecutive mating opportunities had the lowest prepubertal plasma AMH concentrations. As far as we know, this is the first time that a relationship between AMH and early fertility has been established in a domestic species. Our study reinforces the prognostic value of AMH, even when the physiological status is not fully known. Both in the human and bovine, AMH levels are almost independent of the phase of the ovarian cycle, explaining why a single AMH measurement is usually sufficient \[[@B21],[@B22]\]. This could be due to the fact that AMH is not involved in feed-back mechanisms of the hypothalamo-pituitary-gonadal axis \[[@B21]\]. Previous studies in the human have demonstrated a relationship between AMH and fertility. Very low concentrations of AMH are associated with infertility in the case of pre-menopause and premature ovarian failure, and very high AMH concentrations with an absence of ovulation in women with polycystic ovarian syndrome. Low or undetectable levels of AMH is explained by decrease or absence of follicular growth due to depletion of the ovarian reserve of primordial follicles. Abnormally high levels of AMH are explained by an excess of antral follicles whose growth is blocked prior to ovulation \[[@B21]\]. Recent results have also evidenced that AMH seems to be an excellent marker of ovarian function in girls and adolescents. In this way, elevated serum AMH levels have been found in girls with a precocious maturation of the adrenal cortex, suggesting that follicular development has proceeded to a more advanced developmental stage than expected for chronological age \[[@B23]\]. It has been demonstrated in adolescent patients with inherited reproductive problems (Turner syndrome) that their AMH concentrations correlate significantly with ovarian function at time of AMH measurement, showing an excellent sensitivity and specificity as a screening test of premature ovarian failure. Furthermore, it has been suggested that variations in AMH concentrations during childhood may theoretically predict the duration of any given girl's reproductive lifespan \[[@B24]\], as AMH seems to reflect the continuous decline of the follicle pool with age \[[@B25]\]. In the case of ewe lambs, the observed AMH values would be rather a reflection of the beginning of follicular growth in prepubertal animals, showing an ovarian activity quantitatively different between animals of the same age. It is possible that ewe lambs in our study with the highest ovarian activity at 3.6 months were also the most precocious in terms of early puberty, resulting in the improved fertility observed when testing fertility of different ewes at the same age. The lambs with the earliest follicular development would be fertile at an early age. All these findings highlight the importance of individual differences in chronological follicular development and in the follicular reserve as being responsible for future reproductive life. In sheep, recent in vitro studies have demonstrated that prepubertal oocytes isolated from ovaries with a high number of follicles larger than 2 mm are more competent for in vitro development, presenting a higher cleavage rate, developmental kinetics and cell number at the blastocyst stage \[[@B26]\]. As it has been established that the granulosa cells of large preantral and small antral healthy follicles produce the highest AMH amounts in rat \[[@B27]\] and cow \[[@B15],[@B22]\], prepubertal ewe lambs with higher plasma AMH concentrations are supposed to have a higher number of healthy follicles in these stages, suggesting that the quality of the oocytes and embryos coming from these ewes may be higher. Although further studies are needed to confirm this hypothesis, it could partially explain the increased fertility at first mating of ewe lambs with the higher concentrations in plasma AMH during their prepubertal phase. Concerning the ability of AMH to be used as a screening test to select the most precocious ewes in terms of fertility at first mating, in the present study the AUC value for AMH (0.698) was very close to 0.7, the limit value to classify tests as \"moderately accurate\" according to an arbitrary guideline. This guideline distinguishes between non-informative (AUC = 0.5), less accurate (0.5 \< AUC \< 0.7), moderately accurate (0.7 \< AUC \< 0.9), highly accurate (0.9 \< AUC \<1) and perfect tests (AUC = 1) \[[@B28]\]. The values obtained by cross-validation methods were somewhat lower and would correspond to a "less accurate" test. This is not strange as it its well known that the establishment and maintenance of pregnancy depends upon several genetic and environmental factors. Previous reports in the human have shown that while high AMH levels prior to the initiation of IVF treatments is a good predictor of the ovarian response, it has not proven to be as predictive of successful pregnancy. This indicates that factors other than the quantitative aspects of ovarian reserve may influence the fate of pregnancy \[[@B29]\]. Nevertheless, although not being highly accurate, we have shown that fertility of replacement ewes at first mating can be increased up to about 20 percentage points by using plasma AMH concentration as a screening test. It remains to be assessed at which age and weight the diagnostic test based on the measurement of AMH concentrations on ewe lambs is the most efficient for accurate prediction. Concerning fertility at the second mating and fertility after two consecutive service periods, we have shown that the differences between ewes with AMH concentration above and below the cutoff point are also significant and of great practical importance (+21.5 and +20.6 percentage points, respectively). This would indicate that AMH could also be a predictor of fertility in adulthood, but this needs further confirmation. Conclusions =========== In conclusion, a single AMH measurement in the plasma of prepubertal ewe lambs may be useful to select, in a very early age, the replacement ewes with a high predicted fertility at first mating. Methods ======= All experimental procedures were performed in accordance with the guidelines of the European Union (2003/65/CE) and Spanish regulations (RD 1201/2005, BOE 252/34367-91) for the use and care of animals in research. Experimental design ------------------- Prepubertal ewe lambs (aged around 3.6 months) were sampled for plasma AMH determination. In order to test the presence of gonadotropin-sensitive follicles at this age, an injection of eCG was applied, and 6 days after, the induced ovulation (number of ovulations) was recorded by laparoscopy. The relationship between AMH concentration and the ovarian response to eCG was investigated. Seven months later, the same young ewes were first joined to rams twice, 4 months apart. The relationship between fertility at these mating opportunities and plasma AMH concentration before puberty was studied. A ROC curve for AMH was analyzed to find the best cutoff value to predict fertility at first mating. Plasma sampling and determination of eCG-induced ovulation in prepubertal ewe lambs ----------------------------------------------------------------------------------- The experiment started in July (beginning of the breeding season) in the facilities of the research center (CITA). A total of 76 Rasa Aragonesa prepubertal ewe lambs aged 109 ± 18 days (mean ± SD) and weighing of 25.0 ± 2.4 kg (mean ± SD) were used. Lamb ewes were weaned at 50 days postpartum and fed at libitum with concentrate until the beginning of the experiment. Ewe lambs received eCG (600 IU im; Sincropart PMSG, CEVA Salud Animal S.A., Barcelona, Spain), and simultaneously, blood samples were collected by jugular puncture using 5 mL vacuum tubes with lithium heparin that were immediately centrifuged at 2100 × g for 25 min. Plasma was stored at −20°C until analyzed for AMH. Six days after eCG treatment, ovulation was recorded by laparoscopy as follows. Ewe lambs were sedated with propofol (6 mL iv; Propofol Lipuro 1%, Braun, Spain) and local anesthesia with lidocaine was applied. An endoscope was inserted into the abdominal cavity through a 1-cm incision, approximately 10 cm cranial to the udder, and an atraumatic grasping forceps was introduced through a contralateral 1-cm incision. After total visualization of both ovaries, the number of healthy corpora lutea was recorded. The incisions were closed with staples and sprayed with topical chlortetracycline hydrochloride, and 1 mL/10 kg oxytetracyclinedihydrate (200 mg/mL; Oxycen-200 L.A., s.p.® veterinaria, s.a., Tarragona, Spain) was administered intramuscularly. After that, animals were fed at libitum under grazing conditions. Evaluation of the fertility of the ewes at first and second mating ------------------------------------------------------------------ In this study, fertility was defined as the ability of a ewe to become pregnant in a given mating period. In January (end of the breeding season), when ewes were 312 ± 18 days old (mean ± SD) and weighing 33.1 ± 3.1 kg (mean ± SD), they were first joined to four adult Rasa Aragonesa rams of proven fertility. The minimum ram-ewe lamb ratio was around 5% (first mating: 4/76; second mating: 4/54). This ratio was similar or slightly greater than those proposed by other authors \[[@B30],[@B31]\]. Males were removed after 34 days. Fertility at first mating was calculated as the percentage of ewes lambing (and thus successfully mated) over ewes exposed to rams (n = 76) in this first service period. Four months later (in May), the non-pregnant ewes were mated again for 34 days, according to an accelerated lambing program of three lambings in 2 years. Fertility at second mating was calculated as the percentage of ewes lambing (and thus successfully mated) over ewes exposed to rams (not considering pregnant ewes at first mating) in this second service period. Overall fertility after the two consecutive mating periods was calculated as the percentage of ewes lambing after either first or second mating periods, over all the ewes exposed to rams (n = 76). Ten and 30 days after the rams' removal, ultrasound diagnosis of pregnancy was performed. Lambing took place in individual pens and offspring were immediately identified. AMH assay --------- Plasma concentrations of AMH were measured with the Active MIS/AMH ELISA Kit (Beckman Coulter France, Roissy CDG, France), as described previously \[[@B15],[@B32]\]. Just before the assay, the frozen plasma samples were thawed in a warm water bath, vortexed and centrifuged (3200 g, 10 min, 4°C) to remove any cell fragments that could interfere with the reagents of the assay. AMH concentrations were determined in 50-μL undiluted plasma samples. The samples were incubated overnight at 4°C in the presence of the primary antibody, then for 1.5 h at room temperature in the presence of the secondary antibody. Dilutions of the points of the standard curve were made in steer plasma. With these conditions, the limit of detection of the assay was found to be 15 pg/mL. All the plasma samples were analyzed in the same assay. The intra-assay coefficients of variation were all lower than 5 % for the four quality control plasma samples that were tested, containing 27, 112, 195 and 274 pg/mL of AMH. For further validation of the assay, serial dilutions of different ovine plasma and follicular fluid samples in steer plasma were analyzed. Results showed that dilution curves were linear and parallel to the standard curve (Figure [4](#F4){ref-type="fig"}). {#F4} Evaluation of plasma AMH concentration in prepubertal age as a predictor of fertility of the ewe at first and second mating --------------------------------------------------------------------------------------------------------------------------- An ROC analysis was performed to evaluate the predicting ability of plasma AMH concentration to detect the ewes with a high fertility at first mating. The sensitivity was defined as the ratio between the ewes correctly predicted as pregnant above an AMH value and all the pregnant ewes. The specificity was the ratio between the ewes correctly predicted as non-pregnant below an AMH value and all the non-pregnant ewes. The PPV was the ratio between the ewes correctly classified as pregnant and those predicted as pregnant (in other words, the fertility of the selected ewes above an AMH value). The NPV was the ratio of ewes correctly classified as non-pregnant and those predicted as non-pregnant (100 minus fertility of non-selected ewes, below an AMH value). Statistical analysis -------------------- Correlations among the variables: occurrence of ovulation (1 = ovulating; 0 = non-ovulating), number of ovulations in ovulating ewes, plasma AMH concentration, age, live weight and fertility (1 = pregnant; 0 = non-pregnant) were assessed by the Spearman correlation coefficient (ρ). Continuous variables (age, live weight, plasma AMH concentration) were tested by general linear models (ANOVA). Normality of residuals was tested by the Jarque-Bera test, applying a correction for finite samples \[[@B33]\]. Homogeneity of within-groups variances (homoscedasticity) was tested by the Brown-Forsythe test. As plasma AMH concentration failed to accomplish the requirements of residuals normality and homoscedasticity, its values (expressed in pg/mL) were transformed to decimal logarithms (log~10~\[AMH + 1\]) prior to ANOVA tests. Pairwise comparisons between means were carried out by the least significant difference test, applying the "false discovery rate" approach \[[@B34]\] to adjust probabilities for multiple comparisons. For plasma AMH concentrations, although significance shown in tables was obtained through the transformed data, arithmetic means and SEM of the original, non-transformed data are presented. Dichotomous variables, presented as percentages in text and tables, were analyzed by chi-square tests. When the number of ovulations was used as independent variable in ANOVA, ewes with numbers of ovulations higher than two (from 3 to 7) were grouped into one category owing to the small number of ewes in these categories. In order to evaluate plasma AMH concentration as a predictor test of the fertility at first mating, the ROC curve was analyzed. The AUC was used as an indicator of the global performance of the test. The values of AUC were compared with 0.5 (AUC for a "random test"), using the non-parametric approach for testing statistical significance \[[@B28]\]. As AUC values computed from the data set used for model fitting tend to indicate better accuracy than the actual model allows in practice, AUC values estimated by the leave-one-out validation method and by 10-fold cross-validation method are also presented \[[@B35]\]. The 10-fold cross-validation was repeated 10 times and the average AUC of these 10 replicates ± SD was calculated. The cutoff value of AMH to discern between fertile and non-fertile ewes at first mating was chosen as the corresponding point of the ROC curve that maximized the Youden index, a prevalence-independent criterion \[[@B28]\]. All statistical analyses were performed using SAS software \[[@B36]\]. Otherwise specified, all data presented in text are means ± standard errors. Authors' contributions ====================== BL, JLA, PM and JF designed research. BL and JF administered treatments, performed endoscopy and collected samples and data. DM and PM analyzed samples. JLA carried out statistical analyses. BL, JLA, DM and JF have written the manuscript. All authors have contributed to the discussion. All authors read and approved the final manuscript. Acknowledgments =============== This work was supported by INIA (Ministerio de Ciencia e Innovación, Spain; RTA-2011-128 and RZP-2010-002 projects and fellowship for B. Lahoz). The authors would also like to thank Pilar Sánchez and Elías Echegoyen for their technical support.

{.67} {.68} {.69} {.70}

\section{INTRODUCTION} In recent years, due to the increasing need for the robotic system to interact with the human and unconstructed environment, the soft robots which are compliant and flexible attract scientists’ attention\cite{laschi2016soft}. Therefore, soft actuators are widely used in different robotic systems. To ensure safety during the interaction with humans, soft manipulators\cite{jiang2021hierarchical,sozer2020pressure,chen2021soft}, soft rehabilitation exoskeletons\cite{o2020inflatable,polygerinos2015soft}, and soft prosthetic hands\cite{gu2021soft} were proposed to replace the rigid ones. Meanwhile, in the field where the robots need to deal with the complexity of contacting subjects like grasping\cite{brown2010universal,li2021dual,liu2021soft,abondance2020dexterous} or climbing\cite{zhang2021inchworm,rafsanjani2018kirigami}, the soft actuators offer an adaptive solution without a complicated control algorithm. \begin{figure}[thpb] \centering \includegraphics[scale=0.95]{FIG/Fig1.eps} \caption{The prototype of MSTA.} \label{fig1} \end{figure} However, due to the compliance of soft material, the external load of the soft actuator which may vary in different situations may cause undesired destabilization. For this reason, the tuneable stiffness actuator is an effective solution due to its advantages for both rigid robots and soft robots. In this framework, the active and semiactive stiffness regulation methods were proposed in the past research\cite{manti2016stiffening}. The active method including fluidic actuators\cite{robertson2017soft}, shape memory materials\cite{she2016modeling}, and tendon-driven actuators\cite{manti2015bioinspired} aims at stiffening through the antagonistic manner. The semiactive modulation represents that the stiffness is tuned by changing the property of intrinsic rigidity, such as material jamming\cite{brown2010universal}, magnetorheological(MR)\cite{majidi2010tunable}, and low melting point materials\cite{shintake2015variable}. Although the shape memory materials and low melting point material are widely used to control the stiffness, the practicability is limited by their long response time. Therefore, the jamming-based mechanisms offer a solution to achieve a considerable stiffness change with a short response time. For example, Jiang et al.\cite{jiang2019chain} designed a mechanism based on chain-like granular jamming to enhance the stiffness greatly with a short response time. However, the range of tunable stiffness is small, which means that it has poor flexibility and compliance. To address this challenge, the gas-ribbon-hybrid driven actuator was proposed by Zhang et al.\cite{zhang2022gas}. The stiffness of the actuator is modulated in an antagonistic way. However, because of the low original stiffness of the actuator, the destabilization in the non-bending direction especially in the direction perpendicular to bending may occur during application. To be more specific, for example, the grasping may fail when the gripper changes the orientation from the vertical plane to the horizontal plane. \begin{figure*}[thpb] \centering \includegraphics[scale=0.9]{FIG/Fig2.eps} \caption{The design details of the proposed MSTA. (a) The sectional view of the main body and the details of the soft-rigid connection. (b) The overall structure (left) and back view (right) of MSTA. (c) The details of the bone-like structure.} \label{fig2} \end{figure*} Considering all the challenges, to maintain a larger range of the tunable stiffness and prevent the undesired destabilization, we proposed a new soft-rigid hybrid actuator with multi-direction tunable stiffness property (MTSA), as Fig. 1 displays. And the property is achieved through the combination of tendon-air driven solution and a novel bone-like structure (BLS). The main contribution of this paper is to propose the design of MTSA, and the fabrication process that can improve the airtightness. Then the working principle of stiffening of MTSA was analyzed from two orthogonal directions, the bending direction, and the direction perpendicular to bending. Finally, the performances including the fingertip force and the stiffness in different directions were tested. \section{STRUCTURE DESIGN AND FABRICATION} \subsection{Structure Design} The main structure of MTSA, as Fig. 2b shows, is composed of three major parts, the main body, bone-like structures (BLSs), and the rigid connector. The two BLSs are attached to each side of the main body, connecting to the rigid connector to form a soft-rigid hybrid structure. The MTSA is driven by a mixture of tendons and gas. Fig. 2a displays the sectional view of the main body and the details of the soft-rigid connection. The main body includes three components: a soft contact end, the soft chamber, and the actuation tendons. To achieve compliance, the Smooth-On 0020 silicone, with a 20-shore hardness is chosen to fabricate the soft chamber with a nylon mesh inserted to limit the elongation. According to Abondance et al.\cite{abondance2020dexterous}, the contact pad will help to increase the fingertip force. Therefore. the soft contact end, which is cylindrical is made of a harder silicone (Smooth-On 0030) and glued in the front of the soft chamber to contact with the environment and objects. The actuation tendons (Kevlar) cross through the holes in both sides of the actuator and tie to the rope fittings at the end of the actuator. These holes’ centers are in the same plane as the symmetrical plane of the rigid chain. The back of the soft chamber is cast into the connector. Due to the porous structure of the connector, the airtightness of the soft-rigid connection can be enhanced. The quick connector will be connected to the rigid connector through a pipe thread. \begin{figure}[thpb] \centering \includegraphics[scale=0.95]{FIG/Fig3.eps} \caption{The fabrication process of the proposed MSTA. (a) The first step is the fabrication of the soft chamber. (b) The second step, fit the BLSs. (c) The third step, assemble other components.} \label{fig3} \end{figure} The design details of bone-like structures are shown in Fig. 2c. There are three major parts in each BLS, rigid chain, pre-tension mechanism, and plate, connected through a fishline go across them. The fishline is fixed by two set screws in the distal segment and its path is shown in Fig. 2c in yellow. Three types of segments (PLA) constitute the rigid chain: distal segment, common segment, and proximal segment. The bulge and concavity of the common segment are semicylindrical, and the holes are perforated along the tangent of the cylinder. The segments are designed as a bone to match the size of the actuator. The pre-tension mechanism, whose components are 3D printed by nylon is to offer the pre-tightening force of BLS. The two pulleys (nylon) on each side of BLS provide a path to guide the fishline and form a closed-loop path. One of the pulleys is connected to the slider (PLA) placed in the chute of the backward plate (PLA). By adding specific sizes of PLA blocks to change the position of the slider, the pre-tightening force of the fishline can be regulated, meaning that the stiffness of BLS can be adjusted, either. The plate, pre-tension mechanism, and rigid connector are connected by nuts and screws. \subsection{Fabrication} In this work, the MTSA is fabricated based on combining the porous connector with the lost-wax method. The lost-wax method is utilized to enhance the airtightness so that the allowable input pressure range of the actuator can be extended, enlarging the tuneable stiffness range. The porous sidewall of the connector reinforced the airtightness in the soft-rigid-connection boundary. The fabrication process, which is illustrated in Fig. 3 includes three steps: fabrication of the soft chamber, fitting the BLSs, and assembling other components. All the molds for casting were 3D printed with ABS. Fabrication of the soft chamber. As Fig. 3a shows, the wax core was fabricated by pouring the melted wax into molds and demolding it after the wax was cured. To make the lost-wax process more accessible, the low melting point wax (about 70℃) was chosen. Then, the connector with the wax core fitted to it through the cylinder hole was placed in the mold. The fluid silicone was poured into the pre-closed mold. Before the silicone was cured, a nylon mesh was embedded into the bottom layer and two metal sticks were inserted through the holes in the mold. Then, the soft chamber with a wax core in it was obtained after it was cured. \begin{figure}[thpb] \centering \includegraphics[scale=0.95]{FIG/Fig4.eps} \caption{(a) Bending schematic of MSTA. (b) The image of the evaluation function. (c) The analysis model for the bending deformation state. (d) The analysis model for the combined deformation state.} \label{fig4} \end{figure} Fitting the BLSs. As Fig. 3b displays, the soft chamber with wax core had been heated in the water bath to 85℃ for 15 minutes. Squeezed out all the molten wax from the cylinder hole in the connector, then the soft chamber was assembled with two BLSs and the plate. Subsequently, two cover layers were cast and bonded with the soft chamber by uncured silicone. Assembling other components. In this process, as presented in Fig. 3c, the sealing layer of MTSA and the contact end were fabricated. The sealing layer can be omitted since it was not an essential structure to improve the airtightness. The soft-rigid connection in porous connector has played the main role in reinforcing the airtightness. The actuation tendons crossed through the reserved holes and were fixed in the rope fitting. After connecting the quick connector and gluing the contact end, the MTSA was fabricated. \section{WORKING PRINCIPLE} The working principle of the multi-direction stiffness property, as Fig. 4 shows will be discussed from two directions of MTSA. \subsection{Bending direction} In the bending direction (B direction), as Fig. 4a illustrates the curving deformation of MTSA is driven by pressuring the chambers while keeping the tendons in a relaxed state. The bending of the MTSA is considered as a constant curvature and the bending angle can be calculated as , where is the bending angle of a pair of chambers and denotes the number of chambers. In this state, the actuator will remain the high compliance. When the stiffness enhancement is required, the tightened tendons will limit the bending trend of MTSA while the air pressure is still rising. The increased gas pressure will improve the stiffness of the actuator in the bending direction. \subsection{Direction perpendicular to bending} In the direction perpendicular to bending (PB direction), as Fig. 4c and 4b display, there are two different states for stiffness modulation, the bending deformation state, and the combined deformation state, depending on the bending angles of the actuator. In general, for these two states, the main principle can be concluded that the enhanced stiffness is achieved through solid components jamming and can be tuned by changing the pulling force to rope. As shown in Fig. 4c, the BLS whose components are connected through a close-loop rope can be treated as a beam with revolute joints. The cross section of the beam is equivalent to the cross section of the bulge of common segments. The rope provides the tension to keep the BLS tightened. Except for the fracture of BLS, the BLS might break due to the granular separation, as Fig. 4c illustrates, which will lead to a sharp decrease in the stiffness. According to the research done by \cite{jiang2019chain}, the BLS will operate normally under the condition that follows the below equation: $$ F_{ext}\leqslant \frac{F_Tc}{L}\eqno{(1)} $$ where $F_t$ is the pulling force provided by rope and $F_{ext}$ denotes the external force applied to the end of BLS, $c$ represents the thickness of BLS, and $L$ is the length of the rigid chain. When the bending angle equals zero, that the BLS is in the bending deformation state, the BLSs assembled on each side of MTSA can be treated as a cantilever beam, as Fig. 4c shows. Therefore, the PB-direction stiffness can be enhanced. And the stiffness $k$ of a single BLS can be estimated as the following equation, according to Euler–Bernoulli beam theory: $$ k\propto \frac{3EI}{L^3}\eqno{(2)} $$ where $E$ is Young’s modulus, and $I$ refers to the area moment of inertia. The area moment of inertia $I$ of BLS can be described as $$ I=\frac{hb^3}{12}\eqno{(3)} $$ where $d$ is the width of the equivalent beam of BLS. According to equitation (2), The stiffness in the bending deformation state completely depends on the moment of inertia $I$, since Young’s modulus $E$ and length of BLS $l$ are constants. When the BLS is in the combined deformation state, the bending is performed, and the equivalent beam model turns into the curved cantilever beam as it is illustrated in Fig. 4c. To evaluate the factors influencing the stiffness, the curved cantilever beam can be simplified into two connecting rods as Fig. 4d shows. And Their angle of intersection is described as $\theta $. In this state, when the external force $F_{ext}$ is applied, the flexural-torsional deformation will occur. With small deformation, the deflection of the curved beam can be analyzed separately from the former rod and latter rod as follows: $$ \omega =\omega _{bend1}+\omega _{bend2}+\omega _{tor2}\eqno{(4)} $$ where $\omega $ denotes the total deflection at the distal end, $w_{bend1}$ is the deflection caused by the bending of the former rod, and $\omega _{bend2}$ with $\omega _{tor2}$ refers to the deflection of the flexural-torsional deformation of the latter rod. For the bending deformation of two rods, the deflection can be given as $$ \begin{cases} \omega _{bend1}=\frac{F_{ext}L_{1}^{3}}{3EI}\\ \omega _{bend2}=\frac{M_1L_{2}^{2}}{2EI}\\ M_1=F_{ext}\left( L_1\cos \theta +L_2 \right)\\ \end{cases}\eqno{(5)} $$ where $L_1$ is the length of the former rod, $L_2$ is the length of the rod, and $M_1$ refers to the bending moment applied to the latter rod. Then the deflection caused by the torsion of the latter rod is analyzed. To facilitate the evaluation, the torsional deformation is simplified to the geometrical model as shown in Fig. 4d, and the geometrical relationship can be described as $$ \omega _{tor2}=L_1\sin \theta \sin \varphi\eqno{(6)} $$ where the $\varphi $ is the torsional angle, which can be calculated as $$ \begin{cases} \varphi =\frac{TL_2}{G\beta hb^3}\\ T=F_{ext}L_1\sin \theta\\ \end{cases}\eqno{(7)} $$ where $G$ is the shear modulus, $\beta $ is a correction parameter for rectangular section torsion that is related to $h$ and $b$, and $T$ is the torque. Based on Equitation (4)-(7), the relationship between $\omega $ and $F_{ext}$ can be obtained: $$ \omega =\frac{F_{ext}}{hb^3}\left( L_1\sin \theta \sin \left( A\sin \theta \right) +B\cos \theta +C \right)\eqno{(8)} $$ $$ \begin{cases} A=\frac{L_1L_2}{G\beta}\\ B=\frac{6L_1L_{2}^{2}}{E}\\ C=\frac{4L_{1}^{3}+6L_{2}^{3}}{E}\\\end{cases} $$ To better analyze the stiffness, the parameter $\beta $ is regarded as a constant and is chosen as the value when ${{h}\//{b}}=1$. Based on the equitation (8), it can be deduced that the moment of inertia plays the same role in both the bending deformation state and combined deformation state, the higher moment of inertia $I$ brings the larger stiffness $k$. While in the combined deformation state, the stiffness will change with $\theta $. To estimate the influence, the parameters of PLA properties were chosen to draw the image of the following function as Fig. 4b: $$ f\left( \theta \right) =L_1\sin \theta \sin \left( A\sin \theta \right) +B\cos \theta +C.\eqno{(9)} $$ Because of $k\propto {{1}\//{f\left( \theta \right)}}$, the stiffness in PB direction will decrease firstly and then increase with the bending angle changing from $0\degree$ to $180\degree$. \section{EXPERIMENTS AND RESULTS} \begin{figure}[thpb] \centering \includegraphics[scale=1.4]{FIG/Fig5.eps} \caption{The testing results of the finger.} \label{fig5} \end{figure} To verify the design and characterize the ability of MTSA, the fingertip force and the tunable stiffness ability in the B direction and PB direction were tested. \subsection{Fingertip Force} To estimate the strength of MTSA, the fingertip force of MTSA was tested. The actuator connecting with air tubes was fixed on the aluminum frames and the distal end was aligned with the surface of the force gauge (HP-200, HANDPI Ltd., scale 0 - 200 N, accuracy $\pm 0.5\%$). From our testing, the maximum input pressure is 90 kPa. Therefore, all the experiments including fingertip force testing were conducted under the safety input pressure, ranging from 0 kPa to 65kPa. The fingertip force was tested at incremental steps of 5 kPa and the average data were obtained after five repeated tests. And all the results were shown in Fig. 5. The maximum fingertip force is 7.83N in 65kPa, indicating that a single actuator as Fig. 5 illustrates. Compared with other actuators in the same pressure, 7.83N is a relatively high value \cite{abondance2020dexterous,zhang2022gas,chen2017reconfigurable,low2020bidirectional,park2018development}. The fabrication process which is based on the combination of the lost-wax method and porous connector ensures airtightness, increasing the range of allowable input pressure and thus resulting in a larger fingertip force. \subsection{Stiffness in the B Direction} The MTSA was placed in the same frame mentioned in the fingertip force section. Considering that the soft actuators, which are used for grasping usually interact with subjects through the contact end directly, the extension direction of MTSA was chosen as the evaluation direction of bending stiffness, as Fig. 6a displays. A carbon fiber tube with a 26mm diameter was used to generate the pulling force by being connected to the force gauge which was assembled on the sliding table through ropes. The sliding table was programmed to move 1mm rightward each time until the total distance reached 10mm and the corresponding pulling force during this process would be recorded. Then, the B-direction stiffness can be characterized by calculating the incremental ratio of the force and displacement. According to the working principle, the B-direction stiffness can be tuned in different angles, Hence, the experiments of B-direction stiffness were conducted in four groups of angle, $45\degree$, $90\degree$, $135\degree$, and $180\degree$. For each angle, four groups of different incremental pressures were chosen. And every single set was repeated three times to obtain the average result. Fig. 6a draws the experimental result of the group $90\degree$. As it shows, the relationship between pulling force and displacement is nearly linear. Therefore, all the stiffnesses were calculated as the slope of the fitted line to pulling force and displacement. To determine the magnitude of the change in stiffness more intuitively, the ratio of stiffness variation for each group was shown in Fig. 6b. From Fig. 6b, for all the four different bending angles, the stiffness was enhanced from 1.5 to 4 times, with a maximum stiffness of 1.29 N/mm, indicating that the stiffness in the bending direction is tuned successfully through our proposed mechanism. The result also showed that the tunable stiffness ranges when the bending angles were $45\degree$ and $90\degree$ were larger than when the bending angle is $135\degree$ or $180\degree$. This indicated that the range of tunable stiffness is determined by the increased pressure since when the bending angle is $135\degree$ or $180\degree$, the input pressure of MTSA is close to the maximum allowable pressure which limited their incremental pressure. \begin{figure*}[thpb] \centering \includegraphics[scale=0.95]{FIG/Fig6.eps} \caption{The experimental results of stiffness testing. (a) Measurement of the relationship between pulling force and displacement in the B direction when the bending angle is . (b) Aggregation of the stiffness in different bending angles with different increased pressure. (c) Measurement of the PB-direction stiffness when the bending angle is and the input pressure is 0 kPa. (d) Aggregation of PB-direction stiffness in different bending angles and different weights to pull the rope. The lower plane on each graphic is the result of the stiffness of the actuator without BLSs.} \label{fig6} \end{figure*} \subsection{Stiffness in the PB Direction} To verify the multi-direction tunable stiffness property in the PB direction, the same platform mentioned before was utilized. And the MTSA ’s side contacted with the force gauge. Based on the working principle, PB-direction stiffness can be modulated by changing the tension of the ropes of BLSs. Therefore, two weights were connected through pulleys to each BLS to quantify the ability of variable stiffness as Fig. 6c illustrates. Two kinds of weights, 0.5kg one and 1kg one, were available to combine into different weights. Three groups of bending angles, $0\degree$, $45\degree$, and $90\degree$ were chosen to estimate the PB-direction stiffness. It is worth stating that although in the working principle section, the effect of enhanced PB-direction stiffness brought by BLS was mainly discussed, the stiffness in the PB direction is influenced by the coupling of the input pressure and BLSs. Hence, the experiment for each bending angle included the different increased pressures and weights to pull the rope. Five groups of weights 0kg, 0.5kg, 1kg, and 2kg and at least three groups of pressure at incremental steps of 10 kPa were conducted for each bending angle. Moreover, to compare the ability of BLS, the same input air pressure for actuator without BLSs were tested. The sliding table moved at a 2-mm interval from 0 to 40mm to control the deflection. A series of data when bending angle is $0\degree$ and the increased pressure is 0kPa is shown in Fig. 6c. From our analysis, the slope of all curves begins to decrease when the displacement is larger than 20mm because torsion deformation generated. Therefore, the PB-direction stiffnesses were calculated within the range of data that had linear characteristics. Then, the relationship among stiffness, weights loaded to the ropes, and increased pressure was shown on Fig. 6d for $0\degree$, $45\degree$, and $90\degree$ Respectively. As Fig. 6d illustrates, comparisons between the PB-direction stiffness of actuator with BLSs and without BLSs confirms the enhanced stiffness property of BLS. The stiffness of actuator with BLSs is enhanced 4.2, 3.5, and 2.2 times for $0\degree$,$45\degree$, and $90\degree$ group than actuator without BLSs. Meanwhile, as can be seen, the stiffness increase with the weights to pull the rope and air pressure and the range of tunable stiffness are about 1.5, 1.2, and 1.2 times for $0\degree$,$45\degree$, and $90\degree$ group. These results proved that the PB-direction stiffness was tuned decoupling successfully. In addition, the largest stiffness when no increased pressure was input, increases when the bending angle varied from $0\degree$ to $45\degree$ , and decreases when the bending angle is $90\degree$. This trend follows the analysis proposed in the working principle section. However, in other conditions, the stiffness did not share the same trends. This indicated the stiffness analysis model needs to be revised further. And the loose BLSs caused by the uncontrollable axial elongation of MTSA may contribute to the results. Based on the results of stiffness tests in the B direction and PB direction, it can be concluded that the multi-direction tunable stiffness property is achieved through the combination of tendon-gas-hybrid actuation and bone-like structures. \section{CONCLUSIONS} This study introduces a soft-rigid hybrid actuator with multi-direction tunable stiffness property and a specific fabrication strategy to improve the airtightness. Then the working principle of the multi-direction tunable stiffness was analyzed. To characterize its performance, several experiments including fingertip force and stiffness tests were conducted. Overall, the stiffness in both the B direction and PB direction can be decoupled for modulation in a wide range, which achieves the design concept of our proposed mechanism. However, during our experiments and theoretical analysis, a few drawbacks of MTSA still exist. Firstly, although the working principle section has analyzed that one of the important factors is the moment of inertia, a better geometric dimension was not evaluated due to geometrical limitations brought by the soft chamber. Second, the simplification of the mathematical model of the PB-direction stiffness of BLS is rough and cannot predict the trend of changes in the experiment well. Third, though our fabrication method has improved the airtightness of the connection boundary between the solid and soft chamber, the lateral wall of the front chamber is easy to explode during pressurizing. The excessively thin wall thickness was caused by the displacement that happened at the distal end of the lost wax core since it only has one simple support. Therefore, in our future work, the fabrication method is due to be revised and the optimization solutions for designing the MTSA considering the coupling effect of the geometrical size of the soft chamber and BLSs will be conducted. Meanwhile, the MTSA will be utilized in different practical scenarios to perform its ability to tune the stiffness decoupling, like versatile gripper and soft crawling robot. \bibliographystyle{IEEEtran}

\section{Introduction} In the words of Hermann Weyl: ``{\em A thing is symmetrical if there is something you can do to it so that after you have finished doing it, it looks the same as before.}'' \smallskip Symmetries can be discrete or continuous. For instance, the polynomial $f(x_1,\ldots, x_n):=x_1^2 + \cdots + x_n^2$ is symmetric under any permutation of its variables. The associated set of symmetries is the symmetric group $S_n$. The Boolean cube $\{0,1\}^n$ also remains invariant under the permutation of coordinates by $S_n$. A square remains unchanged if we rotate it by multiples of $90$ degree around its center. The integer lattice $\mathbb{Z}^n$ remains invariant under translation by any vector with integer entries. These are examples where the set of actions that preserve the object in question is discrete -- finite or infinite. An infinite line, e.g., $\mathbb{R}$ can be translated by any real number and it remains the same. A sphere can be rotated arbitrarily and it remains the same. An $n \times n$ complex matrix $A$ can be conjugated by a unitary $U$ as $UAU^*$ and its eigenvalues remain the same. In all of these cases, the set of actions that preserves the object is continuous and hence, infinite. Symmetries, both discrete and continuous, play a central role in physics and mathematics. In physics, symmetries have long served as a guiding principle to search for the laws of nature \cite{Gross14256}. Physical scenarios are encoded using numbers, vectors, and matrices that depend on the choice of a ``reference frame,'' and we expect the laws of nature to be invariant to changes in the reference frame. For instance, via the addition of a new structure -- the curvature of spacetime -- the equations of general relativity makes it possible to (approximately) describe the universe from any reference frame or point of view using the same equations. Interestingly, each global continuous symmetry of a physical system imposes the conservation of a quantity: symmetry under translation implies conservation of momentum, symmetry under rotation implies conservation of angular momentum, and symmetry under time implies conservation of energy. This is the content of Noether's theorem \cite{Noether1918}. In mathematics, symmetries are important in almost all branches. In particular, Lie groups that encode continuous symmetries are central to analysis, topology, algebraic geometry, differential geometry, number theory, and Riemannian geometry. Moreover, as we explain in Sections \ref{sec:Weyl} and \ref{sec:classification}, there is deep connection between several discrete groups and Lie groups, leading to their applications in discrete mathematics and combinatorics. Symmetries -- both discrete and continuous -- also arise in optimization and sampling problems and have been used in the design of efficient algorithms. While there is a growing list of problems and works where the lens of symmetry has been helpful (see \cite{Barvinok05convexgeometry,sanyal2011,Waterhouse:1983:DSP,Edelman1999,Parrilo,GargGOW16,GargGOW17,SraVY18,Allen-ZhuGLOW18,BurgisserFGOWW19,LeakeV20,LeakeV20b,LMV2021} and the references therein), unlike mathematics and physics where understanding symmetries is an essential part of the basic toolkit of a student, a systematic treatment of symmetries from the point of view of optimization is lacking. The goal of this article is not to fill this void, but 1) to entice the reader enough to appreciate continuous symmetries and 2) to present an introduction to Lie theory from scratch. We start with the question of why continuous symmetries might arise in optimization and sampling problems? The answer is often similar to the same question in physics. Sometimes, the function that we would like to optimize corresponds to ``physical'' or geometric quantities such as distances, inner products, volumes, or curvature associated with points in a domain involving vectors, matrices, or tensors. These geometric quantities often do not depend on the reference frame used to represent the underlying structures such as vectors, matrices, and tensors and, hence, continuous symmetries arise. For instance, inner products, determinants, and trace, all have well-known symmetries. Moreover, this set of symmetries ``compose'' with each other in a natural manner. \newpage Consider the function $ f(x_1,\ldots,x_k):= \sum_{i,j=1}^k \|x_i-x_j\|^2$, that is the sum of squared Euclidean distances between $n$ points with positions $x_1,\ldots,x_k \in \mathbb{R}^n$. This function is invariant to translating all the points by the same vector $c \in \mathbb{R}^n$: $$f(x_1+c,\ldots, x_n+c)=f(x_1,\ldots,x_n).$$ This set of translations corresponds to $\mathbb{R}^n$ and the translations compose as ``addition:'' Translating by $c_1$ and then by $c_2$ has the same effect as translating by $c_1+c_2$, endowing the Euclidean space with an abelian group structure. The function $f$ is also invariant under rotations. Let $Q$ be an $n \times n$ orthogonal matrix, meaning that $QQ^\top = Q^\top Q = I$. Then, $$f(Qx_1,\ldots,Qx_n)=f(x_1,\ldots,x_n).$$ This is because the orthogonality of $Q$ implies that $$ \|Qx_i-Qx_j\|^2 = \langle Q(x_i-x_j), Q(x_i-x_j) \rangle = (x_i-x_j)^\top Q^\top Q (x_i-x_j)=(x_i-x_j)^\top(x_i-x_j) = \|x_i-x_j\|^2.$$ Note that the set of $n \times n$ orthogonal matrices forms a group under matrix multiplication; however, this group is nonabelian. Moreover, sometimes, the domain itself might be symmetric. Consider the complex unit sphere $S_\mathbb{C}^n:= \left\{ (x_1,\ldots, x_n)\in \mathbb{C}^n: \sum_{i=1}^n |x_i|^2 = 1\right\}.$ For an $n \times n$ unitary matrix $U$, we have that $$U \circ S_\mathbb{C}^n := \{Ux: x \in S_\mathbb{C}^n\}= S_\mathbb{C}^n$$ because $\|Ux\|^2=\|x\|^2$. A matrix $U \in \mathbb{C}^{n \times n}$ is said to be unitary if $UU^* = I$. Hence, the action of such a unitary matrix does not change the domain. Moreover, just like orthogonal matrices, the set of unitary matrices also forms a group under matrix multiplication. What is special about the three groups we have seen so far? It is that one can do calculus on them (and hence optimize functions) and they come with natural measures associated with them and hence, we can define sampling problems on them. In the first part of this article, we focus on two concrete examples -- one involving an optimization problem and one involving a sampling problem. Both these results have multiple elementary proofs. We also present elementary and self-contained proofs of these results. The proofs of both results rely on first establishing a connection between a Lie group and a convex polytope. And then, going from this convex polytope to a discrete group. These proofs have been carefully chosen to set the stage for the introduction of Lie theory and, finally, presentation of deep generalizations of both these results. \paragraph{Example 1: Minimum eigenvalue of a matrix.} The first problem we consider is the following well-known optimization characterization of the smallest eigenvalue of a Hermitian matrix.\footnote{An $n \times n$ matrix $A$ with entries in $\mathbb{C}$ is Hermitian if $A=A^*$.} \begin{theorem}\label{thm:eigenvalue} Let $A$ be an $n \times n$ Hermitian matrix with smallest eigenvalue $\lambda_1$, then \begin{equation}\label{eq:eigenvalue} \min _{v \in \mathbb{C}^n: \|v\|=1} v^* A v = \lambda_1. \end{equation} \end{theorem} \noindent Neither the objective function nor the domain in Equation \eqref{eq:eigenvalue} is convex. We start by rewriting the above problem as an optimization problem over the following (nonconvex) manifold: \[ \mathcal{P}_1:= \{ X: X \in \mathbb{C}^{n \times n}, \Tr(X)=1, X=X^*, X^2=X\}. \] The reason is that $\mathcal{P}_1$ consists of rank one projection matrices and the only rank one projections are of the type $vv^*$ for a vector $v$ such that $\|v\|=1$. Over this manifold, the objective function of Equation \eqref{eq:eigenvalue} becomes linear: $$ \langle A,X\rangle_F.$$ Here $\langle \cdot, \cdot \rangle_F$ is the Frobenius inner product in the space of matrices. The first observation is that while $\mathcal{P}_1$ is nonconvex, it has symmetries: Any $n \times n$ unitary matrix $U$ acting on $\mathcal{P}_1$ by conjugation leaves $\mathcal{P}_1$ invariant: $$ U \mathcal{P}_1 U^* = \{ UXU^*: X \in \mathcal{P}_1 \} = \mathcal{P}_1.$$ The second observation is that the set of all $n \times n$ unitary matrices is a group under matrix multiplication: the {unitary group} $\U(n)$, which is also a manifold. Thus, $$\mathcal{P}_1 = \{ Ue_1e_1^*U^*: U \in \mathrm{U}(n)\}.\footnotemark{}$$ Hence, the problem in Equation \eqref{eq:eigenvalue} is not just any nonconvex problem, it can be paramterized as a linear optimization problem over a continuous group that is also a manifold; in fact, a Lie group. This structure is implicitly or explicitly used in any proof of Theorem \ref{thm:eigenvalue}. In particular, this structure implies a ``convexity'' which, in turn, implies Theorem \ref{thm:eigenvalue}. \footnotetext{$e_1$ is the vector that is one in the first coordinate and zero elsewhere.} In Section \ref{sec:eigenvalue}, we present a proof of this theorem from this point of view. Where we show how this optimization problem can be reduced to a linear optimization problem over a polytope and, consequently, over its vertices. The reader might wonder why should we care about this point of view and such a proof. Especially since Theorem \ref{thm:eigenvalue} can be proved without a reference to a group or manifold. In short, the answer is that this viewpoint leads to a far-reaching generalization: Any linear optimization problem over such a symmetric object reduces to a linear optimization problem over an associated convex polytope; see Theorem \ref{thm:Kostant} and Corollary \ref{cor:linearopt}. However, to state the latter result formally and understand its proof, one needs the language of Lie theory, which is discussed in Section \ref{sec:Lie}. \paragraph{Example 2: Sampling from an exponential density on the complex unit sphere.} The second example we consider is the ``sampling'' version of the first example. Given an $ n \times n$ Hermitian matrix $A$, sample from the probability distribution where a unit vector $v$ on the complex unit sphere is picked with probability proportional to $e^{-v^*Av}$. The complex unit sphere in $n$ dimensions is the set $S_\mathbb{C}^n := \{(z_1,\ldots,z_n) \in \mathbb{C}^n: \sum_i |z_i|^2=1\}.$ First, to formally define this problem, we note that the complex unit sphere has a Haar measure $\mu_{S_\mathbb{C}^n}$ with respect to which we can define the above density. And second, the density $e^{-v^*Av}$ favors vectors with lower values of $v^*Av$, making this problem as a ``robust'' version of the optimization problem. Finally, note that the density can be viewed as an exponential density $e^{-\langle A,X\rangle_F}$ where $X \in \mathcal{P}_1$. Let $\mu_1$ be the pushforward measure of the map from the complex unit sphere to $\mathcal{P}_1$ that maps $v \mapsto vv^*$. $\mu_1$ is the unique unitarily invariant measure on $\mathcal{P}_1$. Thus, we have not only a group structure, a manifold structure (as discussed in the optimization example), but also an invariant measure. Such exponential densities arise under the name \emph{matrix Langevin} or {\em matrix Bingham} in statistics \cite{ChikusePaper,ChikuseBook}, in quantum inference \cite{LMV2021}, and in the context of differentially private rank-$1$ approximation of a given matrix \cite{Kamalika,KTalwar}. Unlike the eigenvalue (optimization) problem discussed earlier, efficient sampling algorithms from such densities were only recently discovered \cite{LMV2021}, and they crucially relied on the ``symmetry viewpoint.'' In fact, a very related and almost equivalent problem is that of computing the following integral or the partition function of this exponential density: $$ \int_{\mathcal{P}_1} e^{-\langle A,X\rangle_F}d\mu_1(X).$$ Here, we focus on this problem instead of the sampling problem. We prove the following formula for this integral that immediately implies an efficient algorithm to compute it. \begin{theorem}\label{thm:partition} Let $A$ be an $n\times n$ Hermitian matrix with distinct eigenvalues $\lambda_1 < \cdots < \lambda_n$. % Then $$ \int_{\mathcal{P}_1} e^{-\langle A,X\rangle_F}d\mu_1(X) = (n-1)! \sum_{i=1}^n \frac{e^{-\lambda_i }}{\prod_{j \neq i} (\lambda_j-\lambda_i)}. $$ \end{theorem} It should be surprising that this integral (over the unitary group) reduces to such a formula involving a finite sum with a small number of terms. While it is not obvious, we point out that the right hand side is in fact a determinant (of a Vandermonde-type) matrix which, in turn, is a summation over $S_n$. Thus, an integral over $\U(n)$ reduces to a sum over the symmetric group $S_n$. We present a self-contained proof of this theorem in Section \ref{sec:ex2} that relies on a connection between two highly symmetric objects -- $\mathcal{P}_1$ and (another convex polytope) the probability simplex. Once we see a proof for the case when the eigenvalues are all distinct, it is not too difficult to come up with a similar formula in the case when eigenvalues are repeated. Moreover, while we do not present it here, the ideas that go in the proof of this theorem can be converted into an efficient sampling algorithm; see \cite{LMV2021}. The reader would have every reason to suspect that this result is also a tip of the iceberg: This formula is a very special case of the Harish-Chandra integral formula; see Theorems \ref{thm:HC} and \ref{thm:HCIZ}. While one needs the toolkit of Lie theory to formulate and prove these results, let us conclude by mentioning that these results are remarkable: They convert certain integrals over continuous groups to summations over discrete groups -- Weyl groups. \section{Optimization Characterization of the Minimum Eigenvalue of a Matrix}\label{sec:eigenvalue} In this section we prove Theorem \ref{thm:eigenvalue}. Recall that we are given an $n \times n$ Hermitian matrix $A$ and our goal is to prove that $$ \min_{X \in \mathcal{P}_1} \langle A,X \rangle_F = \min_{U \in \U(n)} \langle A,Ue_1e_1^\top U^* \rangle = \lambda_1.\footnotemark{}$$ Let $A=WDW^*$ where $W$ is a unitary matrix and $D=\diag(\lambda)$ is the diagonal matrix consisting of eigenvalues of $A$. Then we obtain\footnotetext{$e_1e_1^\top$ is a diagonal matrix with $1$ in the $(1,1)$ entry and $0$ everywhere else.} $$ \min_{U \in \U(n)} \langle WDW^*,Ue_1e_1^\top U^* \rangle_F = \min_{U \in \U(n)} \langle U^*WDW^*U,e_1e_1^\top \rangle_F = \min_{V \in \U(n)} \langle VDV^*,e_1e_1^\top \rangle_F.$$ Thus, the answer to this minimization problem is the smallest entry in the diagonal of $VDV^*$ when $V$ varies over $\U(n)$. \paragraph{From a unitary orbit to a polytope.} We compute the diagonal entries of $A = VDV^*$ via: \[ a_{ii} = \sum_{j=1}^n v_{ij} \lambda_j v_{ij}^* = \sum_{j=1}^n |v_{ij}|^2 \lambda_j = [(V \odot V^*)\lambda]_i. \] That is, the diagonal of $A$ is given by $(V \odot V^*) \lambda$, where $\odot$ denotes the entrywise product. % Let $Q := V \odot V^*$. % All entries of $Q$ are nonnegative. % Since $V$ is a unitary matrix, we have that for all $1\leq j \leq n$, $$ \sum_{i=1}^n Q_{ij} = \sum_{i=1}^n |v_{ij}|^2 = 1.$$ Similarly, for all $1\leq i \leq n$, $$ \sum_{j=1}^n Q_{ij} = \sum_{j=1}^n |v_{ij}|^2 = 1.$$ Thus, $Q$ is doubly stochastic. % The set of all doubly stochastic matrices is convex and, in fact, a polytope -- the Birkhoff polytope \cite{birk:46}. % The well-known (and easy to prove) Birkhoff-von Neumann theorem states that the Birkhoff polytope is a convex hull of $n \times n$ permutation matrices. % Thus, $Q$, which is doubly stochastic can be written as a convex combination of permutation matrices. % Hence, for any unitary $V$, the diagonal of $A=VDV^*$ which is $Q\lambda$, can be written as a convex combination of permutations of $\lambda$. Thus, we have made a connection between the set of all Hermitian matrices with eigenvalues $\lambda=(\lambda_1,\ldots, \lambda_n)$ -- the unitary ``orbit'' of $\diag(\lambda)$ -- and the ``permutation polytope'' $P_\lambda$ of $\lambda$ defined as the convex hull of all the permutations of the vector $\lambda$: $$ P_\lambda := \mathrm{hull} \{ (\lambda_{\pi(1)},\ldots,\lambda_{\pi(n)}): \pi \in S_n \}. $$ We have argued that the diagonals of the matrices in the unitary orbit of $\diag(\lambda)$ lie in $P_\lambda$. $$ \min_{V \in \U(n)} \langle VDV^*,e_1e_1^\top \rangle_F \geq \min_{v \in P_\lambda} \langle v ,e_1 \rangle= \lambda_1.$$ Moreover, plugging in the unitary $U$ which diagonalizes $A$ and ensures that the smallest eigenvalue of $A$ appears as the $(1,1)$ entry in $VAV^*$. Thus, we get that $$ \lambda_1 = \langle UAU^*, e_1e_1^\top \rangle_F \geq \min_{V \in \U(n)} \langle UDU^*,e_1e_1^\top \rangle_F.$$ This completes the proof of Theorem \ref{thm:eigenvalue}. \paragraph{The Schur-Horn theorem.} The fact that, for a given vector $\lambda$, the set of diagonal vectors of matrices in the unitary orbit of $\diag(\lambda)$ lie in $P_\lambda$ is often known as the ``Schur'' part of the Schur-Horn theorem \cite{schur1923,horn1954}. The converse, known as the ``Horn'' part, is also true, giving us the Schur-Horn theorem; the proof below is adapted from \cite{Kadison4178}. \begin{theorem}[Schur-Horn] \label{prop:schur_horn} If $D := \diag(\lambda)$ is an $n \times n$ real diagonal matrix and $U$ is an $n \times n$ unitary matrix, then the diagonal of $UDU^*$ is in $P_\lambda$. Conversely, given any vector $v \in P_\lambda$, there exists a unitary matrix $U$ such that $UDU^*$ has diagonal vector $v$. \end{theorem} \begin{proof}{\bf (Sketch)} We have already proved the first part of this theorem. % For the other direction it suffices to show: If $w$ is the diagonal vector of some $UDU^*$ and $\tau$ is any transposition of $i,j \in [n]$, then $(1-t)w + t(\tau \cdot w)$ is the diagonal vector of $VDV^*$ for some unitary $V$ for any $t \in [0,1]$, where $\tau \cdot w$ indicates the corresponding permutation of the entries of $w$. % Since $w \mapsto (1-t)w + t(\tau \cdot w)$ only alters 2 entries of $w$, it suffices to demonstrate the desired claim on $2 \times 2$ matrices. To show this, we first have from the previous paragraph that $$\diag(UDU^*) = (|u_{11}|^2\lambda_1 + |u_{12}|^2\lambda_2, |u_{21}|^2\lambda_1 + |u_{22}|^2\lambda_2).$$ % Choosing $U := \left[\begin{smallmatrix} \sqrt{1-t} & \sqrt{t} \\ -\sqrt{t} & \sqrt{1-t} \end{smallmatrix}\right]$ for $t \in [0,1]$ gives $$\diag(UDU^*) = \left((1-t) \cdot \lambda_1 + t \cdot \lambda_2, t \cdot \lambda_1 + (1-t) \cdot \lambda_2\right).$$ % Letting $t$ vary over $[0,1]$ then gives all possible length-2 vectors majorized by $\lambda$. \end{proof} \noindent Thus, if we define a mapping $$ \phi: \{U\diag(\lambda) U^*: U \in \U(n)\} \to P_\lambda$$ such that $$ \phi(U\diag(\lambda) U^*)=\diag(U\diag(\lambda) U^*),$$ then the range of this map is the convex polytope $P_\lambda$. Note that $P_{(1,0,\ldots,0)}$ is the probability simplex. \paragraph{A generalization of Theorem \ref{thm:eigenvalue}: From $ \U(n)$ to $S_n$.} We now show how to use the Schur-Horn theorem to reduce the problem of optimizing a linear function over the infinite unitary group to the finite symmetric group. Given two real diagonal $n \times n$ matrices $D$ and $D'$, we have that: \begin{equation}\label{eq:Sn} \min_{U \in \U(n)} \langle UDU^*, D' \rangle_F = \min_{\sigma \in S_n} \langle \sigma D \sigma^*, D' \rangle_F. \end{equation} To see this, let $\sigma_0$ be the permutation matrix which minimizes $\langle \sigma D \sigma^*, D' \rangle_F$ over all permutation matrices $\sigma$.\footnote{Note that all permutation matrices are unitary.} % By the Schur-Horn theorem, for any $U$, the diagonal of $UDU^*$ can be written as a convex combination of the permutations of the diagonal of $D$. % By linearity of $\langle \cdot, D' \rangle_F$, the value of $\langle UDU^*, D' \rangle_F$ must then be at least the value of $\langle \sigma_0 D \sigma_0^*, D' \rangle_F$. \section{Sampling from an Exponential Density on the Complex Sphere}\label{sec:ex2} In this section we prove Theorem \ref{thm:partition}. This proof is adapted from \cite{LeakeV20}; see also \cite{Vergne1996}. \paragraph{From $\mathcal{P}_1$ to the simplex.} The first step is to observe the following remarkable equality for real $A = \diag(\lambda)$, where $m$ is the Lebesgue measure on the simplex $\Delta_1:=\{p \in \mathbb{R}_+^n: \sum_{i=1}^n p_i=1\}$: \[ \int_{\mathcal{P}_1} e^{-\langle A, X \rangle_F} d\mu_1(X) = \int_{\Delta_1} e^{-\langle \lambda, x \rangle} dm(x). \] Put another way, exponential measures on $\mathcal{P}_1$, a nonconvex manifold, correspond to exponential measures on $\Delta_1$, a convex polytope. To see this, first note the following equalities for any nonegative integers $m_1,\ldots,m_n$. The first equality is the Bombieri inner product formula (Lemma \ref{lem:bombieri}), and the second equality is a basic induction after a change of variables: \[ \int_{\mathcal{P}_1} X_{11}^{m_1} \cdots X_{nn}^{m_n} d\mu_1(X) = \frac{m_1! \cdots m_n! (n-1)!}{(m_1+\cdots+m_n+n-1)!} = \int_{\Delta_1} x_1^{m_1} \cdots x_n^{m_n} dm(x). \] The exponential equality then follows from taking limits, since $\mathcal{P}_1$ and $\Delta_1$ are compact and $e^{-\langle A, X \rangle_F}$ and $e^{-\langle \lambda, x \rangle}$ are limits of polynomials. We note that this argument also implies the more general fact that $m$ is the pushforward of $\mu_1$ through the map $\phi: X \mapsto \diag(X)$: \[ \int_{\mathcal{P}_1} f(\phi(X)) d\mu_1(X) = \int_{\Delta_1} f(x) dm(x). \] \paragraph{From the simplex to a finite sum.} This transfer to the simplex now leads to an explicit computation for the integral when $A$ is a diagonal matrix $A= \diag(\lambda)$ as a finite sum. The assumption that $A$ is diagonal is without loss of generality because conjugating $A$ by any unitary does not change the integral $\int_{\mathcal{P}_1}e^{-\langle A,X\rangle_F} d\mu(X).$ By making a change of variables, the simplex integral becomes an iterated convolution: \[ \int_{\Delta_1} e^{-\langle \lambda, x \rangle} dm(x) = (n-1)!\int_0^1 \int_0^{1-x_1} \cdots \int_0^{1-x_1-\cdots-x_{n-2}} e^{-\langle \lambda, x \rangle} dx = \left.(e^{-\lambda_1 t} * \cdots * e^{-\lambda_n t})\right|_{t=1}.\footnotemark{} \] This is stated formally in Lemma \ref{lem:laplace_integral}. Applying the Laplace transform $\mathcal{L}$ converts this convolution into a partial fraction decomposition problem for distinct values of $\lambda_i$: \footnotetext{Here, $*$ denotes the usual integral convolution: Given functions $f, g\colon \R\to \R$, $(f*g)(t)$ is $\int_{-\infty}^\infty f(t-\tau)g(\tau)d\tau$.} \[ \left.(e^{-\lambda_1 t} * \cdots * e^{-\lambda_n t})\right|_{t=1} = \mathcal{L}^{-1}\left[\frac{1}{\prod_i (s+\lambda_i)}\right](1) = \mathcal{L}^{-1}\left[\sum_i \frac{c_i}{s+\lambda_i}\right](1) = \sum_i c_i e^{-\lambda_i}. \] Computing the values of $c_i$ via a standard partial fractions formula gives: \[ \frac{1}{(n-1)!} \int_{\mathcal{P}_1} e^{-\langle Y, X \rangle_F} d\mu_1(X) = \sum_{i=1}^n \frac{e^{-\lambda_i}}{\prod_{j \neq i} (\lambda_j-\lambda_i)}. \] This is stated formally in Proposition \ref{prop:evaluation_formula}. \paragraph{Proof of Theorem \ref{thm:partition}.} We now state a lemma which gives the most basic result about integrals on $\mathcal{P}_1$. Specifically, we state a well-known result for integrals of polynomial-like functions. This proof is very related to the unitarily invariant inner product on homogeneous polynomials, which has many names in the literature: Bombieri-Weyl inner product, Fischer-Fock inner product, Segal-Bargmann inner product, etc. The following lemma is standard; see e.g. Lemma 3.2 of \cite{pinasco2012}. \begin{lemma}[Bombieri inner product formula] \label{lem:bombieri} For $\alpha \in \{0,1,2,\ldots\}^n$ such that $\sum_i \alpha_i = d$, we have: \[ \int |v|^{2\alpha} d\mu_{S_\mathbb{C}^n}(v) = \int \prod_i |v_i|^{2\alpha_i} d\mu_{S_\mathbb{C}^n}(v) = \binom{d}{\alpha}^{-1} \binom{d+n-1}{n-1}^{-1} = \frac{\alpha_1! \cdots \alpha_n!(n-1)!}{(d+n-1)!}. \] Here, $\binom{d}{\alpha}$ is the multinomial coefficient and $\binom{d+n-1}{n-1}$ is the binomial coefficient. \end{lemma} \noindent The next lemma relates the integrals we want to compute and the Laplace transform. \begin{lemma} \label{lem:laplace_integral} For $\lambda_1,\ldots,\lambda_n \in \R$ and $x_n := 1-x_1-\cdots-x_{n-1}$, we have the following where $*$ denotes the usual integral convolution: \[ \int_0^1 \int_0^{1-x_1} \cdots \int_0^{1-x_1-\cdots-x_{n-2}} e^{-\langle \lambda, x \rangle} dx_{n-1} \cdots dx_1 = \left.(e^{-\lambda_1 t} * \cdots * e^{-\lambda_n t})\right|_{t=1}. \] If $\lambda_1 < \lambda_2 < \cdots < \lambda_n$, then we further have: \[ \int_0^1 \int_0^{1-x_1} \cdots \int_0^{1-x_1-\cdots-x_{n-2}} e^{-\langle \lambda, x \rangle} dx_{n-1} \cdots dx_1 = \sum_{i=1}^n \frac{e^{-\lambda_i}}{\prod_{j \neq i} (\lambda_j-\lambda_i)} \] \end{lemma} \begin{proof} We first compute: \[ \begin{split} \int_0^1 \int_0^{1-x_1} &\cdots \int_0^{1-x_1-\cdots-x_{n-2}} e^{-\langle \lambda, x \rangle} dx_{n-1} \cdots dx_1 \\ &= \int_0^1 e^{-\lambda_1x_1} \cdots \int_0^{1-x_1-\cdots-x_{n-2}} e^{-\lambda_{n-1}x_{n-1}} e^{-\lambda_n(1-x_1-\cdots-x_{n-1})} dx_{n-1} \cdots dx_1 \\ &= \int_0^1 e^{-\lambda_1x_1} \cdots \int_0^{1-x_1-\cdots-x_{n-3}} e^{-\lambda_{n-2}x_{n-2}} \left.(e^{-\lambda_{n-1} t} * e^{-\lambda_n t})\right|_{t=1-x_1-\cdots-x_{n-2}} dx_{n-2} \cdots dx_1 \\ &=\cdots = \int_0^1 e^{-\lambda_1x_1} \left.(e^{-\lambda_2 t} * \cdots * e^{-\lambda_n t})\right|_{t=1-x_1} dx_1 = \left.(e^{-\lambda_1 t} * \cdots * e^{-\lambda_n t})\right|_{t=1}. \end{split} \] Using the Laplace transform, we have $\mathcal{L}[e^{-\lambda_i t}](s) = \frac{1}{s+\lambda_i}$ which implies: \[ e^{-\lambda_1 t} * \cdots * e^{-\lambda_n t} = \mathcal{L}^{-1}\left[\frac{1}{(s+\lambda_1)(s+\lambda_2) \cdots (s+\lambda_n)}\right](t). \] Assuming $\lambda_1 < \lambda_2 < \cdots < \lambda_n$, we can use Lagrange interpolation to compute: \[ \mathcal{L}^{-1}\left[\frac{1}{(s+\lambda_1)(s+\lambda_2) \cdots (s+\lambda_n)}\right](t) = \mathcal{L}^{-1}\left[\frac{1}{Q'(-\lambda_1) \cdot (s+\lambda_1)} + \cdots + \frac{1}{Q'(-\lambda_n) \cdot (s+\lambda_n)}\right](t). \] Here, $Q(s) := (s+\lambda_1)(s+\lambda_2) \cdots (s+\lambda_n)$. With this we have: \[ \mathcal{L}^{-1}\left[\frac{1}{Q'(-\lambda_1) \cdot (s+\lambda_1)} + \cdots + \frac{1}{Q'(-\lambda_n) \cdot (s+\lambda_n)}\right](t) = \frac{e^{-\lambda_1 t}}{Q'(-\lambda_1)} + \cdots + \frac{e^{-\lambda_n t}}{Q'(-\lambda_n)} = \sum_{i=1}^n \frac{e^{-\lambda_i t}}{\prod_{j \neq i} (\lambda_j-\lambda_i)}. \] Plugging in $t=1$ gives the equality in the second statement. \end{proof} \noindent We now state and prove the full evaluation formula for $\mathcal{P}_1$ in full generality. \begin{proposition}[Generalization of Theorem \ref{thm:partition}] \label{prop:evaluation_formula} Fix $n \in \N$, and let $\mu_{S_\mathbb{C}^n}, \mu_1,$ and $\mu_{\Delta_1}$ be the uniform probability distributions on the complex unit sphere in $\mathbb{C}^n$, on $\mathcal{P}_1$, and on the standard simplex in $\R^n$, respectively. For a given analytic function $f$ on the standard simplex the following expressions are equal: \begin{enumerate} \item $\displaystyle \int_{S_\mathbb{C}^n} f(|v_1|^2,\ldots,|v_n|^2) d\mu_{S_\mathbb{C}^n}(v)$, \item $\displaystyle \int_{\mathcal{P}_1} f(\diag(X)) d\mu_1(X)$, \item $\displaystyle \int_{\Delta_1} f(x) d\mu_{\Delta_1}(x)$, \item $\displaystyle (n-1)! \int_0^1 \int_0^{1-x_1} \cdots \int_0^{1-x_1-\cdots-x_{n-2}} f(x_1,\ldots,x_{n-1},1-x_1-\cdots-x_{n-1}) dx_{n-1} \cdots dx_1$. \end{enumerate} \end{proposition} \begin{proof} First, for the equality of $(1)$ and $(2)$, note that $\mu_1$ is the pushforward measure of $\mu_{S_\mathbb{C}^n}$ through the map $\psi: S_\mathbb{C}^n \to \mathcal{P}_1$ given by $\psi: v \mapsto vv^*$. (To see this, note that $\psi$ is unitarily invariant and $\mu_{S_\mathbb{C}^n}$ and $\mu_1$ are the unique unitarily invariant measures on their domains.) With this, we have: \[ \int_{\mathcal{P}_1} f(\diag(X)) d\mu_1(X) = \int_{S_\mathbb{C}^n} f(\diag(\psi(v))) d_{S_\mathbb{C}^n}(v) = \int_{S_\mathbb{C}^n} f(|v_1|^2,\ldots,|v_n|^2) d_{S_\mathbb{C}^n}(v). \] That is, $(1)$ and $(2)$ are equal. Next, the equality of $(3)$ and $(4)$ follows from the fact that the map between the two domains of integration (both of which are simplices) is affine. Therefore the determinant of the Jacobian is a constant, and thus we only need to integrate over a constant function to determine that constant. A simple induction shows that it is $(n-1)!$. To prove the equality of $(1)$ and $(4)$, we compute the integrals on a given monomial $x^m := x_1^{m_1} \cdots x_{n-1}^{m_{n-1}} (1-x_1-\cdots-x_{n-1})^{m_n}$. First, by Lemma \ref{lem:bombieri} we have: \[ \int |v|^{2m} d\mu_{S_\mathbb{C}^n}(v) = \frac{m_1! \cdots m_n! (n-1)!}{(|m|+n-1)!} = \binom{|m|+n-1}{m_1,\ldots,m_n,n-1}^{-1}. \] Now, note: \[ \int_0^1 x^l (1-x)^m dx = \sum_{k=0}^m \binom{m}{k} (-1)^k \int_0^1 x^{l+k} dx = \frac{l!m!}{(l+m+1)!} \sum_{k=0}^m \frac{\prod_{j \neq k} (l+j+1)}{(-1)^k k! (m-k)!} = \frac{l!m!}{(l+m+1)!}. \] The last equality is due to Lagrange interpolation, considering the sum as a function of $n$. Further: \[ \frac{l!m!}{(l+m+1)!} = \binom{l+m+2-1}{l,m,2-1}^{-1}. \] That is, we have equality whenever $n=2$, proving the base case. The rest of the proof goes by induction. First we compute for $\alpha = 1-x_1-\cdots-x_{n-2}$: \[ \int_0^\alpha x_{n-1}^j (\alpha-x_{n-1})^k dx_{n-1} = \int_0^1 (\alpha u)^j (\alpha - \alpha u)^k \alpha du = \alpha^{j+k+1} \int_0^1 u^j (1-u)^k du = \frac{\alpha^{j+k+1} \cdot j!k!}{(j+k+1)!}. \] With this, we then compute the following by induction, letting $\beta = 1-x_1-\cdots-x_{n-3}$: \[ \begin{split} (n-1)! \int_0^1 &\cdots \int_0^\alpha x_1^{m_1} \cdots x_{n-1}^{m_{n-1}} (\alpha-x_{n-1})^{m_n} dx_{n-1} \cdots dx_1 \\ &= \frac{m_{n-1}!m_n!}{(m_{n-1}+m_n+1)!} \cdot (n-1)! \int_0^1 \cdots \int_0^{\alpha+x_{n-2}} x_1^{m_1} \cdots x_{n-2}^{m_{n-2}} \alpha^{m_{n-1}+m_n+1} dx_{n-2} \cdots dx_1 \\ &= \frac{m_{n-1}!m_n!(n-1)}{(m_{n-1}+m_n+1)!} \left[(n-2)! \int_0^1 \cdots \int_0^\beta x_1^{m_1} \cdots x_{n-2}^{m_{n-2}} (\beta-x_{n-2})^{m_{n-1}+m_n+1} dx_{n-2} \cdots dx_1\right] \\ &= \frac{m_{n-1}!m_n!(n-1)}{(m_{n-1}+m_n+1)!} \cdot \frac{m_1! \cdots m_{n-2}! (m_{n-1}+m_n+1)! (n-1-1)!}{(m_1+\cdots+m_n+1+n-1-1)!} \\ &= \frac{m_1! \cdots m_n! (n-1)!}{(m_1+\cdots+m_n+n-1)!}. \end{split} \] This completes the proof of equality of $(1)$ and $(4)$. \end{proof} \section{Groups, Manifolds, and a Brief History of Lie Theory} \label{sec:symmetries_revisited} The two examples of the previous sections make it clear that understanding the symmetries of a particular problem are crucial to understanding how that problem is solved. The optimization and sampling problems on the unitary group posed earlier were able to be simplified by reducing down to the discrete symmetric group. In this and the next sections, we observe and discuss the deeper connections between such continuous and discrete groups for the special class of Lie groups. By understanding these connections, we hope to convince the reader that the structure of the example problems discussed above goes much deeper than those examples might initially suggest. Before moving on, we first need to define formally the fundamental objects with which we will be working: groups and manifolds. A \textbf{group} $G$ is a set with a distinguished element $e \in G$ called the \textbf{identity}, equipped with a product operation $*$ and an inversion operation $\cdot^{-1}$. The product operation should have the property that $g * e = e * g = g$ for all $g \in G$, the inversion operation should have the property that $e^{-1} = e$, and together they should satisfy $g * g^{-1} = e$. Typical examples of groups are the symmetric group $S_n$ and the general linear group ${\rm GL}}\def\SL{{\rm SL}}\def\Sp{{\rm Sp}_n(\R)$ of invertible $n \times n$ real matrices. Informally, a \textbf{(differentiable) manifold} $M$ is a topological space which locally looks like $\R^n$ for some $n$ at every point. Formally, this means that in some open neighborhood $U_x$ about any point $x \in M$ there is a homeomorphism $\phi_x: U_x \to V_x \subset \R^n$, such that if $U_x \cap U_y \neq \varnothing$ then $\phi_x \circ \phi_y^{-1}: V_y \to V_x$ is a smooth map (see \cite{Lee00,VGeodesic} for more on manifolds). {Note that one can also replace $\R^n$ by $\mathbb{C}^n$ and ``smooth'' by ``holomorphic'' in this definition to obtain the notion of a \textbf{complex manifold}.} Given a group $G$ and a manifold $M$, the \textbf{action} of $G$ on $M$ is a function $\phi: G \times M \to M$ for which we write $(g,x)$ as $g \cdot x$. A \textbf{group action} must further satisfy the properties that $g \cdot (h \cdot x) = (gh) \cdot x$ and $e \cdot x = x$. An \textbf{orbit} of $x \in M$ is the set $\mathcal{O}_x := \{y \in M : y = g \cdot x \text{ for some } g \in G\}$. The \textbf{stabilizer} of $x \in M$ is the set $\Stab_x := \{g \in G : g \cdot x = x\}$. A \textbf{group representation} is a vector space $V$ that is acted upon by a group action of $G$. Another way to think of this is to consider a map $\rho$ from an element of the group $g \in G$ to an invertible linear transformation $\rho(g)$ of the vector space $V$. Enforcing $\rho$ to be a group homomorphism (i.e., for any $g_1,g_2 \in G$ we have $\rho(g_1 g_2)=\rho(g_1)\rho(g_2)$) implies the action $g \cdot v := \rho(g) v$ is a group action. The symmetries of an object are given by a group acting on that object. By finding properties that are preserved by these symmetries, we are finding coordinate-free and canonical properties of the underlying object. This allows us to get at the heart of the object without having to deal with the technicalities that often come with a particular choice of coordinates or parameterization of the object. Let us consider the simple example of the vector space $\R^n$. This vector space can be considered to have two continuous groups that act upon it, given by translations and rotations. For translation, we have that the additive group $(\R^n,+)$ acts on $\R^n$ via $(x_1,\ldots,x_n) \mapsto (x_1+c_1,\ldots,x_n+c_n)$ for any $c$ in the group $(\R^n,+)$. This action is abelian or commutative, since changing the order of mutliple translations does not change the effect of the action. For rotation, the group $\SO_n(\R)$ of orthogonal rotation matrices acts on $\R^n$ via $x \mapsto Ox$ for any $O \in \SO_n(\R)$. This action is \emph{not} abelian; different orderings of rotations of $\R^n$ can result in different actions. Given a subset $S \subset \R^n$, the translation and rotation actions preserve many important properties that one might want to know about $S$: volume, surface area, width, distances, and angles. The translation action also preserves some properties that rotation does not, for example the slope of a line. The rotation action similarly preserves other properties as well, like when a set $S$ is centrally symmetric. The key point is that these symmetries bring to the forefront various properties that we find important or interesting about a set $S$. The symmetries serve to remove the coincidental data regarding how a particular set is embedded in $\R^n$, leaving behind only the important information about the set. To see how such continuous symmetries play a role in the development of algorithms, we now turn to Lie theory. As discussed above, Lie theory is be used to generalize and shed light on the two example problems explored in the previous sections. And before discussing the basics of Lie theory in general, we first discuss a number of important properties of the unitary group $\U(n)$. The unitary group $\U(n)$ is essentially the complex vector space equivalent of the group $\SO(n)$ of rotations of a real vector space. The introduction of the complex numbers here serves to simplify the theory, and this is a common theme in Lie theory. This example of the unitary group $\U(n)$ serves as a running example for the remainder of our discussion. Before moving on, we mention a bit of the history of Lie theory. Lie theory is a vast subject that has been studied for around 150 years by some of the greatest mathematicians of that time period. It has had a close association with physics starting with the work of Emmy Noether \cite{Noether1918} and, till today, plays an important role in the search of fundamental laws of nature. The subject is named after Sophus Lie, who was formally initiated the study of infinitesimal group actions on a manifold. This gave rise to the Lie algebra of a Lie group, and Wilhelm Killing then extensively studied the problem of the classification of Lie algebras \cite{killing1889zusammensetzung}. Killing gave the correct classification, albeit with a few incomplete or incorrect proofs, and Elie Cartan gave a rigorous proof based on Killing's work (e.g., see \cite{chern1952elie}). For a more detailed account of the history of this classification and the emergence of Lie theory, most of which occurred in the 19th century, see \cite{hawkins2012emergence}. Later works of E. B. Dynkin \cite{dynkin1947structure} and Nathan Jacobson \cite{jacobson1979lie} made this classification and the surrounding results more accessible, and their results and exposition often serve as the basis of how Lie theory is taught today. While a classification of Lie algebras gave way to a number of results, mathematicians of the 20th century continued searching for more underlying structure which would enable proofs that did not rely on the classification. To this end, Hermann Weyl (see \cite{weyl1968gesammelte}), Claude Chevalley \cite{chevalley1948theorie}, Harish-Chandra \cite{harish1951some}, Hendrik Casimir and Bartel Leendert van der Waerden \cite{casimir1935algebraischer}, and Jean-Pierre Serre (see \cite{serre2012complex}) all played significant roles in the development of this theory, among countless others. Many of their findings are the underpinnings of the Lie theory that we present here. \section{The Unitary Group} \label{sec:unitary_group} Given $n \in \N$, the \textbf{unitary group} $\U(n)$ is the group of $n \times n$ unitary matrices under matrix multiplication. This is defined more explicitly by \[ \U(n) := \{U \in \mathbb{C}^{n \times n} : UU^* = I\}, \] where $\mathbb{C}^{n \times n}$ is the vector space of all complex $n \times n$ matrices, $U^*$ denotes the conjugate transpose of $U$, and $I$ is the identity matrix. Because this group in embedded in the vector space of matrices, it can be equipped with a real\footnote{Even though the unitary group is embedded in the complex vector space $\mathbb{C}^{n \times n}$, it cannot be equipped with a complex manifold structure. However, it can be equipped with a real manifold structure by viewing $\mathbb{C}^{n \times n} \cong (\R^{n \times n})^2$. This is similar to the fact that the space of Hermitian matrices is a real vector space, even though the matrices have complex entries.} manifold structure coming from the vector space. As a manifold, $\U(n)$ is compact. Further, the group operations of multiplication and inversion in $\U(n)$, given by $V \mapsto UV$ for fixed $U \in \U(n)$ and $V \mapsto V^{-1}$, are smooth maps with respect to the manifold structure. That is, the algebraic group structure and the analytic manifold structure are compatible. As a manifold, $\U(n)$ has a tangent space over every point $U \in \U(n)$. These tangent spaces are isomorphic as vector spaces, so let us in particular look at the tangent space over a canonical point, the identity matrix $I \in \U(n)$. We compute the elements of this tangent space in a standard way: by considering derivatives of paths on the manifold through the identity matrix. That is, fix any smooth function $f: (-\epsilon, \epsilon) \to \U(n)$ such that $(-\epsilon, \epsilon) \subset \R$ is a small real interval and $f(0) = I$. An element $X$ of the tangent space of $\U(n)$ at the identity, denoted $T_I \U(n)$, is given by \[ X = \left.\frac{d}{dt}\right|_{t=0} f(t), \] and in fact every element of the tangent space of $\U(n)$ can be constructed in this way. So far this construction has only used the manifold structure of $\U(n)$, but we now employ the group structure to get a better handle on what $X$ can actually be. Since $f(t) \in \U(n)$ for all $t$ near 0, we further have that $f(t) f(t)^* = I$ for all such $t$. By the derivative Leibniz rule, this implies \[ 0 = \left.\frac{d}{dt}\right|_{t=0} \left[f(t) f(t)^*\right] = \left[\left.\frac{d}{dt}\right|_{t=0} f(t)\right] f(0)^* + f(0) \left[\left.\frac{d}{dt}\right|_{t=0} f(t)\right]^* = X + X^*. \] In fact, this precisely describes the tangent space $T_I \U(n)$: \[ T_I \U(n) = \{X \in \mathbb{C}^{n \times n} : X^* = -X\}. \] That is, $T_I \U(n)$ is the (real) vector space of all $n \times n$ skew-Hermitian matrices. Beyond describing the tangent space, the group structure on $\U(n)$ also provides extra structure on the tangent space $T_I \U(n)$. For example, the conjugation action of $\U(n)$ on $\mathbb{C}^{n \times n}$ preserves the space of skew-Hermitian matrices. Specifically, for any $U \in \U(n)$ and any $X \in T_I \U(n)$ we have \[ (UXU^{-1})^* = U^{-*}X^*U^* = UX^*U^{-1} = -UXU^*, \] which precisely says that $UXU^{-1}$ is skew-Hermitian. Another interesting preserving ``action'' of $T_I \U(n)$ on itself is given by the matrix commutator, $[X,Y] := XY-YX$. For any $X,Y \in T_ \U(n)$, we have \[ (XY-YX)^* = Y^*X^*-X^*Y^* = (-Y)(-X)-(-X)(-Y) = -(XY-YX), \] which precisely says that $XY-YX$ is skew-Hermitian. And finally, a general vector space can be equipped with an inner product in any number of ways, but the tangent space $T_I \U(n)$ has a very special inner product $\langle X, Y \rangle := -\Tr(XY)$. (Note that the negation is required for positive definiteness, since $X,Y$ are \emph{skew}-Hermitian.) What makes this inner product so special is its connection to the above actions we described. Specifically, \[ \langle UXU^{-1}, UYU^{-1} \rangle = -\Tr(UXU^{-1}UYU^{-1}) = -\Tr(UXYU^{-1}) = -\Tr(XY) = \langle X, Y \rangle, \] and \[ \begin{split} \langle [X,Y], Z \rangle &= -\Tr((XY-YX)Z) = -\Tr(XYZ-YXZ) = \Tr(YXZ-XYZ) \\ &= \Tr(YXZ-YZX) = \Tr(Y(XZ-ZX)) = -\langle Y, [X,Z] \rangle. \end{split} \] That is, this inner product is \textbf{invariant} under these natural actions of conjugation and commutation. With this, we catch a glimpse of the symmetric nature of the unitary group $\U(n)$. The above discussion has made it clear that the tangent space of $\U(n)$ is more than just a vector space: The unitary group acts on it in a natural way, it acts on itself in a natural way, and these actions are very compatible with a natural inner product. The obvious next questions are as follows: What can we gain from all this structure, in terms of optimization and sampling? And how specific to the unitary group is all of this structure? \section{Lie Theory Basics}\label{sec:Lie} The structure of the unitary group and its tangent space is not specific to the unitary group, but in fact generalizes to \textbf{Lie groups} and their associated \textbf{Lie algebras}. While the features of the unitary group discussed above may have seemed somewhat ad hoc, we see how they can be derived from general principles in the context of Lie theory. After discussing the basics of Lie theory, we then utilize this theory and demonstrate its importance in the context of optimization and sampling. All of this basic material is standard, and can be found in standard references, e.g., \cite{baker2003matrix,knapp2013,hall2003lie}. Chapter 4 of \cite{knapp2013} is a particularly good reference for compact Lie groups. For a view more towards physics and engineering, see also \cite{gilmore2008lie}. As a final note, while we discuss a number of basic Lie theoretic topics, we do not discuss more advanced topics which would be considered to be crucial to a complete discussion of Lie theory. This includes Lie group and Lie algebra representations, roots, weight theory, Dynkin diagrams, nilpotent Lie algebras, the exponential map, and likely other topics. \subsection{Lie Groups} A \textbf{Lie group} $G$ is a group equipped with a manifold structure which is compatible with the group structure. Compatible here means that the operations of multiplication $g*h$ and inversion $g^{-1}$, given by \[ *: G \times G \to G \qquad \text{and} \qquad \cdot^{-1}: G \to G, \] are smooth (differentiable) maps between manifolds. The typical examples of Lie groups are \textbf{matrix groups}, which are groups of matrices equipped with the manifold structure coming from the fact that a matrix can be considered as a point in $\mathbb{C}^{n \times n}$. Some examples are: \begin{itemize} \item $\U(n)$, the group of $n \times n$ unitary matrices. \item ${\rm SU}}\def\U{{\rm U}}\def\SO{{\rm SO}}\def\O{{\rm O}}\def\USp{{\rm USp}(n)$, the group of $n \times n$ unitary matrices with determinant 1. \item ${\rm GL}}\def\SL{{\rm SL}}\def\Sp{{\rm Sp}_n(\mathbb{C})$ and ${\rm GL}}\def\SL{{\rm SL}}\def\Sp{{\rm Sp}_n(\R)$, the groups of invertible $n \times n$ matrices. \item ${\rm GL}}\def\SL{{\rm SL}}\def\Sp{{\rm Sp}(V)$, the group of invertible linear maps from the vector space $V$ to itself. \item $\SL_n(\mathbb{C})$ and $\SL_n(\R)$, the groups of invertible $n \times n$ matrices with determinant 1. \item $\SO_n(\mathbb{C})$ and $\SO_n(\R)$, the groups of orthogonal $n \times n$ matrices with determinant 1. \item $\Sp_{2n}(\mathbb{C})$, the group of $2n \times 2n$ symplectic matrices. \item $\USp(n)$, the group of $2n \times 2n$ symplectic unitary matrices. \end{itemize} Note that even though we often consider matrix groups contained in $\mathbb{C}^{n \times n}$, certain matrix groups may still be considered as either real Lie groups or complex Lie groups. In the case of a complex Lie group, $G$ is a complex manifold and the multiplication and inversion maps must actually be holomorphic. In the case of a real Lie group, we implicitly identify $\mathbb{C}^{n \times n} \cong (\R^{n \times n})^2$. Note that for some matrix groups $G$, like the unitary group $G = \U(n)$, we are forced to consider $G$ as a real Lie group even though elements of the group have complex entries. This situation occurs when there is no way to endow our Lie group $G$ with the structure of a complex manifold, and one important example of such groups are the compact Lie groups. This point is not very important to our exposition, but it is still worth making to avoid certain confusions. Given a Lie group $G$, one can define the notion of a \textbf{Lie subgroup}. A Lie subgroup $H$ of $G$ is a subgroup of $G$ which is embedded within $G$ as a submanifold. A theorem of Cartan \cite{cartan1952theorie} says that if $H$ is a closed subgroup of a real Lie group $G$, then $H$ is in fact a Lie subgroup of $G$. A wealth of examples of Lie subgroups come from the matrix Lie groups listed above. All of them are Lie subgroups of either ${\rm GL}}\def\SL{{\rm SL}}\def\Sp{{\rm Sp}_n(\mathbb{C})$ or ${\rm GL}}\def\SL{{\rm SL}}\def\Sp{{\rm Sp}_n(\R)$. Further, ${\rm SU}}\def\U{{\rm U}}\def\SO{{\rm SO}}\def\O{{\rm O}}\def\USp{{\rm USp}(n)$ is a Lie subgroup of $\U(n)$. Another important example is the \textbf{compact torus} $\U(1)^n$, where $\U(1)$ is isomorphic to the unit circle. There are a number of ways to view $\U(1)^n$ as a Lie subgroup of $\U(n)$, but the typical way of embedding of $\U(1)^n$ into $\U(n)$ is as the set of all diagonal matrices in $\U(n)$. \subsection{The Tangent Space at the Identity Element} Since a Lie group $G$ is a manifold, it has a tangent vector space at every point. Moreover, since these tangent spaces are all isomorphic as vector spaces, we restrict to a specific canonical tangent space; that is, we restrict to the tangent space of the identity element of the group, denoted $T_e G$. Using only the manifold structure of a Lie group $G$, there is a standard way to characterize the tangent space of $G$ at a given point, and we now describe this for $T_e G$. This is precisely what we did for the group $\U(n)$ above. Let $f: (-\epsilon,\epsilon) \to G$ be an injective smooth map from a small real interval about 0 to the Lie group $G$ such that $f(0) = e$, the identity element. Such an $f$ is equivalently defined as a short path in $G$ which passes through the identity element at $t=0$. The expression \[ X = \left.\frac{d}{dt}\right|_{t=0} f(t) \] then gives rise to an element $X$ of the tangent space $T_e G$ according to the direction of the path at the identity. Further, every element of the tangent space can be formed in this way, by choosing an appropriate smooth path through the identity element of $G$. \begin{example}[${\rm GL}}\def\SL{{\rm SL}}\def\Sp{{\rm Sp}_n(\mathbb{C})$] \label{ex:GL_n} Beyond the unitary group $\U(n)$ above, the simplest example is ${\rm GL}}\def\SL{{\rm SL}}\def\Sp{{\rm Sp}_n(\mathbb{C})$ (and the same argument works for ${\rm GL}}\def\SL{{\rm SL}}\def\Sp{{\rm Sp}_n(\R)$). % Since this Lie group is embedded in the space of $n \times n$ complex matrices $\mathbb{C}^{n \times n}$, the elements of the tangent space $T_I {\rm GL}}\def\SL{{\rm SL}}\def\Sp{{\rm Sp}_n(\mathbb{C})$ are $n \times n$ complex matrices. % Since the determinant of a matrix is continuous in its entries, it is straightforward to see that ${\rm GL}}\def\SL{{\rm SL}}\def\Sp{{\rm Sp}_n(\mathbb{C})$ is an $n^2$-dimensional complex manifold. % Thus the tangent space $T_I {\rm GL}}\def\SL{{\rm SL}}\def\Sp{{\rm Sp}_n(\mathbb{C})$ is an $n^2$-dimensional subspace of the vector space of all complex $n \times n$ matrices, and therefore $T_I {\rm GL}}\def\SL{{\rm SL}}\def\Sp{{\rm Sp}_n(\mathbb{C})$ must in fact be equal to the whole vector space of such matrices. \end{example} \subsection{The (Group) Adjoint Action} We now incorporate the group structure of $G$ by defining an action of the group $G$ on the tangent space $T_e G$. Given $X \in T_e G$, let $f: (-\epsilon, \epsilon) \to G$ be a short path in $G$ such that $f(0) = e$ and $\left.\frac{d}{dt}\right|_{t=0} f(t) = X$. We now act on $X$ by $g \in G$ via the \textbf{(group) adjoint action}, given by \[ \Ad_g(X) := \left.\frac{d}{dt}\right|_{t=0} \left[g \cdot f(t) \cdot g^{-1}\right]. \] Note that by the compatibility properties of the group and manifold structures of $G$, we have that $g \cdot f(t) \cdot g^{-1}$ is a smooth path such that $\left[g \cdot f(t) \cdot g^{-1}\right](0) = g \cdot e \cdot g^{-1} = e$. Therefore the above expression for $\Ad_g(X)$ gives a well-defined element of the tangent space $T_e G$. It is then technical but straightforward that this action does not depend on the choice of $f$, and in turn that this action gives rise to a well-defined group action on $T_e G$. In the case of matrix Lie groups, such as those listed above, there is a straightforward way to define this (group) adjoint action. This is given by matrix conjugation, as described in the case of the unitary group in Section \ref{sec:unitary_group}. Specifically, if $G$ is a matrix group, then the tangent space $T_e G$ can be considered as a vector space of matrices. The (group) adjoint action of $G$ on $T_e G$ is given by \[ \Ad_g: X \mapsto gXg^{-1} \] for all $g \in G$ and $X \in T_e G$. For essentially everything we discuss here, the reader can keep this conjugation action in mind for the (group) adjoint action of a Lie group $G$ on its tangent space $T_e G$. \begin{remark}[Adjoint orbits] \label{rem:adjoint_orbits_KKS} The (group) adjoint action of $G$ on its tangent space $T_e G$ is a group action, and thus this action partitions the vector space $T_e G$ into orbits, referred to as the \textbf{adjoint orbits} of $G$. % Given $X \in T_e G$, we denote the adjoint orbit of $X$ by $\mathcal{O}_X := \{\Ad_g(X) : g \in G\}$. In the case of the unitary group, the adjoint orbits are determined precisely by the eigenvalues of the matrices in $T_I \U(n)$. That is, the adjoint orbits of $\U(n)$ can be defined via \[ \mathcal{O}_X =\{Y \in T_I \U(n) : \eig(Y) = \eig(X)\}. \] For compact groups $G$, we note that such adjoint orbits can be given the structure of a symplectic manifold via the \textbf{Kirillov-Kostant-Souriau (KKS) symplectic form}. % We do not discuss this further here, but instead direct the interested reader to \cite{kirillov2004lectures}. % We also note that this KKS form typically refers to \emph{co}adjoint orbits rather than adjoint orbits (as in \cite{kirillov2004lectures}), but these are essentially equivalent in our case, see Remark \ref{rem:coadjoint}. \end{remark} \subsection{The (Algebra) Adjoint Action} \label{sec:algebra_adjoint_action} When considering the unitary group, we constructed a certain action of the tangent space on itself, given by the matrix commutator. It is easy to see that this operation is bilinear, and as we saw in Section \ref{sec:unitary_group}, it preserves the tangent space $T_I \U(n)$ of the unitary group. Because of this, the operation of matrix conjugation can be thought of as a product-like operation on $T_I \U(n)$. In a general Lie group, this action, called the \textbf{(algebra) adjoint action}, can be constructed from the group adjoint action discussed in the previous section. To do this, we take the derivative of the action of $\Ad_g$. More formally, the action of $G$ on $T_e G$ can be thought of as a group homomorphism, given by \[ \phi: G \to {\rm GL}}\def\SL{{\rm SL}}\def\Sp{{\rm Sp}(T_e G) \qquad \text{via} \qquad \phi: g \mapsto \Ad_g. \] The group ${\rm GL}}\def\SL{{\rm SL}}\def\Sp{{\rm Sp}(T_e G)$ is in fact a Lie group, and thus a manifold, and the above map $\phi$ is a smooth group homomorphism. Thus we can consider the differential of the map $\phi$. By definition, this differential gives rise to linear maps from the tangent spaces of $G$ to the tangent spaces of ${\rm GL}}\def\SL{{\rm SL}}\def\Sp{{\rm Sp}(T_e G)$. When restricted to the tangent space $T_e G$, this gives rise to a map \[ d\phi_e: T_e G \to T_I \left[{\rm GL}}\def\SL{{\rm SL}}\def\Sp{{\rm Sp}(T_e G)\right] \qquad \text{via} \qquad d\phi_e: X \mapsto \ad_X. \] That is, we have defined $\ad_X$ to be the element of $T_I \left[{\rm GL}}\def\SL{{\rm SL}}\def\Sp{{\rm Sp}(T_e G)\right]$ given by $d\phi_e(X)$ for $X \in T_e G$. It not yet clear precisely what this $\ad_X$ is, and we describe this further now. We have already described $T_I \left[{\rm GL}}\def\SL{{\rm SL}}\def\Sp{{\rm Sp}(T_e G)\right]$ as the vector space of all (not necessarily invertible) linear maps from $T_e G$ to itself, see Example \ref{ex:GL_n}. Thus $\ad_X$ is a linear map on the tangent space $T_e G$ of $G$. We then refer to this as the \textbf{(algebra) adjoint action} of $T_e G$ on itself. In the case that $G$ is a matrix group, we now compute explicitly this (algebra) adjoint action. Given a short path $f: (-\epsilon,\epsilon) \to G$ such that $f(0) = I$ and $\left.\frac{d}{dt}\right|_{t=0} f(t) = X$, and any $Y \in T_I G$, the differential is defined via \[ \begin{split} d\phi_I\left(\left.\frac{d}{dt}\right|_{t=0} f(t)\right) Y &= \left.\frac{d}{dt}\right|_{t=0} \phi(f(t)) Y = \left.\frac{d}{dt}\right|_{t=0} \left[f(t) Y f(t)^{-1}\right] \\ &= \left(\left.\frac{d}{dt}\right|_{t=0} f(t)\right) Y f(0)^{-1} + f(0) Y \left(\left.\frac{d}{dt}\right|_{t=0} f(t)^{-1}\right) \\ &= XY + Y\left(-f(0)^{-2} \left.\frac{d}{dt}\right|_{t=0} f(t)\right) = XY - YX. \end{split} \] That is, $\ad_X(Y) = XY-YX$ is the matrix commutator for all matrix groups $G$, including the unitary group. \subsection{Lie Algebras} \label{sec:lie_alg_via_group} With the algebra adjoint action in hand, we are ready to define the Lie algebra of a Lie group. Given a Lie group $G$, its \textbf{Lie algebra}, denoted $\mathfrak{g}$, is the tangent vector space $T_e G$ equipped with a product-like operation called the \textbf{Lie bracket}. The Lie bracket on $\mathfrak{g}$ is defined via \[ [\cdot, \cdot]: \mathfrak{g} \times \mathfrak{g} \to \mathfrak{g} \qquad \text{via} \qquad [X,Y] := \ad_X(Y). \] The Lie bracket has a few important properties: \begin{enumerate} \item \textbf{Bilinearity:} $[aX+bY,Z] = a[X,Z] + b[Y,Z]$ and $[X,bY+cZ] = b[X,Y] = c[X,Z]$. \item \textbf{Anti-symmetry:} $[X,Y] = -[Y,X]$. \item \textbf{Jacobi identity:} $[X,[Y,Z]] + [Y,[Z,X]] + [Z,[X,Y]] = 0$. \end{enumerate} While the Lie bracket is referred to as a product-like operation, it is not a typical product because it is not associative. However, one can rewrite the Jacobi identity to view it as an ``associativity error'' formula as follows: \[ [[X,Y],Z] - [X,[Y,Z]] = [Y,[Z,X]]. \] In the case of Lie algebras associated to matrix groups, the Lie bracket is given by the commutator $[X,Y] = XY-YX$ as discussed above. It is straightforward to see that the matrix commutator satisfies these three properties. As with Lie subgroups of Lie groups, there is a notion of a \textbf{Lie subalgebra} of a Lie algebra. A Lie subalgebra $\mathfrak{h} \subset \mathfrak{g}$ is a vector subspace with Lie bracket inherited from $\mathfrak{g}$ such that $[X,Y] \in \mathfrak{h}$ for all $X,Y \in \mathfrak{h}$. According to the Lie group-Lie algebra correspondence, every such Lie subalgebra $\mathfrak{h} \subset \mathfrak{g}$ is the Lie algebra of a Lie subgroup $H \subset G$. There are various categories that different Lie algebras can be put into, but the most important ones for us are \textbf{compact Lie algebra}, \textbf{abelian Lie algebra}, and \textbf{semisimple Lie algebra}. A compact Lie algebra is a Lie algebra associated to a compact Lie group. For example, the Lie algebra $\mathfrak{u}_n$ associated to the Lie group $\U(n)$ is a compact Lie algebra. An abelian Lie algebra is a Lie algebra with trivial Lie bracket: $[X,Y] = 0$ for all $X,Y \in \mathfrak{g}$. A semisimple Lie algebra is a Lie algebra with trivial \textbf{center}; i.e., for which there is no nonzero $X \in \mathfrak{g}$ such that $[X,Y] = 0$ for all $Y \in \mathfrak{g}$. Note that $\mathfrak{su}_n$ is semisimple, but $\mathfrak{u}_n$ is not because the matrix $i \cdot I$ is skew-Hermitian and commutes with all matrices in $\mathfrak{u}_n$. \begin{example}[Lie algebra of the unitary group] In the case of the unitary group $\U(n)$, the associated Lie algebra $\mathfrak{u}_n$ is the (real) vector space of all skew-Hermitian matrices, as discussed in Section \ref{sec:unitary_group}. As stated above, the Lie bracket of $\mathfrak{u}_n$ is the commutator since $\U(n)$ is a matrix group. To see that this is well-defined, we compute the following for any skew-Hermitian matrices $X,Y$: \[ [X,Y]^* = (XY-YX)^* = Y^*X^* - X^*Y^* = -XY + YX = -[X,Y]. \] That is, we have that $[X,Y]$ is itself skew-Hermitian, and thus $[\cdot,\cdot]: \mathfrak{u}_n \times \mathfrak{u}_n \to \mathfrak{u}_n$ is well-defined. \end{example} \begin{example}[Lie algebra of the compact torus] The subspace of diagonal matrices in $\mathfrak{u}_n$, all of which have purely imaginary diagonal entries, is an abelian Lie subalgebra of the compact Lie algebra $\mathfrak{u}_n$. This Lie subalgebra is then the Lie algebra of the compact abelian subgroup $\U(1)^n \subset \U(n)$, given by the diagonal matrices in $\U(n)$. \end{example} Although we have constructed Lie algebras here via the tangent spaces of Lie groups, one can also define an \textbf{abstract Lie algebra} as any vector space with a Lie bracket operation which satisfies the above three properties. Such a Lie algebra is not a priori associated to a Lie group, but a classical classification theorem shows that nice Lie algebras defined in this way can be associated to matrix Lie groups. We discuss this classification further in Section \ref{sec:classification}. \section{Lie Theory for Optimizing Over Orbits} Having discussed the basics of Lie theory, we were able to answer one of our questions; i.e., the structure we found in the unitary group and its tangent space is not at all specific to the unitary group. Given a Lie group $G$ in general, one can construct the associated Lie algebra $\mathfrak{g}$ using the tangent space at the identity, and the conjugation action and commutator of the unitary group $\U(n)$ and its Lie algebra $\mathfrak{u}_n$ generalize to the adjoint actions $\Ad_g$ and $\ad_X$ and the Lie bracket $[\cdot, \cdot]$ of any $G$ and $\mathfrak{g}$. This leaves the other question: What can we gain from all this structure, in terms of optimization and sampling? To answer this question, we first discuss a simple optimization problem that one can associate to the adjoint orbits of a Lie group $G$. To get a sense of this optimization problem, we first consider a related problem for the unitary group. Let $\mathcal{H}(n)$ denote the (real) vector space of all $n \times n$ Hermitian matrices, let $Z \in \mathcal{H}(n)$ be any real diagonal matrix, let $X = e_1 e_1^\top \in \mathcal{H}(n)$, and let $\mathcal{O}_X$ denote the conjugation orbit of $X$ with respect to the group $\U(n)$. We consider the optimization problem \[ \min_{Y \in \mathcal{O}_X} \langle Y, Z \rangle_F, \] where $\langle Y, Z \rangle_F = \Tr(YZ)$ is the Frobenius inner product on Hermitian matrices. Another way to write this is as \[ \min_{Y \in \mathcal{O}_X} \langle Y, Z \rangle_F = \min_{\|v\|_2 = 1} v^* Z v. \] It is well-known that this problem is equivalent to computing the minimum eigenvalue of $Z$, which is precisely the problem studied in Section \ref{sec:eigenvalue} above. Since $Z$ is diagonal, this problem is further equivalent to computing the minimum entry of the diagonal of $Z$. This can be written as \[ \min_{Y \in \mathcal{O}_X} \langle Y, Z \rangle_F = \min_{\sigma \in S_n} (\sigma \cdot e_1)^\top Z (\sigma \cdot e_1), \] where $S_n$ is the symmetric group and $\sigma \cdot v$ acts on the vector $v$ by permuting its entries. The next question then is: Why study further such a simple optimization problem? What we see in the rest of this section is that optimization problem, while simple, has a Lie-theoretic interpretation which demonstrates an interesting connection between the nonconvex adjoint orbits of a Lie group and certain associated convex polytopes. To do this, we need to generalize some of the notions used in the above simple optimization problem: specifically, the Frobenius inner product, the notion of a diagonal matrix, and the symmetric group $S_n$. \subsection{The Killing Form} In the above optimization problem on the unitary group, we considered the conjugation action of $\U(n)$ on $\mathcal{H}(n)$, the space of Hermitian matrices equipped with a particular inner product. To generalize this to other Lie groups beyond $\U(n)$, an easy observation is that $\mathcal{H}(n) = i \cdot \mathfrak{u}_n$, and thus up to this scaling by $i$ we can view $\mathcal{H}(n)$ as essentially equal to the Lie algebra $\mathfrak{u}_n$ of $\U(n)$. The next question is: What then is the Lie-theoretic equivalent of the Frobenius inner product? To answer this, let us recall two properties of the Frobenius inner product which we discussed above. Specifically, for any $X,Y \in \mathfrak{u}_n$ and any $U \in \U(n)$, we have \[ \langle \Ad_U(X), \Ad_U(Y) \rangle_F = \Tr(UXU^{-1}UYU^{-1}) = \Tr(UXYU^{-1}) = \Tr(XY) = \langle X, Y \rangle_F \] and \[ \begin{split} \langle \ad_X(Y), Z \rangle_F &= \Tr((XY-YX)Z) = \Tr(XYZ-YXZ) = \Tr(YXZ-XYZ) \\ &= -\Tr(YXZ-YZX) = -\Tr(Y(XZ-ZX)) = -\langle Y, \ad_X(Z) \rangle_F. \end{split} \] That is, the Frobenius inner product is invariant under the group and algebra adjoint actions of $\U(n)$ and $\mathfrak{u}_n$ on $\mathfrak{u}_n$. Note that the Frobenius ``inner product'' is actually \emph{negative} definite when applied to $\mathfrak{u}_n$ since $\mathfrak{u}_n$ consists of \emph{skew}-Hermitian matrices. How can we generalize this to other Lie groups and algebras? In the general case, a Lie algebra $\mathfrak{g}$ is not necessarily a vector space of matrices, and so we cannot apply the trace. However, we know that $\ad_X$ is a linear operator on $\mathfrak{g}$, and thus can be considered as a matrix according to some chosen basis. With this, we can define a Frobenius-like symmetric bilinear form $B$ on $\mathfrak{g}$ in general, via: \[ B(X, Y) := \Tr(\ad_X \circ \ad_Y). \] We now attempt to prove some invariance properties similar to that of the Frobenius inner product. First, let us determine how $\Ad_g$ and $\ad_Z$ act on the linear map $\ad_X$. First we have \[ \begin{split} \ad_{\Ad_g(X)}(Y) &= [gXg^{-1}, Y] = gXg^{-1}Y - YgXg^{-1} \\ &= g(Xg^{-1}Yg - g^{-1}YgX)g^{-1} = [\Ad_g \circ \ad_X \circ \Ad_g^{-1}](Y), \end{split} \] which says that $\ad_{\Ad_g(X)} = \Ad_g \circ \ad_X \circ \Ad_g^{-1}$. And second we have \[ \ad_{\ad_Z(X)}(Y) = [[Z,X],Y] = [Z,[X,Y]] - [X,[Z,Y]] = [\ad_Z \circ \ad_X - \ad_X \circ \ad_Z](Y), \] which says that $\ad_{\ad_Z(X)} = \ad_Z \circ \ad_X - \ad_X \circ \ad_Z$, the matrix commutator of $\ad_Z$ and $\ad_X$. This says that $\Ad_g$ and $\ad_Z$ act on $\ad_X$ via the adjoint actions of ${\rm GL}}\def\SL{{\rm SL}}\def\Sp{{\rm Sp}_n$ and $\mathfrak{gl}_n$ on $\ad_X \in \mathfrak{gl}_n$. The same arguments as in the case of $\U(n)$ and $\mathfrak{u}_n$ then imply the desired invariance properties. Specifically, we have \[ B(\Ad_g(X), \Ad_g(Y)) = B(X,Y) \qquad \text{and} \qquad B(\ad_Z(X), Y) = -B(X, \ad_Z(Y)). \] This symmetric bilinear form is called the \textbf{Killing form}. For semisimple matrix groups, this Killing form can be explicitly expressed with reference to the Frobenius inner product, via \[ B(X, Y) = C_{\mathfrak{g}} \cdot \Tr(X, Y), \] where $C_{\mathfrak{g}}$ is some positive constant which depends on the Lie algebra $\mathfrak{g}$ being considered. \begin{remark} The expression $B(X, Y) = C_{\mathfrak{g}} \cdot \Tr(X, Y)$ holds for the semisimple Lie algebras $\mathfrak{su}_n$, $\mathfrak{sl}_n$, $\mathfrak{so}_n$, and $\mathfrak{sp}_{2n}$, among others. These expressions do not quite hold for $\mathfrak{u}_n$ and $\mathfrak{gl}_n$ because they have nontrivial center. % In fact, the Killing form is identically zero on the center of $\mathfrak{g}$, but this point is not be very important to our discussion, see Remark \ref{rem:Killing_u_n}. \end{remark} As we saw with $\mathfrak{u}_n$, this bilinear form is not an inner product on $\mathfrak{g}$. However, for a compact Lie algebra the Killing form is always negative semidefinite. And further, for a compact semisimple Lie algebra the Killing form is negative definite, and thus the negation of the Killing form gives rise to an $\Ad_g$-invariant and $\ad_Z$-invariant inner product on $\mathfrak{g}$ in this case. \subsection{Maximal Tori and Cartan Subalgebras} The Killing form defined above gives us a way to generalize the Frobenius inner product with its invariance properties to other Lie algebras. We next generalize the notion of diagonal matrix to a general Lie algebra. Given a Lie algebra $\mathfrak{g}$, there exist certain subalgebras $\mathfrak{t} \subset \mathfrak{g}$ called Cartan subalgebras. While such subalgebras are typically defined for any Lie algebra, we only define them for compact Lie algebras here. \begin{definition}[Cartan subalgebra] Given a compact Lie algebra $\mathfrak{g}$, a \textbf{Cartan subalgebra} or \textbf{maximal abelian subalgebra} $\mathfrak{t} \subset \mathfrak{g}$ is an abelian subalgebra of $\mathfrak{g}$ which is not properly contained in any other abelian subalgebra of $\mathfrak{g}$. \end{definition} In the literature, the term Cartan subalgebra is typically used in reference to complex Lie groups, whereas often the term \textbf{maximal abelian subalgebra} is used in reference to compact Lie groups. We also note here a standard result regarding Cartan subalgebras of compact Lie algebras. \begin{proposition}[Prop. 4.30 of \cite{knapp2013}] \label{prop:Cartan_torus} If $G$ is a compact group with compact Lie algebra $\mathfrak{g}$, then $\mathfrak{t} \subset \mathfrak{g}$ is a Cartan subalgebra if and only if $\mathfrak{t}$ is the Lie subalgebra associated to a maximal torus $T \subset G$. \end{proposition} To get a more concrete feel for the Cartan subalgebra, let us consider $\mathfrak{t} \subset \mathfrak{u}_n$ to be the set of all diagonal matrices in $\mathfrak{u}_n$. For all $X,Y \in \mathfrak{t}$, we have that \[ [X,Y] = XY-YX = 0 \] since $X$ and $Y$ are diagonal matrices. This implies $\mathfrak{t}$ is an abelian subalgebra of $\mathfrak{u}_n$. Further, it is straightforward to see that for all $Z \in \mathfrak{u}_n$ which is not diagonal, there is some diagonal $X \in \mathfrak{u}_n$ for which $[X,Z] \neq 0$. This implies $\mathfrak{t}$ is in fact a maximal abelian subalgebra of $\mathfrak{u}_n$. The Lie subgroup $T \subset \U(n)$ associated to $\mathfrak{t}$ as described by Proposition \ref{prop:Cartan_torus} above is then the compact torus of all diagonal matrices in $\U(n)$. \begin{remark} The usual purpose of a Cartan subalgebra $\mathfrak{h}$ of a Lie algebra $\mathfrak{g}$ in the complex semisimple case is for the classification of semisimple Lie algebras, along with their representations and the representations of their associated Lie groups. % The notion of a group representation was defined in Section \ref{sec:symmetries_revisited}, and a Lie algebra representation can be defined via the differential of a group representation. % The main point is that the linear operators $\ad_H$ for $H \in \mathfrak{h}$ are all simultaneously diagonalizable. % From this, one can construct the \textbf{root system} of the Lie algebra $\mathfrak{g}$, which is a combinatorial construction which helps with Lie algebra classification and which we do not describe any further here. % When considering a representation $V$ of a given Lie algebra $\mathfrak{g}$, the action of the Cartan subalgebra $\mathfrak{h}$ via linear operators on $V$ is also simultaneously diagonalizable. % The associated eigenvalues can then be used to determine the decomposition of $V$ into irreducible subrepresentations via \textbf{highest weight theory}. % We do not explore any of this further here, but the interested reader can find detailed explanations in any standard reference, e.g., \cite{baker2003matrix,knapp2013,hall2003lie}. \end{remark} \subsection{The Weyl Group}\label{sec:Weyl} The Killing form defined above gives us a way to generalize the Frobenius inner product with its invariance properties to other Lie algebras, and the Cartan subalgebras of the previous section generalize the notion of diagonal matrices. The last thing we need to discuss and generalize is the symmetric group that appeared in our simple optimization problem above. Associated to every semisimple or compact Lie group $G$ is a particular finite group $W$ called its Weyl group, and in the case of $\U(n)$, this is precisely the symmetric group $S_n$. For us, the key intuition behind this notion is that the Weyl group connects the continuous symmetries of a Lie group $G$ and its Lie algebra $\mathfrak{g}$ to underlying and hidden discrete finite symmetries of $G$ and $\mathfrak{g}$. This is made very clear when we reinterpret our simple optimization problem in the language of Lie theory in the next section. There are different ways to define the Weyl group of a given Lie group, but we describe just one here which specifically pertains to connected compact Lie groups. \begin{definition}[Weyl group] Let $G$ be a connected compact Lie group, and let $T \subset G$ be a maximal torus in $G$. Let the normalizer of $T$ in $G$ be defined as usual, via $N(T) := \{g \in G : gtg^{-1} \in T, ~ \forall t \in T\}$. % The \textbf{Weyl group} of $G$ is defined to be the quotient group $W := N(T)/T$. \end{definition} Although it is not immediately clear from the definition, the definition of the Weyl group is independent of the choice of maximal torus $T \subset G$, and additionally the Weyl group is always a finite group. To give intuition for the fact that $W$ is finite, we explicitly construct it in the case of the unitary group as follows. First, we let $T \subset \U(n)$ be the subgroup of all diagonal unitary matrices; we know this is a maximal torus in $\U(n)$ for the discussion of the previous section. The normalizer $N(T)$ of $T$ is then equal to the set of matrices of the form $PD$, where $P$ is a permutation matrix and $D$ is a diagonal unitary matrix. From this description, it is straightforward to see that $N(T)/T$ is the set of permutation matrices, which is isomorphic to the symmetric group $S_n$. \subsection{The Kostant Convexity Theorem} We now have all the Lie-theoretically generalized pieces from our simple optimization problem, and we can now rewrite it in these generalized terms. Let $G$ be a connected compact Lie group with compact semisimple Lie algebra $\mathfrak{g}$, let $T \subset G$ be a maximal torus with associated Lie algebra $\mathfrak{t}$, and let $B$ be the Killing form of $\mathfrak{g}$. Our optimization problem can then be generalized as follows. Given $X,Z \in \mathfrak{t}$ with $\mathcal{O}_X$ being the adjoint orbit of $X$ in $\mathfrak{g}$, we want to optimize: \[ \min_{Y \in \mathcal{O}_X} -B(X, Z). \] Recall that the Killing form $B$ is negative definite since $G$ is compact semisimple. But how can we solve this optimization problem? In the original example, the explicitness of the formulation gave us a few simple answers to this question, via the diagonal entries and the eigenvalues. In order to generalize these techniques to the above Lie-theoretic version of the problem, we need the following theorem which demonstrates the hidden convexity of the adjoint orbits of the Lie group $G$. \begin{theorem}[Kostant convexity theorem, \cite{Kostant1973}; see also \cite{ziegler1992kostant}]\label{thm:Kostant} Let $G$ be a connected compact Lie group with compact semisimple Lie algebra $\mathfrak{g}$. % Let $T \subset G$ be a maximal torus with associated Lie algebra $\mathfrak{t}$. % And let $\pi: \mathfrak{g} \to \mathfrak{t}$ be the orthogonal projection induced by the Killing form $B$ (which is negative definite since $\mathfrak{g}$ is semisimple). % Then for every $X \in \mathfrak{t}$, the (group) adjoint orbit $\mathcal{O}_X \subset \mathfrak{g}$ of $G$ intersects $\mathfrak{t}$ in a Weyl group orbit $W \cdot X$, and $\pi(\mathcal{O}_X) = \hull(W \cdot X)$. \end{theorem} \noindent We now prove one direction of the Kostant convexity theorem for the unitary group. The other direction relies on the weight theory of Lie algebras which we have not discussed in detail here. A short proof using weight theory can be found in \cite{ziegler1992kostant}. Note that although technically $\mathfrak{u}_n$ is not semisimple, we can still prove the result by using the slightly tweaked Killing form given by $\Tr(XY)$. (See Remark \ref{rem:Killing_u_n} for more discussion on the Killing form of $\mathfrak{u}_n$.) Although this proof is specifically for the unitary group and its Lie algebra, it captures the essence of the general proof for compact Lie groups and their Lie algebras. \begin{corollary}[Kostant convexity theorem for $\U(n)$] \label{cor:kostant_unitary} Let $\mathfrak{u}_n$ be the Lie algebra of the Lie group $\U(n)$, let $\mathfrak{t} \subset \mathfrak{u}_n$ and $T \subset \U(n)$ be the respective subsets of diagonal matrices, and let $\pi: \mathfrak{u}_n \to \mathfrak{t}$ be the orthogonal projection induced by the inner product $-\Tr(XY)$. Then for every $X \in \mathfrak{t}$, the (group) adjoint orbit $\mathcal{O}_X \subset \mathfrak{u}_n$ of $\U(n)$ intersects $\mathfrak{t}$ in a Weyl group orbit $S_n \cdot X$, and $\pi(\mathcal{O}_X) \subseteq \hull(S_n \cdot X)$. \end{corollary} \begin{proof} First we show that $\mathcal{O}_X \subset \mathfrak{u}_n$ of $\U(n)$ intersects $\mathfrak{t}$ in a Weyl group orbit $S_n \cdot X$. % This follows from the fact that the diagonal matrices in the orbit $\mathcal{O}_X$ are precisely those diagonal matrices with the eigenvalues of $X$ along the diagonal. % Since $X$ is an element $\mathfrak{t}$ and thus is itself diagonal, we have that $\mathcal{O}_X \cap \mathfrak{t} = S_n \cdot X$. Since $S_n \cdot X \subset \mathcal{O}_X$, we further have that the extreme points of $\hull(S_n \cdot X)$ are contained in $\mathcal{O}_X$ and also in $\pi(\mathcal{O}_X)$, since $\pi$ acts by sending all off-diagonal entries to 0. % The statement $\pi(\mathcal{O}_X) \subseteq \hull(W \cdot X)$ is thus equivalent to saying that every linear functional $\ell(Y)$ on $\pi(\mathcal{O}_X) \subset \mathfrak{t}$ is minimized at a point of $W \cdot X$. % Further, any linear functional on $\mathfrak{t}$ can be written as $\ell(Y) := \Tr(YL)$ for some $L \in \mathfrak{t}$, and for such $L$ we also have that $\Tr(YL) = \Tr(\pi(Y)L)$ for all $Y \in \mathcal{O}_X$. Thus it suffices to show that $\ell(Y) := \Tr(YL)$ is minimized over $\mathcal{O}_X$ at a point of $S_n \cdot X$ for all $L \in \mathfrak{t}$. % In fact, it suffices to show this for all $L \in \mathfrak{t}$ with distinct diagonal entries, by limiting. To show this, let us now consider $\mathcal{O}_X$ as a manifold acted upon by $\U(n)$ via the (group) adjoint action. % The tangent space of $\mathcal{O}_X$ at a point $Y$ can then be determined by considering the differential of the group adjoint action, which as we saw above gives the algebra adjoint action (see Section \ref{sec:algebra_adjoint_action}). % That is, for any $Y \in \mathcal{O}_X$ the tangent space of $\mathcal{O}_X$ at $Y$ is given by \[ T_Y \mathcal{O}_X := \{\ad_Z(Y) = [Z,Y] : Z \in \mathfrak{u}_n\}. \] Further, $Y$ is a critical point of $\ell(Y) = \Tr(YL)$ in $\mathcal{O}_X$ if and only if for all paths $f: (-\epsilon,\epsilon) \to \mathcal{O}_X$ such that $f(0) = Y$ we have $\left.\frac{d}{dt}\right|_{t=0} \ell(f(t)) = 0$. Since $\ell$ is linear, we have that \[ \left.\frac{d}{dt}\right|_{t=0} \ell(f(t)) = \ell\left(\left.\frac{d}{dt}\right|_{t=0} f(t)\right) = \ell(\ad_Z(Y)) \] for some $\ad_Z(Y) \in T_Y \mathcal{O}_X$, and every element of $T_Y \mathcal{O}_X$ is the derivative of a path $f$ at 0 in this way. % Thus, $Y$ is a critical point of $\ell(Y) = \Tr(YL)$ in $\mathcal{O}_X$ if and only if \[ 0 = \ell(\ad_Z(Y)) = \Tr(\ad_Z(Y)L) = -\Tr(\ad_Y(Z)L) = \Tr(Z\ad_Y(L)) \] for all $Z \in \mathfrak{u}_n$. % Since $-\Tr(XY)$ is an inner product on $\mathfrak{u}_n$, this implies $\ad_Y(L) = YL-LY = 0$. % By assuming that $L \in \mathfrak{t}$ has distinct diagonal entries as discussed above, this further implies that $Y$ must be diagonal by a straightforward computation. % Therefore, $Y \in \mathcal{O}_X \cap \mathfrak{t} = S_n \cdot X$, and this completes the proof. \end{proof} \noindent The key takeaway of the Kostant convexity theorem is the fact that the image of a nonconvex adjoint orbit of $\mathfrak{g}$ under a linear projection map is a convex polytope. And a key corollary to this is the fact that linear optimization over the orbitope generated by the adjoint orbit $\mathcal{O}_X$ is equivalent to linear optimization over an associated convex polytope. An \textbf{orbitope} is defined to be the convex hull of an orbit of some group action on a vector space, and these objects were extensively studied in \cite{sanyal2011}. It is also worth noting that these objects are sometimes spectrahedral, especially in the context of Lie theory, see \cite{Kobertphdthesis}. We now state formally the key corollary regarding transferring linear optimization from the orbitope of an adjoint orbit $\mathcal{O}_X$ to the associated polytope guaranteed by the Kostnat convexity theorem. \begin{corollary}\label{cor:linearopt} Let $G$, $T$, $\mathfrak{g}$, $\mathfrak{t}$ be as in the Kostant convexity theorem with $\mathfrak{t} \subset \mathfrak{g}$, and suppose $\mathfrak{g}$ is semisimple with negative definite Killing form $B$. Then for any $X,Z \in \mathfrak{t}$, we have \[ \min_{Y \in \mathcal{O}_X} -B(Y, Z) = \min_{Y \in W \cdot X} -B(Y, Z), \] where $\mathcal{O}_X$ is the adjoint orbit of $X$ and $W \cdot X$ is the Weyl group orbit of $X$. That is, optimization of a linear functional can be restricted to $W \cdot X = \mathcal{O}_X \cap \mathfrak{t}$ in this case. \end{corollary} \begin{proof} Let $\pi: \mathfrak{g} \to \mathfrak{t}$ be the orthogonal projection induced by the Killing form $B$, as in the Kostant convexity theorem. Since $Z \in \mathfrak{t}$, we have \[ \min_{Y \in \mathcal{O}_X} -B(Y, Z) = \min_{Y \in \mathcal{O}_X} -B(\pi(Y), Z) = \min_{Y \in \pi(\mathcal{O}_X)} -B(Y, Z). \] The Kostant convexity theorem then implies \[ \min_{Y \in \pi(\mathcal{O}_X)} -B(Y, Z) = \min_{Y \in \hull(W \cdot X)} -B(Y, Z). \] Since a linear functional is always optimized over a polytope at one of its extreme points, we finally have \[ \min_{Y \in \hull(W \cdot X)} -B(Y, Z) = \min_{Y \in W \cdot Z} -B(Y,Z). \] This completes the proof. \end{proof} \begin{remark} \label{rem:Killing_u_n} Note that although we restrict to compact semisimple Lie algebras in many the above results, the same results can also be stated for more general compact Lie algebras by adjusting the Killing form $B$ on $\mathfrak{g}$ so that it is strictly negative definite but still $\Ad$-invariant and $\ad$-invariant. % To do this, one simply needs to construct any negated inner product $B_0$ on the abelian part of $\mathfrak{g}$, and consider $B + B_0$. % The invariance properties are then trivially satisfied by $B_0$ and thus also by $B + B_0$. This is precisely what was done in Corollary \ref{cor:kostant_unitary} to obtain the negated inner product $\Tr(XY)$ on $\mathfrak{u}_n$, since $\mathfrak{u}_n$ is not semisimple but has one-dimensional center given by purely imaginary multiples of the identity matrix. See \cite[Prop.~4.24]{knapp2013} for more information. \end{remark} \begin{remark} Applying Corollary \ref{cor:linearopt} to the unitary group via Corollary \ref{cor:kostant_unitary} gives a result about optimizing $\Tr(YZ)$ over the adjoint orbits contained in $\mathfrak{u}_n$, the space of \emph{skew}-Hermitian matrices. However, the goal of the optimization problem of Section \ref{sec:eigenvalue} is to optimize over orbits of the unitary group within the space of (non-skew) Hermitian matrices. To see that these problems are equivalent, note that $\mathfrak{u}_n = i \cdot \mathcal{H}_{n \times n}$ where $\mathcal{H}_{n \times n}$ is the (real) vector space of Hermitian matrices. And further, the subalgebra $\mathfrak{t}$ of diagonal skew-Hermitian matrices is the set of diagonal matrices with purely imaginary diagonal entries. Therefore optimizing $\Tr(YZ)$ over $Y \in \mathcal{O}_X \subset \mathfrak{u}_n$ for $Z \in \mathfrak{t}$ is equivalent to optimizing $\Tr(iY \cdot iZ) = -\Tr(YZ)$ where $Y$ lies in some conjugation orbit of Hermitian matrices and $Z$ is a Hermitian diagonal matrix. Since minimizing $\Tr(YZ)$ is equivalent to maximizing $-\Tr(YZ)$ and vice versa, this shows that the two optimization problems are equivalent. \end{remark} Beyond the Kostant convexity theorem, there are actually other convexity theorems regarding various types of manifolds. For example, the \textbf{ Atiyah-Guillemin-Sternberg theorem} \cite{Atiyah1982,guillemin1982convexity} gives a similar result in the case of a symplectic manifold and its corresponding moment map. In the case of a symplectic manifold being acted upon by a compact Lie group, the \textbf{Kirwan convexity theorem} \cite{kirwan1984convexity} gives a result similar to that of the Kostant convexity theorem. In fact this generalizes the Kostant convexity theorem, since the KKS symplectic structure of Remark \ref{rem:adjoint_orbits_KKS} shows that adjoint orbits of a compact Lie group are symplectic manifolds. \begin{remark} \label{rem:coadjoint} The Kostant convexity theorem is actually originally stated for \emph{co}adjoint orbits in the Lie algebra dual space $\mathfrak{g}^*$. In our case, the invariant Killing form allows us to transfer this statement to adjoint orbits in order to state the theorem as we have above. \end{remark} \section{Lie Theory for Sampling over Orbits} We now explore how symmetries produced by Lie theory can help with developing sampling algorithms on manifolds. A standard hurdle in the development of sampling algorithms on finite discrete domains is the computation of the partition function. Given a weighting of the points in a discrete set, the partition function is the sum of all weights, giving the normalization constant with which one can turn the weighting into a probability distribution. This computation often bounds the computational complexity of a sampling problem on the domain. Computation of the partition function can sometimes be easy; for example in the case of the uniform distribution on the set of spanning trees on a connected graph, computation of the partition function is given by a determinant of a matrix with rows on the order of the number of vertices, even when the total number of trees is exponentially large (e.g., see \cite{feder1992balanced}). On the other hand, computing the partition function of the uniform distribution on the set of perfect matchings of a bipartite graph is known to be equivalent to computing the permanent of a 0-1 matrix, which is a \#P-hard problem (e.g., see \cite{jerrum1989approximating}). What if the sampling domain is a continuous manifold? In this case, the partition function becomes an integral rather than a sum, and the efficient computability of such a partition function becomes more difficult to obtain or understand. Thus, if we want to sample from continuous Lie-theoretic domains like adjoint orbits, we need to be able to compute integrals over such domains. We soon see how Lie theory and symmetry make this possible. \subsection{The Haar Measure} Often computational questions regarding manifolds (such as integration, expectation, and sampling) require an appropriate probability measure on the manifold. In the discrete case or on convex bodies, there is an implicit probability measure that comes from the fact that there is a well-defined uniform measure. For manifolds in general, there is no clear way to define such a measure. In the case of compact Lie groups however, there is a natural invariant measure which we call the uniform measure, and we define it now. \begin{theorem} If $G$ is a compact Lie group, then there exists a unique probability measure $\mu$ on $G$ called the \textbf{Haar measure} or \textbf{uniform measure} which has the following properties: \begin{enumerate} \item \textbf{Probability measure:} $\int_G 1 \, d\mu(g) = 1$. \item \textbf{$G$-invariance:} For all $h \in G$ and all $\mu$-measurable functions $f$, we have: \[ \int_G f(hg) \, d\mu(g) = \int_G f(g) \, d\mu(g) = \int_G f(gh) \, d\mu(g). \] \end{enumerate} \end{theorem} \begin{remark} Haar measures can actually be defined for Lie groups in general, but the definition is more delicate. % In particular, there are potentially two different Haar measures: one corresponding to invariance respect to left multiplication and the other with respect to right multiplication. % For compact groups these two notions coincide, and so we restrict to this case here throughout. \end{remark} The Haar measure on a compact Lie group $G$ also induces Haar measures on more interesting manifolds which arise via certain constructions on $G$. For example, if a connected compact Lie group $G$ acts smoothly on some real manifold $M$, then the Haar measure on $G$ induces a $G$-invariant measure on any given orbit of the action of $G$ on $M$. This is formalized in the following. \begin{lemma}[Theorem 1.9 of \cite{helgason1984}] \label{lem:quotient_integral} Let $G$ be a compact Lie group with compact Lie subgroup $H$. There exists a $G$-invariant measure on the manifold $G/H$ such that \[ \int_G f(g) d\mu(g) = \int_{G / H} \left(\int_H f(gh) d\nu(h)\right) d\mu_x(g) \] for all $\mu$-measurable functions $f$ on $G$. \end{lemma} \begin{corollary} \label{cor:orbit_Haar_measure} Let $G$ be a compact Lie group, and let $G$ act smoothly on a manifold $M$. % Given $x \in M$, let $\mathcal{O}_x \subset M$ denote the orbit of $x$ in $M$ under the action of $G$. % There exists a unique $G$-invariant probability measure on $\mathcal{O}_x$ called the \textbf{Haar measure} or \textbf{uniform measure}. \end{corollary} \begin{proof} Since $G$ acts smoothly on $M$, the stabilizer $\Stab_x$ of $x$ in $G$ is a closed subgroup of $G$, and thus a closed Lie subgroup by Cartan's closed-subgroup theorem. % Therefore $\Stab_x$ is in fact a compact Lie group and thus has a Haar measure $\nu$, and we also have that $\mathcal{O}_x \cong G / \Stab_x$. % The previous lemma then implies there exists a unique $G$-invariant measure $\mu_x$ on $\mathcal{O}_x \cong G / \Stab_x$ such that \[ \int_G f(g) d\mu(g) = \int_{G / \Stab_x} \left(\int_{\Stab_x} f(gh) d\nu(h)\right) d\mu_x(g) = \int_{\mathcal{O}_x} \left(\int_{\Stab_x} f(gh) d\nu(h)\right) d\mu_x(g). \] This measure $\mu_x$ is then precisely the claimed Haar measure on $\mathcal{O}_x$. \end{proof} The above theorem not only guarantees Haar measures on more general manifolds beyond compact Lie groups, but the proof also tells us how to write certain types of integrals on those manifolds in terms of integrals on the associated compact Lie group. This essentially says that computing integrals over such manifolds (adjoint orbits for example) boils down to computing a similar integral over the compact Lie group acting on the manifold. \subsection{The Harish-Chandra and Harish-Chandra--Itzykson--zuber Integral Formulas} We now have a notion of uniform measure with respect to which we can write down integrals over Lie groups and adjoint orbits. But, this seems to get us no closer to actually computing such integrals. This is where the symmetries of the uniform measure and the Lie group itself come into play, yielding two of the most important results (from a computational perspective) in the theory of compact Lie groups: the \textbf{Harish-Chandra (HC) integral formula}, along with its corollary, the \textbf{Harish-Chandra--Itzykson--Zuber (HCIZ) integral formula}. \begin{theorem}[HC integral formula, \cite{HC}]\label{thm:HC} Let $G$ be a compact connected semisimple Lie group, $\mathfrak{g}$ the associated Lie algebra, and $\mathfrak{t} \subset \mathfrak{g}$ a Cartan subalgebra. % For any $H_1,H_2 \in \mathfrak{t}$, we have \[ \int_G e^{B(H_1, \Ad_g(H_2))} \, d\mu(g) = C_G \cdot \frac{\sum_{w \in W} \epsilon(w) e^{B(H_1, w \cdot H_2)}}{\Pi(H_1) \Pi(H_2)}, \] where $B$ is the Killing form of $\mathfrak{g}$, $\mu$ is the Haar measure on $G$, $W$ is the Weyl group of $G$, $\Pi$ is an efficiently computable function on $\mathfrak{t}$, $C_G$ is an efficiently computable constant, and $\epsilon(w)$ is the sign of $w$. (We do not define the notion of sign here, but note that it is analogous to the notion of the sign of a permutation in $S_n$.) \end{theorem} \noindent A priori, without the HC formula, it is surprising that the integral on the left hand side can be written in terms of determinants. However, with the expression in the HC formula and the knowledge of the Weyl group, one might guess that the right hand side may be written in terms of determinants. We now state the HCIZ formula, which can be viewed as a special case of the HC formula, where the Lie group is chosen to be the unitary group $\U(n)$. \begin{theorem}[HCIZ integral formula, \cite{HC,IZ1980}]\label{thm:HCIZ} Given real diagonal $n \times n$ matrices $Y = \diag(y)$ and $\Lambda = \diag(\lambda)$ with diagonal entries $y_1 > y_2 > \cdots > y_n$ and $\lambda_1 > \lambda_2 > \cdots > \lambda_n$ respectively, we have \[ \int_{\U(n)} e^{-\langle Y, U \Lambda U^* \rangle_F} \, d\mu(U) = \left(\prod_{p=1}^{n-1} p!\right) \frac{\det([e^{-y_i\lambda_j}]_{1 \leq i,j \leq n})}{\prod_{i<j} -(y_i-y_j)(\lambda_i-\lambda_j)}, \] where $\langle \cdot, \cdot \rangle_F$ is the Frobenius inner product and $\mu$ is the Haar measure on $\U(n)$. \end{theorem} \noindent By expanding the determinant in the HCIZ formula, one obtains \[ \det([e^{-y_i\lambda_j}]_{1 \leq i,j \leq n}) = \sum_{\sigma \in S_n} \epsilon(\sigma) \prod_{i=1}^n e^{-y_i \lambda_{\sigma(i)}} = \sum_{\sigma \in S_n} \epsilon(\sigma) e^{-\langle y, \sigma \cdot \lambda \rangle} = \sum_{\sigma \in S_n} \epsilon(\sigma) e^{-\langle Y, \sigma \cdot \Lambda \rangle_F} \] since $Y$ and $\Lambda$ are diagonal, where $\epsilon(\sigma)$ denotes the sign of the permutation $\sigma$. This then implies \[ \int_{\U(n)} e^{-\langle Y, U \Lambda U^* \rangle_F} \, d\mu(U) = \left(\prod_{p=1}^{n-1} p!\right) \frac{\sum_{\sigma \in S_n} \epsilon(\sigma) e^{-\langle Y, \sigma \cdot \Lambda \rangle_F}}{\prod_{i<j} -(y_i-y_j)(\lambda_i-\lambda_j)}, \] which is essentially the HC formula. Note that formally, the HCIZ formula is not an immediate corollary of the HC formula because $\U(n)$ is not semisimple. However, this derivation can be made formal; see \cite{mcswiggen2019}. \begin{remark} The difference in various signs between the HC and HCIZ formulas comes from the fact that in the HCIZ formula we deal with Hermitian matrices, while in the HC formula for $G = \U(n)$ we deal with the Lie algebra $\mathfrak{u}_n$ consisting of \emph{skew}-Hermitian matrices. \end{remark} Harish-Chandra originally studied such integrals in the `50s in order to develop a theory of Fourier analysis on semisimple Lie algebras. Since then, these formulas have acquired far-reaching applications in other areas. The integral expression given above in the HCIZ formula is a partition function for a distribution on the set of unitary matrices $\U(n)$, which we call \textbf{HCIZ densities}. Such unitary matrices are distributed according to the density function $e^{-\langle Y, U \Lambda U^* \rangle_F}$, and they are important in various settings in physics and random matrix theory. For instance, they appear in multi-matrix models in quantum field theory and string theory \cite{IZ,DGZ}, and they are also related to models of coupled Gaussian matrices \cite{IZ} that have been used to solve the Ising model on a planar random lattice \cite{Kaz, BouKaz}. Further, the moments of such HCIZ-distributed matrices are useful for computing correlation functions for matrix models of gauge theories which have been studied for over 20 years \cite{Morozov, Sha, Eynard-note, E-PF, PEDZ}. Beyond HCIZ-distributed random matrices, the integral itself occurs in expressions for joint spectral densities of certain matrix ensembles such as Wishart matrices and off-center Wigner matrices \cite{AGsurvey}. The integral is also important in representation theory, where the corresponding finite sum formula is closely related to the Weyl character formula which describes the characters of irreducible representations of compact Lie groups (see \cite{knapp2013}). The algorithmic significance of these results is that they give discrete sum formulas for continuous partition functions on compact Lie groups. Conceptually, the Weyl group of a Lie group $G$ acts as a ``quadrature group'' for $G$; that is, integrals of certain exponential functions over the group $G$ can be computed by instead summing those functions over the finite Weyl group $W$. Unfortunately the Weyl group $W$ is often related to the symmetric group, and so it is a priori too large for the HC formula to be efficiently computable. For many matrix groups however, the formula can be rewritten in terms of a small sum of determinants, as in the HCIZ formula, and this leads to something which can be efficiently computed, see \cite{LeakeV20b}. See \cite{mcswiggen2018harish} for more details on the formula itself, as well as its history. We now have formulas for integrals over a compact Lie group, but how do these HCIZ and HC densities and formulas relate to integral formulas on the adjoint orbits of a Lie group? We now answer this explicitly in the case of unitary orbits. Once we have the HCIZ formula, we can apply Lemma \ref{lem:quotient_integral} and Corollary \ref{cor:orbit_Haar_measure} to immediately obtain a similar integral formula on the adjoint orbits of $\U(n)$, given as follows. \[ \int_{\mathcal{O}_\Lambda} e^{-\langle Y, X \rangle_F} \, d\mu_\Lambda(X) = \int_{\U(n)} e^{-\langle Y, U \Lambda U^* \rangle_F} \, d\mu(U) = \left(\prod_{p=1}^{n-1} p!\right) \frac{\det([e^{-y_i\lambda_j}]_{1 \leq i,j \leq n})}{\prod_{i<j} -(y_i-y_j)(\lambda_i-\lambda_j)}. \] These log-linear densities on the adjoint obrits of $\U(n)$ are also well-studied. They arise for example as solutions to certain maximum entropy problems over the adjoint orbits, with one specific application being the computation of the entropy-maximizing representation of a quantum density matrix as an ensemble of pure states \cite{Band1976,Park1977,SlaterEntropy1}; see also \cite{LeakeV20}. When the adjoint orbit is given by rank-$k$ PSD projections, such distributions give rise to the exponential mechanism for differentially private low-rank approximation \cite{MTalwar,Kamalika,KTalwar}. In the context of statistics, these distributions go by the name \textbf{matrix Langevin} and \textbf{matrix Bingham} distributions, except that typically one considers adjoint orbits of the orthogonal group rather than the unitary group \cite{ChikusePaper,ChikuseBook}. And finally, when considering a convex body instead of an adjoint orbit, such entropy-maximizing distributions arose in the work of Klartag (inspired by a work of Gromov) on the isotropic constant \cite{Klartag2006,Gromov1990}, and these distributions also appear connection to the works of G\"uler, Bubeck and Eldan on barrier functions for interior point methods \cite{Guler1997,Guler1998,BubeckEldan}. These HC, HCIZ, and orbit formulas immediately prompt a sampling question. That is, we have explicit expressions for the partition functions of these densities on a compact Lie group; can we efficiently sample from such distributions? In the case of the unitary group, the problem of sampling according to a given HCIZ density and the problem of sampling from to a specific adjoint orbit are equivalent. The sampling question in this case has been answered affirmatively, and an efficient algorithm for sampling has been recently given by \cite{LMV2021}. The algorithm makes crucial use of Lie theory: the idea of obtaining a convex polytope from projecting a given adjoint orbit, as in the Kostant convexity theorem, is an essential part of the algorithm. Generally speaking, one first samples from the polytope, and then samples from the adjoint orbit by sampling from the corresponding fiber of the projection map. (Though it should be mentioned that the polytope and projection are slightly different; they are the \textbf{Gelfand-Tsetlin polytope} and the \textbf{Rayleigh map}. See \cite{LMV2021} for more discussion.) Beyond the unitary group, the question of efficient sampling remains open, even in the case of adjoint orbits of the real orthogonal group. That is, efficiently sampling from matrix Langevin and matrix Bingham distributions as they are traditionally defined is an open problem. \begin{remark} The expression above for the integral of a log-linear function over $\mathcal{O}_\Lambda$ looks different than the expression of Theorem \ref{thm:partition}. This is because here we have assumed that the eigenvalues of $\Lambda$ are distinct, while the eigenvalues of $X = e_1e_1^\top$ are not distinct. % One way to make the connection between these two cases is via a limiting argument and applying L'H\^opital's rule. % The details can be found in \cite{LeakeV20}. \end{remark} \subsubsection{Proof (Sketch) of the HCIZ Formula} In this section, we sketch a proof of the HCIZ formula (Theorem \ref{thm:HCIZ}) due to \cite{faraut2015rayleigh}. Recall that we want to prove \[ \int_{\U(n)} e^{-\langle Y, U \Lambda U^* \rangle_F} \, d\mu(U) = \left(\prod_{p=1}^{n-1} p!\right) \frac{\det([e^{-y_i\lambda_j}]_{1 \leq i,j \leq n})}{\prod_{i<j} -(y_i-y_j)(\lambda_i-\lambda_j)}. \] First recall that by Lemma \ref{lem:quotient_integral} and Corollary \ref{cor:orbit_Haar_measure}, we have that \[ \int_{\mathcal{O}_\Lambda} e^{-\langle Y, X \rangle_F} \, d\mu_\Lambda(X) = \int_{\U(n)} e^{-\langle Y, U \Lambda U^* \rangle_F} \, d\mu(U). \] We now define the measure $\mu_\Lambda^{(k)}$ to be the measure obtained by projecting $\mu_\Lambda$ through the map $\pi_k$ which maps a matrix to its leading $k \times k$ principal submatrix. Letting $\mathcal{O}_\Lambda^{(k)} := \pi_k(\mathcal{O}_\Lambda)$, this definition implies the formula \[ \int_{\mathcal{O}_\Lambda^{(k)}} f(Y) \, d\mu_\Lambda^{(k)}(Y) = \int_{\mathcal{O}_\Lambda} f(\pi_k(X)) \, d\mu_\Lambda(X) \] for any function $f$ on $\mathcal{O}_\Lambda^{(k)}$. The support $\mathcal{O}_\Lambda^{(k)}$ of the measure $\mu_\Lambda^{(k)}$ can be described explicitly, using the following result often attributed to Cauchy or Rayleigh. \begin{proposition}[Rayleigh-Cauchy] Let $\Lambda$ be a Hermitian matrix with eigenvalues $\lambda_1 \geq \lambda_2 \geq \cdots \geq \lambda_n$, and let $\mu_1 \geq \cdots \geq \mu_{n-1}$ be the eigenvalues of the leading $(n-1) \times (n-1)$ principal submatrix of $\Lambda$. % Then the following interlacing relations hold: \[ \lambda_1 \geq \mu_1 \geq \lambda_2 \geq \mu_2 \geq \cdots \geq \lambda_{n-1} \geq \mu_{n-1} \geq \lambda_n. \] \end{proposition} \noindent This fact can actually be strengthened: in fact, these interlacing relations completely determine the support of the measure $\mu_\Lambda^{(n-1)}$: \[ \mathcal{O}_\Lambda^{(n-1)} = \{M \in \mathbb{C}^{(n-1) \times (n-1)} : \eig(M) \text{ interlace } \eig(\Lambda)\}. \] That is, every $(n-1) \times (n-1)$ Hermitian matrix with eigenvalues interlacing those of $\Lambda$ can appear as the principal submatrix of a unitary conjugate of $\Lambda$. Further, the set of possible eigenvalues of elements of $\mathcal{O}_\Lambda^{(n-1)}$ form a convex scaled hypercube given by the above inequalities, which we denote by $P_\Lambda$. By induction, we obtain similar results for $\mathcal{O}_\Lambda^{(k)}$ for all $k$. Beyond this expression for $\mathcal{O}_\Lambda^{(n-1)}$, Barychnikov has given a formula for the probability distribution on the eigenvalues of matrices which can appear in $\mathcal{O}_\Lambda^{(n-1)}$. This is given explicitly as follows. \begin{theorem}[Barychnikov, see Theorem 1.3 of \cite{faraut2015rayleigh}] Fix a Hermitian matrix $\Lambda$, and let $X$ be sampled from $\mathcal{O}_\Lambda$ according to the uniform probability measure $\mu_\Lambda$. % Then $\eig(\pi_{n-1}(X))$ is a vector sampled from the convex scaled hypercube $P_\Lambda$ of vectors which interlace $\eig(\Lambda)$, according to a probability measure which is a density function $g_\Lambda$ times Lebesgue measure restricted to $P_\Lambda$, and the $g_\Lambda$ is explicitly given by \[ g_\Lambda(x) = (n-1)! \frac{V_{n-1}(x)}{V_n(\eig(\Lambda))}, \] where $V_n(x)$ is the Vandermonde determinant $V_n(x) = \prod_{i<j} (x_i-x_j)$. \end{theorem} \noindent This then further implies another equivalent expression for orbital integrals, given by \[ \int_{\mathcal{O}_\Lambda^{(n-1)}} f(Y) \, d\mu_\Lambda^{(n-1)}(Y) = \int_{a \in P_\Lambda} \left[\int_{\mathcal{O}_{\diag(a)}} f(X) \, d\mu_{\diag(a)}(X)\right] g_\Lambda(a) \, da. \] That is, we have split the integral by first selecting a collection of eigenvalues $a$ according to $f_\Lambda(a) da$ on $P_\Lambda$, where $f_\Lambda$ is as defined in the previous theorem, and then integrating over the unitary orbit of $\diag(a)$ which is contained within $\mathcal{O}_\Lambda^{(n-1)}$. We now have the necessary tools to prove an inductive formula for the desired orbital intergal over $e^{-\langle Y, X \rangle_F}$. We first give an alternative expression for the exponent $-\langle X, Y \rangle_F$ in the orbital integral. For this, we now assume that $Y = \diag(y)$ is diagonal, which is without loss of generality by unitary conjugation. Letting $X'$ and $Y'$ denote the leading principal submatrices of $X$ and $Y$ respectively, we then have \[ -\langle Y, X \rangle_F = -\Tr(YX) = -\Tr(Y'X') - y_n (\Tr(X)-\Tr(X')) = -\langle Y', X' \rangle_F - y_n (\Tr(\Lambda)-\Tr(X')). \] We now plug this expression into the orbital integral to obtain \[ \int_{\mathcal{O}_\Lambda} e^{-\langle Y, X \rangle_F} \, d\mu_\Lambda(X) = e^{-y_n \Tr(\Lambda)} \int_{\mathcal{O}_\Lambda} e^{y_n \Tr(X')} e^{-\langle Y', X' \rangle_F} \, d\mu_\Lambda(X). \] The integrand is then a function only of $X' = \pi_{n-1}(X)$ and not of $X$ itself. Therefore we can apply the above formulas for integration over $\mathcal{O}_\Lambda^{(n-1)}$ to obtain \[ \begin{split} \int_{\mathcal{O}_\Lambda} e^{-\langle Y, X \rangle_F} \, d\mu_\Lambda(X) &= e^{-y_n \Tr(\Lambda)} \int_{\mathcal{O}^{(n-1)}_\Lambda} e^{y_n \Tr(Z)} e^{-\langle \pi_{n-1}(Y), Z \rangle_F} \, d\mu_\Lambda^{(n-1)}(Z) \\ &= e^{-y_n \Tr(\Lambda)} \int_{\mathcal{O}^{(n-1)}_\Lambda} e^{y_n \Tr(Z)} e^{-\langle \pi_{n-1}(Y), Z \rangle_F} \, d\mu_\Lambda^{(n-1)}(Z) \\ &= \int_{P_\Lambda} e^{y_n(a_1+\cdots + a_{n-1} - \Tr(\Lambda))} \left[\int_{\mathcal{O}_{\diag(a)}} e^{-\langle \pi_{n-1}(Y), Z \rangle_F} \, d\mu_{\diag(a)}(Z)\right] g_\Lambda(a) \, da. \end{split} \] That is, if $\mathcal{I}_\Lambda(Y) := \int_{\mathcal{O}_\Lambda} e^{-\langle Y, X \rangle_F} \, d\mu_\Lambda(X)$ is the integral we want to compute, we have the inductive formula \[ \mathcal{I}_\Lambda(Y) = (n-1)!\frac{ e^{-y_n \Tr(\Lambda)}}{V_n(\eig(\Lambda))} \int_{P_\Lambda} \mathcal{I}_{\diag(a)}(\pi_{n-1}(Y)) \cdot e^{y_n(a_1+\cdots + a_{n-1})} V_{n-1}(a) \, da. \] Our candidate formula for $\mathcal{I}_\Lambda(Y)$ is then of course given by the desired HCIZ formula, where $\lambda$ is the vector of eigenvalues of $\Lambda$: \[ \mathcal{I}_\Lambda(Y) = \left(\prod_{p=1}^{n-1} p!\right) \frac{\det([e^{-y_i\lambda_j}]_{1 \leq i,j \leq n})}{\prod_{i<j} -(y_i-y_j)(\lambda_i-\lambda_j)} = \left(\prod_{p=1}^{n-1} p!\right) \frac{\det([e^{-y_i\lambda_j}]_{1 \leq i,j \leq n})}{V_n(-y) V_n(\eig(\Lambda))}. \] With this, one then can prove relatively straightforwardly that this expression holds inductively. The key idea towards this is to note that since $P_\Lambda$ is a convex scaled hypercube, the integral $\int_{P_\Lambda} f(a) \, da$ can be written as a product of integrals over each of the $n-1$ coordinates of $a \in P_\Lambda \subset \R^{n-1}$. We do this as follows, where $\lambda$ is the vector of eigenvalues of $\Lambda$ and $y' := (y_1,\ldots,y_{n-1})$ is the vector of eigenvalues of $\pi_{n-1}(Y)$ since $Y$ is diagonal: \[ \begin{split} \int_{P_\Lambda} &\mathcal{I}_{\diag(a)}(\pi_{n-1}(Y)) \cdot e^{y_n(a_1+\cdots + a_{n-1})} V_{n-1}(a) \, da \\ &= \int_{\lambda_2}^{\lambda_1} \int_{\lambda_3}^{\lambda_2} \cdots \int_{\lambda_n}^{\lambda_{n-1}} \mathcal{I}_{\diag(a)}(\pi_{n-1}(Y)) \cdot e^{y_n(a_1+\cdots + a_{n-1})} V_{n-1}(a) \, d_{a_{n-1}} \cdots da_2 \, da_1 \\ &= \int_{\lambda_2}^{\lambda_1} \int_{\lambda_3}^{\lambda_2} \cdots \int_{\lambda_n}^{\lambda_{n-1}} \left(\prod_{p=1}^{n-2} p!\right) \frac{\det([e^{-y_i a_j}]_{1 \leq i,j \leq n-1})}{V_{n-1}(-y') V_{n-1}(a)} \cdot e^{y_n(a_1+\cdots + a_{n-1})} V_{n-1}(a) \, d_{a_{n-1}} \cdots da_2 \, da_1 \\ &= \frac{\prod_{p=1}^{n-2} p!}{V_{n-1}(-y')} \int_{\lambda_2}^{\lambda_1} \int_{\lambda_3}^{\lambda_2} \cdots \int_{\lambda_n}^{\lambda_{n-1}} \det([e^{(y_n-y_i) a_j}]_{1 \leq i,j \leq n-1})\, d_{a_{n-1}} \cdots da_2 \, da_1 \\ \end{split} \] Note now that each integration is a linear operation, and that the determinant in the proposed expression for $\mathcal{I}_{\diag(a)}(\pi_{n-1}(Y))$ is multilinear in the coordinates of $a$. These integrals can be passed into the determinant to be applied separately to each of the rows of the input matrix. Combining this with the above inductive expression for $\mathcal{I}_\Lambda(Y)$ then yields the desired HCIZ formula. \section{Classification Theorems for Lie Algebras} \label{sec:classification} In Section \ref{sec:Lie}, we showed how a Lie algebra could be defined as the tangent space of a Lie group $G$ as the identity element $e$, equipped with a certain product-like operation called the Lie bracket. We now define the notion of a Lie algebra more abstractly, and show how this notion coincides with that of Section \ref{sec:lie_alg_via_group}. The power of this abstract definition is that it makes no reference to Lie groups. Powerful classification theorems of Lie algebras, a priori having nothing to do with groups, then extend immediately to certain important classes of Lie groups. Before moving on, we make one caveat. An abstract Lie algebra $\mathfrak{g}$ is a vector space equipped with a Lie bracket which we discuss below. The vector space structure of $\mathfrak{g}$ requires a choice of field, and for us this field is either $\mathbb{C}$ or $\R$. Many of the nice classification results for Lie algebras use $\mathbb{C}$ because they rely on the fact that $\mathbb{C}$ is algebraically closed. However, compact Lie groups, like the unitary group, have associated Lie algebra which is necessarily real. (Even though $\mathfrak{u}_n$ consists of skew-Hermitian matrices with complex entries, there is no way to consider of the space of skew-Hermitian matrices as a complex vector space. For example, multiplication by $i$ does not preserve $\mathfrak{u}_n$.) This point is not very important to our exposition, but it is still worth making to avoid certain confusions. \subsection{Abstract Lie Algebras} An abstract Lie algebra $\mathfrak{g}$ is a vector space equipped with a product-like operation $[\cdot, \cdot]: \mathfrak{g} \times \mathfrak{g} \to \mathfrak{g}$ called the \textbf{Lie bracket}. We also often write $\ad_X(Y) = [X,Y]$, where $\ad_X$ is called the (algebra) adjoint action of the Lie algebra on itself, as discussed above. The Lie bracket has three defining properties, given as follows: \begin{enumerate} \item \textbf{Bilinearity:} $[aX+bY,Z] = a[X,Z] + b[Y,Z]$ and $[X,bY+cZ] = b[X,Y] = c[X,Z]$. \item \textbf{Anti-symmetry:} $[X,Y] = -[Y,X]$. \item \textbf{Jacobi identity:} $[X,[Y,Z]] + [Y,[Z,X]] + [Z,[X,Y]] = 0$. \end{enumerate} In the case of Lie algebras associated to matrix groups, the Lie bracket is given by the commutator $[X,Y] = XY-YX$. It is straightforward to see that this satisfies the defining properties listed above. \subsection{Simple, Semisimple, and Reductive Lie Algebras} We now discuss a number of important notions regarding a Lie algebra $\mathfrak{g}$, some of which we have already seen in the above discussion. This helps us in our attempt to classify the Lie algebras we hope to consider. Here are some of them: \begin{itemize} \item \textbf{Abelian:} A Lie algebra $\mathfrak{g}$ for which $[X,Y] = 0$ for all $X,Y \in \mathfrak{g}$. \item \textbf{Subalgebra:} A subspace $\mathfrak{h} \subset \mathfrak{g}$ for which $[X,Y] \in \mathfrak{h}$ for all $X,Y \in \mathfrak{h}$. \item \textbf{Ideal:} A subalgebra $\mathfrak{h} \subset \mathfrak{g}$ for which $[X,Y] \in \mathfrak{h}$ for all $X \in \mathfrak{g}$ and $Y \in \mathfrak{h}$. \item \textbf{Direct sum:} A direct sum of vector spaces $\mathfrak{g} = \bigoplus_{i=1}^m \mathfrak{g}_i$ with Lie bracket given by \[ [(X_1,X_2,\ldots,X_m),(Y_1,Y_2,\ldots,Y_m)] := ([X_1,Y_1],[X_2,Y_2],\ldots,[X_m,Y_m]). \] \item \textbf{Simple:} A non-abelian Lie algebra which has no nontrivial ideals. \item \textbf{Semisimple:} A direct sum of simple Lie algebras. \item \textbf{Reductive:} A direct sum of a semisimple Lie algebra and an abelian Lie algebra. \end{itemize} As a note, abelian Lie algebras are immediately classified by the dimension of the underlying vector space. The classification of semisimple and reductive Lie algebras is more complicated, and we describe it in the next section. \subsection{Classification of Complex Reductive Lie Algebras} One of the most important features of Lie theory is that there is a bijection between a certain class of nice Lie groups and the important class of semisimple Lie algebras. That is, not only can one construct a Lie algebra from such a Lie group, but there is also an inverse map from abstractly-defined Lie algebras to nice Lie groups. Therefore such Lie groups can be classified by classifying Lie algebras, without any reference to the fact that these algebras come from groups. This is a powerful fact since Lie algebras are linear-algebraic objects, and linear algebra is typically much easier than the geometry that would come with studying the groups directly. Broadly speaking, the classification relies upon the fact that a given Lie algebra can be associated to certain \emph{discrete} combinatorial objects called Dynkin diagrams. By classifying the valid combinatorial objects, one then obtains a classification of the associated Lie algebras. We do not go through such a proof of the classification, and the interested reader can find various proofs in standard references. Instead, we simply state the classification theorem. \begin{theorem}[Classification of complex simple Lie algebras, see \S VI.10 of \cite{knapp2013}] Every complex simple Lie algebra fits into one of the following categories: \begin{itemize} \item $A_n \colon \mathfrak{sl}_{n+1}$, the Lie algebra associated to $\SL_n(\mathbb{C})$. \item $B_n \colon \mathfrak{so}_{2n+1}$, the odd-dimensional Lie algebra associated to $\SO_{2n+1}(\mathbb{C})$. \item $C_n \colon \mathfrak{sp}_{2n}$, the Lie algebra associated to $\Sp_{2n}(\mathbb{C})$. \item $D_n \colon \mathfrak{so}_{2n}$, the even-dimensional Lie algebra associated to $\SO_{2n}(\mathbb{C})$. \item $E_6, E_7, E_8, F_4, G_2 :$ the so-called \textbf{exceptional Lie algebras}. \end{itemize} Note that this also classifies complex semisimple and reductive Lie algebras by taking direct sums. \end{theorem} \noindent We do not go into any more detail on the exceptional Lie algebras mentioned in the classification theorem. And as a final note, the notation used for the here ($A_n$, $B_n$, etc.) comes from the notation used to represent the \textbf{Dynkin diagrams} associated to the respective Lie algebras. A Dynkin diagram gives a way to encode the information of a Lie algebra in a purely combinatorial way, and classifying Dynkin diagrams is at the heart of most proofs of the above classification. We do not discuss Dynkin diagrams any further here. \subsection{Classification of Compact Lie Algebras} Although we have classified the complex semisimple Lie algebras above, this is not quite enough for our purposes. We actually want to classify Lie algebras associated to compact Lie groups. And as mentioned above, compact Lie groups must be considered as real manifolds; they cannot in general be given the structure of a complex manifold. Because of this, their associated Lie algebras are real instead of complex, and so they do not fit into the above classification. That said, recall that we defined a compact Lie algebra to be a Lie algebra which is derived from a compact Lie group. Although this is an extrinsic definition, compactness of a Lie algebra $\mathfrak{g}$ can be defined intrinsically via properties of the Killing form of $\mathfrak{g}$, without any reference to an associated Lie group. That is, a compact Lie algebra is a Lie algebra from which the Killing form is negative semidefinite. We now extend the above classification to cover compact Lie algebras. This classification requires the notion of a \textbf{real form} of a complex Lie algebra, which is a real Lie algebra whose complexification is the given complex Lie algebra. In fact, every semisimple complex Lie algebra has a unique compact real form, and the converse is also true, as stated in the following result. \begin{theorem}[Classification of compact simple Lie algebras, see \S VI.10 of \cite{knapp2013}] Every compact simple Lie algebra fits into one of the following categories: \begin{itemize} \item $A_n \colon \mathfrak{su}_{n+1}$, the Lie algebra associated to ${\rm SU}}\def\U{{\rm U}}\def\SO{{\rm SO}}\def\O{{\rm O}}\def\USp{{\rm USp}(n+1)$. \item $B_n \colon \mathfrak{so}_{2n+1}$, the odd-dimensional Lie algebra associated to $\SO_{2n+1}(\R)$. \item $C_n \colon \mathfrak{usp}_n$, the Lie algebra associated to $\USp(n)$. \item $D_n \colon \mathfrak{so}_{2n}$, the even-dimensional Lie algebra associated to $\SO_{2n}(\R)$. \item $E_6, E_7, E_8, F_4, G_2 :$ the compact forms of the exceptional Lie algebras. \end{itemize} Note that this also classifies compact semisimple and reductive Lie algebras by taking direct sums. \end{theorem} Note that some of the notation in the complex and compact classifications results is overloaded. These are distinguished by the fact that the underlying vector spaces are complex and real vector spaces respectively, and this distinction is often clear from context. The final point to make to classify compact Lie algebras completely is to handle the case of compact Lie algebras which are not reductive. While there are many complex Lie algebras which are not reductive, the opposite is true in the compact case. \begin{proposition}[Cor. 4.25 of \cite{knapp2013}] Every compact Lie algebra is reductive. \end{proposition} \noindent This last result then completes the classification of compact Lie algebras. \section{Summary and Conclusion} This article provides both the motivation and the basics to leverage symmetries in the design of efficient algorithms. In the first part, we discussed two basic and important problems. The first was an optimization characterization of the minimum eigenvalue problem (Theorem \ref{thm:eigenvalue}). For an $n \times n$ Hermitian matrix, we showed that \[ \min_{X \in \mathcal{P}_1} \langle A, X \rangle_F = \min_{U \in \U(n)} \langle A, Ue_1e_1^\top U^* \rangle_F = \lambda_1. \] This is a standard result, but we demonstrated how it follows from the Schur-Horn theorem. The Schur-Horn theorem says that the projection onto the diagonal of all matrices in the conjugation orbit of a Hermitian matrix yields a convex polytope. In the case of the orbit of $e_1e_1^\top$ considered above, this polytope is precisely the standard simplex corresponding to the convex hull of $\{e_{\sigma(1)} : \sigma \in S_n\}$. We further saw that this polytope is more generally given by a similar expression; that is, given a Hermitian matrix $A = \diag(v)$, the polytope is given by $\{\sigma \cdot v : \sigma \in S_n\}$ where $\sigma \cdot v$ acts by permuting the entries of $v$. The second problem we considered was that of sampling from a certain exponential density on the complex unit sphere. We showed that this problem is connected to the first one: computing integrals and sampling from the unit sphere is essentially equivalent to doing the same things on $\mathcal{P}_1$, the conjugation orbit of $e_1e_1^\top$ by the unitary group. We then discussed the partition function associated to such a density, which is given by the integral of an exponential function over the orbit $\mathcal{P}_1$. Computing such partitions functions is often the first step in constructing a sampling algorithm, and usually the complexity of this computation determines the complexity of the sampling algorithm. In this case of the partition function being an integral over a nonconvex manifold, it is not immediately clear how to compute such a partition function, and even less so if one wants to do it efficiently. In spite of this, we were able to show that not only could this partition function be efficiently computed in this case, but it in fact has a closed-form formula (Theorem \ref{thm:partition}): \[ \frac{1}{(n-1)!} \int_{\mathcal{P}_1} e^{-\langle Y, X \rangle_F} d\mu_1(X) = \sum_{i=1}^n \frac{e^{-\lambda_i}}{\prod_{j \neq i} (\lambda_j-\lambda_i)}, \] where $\lambda_i$ are the eigenvalues of $Y$. We gave a proof of this fact, which used the key fact that computing certain integrals over the nonconvex manifold $\mathcal{P}_1$ is equivalent to computing similar integrals over the standard simplex. In the second part, we focused on understanding very important groups which arise when there are continuous symmetries. These important groups were the Lie groups, which are special manifolds that have a compatible group structure. A crucial fact regarding Lie groups is that the tangent space at the identity recovers a large amount of information about the Lie group. This tangent space is a vector space equipped with a product-like operation coming from the group product, and it is called the Lie algebra. The Lie algebra of a Lie group is linear-algebraic in nature, while the Lie group is more geometric. Thus the Lie algebra gives access to more linear-algebraic methodologies and algorithms. The symmetries of a Lie group extend to the Lie algebra, and one way this is formalized is in the notion of the adjoint action of a Lie group on its Lie algebra. In the case of matrix groups, this adjoint action is precisely given by the natural action of matrix conjugation. This action allows Lie algebras to be viewed as a representation of the corresponding Lie group - called the ``adjoint representation.'' While we did not discuss it here, one can study more general representations of the Lie group via connections to the Lie algebra. The orbits of this adjoint action on the Lie algebra, called the adjoint orbits, are important Lie theoretic objects that we then studied further. These orbits are nonconvex manifolds, but there are interesting convex bodies associated to these objects. Convex hulls of adjoint orbits are called orbitopes \cite{sanyal2011}, and these are sometimes spectrahedral \cite{Kobertphdthesis}. But more interestingly, we saw that the image of a given adjoint orbit under the projection map to a Cartan subalgebra is a convex polytope, by the Kostant convexity theorem. In the case of the unitary group $\U(n)$, this yields the Schur-Horn theorem, which says that the set of all diagonal vectors of Hermitian matrices with given eigenvalues form a convex polytope. We then explored how certain optimization problems on nonconvex adjoint orbits could be reduced to optimization problems on the corresponding convex polytope. And, we also briefly mentioned that moment maps arise from symplectic structure via the KKS form on the adjoint orbits, which allows for Hamiltonian dynamics on these orbits. Finally, we discussed the HC and HCIZ formulas for certain exponential integrals over compact Lie groups. These important results give formulas for partition functions of certain measures over compact Lie groups and their adjoint orbits. The key surprising fact regarding these formulas is that they serve to reduce integrals over a group to a finite sum, or even a few determinants. That is, they reduce continuous symmetries to discrete symmetries. To conclude, computational problems that arise in nonconvex optimization and sampling can benefit when viewed through the lens of symmetries. This point of view has seen a recent surge and found several applications in areas such as complexity theory, fast graph algorithms, statistics, machine learning, quantum inference, and differential privacy (see, e.g., \cite{GargGOW16,BurgisserFGOWW19,LeakeV20,LMV2021} and the references therein). % However, the full potential of symmetries in optimization and sampling remains to be explored. \section*{Acknowledgments} The authors would like to thanks Anay Mehrotra and Yikai Wu for useful comments. NV would like to acknowledge the support of NSF CCF-1908347. \bibliographystyle{alpha}

1. Introduction {#sec1-materials-12-03379} =============== Negative thermal expansion (NTE) structures have been successfully used in aerospace engineering and precision devices where the structures experience large temperature fluctuation \[[@B1-materials-12-03379],[@B2-materials-12-03379]\]. For the new generation temperature-sensitive devices with lightweight, high-sensitivity and large measurement range, large NTE structures are especially required \[[@B3-materials-12-03379]\]. Bi-material triangular lattice is the most used NTE unit, due to its tailorable coefficient of thermal expansion (CTE) and lightweight. It requires extreme structural parameters to achieve large NTE. Its range of CTE is very narrow, considering the CTEs of the available materials and manufacturing capacity. This severely limits the development of the high-performance temperature-sensitive devices. Hence, proposing a larger NTE lattice is meaningful to improve the performance of temperature-sensitive devices. Much work has focused on expending the magnitude of NTE \[[@B4-materials-12-03379],[@B5-materials-12-03379],[@B6-materials-12-03379]\]. Many solid materials with NTE have been fabricated, and the mechanisms of NTE have been researched, including tungstates \[[@B7-materials-12-03379],[@B8-materials-12-03379],[@B9-materials-12-03379]\], perovskites \[[@B10-materials-12-03379],[@B11-materials-12-03379],[@B12-materials-12-03379]\], magnetic nanocrystals \[[@B13-materials-12-03379]\], cyanide-bridged framework material \[[@B14-materials-12-03379]\] and samarium fulleride \[[@B15-materials-12-03379]\]. However, the densities of solid NTE materials are large. Another prevailing method is utilizing lightweight lattices to obtain large NTE, including topology optimization composites \[[@B2-materials-12-03379],[@B16-materials-12-03379],[@B17-materials-12-03379]\] and bi-material lattices \[[@B18-materials-12-03379],[@B19-materials-12-03379],[@B20-materials-12-03379],[@B21-materials-12-03379],[@B22-materials-12-03379],[@B23-materials-12-03379],[@B24-materials-12-03379]\]. These structures could achieve shrinkage in length or volume with temperature increase to obtain NTE. This NTE is just equivalent property, not a real material property. Among these structures, bi-material lattices are widely used in engineering, due to their excellent mechanical properties and a large range of CTE. These can be divided into two categories: Bending dominated lattice \[[@B25-materials-12-03379],[@B26-materials-12-03379],[@B27-materials-12-03379],[@B28-materials-12-03379],[@B29-materials-12-03379],[@B30-materials-12-03379]\] and stretch-dominated lattice. The bending dominated lattice is based on the bi-material beam with initial curvature. The large stress between the joint interfaces and the complex manufacturing process limit the application of the lattices. The other category is stretch-dominated lattice, including planar lattices \[[@B19-materials-12-03379],[@B20-materials-12-03379],[@B21-materials-12-03379],[@B22-materials-12-03379],[@B31-materials-12-03379],[@B32-materials-12-03379],[@B33-materials-12-03379],[@B34-materials-12-03379],[@B35-materials-12-03379]\], cylindrical shell \[[@B36-materials-12-03379]\] and 3D bi-material lattices \[[@B37-materials-12-03379],[@B38-materials-12-03379],[@B39-materials-12-03379],[@B40-materials-12-03379]\]. It uses a large thermal expansion of one material to generate an angle changes to generate NTE. Triangular lattice is the most used unit. Theoretically, it could obtain unbounded NTE. However, it requires extreme structural parameters (small height and large CTE ratio of the two constituents) to achieve large NTE. When the height of the triangular lattice becomes small, the number of lattices for the whole structure will increase. This makes the assembly difficult. On the other hand, the large CTE ratio of the two constituents will make large thermal stress in the joints. Thus, its range of CTE is very narrow, considering the CTEs of the available materials and assembly process. Fortunately, a promising approach employs the hierarchical structure to achieve large NTE \[[@B3-materials-12-03379],[@B41-materials-12-03379],[@B42-materials-12-03379]\]. The hierarchical structure uses small scale bi-material lattices to replace the elements in the origin lattice. However, the small scale bi-material lattice is hard to fabricate and assembly. Using the hierarchical structure alone can only enlarge the NTE for a small magnitude. In summary, design a novel lattice with larger NTE is challenging and meaningful. In this paper, we report an ITL to overcome the limited NTE of conventional structures. The lattice has larger NTE compared with the bi-material triangular lattice with the same height and hypotenuse angle. In [Section 2](#sec2-materials-12-03379){ref-type="sec"}, we theoretically show that large inner NTE can be obtained by a composite rod. Based on the composite rod, we propose the ITL. In [Section 3](#sec3-materials-12-03379){ref-type="sec"}, finite element method analysis is conducted to reveal the influences of the structure parameters on the CTE and the maximum stress of the ITL. Moreover, comparisons of the ITL with the bi-material triangular lattice are conducted. In [Section 4](#sec4-materials-12-03379){ref-type="sec"}, CTE measurement by interferometry of the proposed lattice is performed to verify its excellent performance of large NTE. Discussions and conclusions are presented in [Section 5](#sec5-materials-12-03379){ref-type="sec"} and [Section 6](#sec6-materials-12-03379){ref-type="sec"}, respectively. 2. The Composite Rod and the ITL {#sec2-materials-12-03379} ================================ 2.1. Composite Rod {#sec2dot1-materials-12-03379} ------------------ ### 2.1.1. Structure of The Composite Rod {#sec2dot1dot1-materials-12-03379} The structure the composite rod is briefly presented here. It incorporates two constituents with different CTE. Constituent a with the lower CTE, α~a~, forms a long rod. Constituent b having the larger CTE, α~b~, is arranged as discontinuous short rod within the long rod, as shown in [Figure 1](#materials-12-03379-f001){ref-type="fig"}. The part between the inner ends of the two short rods is the 'virtual rod'. The 'virtual rod' could achieve large NTE if the length ratio L/r changes. It can be used as NTE rod for large NTE structure design. ### 2.1.2. Thermal Expansion of the 'Virtual Rod' {#sec2dot1dot2-materials-12-03379} The thermal expansion of the virtual rod is determined by the origin length, the length variation and the temperature variation. The origin length of the structure is fixed. For a given temperature increment, the length change decides the equivalent thermal expansion of the virtual rod. In the composite rod, virtual rod's CTE is decided by the thermal deformations of the long rod and the two short rods. For homogeneous temperature increment, the elongation of the long rod is smaller than that of the short rod. It generates a shrinkage of the inner ends of the two short rods. This will make the virtual rod with NTE. The 'virtual rod' has thermal expansion $\overline{\alpha}$ defined such that an increment in temperature ΔT causes a length change. With pin-jointed assumption, the detailed derivation of $\overline{\alpha}$ is given in [Section S1 (Supplementary Materials)](#app1-materials-12-03379){ref-type="app"}, $$\overline{\alpha} = \frac{\alpha_{b}r - (\alpha_{b} - \alpha_{a})L}{r}.$$ This relation between the thermal expansion ratio $\overline{\alpha}$/α~b~ and the length ratio L/r is plotted in [Figure 2](#materials-12-03379-f002){ref-type="fig"}. For a given thermal expansion ratio, thermal expansion of the 'virtual rod' is increasing if the length ratio increases. It is apparent that the 'virtual rod' has negative net thermal expansion within a realizable window of length ratio and thermal expansion ratio. Specifically, when the parameters on the red line where α/α~b~ = 0, zero thermal expansion emerges. There are two ways to achieve large inner NTE of the composite rod: Increasing the thermal expansion ratio and enlarging the length ratio. Large thermal expansion ratio will generate large thermal stress in the joints. On the contrary, it is easy to change the length ratio. The virtual rod's NTE could be extremely large if the length ratio increases. The composite rod is like a negative multiplier. It converts the positive thermal expansion ratio, α~b~/α~a~, to negative thermal expansion ratio, α/α~b~. The large inner NTE composite rod lays the foundation for the ITL's design. 2.2. ITL {#sec2dot2-materials-12-03379} -------- The large NTE of the composite rod is only achieved at its inner 'virtual rod'. The inner NTE is rarely used in engineering. Thus, we need to convert the inner NTE to the outer node. In this part, we design an ITL with large NTE in its vertical direction using the composite rod. It realizes this conversion. ### 2.2.1. Structure of the ITL {#sec2dot2dot1-materials-12-03379} The following features characterize the ITL. As shown in [Figure 3](#materials-12-03379-f003){ref-type="fig"}a, the base is the composite rod. Constituent 'a' with the lower CTE, forms the hypotenuse. Constituent 'b' with the larger CTE, forms the upper base. The two hypotenuses, upper base and the 'virtual rod' form an inverted isosceles trapezoid. The topology could be connected by pin-jointed, welding or press-fit. It should enable the length changes to be accommodated by a rotation (angle change) at the nodes or deformation of the rods. ### 2.2.2. NTE Mechanism of the ITL {#sec2dot2dot2-materials-12-03379} For homogeneous temperature increment, thermal deformation of the lattice is determined by the thermal expansion of all the rods. As shown in [Figure 3](#materials-12-03379-f003){ref-type="fig"}a, the arrows show the deformation mechanism of the ITL after heating. The base of the single ITL is the composite rod. It is negative deformation. The lower ends of the two hypotenuses move inward. While the upper base is positive thermal deformation. The upper ends of the two hypotenuses move outward. Then, the left hypotenuse rotates anticlockwise, and the right hypotenuse rotates clockwise. Thus, the inverted trapezoid structure converts the inner NTE of the composite rod to an extra vertical direction contraction through this rotation. It could enlarge the magnitude of NTE, due to the extra vertical contraction. The assembly method must be decided ahead. The pin-jointed lattice allows free rotation at the nodes. It makes no stress in the lattice. However, the joint gaps at the nodes will generate remarkable CTE error. The rigid connection could overcome the former disadvantage. It could make the two constituents connect with no gap. This will decrease the CTE error of the ITL. Hence, the ITL lattices used in following are rigidly connected. With rigid connection assumption, the deformations of the rods have both thermal deformation and bending deformation. It is hard to theoretically figure out the thermal expansion of the ITL. Thus, the CTE and stress of the ITL are explored through finite element analysis. 3. Effects of The Basic Structure Parameters {#sec3-materials-12-03379} ============================================ In this section, finite element analysis was performed to explore the effects of the basic structural parameters of ITL by using COMSOL Multiphysics. Schematic illustration of the ITL is shown in [Figure 3](#materials-12-03379-f003){ref-type="fig"}. With rigid connection assumption, the deformations of the rods are more complicated. The asymmetry of the structure will cause additional bending deformation. For a single ITL, as shown in [Figure 3](#materials-12-03379-f003){ref-type="fig"}a, the upper base and the short rods of the base will have an extra bending deformation when the temperature rises. This is induced by the bending of the hypotenuse. The bending deformation will decrease the total NTE of the ITL. The X-symmetric ITL, as shown in [Figure 3](#materials-12-03379-f003){ref-type="fig"}b, could overcome the extra bending deformation of the short rods of the composite rod. However, the two upper bases still have an extra bending deformation when the temperature rises. However, the ITL array could increase the structural symmetry. It could decrease the extra bending deformation. Thus, an ITL array model with three X-symmetric ITLs along the Y-direction was used in the following simulation. The geometrical model of an ITL array used in the numerical simulation is illustrated in [Figure 4](#materials-12-03379-f004){ref-type="fig"}. The boundary conditions are illustrated in [Figure 4](#materials-12-03379-f004){ref-type="fig"}a. It has three X-symmetric ITLs along the Y-direction. The two constituents are aluminium alloy and titanium, and the CTE ratio of the two constituents is 'α~b~/α~a~'. Half the length of the long rod is 'L', half the length of the 'virtual rod' is 'r', the length of the hypotenuse of the isosceles trapezoid is 'a', half the length of the upper base is 'b', the height of the ITL is 'h'. The 3D view of the geometrical model is show in [Figure 4](#materials-12-03379-f004){ref-type="fig"}c. The width of all the rods are 'w'; the thickness of all the rods are 't'. Due to the symmetry about y axis, the x-direction translation degree of freedom U for nodes A and B was restricted. Due to the symmetry about x axis, the y-direction translation degree of freedom V for nodes C and D was restricted. Besides, to avoid the z-direction translation, the z-direction translation degree of freedom W for nodes A, B, C and D was also constrained. With these constraints, the model was statically determined. In the simulation, the initial temperature was 20 ℃, and the final temperature was set as 220 ℃. As shown in [Figure 4](#materials-12-03379-f004){ref-type="fig"}b, the members are meshed by free hexahedral elements. The maximum element size is 2.5 mm; the minimum element size is 0.25 mm. The number of the vertex elements is 262, the number of edge elements is 13988, the number of the boundary elements is 52620, the number of elements is 40584. In order to ensure a mesh independent result, the stress concentration areas are meshed by smaller element size to ensure quite dense mesh density, as shown in [Figure 4](#materials-12-03379-f004){ref-type="fig"}d. [Figure 4](#materials-12-03379-f004){ref-type="fig"}e shows the deformation and stress of an ITL array (r = 24 mm, b = 60 mm, h = is 30 mm, L/r = 5, α~a~ = 8.6 × 10^−6^ K^−1^, α~a~ = 23 × 10^−6^ K^−1^, t = w = 5 mm) induced by a temperature increase of 200 °C. The legend bars indicate overall deformation and stress distribution, respectively. The vertical red line is the center distance of the long rods on the first and third X-symmetric ITL. The equivalent NTE of the ITL was calculated according to the deformation of the center distance. [Figure 4](#materials-12-03379-f004){ref-type="fig"}f provides the y component of deformation and stress along the hypotenuse. It was plotted according to the simulation results of the tilted red line in [Figure 4](#materials-12-03379-f004){ref-type="fig"}e. The maximum stress of the bi-material ITL was calculated according to the curve of stress along the hypotenuse. The following mainly studied the effects of the ITLs' basic structural parameters, including the length ratio of the long rod and the 'virtual rod', 'L/r', the CTE ratio of the two constituents, 'α~b~/α~a~'; and the height of the ITL, 'h'. The two constituents used in the numerical simulation were aluminum alloy and titanium. Properties of the constituents are presented in [Table 1](#materials-12-03379-t001){ref-type="table"}. The properties of the materials did not change with the temperature during the simulation. 3.1. The Length Ratio of The Long Rod and The Virtual Rod, L/r {#sec3dot1-materials-12-03379} -------------------------------------------------------------- The length ratio of the long rod and the virtual rod determined the CTE of the composite rod. Therefore, the length ratio of the two rods could greatly affect the CTE of the bi-material ITL. In this part, the effects of the length ratio on the CTE and the maximum stress of the bi-material ITL were presented. An instructive comparison of the bi-material ITL and the traditional triangular lattice was made as well. When the length ratio was changed from 2 to 10, CTE and stress of the ITLs were calculated. CTE was calculated by the Y-direction of the ITL, and the maximum stress was calculated without outer forces. The materials used in the numerical simulation were aluminum alloy and titanium. The other structural parameters except for the length of the long rod were fixed: The virtual rod length was 24 mm, the upper base length was 120 mm, the bi-material ITL height was 30 mm, the width of all the rods was 5 mm, the thickness of all the rods was 5 mm. The length ratio was changed by varying the length of the long rod. When the length ratio decreased to 1, the composite rod became only the long rod. Then, the two lower ends of the two hypotenuses will connect by the long rod. Thus, if L/r = 1, the Y-direction CTE the bi-material ITL would be equal to the Y-direction CTE of the traditional triangular lattice with the same height and hypotenuse angle. The Y-direction CTE and the maximum stress of the bi-material ITL are provided in [Figure 5](#materials-12-03379-f005){ref-type="fig"}. The Y-direction CTE of the ITL linear decreased when the length ratio increases gradually. This is due to the CTE diminution of the composite rod. When the length ratio of the composite rod increases, the inner NTE of the composite rod will increase. The inner end of the short rod will push the hypotenuse to generate a larger angle of rotation. It generates a contraction along the vertical direction. Thus, the equivalent Y-direction CTE of the ITL linear decreases. In addition, the Y-direction CTE of the ITL is larger than that of the triangular lattice. This because the inner end of the short rod could push the hypotenuse to generate an extra rotation. It indicates the composite rod could enlarge the NTE of the ITL. On the other hand, the maximum stress of the ITL is linear increases when the length ratio increases. This is caused by the large deformation of the rod induced by the large NTE. It can be inferred that the ITL could achieve larger CTE compared with the triangular lattice when their height is equal. 3.2. The CTE Ratio of The Two Constituents {#sec3dot2-materials-12-03379} ------------------------------------------ The CTE Ratio of the two constituents is the primary factors that must be decided at the first stage of the ITL designing. It is of great importance to analyze how the CTE ratio affects the Y-direction CTE and maximum stress of the ITL. CTE and maximum stress of the ITL were calculated when the CTE ratio of the two constituents ranges from 2 to 10. The CTE was calculated by the Y-direction of the ITL. In this case, we used the properties of aluminium alloy and titanium in the numerical simulation. The CTE of the titanium was fixed during the numerical simulation. The desired CTE ratio was achieved by varying the CTE setting of the aluminium alloy. The ratio of L/r was 5; the virtual rod length was 24 mm, the upper base length was 120mm, the bi-material ITL height was 30mm, the width of all the rods was 5 mm, the thickness of all the rods was 5 mm. The results are presented in [Figure 6](#materials-12-03379-f006){ref-type="fig"}. [Figure 6](#materials-12-03379-f006){ref-type="fig"} shows the Y-direction CTE and the maximum stress of the ITLs with different CTE ratio of the two constituents. The Y-direction CTE of the ITLs decreases as the CTE ratio of the two constituents increases. The gradient of the Y-direction CTE of the ITLs also decreases when the CTE ratio increases. The trend the Y-direction CTE of the triangular lattice is the same as that of the ITL. Large CTE ratio will generate large elongation of high CTE constituents. The large elongation will increase the rotation of the hypotenuse to enlarge the NTE. Moreover, the CTE of the ITL is more than two times compared with the CTE of the triangular lattice. This is not caused by the variation of the CTE ratio, but the composite rod. On the other hand, the maximum stress of the ITL increases when the CTE ratio increases. This is because the deformation of each rod increases as the CTE ratio increasing. Extremely large CTE ratio may lead to the maximum stress over the yield stress of the materials. Hence, it is advisable to choose a proper CTE ratio. 3.3. Height of The Lattice {#sec3dot3-materials-12-03379} -------------------------- The height of the lattice was another important parameter. By changing the length of the hypotenuse and keeping the length of others rods constant, the height of the lattice changed. It had the same effect as changing the angle of the hypotenuse. It not only affected the CTE of the lattice, but also affected the stress and equivalent density. The Y-direction CTE and the maximum stress of the ITLs with the height varying from 6 to 30 mm were calculated. The properties of materials used in the simulation are provided in [Table 1](#materials-12-03379-t001){ref-type="table"}. Meanwhile, the ratio of L/r (5), the virtual rod length (24 mm), the upper base length (120 mm), the rod width (5 mm), the thickness of all the rods (5 mm) were fixed. The results are shown in [Figure 7](#materials-12-03379-f007){ref-type="fig"}. The Y-direction CTE and the maximum stress of the ITLs with different height are shown in [Figure 7](#materials-12-03379-f007){ref-type="fig"}. The Y-direction CTE of the ITL and that of the triangular lattice both decrease when the height decreases. This is because the slant angle of the hypotenuse becomes larger if the height decreases. For the same displacement of the hypotenuse end, the larger slant angle, the larger vertical contraction. In addition, the Y-direction CTE of the ITL is over twice compared with that of the triangular lattice. This is induced by the composite rod. On the other hand, the maximum stress of the ITL and that of the triangular lattice both decrease when the height increases. This is because the Y-direction CTE become small. It makes the deformation of all the rods become small. While the maximum stress increases when the height decreases. This is due to the large rod deformation caused by the large CTE. It can be concluded that a reasonable height should be chosen. When the height decreases, the Y-direction CTE of the ITL decreases. Whereas, the equivalent density of the ITL increases. If the height of the lattice decreases to near zero, there will be not enough space for the negative thermal deformation. When the height is less than 6 mm, continuing to diminish the height would not be any good and the maximum stress increases quickly. It must ensure that the maximum stress less than the yield stress of the materials. Through this simulation, the effects of the length ratio, CTE ratio and height are provided. These three parameters can change the NTE of the structure. As the height decreases, the NTE of the ITL increases sharply. However, the angle of the hypotenuse decreases with the height decreases. This will lead to a decrease in the bearing capacity of the ITL and large weight. The NTE increases faster when the CTE ratio range from 1--6, and the growth speed becomes slower and slower after it is greater than 6. The CTE ratio must be decided at the first stage of the ITL designing. It cannot be changed easily in following the process. These two parameters can be found in the traditional triangle lattice. The increase of length ratio makes the NTE almost linear increase. It makes the NTE of lattice larger and the design more flexible. The stress of structure is related to deformation. The greater the NTE, the greater the stress. 4. Experiment and Results {#sec4-materials-12-03379} ========================= In this section, large CTE of the ITL was verified through experiments. Two constituents were selected. They were the titanium (TA2) as constituent 'a' and the aluminium alloy (7075-T6) as constituent 'b'. Over the range 20--200 °C, the average thermal expansions were 8.6 and 23.1 × 10^−6^ K^−1^ for the titanium and aluminium alloy. A combination of the two constituents fabricated the ITL array sample, as shown in [Figure 8](#materials-12-03379-f008){ref-type="fig"}. The sample was the same as the model in the numerical simulation. It contained three X-symmetric ITL array along the Y-direction. The sample was manufactured by electric discharge machining from 5 mm thick plate of the constituents. The structure parameters of the sample are as follow. The ratio of L/r is 5; and the virtual rod length is 24 mm, the upper base length is 120 mm, the height is 30 mm, the rod width is 5 mm, the thickness of all the rods are 5 mm. As shown in [Figure 8](#materials-12-03379-f008){ref-type="fig"}, it employs an interference fit with interference of 20 µm along X-direction. [Figure 8](#materials-12-03379-f008){ref-type="fig"}b provides a photograph of the joint. Screw fastening was used to ensure the reliability of the connection. The displacement was measured between the long rods on the first and third X-symmetric ITLs, as shown in [Figure 9](#materials-12-03379-f009){ref-type="fig"}b. It was measured by the Lenscan 600 (LS600, Nimes, France). It can measure the center thickness of optical elements and air gaps along the optical axis based on low coherence interferometry. Its measurement range was 600 mm, with an accuracy of ±1 µm. Two little lenses were assembled on the center of the long rods, the angle of the lenses could be adjusted by the adjusting screws, as shown in [Figure 9](#materials-12-03379-f009){ref-type="fig"}b. Adjusted the height and position of the adjusting bracket of the laser to make the laser beam light on the lenses vertically. Then, adjusted the angle of the lenses to make the two reflected light spots coincide with the laser beam light. Thus, the light adjustment was completed. The specimen was heated slowly from 20 to 200 ℃ on a homemade hot plate. The specimen and the hot plate were put in an oven (101-0BS, Shanghai, China). The temperature of the hot plate was adjusted by the temperature control switch (HS-618F, Shanghai, China). The temperature of the specimen was measured using thermocouple probes located on the specimen in different positions, as shown in [Figure 9](#materials-12-03379-f009){ref-type="fig"}b. The measurement value of the two thermocouple probes will indicate whether the specimen temperature is uniform or not. First, we set the oven and the hot plate to the target temperature. Second, when the temperature of the oven and the hot plate reach the target temperature, keep them at the target temperature for 30 min. Third, observe the measurement value of the two thermocouple probes. When the difference is less than 0.5 °C, measure the air distance between the two lenses. Then, increase the target temperature by 20 °C, repeat nine times. Finally, the Y-direction CTE of the array was calculated. The results are plotted in [Figure 10](#materials-12-03379-f010){ref-type="fig"}, along with the numerical simulation predictions. The air distance is the distance between the two lenses fixed on the center of the long rods. As shown in [Figure 10](#materials-12-03379-f010){ref-type="fig"}, the air distance is decreasing when the temperature rises. It proves that the ITL array could achieve a Y-direction NTE. The simulation CTE prediction 'α~Si~' is shown as a dotted line, while the experimental CTE 'α~Exp~' is the solid line. The mean NTE of the sample is −74.4 × 10^−6^ K^−1^ from 20--200 °C. The experiment results indicate the ITL could achieve large value of NTE. Moreover, [Figure 10](#materials-12-03379-f010){ref-type="fig"} shows that the NTE of the sample increases gradually with the temperature rising. This because the Young's modulus of the materials will decrease with the temperature rising. It will make the hypotenuse more easily to bending. Thus, the NTE of the sample increases gradually. 5. Discussion {#sec5-materials-12-03379} ============= According to the formula derived by Miller \[[@B21-materials-12-03379]\], the theoretical CTE of the triangular lattice was about −28 × 10^−6^ K^−1^ when the height and the angle of the hypotenuse were the same with the sample (see in [Section S2 of Supplementary Materials](#app1-materials-12-03379){ref-type="app"}). The experiment results showed the ITL based on the composite rod could achieve 2.6 times the value of NTE compared with the traditional triangular lattice. What is more, the factor will reach to 10 times when the length ratio of the L/r was 28 (see in [section S2 of Supplementary Materials](#app1-materials-12-03379){ref-type="app"}). The composite rod has great potential to be utilized to enlarge the magnitude of NTE of ITL. Moreover, the large NTE is independent of the size scale. By addressing the problem of isotropic NTE, connection and machining process, the composite rod and ITL can be used as elements for NTE structure design conveniently. The triangular lattice has been widely used to design various lattices. It could achieve a wide range of CTE and high stiffness. The prime advantage of our structure is utilizing a composite structure to enlarge the NTE when the constituents and height are restricted. Compared with the triangular lattice, the stiffness of the ITL is not enough. We will analyze and improve the stiffness of the ITL in future work. The properties of the materials are not changed during the simulation. However, the Young modulus and the Poisson's ratio could be temperature and stress-dependent. This will make the simulation results are inconsistent with the experimental results. The results in [Figure 10](#materials-12-03379-f010){ref-type="fig"} prove this different. Thus, the influences of temperature and stress on the properties of the materials should be considered in future work. Following are considerations about future improving the measurement accuracy. The residual stress of the sample has an important effect on the deformation measurement. During the machining process, we find that the remaining part of the plate after machining having a buckling deformation. The buckling deformation of the residual part of the plate indicates that there is residual stress on the sample. When the sample is heated, the residual stress on the specimen will redistribute. This will generate extra deformation and influence the measurement results. Heat treatment could obviously decrease the residual stress. Hence, we decrease the measurement error thought heat treatment.The temperature homogeneity is another factor which influent the thermal deformation of the sample. When the sample is heated, the temperature of the part, which is close to the heat source, is higher compared with the other part. In the experiment, to obtain an approximate homogeneous temperature, we used the thermocouple probes to monitor the temperature of the sample in different positions. It can only provide the temperature of the measurement point. This is not consistent with the homogeneous temperature hypothesis. It will generate a large measurement error. In our future work, we will use the thermal imager to monitor the temperature of the sample. It will provide a global temperature distribution of the sample. This will decrease the temperature differences of the different part of the sample, and improve the accuracy of the measuring results.The air refractive index also affects the measurement results. It needs the air refractive index to calculate the optical distance of the two parallel lenses. During the measurement, the air refractive index is set as a constant, but it changes when the temperature increases actually. It should be considered in order to obtain a more accurate results in our future work. 6. Conclusions {#sec6-materials-12-03379} ============== In this paper, we reported an ITL based on the composite rod. The inverted trapezoid structure realized that the inner NTE of the composite rod converted to an extra vertical NTE of the ITL. The magnitude of NTE of the proposed lattice was analyzed numerically and experimentally. The finite element simulations showed that larger NTE could be achieved by increasing the length ratio, increasing the CTE ratio of the two constituents, or decreasing the height of the ITL. The NTE of ITL was better compared with the traditional triangular lattice. It provided an explicit guideline for designing an ITL with large NTE. The ITL specimen was fabricated by the electric discharge machining. The CTE measurement results by the laser interferometry successfully demonstrated its large NTE. NTE of the ITL is 2.6 times, compared with the traditional triangular lattice. Theoretically, the ITL could improve the NTE by order of magnitude if the length ratio of the composite rod was large enough. This structure is especially useful for temperature-sensitive sensors with high-sensitivity and thermal actuator with a large stroke. The authors would like to thank Kai Zhang, Renxiu Han, and Tian Lan for their support. The authors also would like to thank Qiang Xu of National University of Defense and Technology Microsystem Laboratory for his support. The following are available online at <https://www.mdpi.com/1996-1944/12/20/3379/s1>, Section S1: Thermal expansion of the composite rod, Section S2: Thermal expansion calculation. ###### Click here for additional data file. Conceptualization and methodology, W.L., S.X. and C.Z.; finite element simulation, W.L.; experiment and investigation, W.L. and S.X.; writing---original draft preparation, W.L.; writing---review and editing, S.X., M.Z. and G.L.; project administration, G.L. This research received no external funding. The authors declare no conflict of interest. {#materials-12-03379-f001} {#materials-12-03379-f002} {#materials-12-03379-f003} {ref-type="fig"} (**e**).](materials-12-03379-g004){#materials-12-03379-f004} {#materials-12-03379-f005} {#materials-12-03379-f006} {#materials-12-03379-f007} {#materials-12-03379-f008} {#materials-12-03379-f009} {#materials-12-03379-f010} materials-12-03379-t001_Table 1 ###### Material properties for the aluminum alloy \[[@B43-materials-12-03379]\], and titanium \[[@B44-materials-12-03379]\]. ------------------------------------------------------------------------------------------ Material Young's Modulus E (GPa) Coefficient of Thermal Expansion α (×10^−6^ K^−1^) ----------- ------------------------- ---------------------------------------------------- Al alloy\ 72 23.1 7075-T6 Ti\ 107.9 8.6 TA2 ------------------------------------------------------------------------------------------

1. Introduction {#sec1-polymers-09-00708} =============== Electrochromism refers to the alternation of optical absorption or color of an electroactive species by electrochemically induced redox reactions \[[@B1-polymers-09-00708]\]. Materials capable of such color switching have found use in many applications such as electronic papers \[[@B2-polymers-09-00708],[@B3-polymers-09-00708]\], automatic-dimming mirrors, smart windows \[[@B4-polymers-09-00708]\], adaptive camouflage \[[@B5-polymers-09-00708],[@B6-polymers-09-00708]\], eyewear \[[@B7-polymers-09-00708],[@B8-polymers-09-00708]\], and energy storage devices \[[@B9-polymers-09-00708],[@B10-polymers-09-00708]\]. A number of organic and inorganic materials have been used to construct electrochromic devices, such as transition metal oxides, inorganic coordination complexes, organic dyes and polymers, and organic-metallic hybrid polymers \[[@B11-polymers-09-00708],[@B12-polymers-09-00708],[@B13-polymers-09-00708],[@B14-polymers-09-00708],[@B15-polymers-09-00708]\]. Among the different types of electrochromic materials, organic polymers attract much interest because of several advantages such as mechanical flexibility, enhanced processability, easy color tuning, rapid switching and high coloration efficiency \[[@B16-polymers-09-00708],[@B17-polymers-09-00708],[@B18-polymers-09-00708],[@B19-polymers-09-00708],[@B20-polymers-09-00708]\]. Triphenylamine (TPA) is a promising core unit for opto- and electro-active materials because of the facile formation of radical cation and its charge transport ability. Many triphenylamine and triarylamine starbursts, dendrimers, and polymers have been used as electron-blocking and hole-transporting layers in organic electronic devices such as organic light-emitting diodes (OLEDs), organic field-effect transistors (OFETs), and dye-sensitized solar cells \[[@B21-polymers-09-00708],[@B22-polymers-09-00708],[@B23-polymers-09-00708],[@B24-polymers-09-00708],[@B25-polymers-09-00708]\]. Since 2005, Liou's and our groups have extensively studied non-conjugated triarylamine-based polymers for electrochromic applications \[[@B26-polymers-09-00708],[@B27-polymers-09-00708],[@B28-polymers-09-00708],[@B29-polymers-09-00708],[@B30-polymers-09-00708],[@B31-polymers-09-00708],[@B32-polymers-09-00708],[@B33-polymers-09-00708],[@B34-polymers-09-00708]\]. We prepared and characterized various structurally different polyimides and polyamides with triarylamine units and aromatic or non-aromatic spacers in the polymer backbone. The obtained polymers are generally soluble in most of the polar aprotic solvents allowing preparing amorphous thin films via solution casting. In general, these films show strong coloration change upon oxidation and exhibit high optical contrast and cycling stability. The introduction of electron-donating substituents such as methoxy groups into the electrochromic unit in the reactive sites could decrease the oxidation potential and allow the formation of more stable triarylamine radical cations \[[@B35-polymers-09-00708]\]. Thus, a further improvement on the electrochromic performance of the polymers could be obtained using 4,4′-diamino-4′′-methoxytriphenylamine (MeOTPA-(NH~2~)~2~) as a functional diamine monomer \[[@B35-polymers-09-00708]\]. Unsubstituted TPA can be readily electrochemically oxidized to a TPA radical cation. The formed radical cation dimerizes and upon abstraction of two protons yields tetraphenylbenzidine (TPB) \[[@B36-polymers-09-00708],[@B37-polymers-09-00708]\]. The electrochemical coupling reaction of TPA has been used as a convenient and effective method to prepare electroactive polymers for optoelectronic applications \[[@B38-polymers-09-00708],[@B39-polymers-09-00708],[@B40-polymers-09-00708],[@B41-polymers-09-00708],[@B42-polymers-09-00708],[@B43-polymers-09-00708]\]. Systematic studies on the electrochemical oxidative coupling of carbazole have been reported by Ambrose and Nelson \[[@B44-polymers-09-00708],[@B45-polymers-09-00708]\]. By applying oxidative potential, carbazole radical cations are generated that undergo coupling to 3,3′-biscarbazole. Many redox-active and electrochromic polymers have been synthesized via the cabazole-based oxidative coupling polymerization \[[@B46-polymers-09-00708],[@B47-polymers-09-00708],[@B48-polymers-09-00708],[@B49-polymers-09-00708],[@B50-polymers-09-00708],[@B51-polymers-09-00708],[@B52-polymers-09-00708]\]. Additionally, several functional, microporous thin films with high surface areas have been produced by this approach \[[@B53-polymers-09-00708]\]. Electrochemical polymerization allows for simultaneous polymer synthesis and film deposition in short times and under mild and metal- or catalyst-free conditions. In this work, two novel electropolymerizable monomers MeOTPA-(TPA)~2~ and MeOTPA-(NPC)~2~ were synthesized from the condensation reactions of 4,4′-diamino-4′′-methoxytriphenylamine (MeOTPA-(NH~2~)~2~) with 4-carboxytriphenylamine (TPA-COOH) and *N*-(4-carboxyphenyl)carbazole (NPC-COOH), respectively. Electroactive and electrochromic poly(amide-amine) films could be easily built-up on the electrode surface via TPA- or carbazole-based electrochemical oxidative coupling. In comparison with other monomers without redox-active centers \[[@B54-polymers-09-00708]\], the present MeOTPA-centered monomers would result in poly(amide-amine)s with an additional, low-oxidation-potential MeOTPA redox center. Thus, the prepared electroactive polymer films are expected to exhibit lowered onset operating voltages when they are used as active layers in electrochromic devices. Better coloration efficiency and multiple electrochromic switching are also anticipated. 2. Results and Discussion {#sec2-polymers-09-00708} ========================= 2.1. Monomer Synthesis {#sec2dot1-polymers-09-00708} ---------------------- The MeOTPA-cored diamide monomers MeOTPA-(TPA)~2~ and MeOTPA-(NPC)~2~ and model compound MeOTPA-(Ph)~2~ were prepared by condensation of MeOTPA-(NH~2~)~2~ with 4-carboxytriphenylamine, *N*-(4-carboxyphenyl)carbazole, and benzoic acid, respectively, using triphenyl phosphite (TPP) and pyridine as condensing agents. Synthetic details and characterization data are included in the [Supplementary Materials](#app1-polymers-09-00708){ref-type="app"}. The synthetic route is shown in [Scheme 1](#polymers-09-00708-sch001){ref-type="scheme"}, and all the target compounds were characterized by FTIR and ^1^H NMR spectroscopy. [Figure S1 (Supplementary Materials)](#app1-polymers-09-00708){ref-type="app"} illustrates FT-IR spectra of all the synthesized compounds. The nitro groups of MeOTPA-(NO~2~)~2~ show the characteristic absorptions at 1575 and 1318 cm^−1^ (--NO~2~ asymmetric and symmetric stretching). After reduction, the characteristic absorptions of nitro group disappear and MeOTPA-(NH~2~)~2~ shows the typical --NH~2~ stretching absorption pair at 3334 and 3216 cm^−1^. All of the synthesized compounds MeOTPA-(Ph)~2~, MeOTPA-(TPA)~2~ and MeOTPA-(NPC)~2~ show the characteristic amide absorption bands at around 3267--3279 cm^−1^ (N--H stretching) and 1635--1650 cm^−1^ (amide C=O stretching). The ^1^H and H--H correlation spectroscopy (COSY) NMR spectra of MeOTPA-(TPA)~2~, MeOTPA-(NPC)~2~ and MeOTPA-(Ph)~2~ illustrated in [Figure 1](#polymers-09-00708-f001){ref-type="fig"}, [Figure 2](#polymers-09-00708-f002){ref-type="fig"} and [Figure S2 (Supplementary Materials)](#app1-polymers-09-00708){ref-type="app"} agree well with the proposed molecular structures. In addition, mass spectroscopy results of monomers MeOTPA-(TPA)~2~ and MeOTPA-(NPC)~2~ are consistent with their calculated values ([Figure S3, Supplementary Materials](#app1-polymers-09-00708){ref-type="app"}). 2.2. Electrochemical and Spectroelectrochemical Properties of Model Compound {#sec2dot2-polymers-09-00708} ---------------------------------------------------------------------------- The electrochemical behavior of a solution 2 × 10^−4^ M model compound (MeO)~2~TPPA-(Ph)~2~ in 0.1 M Bu~4~NClO~4~/CH~2~Cl~2~ electrolyte was examined by cyclic voltammetry (CV) at 50 mV·s^−1^ scan rate. As shown in [Figure 3](#polymers-09-00708-f003){ref-type="fig"}, MeOTPA-(Ph)~2~ showed one pair of reversible redox waves with an oxidation peak at around 0.75 V (*E*~1/2~ = 0.71 V) in the first CV scan. The reversible wave can be attributed to the oxidation process occurring in the MeOTPA core. This model compound showed very stable electrochemical oxidation processes in the first 50 repetitive CV scans between 0 and 0.9 V. [Figure 4](#polymers-09-00708-f004){ref-type="fig"} shows the optical absorption and color changes of MeOTPA-(Ph)~2~ on a Pt grid in an OTTLE cell. In the neutral form, the compound exhibited a strong absorption at λ~max~ = 336 nm. When the applied potential was gradually raised from 0 to 1.0 V, the absorbance at 336 nm slightly decreased and new bands appeared at 390, 656 and 789 nm. Meanwhile, the color changed to turquoise. The spectral and color changes are apparently attributed to the formation of cationic radical species of this compound caused by oxidation of the MeOTPA core. 2.3. Electrochemical Polymerization of Monomers {#sec2dot3-polymers-09-00708} ----------------------------------------------- [Figure 5](#polymers-09-00708-f005){ref-type="fig"}a presents cyclic voltammetry (CV) scans of 2 × 10^−4^ M MeOTPA-(TPA)~2~ in 0.1 M Bu~4~NClO~4~/CH~2~Cl~2~ at a scan rate of 50 mV·s^−1^ between 0 and 1.2 V. After a ten cycles scan, no new peaks were detected. When the voltage was scanned between 0 and 1.4 V, two oxidation peaks at 0.71 and 1.26 V were observed in the first scan. New oxidation peaks appeared at about 0.95 and 1.08 V in the subsequent scan, indicating the occurrence of the coupling reactions between terminal TPA cationic radicals forming the tetraphenylbenzidine (TPB) segments. Upon successive scanning, the redox wave current intensities gradually increased, which indicates the formation of the electrochemically active polymeric film (coded as P1, [Scheme 2](#polymers-09-00708-sch002){ref-type="scheme"}) on the electrode surface ([Figure 5](#polymers-09-00708-f005){ref-type="fig"}b). Similarly, the NPC-based monomer MeOTPA-(NPC)~2~ did not undergo electropolymerization when the potential was scanned between 0 and 1.2 V ([Figure 6](#polymers-09-00708-f006){ref-type="fig"}a). As the potential was scanned between 0 and 1.4 V, the carbazole groups started to oxidize and coupling reactions occurred forming the biscarbazole units. As the CV scanning continued, the redox currents intensified ([Figure 6](#polymers-09-00708-f006){ref-type="fig"}b) and a polymer film (coded as P2) gradually deposited on the electrode surface. As a typical example, the IR spectra of MeOTPA-(TPA)~2~ and its electro-generated P1 film are shown in [Figure 7](#polymers-09-00708-f007){ref-type="fig"}. The P1 film shows the characteristic amide absorption bands at around 3285 cm^−1^ (N--H stretching) and 1654 cm^−1^ (amide C=O stretching). As compared to that of monomers, the P1 film shows very similar IR absorptions in the range of 1200--1700 cm^−1^ being associated with amide C=O, aromatic ring and C−N stretch. The change in the low frequency region 500--1000 cm^−1^, peculiar to benzene ring C--H out-of-plane bending, can be attributed to the TPA coupling reaction. 2.4. Optical and Electrochemical Properties of Polymers {#sec2dot4-polymers-09-00708} ------------------------------------------------------- The UV-vis absorption spectra of monomers MeOTPA-(TPA)~2~ and MeOTPA-(NPC)~2~ in CH~2~Cl~2~ and their derived polymer films of P1 and P2 on an ITO-glass substrate are shown in [Figure S4 (Supplementary Materials)](#app1-polymers-09-00708){ref-type="app"}. The spectra of the monomers show absorption bands with maximum peaks at 350 and 336 nm and absorption onsets at 408 and 426 nm, respectively. The polymer films of P1 and P2 show absorption maxima at 361 and 338 nm and absorption onsets at 454 and 449 nm, respectively. The red-shift of absorption maximum and onset of the polymer films compared to the monomers implies an extended π-conjugation length. The electrochemical behavior of the electro-synthesized polymer films on the ITO-glass substrate were investigated by cyclic voltammetry in a 0.1 M Bu~4~NClO~4~/CH~2~Cl~2~ solution without the monomer. The optical and electrochemical properties of the polymers are summarized in [Table 1](#polymers-09-00708-t001){ref-type="table"}. As illustrated in [Figure 8](#polymers-09-00708-f008){ref-type="fig"}, the polymer films exhibited three reversible redox couples, at the half-wave potentials (*E*~1/2~) of 0.61, 0.93 and 1.04 V for P1 from MeOTPA-(TPA)~2~, and 0.62, 0.80 and 1.07 V for P2 from MeOTPA-(NPC)~2~. The oxidation onset potential (*E*~onset~) of P1 and P2 were recorded at 0.47 and 0.49 V (vs. Ag/AgCl), respectively. Redox waves at lower potentials of these two polymers are apparently attributable to reversible electrochemical oxidation processes of the MeOTPA center. The energy levels of the highest occupied molecular orbital (HOMO) and lowest unoccupied molecular orbital (LUMO) of the corresponding polymers estimated from the *E*~onset~ values and UV-vis absorption edge are also included in [Table 1](#polymers-09-00708-t001){ref-type="table"}. 2.5. Spectroelectrochemical Properties of Polymers {#sec2dot5-polymers-09-00708} -------------------------------------------------- Spectroelectrochemical behaviors of the polymeric films on ITO glass were investigated. [Figure 9](#polymers-09-00708-f009){ref-type="fig"} illustrates the changes in the electronic absorption spectrum of the P1 film upon changing the applied potentials. The electronic absorption profile of neutral form of the film exhibits absorption maximum at around 362 nm, which is due to π−π\* transitions for P1. When the applied potential was gradually raised to 0.8 V, the absorption intensity at 362 nm slightly dropped and new band appeared at 775 nm, and the color changed to pale green (L\*: 77; a\*: −9; b\*: 13), which could be attributed to the formation of MeOTPA radical cation. Upon further oxidation at applied potential to 1.1 V, a new absorption peak at 502 nm was observed and the film changed into an orange color (L\*: 50; a\*: 17; b\*: 15). The spectral change should be originated from the tetraphenylbenzidine radical cation (TPB+•) moieties. When the applied potential was adjusted to 1.3 V, the absorption peak at 502 nm decreased, with a formation of new absorption band at around 716 nm and a color change to blue (L\*: 21; a\*: 6; b\*: −26). This result indicated the formation of TPB dicationic species. A similar spectral change was observed for the P2 film at the early stage of the oxidation ([Figure 10](#polymers-09-00708-f010){ref-type="fig"}). Upon oxidation at 0.8 V, an absorption band at 341 nm started to decrease and a new band at 779 nm gradually increased in intensity, and the film changed into a pale green color (L\*: 65; a\*: −8; b\*: 7). When the applied voltage was increased to 1.3 V, absorption band at 779 nm decreased in intensity and new absorption bands intensified at 550 nm (L\*: 39; a\*: 22; b\*: −4). 2.6. Electrochromic Switching {#sec2dot6-polymers-09-00708} ----------------------------- Double potential step chronoamperometry was carried out to estimate the response time and cycling stability of the electro-generated films. The potential was stepped between neutral and oxidized states of the polymer film with a suitable residence time. During the potential switching, the transmittance percentage (% *T*) at the wavelength of maximum contrast was measured using a UV-vis spectrophotometer. The measurements were conducted using a single wavelength with obvious absorption intensity change. As shown in [Figure 11](#polymers-09-00708-f011){ref-type="fig"}a, for P1 optical contrast (Δ% *T*) and switching time (green coloring, at 90% full-transmittance change) were measured as 30% and 6 s for 775 nm by stepping the potential between 0 and 0.87 V with a residence time of 14 s. As the applied voltage was stepped from 0 to 1.12 V, the film exhibited Δ% *T* of 41% at 509 nm for the oxidized orange state which required 10.7 s for the coloring step and 4.9 s for the bleaching step as shown in [Figure 11](#polymers-09-00708-f011){ref-type="fig"}b. When the applied voltage was switched between 0 and 1.65 V, the film required 11.9 s for the coloring step and 9.5 s for the bleaching step, and Δ% *T* was 77% at 716 nm for the blue coloring ([Figure 11](#polymers-09-00708-f011){ref-type="fig"}c). Coloration efficiency (*CE*) is another important parameter for the electrochromic applications. On the basis of the equation *CE* = Δ*OD*/*Q*~d~, where Δ*OD* is the optical absorbance change, and *Q*~d~ (mC/cm^2^) is the inject/ejected charge during a redox step ([Supplementary Materials Figures S5 to S9](#app1-polymers-09-00708){ref-type="app"}). The *CE* values of P1 were estimated to be 73 cm^2^/C at 775 nm, 57 cm^2^/C at 509 nm and 65 cm^2^/C at 716 nm ([Table 2](#polymers-09-00708-t002){ref-type="table"}). The switching stability of P2 is illustrated in [Figure 12](#polymers-09-00708-f012){ref-type="fig"} at 779 and 550 nm with a pulse time of 13 and 16 s, respectively, for 20 cycles. The P2 film exhibited 6 s switching time at 779 nm between its neutral and the first oxidized state. The *CE* value for the green coloring was estimated as 123 cm^2^/C at 779 nm. The switching time and Δ% *T* values for the P2 film were measured as 7.1 s and 33%, respectively, while stepping the potential between 0 and 1.39 V at 550 nm. The high electrochromic switching stability of these two polymers at the first oxidation state could be attributed to their high electrochemical stability of the MeOTPA core. 2.7. Electrochromic Devices {#sec2dot7-polymers-09-00708} --------------------------- Based on the above-mentioned results, it can be concluded that the electro-generated polymers may find use in the construction of electrochromic devices (ECD) and optical display. Therefore, we fabricated single layer electrochromic cells as preliminary investigations. The schematic illustration of the structure of the ECD is shown in [Figure 13](#polymers-09-00708-f013){ref-type="fig"}a. The ECD using P1 as active layer was fabricated ([Figure 13](#polymers-09-00708-f013){ref-type="fig"}b). The ECD is colorless in its neutral state, and it slowly switched to pale green, maroon, and blue color at an increasing oxidation voltage of 1.8, 2.1, and 2.5 V, respectively. 3. Conclusions {#sec3-polymers-09-00708} ============== Two triphenylamine-based compounds MeOTPA-(TPA)**~2~** and MeOTPA-(NPC)**~2~** composed of a central MeOTPA core connected through amide linkers to terminal triphenylamine or *N*-phenylcarbazole units were developed as electroactive monomers to prepare electrochromic poly(amide-amine) films on the ITO electrode surface through electropolymerization. The electro-generated polymer films exhibited reversible redox processes and multi-colored electrochromic behaviors upon electro-oxidation. In general, the electrochromic films possessed moderate coloration efficiency and good cycling stability. The electrochemically synthesized polymers might be promising candidates for electrochromic materials. This work was financially supported by the Ministry of Science and Technology, Taiwan (Grant MOST 105-2221-E-027-131-MY3). ###### Click here for additional data file. Synthetic details and characterization data of the synthesized compounds are available online at [www.mdpi.com/2073-4360/9/12/708/s1](www.mdpi.com/2073-4360/9/12/708/s1). Sheng-Huei Hsiao planned this project. Hsing-Yi Lu performed the experiments. Hsing-Yi Lu and Sheng-Huei Hsiao analyzed the data and co-wrote the paper. The authors declare no conflict of interest. Figures, Schemes and Tables =========================== {#polymers-09-00708-f001} {#polymers-09-00708-f002} {#polymers-09-00708-sch001} {#polymers-09-00708-f003} {#polymers-09-00708-f004} {#polymers-09-00708-f005} {#polymers-09-00708-f006} {#polymers-09-00708-f007} {#polymers-09-00708-sch002} {#polymers-09-00708-f008} {#polymers-09-00708-f009} {#polymers-09-00708-f010} ###### Percentage transmittance changes of the cast film of P1 on the ITO-glass slide (coated area \~1 cm^2^) (in CH~2~Cl~2~ with 0.1 M Bu~4~NClO~4~ as the supporting electrolyte) by applying a potential at (**a**) 0.0 ↹ 0.87 V for 50 cycles, (**b**) 0.0 ↹ 1.12 V for 40 cycles, and (**c**) 0.0 ↹ 1.65 V for 20 cycles.   {#polymers-09-00708-f012} {#polymers-09-00708-f013} polymers-09-00708-t001_Table 1 ###### Optical and electrochemical properties of the electrochemically synthesized polymers. Polymers UV-Vis Absorption (nm) ^a^ Oxidation Potential (V) ^b^ *E*~g~ (eV) ^c^ HOMO (eV) ^d^ LUMO (eV) ^d^ ---------- ---------------------------- ----------------------------- ----------------- --------------- --------------- ------ ------ ------ ------ P1 361 454 0.47 0.61 0.93 1.04 2.73 4.83 2.10 P2 338 449 0.49 0.62 0.80 1.07 2.76 4.85 2.09 ^a^ UV-vis absorption maximum and onset wavelengths for the polymer films. ^b^ Calculated from first cyclic voltammetry (CV) scans, versus Ag/AgCl in CH~2~Cl~2~ at a scan rate of 50 mV·s^--1^. ^c^ Optical bandgap calculated from absorption edge of the polymer film: *E*~g~ = 1240/λ~onset~. ^d^ The highest occupied molecular orbital (HOMO) and lowest unoccupied molecular orbital (LUMO) energy levels were calculated from *E*~onset~ values of CV diagrams and were referenced to ferrocene (4.8 eV relative to vacuum energy level; *E*~onset~ = 0.44 V; *E*~1/2~ = 0.52 V in CH~2~Cl~2~). EHOMO = *E*~onset~ + 4.8 − 0.44 (eV); ELUMO = EHOMO − *E*g. polymers-09-00708-t002_Table 2 ###### Electrochromic properties of the polymer films. Polymers λ~max~ ^a^ (nm) Δ% *T* Response Time ^b^ Δ*OD* ^c^ *Q*~d~ ^d^ (mC/cm^2^) *CE* ^e^ (cm^2^/C) ---------- ----------------- -------- ------------------- ----------- ----------------------- -------------------- ----- P1 775 30 6.0 3.0 0.16 2.21 73 502 41 10.7 4.9 0.26 4.56 57 716 77 11.9 9.5 0.68 10.39 65 P2 779 36 6.0 5.8 0.38 3.09 123 550 33 7.1 6.4 0.18 4.25 42 ^a^ Wavelength of absorption maximum. ^b^ Time for 90% of the full-transmittance change. ^c^ Optical density change (Δ*OD*) = log (*T*~bleached~/*T*~colored~), where *T*~colored~ and *T*~bleached~ are the maximum transmittance in the oxidized and neutral states, respectively. ^d^ *Q*~d~ is ejected charge, determined from the in situ experiments. ^e^ Coloration efficiency (*CE*) = Δ*OD*/*Q*~d~.

INTRODUCTION {#s1} ============ Ionizing radiation causes irreversible molecular changes (i.e. damage) to DNA. Damage is caused by direct ionization or reactions of hydroxyl radicals (^•^OH) from water radiolysis, and includes strand breaks and base lesions. Double-strand breaks (DSBs), consisting of two or more single-strand breaks (SSBs) in close proximity produced in genomic DNA, are thought to be harmful lesions that can cause serious biological effects, such as cell death. Moreover, base lesions can cause cell mutations that may contribute to carcinogenesis. Production of these DNA lesions can be suppressed by some chemical compounds. Dimethyl sulfoxide (DMSO), which is used to limit the formation of DNA lesions under irradiation, acts as an ^•^OH scavenger \[[@RRU079C1], [@RRU079C2]\]. Antioxidants, such as catechins and ascorbic acid, also reduce the formation of DNA lesions. High reactivity between antioxidants and oxidative radicals has been reported by several researchers \[[@RRU079C3]--[@RRU079C10]\]. In addition to acting as radical scavengers, some antioxidants are reported to repair the precursors of DNA lesions (DNA radicals) by chemical reactions \[[@RRU079C11]--[@RRU079C16]\]. This repair process is called 'chemical repair'. As DNA radicals generally have longer lifetimes than ^•^OH in a cell or under cell-mimetic conditions \[[@RRU079C17]--[@RRU079C19]\], chemical repair should be a more efficient radioprotection process than radical scavenging. In a previous report, the authors investigated the chemical repair activity of ascorbic acid and found that the precursors of abasic sites (AP sites) and other base lesions were more easily repaired than those of strand breaks \[[@RRU079C16]\]. Approximately 50--60% of these lesions were chemically repaired by ascorbic acid. On the other hand, \<10% of ^•^OH produced from radiolysis was scavenged by the ascorbic acid concentrations under the experimental conditions. The results showed that the chemical repair activity of ascorbic acid was more efficient than its ^•^OH scavenging activity. It was also suggested that ascorbic acid might be suitable as an agent to reduce mutations. Edaravone (Radicut®, 3-methyl-1-phenyl-2-pyrazolin-5-one; Fig. [1](#RRU079F1){ref-type="fig"}), which is an artificial antioxidant with low toxicity, is generally known as a neuroprotective drug, and it has been developed and clinically used in Japan since 2001. Its medical and pharmacological effects on brain infarctions and cardiovascular diseases have been confirmed by numerous *in vivo* experiments \[[@RRU079C20]--[@RRU079C27]\]. The pharmacological effects of edaravone have mainly been attributed to its reactivity with free radicals, such as ^•^OH and peroxide radicals \[[@RRU079C28]--[@RRU079C31]\]. Because edaravone has recently received much attention for use as a radioprotector \[[@RRU079C32], [@RRU079C33]\], the authors have investigated its reactions with ^•^OH and other oxidative radicals by pulse radiolysis and determined a pattern of reaction for edaravone as an antioxidant, together with the rate constants of the reactions \[[@RRU079C29], [@RRU079C34]\]. Much information about the radical scavenging properties of edaravone was revealed by these experiments. However, to understand the radioprotective effects of edaravone, other effective repair processes, such as chemical repair properties, must also be investigated. In the present study, pulse radiolysis using a DNA model compound, deoxyguanosine monophosphate (dGMP), was conducted to observe the basic reactions of chemical repair and reactivity, and was compared with that of ascorbic acid. Furthermore, to investigate the chemical repair activity of edaravone on DNA lesions, yields of lesions on plasmid DNA in gamma ray-irradiated aqueous solutions containing several concentrations of edaravone were measured by a method recently developed by the authors \[[@RRU079C16]\]. Fig. 1.Chemical structure of edaravone in water. MATERIALS AND METHODS {#s2} ===================== Samples {#s2a} ------- Edaravone (3-methyl-1-phenyl-2-pyrazolin-5-one), Tris, boric acid, ethylenediaminetetraacetic acid (EDTA)·2Na·2H~2~O, sodium hydrate (NaOH), 1% bromophenol blue, Orange G, NaH~2~PO~4~ and Na~2~HPO~4~ were purchased from Wako Pure Chemical Industries, Ltd (Osaka, Japan). Ascorbic acid and deoxyguanosine monophosphate (dGMP) were purchased from Sigma--Aldrich (St Louis, MO). Formamidopyrimidine-DNA glycosylase (Fpg), endonuclease III (Nth), endonuclease IV (Nfo), and reaction buffer solutions used for the treatment of these enzymes were purchased from New England BioLabs Inc. (Ipswich, MA). Stock solutions of the enzymes (concentrations of 10 U μl^−1^ for Nth and Nfo, and 8 U μl^−1^ for Fpg) and the reaction buffers were stored at −20°C. Nth protein mainly excises ring-saturated pyrimidines (e.g. 5,6-dihydrothymine \[DHT\]), thymine glycol, and AP sites. Fpg protein excises mainly 2,6-diamino-4-hydroxy-5-*N*-methyl formamidopyrimidine, 7,8-dihydro-8-oxo-2′-deoxyguanine (8-oxoGua) and AP sites. Nfo protein excises various AP sites \[[@RRU079C35], [@RRU079C36]\]. Water (\>18 MΩ) purified by Milli-Q system (Millipore, Billerica, MA) was used to make all aqueous solutions. Phosphate buffer solutions (pH 6.9) were prepared from NaH~2~PO~4~. 2H~2~O and Na~2~HPO~4~. 12H~2~O. TBE buffer solutions consisted of 8.9 × 10^−2^ mol dm^−3^ Tris, 8.9 × 10^−2^ mol dm^−3^ boric acid, and 2 × 10^−3^ mol dm^−3^ EDTA. The loading buffer solutions for electrophoresis consisted of 2% (w/v) g dm^−1^, bromophenol blue, 2% (w/v) Orange G, 30% (w/v) glycerol, and 1 × 10^−2^ mol dm^−3^ EDTA. The pH of the TBE buffer and the loading buffer were adjusted to 8·0 with 1N NaOH. The deionized water and buffer solutions were sterilized at 120°C for 20 min before sample preparation. Pulse radiolysis {#s2b} ---------------- Pulse radiolysis experiments were carried out at The University of Tokyo using an electron linear accelerator (energy: 35 MeV; pulse width: 10 ns). Details of the experimental setup are described elsewhere \[[@RRU079C29], [@RRU079C34]\]. All irradiated solutions were adjusted to pH 6.9 by 1.0 × 10^−2^ mol dm^−3^ phosphate buffer. Since the sample solutions containing antioxidants could be spontaneously oxidized in an aerated solution, they were freshly prepared and quickly deaerated using N~2~O gas for 15 min. Dosimetry was carried out using an N~2~O-saturated 1 × 10^−2^ mol dm^−3^ KSCN aqueous solution, taking *Gϵ* as 5.2 × 10^−4^ m^2^ J^−1^ at 475 nm \[[@RRU079C37]\]. An absorbed dose of 9 Gy per pulse was used in this experiment. Electron beam irradiation causes water radiolysis, generating many kinds of reactive species. To observe the reactions of chemicals with ^•^OH, other radical species need to be quenched. N~2~O gas is often used to scavenge hydrated electrons, one of the main products of water radiolysis, and convert them to ^•^OH \[[@RRU079C38]\]. The scheme for water radiolysis and the reaction of N~2~O have been described previously \[[@RRU079C34]\]. The conversion of hydrated electrons to ^•^OH is completed within 10 ns. Under these experimental conditions, hydrated electrons react predominantly with N~2~O because the concentration and the rate constant of N~2~O are both much higher than those of the antioxidants \[[@RRU079C39]\]. The yield of ^•^OH has been estimated at 0.59 μmol J^−1^ \[[@RRU079C40]\]. Gamma ray irradiation of DNA solutions {#s2c} -------------------------------------- Plasmid DNA pUC18 (2686 bp) was obtained from *E. coli* JM109 by Qiagen HiSpeed Plasmid Kit (Qiagen, Hilden, Germany) and purified by dialysis using a nitrocellulose membrane of pore size 0.025 μm (Millipore, Billerica, MA). After dialysis, it was confirmed by agarose electrophoresis that \>90% of the plasmids were intact (closed circular form). Plasmids were stored at −20°C at a concentration of 1.88 × 10^−1^ g dm^−3^ dissolved in 2.0 × 10^−2^ mol dm^−3^ phosphate buffer solution (pH 6.9). DNA stock solutions were diluted with 2.0 × 10^−2^ mol dm^−3^ phosphate buffer solution (pH 6.9) to a concentration of 1.0 × 10^−2^ g dm^−3^ for irradiation samples. Edaravone was added to the solutions just before irradiation. The concentrations of edaravone were adjusted to 0, 10, 100 and 1000 μmol dm^−3^. As the number of electrons of the solutes was less than that of water in a unit volume, over 99% of the radiation energy was believed to be deposited onto the water molecules under these conditions. Thus, DNA damage was predominantly induced by free radicals, such as ^•^OH, produced from water radiolysis. Irradiation conditions were the same as in the previous experiment \[[@RRU079C16]\]. For irradiation, 100 μdm^3^ samples were placed into 1.5-ml polypropylene tubes (Eppendorf, Hamburg, Germany). Each tube was placed into 20 ml of icy water maintained at 4°C during irradiation and exposed to ^60^Co gamma rays under aerobic conditions using a gamma ray apparatus (The University of Tokyo, Tokyo, Japan). The dose rate, measured by a Fricke dosimeter, ranged from 0.05--3.2 Gy min^−1^, depending on the position of samples. Each sample was irradiated at least three times. After irradiation the lesions were detected as strand breaks by electrophoresis. To detect base lesions and AP sites, irradiated samples were treated with Nth, Fpg, and Nfo before electrophoresis. The details have been described previously \[[@RRU079C16]\]. The dose response of the residual closed-circular DNA was determined from the logarithmic loss. A *D*~37~ value, which represents the dose of radiation required to give on average one SSB per plasmid molecule, was calculated from the slope of the response curve. The chemical yield of the SSBs (*G*~ssb~) is represented using the value of *D*~37~ as follows: $$G_{ssb} = {\frac{\lbrack\,\text{plasmid}\rbrack}{\rho D_{37}},}$$ where \[plasmid\] is the molar concentration of plasmid DNA molecules in an irradiated solution, and *ρ* is the density of the solution (∼1.0 kg dm^−3^). Yields of base lesions revealed by the enzymatic treatments, *G*~enzyme\ (=\ Nth,\ Fpg\ or\ Nfo)~ were determined by subtracting the yield of SSBs in the incubated samples without the enzyme, *G*~enz−~, from that in the solution incubated with the enzyme, *G*~enz+~. $$G_{enzyme} = G_{{enz} +} - G_{{enz} -}.$$ RESULTS {#s3} ======= Reactivity of antioxidants with dGMP radicals {#s3a} --------------------------------------------- Transient absorption spectra (Fig. [2](#RRU079F2){ref-type="fig"}) were obtained by pulse radiolysis of N~2~O-saturated solution containing 2.0 × 10^−3^ mol dm^−3^ dGMP and 1.0 × 10^−4^ mol dm^−3^ edaravone at pH 6.9. Under these conditions, it is calculated that 94% of ^•^OH was scavenged by dGMP based on a competition reaction between dGMP \[[@RRU079C41]\] and edaravone \[[@RRU079C29]\]. Therefore the spectrum obtained 1 μs after pulse irradiation is due to the absorption spectrum of dGMP hydroxyl radical adducts \[[@RRU079C42]\]. According to a previous report, this broad spectrum consists of absorption bands of two different dGMP hydroxyl radical adducts, G^•^8-OH (∼300 nm) and G^•^4-OH (∼600 nm) \[[@RRU079C43]\]. Absorption spectrum of dGMP hydroxyl radical adducts obtained 80 μs after the pulse irradiation by pulse radiolysis of the N~2~O-saturated solution of 2.0 × 10^−3^ mol dm^−3^ dGMP without edaravone is also shown in the inset of Fig. [2](#RRU079F2){ref-type="fig"}. This spectrum showed that these dGMP radicals were relatively stable in the solution without edaravone. From the time profiles in Fig. [3](#RRU079F3){ref-type="fig"} and the spectrum in the inset of Fig. [2](#RRU079F2){ref-type="fig"}, it is clear that the decay rate of absorbance at 320 and 600 nm depended on the concentrations of edaravone. These profiles show that these dGMP hydroxyl radical adducts were scavenged by edaravone. Rate constants of the reactions were found to be 4.1 × 10^8^ dm^3^ mol^−1^ s^−1^ and 4.3 × 10^8^ dm^3^ mol^−1^ s^−1^ from the absorbance decay at 320 and 600 nm. On the other hand, the absorbance at 350 nm almost remained unchanged by the addition of edaravone (Fig. [3](#RRU079F3){ref-type="fig"}b). It seems that the edaravone radical, the absorption peak of which is ∼350 nm, was produced by these reactions (see Discussion). Fig. 2.Transient absorption spectra obtained 1 μs (closed circles) and 80 μs (open squares) after irradiation by pulse radiolysis of a N~2~O-saturated aqueous solution containing 2 × 10^−3^ mol dm^−3^ dGMP, 1.0 × 10^−4^ mol dm^−3^ edaravone, and 2.0 × 10^−2^ mol dm^−3^ phosphate buffer at pH 6.9. Inset: the transient spectrum obtained 80 μs after irradiation by pulse radiolysis of a N~2~O-saturated aqueous solution containing 2 × 10^−3^ mol dm^−3^ dGMP. Fig. 3.Time profiles of absorbance at (**a**) 320, (**b**) 350 and (**c**) 600 nm of the solutions containing several concentrations of edaravone and 2 × 10^−3^ mol dm^−3^ dGMP at pH 6.9. Transient absorption spectra of a N~2~O-saturated solution containing 2.0 × 10^−3^ mol dm^−3^ dGMP and 1.0 × 10^−4^ mol dm^−3^ ascorbic acid determined 1 and 80 μs after the pulse irradiation were also obtained to compare with that of edaravone (Fig. [4](#RRU079F4){ref-type="fig"}). Under the experimental conditions, the absorption spectrum of dGMP hydroxyl radical adducts was observed 1 μs after pulse irradiation because it was calculated that \>97% of ^•^OH was scavenged by dGMP based on a competition reaction between dGMP and ascorbic acid. The rate constant of the reaction between ascorbic acid and ^•^OH, (4.5 ± 0.3) × 10^9^ dm^3^ mol^−1^ s^−1^, which was determined in this study by measuring the buildup rate of ascorbate radical after pulse irradiation of N~2~O-saturated solutions containing 0.2--1.0 × 10^−3^ mol dm^−3^ ascorbic acid at pH 6.9, was used for the competition reaction. The absorbance derived from dGMP hydroxyl radical adducts decreased, with an increase in a strong absorption band at 360 nm. This spectral alteration, which was also reported previously by O\'Neill \[[@RRU079C42]\], is believed to be due to electron transfer reactions of dGMP hydroxyl radical adducts and ascorbic acid, resulting in the production of an ascorbate radical. The rate constants of the decay at 320 and 600 nm were found to be 4.1 × 10^8^ dm^3^ mol^−1^ s^−1^ and 4.2 × 10^8^ dm^3^ mol^−1^ s^−1^, respectively. These values are similar to the reported value measured by O\'Neill, (4.8 ± 0.5) × 10^8^ dm^3^ mol^−1^ s^−1^ and are almost the same as those of the reactions of edaravone and dGMP hydroxyl radical adducts. Fig. 4.Transient absorption spectra obtained 1 μs (closed circles) and 80 μs (open squares) after irradiation by pulse radiolysis of a N~2~O-saturated aqueous solution containing 2 × 10^−3^ mol dm^−3^ dGMP, 1.0 × 10^−4^ mol dm^−3^ ascorbic acid, and 2.0 × 10^−2^ mol dm^−3^ phosphate buffer at pH 6.9. Gamma ray-irradiation of plasmid DNA solutions containing edaravone {#s3b} ------------------------------------------------------------------- The semi-log plots of the dose--response curves of fractions of closed-circular forms in gamma ray-irradiated pUC18 plasmid DNA (1.0 × 10^−2^ g dm^−3^) aqueous solutions, with or without 100 μmol dm^−3^ edaravone, are shown in Fig. [5](#RRU079F5){ref-type="fig"}. The fractions of closed circular forms of DNA in the solutions incubated at 37°C for 1 h, with or without Nth, are also plotted. The dose--response curve of closed-circular forms in DNA incubated without Nth was slightly steeper than that of closed-circular forms in DNA without incubation, regardless of whether ascorbic acid was added (see the circle and square symbols in Fig. [5](#RRU079F5){ref-type="fig"}). This indicates that heat treatment during incubation induced damage in heat labile sites \[[@RRU079C44]\]. On the other hand, closed-circular forms in DNA incubated with Nth decreased drastically. This indicates that a lot of Nth-sensitive sites are produced in irradiated DNA. The results were similar to the results of the previous experiments in DNA solutions containing ascorbic acid \[[@RRU079C16]\]. Nth-sensitive lesions and strand breaks, which are produced by irradiation, were suppressed by edaravone. The chemical yields of prompt SSBs and Nth-, Fpg- and Nfo-sensitive sites (*G*~SSB~, *G*~Nth~, *G*~Fpg~, *G*~Nfo~) in the absence and presence of 10, 100 and 1000 μmol dm^−3^ edaravone were determined (Table [1](#RRU079TB1){ref-type="table"}). The ratio of the yield of each lesion produced in the presence of edaravone to the yield produced in the absence of edaravone (*G*~0~) was calculated for each concentration (shown in Fig. [6](#RRU079F6){ref-type="fig"} as '*G*/*G*~0~'). The *G*/*G*~0~ ratio decreased with increasing concentrations of edaravone. At the lower concentrations (10 and 100 μmol dm^−3^), the production of Nfo-sensitive sites (AP sites) has a tendency to be more suppressed than that of SSBs. The Nth- and Fpg-sensitive sites (AP sites and other base lesions) also had lower ratios than those of prompt SSBs, though the margin of error for the experimental data was large. Table 1.Chemical yields (*G* values) of strand breaks and of Nth-, Fpg- and Nfo-sensitive sites, and the fraction of repairable precursors of DNA lesions (*p*)\[Eda\] (μmol dm^−3^)*p*0101001000*G* (10^−3^ μmol J^−1^)Nth-sensitive lesions4.7 ± 1.21.5 ± 0.80.29 ± 0.050.056 ± 0.0030.36 ± 0.13Fpg-sensitive lesions7.3 ± 2.22.5 ± 0.60.43 ± 0.150.090 ± 0.0140.35 ± 0.13Nfo-sensitive lesions2.9 ± 0.50.67 ± 0.210.12 ± 0.060.031 ± 0.0070.56 ± 13SSBs6.4 ± 2.22.9 ± 0.34.3 ± 0.80.055 ± 0.0130.13 ± 0.09 Fig. 5.Dose--response curves of the remaining fraction of closed circular plasmid DNA in 2.0 × 10^−2^ mol dm^−3^ phosphate buffer (pH 7) aqueous solution of 1.0 × 10^−2^ g dm^−3^ pUC18 plasmid DNA with 100 μmol dm^−3^ edaravone (open symbols). The results of our previous study in DNA solution without the addition of antioxidants are also plotted (closed symbols) \[[@RRU079C11]\]. Circles: DNA directly analyzed without any chemical treatments after irradiation; diamonds: DNA incubated with Nth treatment after irradiation; squares: DNA incubated without Nth treatment after irradiation. Fig. 6.Ratios of the chemical yields of prompt SSBs and of Nth-, Fpg- and Nfo-sensitive sites obtained in the presence of edaravone at several concentrations (*G*) to the yields obtained without edaravone (*G*~0~). According to our previous paper on ascorbic acid, the enzyme-sensitive lesions with lower *G*/*G*~0~ ratios indicate a higher level of chemical repair activity of ascorbic acid at AP sites and other base lesions. The chemical repair efficiency was estimated by the parameter, *p*, which represents the fraction of chemically repairable yields in all detected lesions, obtained by a simple model of a competition reaction (16). The value of *p* represents the proportion of lesions potentially repairable by an antioxidant. Enzyme-sensitive lesions were more easily repaired by edaravone than SSBs (Fig. [6](#RRU079F6){ref-type="fig"}), although the error margins of the *G*/*G*~0~ ratios of Fpg- and Nth-sensitive lesions were rather large. The value of *p* for edaravone was also determined for each lesion (Table [1](#RRU079TB1){ref-type="table"}). The value of *p* for Nfo-sensitive sites was 0.56 ± 0.13, indicating that about half of the precursors of AP sites could be chemically repaired by edaravone. On the other hand, the *p* values for edaravone to Nth- and Fpg-sensitive sites were ∼0.3, showing that ∼30% of the precursors of these lesions could be chemically repaired. Thus, edaravone can chemically repair precursors of AP sites and other base lesions more easily than those of SSBs. As these enzyme-sensitive lesions can cause cell mutations, this result suggests that edaravone may have a potentially beneficial role in the suppression of cell mutations under irradiation. As mentioned above, two types of dGMP hydroxyl radical adducts, G^•^8-OH and G^•^4-OH, are mainly produced through the reaction of guanine and ^•^OH. These radicals are known to be precursors of 8oxoG, which is an Fpg-sensitive lesion. In this gamma ray irradiation experiment, the inhibition of Fpg-sensitive lesions was observed by adding edaravone. This result was consistent with the results of our pulse radiolysis experiments, in which G^•^8-OH and G^•^4-OH were scavenged by the antioxidants. DISCUSSION {#s4} ========== Reduction reaction of edaravone {#s4a} ------------------------------- The decay rate of absorbance at 320 and 600 nm depends on the concentrations of edaravone (Fig. [3](#RRU079F3){ref-type="fig"}), which indicates that the decay of the dGMP hydroxyl radical adducts is attributable to the reactions with edaravone. However, unlike the result for ascorbic acid (Fig. [4](#RRU079F4){ref-type="fig"}), the definite absorption bands of edaravone radicals that should be produced as a result of the reactions were not observed. In our previous report, one-electron oxidized edaravone radicals, the molecular coefficient of which at 345 nm was (2.60 ± 0.10) × 10^3^ dm^3^ mol^−1^ cm^−1^ \[[@RRU079C29]\], were produced by reactions of edaravone with oxidizing radicals. The molecular coefficient of the edaravone radicals is almost the same as that of dGMP hydroxyl radical adducts, which is calculated from Fig. [2](#RRU079F2){ref-type="fig"} using the yield of ^•^OH in N~2~O-saturated solutions, 0.59 μmol J^−1^ \[[@RRU079C40]\], as the yield of dGMP hydroxyl radical adducts. Therefore, it was presumed these dGMP hydroxyl radical adducts were reduced by edaravone through electron transfer reactions, from which a one-electron oxidized edaravone radical was produced. This pattern of reaction is similar to that of ascorbic acid. Pulse radiolysis experiments showed that edaravone can reduce dGMP hydroxyl radical adducts through electron transfer reactions with high rate constants in the same manner as ascorbic acid. Comparison of chemical repair activity of edaravone with that of ascorbic acid {#s4b} ------------------------------------------------------------------------------ In our previous report, 50--60% of the precursors of the enzyme-sensitive lesions were chemically repaired by ascorbic acid \[[@RRU079C16]\]. The results of the present experiment show that the chemical repair activity of edaravone on enzyme-sensitive lesions is lower than that of ascorbic acid. On the other hand, the values of *p* for both edaravone and ascorbic acid to SSBs were almost zero, indicating that SSBs are seldom repaired by either edaravone or ascorbic acid. As discussed above, the pattern of reaction of edaravone to dGMP hydroxyl radical adducts and the rate constants were similar to those of ascorbic acid. However, the chemical repair activity of edaravone against the precursors of Fpg-sensitive lesions, which include 8oxoG, was relatively low. Several reasons can be suggested to explain this difference. First, the reactivity of edaravone with other precursors of Fpg-sensitive lesions might be different from that of ascorbic acid. This could be verified by pulse radiolysis experiments using other DNA model compounds, such as those used in previous experiments performed by Shi *et al*. \[[@RRU079C45]\]. Another reason could be the difference in accessibility to the precursors of lesions on DNA, which is attributable to the structures of the antioxidants. Measuring the chemical repair activity of derivatives of antioxidants might be a useful way to investigate their accessibility in terms of structure--activity relationships. Some sulfhydryl compounds, such as glutathione, are also known to show chemical repair activity via hydrogen donation reactions \[[@RRU079C12], [@RRU079C41]\]. Applying the methods used in this study to sulfhydryl compounds might help improve our understanding of their chemical repair activity and provide some information about the differences observed in the chemical repair activities between antioxidants and sulfhydryl compounds. CONCLUSIONS {#s5} =========== The conclusions obtained from the pulse radiolysis experiment in dGMP solutions and the gamma ray irradiation experiment in plasmid DNA solutions are summarized as follows. \(i\) Edaravone reduced dGMP hydroxyl radical adducts through electron transfer reactions, from which a one-electron oxidized edaravone radical was produced. The pattern of reaction and the rate constants were similar to those of ascorbic acid. \(ii\) By adding edaravone into plasmid DNA solutions before gamma ray irradiation, Nth-, Fpg- and Nfo-sensitive lesions were suppressed. This phenomenon was attributed to the chemical repair reactions of edaravone. However, the suppression of SSBs was not effective. \(iii\) The efficiency of the chemical repair reaction of edaravone on Nth- and Fpg-sensitive lesions was relatively lower than that of ascorbic acid. Further studies will be required to elucidate the reasons for this difference. FUNDING {#s6} ======= Funding to pay the Open Access publication charges for this article was provided by The University of Tokyo. The authors are grateful to Mr Hiroishi of The University of Tokyo for assistance in operating the ^60^Co gamma ray equipment. The authors also thank Dr Fujii of the Japan Atomic Energy Agency for technical support to prepare the DNA samples. [^1]: Present address: Nuclear Safety Research Center, Japan Atomic Energy Agency, 2--4 Shirakatashirane, Tokai-mura, Naka-gun, Ibaraki 319--1195, Japan

1. Introduction {#sec1-ijerph-13-00257} =============== Hypertension has been a persistent public health challenge due to its high prevalence rate \[[@B1-ijerph-13-00257]\]. The global prevalence rate of hypertension has been increasing for the past two decades \[[@B2-ijerph-13-00257]\]. Hypertension accounts for 3.7% of total disability adjusted life years (DALYS) \[[@B3-ijerph-13-00257]\]. Overweight and obesity are well-known independent risk factors for the incidence of hypertension \[[@B1-ijerph-13-00257],[@B4-ijerph-13-00257],[@B5-ijerph-13-00257],[@B6-ijerph-13-00257]\]. The prevalence of overweight and obesity is increasing worldwide and in China \[[@B7-ijerph-13-00257],[@B8-ijerph-13-00257]\]. Body mass index (BMI), waist circumference (WC), waist:hip ratio (WHR) and waist:stature ratio (WSR) are four indicators commonly used to measure the degree of obesity \[[@B4-ijerph-13-00257],[@B5-ijerph-13-00257]\]. BMI appears to be sufficient for it is collected more often in nutrition and health surveys, and it is collected with universally accepted protocols compared to WC, WHR and WSR \[[@B9-ijerph-13-00257]\]. Evidence has indicated that BMI is the best predictor of incident hypertension \[[@B10-ijerph-13-00257]\]. Previous studies showed that hypertension risk increased with a higher BMI \[[@B5-ijerph-13-00257],[@B11-ijerph-13-00257],[@B12-ijerph-13-00257]\]. However, the association between change in BMI and the incidence of hypertension is still unclear in the Chinese population, and it may be impacted by changes in lifestyle or targeted intervention. Widely used among Western populations, the BMI cut-off value of 25 kg/m^2^ for overweight has been recommended by the World Health Organization (WHO) as an international criterion \[[@B11-ijerph-13-00257]\]. Country-specific and ethnicity-specific BMI cut-off values are suggested, for it is not appropriate to use a single cut-off value to define overweight for all populations due to the potential ethnic variation in body build and composition \[[@B13-ijerph-13-00257],[@B14-ijerph-13-00257]\]. The cut-off level for overweight (24.0 kg/m^2^) has been developed using BMI for the general Chinese population by the Working Group on Obesity in China (WGOC) \[[@B15-ijerph-13-00257],[@B16-ijerph-13-00257],[@B17-ijerph-13-00257]\], and according to the standard of WS/T 428--2013 (China), the optimal BMI cut-off value is 24 kg/m^2^ for Chinese adults. The impact of change in overweight status based on this cut-off value on the incidence of hypertension has not been well characterized. The aim of this study is to investigate change in BMI and its impact on the incidence of hypertension in 18- to 65-year-old Chinese adults utilizing data from the China Health and Nutrition Survey (CHNS). 2. Subjects and Methods {#sec2-ijerph-13-00257} ======================= 2.1. Study Sample {#sec2dot1-ijerph-13-00257} ----------------- Two waves of data collected in 2006 and 2011 by the CHNS with samples drawn from nine provinces in China (Guangxi, Guizhou, Hunan, Hubei, Henan, Jiangsu, Shandong, Liaoning and Heilongjiang) are used in this study. The CHNS is an ongoing study started in 1989, and it examined the influence of social and economic transformation among Chinese communities and society on the nutritional and health status of the Chinese population. Detailed descriptions of the study have been presented elsewhere \[[@B18-ijerph-13-00257]\]. Questionnaires and data sets can be downloaded in the following link \[[@B19-ijerph-13-00257]\]. This research was approved by the Institutional Review committees of the University of North Carolina at Chapel Hill, as well as by the National Institute for Nutrition and Health, Chinese Center for Disease Control and Prevention. All participants provided written informed consent for their participation in the survey. Of the 5370 participants who were involved in both surveys in 2006 and 2011, whose ages were 18 to 65 in 2006, and who were non-pregnant or non-lactating women in 2006 and 2011 or who were men, 3989 (74%) had complete and plausible anthropometric measurements and measurements of blood pressure. These included, for example, a baseline BMI of 15--44 kg/m^2^; 5-year changes in height \<10 cm and in BMI \< 10 kg/m^2^; the difference between any two of the three measurements of systolic blood pressures (SBP) in each survey \<10 mm·Hg, and the difference between any two of the three measurements of diastolic blood pressures (DBP) in each survey \<10 mm·Hg \[[@B11-ijerph-13-00257]\]. Of the 3989 participants, 3215 (81%) with normal blood pressure in the survey in 2006 were included in our sample. The exclusion of samples with implausible or extreme values in blood pressure or anthropometric measures helped us increase the estimate precision. 2.2. Measures {#sec2dot2-ijerph-13-00257} ------------- All participants underwent standardized examinations including collections of SBP and DBP. Three blood pressure measurements were taken on the right arm by trained health workers, and all of the workers followed the standardized procedure using regularly calibrated mercury sphygmomanometers with appropriate-sized cuffs. SBP was measured at the first appearance of a pulse sound (Korotkoff phase 1) and DBP was measured at the disappearance of the pulse sound (Korotkoff phase 5). Three measurements of SBP or DBP were averaged to reduce the effect of measurement error. Hypertension was defined as SBP ≥ 140 mm Hg, DBP ≥ 90 mm Hg or being diagnosed by a doctor previously \[[@B9-ijerph-13-00257]\]. Cumulative incidence was calculated by dividing new hypertension cases over the study period by the total at-risk population in 2006. Anthropometric measures were taken after the participants had removed their shoes and heavy clothing. Height and weight were measured by the trained workers. BMI (kg/m^2^) was calculated based on weight and height which were measured by trained workers who followed standardized procedures. The cut-off level for overweight (24.0 kg/m^2^) has been developed using BMI for the general Chinese population by the WGOC \[[@B15-ijerph-13-00257],[@B16-ijerph-13-00257],[@B17-ijerph-13-00257]\], and according to the standard of WS/T 428--2013 (China), the optimal BMI cut-off value is 24 kg/m^2^ for Chinese adults. Overweight was defined as BMI ≥ 24 kg/m^2^, and non-overweight was defined as BMI \< 24 kg/m^2^. Covariates such as sex, age, smoking habits, alcohol consumption, and place of residence were collected by direct interviews. 2.3. Statistical Analysis {#sec2dot3-ijerph-13-00257} ------------------------- The difference value (D-value) in BMI between 2006 and 2011 was calculated to evaluate the change of BMI from 2006 to 2011. D-value \> 0 meant that the BMI increased from 2006 to 2011; the greater the D-value was, the more the BMI increased. D-value \< 0 meant that the BMI decreased from 2006 to 2011; the lower the D-value was, the more the BMI decreased. Firstly, the participants were divided into two groups according to BMI at baseline: overweight and non-overweight groups. The mean and standard deviation (SD) for continuous variables and percentage for categorical variables were calculated and compared between overweight and non-overweight subjects. The two-tail independent *t* test or chi-square test were used to test differences between overweight subjects and non-overweight subjects in the means or proportions of baseline variables such as age, SBP, DBP, BMI, place of residence, smoking habits, alcohol consumption and incidence of hypertension. The logistic regression model was used to examine the association between change in BMI and incidence of hypertension, odds ratio (OR) and 95% confident interval (95% CI) were calculated. All potential confounding factors in the logistic regression analyses were collected at baseline. Models were initially adjusted for the baseline age, BMI, SBP and DBP (model 1). To observe the influence of lifestyle factors, models were additionally adjusted for smoking habits, alcohol consumption and place of residence at baseline (model 2). The incidence of hypertension stratified by overweight status in 2006 and 2011 was investigated by logistic regression analysis. Additionally, the association between change in SBP and DBP and change in overweight status between 2006 and 2011 was investigated by multivariable linear regression analysis. All reported *p*-values were two-tailed, and those less than 0.05 were considered statistically significant. All statistical analyses were conducted using SAS 9.2 (SAS Institute, Cary, NC, USA). 2.4. Ethics Statement {#sec2dot4-ijerph-13-00257} --------------------- This research has been approved by the Institutional Review committees of the University of North Carolina at Chapel Hill and the National Institute for Nutrition and Health, Chinese Center for Disease Control and Prevention (Ethical Review No: 201524). All participants gave written informed consent for their participation in the survey. 3. Results {#sec3-ijerph-13-00257} ========== In total, 3215 participants (1441 men and 1774 women) were studied, including 606 (18.8%) subjects (312 men and 294 women) who developed hypertension in 2011. At baseline, there were 974 non-overweight and 467 overweight men, and 1146 non-overweight and 628 overweight women. In 2011, there were 845 non-overweight and 596 overweight men, and 1030 non-overweight and 744 overweight women. The baseline characteristics of participants stratified by sex are shown in [Table 1](#ijerph-13-00257-t001){ref-type="table"}. At baseline, the mean SBP, DBP and BMI were higher among overweight men compared with non-overweight men (*p* \< 0.05). The levels of age, SBP, DBP and BMI were higher among overweight women compared with non-overweight women (*p* \< 0.05). The proportions of subjects who live in urban areas were higher in overweight men than in non-overweight men. [Table 2](#ijerph-13-00257-t002){ref-type="table"} shows the association between incidence of hypertension and the BMI D-value. Compared with men with a BMI reduction of 0.4783 kg/m^2^ or more (*i.e.*, the lowest quartile of the BMI D-value), men with a mild reduction or gain (the second quartile of the BMI D-value of −0.4783 to 0.4920 kg/m^2^) had an increased risk for developing hypertension (OR and 95% CI: 1.448, 0.979--2.141, respectively) (model 2, [Table 2](#ijerph-13-00257-t002){ref-type="table"}). The risk for the third BMI D-value quartile rose to 1.138 (95% CI, 0.764--1.696). The risk for the highest BMI D-value quartile rose to 2.303 (95% CI, 1.560--3.401). Compared with women with a BMI reduction of 0.5953 kg/m^2^ or more (*i.e.*, the lowest quartile of the BMI D-value), women with a mild reduction or gain (the second quartile of the BMI D-value of −0.5953 to 0.4732 kg/m^2^) had an increased risk for developing hypertension (OR and 95% CI: 1.095, 0.743--1.613, respectively). The risk for the third BMI D-value quartile rose to 1.152 (95% CI, 0.778--1.706). The risk for the highest BMI D-value quartile rose to 1.745 (95% CI, 1.199--2.540). There was a significant trend of increased incidence of hypertension with an increase in the quartile of the BMI D-value in men (*p* for trend \< 0.001) and women (*p* for trend = 0.004). The participants were further stratified by overweight status in 2006 and in 2011. Compared with non-overweight subjects in 2011, overweight was associated with a 1.469-fold and 1.580-fold increase in the risk of incident hypertension in non-overweight subjects and overweight subjects at baseline, after adjustment for important covariates (*p* \< 0.05) ([Table 3](#ijerph-13-00257-t003){ref-type="table"}). Overweight subjects in 2011 had a significantly higher increase in SBP and DBP compared with non-overweight subjects in 2011 regardless of their overweight status at baseline, after adjustment for important covariates (*p* \< 0.05) ([Table 4](#ijerph-13-00257-t004){ref-type="table"}). 4. Discussion {#sec4-ijerph-13-00257} ============= We measured change in BMI and its impact on the incidence of hypertension, and the association between changes in overweight status based on the BMI cut-off value recommended by the WGOC and according to the standard of WS/T 428--2013 (China) and the incidence of hypertension in 18- to 65-year-old Chinese adults. The results of this study showed that increased BMI was positively associated with the incidence of hypertension after adjustment for some important potential confounding factors, both in men and women. Our findings implied that change in BMI could affect the change of blood pressure (BP), and reducing BMI by modifying lifestyle could prevent and control incidence of hypertension. The association between BMI at baseline and the incidence of hypertension during follow-up has been analyzed in previous studies \[[@B5-ijerph-13-00257],[@B11-ijerph-13-00257],[@B12-ijerph-13-00257]\]. Overweight status might change due to lifestyle modification. The change of BMI may be mainly caused by lifestyle change or targeted intervention. Several studies have reported the impact of such a change on hypertension risk \[[@B1-ijerph-13-00257],[@B20-ijerph-13-00257],[@B21-ijerph-13-00257],[@B22-ijerph-13-00257],[@B23-ijerph-13-00257],[@B24-ijerph-13-00257]\]. A meta-analysis of randomized controlled trials showed that weight loss was very important for the prevention and treatment of hypertension \[[@B20-ijerph-13-00257]\]. BMI changes were independently associated with blood pressure changes for men and women in Tromso's Study in Norway \[[@B21-ijerph-13-00257]\]. Weight loss was found to reduce the risk for hypertension in Huang's study in a cohort of US females \[[@B22-ijerph-13-00257]\]. A prospective study in Japan found that weight gain increased the risk of developing hypertension among Japanese men and women \[[@B23-ijerph-13-00257]\]. Change in BMI and its impact on the incidence of hypertension in the Chinese should be detected because there are potential ethnic variations in body composition and build, and potential ethnic variations in the health risks associated with obesity among populations \[[@B14-ijerph-13-00257]\]. However, there is controversy regarding whether change in BMI impacts the incidence of hypertension in the Chinese \[[@B1-ijerph-13-00257],[@B24-ijerph-13-00257]\]. In Chen's study in a Taiwan community in China, BMI gain is associated with the risk of developing hypertension in men, but the risk in women is influenced by their menopausal status \[[@B1-ijerph-13-00257]\]. In Luo's study in Jiangsu province in China, change in BMI was associated with the incidence of hypertension in both men and women regardless of the women's menopausal status \[[@B24-ijerph-13-00257]\]. Furthermore, abnormal BMI (overweight or obesity) was defined as BMI ≥ 25 kg/m^2^ in Luo's study \[[@B24-ijerph-13-00257]\], and it is the BMI cut-off value for defining overweight recommended by the WHO \[[@B14-ijerph-13-00257]\], but this cut-off value may not be appropriate for Asians \[[@B25-ijerph-13-00257]\], because Asians have a higher body fat percentage than Caucasians at the same BMI level \[[@B26-ijerph-13-00257]\]. There are potential ethnic variations in body composition and build, and variations in the health risk associated with overweight among populations \[[@B14-ijerph-13-00257]\]. The cut-off level for overweight (24.0 kg/m^2^) has been developed using BMI for the general Chinese population by the WGOC \[[@B15-ijerph-13-00257],[@B16-ijerph-13-00257],[@B17-ijerph-13-00257]\], and according to the standard of WS/T 428--2013 (China), the optimal BMI cut-off value is 24 kg/m^2^ for Chinese adults. Whether the risk of incident hypertension of overweight subjects would decrease if they reduce their BMI to less than 24 kg/m^2^ and whether the risk of incident hypertension of subjects with BMI \< 24 kg/m^2^ would increase if their BMI becomes higher than 24 kg/m^2^ should be examined. We studied the relationship between five years' change in BMI and the incidence of hypertension in 18- to 65-year-old Chinese adults with 3215 participants. In our study, out of 2120 individuals whose BMI was less than 24 kg/m^2^ at baseline, 415 persons developed overweight in 2011; out of 1095 overweight individuals at baseline, 170 persons were corrected to less than 24 kg/m^2^ in 2011. There was a significant trend of increased incidence of hypertension with an increase in the quartile of the BMI D-value in men (*p* for trend \< 0.001) and women (*p* for trend = 0.004). During the five-year follow-up, the numbers of new cases of hypertension among people in the highest quartiles of the BMI D-value were different from those in the lowest quartiles in men and women (*p* \< 0.05). The participants were further stratified by overweight status in 2006 and in 2011. Compared with non-overweight subjects in 2011, the ORs of incident hypertension were all significantly higher for overweight subjects, regardless of their overweight status at baseline (*p* \< 0.05). These results indicate that the incident hypertension risk of overweight subjects would decrease if they reduce their BMI to less than 24 kg/m^2^ and that the incident hypertension risk of subjects with a BMI \< 24 kg/m^2^ would increase if they become overweight. Weight loss is a critical lifestyle modification for overweight and obese people to control hypertension. Although the exact mechanism of the effect of weight loss on BP is not clear, there are several plausible biologic pathways. First of all, the renin-angiotensin-aldosterone system is overactivated in obese subjects; meanwhile, aldosterone concentrations and renin activity are higher in obese subjects than in lean subjects \[[@B20-ijerph-13-00257]\]. Secondly, activity of the sympathetic nervous system is increased in hypertensive and obese subjects, and this might induce obesity-related renal effects. Alternatively, there may be inhibition of the natriuretic peptides system, of which the functional results are vasodilatation and natriuresis \[[@B20-ijerph-13-00257]\]. Thirdly, hyperinsulinemia and decreased insulin sensitivity as part of the metabolic syndrome may also form a necessary link between obesity and hypertension, although the interrelation is still not fully known \[[@B20-ijerph-13-00257]\]. Some limitations merit consideration in this study. First of all, data on lipid markers such as levels of low-density lipoprotein cholesterol and triglycerides, data on plasma glucose and data on family history of hypertension are not available in our study, so these variables were not adjusted in our analysis. Secondly, diet and physical activity are important factors for the incidence of hypertension, but these factors were not included in this analysis. Further studies are needed to include these factors in analyses. Thirdly, we cannot determine the time order between the change in BMI and the incidence of hypertension, and prospective studies are needed to determine the causal relationship between change in BMI and the incidence of hypertension. In addition, BMI, WC, WHR and WSR are four indices commonly used to measure the degree of obesity \[[@B4-ijerph-13-00257],[@B5-ijerph-13-00257]\]. In some studies, WC was shown to be the best predictor of hypertension \[[@B27-ijerph-13-00257],[@B28-ijerph-13-00257]\], and in some other studies, WSR was shown to be the best predictor of the incidence of hypertension \[[@B29-ijerph-13-00257],[@B30-ijerph-13-00257]\]. Further studies should assess the association between change in these obesity indices and the incidence of hypertension. 5. Conclusions {#sec5-ijerph-13-00257} ============== In conclusion, the results from this study provide unequivocal evidence that weight loss makes a crucial contribution to the treatment of hypertension. Prevention of weight gain is likely to have a great impact on the incidence of hypertension in Chinese adults. Intervention programs designed to reduce BMI through lifestyle modification may have significant public health significance in reducing the risk of incident hypertension. This work was supported by research grant from National Natural Science Foundation of China (81172666), the NIH (R01-HD30880, DK056350, and R01-HD38700) and the Fogarty International Center. The NIH provided financial support for CHNS data collection from 1989 to 2011 and in future surveys. Qian Ren, Chang Su, Huijun Wang and Bing Zhang conceived the study. Chang Su, Huijun Wang, Zhihong Wang, Wenwen Du and Bing Zhang participated in the data collection. Qian Ren completed all statistical analyses, interpretation of data and drafted the manuscript. Zhihong Wang and Wenwen Du contributed to the discussion. Huijun Wang and Bing Zhang revised the manuscript. The authors declare no conflict of interest. The founding sponsors had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, and in the decision to publish the results. The following abbreviations are used in this manuscript: BMIbody mass indexCHNSChina Health and Nutrition SurveyORodds ratio95% CI95% confident intervalD-valuedifference valueDALYSdisability adjusted life yearsWCwaist circumferenceWHRwaist:hip ratioWSRwaist:stature ratioWHOWorld Health OrganizationWGOCWorking Group on Obesity in ChinaSBPsystolic blood pressureDBPdiastolic blood pressureSDstandard deviationBPblood pressure ijerph-13-00257-t001_Table 1 ###### Baseline characteristics of 3215 study participants ^a^. Variables Men (*n* 1441) Women (*n* 1774) ------------------------------- ---------------- ------------------ -------------- ------------------ Age (years) 45.1 ± 11.4 46.0 ± 10.2 44.6 ± 10.4 47.7 ± 9.1 ^b^ SBP (mm·Hg) 115.5 ± 10.5 119.4 ± 9.2 ^b^ 110.8 ± 11.7 117.0 ± 10.4 ^b^ DBP (mm·Hg) 75.5 ± 7.3 78.8 ± 6.1 ^b^ 72.7 ± 7.8 76.4 ± 7.1 ^b^ BMI (kg/m^2^) 21.2 ± 1.7 26.1 ± 1.8 ^b^ 21.2 ± 1.8 26.4 ± 2.1 ^b^ Urban residence (%) 27.5 33.4 ^b^ 27.8 29.5 Ever smoked cigarettes (%) 66.5 63.2 2.6 3.0 Alcohol drinker (%) 61.2 63.4 9.0 8.8 Incidence of hypertension (%) 16.6 32.1^b^ 12.3 24.4 ^b^ Notes: SBP, systolic blood pressure; DBP, diastolic blood pressure; BMI, body mass index. ^a^ Values are means ± standard deviation or percentages; ^b^ Significantly greater odds, *p* \< 0.05. ijerph-13-00257-t002_Table 2 ###### The incidence of hypertension associated with BMI D-value in men and women. Quartiles of BMI D-Value Range (kg/m^2^) Hypertension Cases (*n*) No. of Subjects (*n*) OR (95% CI) -------------------------- ----------------- -------------------------- ----------------------- -------------------------- -------------------------- Men (*n* 1441) Quartile 1 \<−0.4783 75 360 1.000 1.000 Quartile 2 −0.4783--0.4920 77 360 1.460 (0.989--2.156) 1.448 (0.979--2.141) Quartile 3 0.4921--1.6977 68 361 1.165 (0.783--1.731) 1.138 (0.764--1.696) Quartile 4 ≥1.6978 92 360 2.289 (1.552--3.376) ^c^ 2.303 (1.560--3.401) ^c^ *p* for trend \<0.001 0.004 Women (*n* 1774) Quartile 1 \<−0.5953 73 443 1.000 1.000 Quartile 2 −0.5953--0.4732 68 444 1.091 (0.741--1.606) 1.095 (0.743--1.613) Quartile 3 0.4733--1.5054 66 444 1.157 (0.783--1.711) 1.152 (0.778--1.706) Quartile 4 ≥1.5055 87 443 1.755 (1.207--2.552) ^c^ 1.745 (1.199--2.540) ^c^ *p* for trend \<0.001 0.004 Notes: BMI, body mass index; D-value, difference value between follow-up and baseline; OR, odds ratio; CI, confidence interval. ^a^ Logistic regression models were used, adjusting for age, BMI, SBP and DBP at baseline; ^b^ Logistic regression models were used, adjusting for variables in model 1 as well as smoking habits, alcohol consumption and place of residence at baseline; ^c^ Significantly greater odds, *p* \< 0.05. ijerph-13-00257-t003_Table 3 ###### The incidence of hypertension stratified by overweight status in 2006 and 2011. Overweight Status in 2006 Overweight Status in 2011 Hypertension Cases (*n*) No. of Subjects (*n*) OR (95% CI) --------------------------- --------------------------- -------------------------- ----------------------- -------------------------- -------------------------- Non-overweight Non-overweight 228 1705 1.000 1.000 Non-overweight Overweight 75 415 1.446 (1.067--1.961) ^c^ 1.469 (1.082--1.992) ^c^ Overweight Non-overweight 38 170 1.000 1.000 Overweight Overweight 265 925 1.530 (1.023--2.288) ^c^ 1.580 (1.053--2.370) ^c^ Notes: OR, odds ratio; CI, confidence interval. ^a^ Logistic regression models were used, adjusting for age, SBP and DBP at baseline; ^b^ Logistic regression models were used, adjusting for variables in model 1 as well as smoking habits, alcohol consumption and place of residence at baseline; ^c^ Significantly greater odds, *p* \<0.05. ijerph-13-00257-t004_Table 4 ###### The association between SBP D-value and DBP D-value and change in overweight status between 2006 and 2011. Overweight Status in 2006 Overweight Status in 2011 SBP D-Value DBP D-Value --------------------------- --------------------------- -------------------------- -------------------------- -------------------------- -------------------------- Non-overweight Non-overweight Reference Reference Reference Reference Non-overweight Overweight 2.980 (1.558--4.402) ^c^ 2.933 (1.511--4.356) ^c^ 2.552 (1.589--3.515) ^c^ 2.528 (1.568--3.488) ^c^ Overweight Non-overweight Reference Reference Reference Reference Overweight Overweight 3.152 (0.897--5.408) ^c^ 3.285 (1.032--5.539) ^c^ 2.419 (0.877--3.960) ^c^ 2.550 (1.018--4.082) ^c^ Notes: SBP, systolic blood pressure; DBP, diastolic blood pressure; D-value, difference value between follow-up and baseline. ^a^ Multivariable linear regression models were used, adjusting for age, SBP and DBP at baseline; ^b^ Multivariable linear regression models were used, adjusting for variables in model 1 as well as smoking habits, alcohol consumption and place of residence at baseline; ^c^ *p* \< 0.05.

\section{Introduction} \label{introduction} The ``standard big bang" model of the universe has been recognized by most of the cosmologists. However, there still exist some puzzles to be clarified. The inflation model as a paradigm to solve the cosmological puzzles has made considerable success in understanding the Universe, especially the recent detection of the B-mode from cosmic microwave background (CMB) by the BICEP2 group \cite{ade2014bicep2} strongly consolidates this model. However, the primordial seeds and some other structures should be built in the initial conditions. Different from the inflation model, \citet{albrecht1999time} introduced a model of time-variable speed of light $c$ in the early universe to solve the cosmological horizon, flatness, and cosmological constant problems. This type of model is usually called the varying speed of light (VSL) theory and gradually becomes an alternative to the inflation. In the VSL theory, the radiation also dominates the Universe at early times. Einstein's gravity is also valid. Therefore, the cosmic geometry and expansion factor are the same as in the standard big bang model. However, free falling observers would measure that the local speed of light varies with cosmic time. Since then, this theory was further studied by Moffat \textit{et al}. \citep{1999PhLB..447..246B,barrow1999cosmologies,avelino1999does,moffat2001acceleration,dkabrowski2013regularizing} to figure out some cosmological problems. Beyond the above problems, the effect of the VSL theory also can be evidenced in some fundamental physics. For example, in 1999, \citet{webb1999search} observed the variation of fine-structure constant $\alpha=e^2/\hbar c$ with cosmic time from the quasar absorption spectra. In following works, \citet{webb2001further} and \citet{king2012spatial} even showed that the fine-structure constant may vary along the space. We note that the fine-structure constant $\alpha$ is inversely proportional to the speed of light. Therefore, a variation of speed of light also can explain the change of fine-structure constant \citep{barrow1999cosmologies,avelino1999does}. The VSL theory, no doubt, should deserve more attention. However, there are also some crucial criticisms against the VSL theory. \citet{ellis2007note} argued that VSL theory makes much of modern physics which depends on a constant speed of light to be rewritten, and also pointed out a number of detailed issues due to the VSL theory, such as the VSL theory might contradict Lorentz invariance, and must modify Maxwell's equation. As the research on this theory develops, some problems have been solved. One of those problems is the breakage of the Lorentz invariance \citep{bassett2000geometrodynamics}, and Ref. \citep{jafari2004operational} showed that VSL theories in the Fock-Lorentz \citep{fok1959theory,manida1999fock} and Magueijo-Smolin \citep{magueijo2002lorentz,magueijo2003generalized} transformations are the redescriptions of special relativity. Thus, the Lorentz invariance is not violated. However, the VSL theory also has to face some other serious problems. In addition, there is not any evidence to support, or completely eliminate this theory. The VSL cosmology is still not universally accepted. In summary, the VSL theory can successfully solve some significant problems, but it is still excluded by mainstream physics because it has to face a series of problems. Thus, whether the speed of light is varying cannot be judged by theoretical analysis and should be judged only by investigations of the observational data. This topic is exactly what we want to do. In this paper, we examine VSL theory from the observational data, and from the fundamental physics. The observational data we use here are common Type Ia supernovae (SNIa), cosmic microwave background (CMB), observational $H(z)$ data (OHD), and baryon acoustic oscillations (BAO). In previous works, some of them mainly replaced the dark energy with the effect of variation of the speed of light to explain the cosmic acceleration \citep{2014MPLA...2950103Z}. Now, we should make it clear whether the speed of light is varying from the observational data. Following previous investigations, we select the common form $c(t)=c_0a^n(t)$ with the contribution of dark energy and matter, and consequently construct a varying speed of light dark energy model (VSLDE). In addition, we also use the $Om(z)$ diagnostic to discriminate the varying speed of light dark energy model from the $\Lambda$ cold dark matter ($\Lambda$CDM). The method can be regarded as a criterion to test whether the speed of light is varying. This paper is organized as follows. In Sec. \ref{VSL}, the VSL theory and the specific form of the speed of light are presented. The method for constraining cosmological models with SNIa, CMB, OHD, and BAO is introduced in Sec. \ref{data}. In Sec. \ref{constrain}, we present the constraint results from the observational data. In Sec. \ref{om}, the $Om(z)$ diagnostic is presented and applied to discriminate the VSL model from $\Lambda$CDM. Finally, the summary and discussion are given in Sec. \ref{conclusion}. \section{The varying speed of light theory} \label{VSL} First of all, we can introduce the time variable $c$, i.e., $c \rightarrow c(t)$. Generally, one can introduce the time variable $c(t)$ in the metric or in the Friedmann equations. In Ref. \citep{bassett2000geometrodynamics}, they believe that the introduction of $c(t)$ in the metric does not change the physics by a coordinate transformation $c \, dt_{\textmd{new}} = c(t) \, dt$. Actually, the scalar factor $a(t)$ will be $a(t_{\textmd{new}})$, which will bring a series of confusing results. In addition, as noted in Ref. \citep{ellis2007note}, the speed of light $c$ is not a dimensionless quantity. Consequently, one can change the quantity $c$ to any value wanted with the transformations of coordinates. The new speed of light $c$ in the new coordinate is just the coordinate speed of light. In conclusion, in order not to change the reality of cosmic time, we would like to investigate the behavior of VSL theory in the normal cosmic time frame. Thus, the introduction of the time variable $c(t)$ in the metric is more fundamental than that in the Friedmann equations, which is used by many works about VSL theory. Consequently, more fundamental characteristics of the VSL cosmology will be revealed, which will be discussed in detail in the following. At first, we adopt the cosmological principle that implies a homogenous and isotropic universe, and assume the speed of light is a function of cosmic time. The Friedmann-Robertson-Walker(FRW) metric can be written as \begin{equation} \label{metric} ds^{2}=-c^2(t)dt^{2}+a^{2}\left[\frac{dr^{2}}{1-kr^2}+r^{2}\left(d\theta^{2}+sin^{2}\theta d\phi^{2}\right)\right], \end{equation} where $k$ denotes the curvature of the space. We will just think about a spatial flat FRW universe, namely $k=0$, for simplicity. The Einstein field equation reads \begin{equation}\label{Ein} G_{\mu\nu}=\frac{8\pi G}{c^4(t)}T_{\mu\nu}. \end{equation} By substituting Eq. (\ref{metric}) into Eq. (\ref{Ein}), we can obtain the Friedmann equations with a varying speed of light, \begin{eqnarray}\label{Fried} 3\frac{\dot{a}^2}{a^2} & = & 8\pi G\rho,\\ 2\frac{\ddot{a}}{a}+\frac{\dot{a}^2}{a^2}-2\frac{\dot{a}}{a}\frac{\dot{c}}{c} & = & -\frac{8\pi G}{c^2}p, \label{Fried2} \end{eqnarray} where $\rho$ and $p$ are the total energy density and pressure, respectively. Here the dot denotes the derivative with respect to cosmic time $t$. Obviously, the term $-2\frac{\dot{a}}{a}\frac{\dot{c}}{c}$ is the effect of the varying speed of light compared with common Friedmann equations. Meanwhile, this term is the difference between the introduction of $c(t)$ in the metric and in Friedmann equations. As we mentioned above, such Friedmann Eqs. (\ref{Fried}) and (\ref{Fried2}), should be more fundamental. The significant effect of this term will be revealed in the following calculations. The corresponding conservation equation reads \begin{equation} \label{conservation} \dot{\rho}+3H\left(\rho+\frac{p}{c^2}\right)=\frac{3H^2}{4\pi G}\frac{\dot{c}}{c}, \end{equation} where $H=\dot{a}/a$ is the Hubble parameter. Equation (\ref{conservation}) shows that the energy-momentum conservation is broken for $\dot{c}(t) \neq 0$. And any change in the speed of light is a source of matter creation. In fact, we can also obtain the same conclusion from Eq. (\ref{Ein}) as \begin{equation} \left(\frac{8\pi G}{c^4(t)}T^{\mu\nu}\right)_{;\mu}\neq 0. \end{equation} To solve this problem, the following two solutions have been proposed. First, it is to modify the right side of Eq.(\ref{Ein}). We can add other terms to $T^{\mu\nu}$ \citep{shojaie2006cosmology} or make gravitational constant $G$ vary in time, so that $G(t)c(t)^{-4}=$const \citep{barrow1999cosmologies}. Thus, the energy-momentum conservation can be satisfied. Second, we can neglect the energy-momentum conservation, and regard the variation of the speed of light as a source of matter creation. More related research has been discussed in Refs. \citep{shojaie2006cosmology,shojaie2007varying,clayton1999dynamical}. In this paper, we adopt the latter suggestion. The following form for the speed of light has been widely used \citep{barrow1999cosmologies}: \begin{equation}\label{ct} c(t)=c_0a^n(t)=c_0\left(1+z\right)^{-n}, \end{equation} where $c_0$ is the current value of the speed of light, and $n$ is a constant. It is obvious that the speed of light $c(t)\rightarrow c_0$ with $n\rightarrow 0$. Moreover, we have $\dot{c}/c=n\dot{a}/a$, so the speed of light grows from zero to $c_0$ for $n>0$, and decreases from infinity to $c_0$ for $n<0$. As discussed in Ref. \citep{2014arXiv1406.0150B}, \citeauthor{2014arXiv1406.0150B} investigated the influence of VSL on the luminosity distance. From the cosmographic method, they found that negative $n$ would lead the universe to accelerate stronger. Positive $n$ would lead to the stronger deceleration. Moreover, a precise test of the fine-structure constant \citep{2007MNRAS.378..221M} provides that the constant $n$ is rather small. Therefore, the parameter $n$ is a considerable significant factor that describes the varying of the speed of light or the evolution of the universe. Using Eqs. (\ref{Fried}) and (\ref{ct}), the conservation equation (\ref{conservation}) could be rewritten as \begin{equation} \dot{\rho}+3H\left(\rho+\frac{p}{c^2}\right) = 2n \rho H. \end{equation} We can define the equation of state (EoS) as \begin{equation}\label{es} w \equiv \frac{p}{\rho c^2(t)}=w_0 a^{-2n}=w_0 \left(1+z\right)^{2n}, \end{equation} where $w_0\equiv p/\rho c_0^2$ denotes the current value of EoS. For matter $w_{m0}=0$, and for radiation $w_{r0}=1/3$. And then, the conservation equation is \begin{equation}\label{cons} \dot{\rho}+\rho H\left(3+3w -2n\right)=0. \end{equation} Solving this differential equation, we have \begin{equation}\label{rho} \rho=\rho_0\left(1+z\right)^{-2n}\exp\left[\int^{z}_{0}\frac{3(1+w)}{1+z}dz\right], \end{equation} where $\rho_0$ is the present value of the total energy density. If we only consider the contribution of dark energy and matter, namely, $\rho=\rho_m+\rho_x$, where $\rho_m$ denotes matter density and $\rho_x$ is dark energy density, Eq. (\ref{rho}) can be expanded as \begin{eqnarray} \rho_m &=& \rho_{m0}\left(1+z\right)^{3-2n}, \\ \rho_x &=& \rho_{x0}\left(1+z\right)^{-2n}\exp\left[\int^{z}_{0}\frac{3(1+w_x)}{1+z}dz\right], \end{eqnarray} where $w_x=w_{x0}\left(1+z\right)^{2n}$, and $w_{x0}$ is the current value of the equation of state of dark energy. When $n \rightarrow 0$, all of these equations will reduce to the general dark energy model. Solving the corresponding Friedmann equation (\ref{Fried}), we obtain the Hubble parameter \begin{equation}\label{Ez} \begin{aligned} E^2=&\Omega_{m0}\left(1+z\right)^{3-2n}\\ &+\Omega_{x0}\left(1+z\right)^{-2n}\exp\left[\int^{z}_{0}\frac{3(1+w_x)}{1+z}dz\right], \end{aligned} \end{equation} where $E^2\equiv H^2/H_0^2$, $\Omega_{m0}$ and $\Omega_{x0}$ are the matter and dark energy density parameter today, respectively. From the normalization condition, we have three independent variables, $\Omega_{m0}$, $w_{x0}$, and $n$, which should be confirmed by the observational data. \section{The method for constraining cosmological models and observational data} \label{data} In this section, we would like to introduce the observational data and constraint method. The corresponding observational data we use are distance moduli of SNIa, CMB shift parameter, OHD Hubble parameter, and BAO distance parameter. However, we should notice that the varying speed of light would affect the distance measurements. As found by \citet{barrow2000can}, the variable $c$ does not introduce an intrinsic effect in the spectral line of supernova and also does not dim the standard candle. However, the variable $c$ would correct the Hubble diagram, which induces that the objects in the VSL theory are farther away from us because of the faster speed of light. This effect also can be evidenced in the Taylor series expansion of luminosity distance \citep{barrow2000can,balcerzak2013redshift}. As shown by them, this effect is presented by the introduction of $n$ in the second term. In addition, the variation of the speed of light would affect the evolution of cosmology through the varying EoS of dark energy. \subsection{Type Ia supernovae} \label{SNIa data} As early as 1998, cosmic accelerating expansion was first observed by ``standard candle" SNIa which all have the same intrinsic luminosity. Therefore, the observables are usually presented in the distance modulus, the difference between the apparent magnitude $m$ and the absolute magnitude $M$. The latest version is the Union2.1 compilation \citep{suzuki2012hubble} which includes 580 samples. They are discovered by the Hubble Space Telescope Cluster Supernova Survey over the redshift interval $0.01<z<1.42$. The theoretical distance modulus is given by \begin{equation} \mu_{th}(z)= m-M= 5 \log_{10}D_L(z)+\mu_0, \end{equation} where $\mu_0=42.38-5\log_{10}h$, and $h$ is the Hubble constant $H_0$ in units of 100 km s$^{-1}$Mpc$^{-1}$. The corresponding luminosity distance function $D_L(z)$ in the VSL theory is \begin{eqnarray}\label{dl} D_L(z) = (1 + z) \int_{0}^{z} \frac{ c(t) dz'}{E(z'; \textbf{p})}, \end{eqnarray} where $E(z'; \textbf{p})$ is the dimensionless Hubble parameter given by Eq. (\ref{Ez}), and \textbf{p} stands for the parameter vector of the evaluated model embedded in expansion rate parameter $E(z)$. We note that the parameters in the expansion rate $E(z)$ include the annoying parameter $h$. To be immune from the Hubble constant, we should marginalize over the nuisance parameter $\mu_0$ by integrating the probabilities on $\mu_0$ \citep{pietro2003future,nesseris2005comparison,perivolaropoulos2005constraints}. Finally, we can estimate the remaining parameters without $h$ by minimizing \begin{equation} \label{chi2_SN2} \tilde{\chi}^{2}_{\rm SN}(z,\textbf{p})= A - \frac{B^2}{C}, \end{equation} where \begin{eqnarray} A(\textbf{p}) &=& \sum_{i} \frac{[\mu_{obs}(z) - \mu_{th}(z; \mu_0 = 0, \textbf{p})]^2}{\sigma_{i}^{2}(z)}, \nonumber\\ B(\textbf{p}) &=& \sum_{i} \frac{\mu_{obs}(z) - \mu_{th}(z; \mu_0=0, \textbf{p})}{\sigma_{i}^{2}(z)}, \nonumber\\ C &=& \sum_{i} \frac{1}{\sigma_{i}^{2}(z)} , \nonumber \end{eqnarray} and $\mu_{obs}$ is the observational distance modulus. This approach has been used in the reconstruction of dark energy \citep{wei2007reconstruction}, parameter constraint \citep{wei2010observational}, reconstruction of the energy condition history \citep{wu2012reconstructing}, etc. \subsection{Cosmic microwave background} \label{cmb data} The CMB experiment measures the temperature and polarization anisotropy of the cosmic radiation in the early epoch. It generally plays a major role in establishing and sharpening the cosmological models. In the CMB measurement, the shift parameter $R$ is a convenient way to quickly evaluate the likelihood of the cosmological models. It can be obtained from acoustic oscillations \citep{hu1996small,hinshaw2009five} and contains the main information of the CMB observation. In the VSL theory, it is expressed as \begin{equation} R=\sqrt{\Omega_{m0}}\int^{z_s}_0\frac{ c(t) dz'}{E(z';\textbf{p})}, \end{equation} where $z_s=1090.97$ is the redshift of decoupling. According to the measurement of WMAP-9 \citep{hinshaw2012nine}, we estimate the parameters by minimizing the corresponding $\chi^2$ statistics \begin{equation} \label{cmb constraint} \chi ^2_R = \left( \frac{R-1.728}{0.016} \right)^2. \end{equation} \subsection{Observational $H(z)$ data} \label{ohd} Unlike the distance measurement, the OHD is an indirect measurement of the cosmic expansion history. The $H(z)$ data can be obtained from differential ages of galaxies \citep{simon2005constraints,stern2010cosmic}, from the BAO peaks in the galaxy power spectrum \cite{gaztanaga2009clustering,moresco2012improved} or from the BAO peaks using the Ly$\alpha$ forest of quasars \cite{2013A&A...552A..96B}. In this paper, we will use the new data from Ref. \citep{farooq2013hubble} that contains 28 observational data. The best-fit values of the parameters can be obtained by minimizing \begin{equation} \chi^2_{\rm{OHD}}=\sum_{i}^{28}\frac{[H_0E(z_i)-H_{obs}(z_i)]^2}{\sigma_i^2}. \end{equation} In the calculation, we adopt the WMAP9 result \citep{hinshaw2012nine} $H_0=70\pm 2.2$ km s$^{-1}$Mpc$^{-1}$ as the prior. \subsection{Baryon acoustic oscillation} \label{bao data} The measurement of BAO in the large-scale galaxies has rapidly become one of the most important observational pillars in cosmological constraints. This measurement is usually called the standard ruler in cosmology \citep{eisenstein1998baryonic}. The distance parameter $A$ obtained from the BAO peak in the distribution of sloan digital sky survey luminous red galaxies \citep{eisenstein2005detection} is a significant parameter and is defined as \begin{equation} A_{th}=\Omega_{m0}^{1/2}E(z_1)^{-1/3}\left[\frac{1}{z_1}\int^{z_1}_{0}\frac{c(t) dz'}{E(z';\textbf{p})}\right]^{2/3}. \end{equation} We use the three combined data points in Ref. \citep{addison2013cosmological} that cover $0.1 < z <2.4$ to determine the parameters in evaluated models. The expression of $\chi^2$ statistics is \begin{equation} \chi^2_{A}=\sum_{i} \left(\frac{A_{th}-A_{obs}}{\sigma^2_A}\right)^2, \end{equation} where $A_{obs}$ is every observational distance parameter and $\sigma_A$ is its corresponding error. Since the SNIa, CMB, OHD, and BAO data points are effectively independent measurements, we can simply minimize their total $\chi^{2}$ values as \begin{eqnarray} \label{21}\chi^2(z, \textbf{p})= \tilde{\chi}^{2}_{\rm SN} + \chi^{2}_{R}+\chi^{2}_{ A}+\chi^2_{\rm{OHD}}, \nonumber \end{eqnarray} to determine the parameters in the VSLDE. \section{constraining the parameters of VSL theory with observational data} \label{constrain} \begin{figure} \begin{center} \includegraphics[width=0.48\textwidth]{omwc} \caption{The $68.3\%$ and $95.4\%$ confidence regions in the $w_{x0}$-$\Omega_{m0}$ parameter space for VSLDE, which are constrained by combined observational data of SNIa, CMB, OHD and BAO. The red asterisk indicates the best-fit point. \label{omwc}} \end{center} \end{figure} \begin{figure} \begin{center} \includegraphics[width=0.48\textwidth]{wcn} \caption{Same as Fig. \ref{omwc} but for the $n$-$w_{x0}$ parameter space. \label{wcn}} \end{center} \end{figure} Using the combined observational data sets of SNIa, CMB, OHD, and BAO, we perform the $\chi^2$ statistics. The contour diagram of parameters $w_{x0}$-$\Omega_{m0}$ and $n$-$w_{x0}$ are given in Figs. \ref{omwc} and \ref{wcn}, respectively. It is shown that the combined data can present very compact constraints on the parameters. Considering the degrees of freedom (dof), the result is also pretty good, i.e., $\chi^2_{min}$/dof=0.9602 with parameters $w_{x0}$-$\Omega_{m0}$, and $\chi^2_{min}$/dof=0.9517 with parameters $n$-$w_{x0}$. Marginalizing over the posterior probability, we obtain the matter density $\Omega_{m0}=0.2878^{+0.0153}_{-0.0151}\left(1\sigma\right)$, the current EoS of dark energy $w_{x0}=-1.017^{+0.028}_{-0.030}\left(1\sigma\right)$, and the key constant $n=-0.0033^{+0.0045}_{-0.0045}\left(1\sigma\right)$. As previously mentioned, $n$ is characteristic for the severity of the varying speed of light. First, from the constrained constant $n$, we find that the variation of the speed of light is tiny. The constant $n$ is only in order of 10$^{-3}$. Moreover, the observational data show that nonvariable $c$ cannot be ruled out in the $1\sigma$ confidence level. In Fig. \ref{speed}, we reconstruct $c(z)$ with redshift $z$. We find that the speed of light decreases with decreasing redshift $z$. Considering the error estimation of $c(z)$, we obtain that the speed of light almost has no variation for $z<10^{-1}$. Even in the early epoch, its variation is only in order of 10$^{-2}$. In addition, we also deduce that the high-$z$ data are more sensitive to the VSL model than the lower-$z$ data. Although current data including the CMB cannot remove or admit the VSL model with a high significance, the future observations with a high redshift may be able to test this effect with the improvement of observational precision. Taking the CMB shift parameter as an example, the uncertainty is now 0.9\% from WMAP-9. Besides the parameter $n$, we note that the constrained EoS of dark energy is also interesting. Although the cosmological constant as a candidate of dark energy cannot be ruled out by the data, we find that they favor the phantomlike dark energy with $w_x<-1$ much more. In Fig. \ref{wz}, we show the reconstruction of $w_x (z)$ with the redshift. We find that the best-fit value of EoS is less than $-1$ for $z<10$. For the future, it almost behaves like the phantom. Moreover, the values of the parameters $\Omega_{m0}$ and $w_{x0}$ are pretty consistent with the mainstream physics. \begin{figure} \begin{center} \includegraphics[width=0.48\textwidth]{ct} \caption{The reconstruction of speed of light $c (z)$. The central red solid line is the best-fit value of $c(z)$. Top and bottom lines are the errors of reconstructed $c(z)$ in the $1\sigma$ confidence level. \label{speed}} \end{center} \end{figure} \begin{figure} \begin{center} \includegraphics[width=0.48\textwidth]{wz} \caption{The reconstruction of EoS of dark energy $w_x (z)$. The central red solid line is the best-fit value of $w_x (z)$. The shaded regions are the errors of reconstructed EoS within $1\sigma$ confidence level. \label{wz}} \end{center} \end{figure} \section{Difference between VSLDE and $\Lambda$CDM} \label{om} \begin{figure} \begin{center} \includegraphics[width=0.48\textwidth]{wucha} \caption{The reconstruction of $Om(z)$ diagnostic. The central red solid line is the best-fit curve of $Om(z)$ in the VSLDE. The blue dashed line represents the $Om(z)$ in the spatially flat $\Lambda$CDM. The shaded area shows the 1$\sigma$ error region in the VSLDE. \label{wucha}} \end{center} \end{figure} According to the above analysis, we can see that the VSLDE is very close to $\Lambda$CDM. Now, we apply an effective geometrical diagnostic, $Om(z)$, to directly show the difference between the VSLDE and $\Lambda$CDM. $Om(z)$ is defined as \citep{sahni2008two} \begin{equation} Om(z)\equiv \frac{E^2-1}{(1+z)^3-1}. \end{equation} For a spatially flat radiation vanished $\Lambda$CDM model, we obtain $Om(z)=\Omega_{m0}$, a constant that is independent of the redshift. Therefore, it is used widely to discriminate some exotic dark energy models (e.g., quintessence \citep{sahni2008two}, phantom \citep{sahni2008two}, and Ricci dark energy model \citep{feng2008statefinder}) from $\Lambda$CDM. Investigations in Ref. \citep{sahni2008two} also show that a positive slope of $Om(z)$ indicates a phase of phantom ($w<-1$) while a negative slope represents quintessence ($w>-1$). Using the observational data of SNIa, CMB, OHD, and BAO, we obtain $\Omega_{m0}=0.279$ in the $\Lambda$CDM model. Therefore, $Om (z)$ in the $\Lambda$CDM model is the constant 0.279. In Fig. \ref{wucha}, we reconstruct the geometrical diagnostic $Om(z)$ at $68.3\%$ confidence level, and find that the VSLDE and $\Lambda$CDM cannot be discriminated. Because the key factor $n$ in the varying speed of light theory is so feeble. To some extent, the speed of light may be a constant with high significance. \section{Conclusion and discussion} \label{conclusion} Varying speed of light theory has been put forward to explain some cosmological problems, such as horizon, flatness and cosmological singularity problems. However, the VSL cosmology is excluded by mainstream physics because it makes much of modern physics depending on a constant speed of light to be rewritten. Moreover, it also faces a number of other detailed issues. In the present paper, we devote ourselves to testing whether the speed of light is varying from the observational data. Considering a specific form for the variable $c(t)=c_0a^n(t)$, we deduce the corresponding Friedmann equation and Hubble parameter. In some previous references, some of them introduced the time variable $c(t)$ just in the Friedmann equation. However, we think that the application of $c(t)$ in the metric is much more fundamental and realistic. Using the combined observational data of SNIa, CMB, OHD, and BAO, we obtain $\Omega_{m0}=0.2878^{+0.0153}_{-0.0151}$, $w_{x0}=-1.017^{+0.028}_{-0.030}$, and $n=-0.0033^{+0.0045}_{-0.0045}$ at the $1\sigma$ confidence level. From the reconstructed $c(z)$ in Fig. \ref{speed}, we obtain that the speed of light almost has no variation for $z<10^{-1}$. Even in the early epoch, its variation is only in the order of $10^{-2}$. In Ref. \citep{2003PhRvD..68f3511S}, \citeauthor{2003PhRvD..68f3511S} constructed the phase-space portraits of the VSL models and obtained a global view on their dynamics. The most probability is that the universe can be created with constant $c (n=0)$, if the classical spacetime emerges via the quantum tunneling process. Moreover, the least size $a_{0}$ of the VSL FRW universe probably permits an insignificant $-1<n<0$. The speed of light $c$ decreases with the evolution of the universe. Obviously, our investigation verifies and favors the previous deductions. In addition, we also find that the high-$z$ data are more sensitive to the VSL model than the lower-$z$ data. Applying an effective geometrical diagnostic, $Om(z)$, we find that it is still difficult to differentiate the VSLDE from the $\Lambda$CDM model, because the key factor $n$ is so trivial. Apparently the observational data favor the constant speed of light. However, we should note that high-$z$ observations are more useful for testing this model. With the improvement of observational precision and the introduction of some new observations, the constrained results may be consolidated. For example, the redshift drift proposed by \citet{sandage1962change} would monitor the wavelength shift of a quasar from the Ly$\alpha$ absorption lines \citep{loeb1998direct,liske2008cosmic}. The measurements can extend to the redshift region $z=2-5$ to get the signal in order of about $10^{-11}$. In Ref. \citep{balcerzak2013redshift}, the authors investigate the redshift drift in the VSL theory and find that the redshift drift for 15 years can test $|n|<0.045$. Therefore, we can obtain that more observational baselines are needed to distinguish the VSL models from constant-$c$ cosmology. Besides the secular redshift drift, we note that the future gravitational wave probes will have higher sensitivity. With the detection of the B-mode from CMB by the BICEP2 group \cite{ade2014bicep2}, the gravitational wave probes will open a new window for us. The related research can be found in Ref. \citep{2014arXiv1404.5567M}. In addition, we find from Fig. \ref{wcn} that parameters space $n$ and $w_{x0}$ are related in the considered observational data. Therefore, we also can expect that the inclusion of some other data, such as the gravitational lens and the growth rate, could give a tighter constraint on the VSL model because of different degeneracies. On the other hand, there exist some deficiencies to advance. In the present paper, we adopt a widely used form of VSL, i.e., $c(t) = c_0 a^{n} (t)$. The conclusion is under this assumption. Therefore, we must be careful to extrapolate our conclusions. In the future, we would like to investigate the VSL theory in other forms. \acknowledgments{ We quite appreciate the anonymous referee for his/her suggestions to improve this manuscript. J.-Z. Qi would like to thank Rong-Jia Yang for his useful help. This work is supported by the National Natural Science Foundation of China (Grant No. 11235003, No. 11175019, and No. 11178007).}

INTRODUCTION ============ Fragile X syndrome (FXS), the most common cause of inherited mental retardation with a prevalence of 1:4000 in males and 1:6000 in females, is caused by an expansion of a (CGG)*~n~* sequence in the 5′ UTR of the *FMR1* gene to \>200--2000 repeats ([@B1; @B2; @B3; @B4; @B5]). Expansion of the (CGG)*~n~* tract beyond ∼200 repeats entails hypermethylation of the repeat sequence and of a neighboring CpG island that results in the transcriptional silencing of *FMR1* and absence of the product FMRP protein ([@B3]). Although carriers of *FMR1* premutation alleles that have \>55--200 (CGG) repeats do not develop FXS, some do present various forms of clinical involvement such as minor alterations of physical features and emotional problems ([@B6]). Notably, ∼20% of the female premutation carriers develop premature ovarian failure (POF) ([@B7]) and one-third of the carrier males are affected by fragile X-associated tremor-ataxia syndrome (FXTAS) ([@B8],[@B9]). In contrast to the absence of *FMR1* transcripts in FXS cells, the levels of *FMR1* mRNA in peripheral blood leukocytes at the upper range of FXS (CGG)*~n~* premutation were found to increase by 5- to 10-fold relative to normal subjects ([@B10],[@B11]). However, the amounts of FMRP in these cells remain within the normal range or are somewhat below normal at higher premutation repeat sizes ([@B10],[@B12],[@B13]). A progressively diminishing association of *FMR1* mRNA with polysomes that have increasing (CGG) repeat sizes was observed in lymphoblastoid cell lines of premutation carriers that over-express *FMR1* mRNA ([@B14]). It appears, therefore, that expanded premutation (CGG)*~n~* tracts in *FMR1* mRNA lower the efficiency of its translation *in vivo*. A clue to the underlying mechanism of the declining efficacy of translation was obtained in a study of the impeded translation of *FMR1* mRNA that contained longer (CGG)*~n~* tracts. Although cells of a mildly affected FXS patient with *FMR1* alleles of 57--285 CGG repeats exhibited normal steady-state levels of *FMR1* mRNA, synthesis of FMRP from transcript with more than 200 repeats was markedly decreased and these transcripts were associated with stalled 40S ribosomal subunits ([@B15]). A plausible interpretation is that the longer (CGG)*~n~* sequences form secondary structures that block the migration of the 40S subunit. Thus, (CGG)~55--200~ premutation repeats may conceivably also fold into secondary structure(s) that slow down or obstruct ribosome progression, therefore lowering the efficiency of translation. DNA and RNA (CGG) repeat sequences readily fold into hairpin structures ([@B16; @B17; @B18; @B19]) that may pair to generate stable intermolecular tetraplexes ([@B20; @B21; @B22; @B23]). Quadruplex (CGG)*~n~* structures in DNA templates were reported to hinder the progression *in vitro* of DNA polymerases ([@B22],[@B24]). In a likely analogy, (CGG)*~n~* hairpin and quadruplex structures in mRNA may block polysome formation and impede ribosome progression and productive protein synthesis. If mRNA translation is obstructed by secondary structures of the moderately expanded premutation (CGG)*~n~* tracts, it should be enhanced by their destabilization. Indeed, several proteins and a low molecular size porphyrin were shown to efficiently disrupt *in vitro* tetraplex (CGG)*~n~* structures. The human Werner syndrome helicase was reported to catalyze the unwinding of intermolecular quadruplex structures of (CGG)*~n~* tracts in DNA ([@B24],[@B25]). Member proteins of the hnRNP family; CBF-A ([@B23],[@B26],[@B27]), hnRNP A2 and modified hnRNP A1 ([@B23]) and UP-1 ([@B28],[@B29]) were reported to mediate non-enzymatic destabilization of quadruplex forms of (CGG)*~n~* in both DNA and RNA. Additionally, the quadruplex DNA interacting cationic porphyrin TMPyP4 was shown to disrupt tetrahelical (CGG)*~n~* in DNA ([@B30]). In this work, we report that upstream positioned mid-range premutation (CGG)*~n~* tracts preferentially suppress the *in vitro* and *in vivo* translation of a reporter firefly luciferase gene. Demonstrating that (CGG)*~n~* tract in RNA forms *in vitro* an intramolecular secondary structure which is likely to impede translation, we show that the tetraplex (CGG)*~n~* destabilizing proteins hnRNP A2 and CBF-A enhance *in vivo* the translation of mRNA molecules that contain 62 or 99 (CGG) repeats. MATERIALS AND METHODS ===================== Plasmids -------- *pT7-FMR1-5′-UTR(CGG)~n~-FL*---pSP6-5′-*FMR1*-UTR (CGG)*~n~*-FL (*n* = 0, 30, 62, 99) plasmids were prepared by modifications of previously described constructs ([@B31]). The no-repeat plasmid (high-copy origin) was digested with NaeI and BlpI, while the 30, 62 and 99 CGG-repeat plasmids (low-copy origin) were digested with BglII. The BglII overhang was blunted with mung bean nuclease and digested by BlpI. Thus regions containing the CMV promoter and the 5′-end of the *FMR1* 5′-UTR, up to the BlpI site were removed while creating blunt-end to BlpI-overhang linear vectors. Both high- and low-copy plasmids were ligated to a double-stranded oligomer, 5′-(GGCATTTAGGTGACACTATAGATCAGGCGC)-3′·3′-(CCGTAAATCCACTGTGATATCTAGTCCGCGAGT)-5′ that contained the SP6 promoter and restored the 5′-end of *FMR1*. The formed *FMR1* 5′-UTR sequence starting from transcription site 1 was as previously described ([@B32]) with the following modifications. The first two 5′ AG bases of the 5′-UTR were changed to GA for optimal SP6 transcription, the 3′-end base was C instead of G, and the no CGG-repeat 5′UTR was devoid of both the CGG repeat and the following GC dinucleotide. The (CGG)~30~ repeat sequence had two intervening (AGG) triplets at positions 11 and 21, the (CGG)~62~ repeat contained an (AGG) interruption at the 11th position and the (CGG)~99~ repeat had at least one AGG intervening trinucleotide of undetermined location. Next, bacteriophage T7 promoter was incorporated into the pSP6-5′-*FMR1*-UTR(CGG)*~n~*-FL plasmids. Double-stranded DNA adaptor including the T7 promoter sequence (underlined) 5′-d(TCACGCGTACTT[AATACGACTCACTATAG]{.ul}GCTAGCCGC)-3′**•**3′-d(CGCATGAA[TTATGCTGAGTGATAT]{.ul}CCGATCGGCGAGT)-5′ was ligated into the BlpI site upstream to the plasmid UTR-(CGG)*~n~* sequence. Plasmids without a (CGG) tract or that included a (CGG)~30~ insert were propagated in 500 ml of LB medium at 37°C for 12 h in *Escherihcia coli DH10B* (gift of Dr D. Gidoni, Volcani Center, Israel) and purified (Maxi-prep kit, Qiagen). Due to the instability of the (CGG)~62~ and (CGG)~99~ repeats, plasmids harboring these inserts were grown in multiple 3-ml cultures. Following plasmid purification (Mini-prep kit, Qiagen), lengths of the BlpI and NheI excised inserts were determined by agarose gel electrophoresis and only plasmids that contained validated (CGG)~62~ or (CGG)~99~ tracts were used. *pCMV-FMR1-5′-UTR (CGG)~n~-FL*---pCMV-*FMR1*-5′-UTR(CGG)~99~-FL plasmid ([@B31]) was contributed by Dr P. Hagerman, UC Davies. To construct plasmids with 0, 30 or 62 repeats, the *FMR1*-5′-UTR(CGG)~99~ sequence was excised by BlpI and NheI digestion and the linearized plasmid was ligated with (CGG)~0~, (CGG)~30~ or (CGG)~62~-containg *FMR1*-5′-UTR sequences that were excised from the respective pT7-*FMR1*-5′-UTR(CGG)*~n~*-FL plasmids by BlpI and NheI digestion. Plasmid propagation and verification of correct repeat sizes were performed as described for the pT7-*FMR1*-5′-UTR(CGG)*~n~*-FL plasmids. *pCS107-FMR1-5′-UTR(CGG)~33~*---The *FMR1* 5′ untranslated region was amplified from genomic DNA of EBV transformed B lymphoblastoid cell line of a normal subject (gift of Dr P. Chiurazzi, Catholic University, Italy). The UTR tract which included 33 (CGG) repeats interrupted by two (AGG) triplet units at trinucleotides 11 and 21 was PCR amplified using a 5′-*FMR1* exon forward primer 5′-(CGCGGATCCCGCTCAGCTCCGTTTCGGTTTC)-3′ and *FMR1* 3′-end exon primer 5′-(CCGGAATTCTAGAAAGCGCCATTGGAGCCCC)-3′. The amplified fragment was restricted by BamHI and EcoRI digestion and cloned into SP6 promoter containing pCS107 plasmids that were propagated in *E. coli DH10B* and purified (Maxi-prep, Qiagen). *pCMV2-Flag-hnRNP A2 and pCMV2-Flag-CBF-A*---Plasmid *pCMV2-Flag* was constructed to harbor and express cDNA encoding the quadruplex (CGG)*~n~*-disrupting protein hnRNP A2 or its mutant hnRNP A2 F\[54\]S that lost the destabilizing activity ([@B23]). The plasmids pGEX-2T-*hnRNP A2* (gift of Dr Ralph C. Nichols, Dartmouth School of Medicine, NH) or pGEX-2T-*hnRNP A2 F\[54\]S* ([@B23]) were cut by EcoRI and SalI and the electrophoretically isolated cDNA were ligated into EcoRI and SalI-linearized pCMV2-Flag plasmid (Stratagene). To construct pCMV2-Flag plasmids that express the quadruplex (CGG)*~n~* destabilizing CBF-A protein or its mutant CBF-A ΔRNP1~1~ that lost the tetraplex disrupting activity ([@B26]), pCMV2-Flag was cut by EcoRI and dephosphorylated with calf intestinal phosphatase (New England Biolabs). CBF-A or CBF-A ΔRNP1~1~ encoding sequences that were cleaved by the same enzyme out of *pGEX-2T* plasmids ([@B26]) were ligated to the linearized pCMV2-Flag plasmids. The restriction digestion at the border of the Flag and the CBF-A or CBF-A ΔRNP1~1~ encoding sequences caused loss of the reading frame. A proper reading frame was restored by PFU (Promega) catalyzed PCR amplification of the altered DNA region using a 5′ primer 5′-d(CTTGCGGCCGCGAATCCATGGCCGACC)-3′ and a 3′ primer 5′-d(GGTCGGCCATGGATTCGCGGCCGCAAG)-3′. The modified plasmids were cloned and restoration of the reading frame was verified by DNA sequencing. Coupled *in vitro* transcription--translation --------------------------------------------- Coupled transcription and translation reactions were carried out using the TNT® transcription**--**translation system (Promega). Reaction conditions were essentially according to the manufacturer\'s specifications. Briefly, reaction mixtures contained in a final volume of 25 μl: 1 μl of 25× TNT reaction buffer; 0.5 μg XbaI-linearized pT7-*FMR1*-5′-UTR(CGG)*~n~*-FL plasmid; 12.5 μl 2× TNT rabbit reticulocyte lysate; 0.5 μl TNT T7 RNA polymerase; 5.0 μCi \[α-^32^P\] UTP (3000 Ci/mmol, Amersham); 20 μM of each amino acid; 40 U RNasin ribonuclease inhibitor (Promega). Following incubation at 30°C for 30 min, the samples were rapidly cooled to 4°C and 5 μl aliquots of each reaction mixture were mixed with 100 μl of luciferase assay Reagent II (Promega). Following a 3 s delay, firefly luciferase activity was determined for 30 s (20/20 luminometer, Turner Biosystems). To measure levels of the radiolabeled RNA, gel-loading dye containing SDS to a final concentration of 0.5% was added to 10 μl aliquots of the reaction mixtures and the samples were electrophoresed through 0.8% agarose gel. The identity of the radioactive RNA bands as full-length luciferase gene transcripts was verified by northern analysis using the luciferase 3′ gene probe 5′-(CTTTCCGCCCTTGGCCTTTATGAGGATC)-3′. The relative levels of the labeled luciferase RNA molecules that contained (CGG)*~n~* tracts of different lengths were determined by phosphor imager analysis. Transcription *in vitro* ------------------------ To investigate the effect of different lengths of (CGG)*~n~* tracts on transcription, pT7-*FMR1*-5′-UTR(CGG)*~n~*-FL plasmids (*n* = 0, 30, 62 or 99) were linearized by XbaI restriction digestion and transcribed *in vitro* by T7 RNA polymerase. Reaction mixtures contained in a final volume of 20 μl: 0.5 μg linear plasmid DNA; 1 mM each of the four rNTPs; 3.0 μCi \[α-^32^P\] UTP (3000 Ci/mmol, Amersham); 40 U of ribonuclease inhibitor (Takara); 80 mM potassium acetate; 40 U of T7 RNA polymerase and T7 polymerase reaction buffer (Promega). The reaction mixtures were incubated at 30°C for 30 min and the RNA polymerization reaction was terminated by the addition of 5 μl of 2.5% SDS in gel loading dye. Following resolution of the radiolabeled RNA by electrophoresis in 0.8% agarose, the gels were dried and radioactivity of the product RNA molecules was quantified by phosphor imaging analysis. That the radioactive RNA bands represented full-length transcripts was confirmed by northern analysis as described for the coupled transcription--translation system. Translation *in vitro* ---------------------- Equal amounts of radiolabeled RNA transcripts of the pT7-*FMR1*-5′-UTR(CGG)*~n~*-FL plasmids were translated *in vitro*. To prepare the RNA, 1 μg of each XbaI-linearized plasmid was transcribed at 37°C for 1.5 h in a final volume of 20 μl of Ampliscribe T7 transcription system (Epicenter Technologies) that contained 3.0 μCi \[α-^32^P\] UTP (3000 Ci/mmol, Amersham). Plasmid DNA was removed by digestion at 37°C for 15 min with 1 U of RNase-free DNase (Promega) and the RNA transcripts were extracted with ultra pure phenol:chloroform:isoamyl alcohol (25:24:1, Invitrogen) followed by chloroform extraction. The RNA was precipitated and washed with 70% ethanol and dissolved in RNase-free water. Approximately equal amounts of RNA as estimated by absorption at 260 mμ were resolved by agarose gel electrophoresis and the precise amounts of RNA were normalized by the radioactivity of the full-length RNA transcript bands. Equal amounts (0.5 μg each) of the *FMR1*-5′-UTR(CGG)*~n~*-FL transcripts were translated in rabbit reticulocyte lysate (Flexi® translation system, Promega) as follows. The RNA samples were incubated at 30°C for 30 min in a final volume of 25 μl of reticulocyte lysate translation system that contained 2.0 mM DTT; 1 mM magnesium acetate; 70 mM KCl and 20 U ribonuclease inhibitor (Takara). The translation reaction was terminated by rapid cooling of the samples to 4°C and the luciferase activity was determined in 10 μl aliquots of each reaction mixture as described for the coupled transcription--translation system. Ribonuclease T1 protection analysis ----------------------------------- One microgram of XhoI-linearized pCS107-*FMR1*-5′-UTR(CGG)~33~ plasmid DNA was transcribed in AmpliScribe SP6 system (Epicenter Biotechnologies) according to the manufacturer\'s instructions. The RNA transcript was precipitated and washed with 70% ethanol and resolved by denaturing gel electrophoresis through 10% polyacrylamide, 8 M urea. The gel band that contained the RNA transcript was visualized by UV adsorption, excised and agitated overnight at 4°C in a solution of 0.3 M sodium acetate pH 5.2; 0.5 mM EDTA; 0.1% SDS. Following ethanol precipitation and wash, the eluted RNA transcripts were dephosphorylated at 37°C for 1 h by shrimp alkaline phosphatase (USB). The phosphatase was heat-inactivated at 65°C for 15 min and the RNA was 5′-end labeled using T4 polynucleotide kinase and \[γ-^32^P\] ATP ([@B33]). The labeled RNA was resolved by denaturing gel electrophoresis through 10% polyacrylamide, 8 M urea and recovered and isolated by ethanol precipitation as described above. Duplex regions of RNA were distinguished from unpaired tracts by the resistance of the former to digestion by the single-stranded RNA-specific nuclease RNase T1. Labeled RNA transcripts were dissolved in RNase-free water and supplemented with unlabeled yeast RNA carrier to a final concentration of 8 μM. Digestion of native RNA with a specified amount of RNase T1 (Ambion) was conducted at 20°C for 20 min in a reaction mixture that contained in 10 μl final volume 10 mM MgCl~2~ and 80 mM potassium acetate in 10 mM Tris-HCl buffer, pH 7.4. Digestion of heat-denatured RNA by RNase T1 was conducted at 55°C for 15 min in the same reaction mixture that also contained 3.5 M urea. The reactions were terminated by adding to each mixture 20 μl of inactivation-precipitation buffer as specified by the manufacturer. Products of the ribonuclease digestion reactions were resolved by denaturing gel electrophoresis in 12% polyacrylamide, 8 M urea. Formation and resolution of intramolecular secondary structure of (CGG)~33~ RNA ------------------------------------------------------------------------------- *FMR1*-5′-UTR(CGG)~33~ RNA was transcribed and internally labeled in reaction mixtures that contained in a final volume of 20 μl: 1.0 μg XhoI-linearized pCS107-*FMR1*-5′-UTR(CGG)~33~ plasmid DNA, 10 mM DTT, 40 U RNasin (Promega), 0.5 mM ATP, 3.0 mM CTP, 6.0 mM GTP or 0.5 mM 7-deazaGTP (TriLink Biotechnologies), 8.3 μM UTP, 50 μCi \[α-^32^P\] UTP (800 mCi/mmol, Amersham), 20 U SP6 polymerase and 4 μl 5× SP6 transcription buffer (Promega). Transcription was conducted at 37°C for 60 min, the DNA was digested by 1 U DNase I at the same temperature for 15 min and the product RNA was phenol extracted and precipitated and washed with ethanol. An intramolecular secondary structure of the RNA was generated by incubating at 4°C for 15 min, mixtures that contained in a final volume of 10 μl: \<2.5 nM 5′-^32^P *FMR1*-5′-UTR(CGG)~33~ RNA, 120 mM KCl, 1 mM EDTA, 0.5 mM DTT, 20% glycerol in 20 mM Tris-HCl buffer, pH 7.8. Control reactions were performed under identical conditions except that KCl was omitted from the reaction mixture. The RNA was resolved by electrophoresis in non-denaturing 4% polyacrylamide gel in 0.5× TBE buffer, pH 8.2 that contained 10 mM KCl. Transfection of cultured human cells ------------------------------------ Human Embryonic Kidney 293 cells (HEK293) were seeded in 0.1% gelatin-coated 10 cm plates and grown to 80--90% confluence at 37°C in 5% CO~2~ atmosphere in Dulbecco Modified Eagle\'s Medium (DMEM) supplemented with 4.5 g/l [d]{.smallcaps}-glucose, 5 mM [l]{.smallcaps}-glutamine, 10% fetal calf serum, 83.3 U/ml each Penicillin and Streptomycin and 0.2 μg/ml Amphotericin B (Biological Industries, Israel). Cells that were reseeded at 3.3 × 10^5^ cells per 6 cm gelatin-coated plate, were grown overnight and transiently co-transfected with three plasmids: a vector harboring reporter firefly luciferase (FL) without or with an upstream (CGG) repeat tract; plasmid that encoded normalizing Renilla luciferase (RL) and plasmids expressing quadruplex destabilizing proteins or their inactive mutants. The growth medium was replaced with 4 ml of fresh DMEM to which 100 μl of DMEM were added that contained 6 μl Fugene 6.0 (Roche), 0.2 μg pCMV-*FMR1*-5′-UTR(CGG)*~n~*-FL DNA and 0.02 μg pCMV-RL plasmid (Promega) and 1.8 μg of DNA of one of the following expression plasmids: pCMV2-Flag; pCMV2-Flag-*hnRNP A2* pCMV2-Flag-*hnRNP A2 F\[54\]S*; pCMV2-Flag *CBF-A* or pCMV2-Flag-*CBF-A ΔRNP1~1~*. The cells were harvested 24 h after transfection, washed once with 1.5 ml cold phosphate buffered saline (PBS) and resuspended in 3 ml PBS. Aliquots of each sample were used to determine FL and RL activities and to conduct semi-quantitative RT-PCR measurements of the levels of their mRNA transcripts. Correction for variations in cell viability and transfection efficiency was performed for each experiment as described ([@B31]) except that the RL-corrected FL activity was normalized to FL activity of cells transfected with reporter pCMV-*FMR1*-5′-UTR-FL plasmid with no upstream (CGG) repeat tract. Dual luciferase assay --------------------- FL and RL activities were determined in lysates of transfected HEK293 cells according to the manufacturer\'s instructions using the dual luciferase reporter assay system, (Promega). Briefly, the cells were lysed in passive lysis buffer (Promega) and 20 μl of the cell lysate were added to 100 μl luciferase reagent II. Following a 3 s delay, FL activity was measured for 30 s using the 20/20 luminometer. The reaction was terminated by adding 100 μl Stop and Glo reagent to quench the FL activity and after a 3 s delay, the activity of RL was determined for 30 s. In each sample, triply measured values of FL activity were normalized to RL activity. Semi-quantitative RT-PCR determination of relative mRNA levels -------------------------------------------------------------- RNA was isolated from HEK293 cells using the total RNA isolation kit (Cartagen, USA) and genomic and plasmid DNA were removed by use of the Turbo DNA-free kit (Ambion). Reverse transcription and amplification reactions were conducted with the Reverse-it one-step RT-PCR kit (ABgene, UK). Each reaction mixture contained in a final volume of 25 μl: 0.05 μg total RNA; 0.3 or 0.9 μCi \[α-^32^P\] dCTP (3000 Ci/mmol) for the reverse transcription of FL or RL mRNA, respectively, and 5.0 pmol each of FL forward primer 5′-(CTATGAAGAGATACGCCCTGGTTCCTGG)-3′ and FL reverse primer 5′-(GGCAGTTCTATGAGGCAGAGCGAC)-3′ or RL forward primer 5′-(GGGATGAATGGCCTGATATTGAAGAAG)-3′ and RL reverse primer 5′-(CAATTTGTACAACGTCAGGTTTACCACC)-3′. FL or RL RNA were amplified by RT-PCR for 26 or 29 cycles, respectively, of 94°C for 20 s, 53°C for 25 s and 72°C for 1 min. Based on measurements of the levels of radiolabeled FL and RL reverse transcripts for each cycle between the 20th and 35th cycles, it was determined that 26th or 29th cycles were at the middle of the respective linear range of amplification of FL and RL RNA. Every set of reactions included a negative control of a mixture without RT to verify that all the amplification products were copies of RNA template and not of contaminating DNA. Equal aliquot volumes of the FL or RL RT-PCR ^32^P-labeled DNA products were electrophoresed at 12 V/cm through 4 or 6% non-denaturing polyacrylamide gels in 0.5× TBE buffer (2.4 mM EDTA in 1.08 mM Tris-borate buffer, pH 8.3) until a bromophenol blue marker reached the end of the gel. DNA size markers (T4 PNK end-labeled 5′-^32^P peqGOLD 100 bp DNA ladder, PeqLab Biotechnologies) were used to identify full-length FL and RL RT-PCR DNA product bands. The identity of these DNA products was verified by parallel electrophoresis of radiolabeled PCR products of the original plasmids DNA with the FL or RL primers. Amounts of the electrophoretically resolved DNA samples were quantified in the dried gels by phosphor imaging analysis. Western analysis ---------------- Equal amounts of HEK293 cell lysates protein were resolved by 10% SDS-PAGE. Following transfer to nitrocellulose membrane, the expression of functional or mutant Flag-hnRNP A2 or FLAG CBF-A proteins was detected by use of murine anti-Flag monoclonal antibody (1:1000, Sigma) as the primary antibody and horseradish peroxidase-conjugated goat anti mouse IgG (H + L, 1:10 000, Pierce) as secondary antibody. Horseradish peroxidase activity was detected by use of the Super Signal Wes Pico chemiluminescence substrate (Pierce). RESULTS ======= Premutation size (CGG)*~n~* repeat tract in DNA and RNA selectively diminish translation *in vitro* --------------------------------------------------------------------------------------------------- We first evaluated in model T7 promoter driven coupled and separate *in vitro* transcription and translation systems, the effects of increasing sizes of normal and premutation-range (CGG)*~n~* tracts on FL gene expression. ### Premutation-range (CGG)~n~ repeat sequences selectively depress translation in vitro The effect of increasing lengths of upstream (CGG)*~n~* tracts was first examined in a coupled *in vitro* transcription--translation system using as template linearized pT7-*FMR1*-5′-UTR(CGG)*~n~*-FL plasmids that harbor in the 5′ to 3′ direction bacteriophage T7 promoter, the GC-rich 5′-UTR of *FMR1*, tracts of 0, 30, 62 or 99 (CGG) repeats and an FL reporter gene. Following transcription of the plasmids and coupled translation, FL mRNA that was identified by northern analysis, was resolved by agarose gel electrophoresis and quantified and FL protein activity was determined as described in Materials and Methods section and in the legend to [Figure 1](#F1){ref-type="fig"}. Average results of three independent measurements are summarized in [Figure 1](#F1){ref-type="fig"}. As is evident, transcription driven by the T7 promoter was only minimally affected by the expanding length of the (CGG) repeat sequence ([Figure 1](#F1){ref-type="fig"}, first panel). Relative to FL transcription with no (CGG)*~n~* upstream tract, a cluster of 30 (CGG) trinucleotides that represented the most frequent normal size of the repeat sequence ([@B34]), enhanced on average the accumulation of FL transcripts by ∼1.3-fold whereas 62 and 99 repeats did not change the initial level of FL mRNA. Unlike transcription, the extent of translation was significantly affected by the (CGG)*~n~* stretches. Relative to FL transcripts with no upstream (CGG) repeat tract, a stretch of 30 (CGG) trinucleotides enhanced on average the translation of FL mRNA by 1.6-fold ([Figure 1](#F1){ref-type="fig"}, second panel). By contrast, 62 or 99 (CGG) repeats depressed the relative translation of FL by 1.3- and 1.8-fold, respectively. Moreover, when compared with RNA that contained 30 (CGG) triplets, tracts of 62 or 99 (CGG) triplets lowered the level of FL protein by 2.2- or 3-fold, respectively. The decreased efficiency of translation of FL mRNA molecules as a function of the increasing length of the (CGG)*~n~* tract is best reflected by the ratios of FL protein to mRNA. As shown in the third panel of [Figure 1](#F1){ref-type="fig"}, relative to transcripts devoid of the (CGG) repeat sequence, the efficiency of translation of mRNA that contained 62 or 99 repeats was diminished by 1.4- or 1.75-fold, respectively. Further, when compared to the maximally translated RNA molecules that included 30 (CGG) repeats, the utilization of mRNA molecules with 62 or 99 (CGG) repeats was diminished by 1.7- or 2.2-fold, respectively. Figure 1.Effect of (CGG) repeat size on the synthesis of FL mRNA and protein in a coupled transcription--translation *in vitro* system. Coupled transcription and translation of linearized pT7-*FMR1*-5′-UTR(CGG)*~n~*-FL plasmid DNA (*n* = 0, 30, 62 or 99) were conducted in a TNT® (Promega) reaction mixture containing \[α-^32^P\] UTP as described under Materials and Methods section. Presented are histograms of the measured levels of FL mRNA and protein activity and of their calculated ratios. The inset in sub-panel **A** is a typical phosphor image of agarose gel resolved radiolabeled FL transcripts containing increasing lengths of the (CGG)*~n~* repeat sequence. The control lane has PCR product of the pT7-*FMR1*-5′-UTR(CGG)~0~-FL DNA. RNA and protein were quantified as detailed in the Materials and Methods section. Each data point is an average of the results of three independent experiments. FL protein to mRNA ratios (panel **C**) were calculated by dividing for each (CGG)*~n~* repeat size the average value of FL protein activity by the corresponding average level of FL mRNA. The effect of (CGG) repeat length on the efficiency of mRNA and protein synthesis *in vitro* was next investigated in separate transcription or translation systems. Linearized pT7-*FMR1*-5′-UTR(CGG)*~n~*-FL DNA (*n* = 0, 30, 62 or 99) was transcribed using T7 RNA polymerase and the amounts of full-length FL transcripts as identified by northern analysis were quantified. Results presented in [Figure 2](#F2){ref-type="fig"}A show that the presence of increasing lengths of an upstream (CGG)*~n~* tract upstream to the FL gene lowered the synthesis of mRNA to a limited extent. Relative to DNA with no (CGG)*~n~* tract, 30, 62 or 99 (CGG) repeats decreased the average levels of FL gene transcripts by 9, 27 or 33%, respectively. The effect of an increasing number of mRNA (CGG) repeats on protein synthesis *in vitro*, was measured by translating in a reticulocyte lysate system equal amounts of full-length FL mRNA molecules that contained 0, 30, 62 or 99 (CGG) repeats and determining the activity of the product FL protein (see Material and Methods section). As shown in [Figure 2](#F2){ref-type="fig"}B, relative to the translation of mRNA devoid of a (CGG) repeat sequence, the translation of FL mRNA molecules that included 30, 62 or 99 (CGG) repeats declined on average by 5, 75 or 72%, respectively. Since the translation system was provided with equal amounts of each mRNA type, these results indicated a diminishing efficacy of the translation of mRNA molecules that contained progressively increasing length of the (CGG) repeat sequence. Hence, results obtained *in vitro* with coupled or separate transcription and translation systems both demonstrated preferential decline in the efficacy of mRNA translation as a function of (CGG)*~n~* length. Figure 2.Effect of (CGG) repeat size on T7 polymerase catalyzed FL transcription and on the translation *in vitro* of FL mRNA. (**A**) Levels of *FMR1*-5′-UTR(CGG)*~n~*-FL mRNA (*n* = 0, 30, 62 or 99) transcribed *in vitro* by T7 polymerase. Reaction conditions and resolution and quantification of the radiolabeled transcripts were as described under Materials and Methods section. The height of each column represents an average of the results of four independent experiments with the SD indicated by error bars. (**B**) Levels of the *in vitro* translated FL protein. Equal amounts of pT7-*FMR1*-5′-UTR(CGG)*~n~*-FL RNA transcripts containing the indicated (CGG) repeat sizes were translated *in vitro* in a rabbit reticulocyte lysate Flexi® system (Promega) section and relative FL protein activities were determined as described in the Materials and Methods section. Shown are average results and SD of four independent experiments. (CGG) repeat tract in RNA forms secondary structure --------------------------------------------------- DNA and RNA (CGG) repeat sequences were shown to readily form hairpin ([@B16],[@B17],[@B19],[@B20],[@B22],[@B35]), tetraplex ([@B20; @B21; @B22]) and slipped-strand ([@B36]) secondary structures. A likely cause for the observed declining efficiency of translation as a function of the size of the upstream (CGG)*~n~* mRNA sequence could be the formation of secondary structures by the RNA repeat sequence. To probe the structure of a (CGG)*~n~* tract in RNA, denatured or native 5′-^32^P labeled RNA transcripts of pCS107-*FMR1*-5′-UTR(CGG)~33~ DNA were digested by RNase T1 and products of the nucleolytic digestion were resolved by denaturing polyacrylamide gel electrophoresis (see Materials and Methods section). RNase T1 cleaves phosphodiester bonds in single-stranded RNA 3′ to unpaired guanine residues whereas paired guanines resist digestion. As seen in [Figure 3](#F3){ref-type="fig"}, the *FMR1*-5′-UTR(CGG)~33~ denatured RNA transcript in lane 1 was cleaved by RNase T1 at each and every guanine residue. By contrast, as indicated in lanes 2 and 3, the enzyme did not digest portions of the native RNA GC-rich *FMR1* 5′-UTR sequence and particularly of the (CGG)~33~ repeat. In fact, nearly all the (CGG)~33~ stretch in native RNA-resisted digestion by RNase T1 except for cleavage at the tenth and twentieth (CGG) trinucleotides that, respectively, precede the (AGG)~11~ and (AGG)~21~ intervening triplets and a faint digestion band at the fifteenth (CGG) trinucleotide. Thus, except for selected unpaired (CGG) trinucleotides, the majority of the guanine residues were engaged in base pairing. These data indicated, therefore, that most of the repeat sequence folded into a secondary structure. Notably, a large portion of the *FMR1* 5′ UTR sequence also resisted digestion by RNase T1 indicating that it too was largely in secondary structure. Figure 3.Ribonuclease T1 protection analysis of FMR1-5′-UTR(CGG)~33~ RNA. Linearized pCS107- *FMR1-*5′-UTR(CGG)~33~ was transcribed in AmpliScribe SP6 system (Epicenter Biotechnologies). The RNA transcripts were isolated, purified and \[5′-^32^P\] end labeled as described under Materials and Methods section. Native or denatured radiolabeled RNA samples were digested by the indicated amounts of RNase T1 under conditions specified in the 'Materials and Methods' section. Shown is a representative phosphor image of ribonuclease digests of RNA samples resolved by 12% polyacrylamide, 8 M urea denaturing gel electrophoresis. We next sought to directly demonstrate the ability of the RNA (CGG)~33~ tract to form an intramolecular secondary structure. Aliquots of \<2.5 nM 5′-^32^P labeled *FMR1*-5′-UTR(CGG)~33~ RNA were incubated at 4°C for 15 min without or in the presence of 120 mM KCl and resolved by electrophoresis in non-denaturing polyacrylamide gels that were, respectively, devoid of salt or contained 10 mM KCl. Representative electrophoregrams shown in [Figure 4](#F4){ref-type="fig"} indicated that the 365-bases-long single-stranded (CGG)~33~ RNA migrated in the absence of salt between DNA size markers of 200 and 250 bp. By contrast, in the presence of KCl, the RNA migrated between DNA markers of 100 and 200 bp. Thus, the (CGG)~33~ RNA sequence formed in the presence of K^+^ ions a rapidly migrating compact structure. The zero-order kinetics of the formation of this species was consonant with it being an intramolecular secondary structure that may represent either a hairpin or quadruplex formation (see Discussion section). Being positioned at the 5′ end of *FMR1* mRNA, such structure could well retard its translation. Figure 4.FMR1-5′-UTR(CGG)~33~ RNA forms K^+^-ion-dependent compact intramolecular secondary structure. 5′-^32^P *FMR1*-5′-UTR(CGG)~33~ RNA was incubated at 4°C for 15 min in intramolecular quadruplex formation mixtures that were devoid of salt or contained 120 mM KCl as described under Materials and Methods section. The RNA mixtures were resolved by electrophoresis in non-denaturing 6% polyacrylamide gels in 0.5× TBE buffer, pH 8.2 that were, respectively, devoid of or contained 10 mM KCl. Shown are autoradiograms of electrophoretically resolved RNA without or with KCl. The salt-dependent different electrophoretic mobility of the RNA can be assessed by their different migration relative to 5′-\[^32^P\] labeled DNA size markers (M). ### Repression of translation by (CGG)~n~ tracts in mRNA is alleviated in vivo by tetraplex (CGG)~n~ destabilizing proteins We next examined in HEK293 cells the *in vivo* effects of premutation-range (CGG) repeat tracts on *FMR-1* promoter-driven transcription and translation of an FL reporter gene. In subsequent experiments, we inquired whether tetraplex (CGG)*~n~* destabilizing proteins affect the expression of an FL gene preceded by (CGG) repeats of different lengths. ### Translation in vivo of FL mRNA is repressed by an upstream (CGG)~99~ RNA tract To investigate the effect of different lengths of upstream (CGG)*~n~* tracts on the *in vivo* expression of the FL reporter gene, HEK293 cells were transfected with pCMV-*FMR1*-5′-UTR(CGG)*~n~*-FL (*n* = 0, 30, 62 or 99). Following 24 h growth, the cells were harvested and lysed and the RL-adjusted relative amounts of FL mRNA and protein were determined and normalized to the transfection efficiency as described under Materials and Methods section. Results of these experiments indicated that relative to a pCMV-*FMR1*-5′-UTR(CGG)*~n~*-FL plasmid with no (CGG)*~n~* tract, the amounts of transcribed FL mRNA for plasmids with 30, 62 or 99 (CGG) repeats increased on average by 1.9-, 1.1- or 5.9-fold, respectively ([Figure 5](#F5){ref-type="fig"}A, see also [Table 1](#T1){ref-type="table"}). Parallel comparison indicated that the relative activity of FL protein encoded by plasmids that had 30, 62 or 99 repeats increased by 2.5-, 1.4- or 3.5-fold, respectively ([Figure 5](#F5){ref-type="fig"}B and [Table 1](#T1){ref-type="table"}). Thus, transcription of the FL gene was enhanced *in vivo* in the presence of 30 or 99 (CGG) repeats but was essentially unaffected by 62 repeats. By contrast (CGG)*~n~* tracts of all sizes raised the levels of FL protein to different extents. Most notable, however, was the different effect of different (CGG)*~n~* sizes on the efficacy of FL mRNA utilization. Tracts of 30 or 62 (CGG) repeats increased the relative efficiency of FL translation by 1.3-fold. However, the longer (CGG)~99~ sequence reduced the relative efficacy of FL translation to 0.6-fold ([Figure 5](#F5){ref-type="fig"}C and [Table 1](#T1){ref-type="table"}). Thus, as revealed by comparing [Figures 1](#F1){ref-type="fig"} and [2](#F2){ref-type="fig"} to [Figure 5](#F5){ref-type="fig"} and [Table 1](#T1){ref-type="table"}, the (CGG)~99~ tract lowered the relative efficacy of FL mRNA translation in both *in vitro* and *in vivo* systems. Figure 5.Effect of (CGG)*~n~* repeat size on the synthesis of FL mRNA and protein in HEK293 cells. Cultured HEK293 cells that were co-transfected with FL reporting pCMV-*FMR1*-5′-UTR(CGG)*~n~*-FL (*n* = 0, 30, 62 or 92) and RL normalizing pCMV-RL plasmids were harvested after 24 h, lysed and aliquots of the cell lysates were used to conduct semi-quantitative RT-PCR measurements of FL and RL mRNA levels and to determine the FL protein activity (Materials and Methods section). (**A**) Histogram of the measured relative levels of FL mRNA as a function of the (CGG)*~n~* tract size. Each column represents an average of the results of nine independent experiments. (**B**) Histogram of the measured relative levels of FL protein activity as a function of the (CGG)*~n~* tract size. Each column represents an average of the results of ten independent experiments. (**C**) Histogram of the calculated ratios of FL protein to mRNA. Table 1.Expression of hnRNP A2 in HEK293 cells increases the efficiency of translation of mRNA that contains 62 or 99 (CGG) repeatsExpressed proteinNonehnRNPA2hnRNP A2 F\[54\]S(CGG) repeatsFL mRNAFL proteinFL protein/ mRNAFL mRNAFL proteinFL protein/ mRNAFL mRNAFL proteinFL protein/ mRNA01.01.01.01.01.0**1.0**1.01.01.0301.9 ± 0.6 (9)2.5 ± 1.2 (10)1.31.3 ± 0.7 (4)1.6 ± 1.2 (3)**1.2**1.7 ± 1.1 (3)1.9 ± 0.2 (2)1.1621.1 ± 0.3 (8)1.4 ± 0.7 (7)1.31.0 ± 0.2 (4)2.0 ± 0.8 (4)**2.0**1.3 ± 0.4 (3)1.1 ± 0.1 (3)0.8995.9 ± 1.7 (6)3.5 ± 1.5 (6)0.61.4 ± 0.3 (4)4.3 ± 0.7 (4)**3.1**3.2 ± 1.1 (3)2.6 ± 1.2 (3)0.8[^1] ### The tetraplex (CGG)~n~ destabilizing protein hnRNP A2 alleviates the repression of FL mRNA by (CGG)~n~ RNA tracts It is reasonable that the source for the diminished efficiency of translation *in vivo* by a (CGG)~99~ tract in mRNA ([Figure 5](#F5){ref-type="fig"}) was the formation of an intramolecular secondary structure that impeded the progression of the translation machinery. Member proteins of the hnRNP family; CBF-A ([@B23],[@B26],[@B27]), hnRNP A2 ([@B23]) and UP-1 ([@B28],[@B29]) were shown to destabilize *in vitro* tetraplex structures of the (CGG)*~n~* sequence. We thus inquired whether the tetraplex (CGG)*~n~*-disrupting protein hnRNP A2 affected the (CGG)*~n~*-modulated transcription or translation of FL *in vivo*. The plasmid pCMV2-Flag-hnRNP A2 which expressed the tetraplex RNA and DNA (CGG)*~n~*-unfolding protein hnRNP A2 under the control of a CMV promoter, was co-transfected into HEK293 cells with the reporter plasmid pCMV-*FMR1*-5′-UTR(CGG)*~n~*-FL (*n* = 0, 30, 62 or 99) and the RL encoding vector pCMV-RL. Transcription and translation of FL and RL in hnRNP A2 expressing cells were compared to their levels in cells that were transfected with the control vectors pCMV2-Flag lacking an hnRNP A2 coding sequence or pCMV2-Flag-hnRNP A2 F\[54\]S that harbored F to S mutation which obliterated the tetraplex (CGG)*~n~* destabilizing activity of hnRNP A2 ([@B23]). Following 24 h growth, the cells were harvested and lysed and the expression of hnRNP A2 or hnRNP A2 F\[54\]S was monitored in the cell lysates by western analysis and normalized levels of FL mRNA and protein were measured (see Materials and Methods section). Results of a set of these experiments are summarized in [Figure 6](#F6){ref-type="fig"} and [Table 1](#T1){ref-type="table"}. Representative western blots of hnRNPA2 and hnRNP A2 F\[54\]S and typical agarose gel resolution of RT-PCR amplified FL and RL reverse transcripts are shown in [Figure 6](#F6){ref-type="fig"}A and B, respectively. It is evident that although the amounts of expressed hnRNP A2 F\[54\]S exceeded those of hnRNPA2, the level of each protein was unaffected by the number of (CGG) repeats in the co-transfected FL reporter gene ([Figure 6](#F6){ref-type="fig"}A). Quantified results of multiple similar determinations of RL normalized levels of FL mRNA and protein activity are summarized in [Table 1](#T1){ref-type="table"} and [Figure 6](#F6){ref-type="fig"}C. Figure 6.Effect of hnRNP A2 and its inactive mutant hnRNP A2 F\[54\]S on the transcription and translation of FL reporter gene preceded by different sizes of (CGG)*~n~* repeat tract. Cultured HEK293 cells were co-transfected with FL reporting plasmid pCMV-*FMR1*-5′-UTR(CGG)*~n~*-FL (*n* = 0, 30, 62 or 99) and RL normalizing pCMV-RL vector and with one of the following plasmids: empty control pCMV2-Flag vector, hnRNP A2 expressing plasmid pCMV2-Flag-*hnRNPA2* or pCMV2-Flag-*hnRNPA2 F\[54\]S* plasmid that expresses mutant hnRNP A2 F\[54\]S. The transfected cells were grown and lysed as detailed in the legend to Figure 5. Aliquots of the cell lysates were used to detect by western analysis the expressed hnRNP A2 proteins, to determine the levels of FL mRNA by semi-quantitative RT-PCR analysis and to measure the activity of FL protein as described under Materials and Methods section. (**A**) Representative western analysis of hnRNP A2 or hnRNP A2 F\[54\]S expressed in cells co-transfected with pCMV-*FMR1*-5′-UTR(CGG)*~n~*-FL plasmids bearing different lengths of the repeat tract. (**B**) Typical electrophoretic separation of FL and RL reverse transcript products of semi-quantitative RT-PCR assay. The FL mRNA RNA molecules, bearing (CGG)*~n~* tracts of the indicated sizes, were produced in cells transfected by empty control pCMV2-Flag vector or in cells that expressed hnRNP A2 or hnRNP A2 F\[54\]S. (**C**) Graphic presentation of average levels of FL mRNA (first panel) and FL protein (second panel) that were produced in control cells and in cells that expressed hnRNP A2 or hnRNP A2 F\[54\]S. Calculated average ratios of FL protein to mRNA in control and hnRNP A2 or hnRNP A2 F\[54\]S expressing cells are shown in the third panel. Data presented in [Table 1](#T1){ref-type="table"} and in the first panel of [Figure 6](#F6){ref-type="fig"}C show that the respective levels of FL (CGG)~99~ mRNA relative to FL (CGG)~0~ mRNA were 5.9- or 3.2-fold higher in cells that did not express hnRNP A2 or that produced mutant hnRNP A2 F\[54\]S protein. By contrast, expression of functional hnRNP A2 brought down FL (CGG)~99~ mRNA to its level in the control cells while augmenting the efficiency of its translation. Relative to RNA with no (CGG) repeat tract, FL (CGG)~99~ translation was increased by 3.5- and 2.6-fold without hnRNP A2 or in the presence of mutant hnRNP A2 F\[54\]S ([Table 1](#T1){ref-type="table"}, [Figure 6](#F6){ref-type="fig"}C, second panel). As a result, in cells that did not have hnRNP A2 or that produced inactive hnRNP A2 F\[54\]S, the ratio of FL protein activity to FL (CGG)~62~ or FL (CGG)~99~ mRNA remained essentially unchanged or was lower by 20--40% relative to control. By clear contrast, in cells that expressed the active tetraplex (CGG)*~n~* resolving protein hnRNP A2, translation of FL (CGG)~62~ or (CGG)~99~ mRNA was elevated 1.4- and 1.2-fold and the ratio of FL protein to mRNA increased by 2.0- or 3.1-fold, respectively ([Table 1](#T1){ref-type="table"}, [Figure 6](#F6){ref-type="fig"}C, third panel). In fact, the stimulation of translation by functional hnRNP A2 may be of even higher relative magnitude since the amounts of expressed hnRNP A2 F\[54\]S significantly exceeded those of hnRNP A2 ([Figure 6](#F6){ref-type="fig"}A). ### The tetraplex (CGG)*~n~* destabilizing protein CBF-A also alleviates the repression of translation by (CGG)~n~ tracts Similar to hnRNP A2, CBF-A another member of the hnRNP family, was shown to disrupt *in vitro* tetraplex structures of (CGG)*~n~* ([@B23],[@B26],[@B27]). We inquired, therefore, whether this protein was also capable of alleviating *in vivo* the repression of translation by premutation-range (CGG)*~n~* mRNA tracts. HEK293 cells were co-transfected with reporter plasmids pCMV-*FMR1*-5′-UTR(CGG)*~n~*-FL (*n* = 0, 30, 62 or 99) and a normalizing pCMV-RL vector. A third co-transfected plasmid was pCMV2-Flag-CBF-A that expressed functional CBF-A protein or one of two control plasmids, pCMV2-Flag devoid of a CBF-A encoding sequence or pCMV2-Flag-CBF-A ΔRNP1~1~ that contained ΔRNP1~1~ mutation in the CBF-A gene that abolished its tetraplex (CGG)*~n~* destabilization activity ([@B26]). Western analyses of active CBF-A and mutant CBF-A ΔRNP1~1~ indicated that both proteins were expressed at similar levels which were unaffected by the size of the (CGG)*~n~* tract in the reporter FL gene ([Figure 7](#F7){ref-type="fig"}A). Typical patterns of agarose gel resolved RT-PCR amplified reverse transcripts of FL and RL mRNA are shown in [Figure 7](#F7){ref-type="fig"}B. Results of multiple measurements of RL normalized levels of FL mRNA and protein as a function of the (CGG)*~n~* size and of the expression or lack thereof of functional CBF-A are summarized in [Table 2](#T2){ref-type="table"} and in [Figure 7](#F7){ref-type="fig"}C. Data indicated that relative to FL gene with no (CGG)*~n~* tract, 30 or 62 (CGG) repeats modestly increased the levels of FL mRNA by up to ∼2-fold whether or not CBF-A or its mutant CBF-A ΔRNP1~1~ were expressed. Moreover, the longer (CGG)~99~ tract affected a steeper increase of FL mRNA of 5.9-, 7.2- or 6.4-fold in cells that were, respectively, transfected with empty vector, expressed CBF-A or harbored CBF-A ΔRNP1~1~ ([Table 1](#T1){ref-type="table"}, [Figure 7](#F7){ref-type="fig"}C, first panel). Thus, by promoting FL (CGG)~99~ mRNA transcription, active CBF-A differed from functional hnRNP A2 which brought down its production to the (CGG)~0~ mRNA level ([Table 1](#T1){ref-type="table"}, see Discussion section). Relative to FL gene with no repeat tract, the presence of 30 or 62 (CGG) repeats moderately increased the levels of FL protein whether or not the cells expressed CBF-A or CBF-A ΔRNP1~1~. By contrast, whereas the relative translation of FL (CGG)~99~ mRNA was elevated ∼4-fold in cells without or with CBF-A ΔRNP1~1~, the level of FL protein increased by almost 14-fold in cells that expressed CBF-A ([Table 2](#T2){ref-type="table"} and [Figure 7](#F7){ref-type="fig"}C, second panel). As a result, whereas the relative efficacy of FL mRNA translation as manifested by the FL protein to mRNA ratio was depressed to 0.6- and 0.7-fold in control or CBF-A ΔRNP1~1~ producing cells, respectively, it increased by 1.9-fold in functional CBF-A expressing cells ([Table 2](#T2){ref-type="table"} and [Figure 7](#F7){ref-type="fig"}C, third panel). Hence, despite their different effects on the transcription of FL (CGG)~99~ mRNA, both the tetraplex (CGG)*~n~* destabilizing proteins hnRNP A2 and CBF-A improved the efficiency of its translation. Figure 7.Effect of CBF-A and its inactive mutant CBF-A ΔRNP1~1~ on the transcription and translation of FL reporter gene preceded by different sizes of (CGG)*~n~* repeat tract. Cultured HEK293 cells were co-transfected with FL reporting plasmid pCMV-*FMR1*-5′-UTR(CGG)*~n~*-FL (*n* = 0, 30, 62 or 99) and RL normalizing pCMV-RL vector and with one of the following plasmids: empty pCMV2-Flag vector, CBF-A expressing plasmid pCMV2-Flag-*CBF-A* or pCMV2-Flag-*CBF-A ΔRNP1~1~* plasmid that expressed mutant CBF-A ΔRNP1~1~. The cells were grown and lysed as detailed in the legend to Figure 5. Aliquots of the cell lysates were used to detect by western analysis the expressed CBF-A proteins, to determine the levels of FL mRNA by semi-quantitative RT-PCR analysis and to measure the activity of FL protein as described under Materials and Methods section. (**A**) Representative western analysis of CBF-A or CBF-A ΔRNP1~1~ expressed in cells co-transfected with pCMV-*FMR1*-5′-UTR(CGG)*~n~*-FL bearing different lengths of the repeat tract. (**B**) Typical electrophoretic separation of FL and RL reverse transcript products of semi-quantitative RT-PCR assay. The FL mRNA RNA molecules, bearing (CGG)*~n~* tracts of the indicated sizes, were produced in cells transfected by empty control pCMV2-Flag vector or in cells that expressed CBF-A or CBF-A ΔRNP1~1~. (**C**) Graphic presentation of average levels of FL mRNA (first panel) and FL protein (second panel) that were produced in control cells and in cells that expressed hnRNP A2 or hnRNP A2 F\[54\]S. The third panel is a fit spline plot of the ratios of FL protein to mRNA. Calculated average ratios of FL protein to mRNA in control and CBF-A or CBF-A ΔRNP1~1~ expressing cells are shown in the third panel. Table 2.Expression of CBF-A increases the efficiency of translation in HEK293 cells of mRNA that contains 62 or 99 (CGG) repeatsExpressed proteinNoneCBF-ACBF-A ΔRNP1~1~(CGG) repeatsFL mRNAFL proteinFL protein/ mRNAFL mRNAFL proteinFL protein/ mRNAFL mRNAFL proteinFL protein/ mRNA01.01.01.01.01.0**1.0**1.01.01.0301.9 ± 0.6 (9)2.5 ± 1.2 (10)1.32.1 ± 0.3 (5)5.0 ± 1.2 (5)**2.3**1.6 ± 0.4 (5)3.3 ± 0.8 (3)2.1621.1 ± 0.3 (8)1.4 ± 0.7 (7)1.31.8 ± 0.3 (4)3.6 ± 0.4 (3)**2.0**1.6 ± 0.3 (4)3.1 ± 0.1 (2)1.9995.9 ± 1.7 (6)3.6 ± 1.5 (6)0.67.2 ± 1.7 (4)13.7 ± 3.7 (4)**1.9**6.4 ± 1.9 (5)4.4 ± 0.5 (3)0.7[^2] DISCUSSION ========== Whereas *FMR1* transcription is completely silenced in fragile X full mutation cells ([@B3]), premutation carriers produce 5- to 10-fold higher amounts of *FMR1* mRNA than normal cells ([@B10],[@B11]). Yet, despite the elevated levels of *FMR1* mRNA in carrier cells, they produce FMRP at amounts that are within or below the norm ([@B10],[@B12],[@B13]). It thus appears that *FMR1* translation is repressed by the moderately expanded premutation (CGG) repeat sequence. Indeed, evidence was presented to show that premutation (CGG) repeat sizes are linked to lowered association of *FMR1* mRNA with lymphoblastoid cell polysomes ([@B14]) and that the 40S ribosomal subunits stall along *FMR1* mRNA molecules that contain more than 200 repeats ([@B15]). Data indicated that (CGG)*~n~* tracts in RNA fold into hairpin ([@B16],[@B17],[@B19],[@B35]) and tetraplex ([@B23]) structures. A reasonable conjecture, therefore, is that the diminished efficacy of the translation of premutation *FMR1* mRNA is due to retarded ribosome association and stalled progression at secondary structures of the (CGG)*~n~* tract in premutation *FMR1* mRNA. In this work, we showed that premutation-range (CGG)*~n~* tracts selectively reduced the efficiency of mRNA translation both *in vitro* and *in vivo*. Second, following a demonstration that (CGG)*~n~* sequence in RNA was capable of forming quadruplex structures, we showed that the expression *in vivo* of proteins that untangle tetraplex structures of (CGG)*~n~* increased the efficacy of mRNA utilization. Relative to FL reporter gene with no (CGG)*~n~* upstream tract, the introduction of 30 (CGG) repeats upstream to the gene stimulated its *in vitro* transcription and translation in coupled T7 promoter-driven transcription--translation system ([Figure 1](#F1){ref-type="fig"}) and in *FMR1* promoter-driven transcription and translation *in vivo* ([Figure 5](#F5){ref-type="fig"}). The most frequent normal number of (CGG) repeats in the *FMR1* gene and its transcript in normal subjects is 29--33 ([@B34]). *In vivo* expression of the FL gene ([Figure 5](#F5){ref-type="fig"}) suggested that secondary structures of the (CGG)~29--33~ tracts may promote maximum efficacy of translation *FMR1* mRNA in cells. Such requirement may exert positive selective pressure that results in the preservation of the (CGG)~29--33~ repeat sequence in the human population. As our results showed, 62 or 99 (CGG) repeats that represent the lower range of premutation size slightly repressed ([Figure 2](#F2){ref-type="fig"}) or did not affect ([Figure 1](#F1){ref-type="fig"}) the *in vitro* transcription of the FL gene. Most significantly, both 62 and 99 (CGG) repeats did sharply suppress FL translation both in coupled or in separate *in vitro* systems ([Figures 1](#F1){ref-type="fig"} and [2](#F2){ref-type="fig"}). The observed declining ratio of FL protein to mRNA indicated that the premutation repeat tracts reduced the efficiency of utilization of the FL mRNA. Results presented in [Figures 3](#F3){ref-type="fig"} and [4](#F4){ref-type="fig"} demonstrated that the (CGG)~33~ RNA repeat maintained *in vitro* a secondary structure and that a large portion of the guanine-rich *FMR1* 5′ UTR sequence itself also folded into secondary configuration ([Figure 3](#F3){ref-type="fig"}). It is plausible that the nature and geometry of secondary structures that (CGG)*~n~* RNA tracts formed independently or together with the 5′ UTR tract differed for different sizes of the repeat sequence. Thus, the spatial arrangement of secondary structures of the normal 30 repeats may have promoted the association of mRNA with polysomes and enhanced ribosome progression whereas the different geometries of secondary structures of premutation sized 62 or 99 repeats impeded the translation machinery and limited the efficacy of mRNA utilization. Formation and stabilization of quadruplex nucleic acids require the presence of alkali ion. Thus, the potassium ion dependence of the formation of intramolecular complex by the (CGG)~33~ RNA sequence ([Figure 4](#F4){ref-type="fig"}) was consistent with tetrahelical nature of this secondary structure. In contrast to tetraplex formation, hairpins can be generated in the absence of ions ([@B16; @B17; @B18; @B19]). However, the possibility remained that under our experimental conditions K^+^ ions were critical for the stability of hairpins. Hence, although the data presented in [Figure 4](#F4){ref-type="fig"} were consonant with a tetraplex structure of the (CGG)~33~ RNA sequence, the possibility that the observed secondary structure was a hairpin could not be ruled out. Transcription *in vivo* of *FMR1* promoter-driven FL reporter gene and translation of the transcripts in HEK293 cells respond differently to the presence of upstream premutation (CGG)*~n~* tracts than do T7 promoter-driven *in vitro* transcription and translation. Yet in both the *in vitro* and *in vivo* systems, premutation size repeat tracts ultimately lowered the efficacy of mRNA utilization. Unlike the essentially unaltered extent of transcription *in vitro* ([Figures 2](#F2){ref-type="fig"} and [3](#F3){ref-type="fig"}), the amount of FL (CGG)~30~ mRNA relative to FL (CGG)~0~ mRNA increased 2-fold, FL (CGG)~62~ mRNA remained unchanged and FL (CGG)~99~ mRNA was elevated by nearly 6-fold ([Figure 5](#F5){ref-type="fig"}). It is plausible that transcription efficiency was differentially affected by distinct (CGG)*~n~* configurations in DNA such that different lengths of the repeat sequence formed secondary structures geometries that were better targets for the transcription factors than others. The modulation of transcription by different lengths of the (CGG) repeat in DNA could also reflect different degree of the *in vivo* chromatinization of the different lengths of the repeat tracts and dissimilar nucleosome positioning along these sequences. The observed higher amounts of transcripts that were produced *in vivo* in the presence of 99 (CGG) triplets was in line with the reported increased accumulation of *FMR1* mRNA in leukocytes of premutation carriers ([@B10],[@B11]). The contrasting unchanged or reduced transcript accumulation *in vitro* was likely to be due to the use of T7 in place of the *FMR1* promoter and to the dissimilar reaction conditions of the *in vitro* and *in vivo* transcription systems. The (CGG)*~n~* tracts had also opposing effect on translation *in vitro* and *in vivo*. Whereas the relative extent of translation of FL mRNA molecules that contained 62 or 99 (CGG) repeats was sharply diminished *in vitro* relative to 30 repeats ([Figures 1](#F1){ref-type="fig"} and [2](#F2){ref-type="fig"}), it decreased in HEK293 cells for 62 repeats but was somewhat increased at 99 repeats ([Figure 5](#F5){ref-type="fig"} and [Table 1](#T1){ref-type="table"}). The observed elevated level of protein in the transfected cultured cells was inconsistent with the reported unchanged or slightly reduced amounts of FMRP in cells of premutation carriers ([@B10],[@B12],[@B13]). Although the reason for this discrepancy is not clear, cells of premutation carriers ([@B14]), transfected HEK293 cells ([Figure 5](#F5){ref-type="fig"} and [Table 1](#T1){ref-type="table"}) and *in vitro* translation systems ([Figures 1](#F1){ref-type="fig"} and [2](#F2){ref-type="fig"}) all displayed reduced translation efficiency of premutation mRNA. A likely common denominator for all the systems may, therefore, be the impediment to the translation machinery by tetraplex structures of the intermediately expanded (CGG)*~n~* tract. Results summarized in [Figures 6](#F6){ref-type="fig"} and [7](#F7){ref-type="fig"} and [Tables 1 and 2](#T1 T2){ref-type="table"} indicated that the RNA binding and tetraplex DNA and RNA-disrupting proteins hnRNP A2 and CBF-A increased the efficacy of (CGG)*~n~* mRNA translation *in vivo*. Based on our observation that (CGG)~33~ RNA formed intramolecular secondary structure ([Figures 3](#F3){ref-type="fig"} and [4](#F4){ref-type="fig"}), we assumed that a stable (CGG)~99~ unimolecular folded domain in FL mRNA was responsible for its retarded translation *in vitro* ([Figures 1](#F1){ref-type="fig"}B, C and 2B) and *in vivo* ([Figure 6](#F6){ref-type="fig"}). In such a case, untangling of the secondary structure should increase the efficiency of translation. We demonstrated in the past that intermolecular tetraplex structures of (CGG)*~n~* in DNA were disrupted *in vitro* by members of the hnRNP family CBF-A and hnRNP A2 ([@B23],[@B26],[@B27]). We also showed that mutations in each of the two conserved domains of both proteins, the RNP1~1~ box and the ATP/GTP-binding fold, specifically abolished their *in vitro* quadruplex disruption activity ([@B23; @B26]). Results described here showed that expression of functional hnRNP A2 in HEK293 cells, but not of its inactive hnRNP A2 F\[54\]S mutant, alleviated the repression of FL mRNA translation by a (CGG)~99~ tract ([Figure 6](#F6){ref-type="fig"}C and [Table 1](#T1){ref-type="table"}). Also, active CBF-A, but not its ΔRNP1~1~ mutant, increased the efficiency of the *in vivo* translation of (CGG)~99~ FL mRNA ([Figure 7](#F7){ref-type="fig"}C and [Table 2](#T2){ref-type="table"}). Although the hnRNP A2 F\[54\]S and CBF-A ΔRNP1~1~ mutants lost their capacity to disrupt tetraplex structures of the (CGG)*~n~* sequence, they maintained their nucleic acids binding activity (Cohen, E. and Weisman-Shomer, P., unpublished data). It is likely, therefore, that the enhancement of (CGG)*~n~* mRNA translation by hnRNP A2 and CBF-A was due to their quadruplex destabilizing activity rather than to their RNA-binding capacity. Whereas both hnRNP A2 and CBF-A were shown to disrupt quadruplex structures of (CGG)*~n~* in DNA, only hnRNP A2 but not CBF-A destabilized *in vitro* a bimolecular quadruplex structure of synthetic (CGG)~7~ RNA ([@B23]). Because the *in vitro* and *in vivo* conditions for CBF-A action were different, since the likely RNA structure that formed *in vivo* was intra- rather than intermolecular ([Figure 4](#F4){ref-type="fig"}), and as the CBF-A ΔRNP1~1~ mutant failed to stimulate translation, we assume that similar to hnRNP A2, CBF-A also enhanced the translation of FL (CGG)~99~ mRNA by untangling its monomolecular secondary structure. The observed correlation between the ability of hnRNP A2 or CBF-A to unfold quadruplex structures of RNA or DNA and their capacity to alleviate the blocking of translation by the (CGG)~99~ tract, was consistent with the folded repeat sequence being a tetrahelix. However, as discussed above, the observed intramolecular secondary structure of the (CGG)*~n~* repeat may represent a hairpin rather than quadruplex. Such scenario raises the untested prospect that hnRNP A2 and CBF-A can also destabilize hairpins and that similar to their tetraplex unfolding activity, the unwinding of hairpins also requires that the proteins possess intact RNP1 and ATP/GTP motifs. Taken together, our results lent credence to the notion that secondary structures of the expanded (CGG) repeat tract in fragile X premutation mRNA hindered translation and that their disruption alleviated this impediment. This conclusion gained support recently by the report that a quadruplex motif in the 5′ UTR of the human NRAS oncogene mRNA modulated its translation *in vitro*. Furthermore, bioinformatics analysis predicted that 5′ UTR elements in 2922 other human mRNA species can assume tetraplex structure that may potentially regulate their translation ([@B37]). These results, as well as the present report on the capacity of quadruplex destabilizing proteins to alleviate the impediment to translation that a folded (CGG)*~n~* tract posed, are in line with the notion that tetrahelical motifs in mRNA may modulate its translation. Although functional hnRNP A2 and CBF-A augment the translation of FL (CGG)~99~ mRNA, their mechanisms of action appeared to be dissimilar. Whereas the level of FL (CGG)~99~ mRNA was 5.9-fold higher than that of FL mRNA with no repeat tract, hnRNP A2 lowered the amount of FL (CGG)~99~ mRNA to that of the FL (CGG)~0~ mRNA. At the same time, hnRNP A2 did not affect the amounts of accumulated FL protein that FL (CGG)~99~ mRNA encoded ([Table 1](#T1){ref-type="table"} and [Figure 6](#F6){ref-type="fig"}C). It is possible that hnRNP A2 suppressed transcription by affecting the conformation and chromatinization of the *FMR1* promoter through its direct interaction with the adjacent (CGG)~99~ repeat tract. Yet, although hnRNP A2 lowered the amount of FL (CGG)~99~ mRNA, it increased the relative efficacy of its utilization by 5.2-fold, as reflected by the rising ratio of FL protein to mRNA from 0.6 to 3.1 ([Table 1](#T1){ref-type="table"} and [Figure 6](#F6){ref-type="fig"}C). By contrast, functional CBF-A increased the relative amount of FL (CGG)~99~ transcripts and amount of FL protein by 7.2- and 13.7-fold, respectively ([Table 2](#T2){ref-type="table"} and [Figure 7](#F7){ref-type="fig"}C). As a result of the accumulation of FL protein in excess over FL mRNA, CBF-A augmented the efficacy of mRNA usage and increased the ratio of FL protein to mRNA by 3.2-fold from 0.6 to 1.9 ([Table 2](#T2){ref-type="table"} and [Figure 7](#F7){ref-type="fig"}C). The different patterns of translation enhancement by hnRNP A2 and CBF-A is not surprising. In addition to its quadruplex (CGG)*~n~* untangling activity, CBF-A also binds single-stranded DNA and tetraplex forms of some sequences whereas hnRNP A2 also binds RNA and interacts with some DNA structures. It is plausible, therefore, that because of their binding to different targets in DNA or RNA in the cell in addition to their direct interaction with tetraplex (CGG)*~n~* in RNA, each protein differently affects transcription and translation. The authors are grateful to Dr Paul Hagerman, UC Davies for fruitful discussions. We wish to thank the three expert referees and the Executive Editor for their excellent critique of this article. This study was supported by grants to M.F. from Conquer Fragile X Foundation, USA, the Israel Science Foundation and the United States-Israel Binational Science Foundation. S.K. was supported by an Israel Science Ministry doctoral scholarship. Funding to pay the Open Access publication charges for the article were provided by the Israel Science Foundation and a VATAT doctoral fellowship to S.K. *Conflict of interest statement*. None declared. [^1]: The calculated average ratios of FL protein to mRNA are in bold. [^2]: The calculated average ratios of FL protein to mRNA are in bold.

\section{Introduction} \label{sec:intro} Network tomography, first proposed in \cite{Vardi96}, is the science of inferring spatially localized network behavior using only metrics that are practically feasible to measure. The problems considered in network tomography can be classified into two broad strands: i) traffic demand tomography---determination of source destination traffic volumes via measurements of link volumes and ii) network delay tomography---link parameter estimation based on end to end path level measurements. However, central to both these areas, is the problem of inferring the statistics of $X,$ a vector of independent non-negative random variables, given the measurement model $Y = AX.$ The challenge in these problems stems from the fact that $A$ is usually an ill posed matrix and hence non-invertible. For excellent tutorials and surveys on the state of art, see \cite{Adams00,Coates02b,Castro04} and \cite{Lawrence07}. For sake of definiteness, we consider here the problem of network delay tomography. The proposed method, however, is also applicable to traffic demand tomography. Under delay tomography, the major problems studied include estimation of bottleneck link bandwidths, e.g. \cite{Liu07,Dey11}, link loss rates, e.g. \cite{Caceres99}, link delays, e.g., \cite{Coates01,LoPresti02,Shih03,Tsang03}, etc. For successful estimation, the proposed solutions to these problems crucially require i) end-to-end data generated using multicast probe packets, real or emulated, and ii) the network to be a tree rooted at a single sender with destinations at leaves. These algorithms mainly exploit the correlations in the path measurements, i.e., the packets have the same experience on shared links. Because of this, any divergence in either of the above requirements results in performance degradation. Consequently, there is a need to develop tomography algorithms for networks with arbitrary topologies using only pure unicast probe packet measurements. Mathematically, this is same as addressing the generalized network tomography (GNT) problem, wherein, given the binary matrix $A,$ the objective is to estimate the statistics of $X,$ a vector of independent non-negative random variables, using only IID samples of the components of the random vector $Y = AX.$ In this paper we propose a novel method, henceforth called the distribution tomography (DT) scheme, for the framework of GNT to accurately estimate the distribution of $X$ even when no prior knowledge about the same is available. We rely on the fact that the class of \textit{generalized hyperexponential} (GH) distributions is dense in the set of non-negative distributions (see \cite{Botha86}). Using this, the idea then is to approximate the distribution of each component of $X$ using linear combinations of known exponential bases and estimate the unknown weights. These weights are obtained by solving a set of polynomial systems based on the moment generating function of the components of $Y.$ For unique identifiability, it is only required that every pair of columns of the matrix $A$ be linearly independent, a property that holds true for the routing matrices of all multicast tree networks. The rest of the paper is organized as follows. In the next section, we develop the notation and formally describe the problem. Section~\ref{sec:approxDist} recaps the theory of approximating non-negative distributions using linear combinations of exponentials. In Sections~\ref{sec:GenNT} and \ref{sec:universality}, we develop our proposed method and demonstrate its universal applicability. We give numerical examples in Section~\ref{sec:ExptResults} and end with a short discussion in Section~\ref{sec:ConclFutW}. \section{Model and Problem Description} \label{sec:model} Any cumulative distribution function (CDF) that we work with is always assumed to be continuous with support $(0, \infty).$ The moment generating function (MGF) of the random variable $X$ will be $M_X(t) = \mathbb{E}(\exp(-tX)).$ For $n \in \mathbb{N},$ we use $[n]$ and $S_n$ to represent respectively the set $\{1, \ldots, n\}$ and its permutation group. We use the notation $\mathbb{R}, \mathbb{R}_{+}$ and $\mathbb{R}_{++}$ to denote respectively the set of real numbers, non-negative real numbers and strictly positive real numbers. In the same spirit, for integers we use $\mathbb{Z}, \mathbb{Z}_{+}$ and $\mathbb{Z}_{++}.$ All vectors are column vectors and their lengths refer to the usual Euclidean norm. For $\delta > 0,$ $B(\mathbf{v}; \delta)$ represents the open $\delta-$ball around the vector $\mathbf{v}.$ To denote the derivative of the map $f$ with respect to $\mathbf{x},$ we use $\dot{f}(\mathbf{x}).$ Lastly, all empty sums and empty products equal $0$ and $1$ respectively. Let $X_1, \ldots, X_N$ denote the independent non-negative random variables whose distribution we wish to estimate. We assume that each $X_j$ has a GH distribution of the form \begin{equation} \begin{array}{cccc} F_j(u) & = & \sum_{k=1}^{d + 1} w_{jk} \left[ 1 - \exp\left( -\lambda_{k} u \right) \right], & u \geq 0 \end{array} \label{eq:Dist} \end{equation} where $\lambda_{k}>0,$ $\sum_{k=1}^{d+1} w_{jk} \lambda_{k} \exp(-\lambda_{k}u) \geq 0$ and $\sum_{k=1}^{d + 1}w_{jk} = 1.$ Further, we suppose that $\lambda_1, \ldots, \lambda_{d + 1}$ are distinct and explicitly known and that the weight vectors of distinct random variables differ at least in one component. Let $A \in \{0,1\}^{m \times N}$ denote an a priori known matrix which is $1-$identifiable in the following sense. \begin{definition} A matrix $A$ is $k-$identifiable if every set of $2k$ of its columns is linearly independent. \end{definition} Let $X \equiv (X_1, \ldots, X_N)$ and $Y = AX.$ For each $i \in [m],$ we presume that we have access to a sequence of IID samples of the $i^{th}$ linear combination $Y_i.$ Our problem then is to estimate for each $X_j,$ its vector of weights $\mathbf{w}_j \equiv (w_{j1}, \ldots, w_{jd})$ and consequently its complete distribution $F_j,$ since $w_{j(d+1)} = 1 - \sum_{k = 1}^{d} w_{jk}.$ Before developing the estimation procedure, we begin by making a case for the distribution model of \eqref{eq:Dist}. \section{Approximating Distribution Functions} \label{sec:approxDist} Consider the problem of simultaneously estimating the members of a finite family of arbitrary distributions, say $\mathcal{G}=\{G^1, \ldots, G^N\}.$ A useful strategy is to approximate each member by a GH distribution. The CDF of a GH random variable $X$ is given by \begin{equation} \begin{array}{cc} F_{X}(u) = \sum_{k=1}^{d + 1} \alpha_k \left[ 1 - \exp\left( -\lambda_{k} u \right) \right], & u \geq 0, \end{array} \label{eq:GHDist} \end{equation} where $\lambda_{k}>0,$ $\sum_{k=1}^{d + 1}\alpha_{k}\lambda_{k}\exp(-\lambda_{k}u)\geq0$ and $\sum_{k=1}^{d + 1}\alpha_{k}=1.$ Consequently, its MGF is given by \begin{equation} M_X(t)=\sum_{k=1}^{d + 1}\alpha_{k}\frac{\lambda_{k}}{\lambda_{k} + t}. \label{eq:GHMGF} \end{equation} In addition to the simple algebraic form of the above quantities, the other major reason to use the GH class is that, in the sense of weak topology, it is dense in the set of distributions (see \cite{Botha86}). In fact given a continuous CDF $F$ with MGF $M$, one can explicitly find a sequence of GH distributions, say $F_n,$ that converge uniformly to it. Furthermore, if $M_n$ is the MGF for $F_n,$ then for each $t \geq 0,$ $M_n(t)\rightarrow M(t).$ The idea is based on the following result. \begin{theorem} \cite{Ou97} \label{thm:Ou} For $n,k \in \mathbb{N},$ let $X_{n,k}$ be a nonnegative GH random variable with mean $k/n,$ variance $\sigma_{n,k}^{2}$ and CDF $W_{n,k}.$ Suppose \begin{enumerate} \item the function $\nu:\mathbb{N}\rightarrow\mathbb{N}$ satisfies $\underset{n\rightarrow\infty}{lim}\nu(n)/n=\infty.$ \item there exists $0<s<1$ such that $\underset{n\rightarrow\infty}{lim}n^{1+s} \sigma_{n,k}^{2}/k=0$ uniformly with respect to $k.$ \end{enumerate} Then given any continuous non-negative distribution function $F,$ the following holds: \begin{enumerate} \item the function $F_{n}$ given by \begin{eqnarray*} F_{n}(u) & = & \sum_{k=1}^{\nu(n)}\left\{F(k/n)-F((k-1)/n)\right\} W_{n,k}(u)\\ & & +(1-F(\nu(n)/n))W_{n,\nu(n)+1}(u) \end{eqnarray*} is a GH distribution for every $n\in\mathbb{N}$ and \item $F_{n}$ converges uniformly to $F,$ i.e., \begin{equation*} \underset{n \rightarrow \infty}{\lim} \underset{-\infty < u < \infty}{\sup} \left|F_{n}(u)-F(u)\right| = 0. \end{equation*}\, \end{enumerate} \end{theorem} Observe that for each $n,$ the exponential stage parameters of $F_n$ depend only on the choice of the random variables $\{X_{n,k}\!: 1\leq k\leq \nu(n) +1\}.$ Regarding $\mathcal{G},$ if we fix the random variables $X_{n,k}$ and let $M^i$ denote the MGF of $G^i,$ then this observation and the above result implies that for any given $\epsilon_1, \epsilon_2 >0$ and any finite set $\tau=\{t_1,\ldots ,t_k\} \subset \mathbb{R}_+,$ $\exists n \equiv n(\epsilon_1, \epsilon_2, \tau) \in \mathbb{N}$ such that for each $i\in[N],$ $G^i$ and its $n^{th}$ GH approximation, $F_n^i,$ are $\epsilon_1-$ close in the sup norm and for each $j \in [k],$ $|M^i(t_j)-M_n^i(t_j)|\leq \epsilon_2.$ Further, the exponential stage parameters are explicitly known and identical across the approximations $F_n^i,$ which now justifies our model of \eqref{eq:Dist}. The problem of estimating the members of $\mathcal{G}$ can thus be reduced to determining the vector of weights that characterizes each approximation and hence each distribution. \section{Distribution Tomography Scheme} \label{sec:GenNT} The outline for this section is as follows. For each $i,$ we use the IID samples of $Y_i$ to estimate its MGF and subsequently build a polynomial system, say $H_i(\mathbf{x}) = 0.$ We call this the elementary polynomial system (EPS). We then show that for each $i \in [m]$ and each $j \in p_i := \{j \in [N]: a_{ij} = 1\},$ a close approximation of the vector $\mathbf{w}_j$ is present in the solution set of $H_i(\mathbf{x}) = 0,$ denoted $V(H_i).$ To match the weight vectors to the corresponding random variables, we make use of the fact that $A$ is $1-$identifiable. \subsection{Construction of Elementary Polynomial Systems} \label{subsec:ConstPolySys} Fix $i \in [m]$ and suppose that $|p_i| = N_i,$ i.e., $Y_i$ is a sum of $N_i$ random variables, which for convenience, we relabel as $X_1, \ldots, X_{N_i}$ in some order. Using \eqref{eq:Dist} and \eqref{eq:GHMGF}, observe that the MGF of $Y_i$ is well defined $\forall t \in \mathbb{R}_{++}$ and satisfies the relation \begin{equation} \label{eq:MGFRel} M_{Y_{i}}(t) = \prod_{j = 1}^{N_i} \left\{ \sum_{k = 1}^{d + 1}w_{jk} \left( \frac{\lambda_{k}}{\lambda_{k}+t} \right) \right\}. \end{equation} On simplification, after substituting $w_{j(d+1)} = 1 - \sum_{k=1}^{d}w_{jk},$ we get \begin{equation} \label{eq:modMGFRel} \mu_i(t) = \left\{ \prod_{j = 1}^{N_i} \left[ \sum_{k=1}^{d}w_{jk}\Lambda_k(t) + \lambda_{d + 1}\right] \right\}, \end{equation} where $\Lambda_k(t) = (\lambda_{k} - \lambda_{d + 1})t/(\lambda_{k} + t)$ and $\mu_i(t) = M_{Y_{i}}(t) (\lambda_{d + 1} + t)^{N_i} .$ For now, let us assume that we know $M_{Y_i}(t)$ and hence $\mu_i(t)$ exactly for every valid $t.$ We will refer henceforth to this situation as the ideal case. Treating $t$ as a parameter, we can then use \eqref{eq:modMGFRel} to define a canonical polynomial \begin{equation} f(\mathbf{x};t) = \left\{ \prod_{j = 1}^{N_i} \left[ \sum_{k=1}^{d}x_{jk}\Lambda_k(t) + \lambda_{d + 1}\right] \right\} - \mu_i(t), \label{eq:tempGenPoly} \end{equation} where $\mathbf{x} \equiv (\mathbf{x}_1, \ldots, \mathbf{x}_{N_i})$ with $\mathbf{x}_j \equiv (x_{j1}, \ldots, x_{jd}).$ As this is a multivariate map in $d \cdot N_i$ variables, we can choose an arbitrary set $\tau = \{t_1, \ldots, t_{d \cdot N_i}\} \subset \mathbb{R}_{++}$ consisting of distinct numbers and define an intermediate square polynomial system \begin{equation} \label{eq:tempPolSys} F_\tau(\mathbf{x}) \equiv (f_1(\mathbf{x}), \ldots, f_{d \cdot N_i}(\mathbf{x})) = 0, \end{equation} where $f_k(\mathbf{x}) \equiv f(\mathbf{x}; t_k).$ Since \eqref{eq:tempPolSys} depends on choice of $\tau,$ analyzing it directly is difficult. But observe that i) the expansion of each $f_n$ or equivalently \eqref{eq:tempGenPoly} results in rational coefficients in $t$ of the form $\Lambda^\mathbf{L} \equiv \Lambda_{1}^{L_{1}}(t)\cdots\Lambda_{d}^{L_{d}}(t)\lambda_{d+1}^{L_{d+1}},$ where $L_k \in \mathbb{Z}_{+}$ and $\sum_{k=1}^{d+1}L_k = N_{i},$ and ii) the monomials that constitute each polynomial are identical. This suggests that one may be able to get a simpler representation for \eqref{eq:tempPolSys}. We do so in the following three steps, where the first two focus on simplifying \eqref{eq:tempGenPoly}. \noindent \emph{Step1-Gather terms with common coefficients}: Let \[ \Delta_{d+1,N_{i}}:=\left\{ \mathbf{L}\equiv\left(L_{1},\ldots,L_{d+1}\right)\in\mathbb{Z}_{+}^{d+1}:\sum_{k=1}^{d+1}L_{k}=N_{i}\right\}. \] For a vector $\mathbf{b}\equiv(b_{1},\ldots,b_{N_{i}})\in[d+1]^{N_{1}},$ let its type be denoted by $\Theta(\mathbf{b})\equiv(\theta_{1}(\mathbf{b}),\ldots,\theta_{d+1}(\mathbf{b})),$ where $\theta_{k}(\mathbf{b})$ is the count of the element $k$ in $\mathbf{b}.$ For every $\mathbf{L}\in\Delta_{d+1,N_{i}},$ additionally define the set \[ \mathcal{B}_{\mathbf{L}}=\left\{ \mathbf{b}\equiv(b_{1},\ldots,b_{N_{i}})\in[d+1]^{N_{i}}:\,\Theta(\mathbf{b})=L\right\} \] and the polynomial $g(\mathbf{x};\mathbf{L})=\sum_{\mathbf{b}\in\mathcal{B}_{L}}(\prod_{j=1,\, b_{j}\ne d+1}^{N_{i}}x_{jb{}_{j}}).$ Then collecting terms with common coefficients in \eqref{eq:tempGenPoly}, the above notations help us rewrite it as \begin{equation} f(\mathbf{x};t)=\sum_{\mathbf{L}\in\Delta_{d+1,N_{i}}}g(\mathbf{x};\mathbf{L})\Lambda^{\mathbf{L}}-\mu_{i}(t).\label{eq:afterColComTerms} \end{equation} \noindent \emph{Step2-Coefficient expansion and regrouping}: Using an idea similar to partial fraction expansion for rational functions in $t$, the goal here is to decompose each coefficient into simpler terms. For each $j,$ $k\in[d],$ let \[ \beta_{jk}:= \begin{cases} \frac{\lambda_{j}\left(\lambda_{k}-\lambda_{d+1}\right)}{\left(\lambda_{j}-\lambda_{k}\right)} & \mbox{if }j\ne k\\ 1 & j=k \end{cases}. \] For each $\mathbf{L} \in \Delta_{d + 1, N_i},$ let $\mathcal{D}(\mathbf{L}):= \{k\in[d]: L_{k}>0\}.$ Further, if $\mathcal{D}(\mathbf{L}) \neq \emptyset$ then $\forall k \in \mathcal{D}(\mathbf{L})$ and $\forall q \in [L_k],$ let \begin{eqnarray*} \bar{\Delta}_{kq}(\mathbf{L}) & := & \{\mathbf{s}\equiv(s_{1},\ldots,s_{d})\in\mathbb{Z}_{+}^{d}: s_{r} = 0\\ & &\forall r\in\mathcal{D}(\mathbf{L})^{c} \cup \{k\}; \sum_{n=1}^{d}s_n = L_{k}-q\} \mbox{ and} \end{eqnarray*} \[ \gamma_{kq}(\mathbf{L}):=\prod_{r\in\mathcal{D}(\mathbf{L})}\beta_{kr}^{L_{r}}\left\{\sum_{\mathbf{s}\in\bar{\Delta}_{kq}(\mathbf{L})}\prod_{r\in\mathcal{D}(\mathbf{L})}\tbinom{L_{r}+s_{r}-1}{L_{r}-1}\beta_{rk}^{s_{r}}\right\}, \] where $\tbinom{L_{r}+s_{r}-1}{L_{r}-1}=\frac{\left(L_{r}+s_{r}-1\right)!}{\left(L_{r}-1\right)!s_{r}!}.$ The desired expansion is now given in the following lemma. \begin{lemma} $\Lambda^{\mathbf{L}} = \sum_{k \in \mathcal{D}(\mathbf{L})}\sum_{q=1}^{L_{k}}\frac{\gamma_{kq}(\mathbf{L})}{\lambda_{d + 1}^{N_i - q - L_{d + 1}}}\Lambda_{k}^{q}(t)\lambda_{d+1}^{N_{i}-q}$ if $\mathcal{D}(\mathbf{L}) \neq \emptyset.$ Further, this expansion is unique and holds $\forall t.$ \end{lemma} \begin{proof} See \cite{Gugan12}. \end{proof} Applying this expansion to each coefficient in \eqref{eq:afterColComTerms} and regrouping shows that it can be rewritten as \begin{equation} \label{eq:genCanPol} f(\mathbf{x};t)=\sum_{k=1}^{d}\sum_{q=1}^{N_{i}}h_{kq}(\mathbf{x})\Lambda_{k}^{q}(t)\lambda_{d+1}^{N_{i}-q}-c(t), \end{equation} where $c(t)=\mu_i(t) - \lambda_{d+1}^{N_{i}}$ and \begin{equation} h_{kq}(\mathbf{x})=\sum_{\mathbf{L}\in\Delta_{d+1,N_{i}},\, L_{k}\ge q}\frac{\gamma_{kq}(\mathbf{L})}{\lambda_{d+1}^{N_{i}-q-L_{d+1}}}g(\mathbf{x};\mathbf{L}). \label{eq:elemPolWeiCombo} \end{equation} \noindent \emph{Step3-Eliminate dependence on $\tau$}: The advantage of \eqref{eq:genCanPol} is that apart from $c(t),$ the number of t-dependent coefficients equals $d\cdot N_{i},$ which is exactly the number of unknowns in the polynomial $f$. Further, as shown below, they are linearly independent. \begin{lemma} \label{lem:genTtauMatrix} For $k \in [d\cdot N_i],$ let $b_k = \min \{ j \in [d]:j\cdot N_i \geq k \}.$ Then the matrix $T_{\tau}$, where for $j,k \in [d \cdot N_i]$ \begin{equation} (T_{\tau})_{jk} = \Lambda_{b_k}^{k - (b_k - 1)\cdot N_i}(t_j) \lambda_{d + 1}^{b_k \cdot N_i - k}, \label{eq:coeffMat} \end{equation} is non-singular. \end{lemma} \begin{proof} See \cite{Gugan12}. \end{proof} Now observe that if we let $\mathbf{c_{\tau}} \equiv (c(t_{1}),\ldots,c(t_{d\cdot N_{i}})),$ $\mathcal{E}_k(\mathbf{x}) \equiv (h_{k1}(\mathbf{x}), \ldots, h_{kN_{i}}(\mathbf{x}))$ and $\mathcal{E}(\mathbf{x})$ $\equiv (\mathcal{E}_1(\mathbf{x}),\ldots,$ $\mathcal{E}_d(\mathbf{x})),$ then \eqref{eq:tempPolSys} can be equivalently expressed as \begin{equation} \label{eq:genSymPolSys} T_{\tau}\mathcal{E}(\mathbf{x})-\mathbf{c}_{\tau}=0. \end{equation} Premultiplying \eqref{eq:genSymPolSys} by $\left(T_{\tau}\right)^{-1},$ which now exists by Lemma~\ref{lem:genTtauMatrix}, we obtain \begin{equation} \label{eq:genSecSymPolSys} \mathcal{E}(\mathbf{x})-\left(T_{\tau}\right)^{-1}\mathbf{c}_{\tau}=0. \end{equation} A crucial point to note now is that $\mathbf{w}\equiv(\mathbf{w}_{1},\ldots,\mathbf{w}_{N_{i}})$ is an obvious root of \eqref{eq:tempGenPoly} and hence of \eqref{eq:genSecSymPolSys}. This immediately implies that $\left(T_{\tau}\right)^{-1}\mathbf{c}_{\tau} = \mathcal{E}(\mathbf{w})$ and consequently \eqref{eq:genSecSymPolSys} can rewritten as \begin{equation} H_{i}(\mathbf{x})\equiv\mathcal{E}(\mathbf{x})-\mathcal{E}(\mathbf{w})=0.\label{eq:genElePolSys} \end{equation} Note that \eqref{eq:genElePolSys} is devoid of any reference to the set $\tau$ and can be arrived at using any valid $\tau.$ Furthermore, because of the equivalence between \eqref{eq:tempPolSys} and \eqref{eq:genElePolSys}, any conclusion that we can draw for \eqref{eq:genElePolSys} must hold automatically for the system of \eqref{eq:tempPolSys}. Because of these reasons, we will henceforth refer to \eqref{eq:genElePolSys} as the EPS. \begin{example} Let $N_i = d = 2.$ Also, let $\lambda_1 = 5, \lambda_2 = 3$ and $\lambda_3 = 1.$ Then the map $\mathcal{E}$ described above is given by \begin{equation} \label{eq:egEPS} \mathcal{E}(\mathbf{x})=\left( \begin{array}{c} x_{11} + x_{21} + 5(x_{11}x_{22} + x_{12}x_{21})\\ x_{11}x_{21}\\ x_{12} + x_{22} - 6(x_{11}x_{22} + x_{12}x_{21})\\ x_{12}x_{22}\\ \end{array} \right). \vspace{0.2cm} \end{equation} \end{example} We next describe some special features of the EPS. Let $\mathbf{x}_{\sigma} := (\mathbf{x}_{\sigma(1)}, \ldots, \mathbf{x}_{\sigma(N_i)}),$ $\sigma \in S_{N_i},$ denote a permutation of the vectors $\mathbf{x}_1, \ldots, \mathbf{x}_{N_i}$ and $\pi_{\mathbf{x}}:=\{\mathbf{x}_{\sigma}: \sigma \in S_{N_i}\}.$ \begin{lemma} \label{lem:EPSSymm} $H_i(\mathbf{x}) = H_i(\mathbf{x}_\sigma),$ $\forall \sigma \in S_{N_i}.$ That is, the map $H_i$ is symmetric. \end{lemma} \begin{proof} Observe that $H_i(\mathbf{x}) = (T_\tau)^{-1} F_{\tau}(\mathbf{x})$ and $F_\tau,$ as defined in \eqref{eq:tempPolSys}, is symmetric. The result thus follows. \end{proof} Next recall that if the complement of a solution set of a polynomial system is non-empty then it must be open dense in the Euclidean topology. This fact and the above result help us now to show that the EPS is almost always well behaved. \begin{lemma} \label{lem:EPSWellBeh} There exists an open dense set $\mathcal{R}$ of $\mathbb{R}^{d\cdot N_i}$ such that if $\mathbf{w} \in \mathcal{R}$ then the solution set of the EPS satisfies the following properties. \begin{enumerate} \item $\mathbf{w} \in V(H_i).$ \item If $\mathbf{x}^{*} \in V(H_i),$ then $\pi_{\mathbf{x}^{*}} \subset V(H_i).$ \label{prop:genSym} \item $|V(H_i)| = k \times N_i!,$ where $k \in \mathbb{N}$ is independent of $\mathbf{w} \in \mathcal{R}.$ Further, each solution is non-singular. \end{enumerate} \end{lemma} \begin{proof} See \cite{Gugan12}. \end{proof} We henceforth assume that $\mathbf{w} \in \mathcal{R}.$ Property \ref{prop:genSym} above then suggests that it suffices to work with \begin{equation} \label{eq:redSolSet} \mathcal{M}_i = \{\alpha \in \mathbb{C}^d : \exists \mathbf{x}^* \in V(H_i) \mbox{ with } \mathbf{x}_1^* = \alpha \}. \end{equation} Observe that $\mathcal{W}_i := \{\mathbf{w}_1, \ldots, \mathbf{w}_{N_i}\} \subset \mathcal{M}_i.$ A point to note here is that $\mathcal{I}_i := \mathcal{M}_i \backslash \mathcal{W}_i$ is not empty in general. Our next objective is to develop the above theory for the case where for each $i \in [m],$ instead of the exact value of $M_{Y_i}(t),$ we have access only to the IID realizations $\{Y_{il}\}_{l \geq 1}$ of the random variable $Y_i.$ That is, for each $k \in [N_i],$ we have to use the sample average $\hat{M}_{Y_i}(t_k; L) = \left(\sum_{l = 1}^{L} \exp(-t_k Y_{il})\right)/L $ for an appropriately chosen large $L,$ $\hat{c}(t_k; L) = \hat{M}_{Y_i}(t_k)(\lambda_{d+1} + t_k)^{N_i} - \lambda_{d+1}^{N_i}$ and $\hat{\mathbf{c}}_{\tau, L} \equiv (\hat{c}(t_1; L), \ldots, \hat{c}(t_{d \cdot N_i}; L))$ as substitutes for each $M_{Y_i}(t_k),$ each $c(t_k)$ and $\mathbf{c}_\tau$ respectively. But even then note that the noisy or the perturbed version of the EPS \begin{equation} \label{eq:NoiGenElePolSys} \hat{H}_i(\mathbf{x}) \equiv \mathcal{E}(\mathbf{x}) - (T_{\tau})^{-1}(\hat{\mathbf{c}}_{\tau, L}) = 0. \end{equation} is always well defined. More importantly, the perturbation is only in its constant term. As in Lemma~\ref{lem:EPSSymm}, it then follows that the map $\hat{H}_i$ is symmetric. Next observe that since $\mathcal{R}$ is open (see Lemma~\ref{lem:EPSWellBeh}), there exists a small enough $\bar{\delta}_i > 0$ such that $B(\mathbf{w}; \bar{\delta}_i) \subset \mathcal{R}.$ Using the regularity of solutions of the EPS (see Property 3 of Lemma~\ref{lem:EPSWellBeh}), the inverse function theorem then gives us the following result. \begin{lemma} \label{lem:NoiEPSWellBeh} Let $\delta \in (0, \bar{\delta}_i)$ be such that for any two distinct solutions in $V(H_i),$ say $\mathbf{x}^{*}$ and $\mathbf{y}^{*},$ $B(\mathbf{x}^{*}; \delta) \cap B(\mathbf{y}^{*}; \delta) = \emptyset.$ Then there exists an $\epsilon(\delta) > 0$ such that if $\mathbf{u} \in \mathbb{R}^{d \cdot N_i}$ and $||\mathbf{u} - \mathcal{E}(\mathbf{w})|| < \epsilon(\delta)$ then the solution set $V(\hat{H}_i)$ of the perturbed EPS $\mathcal{E}(\mathbf{x}) - \mathbf{u} = 0$ satisfies the following: \begin{enumerate} \item All roots in $V(\hat{H}_i)$ are regular points of the map $\mathcal{E}.$ \item For each $\mathbf{x}^{*} \in V(H_i),$ there is a unique $\mathbf{z}^{*} \in V(\hat{H}_i)$ such that $||\mathbf{x}^{*} - \mathbf{z}^{*}|| < \delta.$ \item For each $\mathbf{z}^{*} \in V(\hat{H}_i),$ there is a unique $\mathbf{x}^{*} \in V(H_i)$ such that $||\mathbf{x}^{*} - \mathbf{z}^{*}|| < \delta.$ \end{enumerate} \end{lemma} \begin{proof} See \cite{Gugan12}. \end{proof} The above result, in simple words, states that if we can get hold of a close enough approximation of $\mathcal{E}(\mathbf{w}),$ say $\mathbf{u},$ then solving the perturbed EPS $\mathcal{E}(\mathbf{x}) - \mathbf{u} = 0$ is almost as good as solving the EPS of \eqref{eq:genElePolSys}. We now show how to acquire such an approximation of $\mathcal{E}(\mathbf{w}).$ \begin{lemma} \label{lem:ExistL} Let $\delta$ and $\epsilon(\delta)$ be as described in Lemma~\ref{lem:NoiEPSWellBeh}. Then for tolerable failure rate $\kappa > 0$ and the chosen set $\tau,$ $\exists L_{\tau, \delta, \kappa} \in \mathbb{N}$ such that if $L \geq L_{\tau, \delta, \kappa}$ then with probability greater than $1 - \kappa,$ we have $|| (T_\tau)^{-1} \hat{\mathbf{c}}_{\tau, L} - \mathcal{E}(\mathbf{w}) || < \epsilon(\delta).$ \end{lemma} \begin{proof} Note that $\exp(-t_k Y_{il}) \in [0,1]$ $\forall i, l$ and $k.$ The Hoeffding inequality (see \cite{Hoeffding62}) then shows that for any $\epsilon > 0$, $\Pr\{|\hat{M}_{Y_i}(t_k; L) - M_{Y_i}(t_k)| > \epsilon\} \leq \exp(-2 \epsilon^2 L).$ Since $\mathcal{E}(\mathbf{w}) = (T_\tau)^{-1}\mathbf{c}_\tau,$ the result is now immediate. \end{proof} Let us now fix a $L \geq L_{\tau, \delta, \kappa}$ and let $\mathcal{A}_i(\kappa)$ denote the event $|| (T_\tau)^{-1} \hat{\mathbf{c}}_{\tau, L} - \mathcal{E}(\mathbf{w}) || < \epsilon(\delta).$ Clearly, $\Pr\{\mathcal{A}^{c}_i(\kappa)\} \leq \kappa.$ Observe that when $\mathcal{A}_i(\kappa)$ is a success the solution set of \eqref{eq:NoiGenElePolSys}, with $L$ as chosen above, satisfies all properties given in Lemma~\ref{lem:NoiEPSWellBeh}. Because of the symmetry of the map $\hat{H}_i,$ as in \eqref{eq:redSolSet}, it again suffices to work with \begin{equation} \hat{\mathcal{M}}_i = \{\hat{\alpha} \in \mathbb{C}^d : \exists \mathbf{z}^{*} \in V(\hat{H}_i) \mbox{ with } \mathbf{z}_1^{*} = \alpha\}. \end{equation} We are now done discussing the EPS for an arbitrary $i \in [m].$ In summary, we have managed to obtain a set $\hat{\mathcal{M}}_i$ in which a close approximation of the weight vectors of random variables ${X_j}$ that add up to give $Y_i$ are present with high probability. The next subsection takes a unified view of the solution sets $\{\hat{\mathcal{M}}_i: i\in [m]\}$ to match the weight vectors to the corresponding random variables. But before that, we redefine $\mathcal{W}_i$ as $\{\mathbf{w}_j: j \in p_i\}.$ Accordingly, $\mathcal{M}_i, \hat{\mathcal{M}}_i, V(H_i)$ and $V(\hat{H}_i)$ are also redefined using notations of Section~\ref{sec:model}. \subsection{Parameter Matching using 1-identifiability} \label{subsec:parMatch} We begin by giving a physical interpretation for the $1-$identifiability condition of the matrix $A.$ For this, let $\mathcal{G}_{j} := \left\{ i\in[m]:j\in p_{i}\right\} $ and $\mathcal{B}_{j} := [m] \backslash \mathcal{G}_{j}.$ \begin{lemma} \label{lem:linkAsPathIntersection} For a $1-$identifiable matrix $A$, each index $j\in[N]$ satisfies \[ \{j\} = \bigcap_{g\in\mathcal{G}_{j}} p_{g} \cap \bigcap_{b\in\mathcal{B}_{j}} p_{b}^{c} =: \mathcal{D}_{j}. \] \end{lemma} \begin{proof} By definition, $j \in \mathcal{D}_{j}.$ For converse, if $k\in\mathcal{D}_{j},$ $k \neq j,$ then columns $j$ and $k$ of $A$ are identical; contradicting its $1-$identifiability condition. Thus $\{j\}=\mathcal{D}_{j}.$ \end{proof} An immediate result is the following. \begin{corollary} \label{cor:uniqueValuePathValueIntersection} Suppose $A$ is a $1-$identifiable matrix. If the map $u: [N] \rightarrow X,$ where $X$ is an arbitrary set, is bijective and $\forall i \in [m],$ $v_{i} := \{u(j):j\in p_{i}\},$ then for each $j \in [N]$ \[ \{u(j)\} = \bigcap_{g\in\mathcal{G}_{j}} v_{g} \cap \bigcap_{b\in\mathcal{B}_{j}} v_{b}^{c} \] \end{corollary} By reframing this, we get the following result. \begin{theorem} \label{thm:idlCasParMch} Suppose $A$ is a $1-$identifiable matrix. If the weight vectors $\mathbf{w}_1, \ldots, \mathbf{w}_N$ are pairwise distinct then the rule % \begin{equation} \psi:j \rightarrow \bigcap_{g\in\mathcal{G}_{j}} \mathcal{W}_{g} \cap \bigcap_{b\in\mathcal{B}_{j}} \mathcal{W}_{b}^{c}, \label{eq:idlCasAssRule} \end{equation} % satisfies $\psi(j) = \mathbf{w}_{j}$. % \end{theorem} This result is where the complete potential of the $1-$ identifiability condition of $A$ is being truly taken advantage of. What this states is that if we had access to the collection of solution sets $\{\mathcal{W}_i: i \in [m]\}$ then using $\psi$ we would have been able to uniquely match the weight vectors to the random variables. But note that, at present, we have access only to the collection $\{\mathcal{M}_i: i \in [m]\}$ in the ideal case and $\{\hat{\mathcal{M}}_i: i \in [m]\}$ in the perturbed case. Keeping this in mind, our goal now is to show that if $\forall i_1, i_2 \in [m],$ $i_1 \neq i_2,$ \begin{equation} \mathcal{I}_{i_1} \cap \mathcal{M}_{i_2} = \emptyset, \label{eq:inValSolNotRel} \end{equation} a condition that always held in simulation experiments, then the rules (with minor modifications): \begin{equation} \label{eq:idlCasGenAssRule} \psi:j \rightarrow \bigcap_{g\in\mathcal{G}_{j}} \mathcal{M}_{g} \cap \bigcap_{b\in\mathcal{B}_{j}} \mathcal{M}_{b}^{c} \end{equation} for the ideal case, and \begin{equation} \label{eq:PerCasGenAssRule} \hat{\psi}:j \rightarrow \bigcap_{g\in\mathcal{G}_{j}} \hat{\mathcal{M}}_{g} \cap \bigcap_{b\in\mathcal{B}_{j}} \hat{\mathcal{M}}_{b}^{c} \end{equation} in the perturbed case, recover the correct weight vector associated to each random variable $X_j.$ We first discuss the ideal case. Let $\mathcal{S}:=\{j \in [N]: |\mathcal{G}_j| \geq 2\}.$ Because of \eqref{eq:inValSolNotRel} and Theorem~\ref{thm:idlCasParMch}, note that \begin{enumerate} \item If $j \in \mathcal{S},$ then $\psi(j) = \{\mathbf{w}_j\}.$ \item If $j \in \mathcal{S}^{c}, j \in p_{i^{*}}$ then $\psi(j) = \{\mathbf{w}_j\} \cup \mathcal{I}_{i^*}.$ \end{enumerate} That is, \eqref{eq:idlCasGenAssRule} works perfectly fine when $j \in \mathcal{S}.$ The problem arises when $j \in \mathcal{S}^{c}$ as $\psi(j)$ does not give as output a unique vector. To correct this, fix $j \in \mathcal{S}^{c}.$ If $j \in p_{i^*}$ then let $\mathbf{v}^{sub} \equiv (\mathbf{w}_k : k \in p_{i^*} \backslash \{j\}).$ Because of $1-$identifiability, note that if $k \in p_{i^*}\backslash \{j\}$ then $k \in \mathcal{S}.$ From \eqref{eq:tempGenPoly} and \eqref{eq:tempPolSys}, it is also clear that $(\mathbf{v}^{sub}, \alpha) \in V(H_i)$ if and only if $\alpha = \mathbf{w}_j.$ This suggests that we need to match parameters in two stages. In stage 1, we use \eqref{eq:idlCasGenAssRule} to assign weight vectors to all those random variables $X_j$ such that $j \in \mathcal{S}.$ In stage 2, for each $j \in \mathcal{S}^{c},$ we identify $i^* \in [m]$ such that $j \in p_{i^*}.$ We then construct $\mathbf{v}^{sub}.$ We then assign to $j$ that unique $\alpha$ for which $(\mathbf{v}^{sub}, \alpha) \in V(H_{i^*}).$ Note that we are ignoring the trivial case where $|p_{i^*}| = 1.$ It is now clear that by using \eqref{eq:idlCasGenAssRule} with modifications as described above, at least for the ideal case, we can uniquely recover back for each random variable $X_j$ its corresponding weight vector $\mathbf{w}_j.$ We next handle the case of noisy measurements. Let $\mathcal{U}:= \cup_{i \in [m]} \mathcal{M}_i$ and $\hat{\mathcal{U}} := \cup_{i \in [m]} \hat{\mathcal{M}}_i.$ Observe that using \eqref{eq:PerCasGenAssRule} directly, with probability one, will satisfy $\hat{\psi}(j) = \emptyset$ for each $j \in [N].$ This happens because we are distinguishing across the solution sets the estimates obtained for a particular weight vector. Hence as a first step we need to define a relation $\sim$ on $\hat{\mathcal{U}}$ that associates these related elements. Recall from Lemmas~\ref{lem:NoiEPSWellBeh} and \ref{lem:ExistL} that the set $\hat{\mathcal{M}}_i$ can be constructed for any small enough choice of $\delta, \kappa > 0.$ With choice of $\delta$ that satisfies \begin{equation} \label{eq:delCond} 0 < 4\delta < \underset{\alpha, \beta \in \mathcal{U}}{\min} ||\alpha - \beta||, \end{equation} let us consider the event $\mathcal{A} : = \cap_{i \in [m]} \mathcal{A}_i(\kappa/m).$ Using a simple union bound, it follows that $\Pr\{\mathcal{A}^{c}\} \leq \kappa.$ Now suppose that the event $\mathcal{A}$ is a success. Then by \eqref{eq:delCond} and Lemma~\ref{lem:NoiEPSWellBeh}, the following observations follow trivially. \begin{enumerate} \item For each $i \in [m]$ and each $\alpha \in \mathcal{M}_i,$ there exists at least one $\hat{\alpha} \in \hat{\mathcal{M}}_i$ such that $||\hat{\alpha} - \alpha|| < \delta.$ \item For each $i \in [m]$ and each $\hat{\alpha} \in \hat{\mathcal{M}}_i,$ there exists precisely one $\alpha \in \mathcal{M}_i$ such that $||\hat{\alpha} - \alpha|| < \delta.$ \item Suppose for distinct elements $\alpha, \beta \in \mathcal{U},$ we have $\hat{\alpha}, \hat{\beta} \in \hat{\mathcal{U}}$ such that $||\hat{\alpha} - \alpha|| < \delta$ and $||\hat{\beta} - \beta|| < \delta.$ Then $||\hat{\alpha} - \hat{\beta}|| > 2\delta.$ \end{enumerate} From these, it is clear that the relation $\sim$ on $\hat{\mathcal{U}}$ should be \begin{equation} \label{eq:Rel} \hat{\alpha} \sim \hat{\beta} \mbox{ iff } ||\hat{\alpha} - \hat{\beta}|| < 2\delta. \end{equation} It is also easy to see that, whenever the event $\mathcal{A}$ is a success, $\sim$ defines an equivalence relation on $\hat{\mathcal{U}}.$ For each $i \in [m],$ the obvious idea then is to replace each element of $\hat{\mathcal{M}}_i$ and its corresponding $d-$ dimensional component in $V(\hat{H}_i)$ with its equivalence class. It now follows that \eqref{eq:PerCasGenAssRule}, with modifications as was done for the ideal case, will satisfy \begin{equation} \hat{\psi}(j) = \{\hat{\alpha} \in \hat{\mathcal{U}}:||\hat{\alpha}- \mathbf{w}_j|| < \delta\}. \end{equation} This is obviously the best we could have done starting from the set $\{\hat{\mathcal{M}}_i: i \in [m]\}.$ We end this section by summarizing our complete method in an algorithmic fashion. For each $i \in [m],$ let $\{Y_{il}\}_{l\geq 1}$ be the IID samples of $Y_i.$ \begin{algorithm} \textbf{\large Distribution tomography} \label{alg:EstLinkDis} \noindent \textit{\large Phase 1: Construct \& Solve the EPS.} For each $i \in [m],$ \begin{enumerate} \item Choose an arbitrary $\tau = \{t_1, \ldots, t_{d \cdot N_i}\}$ of distinct positive real numbers. \item For a large enough $L\in\mathbb{N}$ and each $t_{j}\in\tau,$ set $\hat{M}_{Y_{i}}(t_{j})=\left(\sum_{l=1}^{L}\exp(-t_jY_{il})\right)/L,$ $\hat{\mu}_{i}(t_{j})=(\lambda_{d+1}+t_{j})^{N_{i}}\hat{M}_{Y_{i}}(t_{j})$ and $\hat{c}(t_{j})=\hat{\mu}_{i}(t_{j}) - \lambda_{d+1}^{N_{i}}.$ Using this, construct $\hat{c}_{\tau}\equiv(\hat{c}(t_{1}),\ldots,\hat{c}(t_{d\cdot N_{i}})).$ \item Solve $\mathcal{E}(\mathbf{x})-T_{\tau}^{-1}\hat{c}_{\tau}=0$ using any standard solver for polynomial systems. \item Build $\hat{\mathcal{M}}_i = \{\alpha \in \mathbb{C}^d : \exists \mathbf{x}^* \in V(\hat{H}_i) \mbox{ with } \mathbf{x}_1^* = \alpha \}.$ \end{enumerate} \noindent \textit{\large Phase 2: Parameter Matching} \begin{enumerate} \item Set $\hat{\mathcal{U}}:=\bigcup_{i\in[m]}\hat{\mathcal{M}}_{i}.$ Choose $\delta>0$ small enough and define the relation $\sim$ on $\mathcal{\hat{U}},$ where $\hat{\alpha} \sim \hat{\beta}$ if and only if $||\hat{\alpha} - \hat{\beta}||_{2} < 2\delta.$ If $\sim$ is not an equivalence relation then choose a smaller $\delta$ and repeat. \item Construct the quotient set $\hat{\mathcal{U}}\backslash\sim.$ Replace all elements of each $\hat{\mathcal{M}_{i}},$ $V(\hat{H}_{i})$ with their equivalence class. \item For each $j\in\mathcal{S},$ set \\$\hat{\psi}(j)=\left(\bigcap_{g\in\mathcal{G}_{j}}\mathcal{\hat{M}}_{g}\right)\cap\left(\bigcap_{b\in\mathcal{B}_{j}}\hat{\mathcal{M}}_{b}^{c}\right).$ \item For each $j\in\mathcal{S}^{c},$ \begin{enumerate} \item Set $i^{*} = i\in[m]$ such that $j\in p_{i^{*}}.$ \item Construct $\mathbf{v}^{sub}\equiv(\psi(k):k\in p_{i^{*}}, k\ne j\}.$ \item Set $\psi(j)=\hat{\alpha}$ such that $(\mathbf{v}^{sub},\hat{\alpha})\in V(\hat{H}_{i}).$ \end{enumerate} \end{enumerate} \end{algorithm} \section{Universality} \label{sec:universality} The crucial step in the method described above was to come up with, for each $i \in [m],$ a well behaved polynomial system, i.e., one that satisfies the properties of Lemma~\ref{lem:EPSWellBeh}, based solely on the samples of $Y_i.$ Once that was done, the ability to match parameters to the component random variables was only a consequence of the $1-$identifiability condition of the matrix $A.$ This suggests that it may be possible to develop similar schemes even in settings different to the ones assumed in Section~\ref{sec:model}. In fact, functions other than the MGF could also serve as blueprints for constructing the polynomial system. We discuss in brief few of these ideas in this section. Note that we are making a preference for polynomial systems for the sole reason that there exist computationally efficient algorithms, see for example \cite{Sommese05, Li97, Morgan89, Verschelde99}, to determine all its roots. Consider the case where $\forall j \in [N],$ the distribution of $X_j$ is a finite mixture model given by \begin{equation} \label{eq:genMixDist} F_j(u) = \sum_{k = 1}^{d_j + 1}w_{jk}\phi_{jk}(u), \end{equation} where $d_j \in \mathbb{N},$ $w_{j1}, \ldots, w_{j(d_l + 1)}$ denote mixing weights, i.e., $w_{jk} \geq 0$ and $\sum_{k = 1}^{d_k + 1}w_{jk} = 1,$ and $\{\phi_{jk}(u)\}$ are some basis functions, say Gaussian, uniform, etc. The MGF is thus given by \begin{equation} M_{X_j}(t) = \sum_{k = 1}^{d_j + 1}w_{jk}\int_{u = 0}^{\infty}\exp(-ut)d\phi_{jk}(u). \end{equation} Note now that if the basis functions $\{\phi_{jk}\}$ are completely known then the MGF of each $Y_i$ will again be a polynomial in the mixing weights, $\{w_{jk}\},$ similar in spirit to the relation of \eqref{eq:MGFRel}. As a result, the complete recipe of Section~\ref{sec:GenNT} can again be attempted to estimate the weight vectors of the random variables $X_j$ using only the IID samples of each $Y_i.$ In relation to \eqref{eq:Dist} or \eqref{eq:genMixDist}, observe next that $\forall n \in \mathbb{N},$ the $n^{th}$ moment of each $X_j$ is given by \begin{equation} \mathbb{E}(X_j^n) = \sum_{k = 1}^{d_j + 1}w_{jk}\int_{u = 0}^{\infty}u^n d\phi_{jk}(u). \end{equation} Hence, the $n^{th}$ moment of $Y_i$ is again a polynomial in the unknown weights. This suggests that, instead of the MGF, one could use the estimates of the moments of $Y_i$ to come up with an alternative polynomial system and consequently solve for the distribution of each $X_j.$ Moving away from the models of \eqref{eq:Dist} and \eqref{eq:genMixDist}, suppose that for each $j \in [N],$ $X_j \sim \exp(m_j).$ Assume that each mean $m_j < \infty$ and that $m_{j_{1}} \neq m_{j_{2}}$ when $j_{1}\neq j_{2}.$ We claim that the basic idea of our method can be used here to estimate $m_1, \ldots, m_N$ and hence the complete distribution of each $X_j$ using only the samples of $Y_i.$ As the steps are quite similar when either i) we know $M_{Y_i}(t)$ for each $i \in [m]$ and every valid $t$ and ii) we have access only to the IID samples $\{Y_{il}\}_{l \geq 1}$ for each $i \in [m]$, we take up only the first case. Fix $i \in [m]$ and let $p_i := \{j \in [N]: a_{ij} = 1\}.$ To simplify notations, let us relabel the random variables $\{X_j: j \in p_i\}$ that add up to give $Y_i$ as $X_1, \ldots, X_{N_i},$ where $N_i = |p_i|.$ Observe that the MGF of $Y_i,$ after inversion, satisfies \begin{equation} \prod_{j=1}^{N_{i}} (1 + t m_{j}) = 1/M_{Y_{i}}(t). \label{eq:momGenRel} \end{equation} Using \eqref{eq:momGenRel}, we can then define the canonical polynomial \begin{equation} f(\mathbf{x};t):=\prod_{j=1}^{N_{i}}(1+tx_{j})-c(t), \label{eq:canPol} \end{equation} where $\mathbf{x} \equiv (x_{1},\ldots,x_{N_{i}})$ and $c(t) = 1/M_{Y_{i}}(t).$ Now choose an arbitrary set $\tau = \{t_{1},\ldots,t_{N_{i}}\} \subset \mathbb{R}_{++}$ consisting of distinct numbers and define \begin{equation} F_{\tau}( \mathbf{x}) \equiv (f_{1}(\mathbf{x}), \ldots, f_{N_{i}}(\mathbf{x})) = 0, \label{eq:IntmdPolySys} \end{equation} where $f_{k}(\mathbf{x})=f(\mathbf{x};t_{k}).$ We emphasize that this system is square of size $N_{i},$ depends on the choice of subset $\tau$ and each polynomial $f_{k}$ is symmetric with respect to the variables $x_{1,}\ldots,x_{N_{i}}.$ In fact, if we let $\mathbf{c}_{\tau}\equiv(c(t_{1}),\ldots,c(t_{N_{i}}))$ and $\mathcal{E} (\mathbf{x}) \equiv (e_{1}(\mathbf{x}), \ldots, e_{N_{i}} (\mathbf{x})),$ where $e_{k}(\mathbf{x})=\sum_{1\leq j_{1}<j_{2}<\ldots<j_{k}\leq N_{i}}x_{j_{1}}\cdots x_{j_{k}}$ denotes the $k^{th}$ elementary symmetric polynomial in the $N_{i}$ variables $x_{1,}\ldots,x_{N_{i}},$ we can rewrite \eqref{eq:IntmdPolySys} as \begin{equation} T_{\tau} \mathcal{E}(\mathbf{x}) - (\mathbf{c}_{\tau} - \mathbf{1}) = 0. \label{eq:firstSymPolSys} \end{equation} Here $T_{\tau}$ denotes a Vandermonde matrix of order $N_{i}$ in $t_{1},\ldots,t_{N_{i}}$ with $(T_{\tau})_{jk}=t_{j}^{k}.$ Its determinant, given by $\det(T_{\tau}) = \left(\prod_{j=1}^{n} t_{j} \right) \prod_{j>i}(t_{j}-t_{i}),$ is clearly non-zero. Premultiplying \eqref{eq:firstSymPolSys} by $T_{\tau}^{-1},$ we obtain \begin{equation} \mathcal{E}(\mathbf{x}) - T_{\tau}^{-1}(\mathbf{c}_{\tau} - \mathbf{1}) = 0. \label{eq:secSymPolSys} \end{equation} Observe now that the vector $\mathbf{m} \equiv (m_{1},\ldots,m_{N_{i}})$ is a natural root of \eqref{eq:IntmdPolySys} and hence of \eqref{eq:secSymPolSys}. Hence $T_{\tau}^{-1}(\mathbf{c}_{\tau} - \mathbf{1}) = \mathcal{E}(\mathbf{m}).$ The EPS for this case can thus be written as \begin{equation} H_{i}(\mathbf{x}) \equiv \mathcal{E}(\mathbf{x}) - \mathcal{E}(\mathbf{m}) = 0. \label{eq:charPolSys} \end{equation} We next discuss the properties of this EPS, or more specifically, its solution set. For this, let $V(H_{i}):=\{\mathbf{x} \in \mathbb{C}^{N_i} : H_{i}(\mathbf{x}) = 0\}.$ \begin{lemma} \label{lem:EPSSolSet} $V(H_{i})=\pi_{\mathbf{m}}:= \{\sigma(\mathbf{m}): \sigma \in S_{N_{i}}\}.$ \end{lemma} \begin{proof} This follows directly from \eqref{eq:charPolSys}. \end{proof} \begin{lemma} \label{lem:EPSRegSol} For every $\mathbf{x}^{*}\in V(H_{i}),$ $\det(\dot{\mathcal{E}}(\mathbf{x^{*}})) \neq 0.$ \label{prop:regSol} \end{lemma} \begin{proof} This follows from Lemma (\ref{lem:EPSSolSet}) and the fact that $\det(\dot{\mathcal{E}}(\mathbf{x}))=\prod_{1\leq j<k\leq N_{i}}(x_{j}-x_{k}).$ \end{proof} Because of Lemma~\ref{lem:EPSSolSet}, it suffices to work with only the first components of the roots. Hence we define \begin{equation} \mathcal{M}_{i}:=\{\alpha^{*}\in\mathbb{C}: \exists \mathbf{x}^{*} \in V(H_i) \mbox{ with } \mathbf{x}^*_{1} = \alpha \}, \end{equation} which in this case is equivalent to the set $\{m_1, \ldots, m_{N_i} \}.$ Reverting back to global notations, note that \begin{equation} \mathcal{M}_{i}=\{m_j : j \in p_i\}. \label{eq:finPathSolSet} \end{equation} Since $i$ was arbitrary, we can repeat the above procedure to obtain the collection of solution sets $\{\mathcal{M}_i: i \in [m]\}.$ Arguing as in Theorem~\ref{thm:idlCasParMch}, it is now follows that if $A$ is $1-$identifiable then the rule \begin{equation} \psi(j) = \bigcap_{g \in \mathcal{G}_j}\mathcal{M}_g \cap \bigcap_{b \in \mathcal{B}_j} \mathcal{M}_b^{c}, \end{equation} where $\mathcal{G}_j = \{i \in [m]: j \in p_i\}$ and $\mathcal{B}_j = [m]\backslash\mathcal{G}_j,$ satisfies the relation $\psi(j) = m_j.$ That is, having obtained the sets $\{\mathcal{M}_i : i \in [m]\},$ one can use $\psi$ to match the parameters to the corresponding random variables. This clearly demonstrates that even if a transformation of the MGF is a polynomial in the parameters to be estimated, our method may be applicable. \addtolength{\textheight}{-3.5cm} \section{Experimental Results} \label{sec:ExptResults} \begin{figure}[ht!] \begin{center} \subfigure[Tree topology] { \label{fig:tree} \includegraphics[width=0.4\columnwidth]{simNet2,3.png} } \subfigure[General topology] { \label{fig:general} \includegraphics[width=0.4\columnwidth]{simNet5.png} } \end{center} \label{fig:simNet1} \caption{Network topologies in simulation experiments.} \end{figure} In order to verify performance, we conducted matlab based simulations using a network with i) tree topology, Fig.~\ref{fig:tree} and ii) general topology, Fig.~\ref{fig:general}. \subsection{Network with Tree Topology} \label{subsec:treeTopology} We work here with the network of Figure ~\ref{fig:tree}. \begin{experiment} \label{exp:expt1} Node $1$ is the source node, while nodes $3$ and $4$ act as sink. The packet delay across link $j,$ denoted $X_j,$ is a hyperexponential random variable---a special case of the \eqref{eq:Dist}. The count of exponential stages in each link distribution equals three. That is, $d = 2.$ The corresponding exponential stage parameters $\lambda_1, \lambda_2$ and $\lambda_3$ are taken to be $5, 3$ and $1$ respectively. For each link, the weight associated with each exponential stage is given in first three columns of Table~\ref{tab:expt1}. \begin{table}[h!] \caption{\label{tab:expt1}Actual and Estimated weights for Expt.~\ref{exp:expt1}} \centering \begin{tabular}{| c | c c c | c c c |} \hline Link & $w_{j1}$ & $w_{j2}$ & $w_{j3}$ & $\hat{w}_{j1}$ & $\hat{w}_{j2}$ & $\hat{w}_{j3}$\\ \hline 1 & 0.17 & 0.80 & 0.03 & 0.15 & 0.82 & 0.02\\ 2 & 0.13 & 0.47 & 0.40 & 0.15 & 0.46 & 0.39\\ 3 & 0.80 & 0.15 & 0.05 & 0.79 & 0.15 & 0.06\\ \hline \end{tabular} \vspace{0.1cm} \end{table} Let $p_1$ be the path connecting the nodes $1, 2$ and $3.$ Similarly, let $p_2$ be the path connecting the nodes $1, 2$ and $4.$ Let $Y_1$ and $Y_2$ denote respectively the end-to-end delay across each of these paths. If we let $Y \equiv (Y_1, Y_2)$ and $X \equiv (X_1, X_2, X_3)$ then it follows that $Y = AX,$ where \[ A = \left ( \begin{array}{ccc} 1 & 1 & 0\\ 1 & 0 & 1\\ \end{array} \right). \] We are now in the framework of Section~\ref{sec:model}. We first focus on path $p_1.$ Observe that its EPS is given by the map of \eqref{eq:egEPS}. We collect now a million samples of its end-to-end delay. Choosing an arbitrary set $\tau = \{1.9857, 2.3782, 0.3581, 8.8619\},$ we run the first phase of Algorithm~\ref{alg:EstLinkDis} to obtain $\hat{\mathcal{M}}_1.$ This set along with its ideal counterpart is given in Table~\ref{tab:expt11}. \begin{table}[h!] \caption{\label{tab:expt11} Solution set of the EPS for path $p_1$ in Expt.~\ref{exp:expt1}} \centering \begin{tabular}{| c | c c |} \hline sol-ID & $\mathcal{M}_1$ & $\hat{\mathcal{M}}_1$\\ \hline 1 & (0.1300, 0.4700) & (0.1542, 0.4558)\\ 2 & (0.1700, 0.8000) & (0.1292, 0.8356)\\ 3 & (3.8304, -2.8410) & (3.8525, -2.8646)\\ 4 & (0.1933, 0.7768) & (0.2260, 0.7394)\\ 5 & (0.1143, 0.4840) & (0.0882, 0.5152)\\ 6 & (0.0058, -0.1323) & (0.0052, -0.1330)\\ \hline \end{tabular} \vspace{0.1cm} \end{table} Similarly, by probing the path $p_2$ with another million samples and with $\tau = \{0.0842, 0.0870, 0.0305, 0.0344\},$ we determine $\hat{\mathcal{M}}_2.$ The sets $\mathcal{M}_2$ and $\hat{\mathcal{M}}_2$ are given in Table~\ref{tab:expt12}. \begin{table}[h!] \caption{\label{tab:expt12} Solution set of the EPS for path $p_2$ in Expt.~\ref{exp:expt1}} \centering \begin{tabular}{| c | c c |} \hline sol-ID & $\mathcal{M}_2$ & $\hat{\mathcal{M}}_2$\\ \hline 1 & (0.8000, 0.1500) & (0.7933, 0.1459)\\ 2 & (0.1660, 0.7775) & (0.1720, 0.8095)\\ 3 & (5.5623, -4.5638) & (5.5573, -4.5584)\\ 4 & (0.1700, 0.8000) & (0.1645, 0.7669)\\ 5 & (0.8191, 0.1543) & (0.8296, 0.1540)\\ 6 & (0.0245, -0.0263) & (0.0246, -0.0259)\\ \hline \end{tabular} \vspace{0.1cm} \end{table} To match the weight vectors to corresponding links, firstly observe that the minimum distance between $\hat{\mathcal{M}}_1$ and $\hat{\mathcal{M}}_2$ is 0.0502. Based on this, we choose $\delta = 0.03$ and run the second phase of Algorithm~\ref{alg:EstLinkDis}. The obtained results, after rounding to two significant digits, are given in the second half of Table~\ref{tab:expt1}. Note that the weights obtained for the first link are determined by taking a simple average of the solutions obtained from the two different paths. The norm of the error vector equals $0.0443.$ \end{experiment} \begin{experiment} \label{exp:expt2} Keeping other things unchanged as in the setup of experiment~\ref{exp:expt1}, we consider here four exponential stages in the distribution of each $X_j.$ The exponential stage parameters $\lambda_1, \lambda_2, \lambda_3$ and $\lambda_4$ equal $5, 4, 0.005$ and $1$ respectively. The corresponding weights are given in Table~\ref{tab:expt2}. But observe that the weights of the third stage is negligible for all three links. We can thus ignore its presence. The results obtained after running Algorithm~\ref{alg:EstLinkDis} are given in the second half of Table~\ref{tab:expt2}. The norm of the error vector equals $0.1843.$ \begin{table}[h!] \caption{\label{tab:expt2}Actual and Estimated weights for Expt.~\ref{exp:expt2}} \begin{center} \begin{tabular}{|c|c c c c|c c c c|} \hline Link & $w_{j1}$ & $w_{j2}$ & $w_{j3}$ & $w_{j4}$ & $\hat{w}_{j1}$ & $\hat{w}_{j2}$ & $\hat{w}_{j3}$ & $\hat{w}_{j4}$\tabularnewline \hline 1 & 0.71 & 0.20 & 0.0010 & 0.08 & 0.77 & 0.20 & 0 & 0.03\\ 2 & 0.41 & 0.17 & 0.0015 & 0.41 & 0.38 & 0.18 & 0 & 0.44\\ 3 & 0.15 & 0.80 & 0.0002 & 0.04 & 0.12 & 0.70 & 0 & 0.18\\ \hline \end{tabular} \end{center} \vspace{0.1cm} \end{table} \end{experiment} \subsection{Network with General Topology} \label{subsec:genTopology} We deal here with the network of Figure~\ref{fig:general}. \begin{experiment} \label{exp:expt3} Nodes $1$ and $2$ act as source while nodes $3$ and $4$ act as sink. We consider here three paths. Path $p_1$ connects the nodes $1, 2$ and $3,$ path $p_2$ connects the nodes $1, 2$ and $4,$ while path $p_3$ connects the nodes $2, 3$ and $4.$ Let $X \equiv (X_1, X_2, X_3, X_4),$ where $X_j$ is the packet delay across link $j.$ Also, let $Y \equiv (Y_1, Y_2, Y_3),$ where $Y_i$ denotes the end-to-end delay across path $p_i.$ They are related by $Y = AX,$ where \[ A = \left( \begin{array}{cccc} 1 & 1 & 0 & 0\\ 1 & 0 & 1 & 0\\ 0 & 1 & 0 & 1\\ \end{array} \right). \] As in Experiment~\ref{exp:expt1}, the random variable $X_j$ is hyperexponentially distributed with values of $d, \lambda_1, \lambda_2, \lambda_3$ kept exactly the same. By choosing again a million probe packets for each path, we run Algorithm~\ref{alg:EstLinkDis}. The actual and estimated weights are shown in Table~\ref{tab:expt3}. \begin{table}[h!] \caption{\label{tab:expt3}Actual and Estimated weights for Expt.~\ref{exp:expt3}} \begin{center} \begin{tabular}{|c|c c c| c c c|} \hline Link & $w_{j1}$ & $w_{j2}$ & $w_{j3}$ & $\hat{w}_{j1}$ & $\hat{w}_{j2}$ & $\hat{w}_{j3}$\tabularnewline \hline 1 & 0.34 & 0.26 & 0.40 & 0.34 & 0.24 & 0.42\\ 2 & 0.46 & 0.49 & 0.05 & 0.45 & 0.50 & 0.05\\ 3 & 0.12 & 0.65 & 0.23 & 0.11 & 0.68 & 0.21\\ 4 & 0.71 & 0.19 & 0.10 & 0.69 & 0.18 & 0.13\\ \hline \end{tabular} \end{center} \vspace{0.1cm} \end{table} \end{experiment} \section{Discussion} \label{sec:ConclFutW} This paper took advantage of the properties of polynomial systems to develop a novel algorithm for the GNT problem. For any arbitrary $1-$identifiable matrix $A,$ it demonstrated successfully how to accurately estimate the distribution of the random vector $X,$ with mutually independent components, using only IID samples of the components of the random vector $Y = AX.$ Translating to network terminology, this means that one can now address the tomography problem even for networks with arbitrary topologies using only pure unicast probe packet measurements. The fact that we need only the IID samples of the components of $Y$ shows that the processes to acquire these samples across different paths can be asynchronous. Another nice feature of this approach is that it can estimate the unknown link level performance parameters even when no prior information is available about the same. \section*{Acknowledgement} I would like to thank my advisors, Prof. V.~Borkar and Prof. D.~Manjunath, for guiding me right through this work. I would also like to thank C.~Wampler, A.~Sommese, R.~Gandhi, S.~Gurjar and M.~Gopalkrishnan for helping me understand several results from algebraic geometry. \bibliographystyle{IEEE}

INTRODUCTION ============ Thyroid nodules are very prevalent -- they are found in approximately 8% of adults by palpation, 41% by means of ultrasound (US), and in 50% in autopsy pathological examination.^([@B1])^ Thyroid malignancy is relatively rare, and it is diagnosed in approximately 10% of all thyroid nodules.^([@B2]-[@B4])^ The appropriate indication of which nodules should undergo fine needle aspiration (FNA) and which can be monitored is still under debate. In the past two decades several controversies regarding the malignant characteristics aroused, but there is no definitive classification yet.^([@B1],[@B5]-[@B7])^ In the last 5 years, some publications have tried to establish a reliable guideline for thyroid nodule sonographic evaluation.^([@B3]-[@B5],[@B8],[@B9])^ Based on the already established *Breast Imaging Reporting and Data System* ^®^ (BI-RADS^®^) for breast nodules,^([@B10])^ some reports suggested a categorization system of US features in thyroid nodules - the Thyroid Imaging Reporting and Data System (TI-RADS). The purpose of TI-RADS is to group the nodules into different categories with a similar percentage of malignancy as in BI-RADS^®^. It is based on thyroid nodules classification, exclusively regarding B-mode ultrasonographic features, to reduce inter-observer and inter-device variability. OBJECTIVE ========= To present the results of our initial experience in the correlation between TI-RADS and the Bethesda system for cytopathology reports of 1,000 thyroid nodules in patients who underwent sonographic evaluation, followed by fine needle aspiration, and classified according to TI-RADS system. METHODS ======= The institutional review board approved this retrospective study, and the requirement to obtain Informed Consent form was waived. From November 2011 to February 2014, US scan of thyroid gland and neck area and US-guided FNA of thyroid focal nodules were performed by experienced physicians, in our intervention center. The Medical Ethics Committee of the organization approved the study under protocol number CAAE: 41699015.8.0000.0071. A total of 1,000 thyroid nodules in 906 patients were analyzed and classified according to TIRADS, without prior knowledge of the cytological results. The US equipment used were the ATL HDI 5000 (Absolute Medical Equipment, Wesley Hills, New York, United States), IU 22 Philips (Philips Healthcare, Andover, Massachusetts, United States), Aplio 500 Platinum (Toshiba American Medical Systems, Tustin, California, United States) and My Lab 75 (Esaote, Genova, Italy), and the acquired images stored in the PACS System (Carestream Health, California, United States). The TI-RADS classification ranged from 1 (negative findings) to 6 (known proved malignancy) and category 4 was further divided into subcategories 4A (low suspicion), 4B (intermediate suspicion) and 4C (moderate suspicion). The higher the grade of the nodule, the greater the risk of malignancy is ([Table 1](#t1){ref-type="table"} and [Figure 1](#f01){ref-type="fig"}). Table 1Thyroid Imaging Reporting and Data System (TI-RADS) classificationTI-RADSDefinitionUltrasound features1NegativeNormal thyroid2BenignBenign features3Probably benignWithout suspicious features4ALow suspicionOne suspicion feature4BIntermediate suspicionTwo suspicion features4CModerate suspicionThree or four suspicion features5High suspicionFive suspicion features6Known proved malignancyConfirmed malignancy Figure 1Examples of thyroid nodules submitted to cytological examination. (A) Nodule classified as TI-RADS 2; (B) TI-RADS 3; (C) TI-RADS 4A; (D) TI-RADS 4B; (E) TI-RADS 4C; (F) TI-RADS 5; cases A, B and C were considered benign; cases D, E and F were consider malignant according to Bethesda system(11) The US features associated to higher malignancy risks were irregular margins, hipoechogenicity, marked hypoechogenicity (defined as solid nodules, without enhancement or spots, with areas of greater hypoechogenicity within the nodule, or in relation to other hypoechoic areas of the thyroid gland; in this situation, we considered both points), morphology longer than wide, and microcalcifications. FNA was performed by freehand technique under US guidance, using a 23-gauge needle attached to a 20cc syringe. Upon aspiration, a negative pressure was maintained until blood appeared in the hub of the syringe. The anesthesia used was the combination of local anesthetic (lidocaine) and ice. Crossing thyroid vessels was avoided to prevent local bleeding; in mixed nodules, solid areas were chosen. In all FNA procedures, prior to patient discharge, a cytologist assessed the sample to avoid unnecessary punctures and insufficient specimens. After this initial evaluation, experienced pathologists evaluated all samples according to Bethesda system ([Table 2](#t2){ref-type="table"}). Table 2The Bethesda System for Reporting Thyroid CytopathologyCategoryMeaningINon-diagnostic or inadequateIIBenignIIIAtypia/follicular lesion of undetermined significanceIVFollicular neoplasm or suspicious for follicular neoplasmVSuspicious for malignancyVIMalignant[^2] The relation between TI-RADS and Bethesda was evaluated through double entry tables, χ^2^ and Pearson's correlation tests. Considering the biopsy results as malignant or benign, to analyze the association we used binary logistic regression models and assessed the odds ratios of malignancy for each TI-RADS category. The odds ratios were expressed using 95% confidence interval (95%CI). The tests were performed using the Statistical Package for Social Science (SPSS) for Windows, version 17.0, and considering the significance level (α) of 5%. RESULTS ======= One thousand examinations of 906 patients were carried out. Of the total of 1,000 examinations, 24 were Bethesda I and were excluded; hence, we had 976 complete examinations. The nodules classified as Bethesda cytology IV, V and VI were considered suspicious for malignancy. Taking into account all nodules in the analysis, a significant association was observed between the TI-RADS and the Bethesda (p\<0.001) classification, and those with TI-RADS rating 2 or 3 were mostly Bethesda 2 (95.5% and 92.5%, respectively). Among those classified as TI-RADS 4C and 5 (68.2% and 91.3%, respectively), the majority was Bethesda 6 ([Table 3](#t3){ref-type="table"}). Table 3Thyroid Imaging Reporting and Data System (TI-RADS) and Bethesda correlationTI-RADS classificationBethesda results23456TotalAll nodulesn (%)n (%)n (%)n (%)n (%)n (%)2120 (96)4 (3.2)0 (0)0 (0)1 (0.8)1253432 (93.3)23 (5)0 (0)0 (0)8 (1.7)4634A192 (73.3)28 (10.7)6 (2.3)1 (0.4)35 (13.4)2624B35 (43.2)11 (13.6)3 (3.7)1 (1.2)31 (38.3)814C3 (13.6)3 (13.6)1 (4.5)0 (0)15 (68.2)2251 (4.3)1 (4.3)0 (0)0 (0)21 (91.3)23Total783 (80.2)70 (7.2)10 (1)2 (0.2)111(11.4)976 There was an unexpected malignancy case in a nodule that had been classified as TI-RADS 2. A retrospective revision of the images showed that the nodule, in fact, should have been classified as 3 ([Figure 2A](#f02){ref-type="fig"}). Probably the nodule was misclassified by wrongly considering the solid part as spongiform and, in fact, it did not contain colloid foci. There was also one nodule classified as TI-RADS 5 that was benign. It was a case of thyroiditis, confirmed in follow-up US exams ([Figure 2B](#f02){ref-type="fig"}). All the remaining percentages of malignancy in cytology were similar to BI-RADS^®^ method, widely accepted and established. Figure 2Thyroid ultrasound with unexpected results. (A) Nodule classified as TI-RADS 2 that showed malignancy in cytological examination. (B) Nodule classified as TI-RADS 5 with granulomatous thyroiditis in cytological examination (Bethesda II) In the analysis considering all nodules, the proportion of malignant nodules classified as TI-RADS 2 was 0.8%, and among those TI-RADS 3, it was 1.7%. We classified as 4A 16%, 43.2% as 4B, 72.7% as 4C classification, and 91.3% as TI-RADS 5. The results of the logistic regression model was p\<0.001, showing a clear association between TI-RADS and biopsy results. The group rated TI-RADS 3 was considered as the reference for the model to be more numerous. The risk of malignancy for patients classified 4A was estimated at 10.86-fold the risk for those rated as 3 (95%CI: 5.0-23.5). For the individuals with 4B classification, the risk of malignancy was estimated to be 43.27-fold that for patients rated 3 (95% CI: 18.95-98.92). The other estimated odds ratios are presented in [table 4](#t4){ref-type="table"}. Table 4Thyroid Imaging Reporting and Data System (TI-RADS) and correlation with risk of malignancyTI-RADS All nodulesMalignancyOR 95%CIp valueBenignMalignantTotaln (%)n (%)n (%)2124 (99.2)1 (0.8)1250.46 (0.06-3,7)0.4643455 (98.3)8 (1.7)463ReferenceReference4A220 (84)42 (16)26210.86 (5-23.52)0.0024B46 (56.8)35 (43.2)8143.27 (18.9-98.82)\<0.0014C6 (27.3)16 (72.7)22151.67 (47-488.68)\<0.00152 (8.7)21 (91.3)23597.19 (119.3-2987.7)\<0.001Total853 (87.4)123 (12.6)976 [^3] DISCUSSION ========== Thyroid US should be performed in the initial assessment of the gland.^([@B12],[@B13])^ The FNA is an inexpensive and useful tool for detecting thyroid cancer, but it is an invasive procedure. In the management of thyroid nodule patients, recommending who should be submitted to FNA is still controversial. In the last decade, the improvement of Doppler US evaluation drew interest in rating thyroid nodules, based on spectrum and speed mapping parameters, which initially proved promising. In this context, the classification proposed by Chammas et al*.* ^([@B5])^ emerged as one of the main methods used. However, the methodology has some limitations, including much variability inter-examiners and inter-devices, which is greater in evaluation by Doppler as compared to B-mode parameters. Similarly, the retrospective analysis of images is greatly compromised. Several classifications based on US features have been proposed in the last decade, in an attempt to facilitate this selection. However a consensus has not been established, given the difficulty of reproducibility of different classifications proposed, or even due to the low correlation between the US reports and cytology results.^([@B3]-[@B5],[@B8],[@B9],[@B12])^ Currently there is a tendency to standardize the imaging evaluation of different organs^([@B10],[@B14],[@B15])^ with reliable and easily reproducible classifications. The main example is the already established BI-RADS^®^ classification for breast nodules. The TI-RADS classification aims to correlate US features to cytological classification, increasingly graduating the risk of a nodule being malignant, according to the number of features present in the US. Among diverse classifications, Horvath et al.,^([@B3])^ by means of a prospective analysis, proposed ten US patterns to be analyzed during the examination and nodule classification from TI-RADS 2 to 6 (category 4 divided into 4A and 4B) and estimated a malignancy risk of 14.1% in TI-RADS 3, 45% in TI-RADS 4, and 89.6% in TI-RADS 6. Kwak et al.^([@B4])^ proposed a TI-RADS classification through retrospective analysis of patients submitted to thyroid US and FNA, considering the risk of malignancy and subdivisions similar to the BI-RADS^®^ classification (that is, with three subdivisions for category 4), using five US criteria that can be added during thyroid evaluation. This article also described that a malignancy risk lower than 3% is expected for TI-RADS 3, a risk of 3.6 to 91.9% for TI-RADS 4, and of 88.7 to 97.9% for TI-RADS 5. The present study has differences in relation to that proposed by Horvath et al.,^([@B3])^ such as being retrospective and with one more subdivision in the category 4, by adding 4C. Besides our purpose was to facilitate the classification process, reducing from ten to only four features in B-mode US considered in our classification. It also differed from the study by Kwak et al.^([@B4])^ in this issue, since these authors used five features in the classification, one more than ours. This difference relied on the nodule composition, that we judged as liable to mistakes in some cases, since many mixed nodules could generate uncertainty about their precise composition in ultrasonographic evaluation. Instead, we considered two points of marked hypoechogenicity, because the nodules with such characteristics have an increased risk of malignancy as compared to those slightly hypoechoic.^([@B16],[@B17])^ Finally, our classification, unlike others proposed, does not use the sum of points for Doppler features. Although Doppler mapping help in thyroid US evaluation,^([@B5],[@B9],[@B18])^ its large-scale reproducibility is compromised due to the inter-examiner and inter-device variability.^([@B19])^ This study has some limitations, such as being retrospective, sonographic assessment performed by different operators and diverse US machines, use of cytology data instead of pathological data despite high sensitivity and specificity of cytology. Another limitation was the lack of uniformity of criteria for indicating puncture of nodules. CONCLUSION ========== TI-RADS can be considered an appropriate classification in the assessment of thyroid nodules, in order to avoid unnecessary fine needle aspirations and to assist in making decision about when it should be performed. This classification improves communication and reduces confusion among physicians and patients. Our experience demonstrated that the TI-RADS classification is highly reproducible, since it is based on B-mode characteristics of the nodules, especially when performed by experienced radiologists, acquainted with its use. [^1]: Conflict of interest: none. [^2]: Source: Cibas et al.^(11)^ [^3]: 95%CI: 95% confidence interval; OR: *odds ratio*.

Introduction ============ Of the malignant tumors of the adult population only 1% are soft tissue sarcomas (SPM)[@JR160417-1]. Of these, 5 to 16% occur in the head and neck[@JR160417-2] [@JR160417-3]. The malignant fibrous histiocytoma (MFH) is a rare tumor and there are few cases described in the literature, and thus difficult to define prognosis and treatment[@JR160417-4]. It was described in the 60s as a distinct type of SPM, and has been the most common histological type diagnosed since. The atypical fibroxanthoma (FA) is one of its variants. The SPM exhibit aggressive behavior, tendency to invasion and metastatic spread early. We present a case of probable recurrence of AF after two years, surgical treatment and a brief review of the literature. Case Report =========== Female patient, 63 years old, came to us complaining of swelling in the face (cheek), left, with six months of evolution and rapid growth in the past two months. She reported no other symptoms and disorders. However, reported history of resection of a skin lesion on the same location for about two years. Whose histopathology was inconclusive on histological lineage, however, assured clear margins and no additional study was conducted for a conclusive diagnosis. Physical examination revealed subcutaneous swelling in the left side of fibro-elastic consistency, adjusting, measuring approximately four by two centimeters, movable with respect to the depth and apparently fixed to the overlying skin, where scar visualized approximately three inches, corresponding prior to surgery. He had no palpable cervical lymph nodes. Computed tomography (CT) revealed an increase in volume in the subcutaneous tissue of the left face, with contrast medium uptake and cleavage plane with the left parotid gland deeply. We performed fine needle aspiration biopsy that showed only atypical cells. During this period there was an increase in lesion volume and apparent infiltrating the epidermis. Due to suspicion of malignancy with rapid evolution, we opted for surgical resection of the lesion accompanied by frozen section intraoperative examination. This has not been able to define the histological lineage, but confirmed that it was invasive malignancy. Resection was performed with safety margins and included the overlying skin. Primary closure was performed and postoperatively the patient developed salivary fistula, treated with pressure dressings and healing after 15 days. Pathology showed undifferentiated malignant neoplasm characterized by proliferation of spindle cells, pleomorphic epithelioid and some involving the dermis and subcutaneous tissue. Observing occasional mitotic figures. The surgical margins were free. The immunohistochemical study revealed for CD10 expression, favoring the diagnosis of atypical fibroxanthoma (malignant fibrous histiocytoma surface). The tumor was staged as N0 M0 pT1a - stage IIB (UICC 2006). It was chosen to carry out radiotherapy or chemotherapy. The patient had no signs of local recurrence or distant metastasis when, after four months was detected on clinical examination new tumor in the topography of the mandibular body left in an area that had no contact with the resected lesion. A biopsy of the tumor showed that it was FA. The patient was referred to a tertiary hospital to give the sequel treatment, since it required at this time a more extensive resection followed by reconstruction of the mandible. Which was done so as radiotherapy. Discussion ========== The SPM represent only 1% of all cancers of the head and neck[@JR160417-5]. They are grouped together because of similarities in their appearance pathological, clinical presentation and natural history, although they have different cellular origins[@JR160417-3] [@JR160417-6]. MFH is a rare tumor, but it can be in the region of the head and neck should be part of the differential diagnosis of primary cervical lesions. One of its variants is the FA. The clinical behavior is characterized by the tendency to invasion of adjacent soft tissues and early metastatic dissemination[@JR160417-7]. Occasionally these tumors are associated with genetic syndromes or prior radiation. Most often there is no clear etiology[@JR160417-2]. Men are more affected than women (3:2). The most common presentation is as painless swelling[@JR160417-5]. Many symptoms may occur depending on the location of the tumor[@JR160417-2]. Physical examination usually reveals sub mucosal or subcutaneous tumor with distortion or destruction of adjacent structures. CT and magnetic resonance imaging (MRI) scans are the choice in most cases and provide information about bone involvement, extension and lymph node metastasis in addition to size and location[@JR160417-2]. These tests can be used in a complementary way, especially in surgical planning[@JR160417-8]. However, for a histological diagnosis in addition to the microscopic characteristics, it is necessary to perform immunohistochemical analysis[@JR160417-9]. The prognostic factors are histology grade, location, size and lymph node involvement[@JR160417-7]. MFH is generally considered a lesion of high-grade malignancy[@JR160417-10] [@JR160417-11]. Regarding location in the head and neck, is considered a worse prognosis with higher recurrence rates and lower disease-specific survival[@JR160417-12]. Tumors larger than three centimeters in diameter negatively affect the prognosis. Literature data show that surgical safety margin is among the most important prognostic factors affecting local control, although its significance as a determinant of survival remains uncertain[@JR160417-13]. In the study of Belal et al.[@JR160417-3], positive microscopic margins were associated with increased local recurrence. This author also found a significant correlation between adequate margins and disease-free survival and overall survival. In the head and neck high rates of positive margins are reflective of the closeness of relationships between adjacent structures. Local recurrence occurs in approximately 20% of cases, but the biggest cause of death is distant metastasis. The survival rate is 40% in the largest series, ranging from 19 to 75%[@JR160417-5]. The authors differ if the local recurrence itself is a cause of distant metastases and increased mortality[@JR160417-5] [@JR160417-14]. The mainstay of treatment is wide surgical resection with clear margins[@JR160417-1]. Some authors consider elective neck dissection is not necessary since regional metastases are uncommon. Postoperative radiotherapy is essential for banks affected[@JR160417-3]. It is also recommended for unresectable tumors and margins slim. The role of adjuvant chemotherapy is controversial[@JR160417-3], its primary indication is for cases of metastatic disease[@JR160417-15]. In the case of AF patients had a facial tumor with no metastases and that probably it was a local recurrence of resected lesion two years ago, but without histopathological diagnosis. We opted for surgical treatment, associated with intraoperative frozen due to great suspicion of malignancy, diagnostic difficulty and rapid progression at the time of preoperative investigation. The intraoperative frozen was not able to define the histological diagnosis, as this can only be determined after immunohistochemistry. After the histological diagnosis of AF (HFM) neck dissection was not performed due to lack of commitment to regional clinical and radiological evaluations. Also we decided not to radiotherapy of believing in a small possibility of recurrence due to tumor staging and the wide margins of resection (minimum of 1 cm) secured by microscopic analysis. The literature is unclear as to the minimum margin of safety. Chemotherapy was not used in view of the controversies reported in the literature. But after four months of surgery was the emergence of new tumor in the left side, which after biopsy showed that it was FA. The patient was referred to a tertiary hospital for treatment to follow. Conclusion ========== The FA is a rare and difficult histological diagnosis. Surgery with wide resection of the lesion is the mainstay of treatment. The identification of tumor histology is of fundamental importance, as well as the postoperative follow-up. In this case the histological diagnosis of the primary lesion would provide better preoperative planning and early surgical intervention after local recurrence. {#FI160417-1}

\section{Statistical model descriptions of particle yields} One of the main objectives of heavy ion physics is to study the collective properties of strongly interacting matter. It's equation of state, transport coefficients, degree of equilibration and phase structure, and the dependence of these on energy and system size. Thus, the natural approach to study soft particle production in heavy ion collisions is through statistical mechanics techniques. Such an approach has a long and illustrious history \cite{Fer50,Pom51,Lan53,Hag65}. However, the systematic quantitative comparison of data to the statistical model is a comparatively recent field; A consensus has developed that the statistical hadronization model can indeed fit most or all particles for AGS,SPS and RHIC energies \cite{jansbook,bdm,cleymans,barannikova,gammaq_energy,gammaq_size,becattini}. This consensus, is, however, somewhat superficial. The models cited in the references above are considerably different w.r.t. each other, beyond their common origin in statistical mechanics. As a result, linking fit results to physics in a model-independent way has proven to be somewhat problematic. For instance, the observation of ``the horn'', a sharp dependence of certain observables, notably the $K^+/\pi^+$ ratio, with energy \cite{horn} has generated a lot of theoretical interest, and is generally seen as remarkably suggestive of a phase transition. However, many completely different explanations were given for this phenomenon, all to some extent based on statistical mechanics: The original $K^+/\pi^+$ ``horn'' dependence was predicted \cite{horn_theory} as a signature of deconfinement; The motivating argument was that the onset of the phase transition should be marked by a sharp change in the strangeness over entropy ratio, tracked by $K/\pi$. This idea has recently been realized \cite{gammaq_energy} in terms of non-equilibrium at freeze-out: The extra strangeness and entropy in the high-temperature phase translate into lack of chemical equilibrium \textit{after} the phase transition, resulting in super-cooling (freeze-out temperature $\sim$ 140 MeV) and phase-space over-saturation (over-abundance of both light and strange quarks w.r.t. equilibrium expectations). Other authors, however, have interpreted the horn without advocating deconfinement; Models have been developed where the sharp structure arises due to the transition to the Grand Canonical limit for strangeness or from freeze-out conditions where mesons rather than baryons carry the bulk of the entropy and conserved quantum numbers \cite{hornthermal}. Another proposed explanation relies on the {\em absence} of equilibration between Kaons and pions \cite{tomasik}. While this variety of different models is a sign of both the theoretical and experimental vitality of our field, it would be very useful to rule some of these models out. A similar ambiguity is present in the interpretation of strangeness enhancement, now firmly observed both at SPS \cite{stenhsps} and RHIC \cite{stenhrhic} energies. It was originally proposed on the grounds that a perturbative gluon-rich QGP is a lot more efficient as a strangeness production mechanism than a hadron gas \cite{jansbook}. However, a considerable amount of recent attention has been given to models where the enhancement is explained by a \textit{suppression} of strange particles in low-multiplicity systems due to deviation from the Grand-Canonical limit \cite{strangeness_canonical}. The continued ambiguity over these models is evidence that further observables need to be accessed before either of the strangeness production mechanisms can be considered ``falsified''. Finally, a topic of considerable interest is the relationship between the statistical model and freeze-out. The RHIC HBT puzzle \cite{hbt,hydroheinz} makes it clear that freeze-out conditions are, at the moment, not well understood. In particular, it is unclear what, if any, is the effect of the evolution between hadronization (the formation of hadrons as effective degrees of freedom) and freeze-out (the moment when all hadrons decouple). As shown in \cite{hydroheinz}, the discrepancy with experiment would lessen if chemical and kinetic freeze-out were to coincide. While, as claimed in \cite{hydroheinz}, such a high freeze-out temperature would spoil agreement with particle spectra, fits based on a single freeze-out model prescription have shown that, provided a correct treatment of resonances is maintained, particle spectra are compatible with simultaneous freeze-out \cite{florkowski,us_sps,us_rhic}. This somewhat invalidates the claims \cite{stagedfo} that ``blast-wave'' fits to spectra (where resonance feed-down is not included) are evidence that freeze-out occurs after a long period of re-interaction, with $\Xi$ s and $\Omega$ s freezing out before $\pi,K$ and protons. A logical step to clarify these scenarios is to directly measure the abundance of resonances. Here, the situation becomes even more ambiguous: As pointed out in \cite{usresonances}, resonance abundance generally depends on two quantities: $m/T$, where $m$ is resonance mass and the chemical freeze-out temperature, as well as $\tau \Gamma$, where $\Gamma$ is the resonance width and $\tau$ is the reinteraction time. Observing two ratios where the two particles have the same chemical composition, but different $m$ and $\Gamma$, such as $\Lambda(1520)/\Lambda$ vs $K^*/K$, or $\Sigma(1385)/\Lambda$ vs $K^*/K$, could therefore be used to extract the magnitude of the freeze-out temperature and the re-interaction time. Studies of this type \cite{fachini,salur} are still in progress; As we will show, $\Lambda(1520)/\Lambda$ and $K^*/K$ seem to be compatible with sudden freeze-out, provided freeze-out happens in a super-cooled over-saturated state, as suggested in \cite{jansbook,gammaq_energy,gammaq_size}. Other preliminary results, such as $\rho/\pi$, $\Delta/p$ \cite{barannikova}, and now $\Sigma^*/\Lambda$ \cite{salur} seem to be produced \textit{in excess} of the statistical model \cite{fachini,barannikova}, both equilibrium and not. It is difficulty to see how a long re-interacting phase would produce such a result: Resonances whose interaction cross-section is small w.r.t. the timescale of collective expansion would generally be depleted by the dominance of rescattering over regeneration processes at the detectable (on-shell) mass range. More strongly interacting resonances would be re-thermalized at a smaller, close to thermal freeze-out temperature. Both of these scenarios would generally result in a suppression, rather than an enhancement, of directly detectable resonances. Transport model studies done on resonances generally confirm this \cite{urqmd1,urqmd2} The observation of an enhanced $\mu^+ \mu^-$ continuum around the $\rho$ peak \cite{na60} has been pointed to as evidence of $\rho$ broadening, which in turn would signify a long hadronic re-interaction phase \cite{na60}. The absence of a broadening in the {\em nominal} peak itself, prevents us from considering this as the unique interpretation of experimental data. Moreover, even a conclusive link of broadening with hadronic re-interactions would still give no indication to the length of the hadronic rescattering period. Nor it would resolve the discrepancies pertaining hadronic resonances encountered in the previous paragraph; The lack of modification in either mass or width, between p-p and Au-Au seen so far \cite{fachini,salur} is only compatible with the NA60 result\footnote{Unless one assumes that at freeze-out the SPS system to be very different from RHIC's. But no soft data so far would lend support to such a hypothesis, and in all statistical models mentioned here conditions at SPS and RHIC are comparable, save for the larger SPS chemical potentials} provided the $\rho$ is {\em very} quick to thermalize, so \cite{fachini} sees only the $\rho$ s formed close to thermal freeze-out. But, as argued in the previous paragraph, that under-estimates the abundance of the $\rho$ and other resonances, since the $\rho/\pi,K^*/K$, and even $\Lambda(1520)/\Lambda$ ratios point to a freeze-out temperature significantly above the 100 MeV, commonly assumed to be the ``thermal freeze-out'' temperature in a staged freeze-out scenario. It is apparent that the data surveyed above is insufficient for a conclusive analysis. One or more missing observables are required. It is desirable that these observables have a different dependence on the non-equilibrium parameters $\gamma_{q,s}$ (see Eq. \ref{chemneq}), system volume, and temperature, so as to decorrelate these parameters and falsify the non-equilibrium picture in \cite{gammaq_energy,gammaq_size} and it's explanation for the ``horn'' and related discontinuities. It is also desirable that these observables be strongly ensemble Dependant, to test the Canonical origin theory of strangeness enhancement. Finally, a direct experimental sensitivity to the re-interaction period would be invaluable to resolve the conundrums of the preceding paragraphs. As we will show, considering both yields and fluctuations within the same analysis could have all three requirements. \section{Fluctuations in the statistical model} Event-by-event particle fluctuations have elicited theoretical~\cite{jeonkochratios,fluctqgp1,fluctqgp2,prcfluct} and experimental interest~\cite{starfluct2,phefluct,supriya,spsfluct}, both as a consistency check for existing models, as a test of hadron gas equilibration and as a way to search for new physics. The ability of the SHM to describe not just averages, but event-by-event multiplicity fluctuations has however not been widely investigated, and its applicability is currently a matter of controversy. In this work, we use the Grand Canonical ensemble (most physically appropriate for describing RHIC data \cite{prcfluct}), implemented in a publicly-available software package \cite{share,sharev2} to calculate fluctuations $\ave{(\Delta N)^2}$ and yields ($\ave{N}$). At freeze-out, these are given by \begin{eqnarray} \label{yield_formula} \langle N_i\rangle &=& gV\int {d^3p\over (2\pi)^3}\, {1\over \lambda_i^{-1} e^{\sqrt{p^2+m_{i}^{2}}/T}\pm 1}\\\langle(\Delta N_i)^2\rangle & = & \lambda_i{\partial \over \partial \lambda_i} \langle N_i\rangle \end{eqnarray} The parameter $\lambda_i$ corresponds to the particle fugacity, related to the chemical potential by $\lambda_i=e^{\mu_i/T}$. Provided the law of mass action holds, it should be given by the product of charge fugacities (flavor, isospin etc.). It is then convenient to parametrize it in terms of equilibrium fugacities $\lambda_i^{\mathrm{eq}}$ and phase space occupancies $\gamma_i$. For a hadron with $q (\overline{q})$ light quarks, $s (\overline{s})$ strange quarks and isospin $I_3$ the fugacity is then \begin{eqnarray} \label{chemneq} \lambda_i = \lambda_i^{\mathrm{eq}} \gamma_q^{q+\overline{q}} \gamma_s^{s+\overline{s}} \phantom{A},\phantom{A} \lambda_i^{\mathrm{eq}}=\lambda_{q}^{q-\overline{q}} \lambda_{s}^{s - \overline{s}} \lambda_{I_3}^{I_3} \end{eqnarray} If the system is in chemical equilibrium then detailed balance requires that $\gamma_q=\gamma_s=1$. Assuming $\gamma_q=1$ and fitting particle ratios gives the $T \sim 160-170$ MeV freeze-out temperature typical of chemical equilibrium freeze-out models at SPS and RHIC \cite{bdm,cleymans,barannikova}. In a system expanding and undergoing a phase transition, however, the condition of chemical equilibrium no longer automatically holds, so one has to allow for the possibility that $\gamma_q \ne 1,\gamma_s \ne 1$. In particular, if the expanding system undergoes a fast phase transition from a QGP to a hadron gas, chemical non-equilibrium \cite{jansbook} and super-cooling \cite{csorgo} can arise due to entropy conservation: By dropping the hadronization temperature to $\sim 140$ MeV and oversaturating the hadronic phase space above equilibrium ($\gamma_q \sim 1.5,\gamma_s \sim 2$), it is possible to match the entropy of a hadron gas with that of a system of nearly massless partons \cite{jansbook}. In such a case, the pion chemical potential is very close to, but not quite reaching the point of Bose-Einstein condensation $\lambda_i \rightarrow e^{m_i/T}$. The onset of these conditions (and of the associated strangeness over-saturation condition $\gamma_s>1$) is the basis of the non-equilibrium statistical model description of the ``horn'' mentioned in the preceding section \cite{gammaq_energy}. This limit should result in enhanced fluctuations \cite{prcfluct}, since higher order terms in the Bose-Einstein expansion are always more significant than for yields: for $\lambda_i e^{m_i/T} = 1-\epsilon$ $\ave{N_i}$ converges but $\ave{\Delta N_i^2}$ diverges as $\sim \epsilon^{-1/2}$. Thus, a measure of both fluctuations and yields could serve to falsify the non-equilibrium and equilibrium statistical model scenarios: $T$ and $\gamma_q$ correlate in yields and anti-correlate in fluctuations. Measuring both yields and fluctuations can be used to determine both temperature and $\gamma_q$ precisely. The final state yield of particle $i$ is computed by adding the direct yield and all resonance decay feed-downs. \begin{eqnarray} \label{resoyield} \langle N_i\rangle_{\rm total} & = & \langle N_i\rangle + \sum_{{\rm all}\;j \rightarrow i} B_{j \rightarrow i} \langle N_j\rangle \\ \ave{(\Delta N_{ i})^2} & = & \sum_{{\rm all}\;j \rightarrow i} B_{j\to i}(1-B_{j\to i})\ave{N_j} + \sum_{{\rm all}\;j \rightarrow i} B_{j\to i}^2 \ave{(\Delta N_j)^2}\phantom{.....} \end{eqnarray} assuming no primordial correlation between $i$ and $j$. The fluctuation of a ratio $N_1/N_2$ can be computed from the fluctuation of the denominator and the numerator \cite{jeonkochratios} ($\sigma_{X}^2=\ave{(\Delta X)^2}/\ave{X}$): \begin{equation} \label{fluctratio} \sigma_{N_1/N_2}^2 = \frac{\ave{(\Delta N_1)^2}}{\ave{N_1}^2} + \frac{\ave{(\Delta N_2)^2}}{\ave{N_2}^2} - 2 \frac{\ave{\Delta N_1 \Delta N_2}}{\ave{N_1}\ave{ N_2}}. \end{equation} This is a a useful experimental observable, since volume fluctuations, difficult to model from first principles, cancel out event-by-event. Note that fluctuations of ratios scale as the inverse of the absolute normalization, $\sigma_{N_1/N_2} \sim (\ave{V}T^3)^{-1}$. Hence, a fit including such a fluctuation needs to have absolute normalization as one of the fit parameters. Fitting a particle multiplicity, $\sim (\ave{V}T^3)$ together with the ratio provides a way to measure the system volume at chemical freeze-out with great precision, due to the very different scaling of the two types of observables. Note also the appearance of a negative correlation term between $N_1$ and $N_2$ stemming from a common resonance feed-down ($\Delta \rightarrow p \pi$ will be a source of correlation between $N_{p}$ and $N_{\pi}$). \begin{equation} \ave{\Delta N_1 \Delta N_2} \approx \sum_j b_{j \rightarrow 1 2} \ave{N_j} \end{equation} This correlation term is an invaluable phenomenological resource, since it is sensitive to resonance abundance at {\em chemical} freeze-out \cite{jeonkochratios}; Further reinteraction, provided resonances are not kicked out of the detector's acceptance region \cite{prcfluct}, preserve the correlation even when the original resonance stops being reconstructible. Thus, a measurement of the correlation together with the directly detectable resonance is a direct measurement of the amount of re-interaction between the thermal and chemical freeze-out. As can be seen, fluctuations measurements fulfill all three of the requirements enumerated in the last section. They are sensitive to the degree of chemical non-equilibrium and the system volume. They can gauge the amount of hadronic re-interaction after hadronization. And, as shown in \cite{nogc2}, they are strongly ensemble-specific. In the next section, we will attempt to interpret current RHIC data in both yields and fluctuations in light of the controversies described in the previous section, and to suggest experimental observables capable of further clarifying the situation. \section{Statistical model comparison with preliminary RHIC data} \begin{figure}[tb] \epsfig{height=7.cm,clip=,figure=statprof.eps} \epsfig{height=7.cm,clip=,figure=pars_gq.eps} \caption{\label{par}Left panel: Statistical significance profile for the 200 GeV fit including (solid line) and excluding (long-dashed line) the $\sigma^{dyn}_{K/\pi}$ data-point. Right panel: Parameter sensitivities for the fits for $T$,$d\ave{V}/dy$ and $\gamma_s$} \end{figure} The main experimental problem with fluctuation measurements is the vulnerability to effects resulting from limited detector acceptance. This difficulty can be lessened, to some extent, by considering ``dynamical'' fluctuations, obtained by subtracting a ``static'' contribution which should be purely Poisson in an ideal detector. \begin{equation} \sigma^{dyn}=\sqrt{\sigma^2-\sigma_{stat}^2} \end{equation} $\sigma_{stat}$, usually obtained through a Mixed event approach \cite{pruneau}, includes a baseline Poisson component, which for a ratio $N_1/N_2$ can be modeled as \begin{equation} \label{poissrat} \sigma_{stat}^2 = \frac{1}{\ave{N_1}}+\frac{1}{\ave{N_2}} \end{equation} as well as a contribution from detector efficiency and kinematic cuts. Provided certain assumptions for the detector response function hold (see appendix A of \cite{pruneau}), subtracting $\sigma_{stat}$ from $\sigma$ should yield a ``robust'' detector-independent observable. \begin{figure}[tb] \epsfig{width=7.cm,clip=,figure=graph_200.eps} \epsfig{width=7.cm,clip=,figure=pred_200.eps} \caption{\label{graph200} Left panel: Fit of preliminary 200 GeV data, including the $K^{\pm}/\pi^{\pm}$ fluctuations and the $K^* (892)$ and $\Lambda^* (1520)$ resonance. Black circles represent a fit where $\gamma_q$ was fitted. Red triangles down, $\gamma_q=1$. Blue squares, $\gamma_q=1$ and fluctuation not fitted. Right panel: Predictions for resonances and fluctuations calculated with best fit values. Filled circles represent full resonance correlation. Hollow circles assume 20 $\%$ of correlations due to resonances survive within the detector acceptance region, compatible with the calculation in \cite{prcfluct} for $\Delta \eta=\pm 0.5$ } \end{figure} We have performed a fit incorporating all ratios from \cite{barannikova}, with the exception of the $\Delta^{++}/p$, which the STAR collaboration has taken back in more recent analyzes \cite{starwhitepaper}. In addition, we have included the preliminary value for the $K/\pi$ event-by-event fluctuation measured by STAR \cite{supriya}, as well as the published yield for $\phi$ \cite{starphi200}, $\pi^-$ \cite{starpi200}, $K^*$ \cite{fachini} , $\Lambda(1520)$\footnote{This was taken at $5\%$ rather than 10$\%$ centrality. The systematic error on normalization should not however be a dominant effect} and $\Sigma^*$ \cite{salur}. As was pointed out in \cite{qm2005kpi}, the statistical model can not at present describe the difference between $\sigma^{dyn}_{K^-/\pi^-}$ and $\sigma^{dyn}_{K^+/\pi^+}$, although each of the data-points can be fitted acceptably on it's own, with a parameter variation of $10\%$ or so. Since these data points do not at present have a systematic error estimate, we do not regard this as evidence against our model. We choose, therefore, to fit the $\sigma^{dyn}_{K^-/\pi^-}$. Our fit parameters include the normalization (hopefully related to the system ``volume'' at chemical freeze-out), temperature, $\lambda_{q,s,I_3}$ and $\gamma_{q,s}$. We also require, by implementing them as data-points, strangeness, charge and baryon number conservation \[\ \ave{s-\overline{s}}=0 \pm 0.01 \phantom{AAAAA}\frac{\ave{Q}}{\ave{B}}=\left[ \frac{\ave{Q}}{\ave{B}}\right]_{Au}=0.4 \pm 0.01\] The results of the analysis are shown in Figs \ref{par} and \ref{graph200}. Fig. \ref{par} (left panel) shows a statistical significance profile in $\gamma_q$. At each point in the abscissa, a fit was performed minimizing all the other parameters ($T,\lambda_{q,s,I3},\gamma_s,d\ave{V}/dy$). The ordinate shows the statistical significance of such a best fit. The right panel shows what the best fit values of each parameter are at each point in $\gamma_q$. The left panel of Fig. \ref{graph200} shows what the best fit to data looks like in three cases: $\gamma_q$ fitted (circles), $\gamma_q=1$ (triangles), and $\gamma_q=1$ and $\sigma_{K/\pi}$ not fitted (hollow squares). As can be seen in both Fig. \ref{par} and \ref{graph200} (left panels), incorporating fluctuations, yields and ratios rules out equilibrium and forces $\gamma_q$ to be well above unity. An equilibrium fit of the particle yields and ratios gives an acceptable description of particle ratios, somewhat over-predicts the $K^*$ and $\Lambda(1520)$, but under-predicts $\sigma_{K/\pi}^{dyn}$ by many standard deviations. Choosing the Canonical ensemble for strangeness would only make this disagreement worse \cite{comingpaper}. If $\sigma_{K/\pi}^{dyn}$ is fitted together with ratios (red triangles down in Fig. \ref{graph200}), it forces the system volume $d \ave{V}/dy$ to be unrealistically small $(\sim 500 fm^3)$, thereby under-predicting particle yields by several standard deviations. It is only the addition of $\gamma_q$ (circles in Fig \ref{graph200}) that allows fluctuations to be driven to a high enough value while maintaining sufficiently high volume to describe the particle multiplicities, and sufficiently high temperature to describe ratios. The correlation between $T$ and $\gamma_q$ given by particle ratios then fixes temperature to $T=140$ MeV, as predicted in a super-cooling scenario \cite{jansbook,csorgo} and $ d \ave{V}/dy$ to $\sim 1000-1500fm^3$, in agreement with earlier estimates \cite{gammaq_size} and combatible with HBT measurements \cite{hydroheinz}. $\gamma_s$ is also required to be significantly above 1, also in agreement with earlier studies \cite{gammaq_energy,gammaq_size} and with a scenario where strangeness enhancement occurs because of the appearance of a phase where equilibrium strangeness content is higher than for an equilibrium hadron gas \cite{jansbook}. Quantitative disagreements w.r.t. \cite{gammaq_energy,gammaq_size} in $ d \ave{V}/dy$,$\gamma_s$, and chemical potentials arise from data choice, and can therefore be considered a systematic error. It is important to underline that both yields and fluctuations contribute to such a precise determination: Equilibrium statistical models can describe most yields and ratios acceptably with $\gamma_q=\gamma_s=1$, but fail to describe the event-by-event fluctuation. Conversely, transport models provide an acceptable description of event-by-event fluctuations \cite{flucttrans}, but fail to describe the yield of multi-strange particles \cite{urqmdstrange}. Unlike what is sometimes asserted, the $\Lambda(1520)$ and $K^*$ are acceptably described by the statistical model. The strongest disagreement arises from $\Sigma^*$ under-prediction, at the level of 1.5 standard deviations. We await for more measurements of resonances such as $\Delta$ and $\rho$ in central collisions before trying to interpret this under-prediction. The acceptable description of the $\Lambda(1520)$ and $K^*$ yield using the same freeze-out temperature as the stable particles, and the {\em under-prediction} of the $\Sigma^*$, makes a case for the proposition that the re-interaction period between hadronization and freeze-out might be not as significant as generally thought. However, the current data is not capable to {\em rule out} such a long-reinteraction period. The forthcoming measurement of ratios such as $\sigma^{dyn}_{K^-/\pi^+}$ and $\sigma^{dyn}_{K^+/\pi^-}$, however, will make such a falsification possible, since, unlike in the $\sigma^{dyn}_{K^-/\pi^-}$ case, a sizable resonance correlation term exists. For example, comparing \begin{eqnarray} \fl \phantom{A}\frac{1}{2} \left[ (\sigma^{dyn}_{K^+/\pi^+})^2 - (\sigma^{dyn}_{K^+/\pi^-})^2 \right] + \frac{1}{2} \left[ (\sigma^{dyn}_{K^-/\pi^-})^2 - (\sigma^{dyn}_{K^-/\pi^+})^2 \right] = \frac{\ave{\Delta K^+ \Delta \pi^-}}{\ave{K^+} \ave{\pi^-}} + \frac{\ave{\Delta K^- \Delta \pi^+}}{\ave{K^-} \ave{\pi^+}} \nonumber \\ \approx \frac{2}{3} \alpha_{K^{*0} \rightarrow K \pi} \left[ \frac{\ave{K^{*0} (892)}}{\ave{K^+}\ave{\pi^-}} + \frac{\ave{\overline{K^{*0}} (892)}}{\ave{K^-}\ave{\pi^+}} \right]_{chemical\phantom{A}f.o.} \end{eqnarray} with the directly measured $\frac{K^{* 0}}{K^-}$ gives the amount of re-interaction $K$s and $\pi$s went through between hadronization and freeze-out. The biggest systematic uncertainty of this measurement is likely to come from the need to estimate $\alpha_{K^* \rightarrow K \pi}$, the probability that the decay products of $K^{*0}$ will be in the detector acceptance region provided the original resonance was. In \cite{prcfluct} we discuss this number, and give some estimates of it's effect. A more precise estimate can be obtained via the Monte-Carlo based algorithms used to estimate the experimental acceptance as a function of phase space for the directly detected $K^*$. We believe that, while experimental acceptance has to be taken into account, it's effect is not strong enough to destroy the possibility of an estimate of $\left[ \frac{K^{* 0}}{K^-} \right]_{chemical\phantom{A}f.o.}$ Fig. \ref{graph200} (right panel) develops further the reasoning above, and uses the best fit parameters obtained using the current fit to calculate both resonances and fluctuations likely to be measured soon. The fluctuation calculation is given with both all $\alpha_{A B \rightarrow C}=1$ for all $A,B,C$ (filled circles) and $\alpha_{A B \rightarrow C}=0.2$ for all $A,B,C$ (hollow circles) to give an estimate of the systematics arising from resonance decays within a small experimental acceptance. Perhaps the measurement most likely to constrain the $\gamma_q >1$ scenario is $\sigma^{dyn}_{K^+/K^-}$, since this data-point is completely independent of pionic observables. As explained in the preceding section, $\gamma_q>1$ is better able to describe fluctuations because the high pion chemical potential gives pion fluctuations an additional boost. Thus, fluctuations of quantities not involving pions should be considerably suppressed since, even at $\gamma_s \sim 2$, B-E contributions for Kaons are too small to lift Kaon fluctuations too much above the Poisson limit. This explains why the lowest value for $\sigma^{dyn}$ in the given sample is $\sigma^{dyn}_{K^+/K^-}$, even through resonance correlations for this ratio are not nearly as numerous as for the case of, for example $\sigma^{dyn}_{p/\pi}$. It will shortly be clear weather the smaller $\sigma^{dyn}_{K^+/K^-}$ is experimentally confirmed. In conclusion, we have used preliminary experimental data to show that, at 200 GeV, the equilibrium model is unable to describe both yields and fluctuations within the same statistical parameters. The non-equilibrium model, in contrast, succeeds in describing almost all of the yields and fluctuations measured so far at RHIC, with the parameters expected from a scenario where non-equilibrium arises through a phase transition from a high entropy state, with super-cooling and oversaturation of phase space. We await more published data to determine weather the non-equilibrium model is really capable of accounting for both yields and fluctuations in all light and strange hadrons produced in heavy ion collisions. Work supported in part by grants from the Natural Sciences and Engineering research council of Canada, the Fonds Nature et Technologies of Quebec, and the Tomlinson foundation. We would like to thank J. Rafelski, C. Gale and S.Jeon for helpful discussions and continued support.

Introduction {#section1-2055116915593970} ============ An intracranial arachnoid cyst (IAC) is a developmental brain disorder in which fluid accumulates within a split of the arachnoid membrane. These structures do not fit the anatomical definition of a cyst and should be referred to as pseudocysts or diverticula. However, we have adopted the widely used term 'arachnoid cyst' in this article. The cavity does not usually communicate with the subarachnoid space ('non-communicating' cyst). However, IACs communicating with the subarachnoid space have been reported in humans.^[@bibr1-2055116915593970]^ In humans, IACs are often incidental findings, but their clinical incidence seems to be more frequent than previously thought in humans, as well as in dogs, although this is hard to confirm.^[@bibr1-2055116915593970]^ Associated clinical signs are usually focal and generalised seizures. IACs represent 1% of all intracranial masses in humans, and they have been sporadically reported in dogs, where they occur most commonly in the quadrigeminal cistern.^[@bibr1-2055116915593970][@bibr2-2055116915593970][@bibr3-2055116915593970][@bibr4-2055116915593970][@bibr5-2055116915593970][@bibr6-2055116915593970]--[@bibr7-2055116915593970]^ Only three cases have been reported in cats; all were located in the caudal fossa at the level of the quadrigeminal cistern.^[@bibr4-2055116915593970],[@bibr8-2055116915593970],[@bibr9-2055116915593970]^ This case describes an IAC with an unusual location treated conventionally by cystoperitoneal shunt (CPS) placement. Case description {#section2-2055116915593970} ================ A 1-year-old male neutered European cat was presented to the small animal neurology department at Vetagro Sup, Lyon, after a 24 h history of focal seizures and one generalised seizure. The cat had about 12 complex partial seizures characterised by orofacial motor signs (chewing, licking, and eye and facial twitching), hypersalivation, obtundation, tremors of the thoracic limbs and urination. The cat was presented after a generalised tonic-clonic episode. At admission, clinical features were unremarkable with the exception of slight obtundation and tachypnoea with ptyalism. Routine biochemistry (glucose, urea, creatinine, alkaline phosphatase, alanine aminotransferase, total proteins, total cholesterol and triglyceride) and bile acid stimulation were normal. Magnetic resonance imaging (MRI) of the brain was performed using a 0.2-Tesla MRI system (E-Scan XQ; Esaote). A large, well-demarcated, extra-axial cystic lesion was present in the left temporal area, mainly beneath the piriform lobe. It measured 1.6 cm in width, 1.7 cm in height and 2.1 cm in length at its largest points. The lesion was mainly supratentorial and was associated with mild mass effect. Rostrally, it extended up to the caudal mesencephalon. The signal intensity of the content was identical to cerebrospinal fluid (CSF; hypointense on T1-weighted and fluid-attenuated inversion recovery images, hyperintense on T2-weighted images) ([Figures 1](#fig1-2055116915593970){ref-type="fig"} and [2](#fig2-2055116915593970){ref-type="fig"}). This lesion did not enhance after administration of gadolinium (Dotarem; Guerbet). {#fig1-2055116915593970} {#fig2-2055116915593970} The degree of parenchymal compression of the forebrain was calculated with a method previously described by Matiasek et al for IAC in dogs.^[@bibr10-2055116915593970]^ The calculated compression was 41.2%. The MRI features were consistent with the diagnosis of a supratentorial left arachnoid cyst with left hemisphere (and particularly temporal lobe) compression. The seizures were controlled with anticonvulsant treatment started after admission (phenobarbital 2.5 mg/kg q12h). Surgical treatment was cystoperitoneal shunt placement ([Figure 3](#fig3-2055116915593970){ref-type="fig"}). {#fig3-2055116915593970} A dorsolateral approach to the skull was chosen, lateral to the sagittal crest. A peritoneal catheter was tunneled to the left hypochondrium and introduced into the peritoneal cavity by mini-laparotomy. After periosteal elevation, a 5 mm hole was drilled into the temporal bone. The dura was incised after preventive hemostasis, and the cystic catheter was placed. It was anchored to the temporal bone and fixed to a low-pressure valve in the neck, under the muscular fascia. The valve was fixed to the peritoneal catheter. Cystic fluid analysis collected during surgery was consistent with normal CSF. MRI performed 3 weeks after surgery showed a total resolution of the cyst. However, persistent modifications of the brain parenchyma were observed in the temporal lobe, including decreased thickness of the temporal lobe compared with the contralateral side and slight signal abnormalities in T2-weighted images ([Figure 3](#fig3-2055116915593970){ref-type="fig"}). Anticonvulsant treatment was ended 17 months after surgery with progressive tapering for 6 months. Four months after treatment discontinuation, epilepsy recurred and the cat presented two focal and one generalised seizures over 24 h. MRI was performed again and showed the same lesions as described 3 weeks after surgery ([Figure 3](#fig3-2055116915593970){ref-type="fig"}). Phenobarbital treatment was restarted at the initial dose (2.5 mg/kg q12h) ([Figure 4](#fig4-2055116915593970){ref-type="fig"}). {#fig4-2055116915593970} Discussion {#section3-2055116915593970} ========== Differential diagnoses for intracranial cystic lesions are extensive but histological analysis of the lesions can give an accurate diagnosis. In our case, no biopsy was performed perioperatively in order to avoid leaking at the catheter site. The lesions considered in the differential diagnoses list in cats, according to veterinary literature, are parasite cysts, cystic neoplasms, abscesses, cysts with a non-neural epithelial lining or ex vacuo-type cysts.^[@bibr7-2055116915593970],[@bibr11-2055116915593970][@bibr12-2055116915593970][@bibr13-2055116915593970][@bibr14-2055116915593970][@bibr15-2055116915593970][@bibr16-2055116915593970][@bibr17-2055116915593970]--[@bibr18-2055116915593970]^ The lack of contrast enhancement, extra-axial position and homogenous CSF-like aspect of this cyst makes all other possibilities than an IAC less likely. To our knowledge, this is the first case of a supratentorial arachnoid cyst reported in a cat. Only three cases of feline IAC have been described in the veterinary literature. The cats were three Persian-breed males ranging in age from 10 weeks to 4 years. The cysts were diagnosed with MRI in two cases or during necropsy in one case. The cyst was in the caudal fossa at the level of the quadrigeminal cistern in all cases,^[@bibr4-2055116915593970],[@bibr8-2055116915593970],[@bibr9-2055116915593970]^ and in one case it extended rostrally to the cerebral hemisphere.^[@bibr8-2055116915593970]^ In dogs, males and small breeds are over-represented.^[@bibr2-2055116915593970],[@bibr4-2055116915593970][@bibr5-2055116915593970][@bibr6-2055116915593970]--[@bibr7-2055116915593970],[@bibr11-2055116915593970]^ Canine IAC tends to occur mainly in the quadrigeminal cistern, with other infratentorial locations sporadically reported (eg, cerebellomedullary angle).^[@bibr7-2055116915593970]^ In humans, IAC are mainly supratentorial, with 50--60% of them occurring in the middle cranial fossa.^[@bibr3-2055116915593970],[@bibr19-2055116915593970],[@bibr20-2055116915593970]^ They occur less often in the caudal fossa (10%), wherein they occur most frequently in the cerebellopontine angle and not the quadrigeminal cistern. In our case, the cystic lesion was mainly in the forebrain region, spreading slightly to the brain stem, and it looked more similar to human than canine IAC. IAC is often an incidental finding in humans, and it has recently been suggested that this may also be true for dogs;^[@bibr2-2055116915593970],[@bibr5-2055116915593970],[@bibr9-2055116915593970],[@bibr10-2055116915593970],[@bibr21-2055116915593970]^ however, numerous canine symptomatic cases have been reported.^[@bibr2-2055116915593970],[@bibr5-2055116915593970][@bibr6-2055116915593970]--[@bibr7-2055116915593970],[@bibr10-2055116915593970],[@bibr21-2055116915593970]^ Matiasek et al evaluated brain compression in 28 dogs with quadrigeminal cysts and concluded that occipital lobe compression of \>14% was always associated with clinical signs (focal seizures, generalised seizures or both).^[@bibr10-2055116915593970]^ No such statistics exist in cats because only three cases of quadrigeminal cysts have been reported, and their clinical significance is difficult to assess. The first case was a 1-year-old cat with a history of uncontrollable seizures.^[@bibr9-2055116915593970]^ Anticonvulsant therapy (diazepam and phenobarbital) stopped the seizures for 5 months, but the cat relapsed and was euthanased. The second reported case was a 4-year-old cat with a 12 month history of episodic collapse and altered mentation that was presented after a severe episode of non-ambulatory tetraparesis and cranial nerve abnormalities.^[@bibr8-2055116915593970]^ It was treated by CPS placement and had not relapsed 10 months after surgery. The third case was a 10-week-old kitten that died after a 24-h period of lethargy with decreasing levels of consciousness.^[@bibr4-2055116915593970]^ In our case, the clinical significance of the cyst, located in the left temporal area, is difficult to assess. The type of seizures (complex partial seizures with predominant orofacial motor signs) is similar to those observed with hippocampal necrosis. Forebrain compression calculated with a method previously described by Matiasek et al was 41.2%.^[@bibr10-2055116915593970]^ This value is given on an indicative basis and cannot be compared to the percentages found in the study by Matiasek et al (12-- 23%) because the location of the cyst was not the same.^[@bibr10-2055116915593970]^ Nonetheless, because of this value, the cyst likely had clinical significance. Recurrence of seizures after resolution of the cyst can be explained by the persistence of lesions in the temporal lobe (atrophy and diffuse hyperintensity in the hippocampal structures in T2-weighted images) ([Figure 3](#fig3-2055116915593970){ref-type="fig"}). Management of IAC in dogs is usually achieved by either cyst fenestration or, more commonly, by CPS placement.^[@bibr2-2055116915593970],[@bibr5-2055116915593970],[@bibr7-2055116915593970],[@bibr22-2055116915593970],[@bibr23-2055116915593970]^ In humans, it is reported that CPS reduces the cyst permanently, while fenestration may require secondary intervention.^[@bibr24-2055116915593970][@bibr25-2055116915593970]--[@bibr26-2055116915593970]^ Arachnoid cysts do not have an epithelial lining and should be described as pseudocysts or diverticula, but the inner membrane is believed to have secretory activity.^[@bibr11-2055116915593970],[@bibr23-2055116915593970],[@bibr27-2055116915593970]^ In veterinary medicine, both fenestration and shunting are described as treatments, but no study reports the therapeutic superiority of any of them. In our case, CPS placement resulted in no complications during surgery and lead to total resolution of the cyst 21 months after surgery (MRI assessment). Shunting was chosen because it is considered as an efficient and low-morbidity method, and is reported in the literature more frequently.^[@bibr2-2055116915593970],[@bibr8-2055116915593970],[@bibr22-2055116915593970],[@bibr23-2055116915593970]^ Although this technique does not allow histological analysis of the cyst, to avoid catheter leaking, it is considered a less invasive method than fenestration. Conclusions {#section4-2055116915593970} =========== To our knowledge, this is the first reported case of a supratentorial arachnoid cyst in a cat, as all other IACs mentioned in the literature were quadrigeminal cysts.^[@bibr4-2055116915593970],[@bibr8-2055116915593970],[@bibr9-2055116915593970],[@bibr22-2055116915593970],[@bibr23-2055116915593970]^ The location of the cyst in the temporal lobe is unusual in cats and dogs but is consistent with the predominant location in humans.^[@bibr19-2055116915593970],[@bibr20-2055116915593970],[@bibr24-2055116915593970][@bibr25-2055116915593970][@bibr26-2055116915593970]--[@bibr27-2055116915593970]^ The semiology of the seizures suggests that the cyst had clinical significance, even if this cannot be definitively proven. This case describes an IAC that was successfully managed by CPS placement without complications. Shunting could be considered a reference method in IAC surgical therapy but too few cases are reported in veterinary medicine to assess the therapeutic superiority of either CPS or fenestration. **Funding:** The authors received no specific grant from any funding agency in the public, commercial or not-for-profit sectors for the preparation of this case report. **Conflict of interest:** The authors do not have any potential conflicts of interest to declare.

\section{Introduction}\label{sec:introduction} The dynamics of neutrino oscillations in the early universe has been extensively studied in the scientific literature and is expected to have produced an efficient mixing of flavor neutrino distributions at a temperature $T_\gamma \sim 1$ MeV, thus shortly before the freezing of the neutron-to-proton density ratio and the onset of Big Bang Nucleosynthesis (BBN). Indeed, while flavor neutrino conversions are suppressed by matter effects at larger temperatures, at $T_\gamma \sim 10$ MeV the atmospheric mass-squared difference $\Delta m^2_{\rm atm}$ triggers efficient $\nu_\mu-\nu_\tau$ mixing, as well as $\nu_x-\nu_e$ ($x=\mu,\tau$) conversions if the angle $\theta_{13}$ is not vanishing and sufficiently large. Finally, at $T_\gamma\lesssim 3$ MeV oscillations driven by $\Delta m^2_{\rm sol}$ set on and are large enough to achieve strong flavor conversions before BBN \cite{Lunardini:2000fy,Dolgov:2002ab,Wong:2002fa,Abazajian:2002qx}. These results have a major impact on the possible values for the cosmological lepton asymmetry stored in each neutrino flavor, which in analogy with the baryon-antibaryon asymmetry parameter $\eta_b=(n_b-n_{\bar{b}})/n_\gamma$, can be parameterized by the number density ratios \begin{equation} \eta_{\nu_\alpha} = \frac{n_{\nu_\alpha}-n_{\bar{\nu}_\alpha}}{n_\gamma}\vv \,\,\, \alpha=e,\mu,\tau \pp \label{etanualpha} \end{equation} Based on the equilibration of lepton and baryon asymmetries by sphalerons in the very early universe, such a neutrino asymmetry should be of the same order of the cosmological baryon number $\eta_b=273.93 \,\Omega_b h^2\, 10^{-10}$, which is restricted to be a few times $10^{-10}$ by present observations, such as 7-year data from the WMAP satellite and other cosmological measurements \cite{Komatsu:2010fb}. Nevertheless, a cosmological neutrino asymmetry orders of magnitude larger than this value is still an open possibility, with implications on fundamental physics in the early universe, such as their potential relation with the cosmological magnetic fields at large scales \cite{Semikoz:2009ye}. In particular, a non-zero lepton asymmetry leads to an enhanced contribution of neutrinos to the energy density in the form of radiation $\rho_r$, which, after the $e^+e^-$ annihilation phase, is usually parameterized as $\rho_r/\rho_\gamma=1 + 7/8 (4/11)^{4/3} N_{\rm eff}$. The parameter $N_{\rm{eff}}$ is the ``effective number of neutrinos" whose standard value is 3 in the limit of instantaneous neutrino decoupling. Interestingly, recent data on the anisotropies of the cosmic microwave background from WMAP \cite{Komatsu:2010fb} and the primordial $^4$He abundance \cite{Izotov:2010ca,Aver:2010wq,Aver:2010wd} seem to favor a value of $N_{\rm{eff}}>3$, although with large errorbars. If neutrinos are in kinetic and chemical equilibrium, their distribution of momenta is parameterized by a temperature and a well defined chemical potential $\mu_{\nu_\alpha}$, and each flavor neutrino asymmetry can be expressed in terms of the corresponding degeneracy parameter $\xi_{\nu_\alpha}\equiv \mu_{\nu_\alpha}/T_{\nu_\alpha}$ as \begin{equation} \eta_{\nu_\alpha} = \frac{1}{12 \zeta(3)} \left(\frac{T_{\nu_\alpha}}{T_\gamma}\right)^3 \left( \pi^2 \xi_{\nu_\alpha}+\xi_{\nu_\alpha}^3 \right) \pp \label{eta-eq} \end{equation} In this case, if flavor oscillations enforce the condition that all $\xi_{\nu_\alpha}$ are almost the same during BBN, the stringent bound on the electron neutrino degeneracy, which directly enters the neutron/proton chemical equilibrium \cite{KangSteigman}, applies to all flavors. The common value of the neutrino degeneracies is restricted to the range $-0.021 \leq \xi_{\nu_\alpha} \leq 0.005$, $ \alpha =e, \mu, \tau$ \cite{Iocco:2008va} (see also \cite{Hansen:2001hi,Barger:2003zg,Barger:2003rt,Cuoco:2003cu,Cyburt:2004yc, Serpico:2005bc,Simha:2008mt,Krauss:2010xg} for other analyses). This in turn also implies that, \emph{if neutrinos indeed, reach perfect kinetic and chemical equilibrium before they decouple}, any large excess in cosmic radiation density, if observed, must be ascribed to extra relativistic degrees of freedom since the additional contribution to radiation density due to non vanishing $\xi_{\nu_\alpha}$ is very small. Assuming that the neutrino distributions are given by their equilibrium form, the BBN bound on the neutrino degeneracies leads to the following upper limit on the excess contribution to $N_{\rm{eff}}$ \cite{Dolgov:2002ab}, \begin{equation} \Delta N_{\rm{eff}} = \sum_{\alpha=e,\mu,\tau} \left[\frac{30}{7}\left( \frac{\xi_{\nu_\alpha}}{\pi} \right)^2 + \frac{15}{7}\left( \frac{\xi_{\nu_\alpha}}{\pi} \right)^4 \right] \lesssim 0.0006 \vv \label{deltan} \end{equation} which is tiny, even compared with the value of $N_{\rm eff}=3.046$ found solving the neutrino kinetic equations in absence of asymmetries \cite{Mangano:2005cc}. Neutrino kinetic and chemical equilibrium is maintained in the early universe by purely leptonic weak processes such as neutrino-neutrino interactions, $\nu$-$e^\pm$ scatterings and pair processes, $\nu \bar{\nu}\leftrightarrow e^+e^-$, whose rates become of the order of the Hubble parameter at $T_\gamma \sim 2-3$ MeV. Baryons play no role in this concern due to their very low density. At lower temperatures weak interactions are no longer effective, neutrinos decouple from the rest of the primeval plasma and their distribution in phase space is frozen out. Flavor oscillations driven by $\Delta m^2_{\rm atm}$ take place when neutrinos are still fastly scattering off the surrounding medium, so that the changes in their distribution due to oscillations are efficiently readjusted into an equilibrium Fermi-Dirac function. Instead, flavor conversions due to $\Delta m^2_{\rm sol}$ and $\theta_{12}$ occur around neutrino decoupling. This implies, at least in principle, that if neutrinos succeed in achieving comparable asymmetries in all flavors before BBN, their distributions might acquire distortions with respect to equilibrium values due to inefficient interactions. This can be easily understood by a simple example. Suppose that at temperatures higher than $2-3$ MeV we start with a vanishing total asymmetry, but $\eta_{\nu_e}^{\rm in}=- 2 \eta_{\nu_x}^{\rm in} \neq 0$ and we artificially switch-off scattering and pair processes. Due to solar-scale oscillations, in the case of maximal mixing asymmetries in each flavor will eventually vanish, but the neutrino distributions will not correspond to equilibrium, since averaging two equilibrium distributions with a different chemical potential does not correspond to a Fermi-Dirac function. Only scatterings and pair-processes can turn it into an equilibrium distribution with, in this case, zero chemical potential. Though this example is quite extreme and unrealistic, nevertheless, it tells us that the interplay of neutrino freeze--out and $\Delta m^2_{\rm sol}$ oscillation phases might deserve a more careful scrutiny, as first discussed in \cite{Pastor:2008ti}. In fact, depending on the initial flavor neutrino asymmetries and the value of $\theta_{13}$, the final neutrino distributions at the onset of BBN might show non-thermal distortions which change the neutron-proton chemical equilibrium due to the direct role played by electron (anti)neutrinos. Moreover, this corresponds to an asymmetry-depending parameter $N_{\rm{eff}}>3$ which is due to inefficient entropy transfer to the electromagnetic plasma and is not given by the equilibrium value of eq.\ \eqref{deltan}. As pointed out in \cite{Pastor:2008ti} these features will prove to be quite important in case of large initial asymmetries and opposite values for $\nu_e$ and $\nu_{\mu,\tau}$ chemical potentials. It was shown that with fine-tuned initial asymmetries the BBN bound could be respected and at the same time an excess radiation density could survive, corresponding to values of $\Delta N_{\rm{eff}}$ of order unity or larger. In the present work we extend the analysis of \cite{Pastor:2008ti} in two ways. First, we consider a wider range of values of the initial neutrino asymmetries and solve their evolution with the corresponding kinetic equations, including both collisions and oscillations. Moreover, the obtained shape of the neutrino distributions is then plugged into the BBN dynamics allowing, by the comparison between the theoretical results and the experimental data on primordial abundances of deuterium and $^4$He, to find more accurate bounds on the total lepton asymmetry stored in the neutrino sector, as well as the way it was distributed at some early stage in the $\nu_e$ and $\nu_x$ flavors. In fact, for all initial values of neutrino asymmetries which are compatible with BBN bounds, the distortions in the neutrino distribution are typically quite small, see Section \ref{sec:dynamics}, so that it is accurate enough for our purposes to parameterize them in terms of a Fermi-Dirac function with two time-dependent parameters, which correspond to the first two moments of the actual distribution: an effective chemical potential $\xi_{\nu_\alpha}$ and an effective temperature $T_{\nu_\alpha}$, or equivalently, the asymmetry in each flavor and the energy density. The paper is organized as follows. After setting in Section \ref{sec:dynamics} the formalism of kinetic equations which rule the evolution of neutrino distributions written in the standard density matrix formalism and showing an example of their dynamics, we then study the BBN constraints on neutrino asymmetries in Section \ref{sec:BBN}. In particular, we discuss the experimental data which are used in our analysis for $^2$H/H and $^4$He mass fraction $Y_p$, as well as the way we have modified the public BBN numerical code $\mathtt{PArthENoPE}$ \cite{Pisanti:2007hk,serpico,parthenope} to track neutrino evolution. Finally, we report the bounds on initial (at $T_\gamma \sim 10$ MeV) neutrino asymmetries or their final values after flavor oscillation phase. In Section \ref{sec:conclusions} we give our concluding remarks. \section{The dynamics of neutrinos in the early universe with primordial asymmetries}\label{sec:dynamics} Our first aim is to calculate the evolution of the three active neutrino distributions in the epoch of the universe right before BBN, when these particles were interacting among themselves and with electrons and positrons. The corresponding weak collision rate decreases very fast with the expansion until neutrinos decouple at $T_\gamma \sim 1$ MeV. At the same time, it was shown in \cite{Lunardini:2000fy,Dolgov:2002ab,Wong:2002fa,Abazajian:2002qx,Gava:2010kz} that for the neutrino mixing parameters currently allowed, flavor oscillations start to be effective at similar temperatures. In such a case the best way to describe neutrino distributions is to use matrices in flavor space \cite{Sigl:1993fn, McKellar:1994ja}. Since we consider only the three active neutrino species, we will need $3\times3$ matrices in flavor space $\varrho_{\bf p}$ for each neutrino momentum ${\bf p}$, where the diagonal elements are the usual flavor distribution functions (occupation numbers) and the off-diagonal ones encode phase information and vanish for zero mixing. Oscillations in flavor space of the three active neutrinos are driven by two mass-squared differences and three mixing angles bounded by the experimental observations in the following ranges: the "solar" $\Delta m^2_{\rm sol}=7.59^{+0.44}_{-0.37} \times 10^{-5}$ eV$^2$, the "atmospheric" $|\Delta m^2_{\rm atm}|= 2.40^{+0.24}_{-0.22} \times 10^{-3}$ eV$^2$ and correspondingly, the large mixing angles $\sin^2 \theta_{12}= 0.32^{+0.04}_{-0.03}$ and $\sin^2 \theta_{23}= 0.50^{+0.13}_{-0.11}$ ($2\sigma$ ranges from \cite{Schwetz:2008er}). On the other hand, the third angle is quite small, $\sin^2 \theta_{13} \leq 0.053$ ($3\sigma)$ \cite{Schwetz:2008er}, even compatible with a vanishing value, though a mild evidence for $\sin^2 \theta_{13}>0$ has been found in global data analyses (see e.g.\ \cite{Fogli:2008jx,GonzalezGarcia:2010er}). The equations of motion (EOMs) for $\varrho_{\bf p}$ are the same as those considered in reference \cite{Pastor:2008ti}, \begin{equation} {\rm i}\,\frac{d\varrho_{\bf p}}{dt} =[{\sf\Omega}_{\bf p},\varrho_{\bf p}]+ C[\varrho_{\bf p},\bar\varrho_{\bf p}]\,, \label{drhodt} \end{equation} and similar for the antineutrino matrices $\bar\varrho_{\bf p}$. The first term on the r.h.s.\ describes flavor oscillations, \begin{equation} {\sf\Omega}_{\bf p}=\frac{{\sf M}^2}{2p}+ \sqrt{2}\,G_{\rm F}\left(-\frac{8p}{3 m_{\rm w}^2}\,{\sf E} +\varrho-\bar\varrho\right)\,, \end{equation} where $p=|{\bf p}|$ and ${\sf M}$ is the neutrino mass matrix (opposite sign for antineutrinos). Matter effects are included via the term proportional to the Fermi constant $G_{\rm F}$, where ${\sf E}$ is the $3\times3$ flavor matrix of charged-lepton energy densities \cite{Sigl:1993fn}. For our range of temperatures we only need to include the contribution of electrons and positrons. Finally, the last term arises from neutrino-neutrino interactions and is proportional to $\varrho-\bar\varrho$, where $\varrho=\int \varrho_{\bf p}\,{\rm d}^3{\bf p}/(2\pi)^3$ and similar for antineutrinos. For the relevant values of neutrino asymmetries this matter term dominates and leads to synchronized oscillations~\cite{Dolgov:2002ab, Wong:2002fa, Abazajian:2002qx}. The last term in eq.\ \eqref{drhodt} corresponds to the effect of neutrino collisions, i.e.\ interactions with exchange of momenta. Here we follow the same considerations of ref.\ \cite{Pastor:2008ti}, where the reader can find more details on the approximations made and related references. In short, the collision terms for the off-diagonal components of $\varrho_{\bf p}$ in the weak-interaction basis are momentum-dependent damping factors, while collisions and pair processes for the diagonal $\varrho_{\bf p}$ elements are implemented without approximations solving numerically the collision integrals as in \cite{Mangano:2005cc}. These last terms are crucial for modifying the neutrino distributions to achieve equilibrium with $e^\pm$ and, indirectly, with photons. We have solved numerically the EOMs for the matrices in flavor space of neutrinos and antineutrinos with non-zero initial asymmetries. The expansion of the universe is taken into account using comoving variables as in \cite{Dolgov:2002ab}, where it was shown that flavor oscillations between muon and tau neutrinos take place at $T_\gamma > 10$ MeV, when interactions are very effective. For any initial values of the muon or tau neutrino asymmetries, the combined effect of oscillations and collisions is able to equilibrate the two flavors and hence leads to $\eta_{\nu_\mu}^{\rm in}=\eta_{\nu_\tau}^{\rm in}\equiv\eta_{\nu_x}^{\rm in}$. Therefore, we start our numerical calculations for each case at $T=10$ MeV and initial parameters $\eta_{\nu_e}^{\rm in}$ and the total asymmetry $\eta_\nu=\eta_{\nu_e}^{\rm in}+2\,\eta_{\nu_x}^{\rm in}$. Note that hereafter by $\eta_\nu$ we denote the initial value of the total asymmetry, which is kept constant until the onset of $e^+-e^-$ annihilations, when it is diluted by the increase of the photon density (as in the case of the baryon asymmetry $\eta_b$). All neutrino mixing parameters, except for $\theta_{13}$, are taken as the best-fit values in \cite{Schwetz:2008er}. Modifying these parameters within their allowed regions does not affect our results. Instead, the value of $\theta_{13}$ plays an important role in the evolution of the neutrino asymmetries \cite{Pastor:2008ti}. We thus consider either $\theta_{13}=0$ or $\sin^2 \theta_{13}=0.04$, a value close to the upper bound from neutrino experiments \cite{Schwetz:2008er,Fogli:2008jx,GonzalezGarcia:2010er}. \begin{figure} \begin{center} \includegraphics[width=0.8\textwidth,angle=0]{evol_lnu.pdf} \end{center} \caption{\label{evol_lnu} Evolution of the flavor neutrino asymmetries when $\eta_\nu=-0.41$, and $\eta_{\nu_e}^{\rm in}=0.82$. The solid curves correspond to vanishing $\theta_{13}$ (outer black lines) and $\sin^2\theta_{13}=0.04$ (inner red lines). The total neutrino asymmetry is constant and equal to three times the value shown (blue dotted line).} \end{figure} Let us describe the evolution of the flavor asymmetries with a specific example. As shown in \cite{Pastor:2008ti}, large initial values of flavor primordial asymmetries could satisfy the BBN bound only if $\eta_{\nu_e}^{\rm in}$ and $\eta_{\nu_x}^{\rm in}$ have opposite signs and the total asymmetry is not very different from zero. A case with $\eta_\nu=0$ was the main example presented in \cite{Pastor:2008ti}. Here we choose instead a benchmark case with non-zero negative total asymmetry, $\eta_\nu=-0.41$, and $\eta_{\nu_e}^{\rm in}=0.82$. The evolution of the flavor asymmetries is found from the numerical solution of the EOMs and is shown in Figure \ref{evol_lnu}. Other choices of the initial asymmetries will lead to different final values, but the overall behavior of the evolution is similar to the case shown here. One can see in Figure \ref{evol_lnu} the effect of flavor oscillations on the evolution of neutrino asymmetries with the universe temperature and the dependence on the value of $\theta_{13}$. If this mixing angle is close to the present experimental upper bound, flavor oscillations are effective around $T_\gamma\sim 8$ MeV (inner red lines) when neutrinos are still in good thermal contact with the ambient plasma.\footnote{Here we consider only the case of normal hierarchy, $\Delta m^2_{\rm atm}>0$. If we choose an inverted hierarchy the results are very similar, except that equilibrium among the flavor asymmetries is reached slightly earlier.} The neutrino distributions evolve keeping an equilibrium form, and the total asymmetry is almost equally distributed among the three flavors. However, the final value of the electron neutrino asymmetry is too different from zero and this case is not allowed by BBN, as we will see in Section \ref{sec:BBN}. Instead, for $\theta_{13}=0$ flavor oscillations begin only at $T_\gamma\lesssim 3$ MeV, when weak interactions are not frequent enough to keep the neutrino spectra in equilibrium. There is no full equipartition of the total $\eta_\nu$ among the three flavors, although collisions lead to final values of the flavor asymmetries closer to $\eta_\nu/3$ than those expected in an MSW matter neutrino conversion. In this case the final value of the electron neutrino asymmetry is close enough to zero at the onset of BBN to lie in the favored region. If $\sin^2\theta_{13}\lesssim 10^{-3}$, the outcome is very close to the case of vanishing $\theta_{13}$. In Figure \ref{drhofinal} we show the final energy spectra of relic electron neutrinos and antineutrinos in arbitrary units for the same case of Figure \ref{evol_lnu} with vanishing $\theta_{13}$. The upper (lower) solid line stands for the spectra of electron neutrinos (antineutrinos) calculated numerically, while the corresponding dotted lines are described by a Fermi/Dirac distribution just characterized by the same effective value of the electron neutrino degeneracy parameter (as used in all analyses before \cite{Pastor:2008ti}). Both cases lead to the same value of the electron neutrino asymmetry but the real calculation shows that an excess of radiation in neutrinos remains. \begin{figure} \begin{center} \includegraphics[width=0.8\textwidth,angle=0]{drhofinal.pdf} \end{center} \caption{\label{drhofinal} The final energy spectra of relic electron neutrinos and antineutrinos in arbitrary units for the same case of Figure \ref{evol_lnu} with vanishing $\theta_{13}$. Upper (lower) solid line stands for electron neutrino (antineutrino) calculated numerically (label ''real''). Upper (lower) dotted line stands for electron neutrino (antineutrino) described by a Fermi/Dirac distribution just characterized by the same effective value of the electron neutrino degeneracy parameter.} \end{figure} \begin{figure} \begin{center} \includegraphics[width=0.8\textwidth,angle=0]{evol_neff.pdf} \end{center} \caption{\label{evol_Neff} Evolution of the neutrino energy density for the same case as in Figure \ref{evol_lnu}. The vertical axis is marked with $N_{\rm eff}$, left before $e^+e^-$ annihilation, right afterwards. The solid curves correspond to vanishing $\theta_{13}$ (upper black line) and $\sin^2\theta_{13}=0.04$ (lower red line). The case without asymmetries is shown for comparison (blue dotted line).} \end{figure} In Figure \ref{evol_Neff} we show the evolution of the ratio of neutrino to photon energy densities, $\rho_\nu/\rho_\gamma$, properly normalized so that it corresponds to $N_{\rm eff}$ at early and late times as in \cite{Pastor:2008ti}. The fast drop of $\rho_\nu/\rho_\gamma$ at $T\sim0.2$~MeV represents photon heating by $e^+e^-$ annihilations. The case without asymmetries (dotted line) ends at late times at $N_{\rm eff}=3.046$ instead of~3 because of residual neutrino heating~\cite{Mangano:2005cc}. We also show (solid lines) the evolution for our main example, where initially $N_{\rm eff}=4.16$ for our choice of neutrino asymmetries. One can see that as soon as oscillations become effective reducing the flavor asymmetries, the excess of entropy is transferred from neutrinos to the electromagnetic plasma, cooling the former and heating the latter, but this process is only very effective for \emph{large} values of $\theta_{13}$. While the final $N_{\rm eff}$ is $3.1$ for $\sin^2 \theta_{13}=0.04$, for negligible $\theta_{13}$ a significant deviation from equilibrium survives and leads to a final enhanced value of $N_{\rm eff}=3.3$. \section{Results from BBN: constraints on total and electron neutrino asymmetries} \label{sec:BBN} As well known, BBN depends upon neutrino distribution functions in two ways. First of all, electron neutrinos and antineutrinos enter directly in the charge current weak processes which rule the neutron/proton chemical equilibrium. A change in the effective temperature of the distribution can shift the neutron/proton ratio freeze out temperature and thus modifies the primordial $^4$He abundance. Similarly, a non-zero $\nu_e-\bar{\nu}_e$ asymmetry also changes chemical equilibrium towards a larger or smaller neutron fraction for negative or positive values of $\xi_{\nu_e}$, respectively. Furthermore, lepton asymmetries in all flavors translate into a positive extra contribution to the neutrino energy density, speeding up the expansion rate given by the Hubble parameter. \begin{figure} \begin{center} \includegraphics[width=0.6\textwidth,angle=0]{evol_Txie_last.pdf} \end{center} \caption{\label{evol_Txie} Evolution of the effective comoving temperatures and degeneracy parameters of electron (solid lines) and muon or tau (dashed lines) neutrinos for the same case as in Figure \ref{evol_lnu}. Both the case of vanishing $\theta_{13}$ (thick black lines) and $\sin^2\theta_{13}=0.04$ (thin red lines) are shown. The effective temperature for the case without asymmetries is shown in the upper panel for comparison (blue dotted line).} \end{figure} Differently than in the approximated standard treatments, where both neutrino asymmetries and the extra contribution to $N_{\rm{eff}}$ due to $\xi_{\nu_\alpha}$, see eq. \eqref{deltan}, are considered as constant parameters, in the present analysis we exactly follow the evolution of the neutrino distribution versus the photon temperature $T_\gamma$, which is our evolution parameter.\footnote{For another example of BBN calculations with arbitrarily-specified, time-dependent neutrino and antineutrino distribution functions, see Ref.\ \cite{Smith:2008ic}.} To this end we have changed the public numerical code $\mathtt{PArthENoPE}$ \cite{Pisanti:2007hk,parthenope} as follows. For any given initial values (at $T_\gamma= 10$ MeV) for the total neutrino asymmetry $\eta_\nu=\sum_\alpha \eta_{\nu_\alpha}$, unchanged by flavor oscillations, and electron neutrino asymmetry $\eta_{\nu_e}^{\rm in}$ we obtain, as described in the previous section, the time dependent neutrino distributions. The latter are then fitted in terms of Fermi-Dirac functions with the two evolving parameters $T_{\nu_\alpha}(T_\gamma)$ and $\xi_{\nu_{\alpha}}(T_\gamma)$. An example of their evolution is shown in Figure \ref{evol_Txie} for the same choice of initial asymmetries as in the case described in Section \ref{sec:dynamics}. Weak rates are then averaged over the corresponding electron (anti)neutrino distribution. The Hubble parameter is also modified to account for the actual evolution of total neutrino energy density. The final abundances of both the ratio $^2$H/H and the $^4$He mass fraction, $Y_p$, are numerically computed as a function of the input parameters $\eta_\nu$ and $\eta_{\nu_e}^{\rm in}$ and compared with the corresponding experimental determinations. The baryon density parameter has been set to the value determined by the 7-year WMAP result, $\Omega_b h^2 = 0.02260 \pm 0.00053$ (68\% C.L.) \cite{Komatsu:2010fb}.\footnote{The allowed region of $\Omega_b$ in extended cosmological models with free $N_{\rm eff}$ does not differ significantly.} To get confidence intervals for $\eta_\nu$ and $\eta_{\nu_e}^{\rm in}$, one can construct the likelihood function \begin{equation} {\mathcal{L}}(\eta_\nu, \eta_{\nu_e}^{\rm in})\propto \exp\left( -\chi^2(\eta_\nu, \eta_{\nu_e}^{\rm in})/2\right) \vv \end{equation} with \begin{equation} \chi^2(\eta_\nu, \eta_{\nu_e}^{\rm in}) = \sum_{ij} [ X_i(\eta_\nu, \eta_{\nu_e}^{\rm in}) - X_i^{obs} ] W_{ij}(\eta_\nu, \eta_{\nu_e}^{\rm in}) [ X_j(\eta_\nu, \eta_{\nu_e}^{\rm in}) - X_j^{obs} ] \pp \label{eq:chi2} \end{equation} The proportionality constant can be obtained by requiring normalization to unity, and $W_{ij}(\eta_\nu, \eta_{\nu_e}^{\rm in})$ denotes the inverse covariance matrix \cite{serpico}, \begin{equation} W_{ij}(\eta_\nu, \eta_{\nu_e}^{\rm in}) = [ \sigma_{ij}^2 + \sigma_{i,exp}^2 \delta_{ij} + \sigma_{ij,other}^2 ]^{-1} \vv \end{equation} where $\sigma_{ij}$ and $\sigma_{i,exp}$ represent the nuclear rate uncertainties and experimental uncertainties of nuclide abundance $X_i$, respectively \cite{serpico}, while by $\sigma_{ij,other}^2$ we denote the propagated squared error matrix due to all other input parameter uncertainties ($\tau_n$, $G_{\rm N}$, $\Omega_b h^2$, \ldots). In our case we consider in eq.\ \eqref{eq:chi2} as $X_i$ the quantities $^2$H/H and $Y_p$ only. Let us now briefly discuss the set of data we have used in our study. The $^2$H/H number density is obtained by averaging seven determinations obtained in different Quasar Absorption Systems, as in \cite{Iocco:2008va} \begin{equation} ^2{\rm H/H}= (2.87 \pm 0.22) \times 10^{-5} \vv \label{dataD} \end{equation} where the quadratic error has been enlarged by the value of the reduced $\chi^2$ to account for the dispersion of measurements for this dataset, $\chi^2_{\rm min}/6=3.6$, see \cite{Iocco:2008va}. For the $^4$He mass fraction we consider two different determinations. One is the result of the data collection analysis performed in \cite{Iocco:2008va}, \begin{equation} Y_p=0.250 \pm 0.003\, . \label{he1} \end{equation} More recently, new studies of metal poor H II regions have appeared in the literature \cite{Izotov:2010ca,Aver:2010wq,Aver:2010wd}. While these groups both agree on a larger central value with respect to the result of eq. \eqref{he1}, which incidentally seems to pin down a value for $N_{\rm{eff}}>3$ at $2\sigma$, a different estimate of possible systematic effects which dominate the total uncertainty budget is quoted in \cite{Izotov:2010ca} and \cite{Aver:2010wq}, with \cite{Aver:2010wq} quoting a larger error, of the order of 4\%. \begin{eqnarray} Y_p & = & 0.2565 \pm 0.0010 {\rm (stat.)} \pm 0.0050 {\rm (syst.)} \,\,\,\,\, \mbox{\cite{Izotov:2010ca}} \vv \\ Y_p & = & 0.2561 \pm 0.0108 \,\,\,\,\, \mbox{\cite{Aver:2010wq}} \pp \end{eqnarray} Finally, in a recent paper \cite{Aver:2010wd}, Markov Chain Monte Carlo method was exploited to determine the $^4$He abundance, and the uncertainties derived from observations of metal poor nebulae finding \begin{equation} Y_p = 0.2573 \pm 0.0033 \, . \label{he2} \end{equation} In the following, we will use the two results of eq.s \eqref{he1} and \eqref{he2}. While their uncertainties are the same, they differ for the central value, actually the smaller and higher of all results reported above, a fact which will produce two different bounds on the electron neutrino asymmetry. \begin{figure}[t] \begin{center} \includegraphics[width=0.49\textwidth,angle=0]{fig1anew.pdf} \includegraphics[width=0.49\textwidth,angle=0]{fig2anew.pdf} \end{center} \caption{\label{bounds_eta} The 95\% C.L.\ contours from our BBN analysis in the $\eta_\nu-\eta_{\nu_e}^{\rm in}$ plane for $\theta_{13}=0$ (left) and $\sin^2 \theta_{13}=0.04$ (right). The two contours correspond to the different choices for the primordial $^4$He abundances of eqs.\ \eqref{he1} (blue) and \eqref{he2} (purple). The (red) dot-dashed line is the set of values of $\eta_\nu$ and $\eta_{\nu_e}^{\rm in}$ which, due to flavor oscillations, evolve towards a vanishing final value of electron neutrino asymmetry $\eta_{\nu_e}^{\rm fin}$. We also report as dashed lines the iso-contours for different values of $N_{\rm eff}$, the effective number of neutrinos after $e^+e^-$ annihilation stage.} \end{figure} \begin{figure \begin{center} \includegraphics[width=0.49\textwidth,angle=0]{He4_D.pdf} \end{center} \caption{\label{He4_D} Bounds in the $\eta_{\nu_e}^{\rm in}$ - $\eta_\nu$ plane for each nuclear yield. Areas between the lines correspond to 95\% C.L. regions singled out by the $^4$He mass fraction (solid lines) of eq.\ \eqref{he1}, and Deuterium (dashed lines) as in eq.\ \eqref{dataD}. } \end{figure} \begin{figure}[t] \begin{center} \includegraphics[width=0.49\textwidth,angle=0]{plot_in_xis_1.pdf} \includegraphics[width=0.49\textwidth,angle=0]{plot_in_xis_2.pdf} \end{center} \caption{\label{bounds_xi} Same results as in Figure \ref{bounds_eta} in the plane of initial flavor degeneracy parameters $\xi_{\nu_x}^{\rm in}$ and $\xi_{\nu_e}^{\rm in}$.} \end{figure} \begin{figure \begin{center} \includegraphics[width=0.49\textwidth,angle=0]{fig2l.pdf \includegraphics[width=0.49\textwidth,angle=0]{fig2r.pdf} \end{center} \caption{\label{bounds_diff} Same results as in Figure \ref{bounds_eta} in the plane of $\eta_{\nu_e}^{\rm fin}$ and the difference $\eta_{\nu_x}^{\rm fin}-\eta_{\nu_x}^{\rm fin}$, where the superscript indicates that the asymmetries are evaluated at the onset of BBN, $T_\gamma \sim 1$ MeV, after flavor oscillations shuffled the initial $\eta_{\nu_\alpha}^{\rm in}$.} \end{figure} The 95\% C.L.\ contours for the total asymmetry $\eta_\nu$ and the initial value of the electron neutrino parameter $\eta_{\nu_e}^{\rm in}$ are shown in Figure \ref{bounds_eta} for the adopted determinations of $^2$H and $^4$He and for two different choices of $\theta_{13}$. In both cases the contours are close to and aligned along the red dot-dashed line which represents the set of initial values for the asymmetries which eventually evolve toward a vanishing final electron neutrino asymmetry, $\eta_{\nu_e}^{\rm fin}\simeq0$, which is preferred by $^4$He data. We recall that $^4$He is strongly changed if neutron/proton chemical equilibrium is shifted by a large value of $\nu_e-\bar{\nu}_e$ asymmetry around the freezing of weak rates ($T_\gamma \sim 0.8$ MeV). For large $\theta_{13}$, oscillations efficiently mix all neutrino flavors and at BBN $\eta_{\nu_\alpha} \sim \eta_\nu/3$, so the bound on $\eta_\nu$ is quite stringent, $-0.1 \lesssim \eta_\nu \lesssim 0.1$, if we adopt the value of eq.\ \eqref{he1} for $Y_p$. Instead, for the choice of eq.\ \eqref{he2} we find $-0.20 \lesssim \eta_\nu \lesssim 0$, i.e. a larger value for $Y_p$ singles out slightly negative values for $\eta_\nu$ (and $\eta_{\nu_e}^{\rm fin}$), since the theoretical prediction for $Y_p$ grows in this case as the neutron-proton chemical equilibrium shift towards a larger neutron fraction at freeze-out . On the other hand, for a vanishing $\theta_{13}$ the contours for $\eta_\nu$ and $\eta_{\nu_e}^{\rm in}$ show a clear anticorrelation, and even values of order unity for both parameters are still compatible with BBN. The allowed regions of the total neutrino asymmetry are summarized in Table \ref{table_etanu}. We stress that for any value of $\theta_{13}$, the data on primordial deuterium, eq.\ \eqref{dataD}, is crucial for closing the allowed region that the $^4$He bound fixes along the $\eta_{\nu_e}^{\rm fin}\simeq0$ line. In fact, though $^2$H is less sensitive than $^4$He to neutrino asymmetries and effective temperature which enter the Universe expansion rate, see e.g. \cite{Iocco:2008va}, yet including it in the analysis breaks the degeneracy between $\eta_{\nu_e}^{\rm in}$ and $\eta_\nu$ which is present when only $^4$He is used. This can be read from Figure \ref{He4_D} where the 95\% C.L. in the $\eta_{\nu_e}^{\rm in}$ - $\eta_\nu$ plane are shown for $^4$He and $^2$H separately, for the case $\theta_{13}=0$. The solid lines bound the region of the plane compatible with the $^4$He measurement as in eq.\ \eqref{he1}, whereas the dashed contours correspond to Deuterium observation, see eq.\ \eqref{dataD}. The different shape of these two regions is due to the different dependence of nuclide abundances on $\eta_{\nu_e}^{\rm in}$ and $\eta_\nu$, thus their combination breaks the degeneracy and leads to a close contour as shown in Figure \ref{bounds_eta}. \begin{table} \begin{center} \begin{tabular}{l|c|c|} {} & $\theta_{13}=0$ & $\sin^2 \theta_{13}= 0.04$\\ \hline $Y_p=0.250 \pm 0.003$ & $-0.66 < \eta_\nu < 0.63$ & $-0.13 < \eta_\nu < 0.07$ \\ \hline $Y_p = 0.2573 \pm 0.0033$ & $-0.71 < \eta_\nu < 0.56$ & $-0.20 < \eta_\nu < 0.02$ \\ \hline \end{tabular} \end{center} \caption{\label{table_etanu} BBN bounds on the initial total neutrino asymmetry at 95\% C.L.} \end{table} It is also interesting to report our results in terms of other variables, as in Figures \ref{bounds_xi} and \ref{bounds_diff}. In the first case, the BBN contours are shown in the plane of initial flavor degeneracy parameters $\xi_{\nu_e}^{\rm in}$ and $\xi_{\nu_x}^{\rm in}$ while, in Figure \ref{bounds_diff}, we consider a new pair of variables: the electron neutrino asymmetry at the onset of BBN $\eta_{\nu_e}^{\rm fin}$, and the difference $\eta_{\nu_x}^{\rm fin}-\eta_{\nu_e}^{\rm fin}$, which in the standard analysis is usually assumed to be vanishing. One can see from this figure that, while the (95\% C.L.) bound on $\eta_{\nu_e}^{\rm fin}$ is independent of the value of $\theta_{13}$ \begin{eqnarray} && -0.03 \leq \eta_{\nu_e}^{\rm fin} \leq 0.02\, , \,\,\,\, {\rm for}~Y_p = 0.250 \pm 0.003 \vv \\ && -0.05 \leq \eta_{\nu_e}^{\rm fin} \leq 0.003 \, , \,\,\,\, {\rm for}~Y_p = 0.2573 \pm 0.0033 \vv \end{eqnarray} the difference between the final $\nu_e$ and $\nu_x$ asymmetries strongly depends upon this yet unknown mixing angle, as expected. In fact, for large $\theta_{13}$ we recover the standard result, $\eta_{\nu_x}^{\rm fin} \sim \eta_{\nu_e}^{\rm fin}$, due to efficient mixing by oscillations and collisions, while for $\theta_{13}=0$ the two asymmetries can be different. The following ranges are in good agreement with BBN data at 95 \% C.L., \emph{independently} of the adopted value for $Y_p$ \begin{eqnarray} -0.3 \leq \eta_{\nu_x}^{\rm fin} - \eta_{\nu_e}^{\rm fin} \leq 0.3\, , && \,\,\,\, \sin^2 \theta_{13}= 0 \vv \\ -0.04 \leq \eta_{\nu_x}^{\rm fin} - \eta_{\nu_e}^{\rm fin} \leq 0.04 \, , && \,\,\,\, \sin^2 \theta_{13}= 0.04 \pp \end{eqnarray} We conclude that, in particular for $\theta_{13}= 0$ oscillations lead to quite different neutrino asymmetries in $e$ and $\mu/\tau$ flavors, still being in good agreement with BBN, differently than what was previously assumed in the literature, see \cite{Lunardini:2000fy,Dolgov:2002ab,Wong:2002fa,Abazajian:2002qx}, and \cite{Iocco:2008va,Hansen:2001hi,Barger:2003zg,Barger:2003rt,Cuoco:2003cu,Cyburt:2004yc, Serpico:2005bc,Simha:2008mt,Krauss:2010xg} for BBN analyses. In Figures \ref{bounds_eta}-\ref{bounds_diff} we also plot iso-contours for the value of the effective number of neutrinos, $N_{\rm{eff}}$, evaluated after $e^+e^-$ annihilations. For large $\theta_{13}$ BBN data bound $N_{\rm{eff}}$ to be very close to the standard value 3.046, since all asymmetries should be very small in this case and flavor oscillations modify the neutrino distributions while neutrinos are still strongly coupled to the electromagnetic bath. Therefore, we do not expect non-thermal features in the neutrino spectra in this case, since scatterings and pair processes allow for an efficient transfer of any entropy excess. On the other hand, for vanishing $\theta_{13}$, larger values of $N_{\rm{eff}}$ are still compatible with BBN data, up to values of the order of 3.4 at 95\% C.L. The dependence of the largest achievable value of $N_{\rm{eff}}$ on the value of the mixing angle $\theta_{13}$, obtained by spanning in the asymmetry parameter plane the region compatible with BBN, is reported in Figure \ref{theta13_Neff}. It is worth noticing that the final values of $N_{\rm eff}$, in particular for large final asymmetries of $\nu_x$, are also slightly larger than the $N_{\rm eff}$ that one would obtain using the equilibrium expression of eq.\ \eqref{eta-eq}. For example, if we take $\eta_{\nu_e}^{\rm fin}=0$ and $\eta_{\nu_x}^{\rm fin}=0.3$, a point on the boundary of the BBN contours (see Figure \ref{bounds_diff}) and compute the corresponding effective chemical potentials, using eq.\ \eqref{deltan} one gets $N_{\rm{eff}} = 3.2$, while the actual value is larger, $N_{\rm{eff}} = 3.4$, a signal that in this case the interplay of solar-like oscillations and neutrino freeze-out has produced indeed, a mild non-thermal distortion in neutrino distributions. \begin{figure \begin{center} \includegraphics[width=0.8\textwidth,angle=0]{fig_max_neff.pdf} \end{center} \caption{\label{theta13_Neff} Largest values of $N_{\rm{eff}}$ from primordial neutrino asymmetries compatible with BBN, as a function of $\theta_{13}$.} \end{figure} Finally, we notice that our BBN results correspond to a minimal scenario with primordial neutrino asymmetries, as we do not consider a possible extra contribution to $N_{\rm{eff}}$ coming from relativistic degrees of freedom other than standard active neutrinos. Their effect is known to produce looser bounds on neutrino asymmetries, as they speed up expansion and thus, can compensate the effect of a positive $\nu_e-\bar{\nu}_e$ asymmetry. We have explicitly checked that, for some choices of primordial asymmetries, the addition of extra radiation does not modify the evolution of flavor neutrino asymmetries. Of course, in such a case the contribution to the energy density of the additional relativistic degrees of freedom adds up to the surviving excess to $N_{\rm{eff}}$ arising from neutrino asymmetries. \section{Conclusions}\label{sec:conclusions} In this paper we have calculated the evolution of neutrinos in the early universe with initial flavor asymmetries, taking into account the combined effect of collisions and oscillations. Our numerical results for the neutrino momentum distributions, which can develop non-thermal features, were used to find the primordial production of light elements, employing a modified version of the $\mathtt{PArthENoPE}$ BBN code. Comparing with the recent data on primordial $^2$H and $^4$He abundances, we have found the allowed ranges for both the total asymmetry and the initial asymmetry in electronic flavor. These BBN limits mostly depend on the value of $\theta_{13}$, the only unknown mixing angle of neutrinos. As can be seen from Figure \ref{theta13_Neff}, for $\sin^2\theta_{13} \lesssim 10^{-3}$ this implies an effective number of neutrinos bound to $N_{\rm eff}\lesssim 3.4$, independently of which of the experimental $^4$He mass fraction of eq.s \eqref{he1} and \eqref{he2} is adopted, whereas for larger values of $\theta_{13}$ the effective number of neutrinos is closely bound to the standard value 3.046 obtained for vanishing asymmetries. In the near future it will be possible to improve our BBN constraints on the lepton number of the universe. On one hand, thanks to the better sensitivity of new neutrino experiments, either long-baseline or reactor, we expect to have a very stringent bound on $\theta_{13}$ or eventually its measurement (see e.g.\ \cite{Huber:2004ug}). Such results would lead to a more restrictive BBN analysis on primordial asymmetries. On the other hand, in the next couple of years data on the anisotropies of the cosmic microwave background from the Planck satellite \cite{Planck} will largely reduce the allowed range of $N_{\rm eff}$, since the forecast sensitivity is of the order of $0.4$ at $2\sigma$ \cite{Bowen:2001in,Bashinsky:2003tk,Hamann:2010pw}. Indeed, suppose that Planck data confirm a value of $N_{\rm eff}> 3$. An excess of radiation up to $0.4-0.5$ could imply the presence of a significative degeneracy in the neutrino sector, but this would depend on the result of $\theta_{13}$ measurement. A vanishing $\theta_{13}$ would allow such possibility, but a measured $\theta_{13}$ in the next generation of experiments {\it would imply} the presence of extra degrees of freedom others than active neutrinos. In case of a much larger result by Planck, namely $N_{\rm eff} > 4$, it would be impossible to explain such a result in terms of primordial neutrino asymmetries only, and alternative cosmological scenarios with additional relativistic species, such as sterile neutrinos (see for instance \cite{Melchiorri:2008gq}) would be strongly favored. \acknowledgments We would like to thank Georg Raffelt for his useful comments on an earlier version of this paper. G.\ Mangano, G.\ Miele, O.\ Pisanti and S.\ Sarikas acknowledge support by the {\it Istituto Nazionale di Fisica Nucleare} I.S. FA51 and the PRIN 2010 ``Fisica Astroparticellare: Neutrini ed Universo Primordiale'' of the Italian {\it Ministero dell'Istruzione, Universit\`a e Ricerca}. S.\ Pastor was supported by the Spanish grants FPA2008-00319 and Multidark CSD2009-00064 (MICINN) and PROMETEO/2009/091 (Generalitat Valenciana), and by the EC contract UNILHC PITN-GA-2009-237920. This research was also supported by a Spanish-Italian MICINN-INFN agreement, refs.\ FPA2008-03573-E and ACI2009-1051.

Although it was not known at the time, the chemokine field started with the identification of a protein by Heparin Sepharose affinity chromatography, platelet factor 4 (PF4) ([@B1]) now called CXCL4 in the systemic nomenclature introduced in 2000. As with many pathways in scientific research, our entry into the field of chemokines and glycosaminoglycans (GAGs) was a series of fortuitous co-incidences. Christine Power, working on the chemokine project in the Glaxo Institute of Molecular Biology (GIMB) under the leadership of Tim Wells, had hired a postdoctoral scientist, Arlene Hoogewerf to clone murine CCR4, and to generate the KO mouse. Arlene had previously studied the role of proteoglycans in regulating the function of other molecules and had also previously published a paper on the enzymatic activity of CTAP-III, and NAP-2, truncated forms of platelet basic protein/CXCL7, to degrade heparin. Since our group was currently working on these chemokines derived from the β-thromboglobulin precursor, and moreover we were all biochemists, our interest was piqued by the relationship between chemokines and GAGs. This interest was of course inspired by the paper recently published by Antal Rot demonstrating the evidence for haptotaxis as opposed to chemotaxis ([@B2]), earning him the title of Godfather of Chemokine--GAG biology. Arlene then followed a dual Post-doc pathway, partially hijacked by Tim to collaborate with him and his Ph.D. student, Gaby Kuschert (now Gaby Campanella) to lay much of the biochemical groundwork of chemokine--GAG interactions, and also fulfilling her goal of creating the CCR4^−/−^ mouse. They were able to show that chemokines demonstrated selectivity in their interaction with GAGs, beyond the obvious electrostatic interactions between basic and acidic molecules, and importantly made the observation that this interaction could trigger oligomerization of chemokines ([@B3], [@B4]). In addition, they were able to define the pharmacophore responsible for GAG binding of the chemokine IL-8/CXCL8 ([@B5]). My lab became more directly involved in this research direction through the serendipitous encounter of a Ph.D. student, Sarah Fritchley, working in Simi Ali's lab in Newcastle, UK, who was interested in expressing the putative GAG binding mutant of RANTES/CCL5 in *E. coli*, but who did not have a viable expression system. Sarah spent a couple of months in the lab under the tuition of our expert chemokine protein chemist, Fred Borlat, successfully producing the 40's mutant, as we colloquially called it. We published it in JBC with its correct biochemical nomenclature, ^44^AANA^47^-RANTES ([@B6]). Perhaps an omen as to the importance of this mutant for us was that it was an exception among most papers we had submitted as it was accepted overnight! We were heavily involved in screening for chemokine receptor antagonists at this time, and Marie Kosco-Vilbois had set up a simple cell recruitment assay to test putative inhibitors *in vivo --* chemokine-induced peritoneal recruitment in mice. Therefore, to investigate the effect of the abrogation of GAG binding *in vivo*, we asked her technician, Suzanne Herren, to test it for us. Being very rigorous, and accustomed to testing compounds for their ability to inhibit chemokine induced recruitment, in this instance RANTES, Suzanne tested it both for its agonist and antagonist activity. I will never forget my amazement in seeing that the mutant was not only unable to recruit cells but actually inhibited the recruitment induced by RANTES. Again serendipity stepped in. I gave a talk at a BALR meeting in the UK, and after dinner at the speaker's table, joined the youngsters at the adjacent table for a post-prandial "relaxation"... (which would never happen in this day and age). There I met Zoë Johnson, whom I learned the next day, was toying with the idea of doing a Ph.D., after having worked for a few years as an *in vivo* pharmacology laboratory assistant in *in vivo* pharmacology in industry. Since I was responsible for the Student program at the Institute, now the Serono Pharmaceutical Research Institute, she emailed me soon after expressing her interest. Tim was by now Director of the Institute, but had kept his interest in the chemokine project close to his heart, and having the flair and ability to make quick opportunistic decisions, immediately approved her appointment to investigate this phenomenon further. We then embarked on a marvelous and exciting 3 years of research during Zoë's thesis as the importance of the chemokine--GAG interaction unveiled itself. This period was also the beginning of my collaboration with Tracy Handel, one of the most enjoyable and mutually fruitful collaborations between two laboratories that I have had the pleasure to be part of, where we shared everything without any trace of competiveness, leading to several "duo" presentations of our joint discoveries at chemokine meetings. The result of these 3 years, during which we had access to other GAG binding mutants, notably those of MCP-1/CCL2 produced by Tracy's lab in Berkeley, CA, USA, and that of MIB-1β from Patti Liwang in Texas, was the demonstration that the immobilization of chemokines on GAGs was essential for their *in vivo* activity, and that moreover, certain chemokines needed to form oligomers in order to exert their property of cellular recruitment *in vivo* ([@B7]). The inter-relationship between these two properties was shown by the failure to include another chemokine--GAG mutant, that of murine MIP-1α, sent to us by Gerry Graham in Glasgow. He included of course the WT control, which in our hands was inactive in recruiting cells *in vivo*, despite his assurance that it was fully active *in vitro*, so we saw no use in testing his GAG binding mutant in our in *in vivo* assay. What he neglected to tell us was that in line with the work carried out by Lloyd Czapeklski at British Biotech some years previously the WT chemokine had been mutated to no longer oligomerize -- obviously an obligate monomer that was inactive *in vivo*, in accordance with our results with the three obligate monomers described above! We then exploited the inter-relationship of GAG binding and oligomerization using our RANTES mutant, now commonly called 004, in our lab, an abbreviation of its company nomenclature used for all biologicals, AS900004, and AANA-RANTES to our collaborators, as a very effective anti-inflammatory tool ([@B8]--[@B10]). However, the fact that it retained agonist activity precluded its development as a biological therapeutic. However, we felt that we were on the right path to discovering a novel set of molecules that would interfere with the chemokine--GAG interaction and would give us a superior niche to differentiate from our competitors who were all targeting the chemokine:receptor interaction. To achieve this, we used two approaches. The first lead by our talented head of chemistry, Matthias Schwarz, was to carry out an approach coined "SAR by NMR" to identify protein binders. Our target protein RANTES/CCL5 had the advantage that (a) it was small and therefore amenable to NMR technology, and (b) its three dimensional structure in complex with a GAG -- a disaccharide -- had been solved by our X-ray crystallographer, Jeffrey Shaw ([@B11]). The aim was to screen a small library of about 200 sulfated compounds by NMR to identify RANTES/CCL5 binders. The first screen yielded a hit, which prevented binding to heparin, and inhibited RANTES-induced peritoneal cell recruitment, despite only having micromolar affinity. The aim was to then identify a second molecule in a second round of screening, this time in the presence of the first compound, and then using the data obtained from the structures of the complexes solved by Jeff, to design a linker to form a dimer, which would have considerably higher affinity. The screening and structural biology arms worked beautifully, and the dimer was synthesized by the chemists -- but the product no longer inhibited cell recruitment *in vivo* -- and much to our chagrin and despair, it even enhanced it. However, we were still believers and decided to follow another lead. Zoë had shown that the minimal repeating unit of heparin that could inhibit RANTES-induced peritoneal recruitment was a tetrasaccharide. We therefore hired another postdoctoral scientist, India Severin, a chemist whose objective was to identify and then synthesize GAG-based mimetics. Despite a very assiduous program in collaboration with a glycobiology group in Australia, led by Deidre Coombe, we had to admit defeat. Although we identified moieties that inhibited GAG binding to RANTES as well as RANTES binding to the receptor CCR1, we did not achieve our aim of identifying a lead candidate for an anti-inflammatory program ([@B12]). To my delight, several years later, Deidre contacted me with the explanation as to why the design of our dimer resulting from our screen by NMR was incorrect. We had performed our crystallization studies at an acidic pH in order to maintain the monomeric form of RANTES, which would crystallize without aggregating. Modeling studies at physiological pH values revealed that our compounds had bound to the protein at acidic pH in a manner different from that predicted by the docking studies at physiological pH, presumably due to their different protonation states ([@B13]). And even more consoling was the publication of a GAG moiety that had anti-inflammatory properties in a model of lung inflammation by preventing T-cell recruitment ([@B14]). However, we still have a long way to go to fully understand the inter-relationship between the two interactions that chemokines have, especially *in vivo*. Chemokine biologists have always talked about gradients, but without defining whether these gradients are in the fluid phase or caused by immobilized chemokines through their interaction with GAGs. Our work showed that chemokines needed to be immobilized but did not address the question of a gradient. This has recently been beautifully demonstrated by Michael Sixt, where he visualized gradients of CCL21 leading to lymphatic vessels ([@B15]). We believed that the active form of the chemokine must be that, which is immobilized on the extracellular surface. However, our recent work at Novimmune, with Nicolas Fischer and Marie Kosco-Vilbois and another very talented Ph.D. student, Pauline Bonvin, characterizing two anti-murine CXCL10 antibodies, has led to revisiting this hypothesis. The mAb that is active in *in vivo* models of disease does not recognize GAG bound chemokine, whereas the mAb that is ineffective does, a result, which contradicts the notion that it is the GAG bound form of the chemokine that is active *in vivo* (Bonvin et al., manuscript in preparation). However, the active mAb inhibits the binding of the chemokine to GAGs, indicating that this interaction does indeed play a role, but the point of intervention appears more subtle that initially thought. Hopefully, more detailed studies of these two antibodies will provide a greater in depth understanding of the role of GAG binding in chemokine-induced cell migration *in vivo*. Conflict of Interest Statement {#S1} ============================== The author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. [^1]: Edited by: Bernhard Moser, Cardiff University, UK [^2]: Reviewed by: James Edward Pease, Imperial College London, UK [^3]: Specialty section: This article was submitted to Chemoattractants, a section of the journal Frontiers in Immunology

\section{Introduction} Fix $s\in(0,1)$ and $n>2s$. We consider the following problem \begin{equation}\label{problem-1} \begin{cases} (-\Delta)^s u+Vu-u^p=0 &\mbox{ in }{\mathbb R}^n,\\ \lim\limits_{|x|\to \infty}u(x)=0,\quad u>0, \end{cases} \end{equation} where $V$ is a non-negative potential, for a supercritical power nonlinearity, i.e., $p>\frac{n+2s}{n-2s}$. Problem \eqref{problem-1} arises when considering standing wave solutions for the nonlinear fractional Schr\"odinger equation \begin{equation}\label{NLS} -i\frac{\partial \psi}{\partial t}=(-\Delta)^s \psi -Q(y)\psi+|\psi|^{p-1}\psi,\end{equation} that is, solutions of the form $\psi(t,y)=\exp{(i\lambda t)}u(y).$ If $u(y)$ is positive and vanishes at infinity, then $\psi$ satisfies \eqref{NLS} if and only if $u$ solves \eqref{problem-1}. The fractional Schr\"odinger equation, a fundamental tool in fractional quantum mechanics, was introduced by Laskin \cite{Laskin1,Laskin2,Laskin3} (see also the appendix in \cite{Davila-DelPino-Dipierro-Valdinoci}) as a result of extending the Feynman path integral from the Brownian-like paths, which yields the standard Sch\"rodinger equation, to L\'evy-like quantum mechanical paths. There is by now a huge literature on the fractional Schr\"odinger equation, see for instance, \cite{Felmer-Quaas-Tan,Secchi,Cheng,Davila-DelPino-Dipierro-Valdinoci,Molica-Radulescu} for a subcritical power, while for the critical case we have \cite{Frank-Lenzmann,Ambrosio-Figueiredo,Guo-He}, but this list is by no means complete. However, none of these deals with the supercritical regime. The present paper is one of the first attempts in this direction. Our main results are the following: \begin{teo}\label{thm1} Assume that $V\geq 0, V\in L^\infty({\mathbb R}^n)$ and $\displaystyle\lim_{|x|\to \infty}|x|^{2s}V(x)=0$. Then, for $$p>\frac{n+2s-1}{n-2s-1},$$ problem \eqref{problem-1} has a continuum of solutions $u_\lambda$ such that $\displaystyle\lim_{\lambda\to 0}u_\lambda=0$ uniformly in ${\mathbb R}^n$. \end{teo} \begin{teo}\label{thm2} Assume $V\geq 0, V\in L^\infty({\mathbb R}^n)$ and $$\frac{n+2s}{n-2s}<p\leq \frac{n+2s-1}{n-2s-1}.$$ Then the result of the previous theorem holds if either: \begin{itemize} \item[(a)] there exist $C>0$ and $\mu>n$ such that $V(x)\leq C|x|^{-\mu}$; or \item[(b)] there exist a bounded non-negative function $f:\mathbb S^{n-1}\to {\mathbb R}$ not identically zero and $n-\frac{4s}{p-1}<\mu\leq n$ such that \begin{equation*} \lim_{|x|\to \infty}\Big(|x|^\mu V-f\Big(\frac{x}{|x|}\Big)\Big)=0. \end{equation*} \end{itemize} \end{teo} Our approach follows closely the local $s=1$ case from \cite{ddmw}. As in their work, we find a new phenomenon for equation \eqref{problem-1} that is different from the subcritical case, the one of \emph{dispersion}. The main idea in the proof is to perturb a carefully chosen approximate solution. More precisely, the building blocks we use are smooth entire radial solutions of the equation \begin{equation}\label{entiresolution} (-\Delta)^s w=w^p \mbox{ in }{\mathbb R}^n,\quad w>0, \end{equation} which satisfy \begin{equation*} w(0)=1, \quad \lim_{|x|\to \infty}w(x)|x|^{\frac{2s}{p-1}}=\beta^{\frac{1}{p-1}}. \end{equation*} Here $\beta$ is a positive constant chosen so that $w_1(r)=\beta^{\frac{1}{p-1}}|x|^{-\frac{2s}{p-1}}$ is a solution to \eqref{entiresolution}. Its precise value is given in \cite{acdfgw} (there the constant is denoted by $A_{n,p,s}$). Note that $w_1$ is singular at the origin. One of our main contributions here is the construction of a smooth, radially symmetric, entire solution $w$. While in the local case one readily obtains existence by simple ODE phase-plane analysis, in the nonlocal regime we need to use an argument due to Schaaf and a bifurcation argument. Then we prove the asymptotic behavior of $w$ using conformal geometry methods. These ideas were introduced in \cite{acdfgw} to handle a nonlocal ODE. There the authors deal with the singular fractional Yamabe problem and, more generally, with problem \eqref{entiresolution} when the exponent $p$ is subcritical; we were able to extend their methods to the supercritical case. In the case particular $s=\frac{1}{2}$, Chipot, Chlebik and Fila \cite{Chipot-Chlebik-Fila} proved the existence of a radial entire solution and Harada \cite{Harada} established the desired asymptotic behavior at infinity. While our manuscript was being prepared we heard of the recent work by Chen, Gui and Hu \cite{cgh}, that have extended the existence result to systems which includes the existence of solutions to the single equation for any $s\in(0,1)$ using Rabinowitz's bifurcation theory. In addition, they prove an upper bound for the solution, i.e. \begin{equation*} w(x)\leq C|x|^{-\frac{2s}{p-1}} \mbox{ for \ all }x\in {\mathbb R}^n. \end{equation*} However, here we need to show the precise asymptotic behavior of the solution $w$ near infinity, this is, \begin{equation}\label{asymptotics-introduction} w=w(|x|)=\beta^{\frac{1}{p-1}}|x|^{-\frac{2s}{p-1}}(1+o(1)) \mbox{ as }|x|\to \infty, \end{equation} and this is our first difficulty. In the classical case, this is reduced to analyze the corresponding ODE for $v(t)$, where $v(t)=r^{\frac{2}{p-1}}w(r)$, and we have denoted $r=|x|$, $t=-\log r$. As we have mentioned above, in the fractional case $v$ is a solution to a non-local ODE. We use the Hamiltonian approach developed in the previous work \cite{acdfgw} in order to prove \eqref{asymptotics-introduction}. Our second main difficulty is the study the linearized operator of \eqref{entiresolution} around the entire solution $w$. Here again we use the arguments developed in \cite{acdfgw}, in particular, the study of the asymptotic behavior of solutions to a non-local linear ODE in terms of the indicial roots of the problem, and the injectivity properties of this linearized operator. Finally, to pass from the linearized equation to the nonlinear problem \eqref{problem-1} we follows the ideas from the local case in \cite{ddmw}. \medskip The paper is organized as follows. In Sections 2-3, we obtain the existence of entire radial solutions to \eqref{entiresolution} and analyze its asymptotic behavior. In Section 4, we consider the linearized operator around the solution $w$ obtained in Section 2. In Section 5, we use the results in Section 4 to study a perturbed linear problem. Sections 6-7 are devoted to the proof of Theorems \ref{thm1}-\ref{thm2}. \section{Existence of radial entire solutions}\label{section:entire} Let $p+1>2^*(s):=\frac{2n}{n-2s}$. We look first for radially symmetric solutions for \begin{equation}\label{eq:LE} (-\Delta)^s{w}=w^p\quad\text{ in }{\mathbb R}^n,\quad w>0. \end{equation} The constant $\beta>0$ is chosen as in the introduction, so that $w_1(r)=\beta^{\frac{1}{p-1}}r^{-\frac{2s}{p-1}}$ is a solution. However, this $w_1$ is singular at the origin. Our main result in this section is the construction of an entire solution: \begin{proposition}\label{prop:existentire} There exists a radially symmetric, decreasing in the radial variable, entire solution $\bar w$ to \eqref{eq:LE} such that \begin{equation*} \bar w(r)\leq Cr^{-\frac{2s}{p-1}} \mbox{ as } r\to+\infty. \end{equation*} \end{proposition} To prove this proposition, we follow the method in \cite{acdfgw}. First we consider the Dirichlet problem \begin{equation}\label{eq:auxprob}\begin{cases} (-\Delta)^s{w}=\lambda(1+w)^p&\text{ in }{B_1},\\ w=0&\text{ in }{\mathbb R}^n\setminus{B_1}. \end{cases}\end{equation} The existence for this equation follows from classical arguments (we refer to \cite{RosOton-Serra2} for details): for $\lambda>0$ small enough, since $(1-|x|^2)_+^s$ is a positive super-solution, there exists a minimal solution $w_\lambda$. Moreover, $w_\lambda$ is non-decreasing in $\lambda$. Thus one can find a $\lambda^*>0$ such that (i) the minimal solution $w_\lambda$ exists for each $\lambda\in(0,\lambda^*)$, and $w_\lambda$ is radially symmetric and decreasing in the radial variable; (ii) for $\lambda>\lambda^*$, \eqref{eq:auxprob} has no solutions. An argument of Schaaf \cite{Schaaf} shows uniqueness: \begin{lemma}[Uniqueness]\label{lem:uniqueness} There exists $\lambda_0>0$ depending only on $n$, $s$ and $p$, such that for all small $\lambda\in[0,\lambda_0)$, $w_\lambda$ is the unique solution to \eqref{eq:auxprob} \end{lemma} The proof this lemma is postponed to the end of the section. \medskip Now we consider the integral formulation of \eqref{eq:auxprob}, \begin{equation}\label{eq:auxprob2} \begin{cases} w=\lambda T(w) &\text{ in } B_1,\\ w=0 &\text{ in } {\mathbb R}^n\setminus B_1, \end{cases} \end{equation} where (see for example \cite{Bucur}), \[T(w)(x)=C_{n,s}\int_{B_1} \left(\int_0^{r_0(x,y)} \dfrac{t^{s-1}}{(t+1)^{\frac{n}{2}}}\,dt \right) \dfrac{(1+w(y))^{p}}{|x-y|^{n-2s}}\,dy, \quad r_0(x,y)=\dfrac{(1-|x|^2)(1-|y|^2)}{|x-y|^2}. \] We perform a bifurcation argument in the Banach space of non-negative and radially symmetric, non-increasing functions supported on $B_1$, \begin{equation*} E=\set{w\in{\mathcal C}({\mathbb R}^n):w(x)=\tilde{w}(|x|),\,\tilde{w}(r_1)\leq \tilde{w}(r_2)\text{ for } r_1>r_2,\,w\geq0\text{ in }{B_1},\,w=0\text{ in }{\mathbb R}^n\setminus{B_1}}. \end{equation*} \begin{lemma}[Bifurcation]\label{lem:bifurcation} There exists a sequence of solutions $(\lambda_j,w_j)$ of \eqref{eq:auxprob} in $(0,\lambda^*]\times{E}$ such that \[\lim_{j\to\infty}\lambda_j=\lambda_\infty\in[\lambda_0,\lambda^*] \quad\text{ and }\quad \lim_{j\to\infty}\norm[L^\infty(B_1)]{w_j}=+\infty,\] where $\lambda_0$ is given in Lemma \ref{lem:uniqueness}. \end{lemma} \begin{proof} The proof follows closely \cite[Lemma 2.8]{acdfgw}. In the integral formulation \eqref{eq:auxprob2}, it is readily checked that $T$ defines a (nonlinear) compact operator. Moreover, the pair $(0,0)$ is a solution. Now, \cite[Theorem 6.2]{Rabinowitz} yields an unbounded continuum of solutions in ${\mathbb R} \times E$ connecting $(0,0)$. Since the minimal solution $w_\lambda$ is unique for $\lambda\in(0,\lambda_0)$ by Lemma \ref{lem:uniqueness} and there does not exist any solution for $\lambda>\lambda^*$, we must have $\lambda\in[\lambda_0,\lambda^*]$ whenever $\norm[L^\infty(B_1)]{w_\lambda}$ is large. \end{proof} Now we are ready to prove Proposition \ref{prop:existentire} using a blow-up argument. \begin{proof}[Proof of Proposition \ref{prop:existentire}] Let $(\lambda_j,w_j)$ be as in Lemma \ref{lem:bifurcation}. Define \[m_j=\norm[L^{\infty}(B_1)]{w_j}=w_j(0) \quad\text{ and }\quad R_j=m_j^{\frac{p-1}{2s}}.\] Then \[W_j(x)=\lambda_j^{\frac{1}{p-1}}m_j^{-1}w_j\left(\frac{x}{R_j}\right)\] satisfies $0\leq{W_j}\leq{W_j(0)}=1$ and \[\begin{cases} (-\Delta)^s{W_j}=\left(\lambda_j^{\frac{1}{p-1}}m_j^{-1}+W_j\right)^p &\text{ in }{B_{R_j}},\\ W_j=0 &\text{ in }{\mathbb R}^n\setminus{B_{R_j}}. \end{cases}\] By elliptic regularity (see for instance \cite{RosOton-Serra2}), $W_j\in\mathcal {C}_{\textnormal{loc}}^{\alpha}({\mathbb R}^n)$ for some $\alpha>0$ and hence, by passing to a subsequence, $W_j\to{\bar w}$ in $\mathcal C_\textnormal{loc}({\mathbb R}^n)$ for some positive and radially decreasing function $\bar w$ which satisfies \eqref{eq:LE}. Moreover, the solution $\bar w$ is radially symmetric and decreasing. In addition, from Lemma 2.6 of \cite{acdfgw} with $\beta=0$ there, one can get that \begin{equation*} \bar w(x)\leq C|x|^{-\frac{2s}{p-1}} \mbox{ for }x\in \mathbb R^n. \end{equation*} \end{proof} We finally go back to uniqueness and the proof of Lemma \ref{lem:uniqueness}: \begin{proof}[Proof of Lemma \ref{lem:uniqueness}] Suppose $w=w_\lambda+v$ is another solution. Then $v$ solves \[\begin{cases} (-\Delta)^s{v}=\lambda{f(v)}&\text{ in }{B_1},\\ v=0&\text{ in }{\mathbb R}^n\setminus{B_1}, \end{cases}\] where $f(v)=(1+w_\lambda+v)^p-(1+w_\lambda)^p$. Let us write $F(v)=\int_0^{v}f$. For any $\sigma\in{\mathbb R}$, the Poho\v{z}aev identity \cite{RosOton-Serra1} reads \begin{equation* \left(\dfrac{1}{2^*(s)}-\sigma\right)\int_{{\mathbb R}^n}v(-\Delta)^s{v}\,dx\leq\lambda\int_{B_1}\left(F(v)-\sigma{v}f(v)\right)\,dx. \end{equation*} The left hand side is estimated using the fractional Sobolev inequality so that \begin{equation* \left(\dfrac{1}{2^*(s)}-\sigma\right)\left(\int_{B_1}v^{2^*(s)}\,dx\right)^{\frac{2}{2^*(s)}}\,dx\leq\lambda{C}\int_{B_1}\left(F(v)-\sigma{v}f(v)\right)\,dx. \end{equation*} For the right hand side, since $\displaystyle\lim_{v\to+\infty}\frac{F(v)}{vf(v)}=\frac{1}{p+1}$, for any $\varepsilon>0$ there exists $M=M(\varepsilon)$ such that for any $v\geq{M}$, $F(v)\leq\frac{1+\varepsilon}{p+1}vf(v)$. Since $p+1>2^*(s)$, there exist $\varepsilon$ and $\sigma$ such that $\frac{1+\varepsilon}{p+1}<\sigma<\frac{1}{2^*(s)}$. In other words, fixing a choice of $\varepsilon$ and $\sigma$, there is an $M$ such that $F(v)-\sigma{v}f(v)<0$ whenever $v\geq{M}$. Now, since $F(v)$ is quadratic in $v$, namely, \[F(v)=v^2\int_0^1\int_0^1pt_1(1+w_\lambda+t_1t_2v)^{p-1}\,dt_1dt_2,\] we have \begin{equation* \begin{split} \left(\dfrac{1}{2^*(s)}-\sigma\right)\left(\int_{B_1}v^{2^*(s)}\,dx\right)^{\frac{2}{2^*(s)}}\,dx &\leq\lambda{C}\int_{B_1\cap\set{v<M}}\left(F(v)-\sigma{v}f(v)\right)\,dx\\ &\leq\lambda{C}\int_{B_1\cap\set{v<M}}v^2\,dx\\ &\leq\lambda{C}\left(\int_{B_1\cap\set{v<M}}v^{2^*(s)}\,dx\right)^{\frac{2}{2^*(s)}}|B_1\cap\set{v<M}|^{1-\frac{2}{2^*(s)}}\\ &\leq\lambda{C}\left(\int_{B_1\cap\set{v<M}}v^{2^*(s)}\,dx\right)^{\frac{2}{2^*(s)}}. \end{split} \end{equation*} Therefore, $v\equiv0$ if $\lambda>0$ is chosen small enough. \end{proof} \section{Conformal geometry results and the asymptotic behavior for the solution} From the last section, we know that there exists a smooth solution $\bar w$ of $(-\Delta)^s w=w^p $ in ${\mathbb R}^n $ which satisfies $\bar{w}(0)=1$ and $\bar{w}(x)\leq C|x|^{-\frac{2s}{p-1}}$. Here we will show that it has a precise asymptotic behavior at infinity. For the rest of the paper, we will drop the bar and simply denote this particular solution by $w$. More precisely, \begin{proposition}\label{prop:asymptotics} Let $w$ be the solution constructed in Proposition \ref{prop:existentire}. Then \begin{equation*} \lim_{|x|\to \infty}w(x)|x|^{\frac{2s}{p-1}}=\beta^{\frac{1}{p-1}}, \end{equation*} for the constant $\beta$ given in the introduction, and which corresponds to the coefficient of the singular solution to \eqref{eq:LE}. \end{proposition} We will basically use the results in Sections 3-4 of \cite{acdfgw}, and follow their notation accordingly. Set $r=|x|$, and let $P_s^{g_0}$ be the conformal fractional Laplacian on the cylinder $\mathbb R\times\mathbb S^{n-1}$. Then, by its conformal properties one has that \begin{equation}\label{conjugation1} P_s^{g_0}(r^{\frac{n-2s}{2}}w)=r^{\frac{n+2s}{2}}(-\Delta)^{s} w. \end{equation} Set also $r=e^{-t}$ and \begin{equation}\label{wv} v=e^{-\frac{2s}{p-1}t}w(e^{-t}), \end{equation} and consider one further conjugation \begin{equation*}\label{tilde-P} \tilde P_s^{g_0}(v):= e^{(\frac{n-2s}{2}-\frac{2s}{p-1})t}P_s^{g_0}\big(e^{(-\frac{n-2s}{2}+\frac{2s}{p-1})t}v\big) =r^{\frac{2s}{p-1}p}(-\Delta)^s w. \end{equation*} Then the equation for $w$ transforms into an equation for $v$ on the cylinder, more precisely, \begin{equation}\label{conformal-laplacian} \tilde P_s^{g_0}(v)=v^p, \quad\text{in } \mathbb R \times \mathbb S^{n-1}. \end{equation} The operator $\tilde P_s^{g_0}$ can be understood as a Dirichlet-to-Neumann for an extension problem in the spirit of the construction of the fractional Laplacian by \cite{Caffarelli-Silvestre,Chang-Gonzalez,Gonzalez:survey,DelaTorre-Gonzalez}. Without being very precise on the extension manifold $X^{n+1}$, with metric $\bar g^*$, and the extension variable $\rho^*$, we recall the following proposition in \cite{acdfgw}: \begin{proposition} \label{prop:divV*} Let $v$ be a smooth function on the cylinder $M=\mathbb R\times\mathbb S^{n-1}$. The extension problem \begin{equation*}\label{divV*} \left\{\begin{array}{@{}r@{}l@{}l} -\divergence_{\bar{g}^*}((\rho^*)^{1-2s}\nabla_{\bar{g}^*}V^*) -(\rho^*)^{-(1+2s)}\left(\tfrac{4\rho}{4+\rho^2}\right)^{2}2\Big(-\frac{n-2s}{2}+\tfrac{2s}{p-1}\Big)\,\partial_t V^*&\,=0\quad &\text{in } (X,\bar g^*), \medskip\\ V^*|_{\rho=0}&\,=v\quad &\text{on }M, \end{array}\right. \end{equation*} has a unique solution $V^*$. Here $\rho=\rho(\rho^*)$ is a positive function of $\rho^*$, $\rho^*\in(0,\rho_0^*)$. Moreover, for its Neumann data, \begin{equation*}\label{Neumann-V*} \tilde P_s^{g_0}(v)=-d_s \lim_{\rho^* \to 0}(\rho^*)^{1-2s} \partial_{\rho^*} (V^*)+\beta v, \end{equation*} for the constant \begin{equation}\label{constant-extension} d_s=\frac{2^{2s-1}\Gamma(s)}{s\Gamma(-s)}, \end{equation} and $\beta$ as above. \end{proposition} Now consider the spherical harmonic decomposition of $\mathbb S^{n-1}$. For this, let $\mu_m$ denote the $m$-th eigenvalue for $-\Delta_{\mathbb S^{n-1}}$, repeated according to multiplicity, and by $E_m(\theta)$ the corresponding eigenfunction. Then any $v$ defined on $\mathbb R\times\mathbb S^{n-1}$ can be written as $v(t,\theta)=\sum_m v_m(t)E_m(\theta)$. Denote by $\tilde P_s^{(m)}$ the projection of the operator $\tilde P_s^{g_0}$ onto each eigenspace, for $m=0,1,\ldots$. Next, for $w$ a radially symmetric function, then $v$ defined as in \eqref{wv} only depends on the variable $t=-\log r$ and thus, after projection onto spherical harmonics, only the zero-th projection survives and equation \eqref{conformal-laplacian} reduces to \begin{equation}\label{problem-vv} \tilde P_s^{(0)}(v)=v^p, \quad\text{in } \mathbb R. \end{equation} Recalling the expression for $\tilde P^{(0)}_\gamma$ from \cite[formula (4.3)]{acdfgw}, one has \begin{equation}\label{convolution-K} \tilde P^{(0)}_s(v)(t)=P.V. \int_{\mathbb R}\tilde{\mathcal K}_0(t-t')[v(t)-v(t')]\,dt'+\beta v(t) \end{equation} for the convolution kernel \begin{equation*} \tilde{\mathcal K}_0(t)=c\,e^{-(\frac{2s}{p-1}-\frac{n-2s}{2})t}\int_0^\pi \frac{(\sin \phi_1)^{n-2}}{(\cosh t-\cos \phi_1)^{\frac{n+2s}{2}}}\,d\phi_1, \end{equation*} where $c$ is a positive constant that only depends on $n$ and $s$. As a consequence of Proposition \ref{prop:divV*}, one obtains the following Hamiltonian quantity for the equation \eqref{problem-vv} (Theorem 4.3 in \cite{acdfgw}), \begin{equation*}\label{Hamiltonian} \begin{split} H^*(t)=&\frac{1}{{d}_{s}} \left(-\frac{\beta}{2}v^2+\frac{1}{p+1}v^{p+1}\right) +\frac{1}{2}\int_{0}^{\rho^*_0} {(\rho^*)}^{1-2\gamma}\left\{-e^*_1(\partial_{\rho^*}V^*)^2 +e^*_2(\partial_t V^*)^2\right\}\,d\rho^*\\ =&:H_1(t)+H_2(t), \end{split} \end{equation*} where $V^*=V^*(t,\rho^*)$ is the extension of $v=v(t)$. Moreover, this Hamiltonian is monotone in $t$. Indeed, \begin{equation}\label{ham_der0} \partial_t\left[ H^*(t)\right]=-2\int_{0}^{\rho^*_0}\left[(\rho^*)^{1-2s}e^* \left(\tfrac{\rho}{\rho^*}\right)^{-2}\left(1+\tfrac{\rho^2}{4}\right)^{-2} \left(-\tfrac{n-2s}{2}+\tfrac{2s}{p-1}\right)\left[\partial_t V^*\right]^2\right]\,d\rho^*, \end{equation} and the functions $e^*$, $e_1^*$ and $e^*_2$ are strictly positive. \bigskip \begin{proof}[Proof of Proposition \ref{prop:asymptotics}] Our aim is to show that \begin{equation*} \lim_{t\to -\infty}v(t)=\beta^{\frac{1}{p-1}}. \end{equation*} Equation \eqref{ham_der0} implies that \begin{equation*}\label{ham_der} \partial_t\left[ H^*(t)\right] \geq 0, \end{equation*} since $\left(-\tfrac{n-2s}{2}+\tfrac{2s}{p-1}\right)<0$ due to the fact that $p>\frac{n+2s}{n-2s}$. Thus in this case the Hamiltonian $H^*(t)$ is increasing in $t$. By Proposition \ref{prop:existentire}, $v$ is uniformly bounded, hence so is $V^*$. By elliptic regularity, $\partial_t V^*, \ \partial_{\rho^*}V^*$ are also bounded. Then $H^*$ is uniformly bounded. Integrating \eqref{ham_der0} on the real line ${\mathbb R}$, one has \begin{equation*} \begin{split} \int_{-\infty}^\infty \partial_t\left[ H^*(t)\right]dt&=-2\int_{-\infty}^\infty\int_{0}^{\rho^*_0}\left[(\rho^*)^{1-2\gamma}e^* \left(\tfrac{\rho}{\rho^*}\right)^{-2}\left(1+\tfrac{\rho^2}{4}\right)^{-2} \left(-\tfrac{n-2s}{2}+\tfrac{2s}{p-1}\right)\left[\partial_t V^*\right]^2\right]\,d\rho^*dt\\ &=H^*(+\infty)-H^*(-\infty), \end{split} \end{equation*} which is finite by the uniform boundedness of $H^*$. Thus we have that $\int_0^{\rho_0^*}g(\rho^*)|\partial_t V^*|^2\,d\rho^*\to 0 $ as $t\to \pm\infty$ for some non-negative function $g$, which implies that $\partial_t V^*\to 0$ as $t\to \pm \infty$. For any sequence $\{t_i\}\to -\infty$, consider $V^*_i(t, \rho^*)=V^*(t+t_i, \rho^*)$. There exists a subsequence $V^*_i(t, \rho^*)\to V_\infty(t, \rho^*)$ in $\mathcal{C}\left([-1, 1]\times [0, \rho_0^*]\right)$. By the argument above, we know that $\partial_t V_\infty(t, \rho^*)=0$, i.e. $V_\infty=V_\infty(\rho^*)$, and this is a solution of the same equation satisfied by $V^*(t, \rho^*)$. Since $V^*$ is the extension of $v(t)$, one has $v(t)\to V_\infty(0)$ as $t\to -\infty$, where $V_\infty(0)$ is a constant. Moreover, $V_\infty(0)$ is a solution of \begin{equation*} \tilde P_s^{(0)}(v)=v^p, \quad\text{in } \mathbb R. \end{equation*} Since $\displaystyle\lim_{t\to +\infty}H^*(t)=0$, $\partial_t H^*(t)\geq 0$ and is not identically zero. One then obtains that $V_\infty(\rho^*)$ is not trivial, so $V_\infty(0)$ is not zero. From this and the expression of $\tilde P^{(0)}_s$ above (formula \eqref{convolution-K}), one can easily get that $V_\infty(0)=\beta^{\frac{1}{p-1}}$, as desired. \end{proof} \section{Linear theory}\label{section:linear} From the last section, we know that there exists a solution $w$ of $(-\Delta)^s u=u^p $ in ${\mathbb R}^n $ that satisfies \begin{equation}\label{w} \left\{\begin{array}{l} w(0)=1,\\ w(x)\sim\beta^{\frac{1}{p-1}}|x|^{-\frac{2s}{p-1}},\quad \text{as } |x|\to \infty. \end{array} \right. \end{equation} We are now interested in the linear problem: \begin{equation}\label{model-linearization} L\phi:=(-\Delta)^s\phi-\frac{\mathcal V(r)}{r^{2s}}\phi=h, \end{equation} where we have defined the (radial) potential \begin{equation*}\label{potential} \mathcal V:=\mathcal V(r)=r^{2s}pw^{p-1}. \end{equation*} The asymptotic behavior of this potential is easily calculated using the asymptotics for $w$ and, indeed, if we define $r=e^{-t}$, \begin{equation}\label{asymptotics-potential} \mathcal V(t)=\begin{cases} O(e^{-2st}) &\mbox{ as }t\to +\infty,\\ p\beta +O(e^{q_1t} ) &\mbox{ as }t\to -\infty, \end{cases}\end{equation} for some $q_1>0$. In the $s=1$ case, solutions to \eqref{model-linearization} are constructed directly by projecting $\phi$ and $h$ onto spherical harmonics and then solving the corresponding ODEs. This approach cannot be directly applied here. Instead, we use the conformal geometry tools developed in \cite{acdfgw}. By the well known extension theorem for the fractional Laplacian \cite{Caffarelli-Silvestre}, equation \eqref{model-linearization} is also equivalent, for $s\in(0,1)$, to the boundary reaction problem \begin{equation*}\label{eqlinear1} \left\{\begin{array}{r@{}l@{}l} \partial_{yy}\Phi+\dfrac{1-2s}{y}\partial_y \Phi+\Delta_{{\mathbb R}^n}\Phi &\,=0 &\quad\mbox{in }{\mathbb R}^{n+1}_+,\medskip\\ - d_s\lim\limits_{y\to0}y^{1-2s}\partial_y \Phi &\,=\dfrac{\mathcal V(r)}{r^{2s}}\phi+h &\quad\mbox{on }{\mathbb R}^n, \end{array}\right. \end{equation*} where we have denoted $\Phi|_{y=0}=\phi$ and the constant $d_s$ is given by \eqref{constant-extension}. Using the spherical harmonic decomposition as above, we can write $\Phi=\sum_{m=0}^\infty \Phi_m(r,y)E_m(\theta)$, $\phi=\sum_{m=0}^\infty \phi_m(r)E_m(\theta)$, $h=\sum_{m=0}^\infty h_m(r)E_m(\theta)$, and $\Phi_m$ satisfies the following: \begin{equation}\label{eqlinear2} \left\{\begin{array}{r@{}l@{}l} \partial_{yy}\Phi_{m}+\dfrac{1-2s}{y}\partial_y\Phi_{m} +\Delta_{{\mathbb R}^n}\Phi_m-\dfrac{\mu_m}{r^2}\Phi_m &\,=0 &\quad\mbox{in }{\mathbb R}^{n+1}_+,\medskip\\ - d_s\lim\limits_{y\to0}y^{1-2s}\partial_y \Phi_m &\,=\dfrac{\mathcal V(r)}{r^{2s}}\phi_m+h_m &\quad\mbox{on }{\mathbb R}^n. \end{array} \right. \end{equation} We will now use conformal geometry to rewrite equation \eqref{model-linearization}, as explained in the previous section. If we define \begin{equation}\label{shift} \psi=r^{\frac{n-2s}{2}}\phi, \end{equation} then by the conformal property \eqref{conjugation1}, we have that equation \eqref{model-linearization} is equivalent to the problem \begin{equation}\label{eq0} \mathcal L \psi:=P_s^{g_0}(\psi)-\mathcal V\psi=r^{2s}h=:\tilde h. \end{equation} Projecting onto spherical harmonics we also have \begin{equation}\label{eq0m} \mathcal L_m \psi_m:=P_s^{(m)}(\psi_m)-\mathcal V\psi_m=\tilde h_m, \quad m=0,1,\ldots. \end{equation} \subsection{Indicial roots}\label{subsection:indicial-roots} Let us calculate the indicial roots for the model linearized operator defined in \eqref{model-linearization} as $r\to 0$ and as $r\to \infty$. One can see that the indicial roots here are similar to the ones in \cite{acdfgw} with the roles at $r=0$ and $\infty$ interchanged. However, recall that in \cite{acdfgw} a subcritical power is taken, $\frac{n}{n-2s}<p<\frac{n+2s}{n-2s}$, while here $p>\frac{n+2s}{n-2s}$ is supercritical. To handle this difference we need to use the result in \cite{lwz}, where the authors study the stability of the singular solution $w=\beta^{\frac{1}{p-1}}r^{-\frac{2s}{p-1}}$. In particular, they show the existence of a threshold dimension $n_0(s)$ such that in any higher dimension $n>n_0(s)$, there exists $p_2>\frac{n+2s}{n-2s}$ such that the singular solution is unstable if $\frac{n+2s}{n-2s}<p<p_2$, and is stable if $p\geq p_2$. For $n\leq n_0(s)$, however, the singular solution is unstable for all $p>\frac{n+2s}{n-2s}$. When $s=1$, this exponent $p_2$ corresponds to the well-known Joseph-Lundgren exponent \cite{gnw}. Define now \begin{equation*} P_{JL}=\left\{\begin{array}{l} p_2, \quad n>n_0(s),\\ \infty, \quad n\leq n_0(s). \end{array} \right. \end{equation*} Using the above result and following similar argument in the proof of Lemma 7.1 in \cite{acdfgw}, one has the following result: \begin{lemma}\label{indicial} For the operator $L$ we have that, for each fixed mode $m=0,1,\ldots,$ \begin{itemize} \item[\emph{i.}] At $r=0$, there exist two sequences of indicial roots \begin{equation*} \Big\{\tilde \sigma_j^{(m)}\pm i\tilde\tau_j^{(m)}-\tfrac{n-2s}{2}\Big\}_{j=0}^\infty\quad \text{and} \quad\Big\{-\tilde\sigma_j^{(m)}\pm i\tilde\tau_j^{(m)}-\tfrac{n-2s}{2}\Big\}_{j=0}^\infty. \end{equation*} Moreover, at the $j=0$ level the indicial roots are real; more precisely, the numbers \begin{equation*} \tilde\gamma_m^{\pm}:=\pm\tilde\sigma_0^{(m)}-\tfrac{n-2s}{2}=-\tfrac{n-2s}{2}\pm \left[1-s+\sqrt{(\tfrac{n-2}{2})^2+\mu_m}\right],\quad m=0,1,\ldots, \end{equation*} form an increasing sequence in $m$ (except for multiplicity repetitions). \item[\emph{ii.}] At $r=\infty$, there exist two sequences of indicial roots \begin{equation*} \Big\{\sigma_j^{(m)}\pm i\tau_j^{(m)}-\tfrac{n-2s}{2}\Big\}_{j=0}^\infty\quad \text{and} \quad\Big\{-\sigma_j^{(m)}\pm i\tau_j^{(m)}-\tfrac{n-2s}{2}\Big\}_{j=0}^\infty. \end{equation*} Moreover, \begin{itemize} \item[\emph{a)}] For the mode $m=0$, if $\frac{n+2s}{n-2s}<p<P_{JL}$ (the \underline{unstable} case), then the indicial roots $\gamma_0^\pm$ are a pair of complex conjugates with real part $-\frac{n-2s}{2}$ and imaginary part $\pm\tau_0^{(0)}$; while if $p\geq P_{JL}$ (the \underline{stable} case), the indicial roots $\gamma_0^{\pm}:=\pm\sigma_0^{(0)}-\frac{n-2s}{2}$ are real with \begin{equation*}\label{gamma0} -(n-2s)+\tfrac{2s}{p-1}<\gamma_0^-<-\tfrac{n-2s}{2}<\gamma_0^+<-\tfrac{2s}{p-1}, \end{equation*} \item[\emph{b)}] In addition, for all $j\geq 1$, \begin{equation*} \sigma_j^{(0)}>\tfrac{n-2s}{2}. \end{equation*} \item[\emph{c)}] For the mode $m=1$, the indicial roots $\gamma_1^{\pm}:=\pm \sigma_0^{(1)}-\tfrac{n-2s}{2}$ take the values $$-\tfrac{2s}{p-1}-1 \quad\mbox{ and }\quad -(n-2s)+\tfrac{2s}{p-1}+1.$$ \end{itemize} \end{itemize} \end{lemma} \subsection{Conformal geometry and invertibility for a Hardy-type operator} First we recall the results in \cite{acdfgw}, adapted to our setting, in order to study the invertibility of a Hardy type operator with fractional Laplacian: \begin{equation*} \mathcal L_\kappa \psi:=P_s^{g_0}(\psi)-\kappa\psi=r^{2s}h=:\tilde h, \end{equation*} where $\kappa$ is any fixed real constant. Projecting onto spherical harmonics this equation is equivalent to \begin{equation}\label{eq-kappa} \mathcal L_{\kappa,m} \psi_m:=P_s^{(m)}(\psi_m)-\kappa\psi_m=\tilde h_m, \quad m=0,1,\ldots. \end{equation} We will consider values of $\kappa$ for which the indicial roots of $\mathcal L_{\kappa,m}$ are those of Lemma \ref{indicial}, except for a shift of $\frac{n-2s}{2}$ due to the change \eqref{shift}. Based on the results in \cite{acdfgw} we have, in both the stable and the unstable cases of Lemma \ref{indicial}, the following \begin{teo}\label{thm:Hardy-potential} Fix $m=0,1,\ldots$, and assume that the right hand side $\tilde h_m$ in \eqref{eq-kappa} satisfies \begin{equation*}\label{decay-h} h_m(t)=\begin{cases} O(e^{\delta t}) &\text{as } t\to-\infty,\\ O(e^{-\delta_0 t}) &\text{as } t\to+\infty, \end{cases} \end{equation*} for some real constants $\delta,\delta_0$. \begin{itemize} \item[\emph{i.}] Assume that $\delta+\delta_0\geq 0$. Then a particular solution to \eqref{eq-kappa} is \begin{equation*}\label{solucion2} \psi_m(t)=\int_{\mathbb R}\tilde h_m(t') {{\mathcal G}}_m(t-t')\,dt', \end{equation*} where $\mathcal G$ is an even $\mathcal C^\infty$ function outside the origin. The exact formula for $\mathcal G_m$ depends on the value of $\delta$ with respect to the indicial roots in Lemma \ref{indicial}. In any case, \begin{equation}\label{decay-phi} \psi_m(t)=O(e^{\delta t}) \quad\text{as} \quad t\to -\infty, \quad \psi_m(t)=O(e^{-\delta_0 t}) \quad\text{as} \quad t\to +\infty. \end{equation} \item[\emph{ii.}] All solutions of the homogeneous problem $\mathcal L_{\kappa,m} \psi_m=0$ are linear combinations of $e^{\sigma_j\pm i\tau_j}$ and $e^{-\sigma_j\pm i\tau_j}$, where $\sigma_j\pm i\tau_j$, $-\sigma_j\pm i\tau_j$, $j=0,1,\ldots$, are the indicial roots for the problem. Thus the only solution to \eqref{eq-kappa} with decay as in \eqref{decay-phi} is precisely $\psi_m$. \end{itemize} \end{teo} \subsection{Study of the linear problem in weighted spaces} In this subsection we come to study the linear problem \eqref{model-linearization}. For this, we will work in weighted $L^\infty$ spaces in which the weight is chosen differently in a bounded set and near infinity. Define the norms \begin{equation}\label{norms} \begin{split} \|\phi\|_*&=\sup_{\{|x|\leq 1\}}|x|^\sigma |\phi(x)|+\sup_{\{|x|\geq 1\}}|x|^{\frac{2s}{p-1}}|\phi(x)|,\\ \|h\|_{**}&=\sup_{\{|x|\leq 1\}}|x|^{\sigma+2s} |h(x)|+\sup_{\{|x|\geq 1\}}|x|^{\frac{2s}{p-1}+2s}|h(x)|, \end{split} \end{equation} where $\sigma\in (0, n-2s)$ is a constant to be determined later. Our aim is to get the following solvability result: \begin{proposition}\label{invertibility} Let $h$ satisfy $\|h\|_{**}<+\infty$. For linear problem \eqref{model-linearization}, we have: \begin{itemize} \item[\emph{i}.] if $p>\frac{n+2s-1}{n-2s-1}$, then there exists a solution $\phi$ and it satisfies \begin{equation}\label{bound} \|\phi\|_*\leq C\|h\|_{**} \end{equation} for some $C>0$; \item[\emph{ii.}] if $\frac{n+2s}{n-2s}<p<\frac{n+2s-1}{n-2s-1}$ and, in addition, \begin{equation*} \int_{{\mathbb R}^n}h\frac{\partial w}{\partial x_i}dx=0, \ i=1,\cdots, n, \end{equation*} then there exists a solution $\phi$ and it satisfies \begin{equation*} \|\phi\|_*\leq C\|h\|_{**}. \end{equation*} \end{itemize} \end{proposition} \begin{remark} It is known that the linear operator $L$ has $n+1$ kernels corresponding to scaling (the mode zero kernel) and translation (the mode one kernels), i.e. \begin{equation*} \begin{split} z_0(x)&=\frac{\partial }{\partial \lambda}\Big( \lambda^{\frac{2s}{p-1}}w(\lambda x)\Big)\Big|_{\lambda=1}=rw'(r)+\frac{2s}{p-1}w, \\ z_i(x)&=\frac{\partial }{\partial x_i}w(x),\quad i=1,\ldots,n. \end{split} \end{equation*} These constitute an obstruction for the solvability of \eqref{model-linearization} and need to be taken into account in the arguments below. \end{remark} In the following, we will always assume that \begin{equation*} \sigma \in \big(0, |\Real(\gamma_0^+)|\big)\subset \Big(0, \frac{n-2s}{2}\Big). \end{equation*} Additional conditions will be given in the proofs below. \medskip We start with a non-degeneracy result: \begin{lemma}\label{non-degeneracy} If $\phi$ is a solution of \begin{equation*} (-\Delta)^s \phi-pw^{p-1}\phi=0 \quad\mbox{in }{\mathbb R}^n \end{equation*} satisfying $\|\phi\|_*<\infty$, then \begin{equation*} \phi=\sum_{i=1}^n c_i\frac{\partial w}{\partial x_i} \end{equation*} for some $c_i\in {\mathbb R}$. \end{lemma} \begin{proof} Consider the spherical harmonic decomposition $\displaystyle\phi=\sum_m \phi_m E_m$ and write $\psi_m=r^{\frac{n-2s}{2}}\phi_m$. {\bf Step 1:} the mode $m=0$. Define the constant $\kappa=p\beta$ and rewrite equation \eqref{eq0m} for $m=0$ as \begin{equation}\label{choice-h} \mathcal L_{\kappa,0}\psi:=P_s^{(0)}(\psi)-\kappa \psi=(\mathcal V-\kappa)\psi=:\tilde h \end{equation} for some $\psi=\psi(t)$, $\tilde h=\tilde h(t)$. We use \eqref{asymptotics-potential} and the definition of $\psi$ to estimate the right hand side, \begin{equation*}\label{asymptotics-h} \tilde h(t)=\begin{cases} O(e^{-(\frac{n-2s}{2}-\sigma)t} )&\mbox{ as }t\to +\infty,\\ O(e^{(q_1+\frac{2s}{p-1}-\frac{n-2s}{2})t})&\mbox{ as }t\to -\infty, \end{cases}\end{equation*} for some $q_1>0$. Note that there could be solutions to the homogeneous problem of the form $e^{(\sigma_j\pm i\tau_j)t}$, $e^{(-\sigma_j\pm i\tau_j)t}$. But these are not allowed by the choice of weights since $\frac{n-2s}{2}-\frac{2s}{p-1}\in (0, \frac{n-2s}{2})$ and $\frac{n-2s}{2}-\sigma>\sigma_0^{(0)}$ (for this, recall statements \emph{a)} and \emph{b)} in Lemma \ref{indicial}). Now we apply Theorem \ref{thm:Hardy-potential} with $\delta=q_1+\frac{2s}{p-1}-\frac{n-2s}{2}<-\sigma_0^{(0)}$ and $\delta_0=-\sigma+\frac{n-2s}{2}>\sigma_0^{(0)}$. Obviously, $\delta+\delta_0>0$ if {$\sigma<\frac{2s}{p-1}$.} Then we can find a particular solution $\psi_0$ such that \begin{equation*} \psi_0(t)=(e^{\delta t}), \quad\text{as}\quad t\to -\infty,\qquad \psi_0(t)=(e^{-\delta_0 t}), \quad\text{as}\quad t\to +\infty, \end{equation*} so $\psi$ will have the same decay. Now, by the definition of $h$ in \eqref{choice-h}, we can iterate this process with $\delta=lq_1+\frac{2s}{p-1}-\frac{n-2s}{2}$, $l\geq 2$, and the same $\delta_0$, to obtain better decay when $t\to -\infty$. As a consequence, we have that $\psi$ decays faster than any $e^{\delta t}$ as $t\to -\infty$, which when translated to $\phi$ means that $\phi=o(r^{-a})$ as $r\to +\infty$ for every $a\in\mathbb N$. By considering the equation satisfied by the Kelvin transform $\hat{\phi}$ of $\phi$, one has \begin{equation*} (-\Delta)^s \hat{\phi}-\frac{\tilde{\mathcal{V}}}{r^{2s}}\hat{\phi}=0, \end{equation*} where $\tilde{\mathcal{V}} $ satisfies \begin{equation*} \tilde{\mathcal{V}}(x)=p|x|^{-2s}w^{p-1}\Big(\frac{x}{|x|^2}\Big)=\left\{\begin{array}{ll} p\beta(1+o(1)) &\mbox{ as }r\to 0,\\ r^{-2s} &\mbox{ as } r\to \infty, \end{array} \right. \end{equation*} and $\hat{\phi}(r)=o(r^{a})$ as $r\to 0$ for every $a\in\mathbb N$. The strong unique continuation result of \cite{Fall-Felli} (stable case) and \cite{Ruland} (unstable case) for the operator $P_s^{(0)}-\hat{\mathcal V}$ implies that $\hat{\phi}$ must vanish everywhere, which yields that also $\phi$ must be zero everywhere. \medskip {\bf Step 2:} the modes $m=1,\ldots,n$. Differentiating equation $(-\Delta)^s w=w^p$ with respect to $x_m$ we get \begin{equation*} L\frac{\partial w}{\partial x_m}=0. \end{equation*} Since $w$ only depends on $r$, we have $\frac{\partial w}{\partial x_m}=w'(r)E_m$, where $E_m=\frac{x_m}{|x|}$. Using the fact that $-\Delta_{\mathbb S^{n-1}} E_m=\mu_m E_m$, the extension for $w'(r)$ to $\mathbb R^{n+1}_+$ solves \eqref{eqlinear2} with eigenvalue $\mu_m=n-1$, and $\psi_1:=r^{\frac{n-2s}{2}}w'$ satisfies \begin{equation}\label{eq} P_s^{(m)} \psi-\mathcal V\psi=0. \end{equation} Note that $w'(r)$ decays like $r^{-(\frac{2s}{p-1}+1)}$ as $r\to \infty$ and decays like $r$ as $r\to 0$. Assume that $\phi_m$ decays like $r$ as $r\to 0$ and decays like $r^{\gamma_m^-}$ as $r\to \infty$. Then $\psi_m=r^{\frac{n-2s}{2}}\phi_m$ is another solution to \eqref{eq}, and we can find a non-trivial combination of $w'$ and $\phi_m$ that decays faster than $r$ at zero. Now we claim that if $\phi=r^{-\frac{n-2s}{2}}\psi$, where $\psi$ is any solution to \eqref{eq}, it cannot decay faster than $r$ at $0$, which yields that $\phi_m=cw'$ for $m=1, \ldots, n$. To show this claim we argue as in Step 1, taking the indicial roots at $0$ (namely $-(n+1-2s)$ and $1$) and interchanging the role of $+\infty$ and $-\infty$ in the decay estimate. More precisely, we use the facts that if such $\phi$ like $r^{\sigma'}$ for some $\sigma'>1$, i.e. $\sigma'+\frac{n-2s}{2}>\frac{n-2s}{2}+1=\tilde{\sigma}_0^{(1)}$ and decays like $r^{\sigma'_1}$ for some $\sigma'_1<-(\frac{2s}{p-1}+1)$, i.e. $\sigma'_1+\frac{n-2s}{2}<\tilde{\sigma}_1^{(1)}$, similarly to Step 1, one can show that the solution is identically zero, and we conclude that $\phi_m=cw'$ for some $c$. \medskip {\bf Step 3:} the remaining modes $m\geq{n+1}$. We use an integral estimate involving the first mode which has a sign, as in \cite{Davila-delPino-Musso-Wei,Davila-delPino-Musso}. We note, in particular, that $\phi_1(r)=-w'(r)>0$, which also implies that its extension $\Phi_1$ is positive. In general, the $s$-harmonic extension $\Phi_m$ of $\phi_m$ satisfies \[\left\{ \begin{array}{r@{}ll} \divergence(y^{1-2s}\nabla\Phi_m) &\,=\mu_{m}\dfrac{y^{1-2s}}{r^2}\Phi_m &\text{ in } {\mathbb R}^{n+1}_+,\medskip\\ -{d}_s\lim\limits_{y\to0}y^{1-2s}\partial_{y}\Phi_m &\,=pu_1^{p-1}\phi_m &\text { on } {\mathbb R}^{n+1}_+. \end{array} \right.\] We multiply this equation by $\Phi_1$ and the one for $m=1$ by $\Phi_m$. Their difference gives the equality \begin{equation*}\begin{split} (\mu_m-\mu_1)\dfrac{y^{1-2s}}{r^2}\Phi_m\Phi_1 &=\Phi_1\divergence(y^{1-2s}\nabla\Phi_m)-\Phi_m\divergence(y^{1-2s}\nabla\Phi_1)\\ &=\divergence(y^{1-2s}(\Phi_1\nabla\Phi_m-\Phi_m\nabla\Phi_1)). \end{split}\end{equation*} Let us integrate over the region where $\Phi_m>0$. The boundary $\partial\set{\Phi_m>0}$ is decomposed into a disjoint union of $\partial^0\set{\Phi_m>0}$ and $\partial^+\set{\Phi_m>0}$, on which $y=0$ and $y>0$, respectively. Hence \begin{equation*}\begin{split} 0&\leq{d}_s(\mu_m-\mu_1)\int_{\set{\Phi_m>0}}\dfrac{y^{1-2s}\Phi_m\Phi_1}{r^2}\,dxdy\\ &=\int_{\partial^0\set{\Phi_m>0}}\left(\phi_1\lim\limits_{y\to0}y^{1-2s}\pnu{\Phi_m}-\phi_m\lim\limits_{y\to0}y^{1-2s}\pnu{\Phi_1}\right)\,dx\\ &+\int_{\partial^+\set{\Phi_m>0}}y^{1-2s}\left(\Phi_1\pnu{\Phi_m}-\Phi_m\pnu{\Phi_1}\right) \,dy. \end{split}\end{equation*} The first integral on the right hand side vanishes due to the equations $\Phi_1$ and $\Phi_m$ satisfy. Then we observe that on $\partial^+\set{\Phi_m>0}$, one has $\Phi_1>0$, $\pnu{\Phi_m}\leq0$ and $\Phi_m=0$. This forces (using $\mu_m>\mu_1$) \[\int_{\set{\Phi_m>0}}\dfrac{y^{1-2s}\Phi_m\Phi_1}{r^2}\,dxdy=0,\] which in turn implies $\Phi_m\leq0$. Similarly $\Phi_m\geq0$ and, therefore, $\Phi_m\equiv0$ for $m\geq{n+1}$. This completes the proof of the lemma. \end{proof} Now we turn to Fredholm properties. Let $L$ be the operator defined in \eqref{model-linearization}. It is actually simpler to consider the conjugate operator $\mathcal L$ defined in \eqref{eq0} (and its projection \eqref{eq0m}), which is better behaved in weighted Hilbert spaces and simplifies the notation in the proof of Fredholm properties. We define weighted $L^2_{\delta,\vartheta}$ function spaces. These contain $L^2_{\text{loc}}$ functions for which the norm \begin{equation}\label{L2-norm} \|\phi\|^2_{L^2_{\delta,\vartheta}(\mathbb R^n)}=\int_{\mathbb R^n\setminus B_1} |\phi|^2r^{ n-1-2s-2\vartheta}\,drd\theta +\int_{B_1}|\phi|^2r^{n-1-2s-2\delta}\,drd\theta \end{equation} is finite. These should be understood after conjugation \eqref{shift}, as \begin{equation*} \|\psi\|^2_{L^2_{\delta,\vartheta}(\mathbb R\times\mathbb S^{n-1})}=\int_{-\infty}^0\int_{\mathbb S^{n-1}} |\psi|^2e^{2\vartheta t}\,d\theta dt+\int_{0}^{\infty}\int_{\mathbb S^{n-1}}|\psi|^2e^{2\delta t}\,d\theta dt. \end{equation*} The spaces $L^2_{\delta,\vartheta}$ and $L^2_{-\delta,-\vartheta}$ are dual with respect to the natural pairing \begin{equation*} \langle\psi_1,\psi_2\rangle_*=\int_{\mathbb R^n} \psi_1 \psi_2, \end{equation*} for $\psi_1\in L^2_{\delta,\vartheta}$, $\psi_2\in L^2_{-\delta,-\vartheta}$. The relation between \eqref{L2-norm} and the weighted $L^\infty$ norms from \eqref{norms} is given by the following simple lemma: \begin{lemma}\label{lemma:inclusions} Assume that the parameters satisfy \begin{equation}\label{parameters} -\delta<-\sigma+\frac{n-2s}{2},\quad -\vartheta>-\frac{2s}{p-1}+\frac{n-2s}{2}, \end{equation} Then if $\|\phi\|_*<\infty$, one has that $\|\phi\|_{L^2_{-\delta,-\vartheta}}$ is also finite and the inclusion is continuous. \end{lemma} \begin{proof}[Proof of Proposition \ref{invertibility}] First note that elliptic estimates imply that $\mathcal L$ is a densely defined, closed graph operator. Moreover, the adjoint of \begin{equation*}\label{operator1} \mathcal L:L^2_{-\delta,-\vartheta}\to L^2_{-\delta,-\vartheta} \end{equation*} is precisely \begin{equation*}\label{operator2} \mathcal L^*=\mathcal L:L^2_{\delta,\vartheta}\to L^2_{\delta,\vartheta}. \end{equation*} Similarly to the arguments in Section 8 of \cite{acdfgw}, one can show that the linear operator $\mathcal L$ satisfies good Fredholm properties and, in particular, $$\Ker(\mathcal L^*)^\bot=\rango( \mathcal L).$$ By checking the proof for Lemma \ref{non-degeneracy}, one can also get that for this linear problem $\mathcal L^*\phi=0$, the mode $0$ and mode $m$ for $m\geq n+1$ all have trivial kernels. For mode $1$, there is a one dimensional solution spanned by $w'(r)$. It is easy to see that $w'(r)\in L^2_{\delta,\vartheta}$ iff $p<\frac{n+2s-1}{n-2s-1}$, if one chooses the parameters as in \eqref{parameters}, since \begin{equation*} \begin{split} \|\phi\|^2_{L^2_{\delta,\vartheta}(\mathbb R^n)}&=\int_{\mathbb R^n\setminus B_1} |\phi|^2r^{ n-1-2s-2\vartheta}\,drd\theta +\int_{B_1}|\phi|^2r^{n-1-2s-2\delta}\,drd\theta\\ &\sim \int_1^\infty r^{-\frac{4s}{p-1}-2}r^{n-1-2s-\frac{4s}{p-1}+n-2s}\,dr+\int_0^1 r^2r^{n-1-2s-2\sigma+n-2s}\,dr<+\infty \end{split} \end{equation*} for suitable small $\sigma>0$ if $p<\frac{n+2s-1}{n-2s-1}$. From the above argument and the Fredholm property we conclude that: \begin{itemize} \item If $\frac{n+2s}{n-2s}<p<\frac{n+2s-1}{n-2s-1}$, then \begin{equation*} L\phi=h \end{equation*} is solvable iff $\int_{\mathbb R^n}h\frac{\partial w}{\partial x_i}\,dx=0$ for $i=1, \cdots,n$. \item If $p>\frac{n+2s-1}{n-2s-1}$, then \begin{equation*} L\phi=h \end{equation*} is always solvable. \end{itemize} Moreover, the Fredholm property yields that \begin{equation}\label{L^2} \|\phi\|_{L^2_{\delta,\vartheta}(\mathbb R^n)}\leq C\|h\|_{L^2_{\delta-2s,\vartheta-2s}(\mathbb R^n)}. \end{equation} We will show that this estimate still holds in weighted $L^\infty$ norm, i.e. \eqref{bound} holds. As in \cite{acdfgw}, combining with the Green's representation formula, one has \begin{equation*} \phi(x)=-\int_{{\mathbb R}^n}G(x,y)pw^{p-1}\phi(y)\,dy-\int_{{\mathbb R}^n}G(x,y)h(y)\,dy. \end{equation*} First, for $|x|\leq 1$, \begin{equation*} \begin{split} \int_{{\mathbb R}^n}G(x,y)h(y)\,dy&\leq C\left[\int_{\{|y|<\frac{|x|}{2}\}}+\int_{\{\frac{|x|}{2}\leq |y|\leq 2|x|\}}+\int_{\{|y|>2|x|\}}\right]\frac{h(y)}{|x-y|^{n-2s}}\,dy\\ &\leq C\left[ \int_{\{|y|<\frac{|x|}{2}\}}\frac{\|h\|_{**}|y|^{-\sigma-2s}}{|x|^{n-2s}}\,dy+\int_{\{|y|\leq 3|x|\}}\frac{\|h\|_{**}|x|^{-\sigma-2s}}{|y|^{n-2s}}\,dy\right.\\ &\quad+\left.\int_{\{2>|y|>2|x|\}} \frac{\|h\|_{**}|y|^{-\sigma-2s}}{|y|^{n-2s}} \,dy+\int_{\{|y|\geq 2\}}\frac{\|h\|_{**}|y|^{-\frac{2s}{p-1}-2s}}{|y|^{n-2s}}\,dy\right]\\ &\leq C|x|^{-\sigma}\|h\|_{**}. \end{split} \end{equation*} Next, by the definition of weighted $L^2$ norm and relations \eqref{parameters} and \eqref{L^2}, \begin{equation*} \begin{split} \int_{{\mathbb R}^n}&G(x,y)pw^{p-1}\phi(y)\,dy\\&\leq C\left[\int_{\{|y|<\frac{|x|}{2}\}}+\int_{\{\frac{|x|}{2}\leq |y|\leq 2|x|\}}+\int_{\{2>|y|>2|x|\}}\frac{\phi}{|x-y|^{n-2s}}\,dy+\int_{\{|y|>2\}}\frac{|y|^{-2s}\phi}{|x-y|^{n-2s}}\,dy \right]\\ &\leq C\Big( \int_{\{|y|<2\}}\phi^2 r^{-2s-2\delta}\,dy\Big)^{\frac{1}{2}}\\ &\quad\cdot \left[ \Big(\int_{\{|y|<\frac{|x|}{2}\}}\frac{|y|^{2s+2\delta}}{|x|^{2(n-2s)}}\,dy\Big)^{\frac{1}{2}} +\Big(\int_{\{|y|<3|x|\}}\frac{|x|^{2s+2\delta}}{|y|^{2(n-2s)}}\,dy\Big)^{\frac{1}{2}} +(\int_{\{2|x|<|y|<2\}} \frac{|y|^{2s+2\delta}}{|y|^{2(n-2s)}}\, dy\Big)^{\frac{1}{2}}\right]\\ &+C\Big(\int_{\{|y|>2\}}\phi^2r^{-2s-2\vartheta}dy\Big)^{\frac{1}{2}}\Big(\int_{\{|y|>2\}} \frac{|y|^{-4s+2s+2\vartheta}}{|y|^{2(n-2s)}}\,dy \Big)^{\frac{1}{2}}\\ &\leq C\|\phi\|_{L_{\delta,\vartheta}^2({\mathbb R}^n)}|x|^{-\sigma}\leq C\|h\|_{L_{\delta-2s,\vartheta-2s}^2({\mathbb R}^n)}|x|^{-\sigma}\\ &\leq C\|h\|_{**}|x|^{-\sigma}, \end{split} \end{equation*} for suitable $\delta, \vartheta$ satisfying \eqref{parameters}. Thus one has \begin{equation*} \sup_{\{|x|\leq 1\}}|x|^{\sigma}|\phi|\leq C\|h\|_{**}. \end{equation*} Similarly, for $|x|\leq e^{R}$, we still obtain \begin{equation*} \sup_{\{1\leq |x|\leq e^R\}}|x|^{\frac{2s}{p-1}}|\phi|\leq C_R\|h\|_{**}, \end{equation*} for some constant depending on $R$. This implies that the weighted $L^\infty$ norm of $\phi$ in any compact set can be bounded by the weighted $L^\infty$ norm of $h$. So one only needs to worry about the norm at infinity. For this, we go back to the projected problems: \begin{equation*} \mathcal L_{\kappa,m} \psi_m:=P_s^{(m)}(\psi_m)-\kappa\psi_m=\tilde{h}_m+(\mathcal V-\kappa)\psi_m, \quad m=0,1,\ldots. \end{equation*} In order to show the estimate for $\phi$ at infinity, it is enough to prove that \begin{equation*} \|e^{\frac{-2s}{p-1}t}\psi_m\|_{L^\infty(\{t\leq -R\})}\leq C\|e^{\frac{-2s}{p-1}t}\tilde{h}_m\|_{L^\infty(\{t\leq 0\})}+\|e^{-\sigma t}\tilde{h}_m\|_{L^\infty(\{t\geq 0\})} \end{equation*} for $R$ large enough. But this follows from Theorem \ref{thm:Hardy-potential} and the expression for $\mathcal V$ in (\ref{asymptotics-potential}) by taking $\delta=\frac{2s}{p-1}$ and $\delta_0=-\sigma$ (See Lemma 6.7 and the proof for Proposition 6.3 in \cite{acdfgw}). Here we use the assumption on $\sigma$ that $0<\sigma<\frac{2s}{p-1}$ such that $\delta+\delta_0>0$. This completes the proof of the Proposition. \end{proof} \section{The operator $(-\Delta)^s \phi+V_\lambda \phi-pw^{p-1}\phi$ in ${\mathbb R}^n$} In this section we study the following linear problem in ${\mathbb R}^n$ in suitable weighted spaces: \begin{equation}\label{lp} \begin{cases} (-\Delta)^s \phi+V_\lambda \phi-pw^{p-1}\phi =h+\sum\limits_{i=1}^n c_iZ_i,\\ \lim\limits_{|x|\to \infty}\phi(x)=0, \end{cases} \end{equation} where $c_i$ are real constants, and $Z_i$ are the kernels defined by \begin{equation}\label{kernel} Z_i(x)=\frac{\partial w}{\partial x_i}, \ i=1,\cdots,n, \end{equation} and \begin{equation*} V_\lambda(x)=\lambda^{-2s}V\Big(\frac{x-\xi}{\lambda}\Big) \mbox{ for }\lambda>0, \ \xi \in {\mathbb R}^n. \end{equation*} Since $V_\lambda$ has a concentration point at $\xi$, we define \begin{equation*} \begin{split} \|\phi\|_{*,\xi}&=\sup_{\{|x-\xi|\leq 1\}}|x-\xi|^\sigma |\phi(x)|+\sup_{\{|x-\xi|>1\}}|x-\xi|^{\frac{2s}{p-1}}|\phi(x)|,\\ \|h\|_{**,\xi}&=\sup_{\{|x-\xi|\leq 1\}}|x-\xi|^{\sigma+2s} |h(x)|+\sup_{\{|x-\xi|>1\}}|x-\xi|^{\frac{2s}{p-1}+2s}|h(x)|. \end{split} \end{equation*} We will take $|\xi|\leq \Lambda$ for some $\Lambda>0$. For the linear theory, it suffices to assume that \begin{equation}\label{assumptionv} V\in L^\infty({\mathbb R}^n), \quad V\geq 0, \quad V(x)=o(|x|^{-2s}) \mbox{ as }|x|\to \infty. \end{equation} Then we have the following solvability result: \begin{proposition}\label{pro1} We have: \begin{itemize} \item If $p>\frac{n+2s-1}{n-2s-1}$, for $\lambda>0$ small enough, equation \eqref{lp} has a solution $\phi:=\mathcal{T}_\lambda(h)$ with $c_i=0$, and it satisfies \begin{equation*} \|\mathcal{T}_\lambda(h)\|_{*,\xi}\leq C\|h\|_{**,\xi}. \end{equation*} \item If $\frac{n+2s}{n-2s}<p<\frac{n+2s-1}{n-2s-1}$, for $\lambda>0$ small enough, equation \eqref{lp} has a solution $(\phi,c_1,\cdots,c_n):=\mathcal{T}_\lambda(h)$, and it satisfies \begin{equation*} \|\phi\|_{*,\xi}+\sum_{i=1}^n|c_i|\leq C\|h\|_{**,\xi}. \end{equation*} The constant $C>0$ is independent of the parameter $\lambda$. \end{itemize} \end{proposition} \begin{proof} The proof follows the argument in Section 3 of \cite{ddmw}. So here we only sketch the proof. We solve the linear problem near the point $\xi$ and away from this point. For this, decompose $\phi=\varphi+\psi$, where $\varphi, \psi$ satisfy \begin{equation}\label{varphi} \begin{cases} \displaystyle(-\Delta)^s\varphi-pw^{p-1}\varphi=p\xi_0w^{p-1}\psi-\xi_1V_\lambda\varphi+\xi_1h+\sum_{i=1}^nc_iZ_i,\\ \lim\limits_{|x|\to \infty}\varphi(x)=0, \end{cases} \end{equation} and \begin{equation}\label{psi} \begin{cases} (-\Delta)^s\psi-p(1-\xi_0)w^{p-1}\psi+V_\lambda\psi=-(1-\xi_1)V_\lambda\varphi+(1-\xi_1)h,\\ \lim\limits_{|x|\to \infty}\varphi(x)=0, \end{cases} \end{equation} where $\xi_0, \xi_1$ are two cut off functions: \begin{equation*} \xi_0(x)=0 \mbox{ for }|x-\xi|\leq R, \quad \xi_0(x)=1 \mbox{ for }|x-\xi|\geq 2R, \end{equation*} and \begin{equation*} \xi_1(x)=0 \mbox{ for }|x-\xi|\leq \varrho, \quad \xi_1(x)=1 \mbox{ for }|x-\xi|\geq 2\varrho, \end{equation*} for $\varrho,R$ two positive constants independent of $\lambda$ to be fixed later and such that $2\varrho\leq R$. Given $\|\varphi\|_{*,\xi}<\infty$, since $\|p(1-\xi_0)w^{p-1}\|_{L^{\frac{n}{2s}}}\to 0$ as $R\to 0$, equation \eqref{psi} has a solution $\psi=\psi(\varphi)$ if $R>0$ is small enough (this is because the homogeneous problem for \eqref{psi} has only the trivial solution). Moreover, $\psi(x)=O(|x|^{-(n-2s)})$ as $|x|\to \infty$, so the right hand side of \eqref{varphi} has finite $\|\cdot\|_{**}$ norm, by Proposition \ref{invertibility}, \eqref{varphi} has a solution when $\psi=\psi(\varphi)$ which we write as $F(\varphi)$. We claim that $F(\varphi)$ has a fixed point in the Banach space \begin{equation*} X=\{\varphi\in L^\infty({\mathbb R}^n), \|\varphi\|_*<\infty\} \end{equation*} equipped with \begin{equation*} \|\varphi\|_X=\sup_{\{|x|\leq 1\}}|\varphi|+\sup_{\{|x|\geq 1\}}|x|^{\frac{2s}{p-1}}|\varphi|. \end{equation*} Following the argument in the proof of Proposition 3.1 in \cite{ddmw}, we establish pointwise estimates for the solution $\psi$ of \eqref{psi}. Then we can find a bound of the $\|\cdot\|_{**}$ norm of the right hand side of \eqref{varphi}. Since the proof is similar, we omit the details and just state the estimates here: \begin{equation*} \begin{split} \|\psi\|_{*,\xi}&\leq (C_\varrho+C\varrho^{n-2s})\|\varphi\|_X+C_\varrho\|h\|_{**,\xi},\\ \|\xi_0w^{p-1}\psi(\varphi)\|_{**,\xi}&\leq C\varrho^{n-2s}\|\varphi\|_X+C_\varrho\|h\|_{**,\xi},\\ \|\xi_1 V_\lambda \varphi\|_{**,\xi}&\leq C\varrho^{-2s}\|\varphi\|_X a\big(\frac{\varrho}{\lambda}\big), \end{split} \end{equation*} where \begin{equation*} a(r)=\sup_{\{|x|\geq r\}}|x|^{2s}V(x), \quad a(r) \to 0\mbox{ as }r\to \infty. \end{equation*} By the linear theory in Section 2, we know that given $\varphi\in X$, the solution $F(\varphi)$ to (\ref{varphi}) satisfies \begin{equation*} \|F(\varphi)\|_{*,\xi}\leq C\|\xi_0w^{p-1}\psi(\varphi)\|_{**,\xi}+C\|\xi_1 V_\lambda \varphi\|_{**,\xi}+C\|\xi_1h\|_{**,\xi}. \end{equation*} But since the right hand side of (\ref{varphi}) is bounded near the origin, by regularity estimates, we derive \begin{equation*} \|F(\varphi)\|_X\leq C\|\xi_0w^{p-1}\psi(\varphi)\|_{**,\xi}+C\|\xi_1 V_\lambda \varphi\|_{**,\xi}+C\|\xi_1h\|_{**,\xi} \end{equation*} and \begin{equation*} \begin{split} \|F(\varphi_1)-F(\varphi_2)\|_X&\leq C\varrho^{n-2s}\|\varphi_1-\varphi_2\|_X+C a(\frac{\varrho}{\lambda})\varrho^{-2s}\|\varphi_1-\varphi_2\|_X\\ &\leq C\big(\varrho^{n-2s}+\varrho^{-2s}a\big(\frac{\varrho}{\lambda}\big)\big)\|\varphi_1-\varphi_2\|_X. \end{split} \end{equation*} By choosing $\varrho>0 $ small enough and $\lambda$ small enough such that $\frac{\varrho}{\lambda}\to \infty$, we can prove that $F(\varphi)$ is a contraction mapping, and we get a fixed point $\varphi\in X$. Moreover, thanks to the linear theory in Section \ref{section:linear}, we have \begin{equation*} \|\varphi\|_X\leq C\big(\varrho^{n-2s}+\varrho^{-2s}a\big(\frac{\varrho}{\lambda}\big)\big)\|\varphi\|_X+C_\varrho \|h\|_{**,\xi}, \end{equation*} which yields \begin{equation*} \|\varphi\|_X\leq C\|h\|_{**,\xi}. \end{equation*} Combining with the estimate for $\psi$, one has \begin{equation*} \|\phi\|_{*,\xi}\leq C\|h\|_{**,\xi} \end{equation*} for some $C>0$ independent of $\lambda $ small. \end{proof} \section{Proof of Theorem \ref{thm1}} Assume that $p>\frac{n+2s-1}{n-2s-1}$. We aim to find a solution to \eqref{problem-1} of the form $u=w+\phi$, where $w$ is the radial entire solution found in Section \ref{section:entire} (recall that $w$ satisfies \eqref{w}). This yields the following equation for the function $\phi$, \begin{equation*} (-\Delta)^s\phi+V_\lambda \phi-pw^{p-1}\phi=N(\phi)-V_\lambda w, \end{equation*} where \begin{equation*} N(\phi)=(w+\phi)^p-w^p-pw^{p-1}\phi. \end{equation*} Using the operator $\mathcal{T}_\lambda$ defined in Proposition \ref{pro1}, we are led to solve the fixed point problem: \begin{equation*}\label{fixedpoint} \phi=\mathcal{T}_\lambda(N(\phi)-V_\lambda w):=\mathcal{A}(\phi). \end{equation*} Following the argument in Section 4 of \cite{ddmw}, one obtains the following estimate for $N(\phi)$ and $V_\lambda w$: \begin{equation*} \|V_\lambda w\|_{**}\leq C\Big(\lambda^\sigma R^{2s+\sigma}\|V\|_{L^\infty}+a(R)+a\big(\frac{1}{\lambda}\big)\Big)\leq Ca(R) \end{equation*} as $\lambda\to 0$. Choosing $R\to \infty$, then as $\lambda\to 0$ one has that $\|V_\lambda w\|_{**}\to 0$. In addition, for the nonlinear term $N(\phi)$ we have the bound \begin{equation*} \|N(\phi)\|_{**}\leq C(\|\phi\|_*^2+\|\phi\|_*^p). \end{equation*} Consider the set \begin{equation*} \mathcal{F}=\{\phi:{\mathbb R}^n\to {\mathbb R}, \|\phi\|_*\leq \rho\} \end{equation*} where $\rho>0$ small is to be chosen later. It is standard to get the following estimates: \begin{equation*} \begin{split} \|\mathcal{A}(\phi)\|_*&\leq C(\|\phi\|_*^2+\|\phi\|_*^p+\|V_\lambda w\|_{**})<\rho,\\ \|\mathcal{A}(\phi_1)-\mathcal{A}(\phi_2)\|_*&\leq C(\|\phi_1\|^{\min\{p-1,1\}}_*+\|\phi_2\|^{\min\{p-1,1\}}_*)\|\phi_1-\phi_2\|_*<\frac{1}{2}\|\phi_1-\phi_2\|_* \end{split} \end{equation*} for $\lambda $ small and $\rho$ small. One can find that for $\rho$ small enough, $\mathcal{A}(\phi)$ is a contraction mapping in $\mathcal{F}$, thus has a fixed point in this set. This finishes the proof of the theorem. \qed \section{The case $\frac{n+2s}{n-2s}<p\leq \frac{n+2s-1}{n-2s-1}$} In this case, because of the presence of the kernels $Z_i$ defined in \eqref{kernel}, one needs to introduce free parameters and rescale around a point $\xi$, to be chosen later. For this reason, we make the change of variable $\lambda^{-\frac{2s}{p-1}}u(\frac{x-\xi}{\lambda})$ and look for a solution of the form $u=w+\phi$, then $\phi$ satisfies \begin{equation*} (-\Delta)^s \phi+V_\lambda \phi-pw^{p-1}\phi=N(\phi)-V_\lambda w, \end{equation*} where \begin{equation*} V_\lambda(x)=\lambda^{-2s}V\big(\frac{x-\xi}{\lambda}\big). \end{equation*} Similarly to the proof of Lemma 5.1 in \cite{ddmw} can show the following result, for which we omit the proof: \begin{lemma} Let $\frac{n+2s}{n-2s}<p<\frac{n+2s-1}{n-2s-1}$ and $\Lambda>0$. Assume that $V$ satisfies \eqref{assumptionv}. Then there exists $\ve_0>0$ such that for $|\xi|<\Lambda$ and $\lambda<\ve_0$, there exists $(\phi, c_1, \cdots, c_n) $ solution of \begin{equation}\label{nonlinear1} \left\{\begin{array}{l} \displaystyle(-\Delta)^s \phi+V_\lambda \phi-pw^{p-1}\phi=N(\phi)-V_\lambda w+\sum_{i=1}^n c_iZ_i,\\ \displaystyle\lim_{|x|\to \infty}\phi(x)=0. \end{array} \right. \end{equation} We have, in addition, \begin{equation*} \|\phi\|_{*,\xi}+\sum_{i=1}^n |c_i|\to 0 \mbox{ as }\lambda\to 0. \end{equation*} If $V$ also satisfies \begin{equation*}\label{assumptionv1} V(x)\leq C|x|^{-\mu} \mbox{ for \ all }x \end{equation*} for some $\mu>2s$, then for $0<\sigma\leq \mu-2s, \sigma<n-2s$, one has \begin{equation*} \|\phi\|_{*,\xi}\leq C_\sigma \lambda^\sigma \mbox{ for \ all }0<\lambda<\ve_0. \end{equation*} \end{lemma} \bigskip Now we are ready for the proof of our second main theorem: \begin{proof}[Proof of Theorem \ref{thm2}.] We have found a solution $(\phi,c_1,\cdots, c_n)$ to equation \eqref{nonlinear1}. By the linear theory in Section 2, this solution satisfies \begin{equation*} \int_{{\mathbb R}^n}\Big(N(\phi)-V_\lambda\phi-V_\lambda w+\sum_{i=1}^n c_iZ_i\Big)\frac{\partial w}{\partial x_j}\,dx=0 \end{equation*} for $j=1,\cdots, n$. So for $\lambda$ small, we need to choose the parameter $\xi$ such that $c_i=0$ for all $i$, that is \begin{equation}\label{reduceproblem} \int_{{\mathbb R}^n}\Big(N(\phi)-V_\lambda\phi-V_\lambda w\Big)\frac{\partial w}{\partial x_j}\,dx=0 \end{equation} because the matrix with coefficients $\int_{{\mathbb R}^n }Z_i\frac{\partial w}{\partial x_j}\,dx$ is invertible. {\bf Case a.} Since $V\leq |x|^{-\mu}$ for $\mu>n$, the dominant term in \eqref{reduceproblem} is \begin{equation*} \lambda^{-2s}\int_{{\mathbb R}^n}V\big(\frac{x-\xi}{\lambda}\big)w\frac{\partial w}{\partial x_j}\,dx=O(\lambda^{n-2s}). \end{equation*} Using the estimates for $\phi$ in the last section, one can get that \begin{equation*} \int_{{\mathbb R}^n}N(\phi)\frac{\partial w}{\partial x_j}\,dx=o(\lambda^{n-2s}). \end{equation*} For the term $\int V_\lambda \phi \frac{\partial w}{\partial x_j}\,dx$, one has \begin{equation*} \begin{split} \int V_\lambda \phi \frac{\partial w}{\partial x_j}\,dx&=\int_{B_{R\lambda}}\cdots+\int_{B_1 \setminus B_{R\lambda}}\cdots+ \int_{{\mathbb R}^n\setminus B_1}\cdots \\ &\leq C\int_{B_{R\lambda}}\lambda R\cdot \lambda^{-2s}\cdot \lambda^\sigma |x|^{-\sigma }\,dx+\int_{B_1\setminus B_{\lambda R}}\lambda^\sigma |x^{-\sigma}|\cdot a(R)|x|^{-2s}\lambda |x|\,dx\\ &\quad +\int_{{\mathbb R}^n\setminus B_1}\lambda^\sigma a\big(\frac{1}{\lambda}\big)|x|^{-(2s+\frac{4s}{p-1}+1)}\,dx\\ &\leq o(\lambda^{n-2s}) \end{split} \end{equation*} as $\lambda \to 0$, where we have used the fact that $\frac{n+2s}{n-2s}<p<\frac{n+2s-1}{n-2s-1}$. We now set \begin{equation*} F_\lambda^{(j)}(\xi)=\lambda^{-2s}\int_{{\mathbb R}^n}V\big(\frac{x}{\lambda}\big)u_\lambda\frac{\partial w}{\partial x_j}(x+\xi)\,dx+\int_{{\mathbb R}^n}N(\phi)\frac{\partial w}{\partial x_j}(x+\xi)\,dx. \end{equation*} Fix $\rho>0$ small. For $|\xi|=\rho$, and $\lambda$ small, one can check that \begin{equation*} \langle F_\lambda(\xi), \xi\rangle \sim \langle \nabla^2w(0) \xi,\xi\rangle<0 \mbox{ for }|\xi|=\rho, \end{equation*} since $0$ is local maximum point of $w$. By degree theory, we deduce that $F_\lambda$ has a zero in $B_\rho $. \medskip {\bf Case {b}(i). } Assume that $\displaystyle\lim_{|x|\to \infty}\Big(|x|^\mu V-f\big(\frac{x}{|x|}\big)\Big)=0$ for $N-\frac{4s}{p-1}<\mu<n, \ f\neq 0.$ Here one can check that the dominant term of \eqref{reduceproblem} is \begin{equation*} \begin{split} &\lambda^{-2s}\int_{{\mathbb R}^n}V(\frac{x-\xi}{\lambda})w\frac{\partial w}{\partial x_j}\,dx\\ &=\lambda^{\mu-2s}\int_{R^n}|x|^{-\mu}f\Big(\frac{x}{|x|}\Big)w(x+\xi)\frac{\partial w}{\partial x_j}(x+\xi)\,dx+o(\lambda^{\mu-2s})\\ &=O(\lambda^{\mu-2s}). \end{split} \end{equation*} Again, using the estimates for $\phi$, one has \begin{equation*} \begin{split} \int_{{\mathbb R}^n}N(\phi)\frac{\partial w}{\partial x_j}\,dx&=O(\lambda^{\mu+\sigma-2s})=o(\lambda^{\mu-2s}),\\ \int_{{\mathbb R}^n}V_\lambda \phi \frac{\partial w}{\partial x_j}\,dx&\leq C\lambda^{\mu+\sigma-2s}. \end{split} \end{equation*} So now define \begin{equation*} \begin{split} \tilde{F}_j(\xi)&:=\frac{1}{2}\int_{R^n}|x|^{-\mu}f\Big(\frac{x}{|x|}\Big)w^2(x+\xi)\,dx\\ &=\frac{1}{2}\beta^{\frac{2s}{p-1}}|\xi|^{-\frac{4s}{p-1}+n-\mu}\int_{{\mathbb R}^n}|y|^{-\mu} f\Big(\frac{y}{|y|}\Big)\Big|y+\frac{\xi}{|\xi|}\Big|^{-\frac{4s}{p-1}}\,dy+o(|\xi|^{n-\mu-\frac{4s}{p-1}}). \end{split} \end{equation*} Similarly, \begin{equation*} \begin{split} \nabla \tilde{F}(\xi)\cdot \xi&=\frac{1}{2}\Big(n-\mu-\frac{4s}{p-1}\Big)\beta^{\frac{2s}{p-1}}|\xi|^{-\frac{4s}{p-1}+n-\mu} \int_{{\mathbb R}^n}|y|^{-\mu} f\Big(\frac{y}{|y|}\Big)\Big|y+\frac{\xi}{|\xi|}\Big|^{-\frac{4s}{p-1}}\,dy+o(|\xi|^{n-\mu-\frac{4s}{p-1}}). \end{split} \end{equation*} Therefore \begin{equation*} \nabla \tilde{F}(\xi)\cdot \xi<0 \mbox{ \ for all } |x|=R \mbox{ \ for large }R. \end{equation*} Using degree theory, we get the existence of $\xi $ in $B_R$ such that $c_i=0$ for all $i$. \medskip {\bf Case b(ii).} $\displaystyle\lim_{|x|\to \infty}\Big(|x|^n V-f\big(\frac{x}{|x|}\big)\Big)=0, f\neq 0.$ In this case, we have \begin{equation*} \begin{split} G_j(\xi)&:=\int_{{\mathbb R}^n}\Big(N(\phi)-V_\lambda\phi-V_\lambda w\Big)\frac{\partial w}{\partial x_j}\,dx\\ &=\lambda^{-2s}\int_{{\mathbb R}^n}V(\frac{x}{\lambda})u_\lambda(x+\xi)\frac{\partial w}{\partial x_j}(x+\xi)\,dx+o(\lambda^{n-2s}). \end{split} \end{equation*} We claim that $\langle G(\xi),\xi\rangle <0$ for all $|\xi|=\rho$ for $\rho>0$ small enough. Once this is true, using degree theory, we conclude that for some $|\xi|<\rho$ we have $G(\xi)=0$, which finishes the proof. In order to prove this claim, note that for $\rho>0$ small, one has for all $|\xi|=\rho$, \begin{equation*} \langle \nabla w(\xi),\xi\rangle<0 \end{equation*} Thus for $\delta>0$ small but fixed, \begin{equation*} \gamma:=\sup_{x\in B_\delta}\langle \nabla w(\xi),\xi\rangle <0 \mbox{ for \ all }|\xi|=\rho. \end{equation*} Then similarly to the proof of part (b.2) in Section 5 of \cite{ddmw}, one has \begin{equation*} \begin{split} \Big|\lambda^{-2s}\int_{{\mathbb R}^n \setminus B_\delta}V\Big(\frac{x}{\lambda}\Big)u_\lambda(x+\xi)\langle \nabla w(x+\xi),\xi \rangle\Big|&\leq C\lambda^{n-2s},\\ \Big|\lambda^{-2s}\int_{B_{\lambda R}}V\Big(\frac{x}{\lambda}\Big)u_\lambda(x+\xi)\langle \nabla w(x+\xi),\xi \rangle\Big|&=O(\lambda^n).\\ \end{split} \end{equation*} Then \begin{equation*} \begin{split} &\Big|\lambda^{-2s}\int_{B_\delta\setminus B_{\lambda R}}V(\frac{x}{\lambda})u_\lambda(x+\xi)\langle \nabla w(x+\xi),\xi \rangle\Big|\leq c\gamma\int_{B_\delta\setminus B_{\lambda R}}V\Big(\frac{x}{\lambda}\Big),\\ &\int_{B_\delta\setminus B_{\lambda R}}V\Big(\frac{x}{\lambda}\Big) =\int_{B_\delta\setminus B_{\lambda R}}f\Big(\frac{x}{|x|}\Big)|x|^{-n}\,dx +\int_{B_\delta\setminus B_{\lambda R}}|x|^{-n}\Big(V(x)|x|^n-f\Big(\frac{x}{|x|}\Big)\Big)\,dx,\\ &\int_{B_\delta\setminus B_{\lambda R}}f\Big(\frac{x}{|x|}\Big)|x|^{-n}\,dx=\log \frac{1}{\lambda}\int_{\mathbb S^{n-1}}f+O(1),\\ &\int_{B_\delta\setminus B_{\lambda R}}|x|^{-n}\Big(V(x)|x|^n-f\Big(\frac{x}{|x|}\Big)\Big)\,dx\leq \ve \log\frac{1}{\lambda}, \end{split} \end{equation*} for $\ve$ small enough if $R$ is large enough. Combining all the above estimates, one obtains that $\langle G(\xi),\xi\rangle <0$ for all $|\xi|=\rho$ small.\\ \end{proof} \begin{remark} For $p=\frac{n+2s-1}{n-2s-1}$, we can get the same solvability result as in case \emph{ii.} of Proposition \ref{invertibility} if one considers the new weighted norm \begin{equation*} \begin{split} \|\phi\|_*&=\sup_{\{|x|\leq 1\}}|x|^\sigma |\phi(x)|+\sup_{\{|x|\geq 1\}}|x|^{\frac{2s}{p-1}+\alpha}|\phi(x)|,\\ \|h\|_{**}&=\sup_{\{|x|\leq 1\}}|x|^{\sigma+2s} |h(x)|+\sup_{\{|x|\geq 1\}}|x|^{\frac{2s}{p-1}+\alpha+2s}|h(x)|, \end{split} \end{equation*} for some $\alpha>0$ small. In this new norm, one can also show that $w'$ is in the kernel space of $L^*$ even when $p=\frac{n+2s-1}{n-2s-1}$. With some minor modifications, we can also obtain existence of solutions. \end{remark} \bigskip \noindent\textbf{Acknowledgements.} M.d.M. Gonz\'alez is supported by Spanish government grants MTM2014-52402-C3-1-P and MTM2017-85757-P, and the BBVA foundation grant for Researchers and Cultural Creators, 2016. The research of H. Chan and J. Wei is supported by NSERC of Canada.

Introduction ============ Dilated cardiomyopathy (DCM), a primary cardiomyopathy with significant enlargement of the ventricle ([@b1-mmr-22-02-0915]), is one of the most intractable diseases in the cardiovascular field ([@b2-mmr-22-02-0915]). DCM can progressively develop into severe congestive heart failure and seriously threatens the survival rate of the patient ([@b3-mmr-22-02-0915]). However, due to its unknown etiology and underlying mechanisms, DCM cannot be efficiently treated by existing therapeutic strategies, except for heart transplantation ([@b4-mmr-22-02-0915]). Thus, identifying the specific molecular mechanism of DCM is important to facilitate its diagnosis and treatment, and could improve the prognosis of patients with DCM. With the development of databases such as The Cancer Genome Atlas and the Gene Expression Omnibus (GEO), genetic research on DCM has become feasible and is currently ongoing ([@b5-mmr-22-02-0915]). Huang *et al* ([@b6-mmr-22-02-0915]) analyzed 102 samples from the GEO database (GSE5406), and identified module and hub genes of differentially expressed genes (DEGs) that are related to the progression of DCM. In addition, Xiao *et al* ([@b7-mmr-22-02-0915]) used RNA-Seq data (GSE116250) and gene annotation of the Ensembl database to find the key module involved in DCM, and identified five co-expression modules that may have important functions in DCM occurrence. Thus, these previous studies provide possible research directions for elucidating the pathogenesis of DCM. However, due to limited sample quantities, insufficient analysis methods and the lack of experimental verification, a number of previous DEG results for DCM may not be accurate ([@b8-mmr-22-02-0915]). Therefore, to define the related genes in DCM, the present study performed an integrated bioinformatics analysis on four DCM microarrays from GEO, and a protein-protein interaction (PPI) network was established to investigate the interactions among DEGs. Then, the present study examined these hub genes *in vivo* using reverse transcription-quantitative PCR (RT-qPCR) in a mouse DCM model, which was established by intraperitoneal injection of doxorubicin (DOX). Materials and methods ===================== ### Raw transcriptional data acquisition and preprocessing 'Dilated cardiomyopathy' and 'human' were used as key words to search the GEO database (<http://www.ncbi.nlm.nih.gov/geo/>), and raw data (CEL files) from five datasets (GSE17800, GSE21610, GSE42955, GSE79962 and GSE3585), consisting of myocardium transcriptional expression data from 89 patients with DCM and 37 healthy controls were downloaded (details in [Table I](#tI-mmr-22-02-0915){ref-type="table"}) for further analysis. Raw data were first normalized using the RMA function in the R package affy (v1.64.0) ([@b9-mmr-22-02-0915]). Probes in GSE17800, GSE21610 and GSE79962 were then screened by the MAS5CALLS algorithm packaged in affy, and probes detected as 'absent' across all samples were filtered out. The R packages hgu133plus2.db (v3.2.3) ([@b10-mmr-22-02-0915]), hugene10sttranscriptcluster.db (v8.7.0) ([@b11-mmr-22-02-0915]) and hgu133a.db (v3.2.3) ([@b12-mmr-22-02-0915]) were used to annotate probes with gene symbols according to the microarray platform. ### Meta-analysis of DEGs The expression matrixes of GSE17800, GSE21610, GSE42955 and GSE79962 were merged by common gene symbols. MetaMA (v 3.1.2) ([@b13-mmr-22-02-0915]) is an R package for microarray DEG meta-analyses that uses moderated effect size and P-value combinations, while RankProd (v3.12.0) ([@b14-mmr-22-02-0915]) uses a non-parametric rank product method. The merged expression matrix was processed using metaMA and RankProd according to the reference manuals, and overlapping genes from the two methods were deemed as DEGs. ### Functional annotation of the DEGs Lists of upregulated DEGs and downregulated DEGs were uploaded to the Database for Annotation, Visualization and Integrated Discovery (DAVID; v6.8) ([@b15-mmr-22-02-0915],[@b16-mmr-22-02-0915]) website for Gene Ontology (GO, v2019-07) ([@b17-mmr-22-02-0915]) and Kyoto Encyclopedia of Genes and Genomes (KEGG, v91.0) ([@b18-mmr-22-02-0915]) pathway enrichment analysis. The results were downloaded from the website and then visualized using the R package ggplot2 (v3.2.0) ([@b19-mmr-22-02-0915]). ### PPI network construction The DEG list was uploaded to the Search Tool for the Retrieval of Interacting Genes/Proteins (STRING; <https://string-db.org/cgi/input.pl>; v11.0) ([@b20-mmr-22-02-0915]) database, and then a PPI network was established with the minimum required interaction score set as the highest confidence (\>0.9). Cytoscape (<https://cytoscape.org>, v3.7.1) was used to visualize the PPI network, and the MCODE (v1.6) ([@b21-mmr-22-02-0915]) plug-in was used to screen important modules of the network. ### Identification of hub genes CytoHubba (v0.1) ([@b22-mmr-22-02-0915]), a plug-in of Cytoscape software, was used to identify hub genes of the PPI network via four different algorithms. The algorithms used for analysis included degree, Maximum Neighborhood Component (MNC), Density of Maximum Neighborhood Component (DMNC) and Maximal Clique Centrality (MCC). A Venn diagram was constructed and consisted of genes ranked in the top 20 of each method, and genes overlapping in the four groups were deemed hub genes. ### DCM animal model establishment All animal studies were approved by the Animal Care and Ethics Committee of Nanjing Medical University. The animal experiments were performed according to the Guide for the Care and Use of Laboratory Animals ([@b23-mmr-22-02-0915]). The DCM mouse model was established by intraperitoneal injection of DOX as previously described ([@b24-mmr-22-02-0915],[@b25-mmr-22-02-0915]). A total of 12 male C57BL/6 mice (age, 8 weeks; weight, 22--24 g) were purchased from Shanghai SLAC Laboratory Animal Co., Ltd. All mice were housed with sterile rodent chow and water available *ad libitum* in the animal holding room maintained at 26°C and 50% relative humidity, under a 12-h light/dark cycle. Mice were randomly divided into two groups: Control (n=6) and DOX (n=6) groups. Mice in the DOX group were intraperitoneally injected with DOX hydrochloride (cat. no. HY-15142; MedChemExpress; 4 mg/kg) once a week for 5 consecutive weeks. The control group received an equal volume (0.1 ml) of normal saline (0.9%). Mice that presented with progressive cardiac functional decline and ventricular dilation were successfully established as the DCM mouse model. ### Transthoracic echocardiography Then, 7 days after last injection, mice were sedated with 5% isoflurane-O~2~ (cat. no. 792632; Sigma-Aldrich; Merck KGaA) balanced mixture (maintained at 1.5%). Transthoracic echocardiography (Vevo 1100; VisualSonics, Inc.) was performed in M-mode using a 30-MHz transducer immediately after anesthetization. The left ventricular (LV) echocardiogram was assessed in both parasternal long-axis and short-axis views. End-systolic and end-diastolic dimensions were defined as the phases corresponding to the electrocardiogram T wave and to the R wave, respectively ([@b26-mmr-22-02-0915]). M-mode LV internal diameter end systole/diastole (LVIDs/d), LV posterior wall end diastole (LVPWd) and interventricular septal end diastole (IVSd) were averaged from 3--5 beats. Ejection fraction and fractional shortening were calculated as previously described ([@b26-mmr-22-02-0915]). ### RNA extraction and RT-qPCR After echocardiography detection all mice were anaesthetized by breathing a 5% isoflurane-O~2~ balanced mixture (maintained at 1.5%). Mice were confirmed to be deeply anesthetized after they were immobile for 1 min. Then, mice were euthanatized by inhalation of 25% CO~2~, until respiratory and cardiac arrest occurred. Hearts were isolated and ground into 1 mm^3^ pieces, and then dissolved in pure TRI Reagent^®^ (cat. no. 93289; Sigma-Aldrich; Merck KGaA). RNA was extracted using the RNAprep Pure Tissue kit (cat. no. DP431; Tiangen Biotech Co., Ltd.) according to the manufacturer\'s instructions. RT was performed using the PrimeScript RT reagent kit with gDNA Eraser (cat. no. RR047A; Takara Bio, Inc.) according to the manufacturer\'s instructions. gDNA was removed at 42°C for 2 min, then cDNA was synthesized at 37°C for 15 min and 85°C for 5 sec. qPCR was conducted using the CellAmp Direct TB Green^®^ RT-qPCR kit (cat. no. 3735A; Takara Bio, Inc.) and an ABI ViiA 7 system (cat. no. 4453536; Thermo Fisher Scientific, Inc.). The following thermocycling conditions were used: Melting at 95°C for 30 sec, annealing at 95°C for 3 sec, and extension at 60°C for 30 sec for 40 cycles. The primers ([Table II](#tII-mmr-22-02-0915){ref-type="table"}) for β-actin, proenkephalin (PENK), chordin like 1 (CHRDL1), calumenin (CALU), insulin-like growth factor binding protein 3 (IGFBP3) and ceruloplasmin (CP) were synthesized by Generay Biotech Co., Ltd. All samples were detected in triplicate, and gene expression values were normalized to the values of β-actin using 2^−ΔΔCq^ method ([@b27-mmr-22-02-0915]). ### Statistical analysis SPSS software (v16.0, SPSS, Inc.) was used for statistical analysis, and GraphPad Prism software (v5, GraphPad Software, Inc.) was used for mapping. Data are presented as the mean ± SD. For two-group comparisons, an independent sample t-test was used. P\<0.05 was considered to indicate a statistically significant difference. Results ======= ### Identification of DEGs between patients with DCM and healthy controls The raw data of GSE17800, GSE21610, GSE42955 and GSE79962, consisting of myocardium transcriptional expression data from 82 patients with DCM and 32 healthy controls, were downloaded from the GEO database. After preprocessing ([Fig. 1A](#f1-mmr-22-02-0915){ref-type="fig"}), the merged expression matrix of these four datasets was used for downstream analysis. The merged matrix was inputted into metaMA and RankProd, which are two R packages designed for meta-analysis of DEGs. A total of 1,943 genes (metaMA DEGs) were found to be differentially expressed by the metaMA package, while 1,390 genes (RankProd DEGs) were found to be differentially expressed by the RankProd package. A total of 898 overlapping genes between the metaMA DEGs and RankProd DEGs were considered DEGs ([Fig. 1B](#f1-mmr-22-02-0915){ref-type="fig"}), among which 424 genes were upregulated and 474 genes were downregulated ([Fig. 1C](#f1-mmr-22-02-0915){ref-type="fig"}). The top 30 upregulated genes and downregulated genes are listed in [Tables III](#tIII-mmr-22-02-0915){ref-type="table"} and [IV](#tIV-mmr-22-02-0915){ref-type="table"}, respectively. The most significantly upregulated genes were natriuretic peptide A (NPPA) and NPPB, while the most significantly downregulated gene was myosin heavy chain 6 (MYH6). In addition, the upregulation of NPPA and NPPB, and downregulation of MYH6 are well-established markers of heart dysfunction ([@b28-mmr-22-02-0915]). ### Validation of DEGs To assess the robustness of the DEGs, GSE79962 was randomly selected for clustering heat map analysis with all 898 DEGs involved ([Fig. 2A](#f2-mmr-22-02-0915){ref-type="fig"}). For further validation, a fifth microarray dataset (GSE3585) out of the training data was downloaded. After normalization, filtering and annotation, the expression matrix of GSE3585 (healthy controls=5; patients with DCM=7) was extracted. In total, 659 genes of the 898 DEGs were found in the GSE3585 expression matrix, and these 659 genes were used to plot a clustering heat map ([Fig. 2B](#f2-mmr-22-02-0915){ref-type="fig"}). These two clustering heat maps demonstrated that the DEGs were differentially expressed in the myocardium from patients with DCM compared with the healthy controls. ### GO annotation of DEGs To identify an overview of the biological functions of the DEGs, upregulated and downregulated DEGs were uploaded to the DAVID database for GO enrichment analysis. GO analysis results indicated that the upregulated genes were mainly involved in the following biological processes: Cell adhesion, negative regulation of cell proliferation, extracellular matrix (ECM) organization and positive regulation of apoptotic process ([Fig. 3A](#f3-mmr-22-02-0915){ref-type="fig"}). In addition, downregulated genes were primarily enriched in the negative regulation of the apoptotic process, potassium ion transport, mitochondrial translational termination and mitochondrial translational elongation ([Fig. 3B](#f3-mmr-22-02-0915){ref-type="fig"}). Furthermore, all these biological processes are closely related to the occurrence and development of DCM. In addition, it was revealed that both the upregulated and downregulated genes were involved in the process of apoptosis regulation, which indicates the importance of apoptosis in DCM. ### KEGG pathway enrichment of the DEGs The present study investigated the KEGG pathways involved in the DEGs. It was demonstrated that upregulated DEGs were mainly enriched in the ECM-receptor interaction, the p53 signaling pathway, focal adhesion and the transforming growth factor-β signaling pathway ([Fig. 3C](#f3-mmr-22-02-0915){ref-type="fig"}). In addition, downregulated DEGs were primarily enriched in cardiac muscle contraction, the hypoxia-inducible factor-1 signaling pathway, adrenergic signaling in cardiomyocytes and the cGMP-protein kinase G signaling pathway ([Fig. 3D](#f3-mmr-22-02-0915){ref-type="fig"}). Therefore, the present results indicated that these pathways may play vital roles in the progression of DCM. ### Construction of the PPI network, and identification of important modules and hub genes To assess the interactions among the proteins encoded by the DEGs, DEGs were uploaded to the STRING database to establish the PPI network ([Fig. 4](#f4-mmr-22-02-0915){ref-type="fig"}). It was demonstrated that there were 300 nodes and 852 edges in the network. Furthermore, among the 300 nodes, 151 nodes were upregulated, while 149 were downregulated. The MCODE plug-in was used to identify key modules of the PPI network, and the top three modules are presented in [Fig. 5A-C](#f5-mmr-22-02-0915){ref-type="fig"}. GO analysis results indicated that DEGs in module 2 were mainly enriched in protein ubiquitination-related biological processes, while the DEGs in module 3 were collagen- and ECM-related ([Fig. 5D](#f5-mmr-22-02-0915){ref-type="fig"}). Using four algorithms, the cytoHubba plug-in was used to detect hub genes of the PPI network. In addition, the top 20 genes were identified by degree, MCC, MNC and DMNC, which were then used for Venn diagram analysis ([Fig. 5E](#f5-mmr-22-02-0915){ref-type="fig"}). It was revealed that PENK, CHRDL1, CALU, apolipoprotein L1 (APOL1), IGFBP3 and CP overlapped among the four groups, and thus were deemed the hub genes that may play important roles in DCM ([Table V](#tV-mmr-22-02-0915){ref-type="table"}). ### Validation of the hub genes in a mouse DCM model Mice receiving repeated DOX injections were used as a DCM *in vivo* model. As previously reported ([@b29-mmr-22-02-0915],[@b30-mmr-22-02-0915]), DOX mice that presented with progressive cardiac functional decline and ventricular dilation were successfully established as the DCM mouse model. Mice injected with cumulative 20 mg/kg DOX exhibited impaired heart function ([Fig. 6A-C](#f6-mmr-22-02-0915){ref-type="fig"}), dilated LV ([Fig. 6D and E](#f6-mmr-22-02-0915){ref-type="fig"}) and thin ventricular walls ([Fig. 6F and G](#f6-mmr-22-02-0915){ref-type="fig"}), which imitated the pathological manifestation of human DCM ([@b31-mmr-22-02-0915]). APOL1 does not have a homologous gene in mice ([@b32-mmr-22-02-0915]), but the mRNA expression levels of PENK, CHRDL1, CALU, IGFBP3 and CP in mouse hearts were detected with RT-qPCR. The present results indicated that the mRNA expression levels of PENK, CHRDL1, IGFBP3 and CP were significantly increased, and the expression of CALU was significantly decreased, in the hearts of mice treated with DOX compared with the saline controls ([Fig. 6H](#f6-mmr-22-02-0915){ref-type="fig"}). In addition, the changes in the expression of these hub genes observed *in vivo* were consistent with the microarray data, thus indicating an potential connection between hub genes and DCM. Discussion ========== The present study analyzed four DCM microarrays from GEO, which to the best of the authors\' knowledge is the largest sample size that has been investigated, compared with the former researches on gene expression of DCM ([@b6-mmr-22-02-0915],[@b7-mmr-22-02-0915],[@b33-mmr-22-02-0915],[@b34-mmr-22-02-0915]), using integrated bioinformatics analysis. With the establishment of a PPI network, six hub genes associated with DCM were identified, including upregulated PENK, CHRDL1, IGFBP3 and CP, as well as downregulated CALU and APOL1. For further assessment of these hub genes, RT-qPCR was performed in a DOX-induced DCM mouse model. In line with the results of bioinformatics analysis, similar gene changes were observed in DOX mouse hearts; except for APOL1, which lacks homologous genes in mice ([@b32-mmr-22-02-0915]). Furthermore, compared with mice in the control group, the expression levels of PENK, CHRDL1, IGFBP3 and CP were significantly increased in DOX mice, while the expression of CALU was significantly decreased. It is speculated that some of the identified hub genes may serve as biological markers for early DCM screening in the clinic. CP, encoding the metalloprotein that binds the copper in plasma, is involved in the peroxidation of Fe(II) transferrin to Fe(III) transferrin ([@b35-mmr-22-02-0915]). In addition, CP is associated with cardiovascular disease by decreasing nitric oxide bioavailability in blood ([@b36-mmr-22-02-0915]). In a previous clinical study, CP was revealed to be independently related to cardiac function in patients with heart failure ([@b37-mmr-22-02-0915]). CALU produces a Ca^2+^-binding protein that is localized in the endoplasmic reticulum (ER) ([@b38-mmr-22-02-0915]). Furthermore, CALU is mainly expressed in the heart, and facilitates protein folding and sorting in the ER ([@b38-mmr-22-02-0915]). During the excitation-contraction coupling process, CALU regulates Ca^2+^ uptake ([@b39-mmr-22-02-0915]) and plays an important role in maintaining normal heart function ([@b38-mmr-22-02-0915]). Thus, aberrant expression levels of CP and CALU could provide a foundation for the progressive decline of cardiac function in DCM, and have potential value in the early screening of DCM ([@b40-mmr-22-02-0915]). In addition, the DEGs identified in the present study may be valuable for the prognosis of DCM. PENK, encoding a preproprotein of pentapeptide opioids, mainly shows biased expression in adrenal tissue ([@b41-mmr-22-02-0915]). Previous studies have also revealed that PENK is an important biomarker for renal dysfunction ([@b42-mmr-22-02-0915]--[@b44-mmr-22-02-0915]). In addition, APOL1, encoding a secreted high-density lipoprotein that binds to apolipoprotein A-I to promote lipid exchange, is reported to be associated with kidney disease when it is mutated ([@b45-mmr-22-02-0915]). With regards to the interaction between the cardiovascular and renal systems, PENK and APOL1 are also of potential value for cardiac disease research ([@b46-mmr-22-02-0915]). Kanagala *et al* ([@b47-mmr-22-02-0915]) revealed a relationship between PENK and mortality rate in patients with heart failure. Furthermore, the aberrant expression of APOL1 has been revealed to play an important role in the development of hypertension ([@b48-mmr-22-02-0915]). In line with these previous studies, the present results identified upregulation of PENK and downregulation of APOL1 in patients with DCM and DCM-induced mice, which indicated that these may provide prognostic information for the outcome in the progression of DCM. Furthermore, the present study may provide a novel therapeutic target for the mechanism and treatment of DCM. CHRDL1 is an important gene that expresses the antagonist of bone morphogenetic protein 4 (BMP4) ([@b49-mmr-22-02-0915]). In cardiac disease, BMP4 expression varies with cardiac function ([@b50-mmr-22-02-0915],[@b51-mmr-22-02-0915]). In addition, Wu *et al* ([@b51-mmr-22-02-0915]) revealed that recombinant BMP4 plays a protective role in mouse cardiomyocytes. The present study found increased expression of CHRDL1 in DCM, which could potentially antagonize the protective effect of BMP4 and accelerate DCM progression. Thus, suppressing the expression of CHRDL1 may be a novel treatment target for attenuating DCM development. IGFBP3 is a member of the IGFBP family, and encodes a protein with an IGFBP domain and a thyroglobulin type-I domain ([@b52-mmr-22-02-0915]). Previous studies have revealed that under hypoxic conditions, IGFBP3 can promote mitochondria-dependent cardiomyocyte apoptosis by inhibiting the IGF1R/PI3K/Akt survival pathway ([@b53-mmr-22-02-0915]). In the present study, an increased expression of IGFBP3 was identified in DCM, indicating that aberrant IGFBP3 expression may be an important mechanism of DCM occurrence. To investigate the identified hub genes, a DCM mouse model was established by intraperitoneal injection of DOX. DCM is the final common response of myocardium to diverse genetic and environmental insults ([@b1-mmr-22-02-0915]). The etiological causes of DCM are various, and it is generally recognized that anthracycline cardiotoxicity is a common cause of DCM ([@b54-mmr-22-02-0915],[@b55-mmr-22-02-0915]). In addition, the clinical diagnosis of DCM depends on the clinical feature of progressively aggravated LV dilatation and systolic dysfunction ([@b54-mmr-22-02-0915]). The DOX mouse model also exhibits progressive cardiac functional decline and ventricular dilation ([@b29-mmr-22-02-0915]), which is consistent with the diagnosis of DCM. In addition, while the specific pathogenesis of DCM and the mechanism of its progressive progression remain unknown, previous studies revealed that the development of DCM may be associated with decreased mitochondrial function ([@b56-mmr-22-02-0915]), abnormal oxidative stress ([@b57-mmr-22-02-0915],[@b58-mmr-22-02-0915]), as well as excessive myocardial apoptosis and necrosis ([@b59-mmr-22-02-0915],[@b60-mmr-22-02-0915]). In line with these possible causes, it has also been reported that the DOX model is accompanied by similar mitochondrial dysfunction ([@b29-mmr-22-02-0915],[@b61-mmr-22-02-0915]), oxidative stress injury ([@b62-mmr-22-02-0915]) and apoptotic changes ([@b63-mmr-22-02-0915]). Thus, the present study used the DOX mouse model to imitate the process of DCM for assessment of the hub genes. Consistent with the bioinformatics analysis results, similar genetic changes were observed in the DOX mouse model. It was determined that the expression levels of PENK, CHRDL1, IGFBP3 and CP were significantly increased in the hearts of DOX mice, while the expression of CALU was significantly decreased, which is consistent with the results of integrated microarray analysis. However, while these hub genes in DCM have been identified, the specific underlying mechanisms of the genes involved in DCM development are not fully understood. Along with the involvement in autophagy, apoptosis and oxidation, which have been previously identified ([@b64-mmr-22-02-0915]), these hub genes may also interact with each other, which further affects the survival and function of myocardial cells. Thus, further studies are still required to establish the causative correlations between these genes and the occurrence of DCM. In conclusion, the present study identified six hub genes related to DCM, including PENK, CHRDL1, IGFBP3, CP, CALU and APOL1. In addition, these hub genes may provide a mechanism for DCM, and could serve as biomarkers for screening and diagnosis in the clinic. Not applicable. Funding ======= The present study was supported by the National Nature Science Foundation of China (grant no. 81770333). Availability of data and materials ================================== All data generated or analyzed during this study are included in this published article. Authors\' contributions ======================= HZ, JH and QS designed the present study, which was performed by JH, HZ and WJ. JH and WJ made substantial contributions to acquisition and analysis of data. HZ and WJ also made contributions to interpretation of data. JH wrote the initial draft of the manuscript. HZ revised it critically for important intellectual content. QS has given final approval of the version to be published. All authors have participated sufficiently in the work to take public responsibility for appropriate portions of the content and approved the manuscript, as well as agreed to be accountable for all aspects of the work. All authors read and approved the final manuscript. Ethics approval and consent to participate ========================================== All animal studies were approved by the Animal Care and Ethics Committee of Nanjing Medical University. The animal experiments were performed according to the Guide for the Care and Use of Laboratory Animals (National Institutes of Health publication, 8th edition, 2011). Patient consent for publication =============================== Not applicable. Competing interests =================== The authors declare that they have no competing interests. {#f1-mmr-22-02-0915} {#f2-mmr-22-02-0915} {#f3-mmr-22-02-0915} {#f4-mmr-22-02-0915} {#f5-mmr-22-02-0915} {#f6-mmr-22-02-0915} ###### Details of the microarrays used. GEO series number Control DCM Tissue Platform ------------------- --------- ----- ------------ -------------------------------------------------------------------------------------------- GSE17800 8 40 Myocardium GPL570,\[HG-U133_Plus_2\] Affymetrix Human Genome U133 Plus 2.0 Array GSE21610 8 21 Myocardium GPL570,\[HG-U133_Plus_2\] Affymetrix Human Genome U133 Plus 2.0 Array GSE42955 5 12 Myocardium GPL6244,\[HuGene-1_0-st\] Affymetrix Human Gene 1.0 ST Array \[transcript (gene) version\] GSE79962 11 9 Myocardium GPL6244,\[HuGene-1_0-st\] Affymetrix Human Gene 1.0 ST Array \[transcript (gene) version\] GSE3585 5 7 Myocardium GPL96,\[HG-U133A\] Affymetrix Human Genome U133A Array GEO, Gene Expression Omnibus; DCM, dilated cardiomyopathy. ###### Primer sequences. Gene Forward (5′→3′) Reverse (5′→3′) -------- ------------------------- ------------------------- PENK GGACTGCGCTAAATGCAGCTA GAAGCCTCCGTACCGTTTCAT CHRDL1 AACCTCCAAGCCAAAACTTTGA CCAGTGCTACTTTTCTGGTTGTC CALU AATGCTGATGGGTTCATTGATCT GGTTCTTATCTCGAAACTCCACG IGFBP3 CACACCGAGTGACCGATTCC GTGTCTGTGCTTTGAGACTCAT CP CTTAGCCTTGGCAAGAGATAAGC GATGACAGGGCCTAAAAACCC Actb GGCTGTATTCCCCTCCATCG CCAGTTGGTAACAATGCCATGT PENK, proenkephalin; CHRDL1, chordin like 1; CALU, calumenin; IGFBP3, insulin-like growth factor binding protein 3; CP, ceruloplasmin; Actb, β-actin. ###### Top 30 upregulated differentially expressed genes. Upregulated gene Fold change Percentage of false prediction ------------------ ------------- -------------------------------- NPPA 5.57 6.41×10^−49^ NPPB 4.60 5.28×10^−40^ SFRP4 2.46 2.61×10^−25^ SMOC2 2.28 4.41×10^−23^ HBB 2.53 1.40×10^−22^ EIF1AY 2.45 1.82×10^−22^ PHLDA1 1.99 4.61×10^−22^ LTBP2 2.06 7.48×10^−20^ THBS4 2.05 2.12×10^−19^ FRZB 2.07 5.42×10^−19^ POSTN 1.91 5.71×10^−19^ MXRA5 1.82 1.01×10^−18^ CCN2 1.96 6.47×10^−18^ OMD 2.00 6.13×10^−18^ TNNT1 1.91 1.12×10^−17^ PRELP 2.06 1.34×10^−17^ ASPN 1.81 5.31×10^−17^ FMOD 1.94 7.19×10^−17^ CCDC80 1.79 1.42×10^−16^ RGS4 1.84 1.04×10^−15^ SERPINE2 1.70 2.58×10^−15^ USP9Y 1.68 2.78×10^−15^ STAT4 1.90 2.77×10^−15^ PLCE1 1.75 2.67×10^−15^ UCHL1 1.80 3.19×10^−15^ OGN 1.74 1.08×10^−14^ HSPA2 1.72 2.10×10^−14^ NAP1L3 1.77 3.88×10^−14^ MFAP4 1.68 4.80×10^−14^ JAK2 1.70 1.27×10^−13^ ###### Top 30 downregulated differentially expressed genes. Downregulated gene Fold change Percentage of false prediction -------------------- ------------- -------------------------------- MYH6 0.32 2.33×10^−38^ ETNPPL 0.39 6.74×10^−29^ SERPINA3 0.38 1.88×10^−27^ DHRS7C 0.45 1.51×10^−25^ AQP4 0.52 1.29×10^−20^ CCL2 0.47 1.89×10^−17^ LYVE1 0.58 2.19×10^−16^ LSAMP 0.53 2.69×10^−16^ DLK1 0.53 6.81×10^−16^ CORIN 0.53 1.04×10^−15^ HOPX 0.55 1.92×10^−15^ CA14 0.55 4.94×10^−15^ RARRES1 0.54 8.04×10^−15^ CD163 0.62 9.97×10^−15^ KCNIP2 0.60 1.01×10^−14^ C1orf105 0.57 1.20×10^−14^ CD14 0.56 3.45×10^−14^ F13A1 0.61 1.53×10^−13^ F5 0.82 2.01×10^−13^ TOGARAM2 0.60 2.15×10^−13^ CPNE4 0.59 2.05×10^−12^ AQP3 0.62 4.89×10^−12^ VSIG4 0.67 1.31×10^−11^ HEY2 0.63 5.73×10^−11^ PLIN2 0.62 6.50×10^−11^ SELE 0.57 1.46×10^−10^ ADAMTS9 0.63 1.71×10^−10^ NAMPT 0.65 1.90×10^−10^ SLCO5A1 0.60 2.63×10^−10^ HINT3 0.60 9.40×10^−10^ ###### Topological properties of hub genes. Hub gene Degree MCC DMNC MNC Fold change Percentage of false prediction Expression change ---------- -------- ------------- --------- ----- ------------- -------------------------------- ------------------- PENK 19 8.72×10^10^ 1.02476 14 1.71 8.04×10^−13^ Upregulated CHRDL1 16 8.72×10^10^ 1.02476 14 1.31 8.04×10^−6^ Upregulated CALU 16 8.72×10^10^ 0.92137 15 0.84 1.85×10^−2^ Downregulated APOL1 15 8.72×10^10^ 0.92137 15 0.77 5.89×10^−3^ Downregulated IGFBP3 15 8.72×10^10^ 0.92137 15 1.37 9.86×10^−6^ Upregulated CP 15 8.72×10^10^ 1.02476 14 1.32 2.98×10^−6^ Upregulated PENK, proenkephalin; CHRDL1, chordin like 1; CALU, calumenin; APOL1, apolipoprotein L1; IGFBP3, insulin-like growth factor binding protein 3; CP, ceruloplasmin; MNC, Maximum Neighborhood Component; DMNC, Density of Maximum Neighborhood Component; MCC, Maximal Clique Centrality. [^1]: Contributed equally

\section{Comparison with Random Search} \begin{figure} \centering \includegraphics[height=6.5cm]{folder_for_appendix/loss.pdf} \caption{Validation accuracy and noise decay versus episode, using spatial decomposition on CIFAR-10 under $2\times$. We explore for 100 episodes with constant noise $\sigma$. Then we exploit for 300 episodes with exponentially decayed noise $\sigma$. Notice that best accuracy from random exploration plateaus for very long time. Random search cannot find good enough results with limited episodes (\textit{better viewed in color)} } \label{fig:loss} \end{figure} Theoretically, after a large number of trials, the random search result will be close to the reinforcement learning result. It is unclear how many trials are necessary for the random search to succeed, especially for small networks on CIFAR-10. As shown in \fig{fig:loss}, we search sparsity policy for spatial decomposition of Plain-20 under $2\times$. Before 100 episodes, our AMC\xspace behaves the same as random search. Both explore with constant noise $\sigma=0.5$. After 100 episodes, our AMC\xspace exploits with exponentially decayed noise $\sigma$ and start learning. It can find a good model quickly; however, random search plateaus. 400 episodes is insufficient for the random search to get good enough results, even for a small network on CIFAR-10. \section{Effectiveness of Layer Embeddings}\label{sec:exp:state} \begin{figure} \centering \includegraphics[height=6.5cm]{folder_for_appendix/state.pdf} \caption{Effectiveness of each layer embedding, evaluated with spatial decomposition for Plain-20 $2\times$. Combining all embeddings together, the performance is even better (\textit{smaller is better})} \label{fig:state} \end{figure} Layer embedding state $s_t$ (Sec.~\ref{sec:state}) is essential for AMC\xspace to distinguish different convolutional layers. As shown in \fig{fig:state}, we evaluate the effectiveness of each layer's embedding, and the combination of all of them with spatial decomposition Plain-20 under $2\times$. If we just use $stride$ as the layer embedding, it is not informative enough to distinguish different layers. The performances of $n, c, H, W$ are similar, since their values change together when $stride$ is 2. $FLOPs[i], reduced, a_{t-1}$ and layer index are most effective, since they dynamically change at each layer. $a_{t-1}$ and $reduced$ resemble the RNN used by other architecture search approaches~\cite{zoph2017learning,cai2017reinforcement}, since $s_t$ is affected by $s_{t-1}$. In the following experiments, we combine all these embeddings, which makes the performance even better. \section{More speed-up results: Samsung Galaxy S7} We provide the speed-up obtained by AMC measured on a Samsung Galaxy S7 phone with Qualcomm Snapdragon 820 SoC in Table~\ref{tab:samsung}. \begin{table} \setlength{\tabcolsep}{3pt} \centering \caption{AMC speeds up MobileNet on Samsung Galaxy S7 with TF-Lite.} \label{tab:samsung} \scalebox{0.9}{ \begin{tabular}{c|c|cc|ccccc} \hline \multirow{2}{*}{} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Million\\ MAC\end{tabular}} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}top-1\\ acc.\\ (\%)\end{tabular}} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}top-5\\ acc.\\ (\%)\end{tabular}} & \multicolumn{5}{c}{Android} \\ & & & & \begin{tabular}[c]{@{}c@{}}conv.\\ (ms)\end{tabular} & \begin{tabular}[c]{@{}c@{}}depth.\\ (ms)\end{tabular} & \begin{tabular}[c]{@{}c@{}}total\\ (ms)\end{tabular} & \begin{tabular}[c]{@{}c@{}}speed\\ \\\end{tabular} & \begin{tabular}[c]{@{}c@{}}memory\\ \\\end{tabular} \\ \hline\hline \begin{tabular}[c]{@{}c@{}}1.0\\ MobileNet\end{tabular} & 569 & 70.9 & 89.5 & 102.3 & 12.0 & 119.0 & \begin{tabular}[c]{@{}c@{}}8.4 fps\\ (1$\times$)\end{tabular} & 20.1MB \\ \begin{tabular}[c]{@{}c@{}}0.75\\ MobileNet\end{tabular} & 325 & 68.4 & 88.2 & 56.0 & 9.9 & 69.5 & \begin{tabular}[c]{@{}c@{}}14.4 fps\\ (1.71$\times$)\end{tabular} & 14.8MB \\ \hline \begin{tabular}[c]{@{}c@{}}$0.5\times$ FLOPs\\ \textbf{AMC (ours)}\end{tabular} & 285 & 70.5 & 89.3 & 51.7 & 9.2 & 64.4 & \begin{tabular}[c]{@{}c@{}}\textbf{15.5 fps}\\ (\textbf{1.85}$\times$)\end{tabular} & 14.3MB \\ \begin{tabular}[c]{@{}c@{}}$0.5\times$ Time\\ \textbf{AMC (ours)}\end{tabular} & 272 & 70.2 & 89.2 & 48.7 & 8.3 & 59.7 & \begin{tabular}[c]{@{}c@{}}\textbf{16.8 fps}\\ (\textbf{2.00}$\times$)\end{tabular} & 13.2MB \\ \hline \end{tabular} } \end{table} \end{document} \section{Introduction} \begin{figure}[t] \centering \includegraphics[width=\linewidth]{figures/first.pdf} \caption{Overview of AutoML for Model Compression~(AMC\xspace) engine. Left: AMC\xspace~replaces human and makes model compression fully automated while performing better than human. Right: Form AMC\xspace\ as a reinforcement learning problem. We process a pre-trained network (\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot, MobileNet) in a layer-by-layer manner. Our reinforcement learning agent (DDPG) receives the embedding $s_t$ from a layer $t$, and outputs a sparsity ratio $a_t$. After the layer is compressed with $a_t$, it moves to the next layer $L_{t+1}$. The accuracy of the pruned model with all layers compressed is evaluated. Finally, as a function of accuracy and FLOP, reward $R$ is returned to the reinforcement learning agent.} \label{fig:net} \end{figure} In many machine learning applications such as robotics, self-driving cars, and advertisement ranking, deep neural networks are constrained by latency, energy and model size budget. Many approaches have been proposed to improve the hardware efficiency of neural networks by model compression~\cite{jaderberg2014speeding,han2015deep,he2017channel}. The core of model compression technique is to determine the compression policy for each layer as they have different redundancy, which conventionally requires hand-crafted heuristics and domain expertise to explore the large design space trading off among model size, speed, and accuracy. The design space is so large that human heuristic is usually sub-optimal, and manual model compression is time-consuming. To this end, we aim to automatically find the compression policy for an arbitrary network to achieve even better performance than human designed, rule-based model compression methods. Previously there were many rule-based model compression heuristics \cite{han2015learning,han_thesis}. For example, prune less parameters in the first layer which extracts low level features and have the least amount of parameters; prune more parameters in the FC layer since FC layers have the most parameters; prune less parameters in the layers that are sensitive to pruning, etc. However, as the layers in deep neural networks are not independent, these rule-based pruning policies are non-optimal, and doesn't transfer from one model to another model. Neural network architectures are evolving fast, and we need an automated way to compress them to improve engineer efficiency. As the neural network becomes deeper, the design space has exponential complexity , which is infeasible to be solved by greedy, rule-based methods. Therefore, we propose AutoML for Model Compression (AMC\xspace) which leverages reinforcement learning to automatically sample the design space and improve the model compression quality. \fig{fig:net} illustrates our AMC\xspace engine. When compressing a network, our AMC\xspace engine automates this process by learning-based policies, rather than relying on rule-based policy and experienced engineers. We observe that the accuracy of the compressed model is very sensitive to the sparsity of each layer, requiring a fine-grained action space. Therefore, instead of searching over a discrete space, we come up with a continuous compression ratio control strategy with a DDPG~\cite{lillicrap2015continuous} agent to learn through trial and error: penalizing accuracy loss while encouraging model shrinking and speedup. The actor-critic structure also helps to reduce variance, facilitating stabler training. Specifically, our DDPG agent processes the network in a layer-wise manner. For each layer $L_t$, the agent receives a layer embedding $s_t$ which encodes useful characteristics of this layer, and then it outputs a precise compression ratio $a_t$. After layer $L_t$ is compressed with $a_t$, the agent moves to the next layer $L_{t+1}$. The validation accuracy of the pruned model with all layers compressed is evaluated without fine-tuning, which is an efficient delegate of the fine-tuned accuracy. This simple approximation can improve the search time not having to retrain the model, and provide high quality search result. After the policy search is done, the best-explored model is fine-tuned to achieve the best performance. We proposed two compression policy search protocols for different scenarios. For \emph{latency-critical} AI applications (\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot, mobile apps, self-driving cars, and advertisement rankings), we propose \emph{resource-constrained} compression to achieve the best accuracy given the maximum amount of hardware resources (\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot, FLOPs, latency, and model size), For \emph{quality-critical} AI applications (\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot, Google Photos) where latency is not a hard constraint, we propose \emph{accuracy-guaranteed} compression to achieve the smallest model with no loss of accuracy. We achieve \emph{resource-constrained} compression by constraining the search space, in which the action space (pruning ratio) is constrained such that the model compressed by the agent is always below the resources budget. For \emph{accuracy-guaranteed} compression, we define a reward that is a function of both accuracy and hardware resource. With this reward function, we are able to explore the limit of compression without harming the accuracy of models. To demonstrate the wide and general applicability, we evaluate our AMC\xspace engine on multiple neural networks, including VGG~\cite{simonyan2014very}, ResNet~\cite{he2016deep}, and MobileNet~\cite{howard2017mobilenets}, and we also test the generalization ability of the compressed model from classification to object detection. Extensive experiments suggest that AMC\xspace offers better performance than hand-crafted heuristic policies. For ResNet-50, we push the expert-tuned compression ratio~\cite{han_thesis} from 3.4$\times$ to \textbf{5$\times$} with no loss of accuracy. Furthermore, we reduce the FLOPs of MobileNet~\cite{howard2017mobilenets} by 2$\times$, achieving top one accuracy of 70.2\%, which is on a better Pareto curve than 0.75 MobileNet, and we achieve a speedup of \textbf{1.53$\times$} on the Titan XP and \textbf{1.95$\times$} on an Android phone. \begin{table}[t] \centering \caption{Comparisons of reinforcement learning approaches for models searching (NAS: Neural Architecture Search~\cite{zoph2017learning}, NT: Network Transformation~\cite{cai2017reinforcement}, N2N: Network to Network~\cite{ashok2017n2n}, and AMC\xspace: AutoML for Model Compression. AMC\xspace distinguishes from other works by getting reward without fine-tuning, continuous search space control, and can produce both accuracy-guaranteed \ and hardware resource-constrained\ models.} \label{tab:compare} \begin{tabular}{lC{1cm}C{1cm}C{1cm}C{1cm}} \toprule & NAS & NT & N2N & \textbf{AMC\xspace} \\ \midrule optimize for accuracy & \checkmark & \checkmark & \checkmark & \checkmark \\ optimize for latency & & & & \checkmark \\ simple, non-RNN controller & & & & \checkmark\\ fast exploration with few GPUs & & \checkmark & \checkmark & \checkmark \\ continuous action space & & & & \checkmark \\ \bottomrule \end{tabular} \end{table} \section{Related Work} \stitle{CNN Compression and Acceleration.} Extensive works~\cite{han2015learning,han2015deep,luo2017thinet,dong2017more,han2016eie,han2017ese} have been done on accelerating neural networks by compression. Quantization~\cite{zhu2016trained,courbariaux2016binarynet,rastegari2016xnor} and special convolution implementations~\cite{mathieu2013fast,vasilache2014fast,lavin2015fast,bagherinezhad2016lcnn} can also speed up the neural networks. Tensor factorization~\cite{lebedev2014speeding,gong2014compressing,kim2015compression,Masana_2017_ICCV} decomposes weights into light-weight pieces, for example~\cite{xue2013restructuring,denton2014exploiting,girshick2015fast} proposed to accelerate the fully connected layers with truncated SVD; Jaderberg~\emph{et al}\onedot~\cite{jaderberg2014speeding} proposed to factorize layers into 1$\times$3 and 3$\times$1; and Zhang~\emph{et al}\onedot~\cite{zhang2016accelerating} proposed to factorize layers into 3$\times$3 and 1$\times$1. Channel pruning~\cite{polyak2015channel,hu2016network,anwar2016compact,molchanov2016pruning} removes the redundant channels from feature maps. A common problem of these methods is how to determine the sparsity ratio for each layer. \stitle{Neural Architecture Search and AutoML.} Many works on searching models with reinforcement learning and genetic algorithms~\cite{stanley2002evolving,real2017large,brock2017smash,miikkulainen2017evolving} greatly improve the performance of neural networks. NAS~\cite{zoph2017learning} aims to search the transferable network blocks, and its performance surpasses many \textit{manually} designed architectures~\cite{szegedy2015going,he2016deep,chollet2016xception}. Cai~\emph{et al}\onedot~\cite{cai2017reinforcement} proposed to speed up the exploration via network transformation~\cite{chen2015net2net}. Inspired by them, N2N~\cite{ashok2017n2n} integrated reinforcement learning into channel selection. In \tbl{tab:compare}, we demonstrate several merits of our AMC\xspace engine. Compared with previous work, AMC engine optimizes for both accuracy and latency, requires a simple non-RNN controller, can do fast exploration with fewer GPU hours, and also support continuous action space. \section{Methodology} We present an overview of our AutoML for Model Compression (AMC\xspace) engine in \fig{fig:net}. We aim to automatically find the redundancy for each layer, characterized by sparsity. We train an reinforcement learning agent to predict the action and give the sparsity, then perform form the pruning. We quickly evaluate the accuracy after pruning but before fine-tuning as an effective delegate of final accuracy. Then we update the agent by encouraging smaller, faster and more accurate models. \subsection{Problem Definition} Model compression is achieved by reducing the number of parameters and computation of each layer in deep neural networks. There are two categories of pruning: fine-grained pruning and structured pruning. \emph{Fine-grained pruning}~\cite{han2015deep} aims to prune individual unimportant elements in weight tensors, which is able to achieve very high compression rate with no loss of accuracy. However, such algorithms result in an irregular pattern of sparsity, and it requires specialized hardware such as EIE \cite{han2016eie} for speed up. \emph{Coarse-grained / structured pruning}~\cite{li2016pruning} aims to prune entire regular regions of weight tensors (\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot, channel, row, column, block, \emph{etc}\onedot} \def\vs{\emph{vs}\onedot). The pruned weights are regular and can be accelerated directly with off-the-shelf hardware and libraries. Here we study structured pruning that shrink the input channel of each convolutional and fully connected layer. Our goal is to precisely find out the effective \emph{sparsity} for each layer, which used to be manually determined in previous studies~\cite{molchanov2016pruning,li2016pruning,he2017channel}. Take convolutional layer as an example. The shape of a weight tensor is $n\times c\times k \times k$, where $n, c$ are output and input channels, and $k$ is the kernel size. For fine-grained pruning, the sparsity is defined as the number of zero elements divided by the number of total elements, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot $\text{\#zeros} / (n \times c \times k \times h)$. For channel pruning, we shrink the weight tensor to $n\times c^\prime\times k \times k$ (where $c^\prime < c$), hence the sparsity becomes $c^\prime / c$. \subsection{Automated Compression with Reinforcement Learning} AMC leverages reinforcement learning for efficient search over action space. Here we introduce the detailed setting of reinforcement learning framework. \subsubsection{The State Space}\label{sec:state} For each layer $t$, we have 11 features that characterize the state $s_t$: \begin{align} (t, n, c, h, w, stride, k, FLOPs[t], reduced, rest, \textit{a}_{t-1}) \end{align} where $t$ is the layer index, the dimension of the kernel is $n\times c\times k\times k$, and the input is $c\times h \times w$. $FLOPs[t]$ is the FLOPs of layer $L_t$. $Reduced$ is the total number of reduced FLOPs in previous layers. $Rest$ is the number of remaining FLOPs in the following layers. Before being passed to the agent, they are scaled within $[0, 1]$. Such features are essential for the agent to distinguish one convolutional layer from another. \subsubsection{The Action Space}\label{sec:action} Most of the existing works use discrete space as coarse-grained action space (\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot, $\{64, 128, 256, 512\}$ for the number of channels). Coarse-grained action space might not be a problem for a high-accuracy model architecture search. However, we observed that model compression is very sensitive to sparsity ratio and requires fine-grained action space, leading to an explosion of the number of discrete actions (Sec.~\ref{sec:exp:vgg}). Such large action spaces are difficult to explore efficiently~\cite{lillicrap2015continuous}. Discretization also throws away the order: for example, 10\% sparsity is more aggressive than 20\% and far more aggressive than 30\%. As a result, we propose to use continuous action space $\textit{a} \in (0, 1]$, which enables more fine-grained and accurate compression. \subsubsection{DDPG Agent}\label{sec:agent} As illustrated in \fig{fig:net}, the agent receives an embedding state $s_t$ of layer $L_t$ from the environment and then outputs a sparsity ratio as action $\textit{a}_t$. The underlying layer is compressed with $\textit{a}_t$ (rounded to the nearest feasible fraction) using a specified compression algorithm (\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot, channel pruning). Then the agent moves to the next layer $L_{t+1}$, and receives state $s_{t+1}$. After finishing the final layer $L_T$, the reward accuracy is evaluated on the validation set and returned to the agent. For fast exploration, we evaluate the reward accuracy without fine-tuning, which is a good approximation for fine-tuned accuracy (Sec.~\ref{sec:acc}). We use the deep deterministic policy gradient (DDPG) for continuous control of the compression ratio, which is an off-policy actor-critic algorithm. For the exploration noise process, we use truncated normal distribution: \begin{align} \mu^\prime(s_t) \sim \text{TN}\left(\mu\left(s_t \mid \theta_t^\mu\right), \sigma^2, 0, 1\right) \end{align} During exploitation, noise $\sigma$ is initialized as $0.5$ and is decayed after each episode exponentially. Following Block-QNN~\cite{zhong2017practical}, which applies a variant form of Bellman's Equation~\cite{watkins1989learning}, each transition in an episode is $(s_t, a_t, R, s_{t+1})$, where $R$ is the reward after the network is compressed. During the update, the baseline reward $b$ is subtracted to reduce the variance of gradient estimation, which is an exponential moving average of the previous rewards~\cite{zoph2016neural,cai2017reinforcement}: \begin{equation} \begin{gathered} Loss = \frac{1}{N} \sum_i\left(y_i - Q\left(s_i, a_i \mid \theta^Q\right)\right)^2 \\ y_i = r_i - b + \gamma Q(s_{i+1}, \mu(s_{i+1})\mid \theta^Q) \end{gathered} \end{equation} The discount factor $\gamma$ is set to 1 to avoid over-prioritizing short-term rewards~\cite{baker2016designing}. \subsection{Search Protocols} \subsubsection{Resource-Constrained Compression}\label{sec:wall} By limiting the action space (the sparsity ratio for each layer), we can accurately arrive at the target compression ratio. Following~\cite{zoph2017learning,baker2016designing,zhong2017practical}, we use the following reward: \begin{align} R_{err} = -Error \end{align} This reward offers no incentive for model size reduction, so we the achieve target compression ratio by an alternative way: limiting the action space. Take fine-grained pruning for model size reduction as an example: we allow arbitrary action $\textit{a}$ at the first few layers; we start to limit the action $\textit{a}$ when we find that the budget is insufficient even after compressing \textbf{all} the following layers with most aggressive strategy. Algorithm~\ref{Alg1} illustrates the process. (For channel pruning, the code will be longer but similar, since removing input channels of layer $L_t$ will also remove the corresponding output channels of $L_{t-1}$, reducing parameters/FLOPs of both layers). Note again that our algorithm is not limited to constraining \textit{model size} and it can be replaced by other resources, such as FLOPs or the actual inference time on mobile device. Based on our experiments (Sec.~\ref{sec:cifarresource}), as the agent receives no incentive for going below the budget, it can precisely arrive at the target compression ratio. \begin{algorithm}[t] \caption{Predict the sparsity ratio $\texttt{action}_t$ for layer $L_t$ with constrained model size (number of parameters) using fine-grained pruning} \label{Alg1} \begin{algorithmic} \LeftComment{Initialize the reduced model size so far} \If {$t$ is equal to $0$} \State $\texttt{W}_\text{reduced} \gets 0 $ \EndIf \State \LeftComment{Compute the agent's action and bound it with the maximum sparsity ratio} \State $\texttt{action}_t \gets \mu^{\prime}(s_t)$ \State $\texttt{action}_t \gets \min(\texttt{action}_t, \texttt{action}_\text{max})$ \State \LeftComment{Compute the model size of the whole model and all the later layers} \State $\texttt{W}_\text{all} \gets \sum_k \texttt{W}_k$ \State $\texttt{W}_\text{rest} \gets \sum_{k=t+1} \texttt{W}_k$ \State \LeftComment{Compute the number of parameters we have to reduce in the current layer if all the later layers are pruned with the maximum sparsity ratio. $\alpha$ is the target sparsity ratio of the whole model.} \State $\texttt{W}_\text{duty} \gets \alpha \cdot \texttt{W}_\text{all} - \texttt{action}_\text{max} \cdot \texttt{W}_\text{rest} - \texttt{W}_\text{reduced}$ \State \LeftComment{Bound $\texttt{action}_t$ if it is too small to meet the target model size reduction} \State $\texttt{action}_t \gets \max(\texttt{action}_t, \texttt{W}_\text{duty} / \texttt{W}_t)$ \State \LeftComment{Update the accumulation of reduced model size} \State $\texttt{W}_\text{reduced} \gets \texttt{W}_\text{reduced} + \texttt{action}_t \cdot \texttt{W}_t$ \State \State \textbf{return} $\texttt{action}_t$ \end{algorithmic} \end{algorithm} \subsubsection{Accuracy-Guaranteed\ Compression}\label{sec:reward} By tweaking the reward function, we can accurately find out the limit of compression that offers no loss of accuracy. We empirically observe that $Error$ is inversely-proportional to $\log(FLOPs)$ or $\log(\#Param)$~\cite{canziani2016analysis}. Driven by this, we devise the following reward function: \begin{align} R_\text{FLOPs} = - Error \cdot \log(\text{FLOPs}) \\ R_\text{Param} = -Error \cdot \log(\text{\#Param}) \end{align} This reward function is sensitive to $Error$; in the meantime, it provides a small incentive for reducing $\text{FLOPs}$ or model size. Based on our experiments in \fig{sec:cifarperform}, we note that our agent automatically finds the limit of compression. \section{Experimental Results} For fine-grained pruning~\cite{han2015deep}, we prune the weights with least magnitude. The maximum sparsity ratio $a_{max}$ for convolutional layers is set to 0.8, and $a_{max}$ for fully connected layer is set to 0.98. For channel pruning, we use \textit{max response} selection (pruning the weights according to the magnitude~\cite{han2015learning}), and preserve Batch Normalization~\cite{ioffe2015batch} layers during pruning instead of merging them into convolutional layers. The maximum sparsity ratios $a_{max}$ for all layers are set to 0.8. Note that the manual upper bound $a_{max}$ is only intended for faster search, one can simply use $a_{max}=1$ which also produces similar results. Our actor network $\mu$ has two hidden layers, each with 300 units. The final output layer is a sigmoid layer to bound the actions within $(0, 1)$. Our critic network $Q$ also had two hidden layers, both with 300 units. Actions arr included in the second hidden layer. We use $\tau = 0.01$ for the soft target updates and train the network with 64 as batch size and 2000 as replay buffer size. Our agent first explores 100 episodes with a constant noise $\sigma=0.5$, and then exploits 300 episodes with exponentially decayed noise $\sigma$. \subsection{CIFAR-10 and Analysis} We conduct extensive experiments and fully analyze our AMC\xspace\ on CIFAR-10~\cite{krizhevsky2009learning} to verify the effectiveness of the 2 search protocols. CIFAR dataset consists of 50k training and 10k testing $32\times32$ tiny images in ten classes. We split the training images into 45k/5k train/validation. The accuracy reward is obtained on validation images. Our approach is computationally efficient: the RL can finish searching within \textbf{1 hour} on a single GeForce GTX TITAN Xp GPU. \begin{table}[t] \setlength{\tabcolsep}{3pt} \centering\small \caption{Pruning policy comparison of Plain-20, ResNets~\cite{he2016deep} on CIFAR-10~\cite{krizhevsky2009learning}. $R_\text{Err}$ corresponds to FLOPs-constrained compression with channel pruning, while $R_\text{Param}$ corresponds to accuracy guaranteed compression with fine-grained pruning. For both shallow network Plain-20 and deeper network ResNets, AMC\xspace outperforms \textit{hand-crafted} policies by a large margin. This enables efficient exploration without fine-tuning. Although AMC\xspace makes many trials on model architecture, we have separate validation and test dataset. No over-fitting is observed.} \begin{tabular}{cccccc} \toprule Model & Policy & Ratio & Val Acc. & Test Acc. & Acc. after FT. \\ \midrule \multirow{4}{*}{\begin{tabular}[c]{@{}c@{}} Plain-20\\(90.5\%) \end{tabular}} & deep (handcraft) & \multirow{4}{*}{\begin{tabular}[c]{@{}c@{}}50\% FLOPs \end{tabular}} & 79.6 & 79.2 & 88.3 \\ & shallow (handcraft) & & 83.2 & 82.9 & 89.2 \\ & uniform (handcraft) & & 84.0 & 83.9 & 89.7 \\ & \textbf{AMC ($R_\text{Err})$} & & \textbf{86.4} & \textbf{86.0} & \textbf{90.2} \\ \midrule \multirow{3}{*}{\begin{tabular}[c]{@{}c@{}}ResNet-56\\ (92.8\%)\end{tabular}} & uniform (handcraft) & \multirow{3}{*}{\begin{tabular}[c]{@{}c@{}}50\% FLOPs \end{tabular}} & 87.5 & 87.4 & 89.8 \\ & deep (handcraft) & & 88.4 & 88.4 & 91.5 \\ & \textbf{AMC ($R_\text{Err})$} & & \textbf{90.2} & \textbf{90.1} & \textbf{91.9} \\ \midrule \begin{tabular}{c} ResNet-50 \\ (93.53\%) \end{tabular} & \textbf{AMC ($R_\text{Param})$} & 60\% Params & \textbf{93.64} & \textbf{93.55} & -\\ \bottomrule \end{tabular} \label{tab:cifarpr} \end{table} \begin{figure}[t] \centering \includegraphics[width=\linewidth]{figures/plainpr} \caption{Comparisons of pruning strategies for Plain-20 under $2\times$. \textit{Uniform} policy sets the same compression ratio for each layer uniformly. \textit{Shallow} and \textit{deep} policies aggressively prune shallow and deep layers respectively. Policy given by AMC\xspace\ looks like sawtooth, which resembles the bottleneck architecture~\cite{he2016deep}. The accuracy given by AMC outperforms hand-crafted policies. (\textit{better viewed in color})} \label{fig:cifarpr} \end{figure} \subsubsection{FLOPs-Constrained Compression.}\label{sec:cifarresource} We conducted FLOPs-constrained experiments on CIFAR-10 with channel pruning. We compare our approach with three empirical policies~\cite{li2016pruning,he2017channel} illustrated in \fig{fig:cifarpr}: \textit{uniform} sets compression ratio uniformly, \textit{shallow} and \textit{deep} aggressively prune shallow and deep layers respectively. Based on sparsity distribution of different networks, a different strategy might be chosen. In \tbl{tab:cifarpr}, we show us using reward $R_{err}$ to accurately find the sparsity ratios for pruning 50\% for Plain-20 and ResNet-56~\cite{he2016deep} and compare it with empirical policies. We outperform empirical policies by a large margin. The best pruning setting found by AMC differs from hand-crafted heuristic (\fig{fig:cifarpr}). It learns a bottleneck architecture~\cite{he2016deep}. \subsubsection{Accuracy-Guaranteed\ Compression.}\label{sec:cifarperform} By using the $R_\text{Param}$ reward, our agent can automatically find the limit of compression, with smallest model size and little loss of performance. As shown in Table~\ref{tab:cifarpr}, we compress ResNet-50 with fine-grained pruning on CIFAR-10. The result we obtain has up to 60\% compression ratio with even a little higher accuracy on both validation set and test set, which might be in light of the regularization effect of pruning. Since our reward $R_{\text{Param}}$ focuses on $Error$ and offers very little incentive to compression in the meantime, it prefers the high-performance model with harmless compression. To shorten the search time, we obtain the reward using validation accuracy without fine-tuning. We believe if reward were fine-tuned accuracy, the agent should compress more aggressively, because the fine-tuned accuracy is much closer to the original accuracy. \subsubsection{Speedup policy exploration.}\label{sec:acc} Fine-tuning a pruned model usually takes a very long time. We observe a correlation between the pre-fine-tune accuracy and the post fine-tuning accuracy~\cite{han2015learning,he2017channel}. As shown in Table~\ref{tab:cifarpr}, policies that obtain higher validation accuracy correspondingly have higher fine-tuned accuracy. This enables us to predict final model accuracy without fine-tuning, which results in an efficient and faster policy exploration. The validation set and test set are separated, and we only use the validation set to generate reward during reinforcement learning. In addition, the compressed models have fewer parameters. As shown in Table~\ref{tab:cifarpr}, the test accuracy and validation accuracy are very close, indicating no over-fitting. \subsection{ImageNet}\label{exp:imagenet} On ImageNet, we use 3000 images from the \textbf{training set} to evaluate the reward function in order to prevent over-fitting. The latency is measured with $224\times224$ input throughout the experiments. \subsubsection{Push the Limit of Fine-grained Pruning.} \begin{figure}[t] \centering \includegraphics[height=4.6cm]{figures/finegrained} \caption{ The pruning policy (sparsity ratio) given by our reinforcement learning agent for ResNet-50. With 4 stages of iterative pruning, we find very salient sparsity pattern across layers: peaks are $1\times1$ convolution, crests are $3\times3$ convolution. \textbf{The reinforcement learning agent automatically learns that $3\times3$ convolution has more redundancy than $1\times1$ convolution and can be pruned more.} } \label{fig:stage4} \end{figure} \begin{figure}[t] \centering \includegraphics[width=0.93\linewidth]{figures/res50ratio.pdf} \caption{Our reinforcement learning agent (AMC\xspace) can prune the model to a lower density compared with human experts without losing accuracy. (Human expert: $3.4\times$ compression on ResNet50. AMC\xspace\ : $5\times$ compression on ResNet50.)} \label{fig:finegrain} \end{figure} Fine-grained pruning method prunes neural networks based on individual connections to achieve sparsity in both weights and activations, which is able to achieve higher compression ratio and can be accelerated with specialized hardware such as~\cite{han2016eie,han2017ese,parashar2017scnn}. However, it requires iterative prune \& fine-tune procedure to achieve decent performance~\cite{han2015learning}, and single-shot pruning without retraining will greatly hurt the prediction accuracy with large compression rate (say $4\times$), which cannot provide useful supervision for reinforcement learning agent. To tackle the problem, we follow the settings in \cite{han_thesis} to conduct 4-iteration pruning \& fine-tuning experiments, where the overall density of the full model is set to [50\%, 35\%, 25\% and 20\%] in each iteration. For each stage, we run AMC\xspace\ to determine the sparsity ratio of each layer given the overall sparsity. The model is then pruned and fine-tuned for 30 epochs following common protocol. With that framework, we are able to push the expert-tuned compression ratio of ResNet-50 on ImageNet from $3.4\times$ to $\mathbf{5\times}$ (see Figure~\ref{fig:finegrain}) without loss of performance on ImageNet (original ResNet50's [top-1, top-5] accuracy=[76.13\%, 92.86\%]; AMC\xspace\ pruned model's accuracy=[76.11\%, 92.89\%]). The density of each layer during each stage is displayed in Figure~\ref{fig:stage4}. The peaks and crests show that the RL agent automatically learns to prune $3\times3$ convolutional layers with larger sparsity, since they generally have larger redundancy; while prunes more compact $1\times1$ convolutions with lower sparsity. The density statistics of each block is provided in Figure~\ref{fig:finegrain}. We can find that the density distribution of AMC\xspace is quite different from human expert's result shown in Table 3.8 of \cite{han_thesis}, suggesting that AMC can fully explore the design space and allocate sparsity in a better way. \subsubsection{Comparison with Heuristic Channel Reduction.}\label{sec:exp:vgg} \begin{table}[t] \centering \caption{Learning-based model compression (AMC) outperforms rule-based model compression. Rule-based heuristics are suboptimal. (for reference, the baseline top-1 accuracy of VGG-16 is 70.5\%, MobileNet is 70.6\%, MobileNet-V2 is 71.8\%).} \label{tab:vggde} \begin{tabular}{c|c|c|c} \hline & policy & FLOPs & $\Delta Acc~\%$ \\ \hline \multirow{5}{*}{\begin{tabular}[c]{@{}c@{}}VGG-16 \end{tabular}} & FP (handcraft)~\cite{li2016pruning} & \multirow{5}{*}{20\%} & -14.6 \\ \cline{2-2} \cline{4-4} & RNP (handcraft)~\cite{lin2017runtime} & & -3.58 \\ \cline{2-2} \cline{4-4} & SPP (handcraft)~\cite{wang2017structured} & & -2.3 \\ \cline{2-2} \cline{4-4} & CP (handcraft)~\cite{he2017channel} & & -1.7 \\ \cline{2-2} \cline{4-4} & \textbf{AMC (ours)} & & -\textbf{1.4} \\ \hline \hline \multirow{4}{*}{MobileNet} & uniform (0.75-224)~\cite{howard2017mobilenets} & 56\% & -2.5 \\ \cline{2-4} & \textbf{AMC (ours)}& 50\% & \textbf{-0.4} \\ \cline{2-4} & uniform (0.75-192)~\cite{howard2017mobilenets} & 41\% & -3.7 \\ \cline{2-4} & \textbf{AMC (ours)}& 40\% & \textbf{-1.7} \\ \hline \hline \multirow{2}{*}{\ MobileNet-V2\ } & uniform (0.75-224)~\cite{sandler2018inverted} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}70\% \end{tabular}} & -2.0 \\ \cline{2-2} \cline{4-4} & \textbf{AMC (ours)}& & \textbf{-1.0} \\ \hline \end{tabular} \end{table} Here we compare AMC with existing state-of-the-art channel reduction methods: FP~\cite{li2016pruning}, RNP~\cite{lin2017runtime} and SPP~\cite{wang2017structured}. All the methods proposed a heuristic strategy to design the pruning ratio of each layer. FP~\cite{li2016pruning} proposed a sensitive analysis scheme to estimate the sensitivity of each layer by evaluating the accuracy with single layer pruned. Layers with lower sensitivity are pruned more aggressively. Such method assumes that errors of different pruned layers can be summed up linearly, which does not stand according to our experiments. RNP~\cite{lin2017runtime} groups all convolutional channels into 4 sets and trains a RL agent to decide on the 4 sets according to input image. However, the action space is very rough (only 4 actions for each layer), and it cannot reduce the model size. SPP~\cite{wang2017structured} applies PCA analysis to each layer and takes the reconstruction error as the sensitivity measure to determine the pruning ratios. Such analysis is conducted based on one single layer, failing to take the correlations between layers into consideration. We also compare our method to the original channel pruning paper (CP in Table~\ref{tab:vggde}), in which the sparsity ratios of pruned VGG-16~\cite{simonyan2014very} are carefully tuned by human experts (\verb|conv5| are skipped, sparsity ratio for \verb|conv4| and remaining layers is $1:1.5$). The results of pruned VGG-16 are presented in Table~\ref{tab:vggde}. Consistent with our CIFAR-10 experiments (Sec.~\ref{sec:cifarresource}), AMC outperforms all heuristic methods by more than \textbf{0.9\%}, and beats human expert by \textbf{0.3\%} without any human labor. Apart from VGG-16, we also test AMC on modern efficient neural networks MobileNet~\cite{howard2017mobilenets} and MobileNet-V2~\cite{sandler2018inverted}. Since the networks have already been very compact, it is much harder to further compress them. The easiest way to reduce the channels of a model is to use uniform channel shrinkage, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot use a width multiplier to uniformly reduce the channels of each layer with a fixed ratio. Both MobileNet and MobileNet-V2 present the performance of different multiplier and input sizes, and we compare our pruned result with models of same computations. The format are denoted as \emph{uniform (depth multiplier - input size)}. We can find that our method consistently outperforms the uniform baselines. Even for the current state-of-the-art efficient model design MobileNet-V2, AMC can still improve its accuracy by $1.0\%$ at the same computation~(Table \ref{tab:vggde}). The pareto curve of MobileNet is presented in Figure~\ref{fig:sub1}. \subsubsection{Speedup Mobile Inference.} \begin{figure}[t] \centering \begin{subfigure}{.5\textwidth} \centering \includegraphics[width=1\linewidth]{figures/mobile_new2.pdf} \caption{Accuracy v.s. MACs} \label{fig:sub1} \end{subfigure}% \begin{subfigure}{.49\textwidth} \centering \includegraphics[width=1\linewidth]{figures/vsNetAdapt.pdf} \caption{Accuracy v.s. Inference time} \label{fig:sub2} \end{subfigure} \caption{ (a) Comparing the accuracy and MAC trade-off among AMC\xspace, human expert, and unpruned MobileNet. AMC\xspace strictly dominates human expert in the pareto optimal curve. (b) Comparing the accuracy and latency trade-off among AMC\xspace, NetAdapt, and unpruned MobileNet. AMC\xspace significantly improves the pareto curve of MobileNet. Reinforcement-learning based AMC surpasses heuristic-based NetAdapt on the pareto curve (inference time both measured on Google Pixel 1).} \label{fig:test} \end{figure} \begin{table}[t] \setlength{\tabcolsep}{3pt} \centering \caption{AMC speeds up MobileNet. Previous attempts using \textit{rule-based} policy to prune MobileNet lead to significant accuracy degradation~\cite{li2016pruning} while AMC\xspace use \textit{learning-based} pruning which well preserve the accuracy. On Google Pixel-1 CPU, AMC\xspace achieves $1.95\times$ measured speedup with batch size one, while saving the memory by 34\%. On NVIDIA Titan XP GPU, AMC\xspace achieves $1.53\times$ speedup with batch size of 50. The input image size is 224$\times$224 for all experiments and no quantization is applied for apple-to-apple comparison.} \scalebox{0.91}{ \begin{tabular}{c|c|cc|cc|ccc} \hline \multirow{2}{*}{} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Million\\ MAC\end{tabular}} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}top-1\\ acc.\end{tabular}} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}top-5\\ acc.\end{tabular}} & \multicolumn{2}{c|}{GPU} & \multicolumn{3}{c}{Android} \\ & & & & \begin{tabular}[c]{@{}c@{}}latency \\\end{tabular} & \begin{tabular}[c]{@{}c@{}}speed \\\end{tabular} & \begin{tabular}[c]{@{}c@{}}latency \\\end{tabular} & \begin{tabular}[c]{@{}c@{}}speed \\\end{tabular} & \begin{tabular}[c]{@{}c@{}}memory \\\end{tabular} \\ \hline\hline \begin{tabular}[c]{@{}c@{}}100\%\\ MobileNet\end{tabular} & 569 & 70.6\% & 89.5\% & 0.46ms & 2191 fps & 123.3ms & 8.1 fps & 20.1MB \\ \begin{tabular}[c]{@{}c@{}}75\%\\ MobileNet\end{tabular} & 325 & 68.4\% & 88.2\% & 0.34ms & 2944 fps & 72.3ms &13.8 fps & 14.8MB \\ \hline NetAdapt \cite{yang2018netadapt} & - & 69.8\% & - & - & - & 70.0ms & 14.3 fps & - \\\hline \begin{tabular}[c]{@{}c@{}}\textbf{AMC} \\ ($50\%$ FLOPs)\\ \end{tabular} & 285 & 70.5\% & 89.3\% & 0.32ms & \begin{tabular}[c]{@{}c@{}}\textbf{3127 fps}\\ (\textbf{1.43}$\times$)\end{tabular} & 68.3ms & \begin{tabular}[c]{@{}c@{}}\textbf{14.6 fps}\\ (\textbf{1.81}$\times$)\end{tabular} & 14.3MB \\ \begin{tabular}[c]{@{}c@{}}\textbf{AMC} \\ ($50\%$ Latency) \end{tabular} & 272 & 70.2\% & 89.2\% & 0.30ms & \begin{tabular}[c]{@{}c@{}}\textbf{3350 fps}\\ (\textbf{1.53}$\times$)\end{tabular} & 63.3ms & \begin{tabular}[c]{@{}c@{}}\textbf{16.0 fps}\\ (\textbf{1.95}$\times$)\end{tabular} & 13.2MB \\ \hline \end{tabular} } \label{tab:android} \end{table} Mobile inference acceleration has drawn people's attention in recent years. Not only can AMC optimize FLOPs and model size, it can also optimize the inference latency, directly benefiting mobile developers. For all mobile inference experiments, we use TensorFlow Lite framework for timing evaluation. We prune MobileNet~\cite{howard2017mobilenets}, a highly compact network consisting of depth-wise convolution and point-wise convolution layers, and measure how much we can improve its inference speed. Previous attempts using \textit{hand-crafted} policy to prune MobileNet led to significant accuracy degradation~\cite{li2016pruning}: pruning MobileNet to 75.5\% original parameters results in 67.2\% top-1 accuracy\footnote{http://machinethink.net/blog/compressing-deep-neural-nets/}, which is even worse than the original 0.75 MobileNet (61.9\% parameters with 68.4\% top-1 accuracy). However, our AMC\xspace pruning policy significantly improves pruning quality: on ImageNet, AMC\xspace-pruned MobileNet achieved 70.5\% Top-1 accuracy with 285\ MFLOPs, compared to the original 0.75 MobileNet's 68.4\% Top-1 accuracy with 325 MFLOPs. As illustrated in Figure~\ref{fig:sub1}, human expert's hand-crafted policy achieves slightly worse performance than that of the original MobileNet under $2\times$ FLOPs reduction. However, with AMC\xspace, we significantly raise the pareto curve, improving the accuracy-MACs trade-off of original MobileNet. By substituting FLOPs with latency, we can change from FLOPs-constrained search to latency-constrained search and directly optimize the inference time. Our experiment platform is Google Pixel 1 with Qualcomm Snapdragon 821 SoC. As shown in Figure~\ref{fig:sub2}, we greatly reduce the inference time of MobileNet under the same accuracy. We also compare our learning based policy with a heuristic-based policy~\cite{yang2018netadapt}, and AMC better trades off accuracy and latency. Furthermore, since AMC uses the validation accuracy before fine-tuning as the reward signal while \cite{yang2018netadapt} needs local fine-tuning after each step, AMC is more sampling-efficient, requiring fewer GPU hours for policy search. We show the detailed statistics of our pruned model in Table~\ref{tab:android}. Model searched with 0.5$\times$ FLOPs and 0.5$\times$ inference time are profiled and displayed. For 0.5$\times$ FLOPs setting, we achieve \textbf{1.81}$\times$ speed up on a Google Pixel 1 phone, and for 0.5$\times$ FLOPs setting, we accurately achieve \textbf{1.95}$\times$ speed up, which is very close to actual 2$\times$ target, showing that AMC can directly optimize inference time and achieve accurate speed up ratio. We achieve 2.01$\times$ speed up for 1$\times$1 convolution but less significant speed up for depth-wise convolution due to the small computation to communication ratio. AMC\xspace compressed models also consumes less memory. On GPUs, we also achieve up to 1.5$\times$ speedup, less than mobile phone, which is because a GPU has higher degree of parallelism than a mobile phone. \subsubsection{Generalization Ability.} \begin{table}[t] \centering \caption{Compressing Faster R-CNN with VGG16 on PASCAL VOC 2007. Consistent with classification task, AMC\xspace also results in better performance under the same compression ratio on object detection task.} \label{tab:detection} \begin{tabular}{c|c|c} \hline & mAP (\%) & mAP [0.5, 0.95] (\%) \\ \hline baseline & 68.7 & 36.7 \\ \hline $2\times$ handcrafted~\cite{he2017channel} & 68.3 (-0.4) & 36.7 (-0.0) \\ \hline $4\times$ handcrafted~\cite{he2017channel} & 66.9 (-1.8) & 35.1 (-1.6) \\ \hline $4\times$ handcrafted~\cite{zhang2016accelerating} & 67.8 (-0.9) & 36.5 (-0.2) \\ \hline $4\times$ AMC (ours) & \textbf{68.8 (+0.1)} & \textbf{37.2 (+0.5)} \\ \hline \end{tabular} \end{table} We evaluate the generalization ability of AMC\xspace on PASCAL VOC object detection task~\cite{pascal-voc-2007}. We use the compressed VGG-16 (from Sec \ref{sec:exp:vgg}) as the backbone for Faster R-CNN~\cite{ren2015faster}. In \tbl{tab:detection}, AMC\xspace\ achieves \textbf{0.7\%} better mAP[0.5, 0.95]) than the best hand-crafted pruning methods under the same compression ratio. AMC\xspace\ even surpasses the baseline by \textbf{0.5\%} mAP. We hypothesize this improvement as that the optimal compression policy found by the RL agent also serves as effective regularization. \section{Conclusion} Conventional model compression techniques use hand-crafted features and require domain experts to explore a large design space and trade off between model size, speed, and accuracy, which is usually suboptimal and labor-consuming. In this paper, we propose AutoML for Model Compression\ (AMC\xspace), which leverages reinforcement learning to automatically search the design space, greatly improving the model compression quality. We also design two novel reward schemes to perform both resource-constrained\ compression and accuracy-guaranteed\ compression. Compelling results have been demonstrated for MobileNet, MobileNet-V2, ResNet and VGG on Cifar and ImageNet. The compressed model generalizes well from classification to detection tasks. On the Google Pixel 1 mobile phone, we push the inference speed of MobileNet from 8.1 fps to 16.0 fps. AMC\xspace facilitates efficient deep neural networks design on mobile devices. \section*{Acknowledgements} We thank Quoc Le, Yu Wang and Bill Dally for the supportive feedback. We thank Jiacong Chen for drawing the cartoon on the left of Figure 1. \clearpage \bibliographystyle{splncs04}

\section{Introduction}\label{sec:int} With almost 100 gravitational wave (GW) detections up to date, we can firmly say that the era of GW astronomy has started~\cite{GWTC1,GWTC2,GWTC3}. A particularly interesting application in GW astronomy is to use standard sirens -- GW events with an electromagnetic (EM) counterpart -- to measure the expansion of the universe~\cite{schutz_1986,holz_hughes_2005,ligo_2017d,chen_fishbach_2018,mukherjee_ghosh_2020,chen_haster_2022}. For such a source, we can use the luminosity distance from GW detection together with the redshift from its EM counterpart to determine the Hubble constant. However, the relation between the Hubble constant, a redshift of the source, and its luminosity distance is derived under the assumption that the source is at rest relative to the expanding universe and thus a motion of the source needs to be corrected when inferring the Hubble constant~\cite{dodelson_2020}. Different astrophysical scenarios suggest that GW sources are moving with velocities ranging from a few $100\,{\rm km\,s^{-1}}$ up to several percent of the speed of light~\cite{lisa_2022}. Three particularly interesting cases that affect most if not all GW sources are (i) gravitational kicks induced by the merger of the source~\cite{pretorius_2005,herrmann_hinder_2007a,herrmann_hinder_2007b,herrmann_hinder_2007c,koppitz_pollney_2007,baker_centrella_2006,baker_boggs_2008,campanelli_lousto_2006,campanelli_lousto_2007a,campanelli_lousto_2007b,healy_herrmann_2009}, (ii) binary black holes (BBHs), binary neutron stars (BNSs), and black hole neutron star binaries (BHNSB) orbiting a supermassive black hole (SMBH) in the center of a galaxy~\cite{antonini_perets_2012,mckernan_ford_2012,bartos_kocsis_2017,stone_metzger_2017,chen_han_2018,chen_li_2019,addison_gracia-linares_2019,tagawa_haiman_2020,pan_lyu_2021,derdzinski_mayer_2022}, and (iii) a peculiar velocity of the host galaxy~\cite{zinn_west_1984,bahcall_1988,girardi_fadda_1996,carlberg_yee_1996,springel_white_2001,ruel_bazin_2014,scrimgeour_davis_2016,colin_mohayaee_2017}. In particular, the peculiar velocity of the host galaxy has been considered in the literature when correcting for the effect of velocity~\cite{wang_wang_2018,howlett_davis_2020,nicolaou_lahav_2020,mukherjee_lavaux_2021}. However, the corrections only focus on the effect the velocity has on the redshift of the EM counterpart but ignore the effect of the velocity on the apparent distance. Different methods have been proposed to measure the velocity of a GW source by either considering a time-dependent phase shift~\cite{gerosa_moore_2016,meiron_kocsis_2017,inayoshi_tamanini_2017,tamanini_klein_2019,wong_baibhav_2019,torres-orjuela_chen_2020,yu_chen_2021,strokov_fragione_2022} or a change in the spherical modes of the GW~\cite{calderon-bustillo_clark_2018,varma_gerosa_2019,varma_biscoveanu_2020,varma_isi_2020,torres-orjuela_chen_2021,torres-orjuela_amaro-seoane_2021}. From the information of the source's velocity from GWs, we can correct the apparent distance for its change due to the relativistic redshift as well as set tighter constraints on the redshift of the EM counterpart. Taking these two corrections into account will greatly reduce the estimation error for the Hubble constant thus helping to ease the Hubble tension~\cite{di-valentino_mena_2021}. Therefore, it is crucial to have a detailed understanding of the effect of a source's velocity on the GW's frequency and the obtained apparent distance in an expanding universe. In \sref{sec:crs}, we start revising the effect of an expansion of the universe on the frequency of GWs and the luminosity distance of the source. We then study the effect of a source's motion on the frequency and the apparent distance of the source in a static universe in \sref{sec:rrs} to then extend our analysis in \sref{sec:trs} to the case of a moving source in an expanding universe. Most remarkable, we find that the apparent distance of a moving source scales with the square of the relativistic redshift, in contrast to the case of cosmological redshift where the distance scales with a linear factor of the redshift. In \sref{sec:obs}, we analyze the implication of our results on the estimated chirp mass and distance of the source, where we show that for realistic astrophysical scenarios the estimation error of the two can go up to several percent. We, further, discuss the square factor of the relativistic redshift in the apparent distance and show that it can be attributed to a velocity-dependent change in the amplitude of GWs. We study the implications of our results for the measurement of the Hubble constant using standard sirens for the case where only the redshift of the EM counterpart is considered and when the velocity of the source is completely ignored. We find that for a typical peculiar velocity of the source the error can already reach 1\,\% for the first case while for the second case the error can go up to several 10\,\% depending on the distance of the source. We finalise drawing conclusions in \sref{sec:con}. Throughout this paper, unless otherwise indicated, we use geometrical units in which the gravitational constant and the speed of light are equal to one (i.e., $G=c=1$). Furthermore, we use bold symbols for three-dimensional spatial vectors and a tilde to mark four-dimensional space-time vectors. \section{Cosmological redshift}\label{sec:crs} We start considering the effect of the cosmological expansion on the frequency of a GW in a Friedmann-Lemaître-Robertson-Walker (FLRW) universe, described by the metric~\cite{misner_thorne_1973} \begin{equation}\label{eq:flrw} \mathrm{d} s^2 = \mathrm{d} t^2 - a^2(t)\left(\frac{\mathrm{d} r^2}{1-kr^2} + r^2\mathrm{d}\theta^2 + r^2\sin^2(\theta)\mathrm{d}\phi^2\right). \end{equation} Here, $a(t)$ is the time-dependent scale factor that describes the expansion of the universe while $k$ describes its curvature, where $k=0$ corresponds to a flat universe, $k>0$ corresponds to a closed universe, and $k<0$ corresponds to an open universe. For simplicity and because observational data suggests that our universe is relatively flat~\cite{planck_2020,spergel_bean_2007}, we restrict the analysis in this paper to the case of a flat universe $k=0$. However, in principle, the methods developed in this paper can be extended to the cases of a close and an open universe with some additional effort. For the redshift of a closed or an open universe when only considering the cosmological expansion see, e.g., Ref.~\cite{maggiore_2008}. For the derivation of the cosmological redshift of GWs, we follow the procedure introduced in Ref.~\cite{maggiore_2008} where we consider a source located at a comoving distance $r_s$ and an observer located at $r=0$. Suppose the source emits a wave with a crest at a time $t_s$ and a second crest at a time $t_s+\Delta t_s$. Further, assume the first cress is detected by the observer at a time $t_o$ while the second crest is detected at a time $t_o+\Delta t_o$. Because in general relativity GWs travel at the speed of light~\cite{misner_thorne_1973}, we can describe the path of the first crest by setting $\mathrm{d} s^2=0$ in \eq{eq:flrw} and solving the integrals \begin{equation}\label{eq:cri1} \int_{t_s}^{t_o}\frac{\mathrm{d} t}{a(t)} = -\int_0^{r_s}\mathrm{d} r, \end{equation} while for the second crest, we have to solve the integrals \begin{equation}\label{eq:cri2} \int_{t_s+\Delta t_s}^{t_o+\Delta t_o}\frac{\mathrm{d} t}{a(t)} = -\int_0^{r_s}\mathrm{d} r. \end{equation} Note, that in both cases the integral over $r$ is the same because in comoving coordinates the coordinate locations of the source and the observer do not change if they are initially at rest~\cite{maggiore_2008}. Solving the integrals, then subtracting the results of \eq{eq:cri1} from the results of \eq{eq:cri2}, and expanding to linear order in $\Delta t_s$ and $\Delta t_o$, we find \begin{equation}\label{eq:ctd1} \Delta t_o = \frac{a(t_o)}{a(t_s)}\Delta t_s. \end{equation} Therefore, we find that in an expanding universe there is a dilation of the time measured by an observer far from the source. Using the usual convention that ${a(t_o)=1}$~\cite{dodelson_2020}, we define the cosmological redshift as \begin{equation}\label{eq:dcrs} (1+z_{\rm cos}) := a^{-1}(t_s), \end{equation} leading to the following relation for the time measured by the observer's and source's clocks \begin{equation}\label{eq:ctd2} \mathrm{d} t_o = (1+z_{\rm cos})\mathrm{d} t_s. \end{equation} Using this result and that the frequency of a wave is the time derivative of the coordinate invariant phase of the wave, we get that the frequency in the observer frame $f_o$ and the frequency in the source frame $f_s$ fulfill \begin{equation}\label{eq:crs} f_o = \frac{f_s}{(1+z_{\rm cos})}. \end{equation} After deriving the effect of cosmological expansion on the frequency of a GW, we analyze how the apparent distance of the source $D_{\rm app}$ is affected by the cosmological redshift. The apparent distance can be determined using the energy flux $F$ of the source that corresponds to the energy $E$ irradiated per unit time $\mathrm{d} t$ per unit area $\mathrm{d} A$ \begin{equation}\label{eq:flux} F := \frac{\mathrm{d} E}{\mathrm{d} A\mathrm{d} t}. \end{equation} The apparent distance of a GW is then defined as the radius of a sphere for which the absolute luminosity of the source $L := \mathrm{d} E_s/\mathrm{d} t_s$ equals the flux in a unit area \begin{equation}\label{eq:dad} F = \frac{L}{D_{\rm app}^2\mathrm{d}\Omega_s}, \end{equation} where we used that for a sphere $\mathrm{d} A = r^2\mathrm{d}\Omega$. In the case of cosmological expansion, the apparent distance is equal to the luminosity distance $D_L$~\cite{maggiore_2008} but we use the term apparent distance to distinguish the two in the more general case where the source is not only affected by cosmological expansion. To get the apparent distance for a GW source in an expanding universe, we use that the energy of a GW transforms in the same way as its frequency~\cite{maggiore_2008} \begin{equation}\label{eq:ecr} E_o = \frac{E_s}{(1+z_{\rm cos})}, \end{equation} that the time is dilated as in \eq{eq:ctd2}, and that $\mathrm{d}\Omega$ is not affected by the cosmological expansion because it is isotropic while the distance between the source and the observer is $a(t_o)r_d$. However, using that $a(t_0)=1$, we have that the distance between the source and the observer is equal to the comoving distance $r_d$. Putting everything together, we obtain \begin{equation}\label{eq:fadc} F = \frac{L}{(1+z_{\rm cos})^2r_d^2\mathrm{d}\Omega_s} \end{equation} and thus find for the apparent distance \begin{equation}\label{eq:adc} D_{\rm app} = (1+z_{\rm cos})r_d. \end{equation} Therefore, we recover the classical result that the apparent distance, which in this case is equal to the luminosity distance $D_L$, equals the comoving distance times the same factor by which the frequency is redshifted~\cite{sathyaprakash_schutz_2009}. \section{Relativistic redshift}\label{sec:rrs} In this section, we derive the relativistic redshift and the change of the apparent distance induced by a relative motion between the source and the observer in a static universe using the same methods as in the previous section. Therefore, we consider a Minkowsky space with the metric~\cite{schutz_2009} \begin{equation}\label{eq:mink} \mathrm{d} s^2 = \mathrm{d} t^2 - \mathrm{d} r^2 - r^2\mathrm{d}\theta^2 - r^2\sin^2(\theta)\mathrm{d}\phi^2. \end{equation} The case of a moving source in a flat but expanding FLRW universe is discussed in the next section. Before we derive the effect of the velocity, we need to fix the coordinates systems (COs) for the observer $O$ and the source $S$. We set the origins of the two COs to agree at the time $t=t'=0$, where we mark quantities in $S$ with a prime while quantities in $O$ remain unprimed. In $O$, the source moves with a velocity $\boldsymbol{v}$ while the observer remains at the origin of the CO and in $S$ the observer moves with a velocity $\boldsymbol{v}'$ while the source remains at the radial distance $r'_s$. To simplify the calculations, we set $O$ to have its $z$-axis parallel to the velocity of the source $\boldsymbol{v}$ and $S$ to have its $z$-axis anti-parallel to the velocity of the observer $\boldsymbol{v}'$. With this setting the two COs are parallel and the polar angle $\theta$ is affected by the relative velocity while the azimuthal angle $\phi$ is not. We consider again two subsequent crests of a GW, where the first is emitted at the event $\hat{r}'_s = (t'_s, r'_s\sin(\theta'_s)\cos(\phi'_s), r'_s\sin(\theta'_s)\sin(\phi'_s), r'_s\cos(\theta'_s))$ and the second at the event $\hat{r}'_{s+\Delta} = (t'_s + \Delta t'_s, r'_s\sin(\theta'_s)\cos(\phi'_s), r'_s\sin(\theta'_s)\sin(\phi'_s), r'_s\cos(\theta'_s))$. The observer detects the first crest at the event $\hat{r}_o = (t_o,0,0,0)$ and the second at the event $\hat{r}_{o+\Delta} = (t_o+\Delta t_o,0,0,0)$. To determine the relativistic redshift we need to solve the integrals we get from setting the metric in \eq{eq:mink} equal to zero. However, this step requires the boundaries of the integrals to be expressed in the same frame and thus we transform $\hat{r}'_s$ and $\hat{r}'_{s+\Delta}$ into the observer frame using the inverse Lorentz transformation, i.e., $\hat{r}_s = \Lambda(-v_z)\hat{r}'_s$ and $\hat{r}_{s+\Delta} = \Lambda(-v_z)\hat{r}'_{s+\Delta}$~\cite{schutz_2009}. In the observer frame, we get for the emission time of the first crest \begin{equation}\label{eq:tce1} t_s = \gamma(t'_s + vr'_s\cos(\theta'_s)) \end{equation} and for the emission time of the second crest \begin{equation}\label{eq:tce2} t_{s+\Delta} = \gamma(t'_s + \Delta t'_s + vr'_s\cos(\theta'_s)), \end{equation} where $\gamma = (1-v^2)^{-1/2}$ is the Lorentz factor. Furthermore, we have that the radial coordinate of the emitter changes from \begin{equation}\label{eq:rce1} r_s = \sqrt{{r'_s}^2\sin^2(\theta'_s) + \gamma^2(r'_s\cos(\theta'_s) + vt'_s)^2} \end{equation} to \begin{equation}\label{eq:rce2} r_{s+\Delta} = \sqrt{{r'_s}^2\sin^2(\theta'_s) + \gamma^2(r'_s\cos(\theta'_s) + v(t'_s + \Delta t'_s))^2}, \end{equation} while the polar angle changes from \begin{equation}\label{eq:pce1} \cos(\theta_s) = \frac{\gamma(r'_s\cos(\theta'_s) + vt'_s)}{r_s} \end{equation} to \begin{equation}\label{eq:pce2} \cos(\theta_{s+\Delta}) = \frac{\gamma(r'_s\cos(\theta'_s) + v(t'_s + \Delta t'_s))}{r_{s+\Delta}}, \end{equation} and the azimuthal angle remains unaffected. Note that to linear order in $\Delta t'_s$, which is the order to which we expand our results, the difference between $\cos(\theta_s)$ and $\cos(\theta_{s+\Delta})$ scales as $1/r'_s$, and hence it becomes negligible for typical GW sources. The integrals we need to consider are \begin{equation}\label{eq:rri1} \int_{t_s}^{t_o}\mathrm{d} t = -\int_0^{r_s}\mathrm{d} r \end{equation} and \begin{equation}\label{eq:rri2} \int_{t_{s+\Delta}}^{t_o + \Delta t_o}\mathrm{d} t = -\int_0^{r_{s+\Delta}}\mathrm{d} r. \end{equation} Solving the integrals, where we use \eqt{eq:tce1}{eq:rce2} for the boundaries, then subtracting the results of \eq{eq:rri1} from \eq{eq:rri2}, and expanding again to linear order in $\Delta t_o$ and $\Delta t'_s$, we get \begin{equation}\label{eq:rtd1} \Delta t_o = \gamma(1-v\cos(\theta_s))\Delta t'_s, \end{equation} where we used \eq{eq:pce1} to simplify the result. Therefore, we recover the classical result for the time dilation due to a relative velocity \begin{equation}\label{eq:rtd2} \mathrm{d} t_o = \gamma(1-v\cos(\theta))\mathrm{d} t_s. \end{equation} Defining the relativistic redshift as \begin{equation}\label{eq:drrs} (1+z_{\rm rel}) := \gamma(1-v\cos(\theta)) \end{equation} and using again that the frequency of a GW is the time-derivative of its phase, we get the well-known expression for the relativistic redshift \begin{equation}\label{eq:rrs} f_o = \frac{f_s}{(1+z_{\rm rel})}. \end{equation} Next, we derive the apparent distance for a moving source starting again from \eq{eq:flux}. For a moving source, the area in which the GWs are radiated $\mathrm{d} A = r^2\mathrm{d}\Omega$ is affected by a change of the radial coordinate $r$ but also due to a change of the solid angle $\mathrm{d}\Omega$. For the transformation of the solid angle, we use the transformation of the direction of the radiation described by the wave vector $\hat{k} = (f_o, f_o\sin(\theta)\cos(\phi), f_o\sin(\theta)\sin(\phi), f_o\cos(\theta_o))$. Using that the wave vector transforms as $\hat{k}' = \Lambda(v_z)\hat{k}$, we find \begin{equation}\label{eq:tpa} \cos(\theta') = \frac{\cos(\theta)-v}{1-v\cos(\theta)}. \end{equation} From this equation follows \begin{equation}\label{eq:tdpa} \mathrm{d}\cos(\theta') = \frac{\mathrm{d}\cos(\theta)}{\gamma^2(1-v\cos(\theta))^2}, \end{equation} while the azimuthal angle is not affected by the velocity ($\mathrm{d}\phi'=\mathrm{d}\phi)$. Using $\mathrm{d}\Omega = \mathrm{d}\phi\mathrm{d}\cos(\theta)$ together with \eq{eq:tdpa}, we get \begin{equation}\label{eq:rsa} \mathrm{d}\Omega = (1+z_{\rm rel})^2\mathrm{d}\Omega', \end{equation} where we used the definition of the relativistic redshift in \eq{eq:drrs}. We find the transformation of the radial coordinate from the vector describing the wavefront of the GW at the time it reaches the observer $\hat{r}_d = (r_d,r_d\sin(\theta)\cos(\phi),r_d\sin(\theta)\sin(\phi),r_d\cos(\theta))$. Using that $\hat{r}'_d = \Lambda(v_z)\hat{r}_d$ and that the radial coordinate equals to the magnitude of the spatial vector, we find \begin{equation}\label{eq:rrc} r_d = \frac{r'_d}{(1+z_{\rm rel})}. \end{equation} Combining \eqs{eq:rsa}{eq:rrc} with \eq{eq:rtd2} and the fact that the energy of a GW transforms analogous to its frequency (cf. \eq{eq:rrs}) to compute the flux as defined in \eq{eq:flux}, we get \begin{equation}\label{eq:fadr} F = \frac{L}{(1+z_{\rm rel})^2{r'_d}^2\mathrm{d}\Omega'}. \end{equation} Therefore, we find for the apparent distance between the source and the observer \begin{equation}\label{eq:adrs} D_{\rm app} = (1+z_{\rm rel})r'_d. \end{equation} It equals the comoving distance at the time the GW reaches the observer times the relativistic redshift, which is in agreement with the result in Ref.~\cite{chen_li_2019}. However, $r'_d$ is the comoving distance in the source frame, and combining \eqs{eq:rrc}{eq:adrs}, we get that the apparent distance in the observer frame equals the comoving distance between times the square of the relativistic redshift \begin{equation}\label{eq:adro} D_{\rm app} = (1+z_{\rm rel})^2r_d. \end{equation} Therefore, we find a difference in the effect of a cosmological and a relativistic redshift on the apparent distance. While for an expanding universe the apparent distance scales with a linear factor of the redshift (cf. \eq{eq:adc}), the apparent distance scales with a factor of the relativistic redshift square for a moving source. \section{Total redshift}\label{sec:trs} In the two previous, sections we derived the cosmological redshift and its effect on the apparent distance in a flat FLRW universe and the relativistic redshift and how it affects the apparent distance in a Minkowsky universe. In this section, we derive the total redshift for a source moving in a flat FLRW universe and how the apparent distance changes. We use the FLRW metric in \eq{eq:flrw}, set up COs for the observer $O$ and the source $S$ as in \sref{sec:rrs} but use comoving Lorentz transformation \begin{equation}\label{eq:clt} \Lambda_c(v_z) = \left(\begin{array}{cccc} \gamma & 0 & 0 & -a(t)\gamma v \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ -\gamma v/a(t) & 0 & 0 & \gamma \end{array}\right). \end{equation} to transform between the two COs to be consistent with the metric. See \aref{app:clt} for a derivation of the comoving Lorentz transformations. We consider again the case of two subsequent crests of a GW and find for the time coordinate of the first crest \begin{equation}\label{eq:ctce1} t_s = \gamma(t'_s + a(t_s)vr'_s\cos(\theta'_s)), \end{equation} while for the time coordinate of the second crest, we have \begin{equation}\label{eq:ctce2} t_{s+\Delta} = \gamma(t'_s + \Delta t'_s + a(t_{s+\Delta})vr'_s\cos(\theta'_s)), \end{equation} where we use the time of the corresponding events in the scale factor. For the radial coordinate of the emitter when the first crest is emitted, we get \begin{equation}\label{eq:crce1} r_s = \sqrt{{r'_s}^2\sin^2(\theta'_s) + \gamma^2(r'_s\cos(\theta'_s) + vt'_s/a(t_s))^2} \end{equation} and for the second crest \begin{align}\label{eq:crce2} \nonumber &r_{s+\Delta} = \\ &\sqrt{{r'_s}^2\sin^2(\theta'_s) + \gamma^2(r'_s\cos(\theta'_s) + v(t'_s + \Delta t'_s)/a(t_{s+\Delta}))^2}. \end{align} The polar angle for the first crest has the form \begin{equation}\label{eq:cpce1} \cos(\theta_s) = \frac{\gamma(r'_s\cos(\theta'_s) + vt'_s/a(t_s))}{r_s} \end{equation} while the polar angle for the second crest is \begin{equation}\label{eq:cpce2} \cos(\theta_{s+\Delta}) = \frac{\gamma(r'_s\cos(\theta'_s) + v(t'_s + \Delta t'_s)/a(t_{s+\Delta}))}{r_{s+\Delta}}, \end{equation} and the azimuthal angle remains unaffected again. We have one more time that to linear order in $\Delta t'_s$, the difference between $\cos(\theta_s)$ and $\cos(\theta_{s+\Delta})$ scales as $1/r'_s$, and hence we can neglect their difference for typical GW sources. We compute the path of the crests in a FLRW universe by setting $\mathrm{d} s^2 = 0$ in \eq{eq:flrw}, to obtain \begin{equation}\label{eq:tri1} \int_{t_s}^{t_o}\frac{\mathrm{d} t}{a(t)} = -\int_0^{r_s}\mathrm{d} r \end{equation} and \begin{equation}\label{eq:tri2} \int_{t_{s+\Delta}}^{t_o+\Delta t_o}\frac{\mathrm{d} t}{a(t)} = -\int_0^{r_{s+\Delta}}\mathrm{d} r, \end{equation} where the boundaries are given in \eqt{eq:ctce1}{eq:crce2}. Solving the integrals, subtracting the results of \eq{eq:tri1} from \eq{eq:tri2}, and expanding to linear order in $\Delta t_o$ and $\Delta t'_s$, we find \begin{equation}\label{eq:ttd1} \frac{\Delta t_o}{a(t_o)} - \frac{\gamma\Delta t'_s}{a(t_s)} = -\gamma \frac{v}{a(t_s)}\cos(\theta_s)\Delta t'_s, \end{equation} where we used \eqs{eq:ctce1}{eq:cpce1} to simplify the expression. Next, we use that $a(t_o)=1$~\cite{dodelson_2020} and that $a(t_s) = (1+z_{\rm cos})^{-1}$, to get \begin{equation}\label{eq:ttd2} \mathrm{d} t_o = (1+z_{\rm cos})\gamma(1 - v\cos(\theta))\mathrm{d} t_s. \end{equation} Using the definition of the relativistic redshift in \eq{eq:drrs} and that the frequency is the time derivative of the phase, we get \begin{equation}\label{eq:trs} f_o = \frac{f_s}{(1+z_{\rm cos})(1+z_{\rm rel})}, \end{equation} thus finding that the total redshift is equal to the product of the cosmological redshift and the relativistic redshift \begin{equation}\label{eq:dtrs} (1+z_{\rm tot}) := (1+z_{\rm cos})(1+z_{\rm rel}). \end{equation} Our result agrees with Ref.~\cite{bonvin_caprini_2017}, who also considered the combined effect of velocity and cosmological expansion on the frequency of GWs, although their approach is restricted to low velocities. Therefore, our result represents an extension of their work. Finally, we derive the combined effect of cosmological expansion and the velocity of the source on the apparent distance of a GW source. To get the transformation of the radial coordinate, we consider again the transformation of the wave vector describing the wavefront. To consider the effect of the cosmological expansion, we need to multiply $r_d$ by $a(t_o)$ while we take $r'_d$ times $a(t_s)$. To include the effect of the velocity, we use the comoving Lorentz transformation. However, note that because we transform the wave vector seen by the observer (cf. before \eq{eq:rrc}) we have to use $t_o$ in the scale factor in the comoving Lorentz transformation. Therefore, using that $a(t_o)=1$, we get \begin{equation}\label{eq:crrc} r_d = \frac{r'_d}{(1+z_{\rm cos})(1+z_{\rm rel})}. \end{equation} where we, further, used that $a(t_s) = (1+z_{\rm cos})^{-1}$. For the solid angle, we use that it is not affected by the cosmological expansion and hence it transforms as in \eq{eq:rsa}. Inserting these results in \eq{eq:flux} together with the transformation of the time in \eq{eq:ttd2} and using that the energy transforms in the same way as the frequency in \eq{eq:trs}, we get \begin{equation}\label{eq:fadt} F = \frac{L(1+z_{\rm cos})^2}{(1+z_{\rm tot})^2{r'_d}^2\mathrm{d}\Omega'} \end{equation} Therefore, we find for the apparent distance \begin{equation}\label{eq:adt} D_{\rm app} = (1+z_{\rm tot})(1+z_{\rm rel})r_d, \end{equation} where we used \eq{eq:crrc} to express the result in the coordinates of the observer. We see that the apparent distance of a moving source in an expanding universe also contains an additional factor of $(1+z_{\rm rel})$. Furthermore, we see that the apparent distance of a source at rest in an expanding universe in \eq{eq:adc} is equal to the one in \eq{eq:adt} for a vanishing velocity while the apparent distance of a moving source in a Minkowsky space in \eq{eq:adro} can be obtained by considering a source at a cosmological redshift of zero, where the FLRW metric and the Minkowsky metric become equal. \section{Observables for high redshift and fast moving sources}\label{sec:obs} As we showed in the previous sections, the velocity of a GW source affects its frequency and apparent distance. In particular, we find that the apparent distance for moving sources needs to be corrected by an additional factor that does not appear in the case of a source at rest relative to the expanding universe at cosmological distances. Therefore, in this section, we analyze how the chirp mass and the apparent distance inferred from observations are affected when the observer ignores the velocity of the source. Before analyzing the effect of the velocity, we discuss the observables of GWs which are the amplitude of the wave $h_o$ (in two polarizations), the frequency of the wave $f_o$, and its time derivative $\dot{f}_o$~\cite{sathyaprakash_schutz_2009,chen_li_2019}. From the measurement of $f_o$ and $\dot{f}_o$ in the inspiral part of the signal, where the two merging compact objects are far apart and can be approximated as point masses, one can determine the chirp mass of the source in the observer's frame as \begin{equation}\label{eq:cmd} \mathcal{M}_o := \left(\frac{5\dot{f}_of^{-11/3}_o}{96\pi^{8/3}}\right)^{3/5}. \end{equation} It theoretically depends on the masses of the two compact objects as $\mathcal{M}_o = (m_{1,o}m_{2,o})^{3/5}(m_{1,o}+m_{2,o})^{-1/5}$ and describes how the frequency increases with time. Combining this result with the amplitude of the wave $h_o$, one can further determine the apparent distance of the source, using \begin{equation}\label{eq:ado} D_{\rm app} = (\pi f_o\mathcal{M}_o)^{2/3}\frac{4\mathcal{M}_o}{h_o}. \end{equation} From Eqs.~(\ref{eq:crs}), (\ref{eq:rrs}), and (\ref{eq:trs}) we see the frequency of a GW transforms $f_o = f_s/(1+z)$ between the observer's and the source's frame, where $(1+z)$ can be either the cosmological, the relativistic, or the total redshift, while from Eqs.~(\ref{eq:ctd2}), (\ref{eq:rtd2}), and (\ref{eq:ttd2}) it follows that $\dot{f}_o = \dot{f}_s/(1+z)^2$. Using these results in \eq{eq:cmd}, we get \begin{equation}\label{eq:rscm} \mathcal{M}_o = (1+z)\mathcal{M}_s, \end{equation} where $\mathcal{M}_s := ((5\dot{f}_sf^{-11/3}_s)/(96\pi^{8/3}))^{3/5}$ is the chirp mass in the source's frame. Therefore, we have that the chirp mass of a source redshifted by a factor $(1+z)$ is changed by the same factor. This behaviour is known as the `mass-redshift degeneracy' of GWs. From \eq{eq:ado}, we get that the observed amplitude of a GW goes as \begin{equation}\label{eq:amp} h_o = (\pi f_o\mathcal{M}_o)^{2/3}\frac{4\mathcal{M}_o}{D_{\rm app}}. \end{equation} Using that a source at rest and at a cosmological redshift ${(1+z_{\rm cos})}$ has an apparent (luminosity) distance $D_{\rm app} = (1+z_{\rm cos})r_d$ (cf. \eq{eq:adc}), the observed chirp mass in \eq{eq:rscm}, and \eq{eq:crs}, we get for the observed amplitude \begin{equation}\label{eq:oac} h_{o,\text{rest}} = (\pi f_s\mathcal{M}_s)^{2/3}\frac{4\mathcal{M}_s}{r_d} =: h_s. \end{equation} Therefore, we recover the classical result that the amplitude of GWs is independent of the cosmological redshift~\cite{schutz_1986}. The amplitude of a GW from a source that is moving can be determined from \eq{eq:amp} using the apparent distance in the observer frame in \eq{eq:adro} or \eq{eq:adt} for a source in the local universe or at a cosmological distance, respectively. In both cases, we get that the observed amplitude has the form \begin{equation}\label{eq:oar} h_{o,\text{moving}} = \frac{h_s}{(1+z_{\rm rel})}. \end{equation} Therefore, we find that the amplitude of a GW from a moving source differs from the amplitude of a source at rest by a factor $(1+z_{\rm rel})^{-1}$ independent of its cosmological redshift. The difference in the amplitude of a GW from a moving source and a source at rest can be understood from the fact that the solid angle is not invariant under (comoving) Lorentz transformations or, to put in a more physical picture, that the aberration of GWs leads to a focusing of the wave rays parallel to the velocity of the source~\cite{torres-orjuela_chen_2019,torres-orjuela_chen_2020,torres-orjuela_chen_2021}. The same effect is well known for light as a part of the so-called `headlight effect' or `relativistic beaming' and leads to an additional factor of $(1+z_{\rm rel})^2$ in the intensity of the light~\cite{johnson_teller_1982}. We have now all tools to analyze the impact of a motion of the source on the chirp mass and the apparent distance inferred from GW detection. Different astrophysical models predict a variety of velocities for GW sources, where we consider the following three particularly interesting cases: (i) gravitational kicks which in average have a velocity of $400\,{\rm km\,s^{-1}}$~\cite{amaro-seoane_chen_2016} while extreme kicks of up to $5000\,{\rm km\,s^{-1}}$ can occur for some special configurations~\cite{pretorius_2005,herrmann_hinder_2007a,herrmann_hinder_2007b,herrmann_hinder_2007c,koppitz_pollney_2007,baker_centrella_2006,baker_boggs_2008,campanelli_lousto_2006,campanelli_lousto_2007a,campanelli_lousto_2007b,healy_herrmann_2009}, (ii) BBHs, BNSs, and BHNSBs orbiting a SMBH in the center of a galaxy~\cite{antonini_perets_2012,mckernan_ford_2012,bartos_kocsis_2017,stone_metzger_2017,chen_han_2018,chen_li_2019,addison_gracia-linares_2019,tagawa_haiman_2020,pan_lyu_2021,derdzinski_mayer_2022}, where the velocity depends on the mass of the central SMBH and the radius of the orbit -- for a SMBH with a typical mass of $10^7\,{\rm M_\odot}$ and for an orbital radius of $5\,{\rm mpc}$ the orbital velocity is around $3000\,{\rm km\,s^{-1}}$ while for an orbital radius of $100\,R_s$ ($R_s = 2GM/c^2$ is the Schwarzschild radius of a BH with mass $M$) we have a velocity of around 7\,\% the speed of light, and (iii) the peculiar velocity of galaxies induced by the gravitational interaction of the host galaxy with its environment which in average has been measured to be around $1500\,{\rm km\,s^{-1}}$~\cite{zinn_west_1984,bahcall_1988,girardi_fadda_1996,carlberg_yee_1996,springel_white_2001,ruel_bazin_2014,scrimgeour_davis_2016,colin_mohayaee_2017}. Note that, in principle, these velocities are time-dependent and that the effect of the velocity on the redshift also depends on its orientation relative to the line-of-sight (LOS). However, we only want to show the principles of this effect, wherefore we assume the velocity to be constant and either parallel or anti-parallel to the LOS. For a moving source where the observer assumes it to be at rest, the error in the chirp mass can be estimated from the relative difference between the chirp mass the source would have if it would be at rest $\mathcal{M}_r = (1+z_{\rm cos})\mathcal{M}_s$ and the chirp mass observed for the moving source $\mathcal{M}_o = (1+z_{\rm tot})\mathcal{M}_s$ \begin{equation}\label{eq:dcm} \delta\mathcal{M} := \left|\frac{\mathcal{M}_r - \mathcal{M}_o}{\mathcal{M}_o}\right| = \left|1 - \sqrt{\frac{1 \mp v}{1 \pm v}}\right|, \end{equation} where the upper sign corresponds to the velocity being anti-parallel to the LOS (relativistic redshift) and the lower sign to the velocity being parallel to the LOS (relativistic blueshift). \fref{fig:ecm} shows the relative error in the measurement of the chirp mass of the source when the observer assumes it is at rest for different velocities. We see that the error is a monotonic function of the velocity and that it can reach 1\,\% for a velocity of $3000\,{\rm km\,s^{-1}}$ corresponding to a binary orbiting a SMBH of the mass $10^7\,{\rm M_\odot}$ at a distance of $5\,{\rm mpc}$ while for lower velocities like the average kick velocity and the peculiar velocity of galaxies the error is below 1\,\%. For extreme kicks of $5000\,{\rm km\,s^{-1}}$ the error is slightly higher than 1.5\,\% and it can go up to around 7\,\% for a binary orbiting a SMBH at $100\,R_s$. We further can see that there is a difference in the error when the velocity and the LOS are anti-parallel (source moving away from the observer) or parallel (source approaching the observer) but the difference is relatively small and only becomes relevant for highly relativistic velocities of several percent the speed of light. \begin{figure}[tpb] \centering \includegraphics[width=0.48\textwidth]{ecm.pdf} \caption{ The relative error in the chirp mass for a source moving at different velocities when the observer assumes the source is at rest. The red dashed-dotted line corresponds to the velocity and the LOS being anti-parallel while the blue dotted line corresponds to the two being parallel. The vertical lines represent typical velocities for the astrophysical scenarios described in the text. } \label{fig:ecm} \end{figure} The error in the apparent distance can be estimated considering the relative difference between the apparent distance $D_{{\rm app},r} = (1+z_{\rm cos})D_s$ ($D_s := (\pi f_s\mathcal{M}_s)^{2/3}(4\mathcal{M}_s/h_s))$ the source would have if it would be at rest and the observed apparent distance for the moving source $D_{{\rm app},o} = (1+z_{\rm tot})(1+z_{\rm rel})D_s$ \begin{equation}\label{eq:ead} \delta D_{\rm app} := \left|\frac{D_{{\rm app},r} - D_{{\rm app},o}}{D_{{\rm app},o}}\right| = \left|\frac{2v}{1 \pm v}\right|, \end{equation} where the plus sign corresponds to the velocity being anti-parallel to the LOS and the minus sign to the velocity and the LOS being parallel. In \fref{fig:ead}, we see that the relative error in the measurement of the apparent distance when the observer assumes the source is at rest is bigger than for the chirp mass. The error is still negligible for average kick velocities of $400\,{\rm km\,s^{-1}}$ but for a peculiar velocity of galaxies of $1500\,{\rm km\,s^{-1}}$ the error can go up to around 1\,\%. For a binary orbiting a SMBH of mass $10^7\,{\rm M_odot}$ at a distance of $5\,{\rm mpc}$ the error is above 2\,\% and for extreme kicks the error goes up to around 3\,\%. For highly relativistic velocities the dependence on the orientation of the velocity relative to the LOS becomes again more prominent so that for a binary orbiting a SMBH at a radius of $100\,R_s$ and the velocity being anti-parallel to the LOS the error in the apparent distances is around 13\,\% while for the velocity and the LOS being parallel the error increases to 15\,\%. \begin{figure}[tpb] \centering \includegraphics[width=0.48\textwidth]{ead.pdf} \caption{ The relative error in the apparent distance for a source moving at different velocities when the observer assumes the source is at rest. The lines represent the same cases as in \fref{fig:ecm}. } \label{fig:ead} \end{figure} We see that ignoring the velocity of the source impacts the measurement of the chirp mass and the apparent distance causing errors of up to 7\,\% in the first and of up to 15\,\% in the latter. Therefore, including a motion of the source is crucial to obtain accurate results and not over- or underestimating its parameters, where the apparent distance is particularly affected due to its $(1+z_{\rm rel})^2$-dependence. Moreover, the apparent distance plays a crucial role in determining the Hubble constant $H_0$ using standard sirens~\cite{schutz_1986}, wherefore we analyze the impact of the velocity in measurements of $H_0$ in the following section. \section{The effect of the source's velocity on the measurement of the Hubble constant}\label{sec:hub} GW sources with EM counterparts have been proposed as standard sirens to measure the Hubble constant $H_0$~\cite{schutz_1986,holz_hughes_2005}. Up to date the two most prominent GW events used for a measurement of $H_0$ are GW170817, the merger of a BNS with a subsequent $\gamma$-ray burst~\cite{ligo_2017d,chen_fishbach_2018}, and GW190521, a merger of a BBH with the event ``ZTF19abanrhr'' as a candidate EM counterpart~\cite{mukherjee_ghosh_2020,chen_haster_2022}. Usually, the sources are assumed to be at rest relative to the expanding universe although there are many astrophysical scenarios where the source needs to be considered as moving. In Refs.~\cite{wang_wang_2018,howlett_davis_2020,nicolaou_lahav_2020,mukherjee_lavaux_2021} the effect of a source velocity, in particular the peculiar velocity of the host galaxy, has been included to correct for this error, however, the effect of the velocity on the amplitude respective the square factor for $(1+z_{\rm rel})$ in the apparent distance has been ignored. In this section, we analyze the impact of the source's velocity on the measurement of the Hubble constant for two cases: (i) the effect of the velocity is considered in the redshift of the source but not for its apparent distance, and (ii) the source is moving but the motion is completely ignored. For a source at a small cosmological distance ($z_{\rm cos} \ll 1$), we have the following relation between its cosmological redshift $z_{\rm cos}$, its luminosity distance $D_L$, and the Hubble constant $H_0$~\cite{maggiore_2008} \begin{equation}\label{eq:rlh} z_{\rm cos} = H_0D_L. \end{equation} Using that $D_L := (1+z_{\rm cos})r_d$, where $r_d$ is the comoving distance between source and observer in the frame of the observer, we further get $z_{\rm cos} = H_0r_d/(1-H_0r_d)$. From \eq{eq:rlh} it follows that $H_0$ can be inferred using the apparent distance of the GW source $D_{\rm app}$ and the redshift $z_{\rm EM}$ of its EM counterpart as \begin{equation}\label{eq:hub} H_{0,i} = \frac{z_{\rm EM}}{D_{\rm app}}. \end{equation} However, if the source is moving the effect of the velocity on the redshift and the apparent distance needs to be corrected, otherwise we would infer a wrong value for $H_0$. Note that in general the velocity of the GW source and its EM counterpart might differ but for simplicity we assume that the two are equal. An observer including the effect of the velocity on the redshift but ignoring its effect on the apparent distance would use $z_{\rm EM} = z_{\rm cos}$ to infer the Hubble constant, thus finding \begin{equation}\label{eq:hiv} H_{0,v} := \frac{z_{\rm tot}}{D_{\rm app}} = \frac{H_0}{(1+z_{\rm rel})^2}, \end{equation} where $(1+z_{\rm rel})$ is in \eq{eq:drrs}. Therefore, we recover the correct value for $H_0$ if the source is at rest but if the source is moving we obtain an erroneous result. The error done by not considering the effect of the velocity on the apparent distance can be estimated as \begin{equation}\label{eq:ehiv} \delta H_{0,v} := \left|\frac{H_{0,v} - H_0}{H_0}\right| = \left|\frac{2v}{1 \pm v}\right|, \end{equation} where we again consider the special cases of the velocity being anti-parallel to the LOS (plus sign) and parallel to the LOS (minus sign). We see in \fref{fig:ehcv}, that the error increases with an increasing magnitude of the velocity and already can reach around 1\,\% for a peculiar velocity of the host galaxy of $1500\,{\rm km\,s^{-1}}$. For a binary orbiting a SMBH of $10^7\,{\rm M_\odot}$ at a distance of $5\,{\rm mpc}$ with a velocity of $3000\,{\rm km\,s^{-1}}$ the error goes up to around 2\,\% and can exceed 3\,\% for extreme kicks of $5000\,{\rm km\,s^{-1}}$. For higher velocities the orientation of the velocity relative to the LOS becomes significant, varying from 13\,\% for the two being anti-parallel to 15\,\% for the two being parallel for a binary orbiting a SMBH at a distance of $100\,R_s$. \begin{figure}[tpb] \centering \includegraphics[width=0.48\textwidth]{ehcv.pdf} \caption{ The relative error in the Hubble constant for a source moving at different velocities when the observer only considers the effect of the velocity on the redshift of the EM counterpart but not the apparent distance of the source respective the amplitude of the GW. The lines correspond to the same cases as in \fref{fig:ecm}. } \label{fig:ehcv} \end{figure} An observer completely ignoring the effect of velocity would infer the Hubble constant using the total redshift for the EM counterpart ($z_{\rm EM} = z_{\rm tot}$) and the apparent luminosity distance, thus obtaining \begin{equation}\label{eq:hir} H_{0,r} := \frac{z_{\rm tot}}{D_{\rm app}} = \frac{(1+z_{\rm cos})z_{\rm rel} + H_0D_L}{(1+z_{\rm rel})^2D_L}. \end{equation} We recover again the correct value for $H_0$ if the source is at rest but, if the source is moving, we obtain an erroneous result and, in particular, the value inferred for $H_0$ depends on the distance of the source. The error done by ignoring the velocity of the source can be estimated by \begin{align}\label{eq:ehir} \nonumber \delta H_{0,r} :=& \left|\frac{H_{0,r} - H_0}{H_0}\right| \\ =& \left|\frac{(1+z_{\rm rel})^2z_{\rm cos} - (1+z_{\rm rel})(1+z_{\rm cos}) + 1}{(1+z_{\rm rel})^2z_{\rm cos}}\right|, \end{align} where we used \eq{eq:rlh} to eliminate the direct dependence on $H_0$. We, furthermore, use that $(1+z_{\rm rel}) = \sqrt{(1\pm v)/(1\mp v)}$ for the velocity and the LOS being either anti-parallel (upper sign) or parallel (lower sign). The error in the measurement of the Hubble constant further increases when the effect of the velocity is not only ignored in the apparent distance but also in the redshift of the EM counterpart. \fref{fig:ehcrv} shows the relative error for a source at a cosmological redshift of $0.09$ (corresponding to the distance of GW150914~\cite{GWTC1}) for different velocities. We see that even for relatively low velocities of only $400\,{\rm km\,s^{-1}}$, corresponding to the average kick velocity, the error can reach around 1.5\,\%. For an average peculiar velocity of the host galaxy of $1500\,{\rm km\,s^{-1}}$ the error already exceeds 5\,\%, and the error becomes bigger than 10\,\% for a source moving with more than $3000\,{\rm km\,s^{-1}}$, corresponding to the orbital velocity of a binary orbiting a SMBH of $10^7\,{\rm M_\odot}$ at $5\,{\rm mpc}$. The dependence on the orientation of the velocity relative to the LOS becomes again more prominent for high velocities thus causing the error to vary between around 60\,\% and 80\,\% for a binary at an orbit of $100\,R_s$ around a SMBH. In \fref{fig:ehcrc}, we fix the velocity of the source to be the typical peculiar velocity of galaxies and vary the cosmological redshift of the source. We see that for this relatively low velocity the error is almost the same for the velocity being either anti-parallel or parallel to the LOS while it decreases for an increasing cosmological redshift. For a source at the distance of GW170817 ($z_{\rm cos}=0.01$)~\cite{GWTC1}, the error in $H_0$ is more than 50\,\%. The error for a source at the distance of GW150914 decreases to still considerable 5\,\% and goes below 1\,\% for sources at cosmological redshifts of more than 0.3. We see that for a source at the distance of GW190521 ($z_{\rm cos}=0.72$)~\cite{GWTC2} the error goes down to around 0.2\,\%. However, it should be noted that \eq{eq:rlh} assumes the cosmological redshift to be much smaller than one and hence the estimation for redshifts close to one are less reliable. In general, we expect the error for high cosmological redshifts to be at least bounded from below by the errors found when only correcting for the effect of the velocity on the redshift of the EM counterpart, which is independent of the cosmological redshift. \begin{figure}[tpb] \centering \includegraphics[width=0.48\textwidth]{ehcrv.pdf} \caption{ The relative error in the Hubble constant for a source at the cosmological distance of GW150914 ($z_{\rm cos}=0.09$) that is moving at different velocities but the observer assumes it to be at rest relative to the expanding universe. The lines correspond to the same cases as in \fref{fig:ecm}. } \label{fig:ehcrv} \end{figure} \begin{figure}[tpb] \centering \includegraphics[width=0.48\textwidth]{ehcrc.pdf} \caption{ The relative error in the Hubble constant for a source moving with a typical peculiar velocity of galaxies of $1500\,{\rm km\,s^{-1}}$ for different cosmological redshifts of the source. The red dashed-dotted line corresponds to the velocity and the LOS being anti-parallel while the blue dotted line corresponds to the two being parallel. The vertical lines correspond to the cosmological redshift for the GW events GW170817, GW150914, and GW190521. } \label{fig:ehcrc} \end{figure} Last, we consider the case of GW170817 using a velocity of roughly $300\,{\rm km\,s^{-1}}$ as used in the analysis conducted by LIGO~\cite{ligo_2017d}. From \eq{eq:ehir} we get that if the motion of the source would be ignored completely the error would be around 100\,\% thus making the result useless. When correcting by the effect of the motion on the redshift of the EM counterpart but ignoring its effect on the apparent distance as done in Ref.~\cite{ligo_2017d}, the error can be estimated using \eq{eq:ehiv} to be 0.2\,\% and hence we confirm that their result is reliable while the error in $H_0$ is dominated by other effects. However, for sources with higher velocities the error when ignoring their effect on the apparent distance increases quickly and hence it needs to be considered. \section{Conclusions}\label{sec:con} We studied the effect of cosmological expansion and the velocity of the source on the frequency of GWs and the apparent distance obtained from them. We recover the classical result that the frequency of a source at rest relative to the expanding universe is redshifted by the cosmological expansion while its apparent distance increases by the same redshift factor. We also show that the total redshift of a moving source in an expanding universe is equal to the product of the cosmological redshift times the relativistic redshift, as usually assumed. However, we find that the apparent distance of moving source changes by a factor of the relativistic redshift square, in contrast to the linear behavior for the cosmological redshift. The additional factor can be attributed to the fact that the amplitude of GWs from a moving source changes by the same factor relative to a source at rest. Using these results we analyze the estimation error in the two main observables of GWs, the chirp mass of the source and its apparent distance when the velocity is ignored. We consider the three astrophysical scenarios of gravitational kicks, the orbital motion of BBHs, BNSs, or BHNSBs around a SMBH, and the peculiar velocity of the host galaxy which lead to velocities varying from a few $100\,{\rm km\,s^{-1}}$ up to several percent of the speed of light. The error in the chirp mass and the apparent distance varies between 0.1\,\% respective 0.25\,\% for average kick velocities and go up to 7\,\% and 15\,\%, respectively, for sources close to a SMBH. Furthermore, the error exceeds 1\,\% in the apparent distance for a source moving due to the peculiar velocity of its host galaxy. Furthermore, we study how the motion of the source impacts the estimation of the Hubble constant using standard sirens when only considering the effect of the velocity on the redshift of the EM counterpart but not on the apparent distance and when completely ignoring the effect of the source's velocity. We find that in the first case, the error in $H_0$ can already reach 1\,\% for a typical peculiar velocity of the host galaxy and go up to 15\,\% for a binary close to a SMBH. When completely ignoring the effect of velocity, the error increases to 5\,\% for a typical peculiar velocity of the host galaxy for a source at the cosmological redshift of GW150912. We, furthermore, find that the error depends on the cosmological redshift of the source where for a typical peculiar velocity of the host galaxy it can go up to over 50\,\% for a source like GW170817 at a cosmological distance of 0.01 while it decreases to under 1\,\% for sources at cosmological redshifts of more than 0.3. \section*{Acknowledgments} ATO acknowledges support from the Guangdong Major Project of Basic and Applied Basic Research (Grant No. 2019B030302001). XC is supported by the National Science Foundation of China (Grants Nos. 11873022 and 11991053).

\section{Introduction} Precise parton distribution functions (PDFs) are an essential ingredient in analysing data obtained at high energy hadron colliders, such as the Tevatron and the LHC. In perturbative QCD the PDFs are now known at next-to-next-to leading order (NNLO). Indeed, with the current precision, it is important to investigate the effect of electroweak corrections in hadron collider physics. In particular, the QED contributions have large logarithmic terms, up to $\alpha$log$(Q^2/m_q^2)$, arising from photons emitted from the incoming quark lines. At high $Q^2$ scales these corrections should be resummed. Fortunately the QCD factorisation theorem also applies to QED, and so the photon-induced logarithms can be absorbed into the PDFs, just as the $\alpha_s$log$Q^2$ terms are summed by DGLAP evolution. As a consequence the normal DGLAP equations are slightly modified and a photon parton distribution function of the proton, $\gamma^p(x,Q^2)$, is generated. Thus, for example, (at leading order in both $\alpha_s$ and $\alpha$) we have an extra equation for the evolution of the photon PDF \be \frac{\partial\gamma(x,Q^2)}{\partial{\rm log}Q^2}=\frac{\alpha}{2\pi}\int^1_x\frac{dy}{y}\left(P_{\gamma\gamma}\otimes \gamma+ \sum_1 e_i^2 P_{\gamma q} \otimes q_i \right), \label{eq:evol} \ee where \be P_{\gamma\gamma}(y)=-\frac{2}{3}\sum_i e_i^2\delta (1-y),~~~~~~~~ P_{\gamma q}=C_F^{-1} P_{gq}. \ee Similarly, the photon PDF, $\gamma$, contributes to the evolution $\partial q_i /\partial$log$Q^2$ via the $P_{q\gamma}$ splitting. \section{Existing determinations of $\gamma^p$ compared to this work} Indeed, with the present level of precision, it has become topical to treat the photon as one of the point-like partons inside the nucleon and to account for this QED effect explicitly in the global parton analysis. This approach was first followed ten years ago by MRST(2004) \cite{MRST}. Recently it has been used by the NNPDF \cite{NNPDF} and CTEQ \cite{CTEQ} groups. The central issue is the choice of input distributions for the photon PDFs of the proton and neutron. In the original MRST study it was assumed that the starting distributions are given by one-photon emission off valence (constituent) quarks in the leading logarithm approximation. For example, for the photon PDF of the proton the starting distribution was taken to be\footnote{Here we write the convolution of quark PDFs and the $P_{gq}$ splitting function explicitly; whereas in \cite{MRST} it was simply denoted by $\otimes$.} \be \gamma^p(x,Q^2_0)~=~\frac{\alpha}{2\pi}\int\frac{dz}{z}\left[ \frac{4}{9}{\rm log} \left(\frac{Q_0^2}{m^2_u}\right)u_0\left(\frac xz\right)+ \frac{1}{9}{\rm log} \left(\frac{Q_0^2}{m^2_d}\right)d_0\left(\frac xx \right) \right] \frac{1+(1-z)^2}{z}, \label{eq:2} \ee where $u_0$ and $d_0$ are the valence-like distributions of the proton, and where the current quark masses were used. The most direct measurement of the photon PDF at the time of the MRST analysis appeared to be wide-angle scattering of the photon by a electron beam via the process $ep \to e\gamma X$, where the final state electron and photon are produced with equal and opposite large transverse momentum. The subprocess is QED Compton scattering\footnote{There are other contributions which should be included. These will be discussed in Section \ref{sec:5}.}, $e\gamma \to e\gamma$, for which the contribution to the cross section is \be \sigma(ep\to e\gamma X)~=~\int dx^\gamma~\gamma^p(x^\gamma,\mu^2)~\hat{\sigma}(e\gamma \to e\gamma), \label{eq:comp} \ee where $\mu$ is the factorization scale. MRST \cite{MRST} predicted a cross section in agreement with the only measurement of this process available at that time \cite{ZEUS}. The NNPDF \cite{NNPDF} and CTEQ \cite{CTEQ} groups use a different approach to MRST. They parametrise the input photon PDFs, $\gamma(x,Q_0^2)$, and attempt to determine the parameters from the global data, along with the quark and gluon PDFs. Unfortunately the present data are not of sufficient accuracy to provide a reasonable determination of the photon input. The NNPDF collaboration \cite{NNPDF} used freely parametrised (without bias) starting distributions, including the photon PDFs, and constrain the photon PDFs mainly from the Drell-Yan (low-mass, on-shell $W$ and $Z$ production and high-mass) LHC data. There is expected to be the most sensitivity to the low-mass Drell-Yan data \cite{lowmass}. However, the uncertainties observed in the resulting photon PDFs are huge, especially at low $x$. The preliminary CTEQ analysis \cite{CTEQ} proceeds differently. CTEQ keep a similar theoretical form of the distributions $\gamma(x,Q_0^2)$ to that proposed by MRST, but with an arbitrary normalisation parameter, which is expressed as the momentum fraction, $p_0(\gamma)$, carried by the input photon. They find that the constraint coming from the energy-momentum sum rule is weak (allowing $p_0(\gamma)$ to range up to 5$\%$), while to fit the updated ZEUS data for $ep \to e\gamma X$ \cite{ZEUS1} requires $p_0(\gamma)\sim 0.1-0.2\%$, using the valence quark induced input (\ref{eq:2}) and allowing for the extra normalisation parameter. \begin{figure} \begin{center} \vspace{-6.0cm} \hspace{-2.0cm} \includegraphics[height=13cm]{photonPDF} \vspace{-3.5cm} \caption{\sf The coherent and incoherent contributions to the photon PDF, $\gamma^p(x,Q^2)$, corresponding, respectively, to photon emission directly from the proton and from a quark.} \label{fig:1} \end{center} \end{figure} Unlike the above analyses, here we emphasize that the major part of the input, $\gamma^p(x,Q^2_0)$, especially at low $x$, comes from the {\it coherent} emission of the photon from the `elastic' proton, which can be calculated theoretically with good accuracy. The process is sketched in Fig. \ref{fig:1}. (The previous analyses are based only on {\it incoherent} emission from individual quarks within the proton.) The uncertainty on our determination $\gamma^p(x,Q^2_0)= \gamma^p_{\rm coh}+\gamma^p_{\rm incoh}$ comes essentially only from the relatively small contribution $\gamma^p_{\rm incoh}$ which, from a hadron viewpoint, actually corresponds to the QED excitations of the proton into higher mass states. However, here, in Section \ref{sec:3} we adopt the quark viewpoint, and calculate this contribution as the incoherent emission of photons from quarks within the proton. This contribution turns out to be relatively small. Therefore, since $\gamma^p_{\rm coh}$ may be calculated with good accuracy, it means that the uncertainty in the theoretically determined photon input distributions is expected to be small; essentially coming from the uncertainty in the `extrapolation' of valence quarks needed to estimate the contribution to $\gamma^p_{\rm incoh}$ coming from the region below the starting scale $Q_0$. We will quantify this uncertainty below. We summarize the discussion of this Section in Table \ref{tab:1}. \begin{table} [h] \begin{center} \begin{tabular}{|l|l|l|}\hline Group & input photon PDF & data \\ \hline MRST \cite{MRST} & model for $\gamma^p_{\rm incoh}$ & predict $ep \to e\gamma X$ \\ NNPDF \cite{NNPDF} & freely parametrised & fit to LHC Drell-Yan \\ CTEQ \cite{CTEQ} prelim. & parametrise with $p_0(\gamma)$ & fit to $ep \to e\gamma X$ \\ this work & calculate $\gamma^p_{\rm coh}$ (dominates) & predict $ep \to e\gamma X$ \\ & ~~~~~~~~~ +~model for $\gamma^p_{\rm incoh}$ & \\ \hline \end{tabular} \end{center} \vspace*{-0.5cm} \caption{\sf An outline of the procedure used by the various groups to determine the photon PDF.} \label{tab:1} \end{table} \section{Improved input distributions for the photon PDFs \label{sec:3}} Here we will follow the MRST approach, but will use much improved starting distributions for the photon PDFs of the proton and neutron. Indeed, we have \be \gamma^N(x,Q^2_0)=\gamma^N_{\rm coh}+\gamma^N_{\rm incoh} \ee where $N=p,n$. As discussed above, and sketched in Fig. \ref{fig:1}, the contribution $\gamma^p_{\rm coh}$ is caused by coherent photon emission from the proton that remains intact, whereas $\gamma_{\rm incoh}$ is due to non-coherent emission from individual quarks. The {\it coherent} emission from the proton is given by \cite{Bud} \be \gamma^p_{\rm coh}(x,Q^2_0)=\frac{\alpha^{\rm QED}}{2\pi}~\frac{[1+(1-x)^2]}{x}\int_0^{|t|<Q^2_0}dq^2_t\frac{q^2_t}{(q^2_t+x^2m^2_p)^2}F^2_1(t)\ , \label{eq:1} \ee where $q_t$ is the transverse momentum of the emitted photon and \be t=-\frac{q^2_t+x^2m^2_p}{1-x}. \ee $F_1$ is the electromagnetic proton form factor corresponding to $\gamma_\mu$ at the vertex. For the neutron we have \be \gamma^n_{\rm coh}(x,Q_0^2)=0. \ee For the {\it non-coherent} emission from individual quark lines we use an improved form of (\ref{eq:2}) \be \gamma^p_{\rm incoh}(x,Q^2_0)=\frac{\alpha}{2\pi}\int_x^1 \frac{dz}{z}\left[\frac 49 u_0(\frac xz)+ \frac 19d_0(\frac xz)\right]~\frac{1+(1-z)^2}z\int^{Q^2_0}_{|t_{\rm min}|}~\frac{dt}{t-m_q^2} ~\left( 1-F^2_1(t)\right)\ , \label{eq:av} \ee where \be t_{\rm min}=-\frac{x}{(1-x)}\left(m^2_\Delta-(1-x)m^2_N\right) \label{eq:tmin} \ee accounts for the fact that the lowest possible proton excitation is the $\Delta$-isobar. The final factor $(1-F^2_1)$ in (\ref{eq:av}) is the probability to have no intact proton in the final state. We have to exclude an intact proton as its contribution is calculated separately in (\ref{eq:1}). In (\ref{eq:av}), $m_q=m_d$ when convoluted with $d_0$, and $m_q=m_u$ when convoluted with $u_0$~\footnote{To be precise, we replace the integral $\int dt/(t-m_q^2)$ by $\int \left[\frac{q^2_t/(1-z)}{t-m_q^2}\right]^2\frac{dq^2_t}{q^2_t}$, where $t=t_{\rm min}-q^2_t/(1-z)$ with $t_{\rm min}$ given by (\ref{eq:tmin}).}. In this contribution we use the {\it current} quark masses. Here the quark distribution $u_0=u_{\rm valence}+2u_{\rm sea}$ is frozen for $Q<Q_0$ at its value at $Q_0$. The same is true for the other quarks - $d,s$. A similar expression holds for $\gamma^n_{\rm incoh}$, with $4/9 \leftrightarrow 1/9$ and $F_1^p\to F_1^n$. In this way we get an {\bf upper} limit for the non-coherent contribution to the photon input distributions. The other extreme is to take for $u_0$ and $d_0$ just the non-relativistic quark model expectation with \be u_0(x)=u_{\rm non-rel}=2\delta(x-1/3)~~~~ {\rm and}~~~~ d_0(x)=d_{\rm non-rel}=\delta(x-1/3) \ee for the proton, and to use {\em constituent} quark masses $m_q=300- 350$ MeV. The optimum estimate of the non-coherent contribution to the photon PDF input is probably to take a, physics-motivated, linear interpolation between the two limits. That is, to use in (\ref{eq:av}) \be q_0(x,|t|)~=~\frac{|t|}{Q^2_0}~q(x,Q^2_0)+\frac{Q^2_0-|t|}{Q^2_0}~q_{\rm non-rel}(x) \label{eq:interpol} \ee with $m_q=m_{\rm current}+m_{\rm eff}(t)$, where the `effective' constituent quark mass is parametrized by a simplified formula of the form \be m_{\rm eff}\simeq m(0)\exp(-b\sqrt{|t|})\ , \ee with $m(0)=345$ MeV and slope $b=1.4$ GeV$^{-1}$ (see, for example, Fig. 4 in ~\cite{DP}, where the light quark in the instanton vacuum was studied). In general, one may also account for the $\Delta$-isobar excitation. In the latter case, we have to add to (\ref{eq:1}) $\gamma^\Delta_{\rm coh}$, which is also of the form of (\ref{eq:1}), but with\footnote{ Here the $F^\Delta(t)$ form factor includes the normalization for $\gamma+p\to \Delta$ cross section, and at small $q_t$ this $p \to \Delta$ transition `form factor' $F^\Delta(t)\propto q_t$ vanishes.} \be F_1(t)/(q^2_t+x^2m^2_p)~~~~{\rm replaced~ by}~~~~ F^\Delta(t)/(q^2_t+x(m^2_\Delta-(1-x)m^2_p). \ee For the $\Delta$ contribution \be \label{eq:4} |t|=\frac{q^2_t+x(m^2_\Delta-(1-x)m^2_N)}{1-x}. \ee Also when including the $\Delta$ contribution we have to replace in (\ref{eq:av}) \be [1-F^2_1(t)] ~~~~~{\rm by}~~~~~ [1-F^2_1(t)]-F^2_\Delta(t)\Theta(|t|(1-x)-x(m^2_\Delta-(1-x)m^2_N), \ee where here $|t|$ is given by (\ref{eq:4}). In addition, it is possible to include a coherent contribution caused by the anomalous magnetic moment of the proton, described by the proton form factor $F_2$. These non-logarithmic corrections will reduce the remaining incoherent contribution and therefore decrease the final uncertainty in the input $\gamma^p(x,Q_0^2)$. However, since they do not change the result noticeably, we do not consider these possibilities here. \section{Results for the photon PDF} \begin{figure} \begin{center} \vspace{-8.0cm} \includegraphics[height=16cm]{phot1} \vspace{-0.5cm} \caption{\sf The photon PDF of the proton at input ($Q_0^2=1$ GeV$^2$) and after evolution up to $Q^2=200$ GeV$^2$. At each $Q^2$ value, the lowest curve is $\gamma^p_{\rm coh}$ and the continuous `interpolating' curve (obtained from (\ref{eq:interpol})) is the effect of adding the $\gamma^p_{\rm incoh}$ contribution. At input, the proton momentum fraction carried by the photon is 0.16$\%$.} \label{fig:2} \end{center} \end{figure} \begin{figure} \begin{center} \vspace{-6.7cm} \includegraphics[height=16cm]{photn1} \vspace{-0.5cm} \caption{\sf The photon PDF of the neutron at input ($Q_0^2=1$ GeV$^2$) and after evolution up to $Q^2=200$ GeV$^2$. At each $Q^2$ value, the three curves correspond to the upper and lower estimates of $\gamma^n_{\rm incoh}$, together with continuous (`interpolating') curve obtained from (\ref{eq:interpol}).} \label{fig:2n} \end{center} \end{figure} \begin{figure} \begin{center} \vspace{-7.0cm} \includegraphics[height=16cm]{part} \vspace{-0.5cm} \caption{\sf The photon PDFs of the proton and neutron compared with the MSTW \cite{MSTW} NLO quark and gluon distributions at $Q^2=20$ GeV$^2$.} \label{fig:3n} \end{center} \end{figure} In Figs. \ref{fig:2} and \ref{fig:2n} we show, respectively, the input distributions of the photon PDF for the proton and neutron at $Q_0^2=1$ GeV$^2$, as calculated in the previous Section, together with the photon PDF determined from the evolution up to $Q^2=200$ GeV$^2$ using (\ref{eq:evol}), but with NLO partons. We see that the major part of the input photon PDF of the proton is generated by coherent emission of the photon from an `elastic' proton, while the non-coherent contribution only enlarges this value by about 20\%. Actually, the figure shows three curves for the inclusion of $\gamma^p_{\rm incoh}$, corresponding to the two limits of the input distribution together with their `interpolated average', shown by the continuous curve. We also note that for a low $x\sim 10^{-2}\ -\ 10^{-3}$ the evolution to $Q^2=200$ GeV$^2$ increases the photon density by about a factor of two; whereas for $x=10^{-4}$ the increase is about a factor of three. Considering the two terms involving $P_{\gamma\gamma}$ and $P_{\gamma q}$ of (\ref{eq:evol}), which contribute to the evolution of the photon PDF, we note that the first term, $P_{\gamma\gamma}$, decreases $\gamma^p$ only slightly (less than 1$\%$ in the evolution up to $Q^2=200$ GeV$^2$), whereas the growth comes from the $P_{\gamma q}$ term -- the photons emitted by quarks. We have the same growth of each curve due to the linear nature of DGLAP evolution. Here we evolve using MSTW\cite{MSTW} NLO quarks\footnote{We find that the use of the updated CPdeut parton set of MMSTWW \cite{MMSTWW} makes a negligible difference.}. At first sight we might expect the contribution generated by gluons, via the gluon-photon splitting, to be important, due to the large gluon PDF, especially at low $x$. However the value of the NLO splitting function $P^{(1)}_{\gamma g}(z)$ is rather small\footnote{We extract the splitting function as the term proportional to $C_F N_f$ from the known $P^{(1)}_{gg} (z)$ splitting \cite{FP}.}. Moreover, $P^{(1)}_{\gamma g}(z)$ is negative at large $z$. As a result, the inclusion of gluons enlarges $\gamma^p(x,Q^2)$ by less than 2$\%$, and we neglect this effect. Recall that for the photon PDF of the neutron we have $\gamma^n_{\rm coh}(x,Q^2_0)=0$, and so the input is given entirely by $\gamma^n_{\rm incoh}(x,Q^2_0)$, see Fig. \ref{fig:2n}. However, the {\it increase} in $\gamma^n$ in the evolution up to $Q^2=200$ GeV$^2$ (which is driven by the final term in (\ref{eq:evol})) is comparable to that for $\gamma^p$. In Fig. \ref{fig:3n} we compare $\gamma^p$ and $\gamma^n$ with the other PDFs at $Q^2=20$ GeV$^2$. \begin{figure} \begin{center} \vspace{-6.0cm} \vspace{-5.0cm} \includegraphics[height=24cm]{photPDF} \vspace{-1.5cm} \caption{\sf The photon PDF at input $Q_0^2=1$ GeV$^2$ and after evolution up to $Q^2=20,~200, $ and $10^4$ GeV$^2$. The notation of the curves are as in Fig. \ref{fig:2}. We have included the prediction of MRST(2004) \cite{MRST} for comparison.} \label{fig:3} \end{center} \end{figure} In Fig. \ref{fig:3} we show how $\gamma^p$ evolves as $Q^2$ increases from input $Q_0^2=1$ to $Q^2=20,~200$ and $10^4$ GeV$^2$, together with the behaviour predicted\footnote{Actually we use the MRST formulation, but with NLO MSTW partons \cite{MSTW}, which make little change to the behaviour of $\gamma^p$.} by MRST(2004) \cite{MRST} input. The comparison of the predictions by MRST \cite{MRST} and of this work, may, at first sight, appear surprising. MRST is purely based on the incoherent contribution, $\gamma^p_{\rm incoh}$, whereas here the prediction comes dominantly from the coherent contribution (dashed-dotted curve) with a small addition from $\gamma^p_{\rm incoh}$. The explanation is as follows. The incoherent contribution determined by MRST should be suppressed by $t_{\rm min}$, (\ref{eq:tmin}), and by $[1-F_1^2(t)]$ of (\ref{eq:av}). In the present work, the coherent emission from the proton is added. The above two effects (that is the suppression of the incoherent contribution and the inclusion of the coherent emission) partly compensate each other. However, indeed at large $x$, where $|t_{\rm min}|$ is large, MRST goes above the present input, while at low $x$, where $|t_{\rm min}| \simeq (xm_N)^2$ is even less than the current quark mass, our input exceeds the MRST curve. \section{Comparison with $ep \to e \gamma X$ data \label{sec:5}} \begin{figure} \begin{center} \vspace{-3.0cm} \vspace{-.0cm} \includegraphics[height=8cm]{LLQQ} \vspace{-1.5cm} \caption{\sf Representative diagrams mediating inclusive electroproduction of isolated photons, $ep\to e\gamma X$: (a) photon emitted by electron, (b) photon emitted by a quark.} \label{fig:LLQQ} \end{center} \end{figure} To probe the photon PDF experimentally we should consider a hard subprocess where the photon distribution dominates. A good example is the inclusive electroproduction of isolated photons observed in the direction of the incoming electron. This reaction is mediated by the $e\gamma \to e\gamma$ hard subprocess and its cross section is described by (\ref{eq:comp}), and sketched in Fig. \ref{fig:LLQQ}(a). It is known that this Compton scattering process is sharply peaked in the backward direction. Therefore the outgoing photon should be observed at high negative rapidity, $\eta^\gamma$, at angles close to the electron beam. Indeed it was already shown \cite{G1,G2} that this component (called LL) already dominates for\footnote{Here we adopt the convention for $\eta^\gamma$ used by the ZEUS collaboration \cite{ZEUS1}.} $\eta^\gamma <-1$. Unfortunately, the present experimental data are quite limited and the lowest rapidity bin is $-0.7<\eta^\gamma<-0.3$. We compare our theoretical prediction for the LL component with the data in Tables \ref{tab:2} and \ref{tab:3}. As can be seen from the comparison in the first bin, the predicted cross section is close to the measured value. The result depends on the choice of factorisation scale. We present results for $\mu=E_T^\gamma$ and $E_T^\gamma /2$ to indicate the scale dependence. \begin{table} [h] \begin{center} \begin{tabular}{|c|r|c|c|}\hline $\eta^\gamma$ range & \multicolumn{3}{c|} {$d\sigma(ep\to e\gamma X) /d\eta^{\gamma}$ ~~(pb)} \\ \hline & experiment & $\mu=E_T^\gamma$ & $\mu=E_T^{\gamma}/2$ \\ \hline \bf $-0.7--0.3$ & $17.4~ \pm 0.9 ^{~+0.5}_{~-0.7}$ &16.4 &13.3\\ $-0.3-0.1$ & $13.0 ~\pm 0.8 ^{~+0.6}_{~-0.3}$ & 7.7&6.3\\ $0.1-0.5$ & $10.7 ~\pm 0.9 ^{~+0.7}_{~-0.4}$ & 2.7 & 2.24\\ $0.5-0.9$ & $8.7 ~\pm 0.9 ^{~+1.1}_{~-0.7}$ & 0.8 &0.65 \\ \hline \end{tabular} \end{center} \vspace*{-0.5cm} \caption{\sf The second column gives the values of the $ep\to e\gamma X$ cross section measured by the ZEUS collaboration \cite{ZEUS1} in different rapidity, $\eta^\gamma$, intervals. The final two columns show the contribution to the cross section arising from the LL process of Fig. \ref{fig:LLQQ}(a) for two different choices of the factorisation scale $\mu$.} \label{tab:2} \end{table} \begin{table} [h] \begin{center} \begin{tabular}{|l|l|c|c|}\hline $E_T^\gamma$ range & \multicolumn{3}{c|} {$d\sigma(ep\to e\gamma X) /dE_T^{\gamma}$ ~~(pb/GeV)} \\ \hline (GeV) & experiment & $\mu=E_T^\gamma$ & $\mu=E_T^{\gamma}/2$ \\ \hline $4-6$ & $4.87~ \pm 0.28 ^{~+0.40}_{~-0.23}$ &2.4 &1.95\\ $6-8$ & $2.40 ~\pm 0.16 ^{~+0.09}_{~-0.11}$ & 1.46&1.22\\ $8-10$ & $1.24 ~\pm 0.11 ^{~+0.03}_{~-0.04}$ & 0.88 & 0.74\\ $10-15$ & $0.55 ~\pm 0.04 ^{~+0.03}_{~-0.03}$ & 0.12 & 0.10 \\ \hline \end{tabular} \end{center} \vspace*{-0.5cm} \caption{\sf The second column gives the values of the $ep\to e\gamma X$ cross section measured by the ZEUS collaboration \cite{ZEUS1} in different $E_T^\gamma$ intervals. The final two columns show the contribution to the cross section arising from the LL process of Fig. \ref{fig:LLQQ}(a) for two different choices of the factorisation scale $\mu$.} \label{tab:3} \end{table} At large $\eta^\gamma$, the contribution of the Compton-induced process decreases rapidly. In this domain, inclusive isolated photons are mainly produced by quarks, see Fig. \ref{fig:LLQQ}(b). Note that in our theoretical calculation of $ep\to e\gamma X$ we have accounted for the angular, the $E_T$ and the other experimental cuts imposed by the ZEUS collaboration \cite{ZEUS1}, but we have no possibility to include the photon isolation criteria. Therefore the observed cross section corresponding to the LL process should be lower than our prediction. In Table \ref{tab:3} we compare our prediction of the LL contribution with $E_T$ dependence of the measured cross section. However, now the data were collected over a large rapidity interval: $-0.7<\eta^\gamma <0.9$. Here the quark contribution is important, and the LL subprocess describes only about half of the cross section. \section{Conclusions} We have demonstrated that the major part of the photon input PDF of the proton (caused by the coherent emission of the photon that does not destroy the proton) can be calculated with good accuracy. This strongly reduces the possible uncertainties in the QED part of an extended global parton analysis which includes the photon as a point-like parton. In this way, we evaluate the expected photon PDFs by DGLAP evolution with LO QED splittings and NLO MSTW quarks. Note that the further step of including the photon-to-quark splitting will introduce a small violation of isospin symmetry in the `singlet' PDF, in particular $u^d \ne d^n$. The resulting photon distributions agree with data for the electroproduction of isolated photons, $ep\to e\gamma X$, at negative rapidities where the cross section is dominated by the $e\gamma \to e\gamma$ hard subprocess. \section*{Acknowledgements} MGR thanks the IPPP at the University of Durham for hospitality. This work was supported by the Federal Program of the Russian State RSGSS-4801.2012.2. We thank Robert Thorne for valuable discussions. \thebibliography{} \bibitem{MRST} A.D. Martin, R.G. Roberts, R.S. Thorne and W.J. Stirling, Eur. Phys. J. {\bf C39} (2005) 155. \bibitem{NNPDF} NNPDF collaboration, R.D. Ball et al., Nucl. Phys. {\bf B877} (2013) issue 2, 290. \bibitem{CTEQ} CTEQ-TEA group: presented by C. Schmidt at the PDF4LHC meeting on May 14th, 2014. \bibitem{ZEUS} ZEUS collaboration: S. Chekanov et al., Phys. Lett {\bf B595} (2004) 86. \bibitem{lowmass} LHCb collaboration, (2012), LHCb-CONF-2012-013. \bibitem{ZEUS1} ZEUS collaboration, S. Chekanov et al., Phys. Lett. {\bf B687} (2010) 16. \bibitem{Bud} V.M. Budnev et al., Phys. Rept. {\bf C15} (1975) 181. \bibitem{DP} D.I. Diakonov and V.Yu. Petrov, Nucl. Phys. {\bf B272} (1986) 457. \bibitem{MSTW} A.D. Martin, W.J. Stirling, R.S. Thorne and G. Watt, Eur. Phys. J. {\bf C63} (2009) 189. \bibitem{MMSTWW} A.D. Martin, A.J.Th.M. Mathijssen, W.J. Stirling, R.S. Thorne, B.J.A. Watt and G. Watt, Eur. Phys. J. {\bf C73} (2013) 2318. \bibitem{FP} W. Furmanski and R. Petronzio, Phys. Lett. {\bf B97} (1980) 437. \bibitem{G1} A. Gehrmann-De Ridder, T. Gehrmann and E. Poulsen, Eur. Phys. J. {\bf C47} (2006) 395. \bibitem{G2} A. Gehrmann-De Ridder, T. Gehrmann and E. Poulsen, Phys. Rev. Lett. {\bf 96} (2006) 132002. \end{document}

\section{Introduction} \subsection{The problem} In this paper, we study the existence of solutions of the following equation: \begin{equation}\label{eq78} \tag{$E_a$} \begin{cases} \diw u + \langle a , u \rangle = f & \textrm{ in } \Omega,\\ u=0 & \textrm{ on } \partial \Omega. \end{cases} \end{equation} Here, \(\Omega\) is a bounded open set in \(\R^n\), \(n\geq 1\), \(a:\Omega \to \R^n\) is a vector field and \(f : \Omega \to \R\) is a function. The notation \(\langle a , u \rangle\) refers to the standard scalar product in \(\R^n\). We look for a solution \(u:\Omega \to \R^n\) in Sobolev spaces or in H\"older spaces. When \(a=0\), the above equation reduces to the classical \emph{divergence equation}, which has attracted considerable attention. Let us just mention for the moment that the expected regularity of a solution naturally depends on the regularity of the data. For instance, assuming that \(\Omega\) is Lipschitz and \(f\in L^{p}(\Omega)\), there exists a solution \(u\) in \(W^{1,p}_{0}(\Omega;\R^n)\) if and only if \begin{equation} \label{eq:int condition a is 0} \int_{\Omega} f=0. \end{equation} The condition \eqref{eq:int condition a is 0} is closely related to the homogeneous Dirichlet boundary condition. When \(f\in C^{0,\alpha}(\Omegabar)\) and \(\Omega\) is \(C^{2,\alpha}\) for some \(\alpha\in (0,1)\), then a solution $u$ exists in \(C^{1,\alpha}(\Omegabar)\) with $u=0$ on \(\partial \Omega\), under the same necessary and sufficient condition \eqref{eq:int condition a is 0}. When \(a\not=0\), the study of the perturbed equation \eqref{eq78} has been initiated in \cite{Csato-Dacorogna}. Quite surprinsingly, it was observed that the lower term \(\langle a , u \rangle\) dramatically modifies the existence theory. Indeed, the condition \eqref{eq:int condition a is 0} does not generalize to some integral condition involving $a,$ \emph{unless $a$ is a gradient}. When \(a\) is a gradient: \(a=\nabla A\) for some function \(A:\Omega\to \R\), then for every \(u\), \[ \diw u + \langle a , u \rangle = e^{-A}\diw (u e^{A}). \] It follows that for every \(f\), the equation \(\diw u + \langle a , u \rangle = f\) is equivalent to \[ \diw (u e^{A}) = e^A f, \] and the classical theory when \(a\equiv 0\) then applies. In particular, the existence of a solution requires that \[ \int_{\Omega}e^{A}f=0. \] The aim of the present paper is to obtain existence and regularity of solutions to the divergence equation with a lower order term \eqref{eq78}, under natural regularity assumptions on the data, in both Sobolev and H\"older spaces, when \(a\) is not a gradient. The existence problem can be formulated in (at least) three different ways that we now detail. Let us debote by \(T_a\) the operator \(\diw + \langle a , \cdot \rangle\). Assuming that the right hand side \(f\) belongs to a given Banach space \(Y\), we look for a function \(u\) in a given Banach space \(X\) such that \(T_a(u)=f\). This leads to the first formulation of the problem: \emph{Is $T_a :X\to Y$ onto ?} If the answer is positive, then the open map theorem implies that for every \(f\in Y\), there exists some \(u\) in \(X\) such that \(\|u\|_X\leq C \|f\|_Y\), where \(C>0\) is a constant which depends only on \(X, Y\) and \(a\). Let us now assume that such a solution \(u\) exists, which is not unique: for a discussion on the kernel \(N(T_a)\) of \(T_a\), see \cite{Csato-Dacorogna}. Therefore, that one can choose \(u\) linearly with respect to \(f\) is not obvious. This is the second way to address the existence problem: \emph{Does there exist a right inverse to} $T_a$ ? If \(T_a\) is surjective, then \(T_a\) admits a right inverse if and only if the kernel \(N(T_a)\) of \(T_a\) admits a complement in \(X\), see \cite[Theorem 2.12]{Brezis functional a}. When \(X=W^{1,p}_0(\Omega;\R^n)\) and \(Y=L^{p}(\Omega)\) for some \(p\in (1,\infty)\), we will obtain a bounded linear operator \(S_a : L^{p}(\Omega)\to W^{1,p}_0(\Omega;\R^n)\) such that \(T_a\circ S_a (f)=f\) for every \(f\in L^{p}(\Omega)\). \emph{A priori}, such an \(S_a\) depends on the exponent \(p\). We will say that a right inverse \(S_a\) to \(T_a\) is \emph{universal} in the scale of Lebesgue spaces if \begin{enumerate} \item the operator \(S_a\) is well defined on \(\bigcup_{1<p<\infty}L^{p}(\Omega)\) with values into the set \(\bigcup_{1<p<\infty} W^{1,p}_0(\Omega;\R^n)\), \item for every \(p\in (1,\infty)\), \(S_a : L^{p}(\Omega) \to W^{1,p}_0(\Omega ; \R^n)\) is continuous. \end{enumerate} We are thus led to the third formulation of the existence problem: \emph{Does there exist a right inverse to $T_a$ which is universal in the scale of Lebesgue spaces ?} Naturally, one can formulate a similar question in the scale of higher order Sobolev spaces \(W^{k,p}\) and H\"older spaces \(C^{k,\alpha}\), with \(k\in \N\), \(p\in (1,\infty)\) and \(\alpha\in (0,1)\). \subsection{The main results} Our first main result answers the three above questions in the scale of Lebesgue spaces: \begin{theorem}\label{th-main} Let \(\Omega\) be a bounded open Lipschitz set. Let \(q>n\) and \(a\in L^{q}(\Omega;\R^n)\) such that \(a\) is not a gradient: there exists no \(A\in W^{1,q}(\Omega)\) such that \(a=\nabla A\). Then there exists a linear operator \[ S_a : \bigcup_{1<p\leq q}L^{p}(\Omega)\to \bigcup_{1<p\leq q} W^{1,p}_0(\Omega;\R^n) \] such that for every \(1<p\leq q\), the map \(S_{a}\) is continuous from \(L^{p}(\Omega)\) into \(W^{1,p}_0(\Omega;\R^n)\) and \[ \forall f\in L^{p}(\Omega), \qquad \diw S_a(f) + \langle a, S_a(f) \rangle =f. \] \end{theorem} The assumption on the exponent \(q>n\) is related to the fact that the lower order term \(\langle a , u \rangle\) is expected to be in \(L^{p}\) for any \(u\in W^{1,p}\). When \(p<n\), the Sobolev embedding \(W^{1,p}\subset L^{p^*}\), \(p^*=\frac{np}{n-p}\) suggests that we should only require \(a\in L^{n}\). Indeed, under such an assumption, \(\langle a , u \rangle\) belongs to \(L^{p}\) for every \(u\in W^{1,p}_0\). Here, we use the fact that \(\frac{1}{p^*}+\frac{1}{n}=\frac{1}{p}\). In the above statement, we require the slightly stronger assumption \(a\in L^{q}\) for some \(q>n\). When \(p>n\), using the Morrey embedding \(W^{1,p}\subset L^{\infty}\), one can see that the assumption \(a\in L^{p}\) is the natural assumption to ensure that \(\langle a , u \rangle\) belongs to \(L^p\). Finally, when \(p=n\), the fact that \(W^{1,p}_0\subset \bigcap _{1\leq r<\infty}L^{r}\) shows that the assumption \(a\in L^{q}\) for some \(q>n\) is natural to ensure that \(\langle a , u \rangle\) belongs to \(L^{p}\). The operator \(S_a\) is a universal construction in the scale of Lebesgue spaces. In particular, for every \(1<p_1<p_2\), \(S_a|_{L^{p_2}}=(S_{a}|_{L^{p_1}})|_{L^{p_2}}\). As a matter of fact, \(S_a\) does not depend on \(q\), in the following sense: \begin{remark}\label{remark-185} If there exists \(r>q\) such that \(a\in L^{r}(\Omega)\), then \(S_a\) maps continuously \(L^{r}(\Omega)\) into \(W^{1,r}_0(\Omega;\R^n)\). \end{remark} The above remark can be seen as a regularization property of the construction given in the proof of Theorem \ref{th-main}. In order to obtain a universal construction in the whole scales of Sobolev and H\"older spaces, we assume that \(\Omega\) is at least of class \(C^2\). We use the following notation to abbreviate higher order Sobolev and H\"older spaces with zero boundary values: $$ W_z^{m,p}=W^{m,p}\cap W_0^{1,p}\quad\text{ and }\quad C_z^{m,\alpha}=\left\{u\in C^{m,\alpha}:\, u=0\text{ on }\delomega\right\}. $$ In particular, when \(m\geq 2\), the space \(W_{z}^{m,p}\) does not agree with $W_0^{m,p}$ which usually denotes the closure of smooth functions with compact support in $\Omega$ in the norm $W^{m,p}$. \begin{theorem}\label{th-universal-property} Let \(\Omega\) be a bounded open set of class \(C^2\). Let \(q>n\) and \(a\in L^{q}(\Omega;\R^n)\) such that \(a\) is not a gradient. Then there exists an operator \[ \overline{S_a}:\bigcup_{1<p\leq q}L^{p}(\Omega)\to \bigcup_{1<p\leq q} W^{1,p}_0(\Omega;\R^n) \] such that \(\diw S_a(f) + \langle a, S_a(f) \rangle =f\) for every \( f\in \bigcup_{1<p\leq q} L^{p}(\Omega)\). Moreover, we have the following additional properties: \begin{enumerate} \item We assume that \(\Omega\) is of class \(C^{m+2}\) and \(a\in W^{m,r}(\Omega;\R^n)\) for some \(m\in \N\) and \(r>\frac{n}{m+1}\). Then for every \(1<p\leq r\), \(\overline{S_a}\) maps continuously \(W^{m,p}(\Omega)\) into \(W^{m+1,p}_z(\Omega;\R^n)\). \item We assume that \(\Omega\) is of class \(C^{m+2, \alpha}\) and \(a\in C^{m,\alpha}(\Omegabar;\R^n)\) for some \(m\in \N\) and \(\alpha \in (0,1)\). Then \(\overline{S_a}\) maps continuously \(C^{m,\alpha}(\Omegabar)\) into \(C^{m+1,\alpha}_z(\Omegabar;\R^n)\). \end{enumerate} \end{theorem} \begin{remark}\label{remark-regularity-omega} The regularity assumptions that we make on the data are essentially sharp, except possibly for the set \(\Omega\). Indeed, in the scale of Sobolev spaces, one expects that the statement holds true for every \(\Omega\) of class \(C^{m+1}\) instead of class \(C^{m+2}\). Similarly, in the scale of H\"older spaces, the conclusion should be correct when \(\Omega\) is merely of class \(C^{m+1,\alpha}\). However, the proof of Theorem \ref{th-universal-property} relies on the inversion of the divergence (when \(a=0\)) and we are not aware of any \emph{universal} construction of such an inverse under these sharper regularity assumptions on the domain \(\Omega\). \end{remark} \subsection{Comparison with previous results} In \cite{Csato-Dacorogna}, the existence of a solution to \eqref{eq78} is proved when the data are smooth: one assumes that \(\Omega\) is \(C^{r+4}\), \(f\) and \(a\) are \(C^{r+3}\) for some \(r\geq 0\). Moreover, one requires that the domain \(\Omega\) is diffeomorphic to a ball. Finally, the vector field \(a\) must satisfy the following condition: \begin{equation} \label{theorem:curl a assmumption} \operatorname{curl}a(x_0)\neq 0\quad\text{ for some }x_0\in \Omega. \end{equation} Under these assumptions, there exists a solution \(u\in C^{r+1}(\overline{\Omega};\R^n)\), see \cite[Theorem 2]{Csato-Dacorogna}. Under a stronger assumption on the vector field \(a\), namely \begin{equation}\label{eq92} \inf_{x\in \partial \Omega}|\textrm{ curl }a(x)| >0, \end{equation} a right inverse to \(T_a\) is constructed in the setting of H\"older spaces, see \cite[Theorem 3]{Csato-Dacorogna}. In the latter statement, the regularity assumptions are sharp for \(f\), but not for \(\Omega\) or \(a\). It also follows from the proof that the construction is universal in the scale of H\"older spaces. Both conditions \eqref{theorem:curl a assmumption} and \eqref{eq92} imply (but are not equivalent to) the fact that \(a\) is not a gradient. In \cite[Theorem 5]{Csato-Dacorogna}, it was observed that solutions to \eqref{eq78} exist in certain cases even if $\operatorname{curl}a$ vanishes everywhere without any integral condition of $f$, as long as $a$ is not a gradient. In view of Theorem \ref{th-main}, the latter turns out to be the natural assumption for the existence theory of \eqref{eq78}. \begin{remark} The two results in \cite{Csato-Dacorogna} are stated for a more general boundary condition, given by a vector field \(u_0\). However, this case easily reduces to the case \(u_0\equiv 0\), up to a modification of the right hand side \(f\) (see the first step of the proof of \cite[Theorem 2]{Csato-Dacorogna}). \end{remark} \subsection{Some ideas of the proof} We follow a totally different approach from the one used in \cite{Csato-Dacorogna}. The proof of Theorem \ref{th-main} relies on the fact that \eqref{eq78} is a compact perturbation of the classical divergence equation \(\diw u=f\). In order to be more precise, we need to give a quick review of the case $a=0$. Thus we shall consider solutions $u$ of the problem \begin{equation}\label{eq:a is zero}\tag{$E_0$} \begin{cases} \diw u = f & \textrm{ in } \Omega,\\ u=0 & \textrm{ on } \partial \Omega. \end{cases} \end{equation} For every \(1<p<\infty\), we define $L^p_{\sharp}(\Omega)$ as $$ L^p_{\sharp}(\Omega)=\left\{f\in L^p(\Omega):\,\intomega f=0\right\}. $$ Fix \(p\in (1,\infty)\). As explained in \cite[Lemma 10]{Auscher-Russ-Tchamitchian}, the operator \(T_0=\diw\) defined on \(W^{1,p}_0\) is surjective onto \(L^{p}_\sharp\), provided that the range \(R(T_{0}^*)\) of the dual \(T_{0}^*\) is closed. Indeed, this condition implies that the range of \(T_{0}\) is equal to \(N(T_{0}^*)^\perp\), see \cite[Theorem 2.19]{Brezis functional a}. This leads to the desired conclusion since \(N(T_{0}^*)^\perp=L^{p}_{\sharp}\). When \(p=2\), the fact that \(R(T_{0}^*)\) is closed can be obtained as a consequence of the following estimate \cite[Chapitre 3, Lemme 7.1]{Necas}: \begin{equation} \label{eq:Necas estimate} \forall f\in L^{2}(\Omega), \qquad \|f\|_{L^{2}} \leq C \left( \|f\|_{W^{-1,2}} + \|\nabla f\|_{W^{-1,2}}\right). \end{equation} This strategy to prove the surjectivity of \(T_{0}\) thus relies on a \emph{duality} estimate. It does not provide a right inverse to \(T_0\). The duality approach is also one of the main features of the proof given by Bourgain and Brezis in \cite[Theorem 2']{Bourgain-Brezis} to establish the existence of a right inverse to \(T_0\) when \(\Omega\) is a Lipschitz set. Indeed, their argument relies on the following abstract result, see \cite[Lemma 8]{Bourgain-Brezis}: \begin{lemma}\label{lemma_fa} Let \(E\) and \(F\) be two Banach spaces and let \(T\) be a bounded linear operator from \(E\) into \(F\) such that \(N(T^*)=\{0\}\). Assume that there exists a bounded linear operator \(\Tilde{S}\) from \(F\) to \(E\) and a compact linear operator \(K\) from \(F\) into itself such that \(T\circ \Tilde{S} = I+K\). Then \(T\) admits a right inverse \(S\). \end{lemma} The above lemma is then applied to $E=W_0^{1,p}(\Omega;\re^n),$ $F=L^{p}_{\sharp}(\Omega)$ and $T=T_0=\diw,$ where the condition $N(T_{0}^{\ast})=\{0\}$ follows from the fact that $T_{0}^{\ast}$ acts on $L^p_{\sharp}$ (and not $L^p$). The strategy adopted in \cite{Bourgain-Brezis} can be adapted in various settings. It can be exploited in any higher order Sobolev spaces or H\"older spaces, see \cite{HPNG-thesis}, to get an existence theory for \eqref{eq:a is zero} under sharp regularity assumptions on the domain \(\Omega\). For the equation \eqref{eq78}, we will heavily rely on a minor adaptation of the proof of Lemma \ref{lemma_fa} (with $T=T_a=\diw\cdot+\langle a,\cdot\rangle$). To the best of our knowledge, the right inverse \(S\) constructed in the proof of \cite[Theorem 2']{Bourgain-Brezis} depends on the exponent \(p\). In order to get a \emph{universal} right inverse to the divergence operator, at least in the scale of Lebesgue spaces, one can rely on the construction due to Bogovski \cite{Bogovski}, see also \cite[Theorem 4.1]{Acosta-Duran-Muschietti}: \begin{theorem}\label{th-Bogovski} Assume that \(\Omega\) is a Lipschitz set. Then there exists a linear operator \[ S_0 : \bigcup_{1<p<\infty} L^{p}_\sharp(\Omega) \to \bigcup_{1<p<\infty}W^{1,p}_0(\Omega;\R^n) \] such that \(T\circ S_0 (f)= f\) for every \(f\in \bigcup_{1<p<\infty} L^{p}_\sharp(\Omega)\). Moreover, for every \(p\in (1,\infty)\), \(S_0\) is continuous from \(L^{p}_\sharp(\Omega)\) into \(W^{1,p}_0(\Omega;\R^n)\). \end{theorem} The proof of Theorem \ref{th-main} relies on the existence of such an \(S_0\). We first observe that \(T_a\circ S_0 = I +K\) where \(I\) is the identity on \(L^{p}_{\sharp}\) and \(K\) is a compact operator from \(L^{p}_{\sharp}\) into itself. We then prove that \(N(T_{a}^*)=\{0\}\). We next apply Lemma \ref{lemma_fa} with \(T=T_a\) and \(\Tilde{S}=S_0\) to get the desired operator \(S_a\). In order to check that \(S_a\) has the universal properties stated in Theorem \ref{th-main}, we exploit the universal property of the operator \(S_0\) given by Theorem \ref{th-Bogovski}. We also need to detail (and slightly adapt) the explicit construction in the proof of Lemma \ref{lemma_fa}, to ensure that the resulting operator \(S_a\) still possesses the universal property in the scale of Lebesgue spaces. The operator \(S_0\) given by Theorem \ref{th-Bogovski} maps \(W^{k,p}\) into \(W^{k+1,p}_0\) for every \(k\in \N\). However, \(W^{k+1,p}_0\), which is the closure of \(C^{\infty}_c\) in \(W^{k+1,p}\), is strictly contained in \(W^{k+1,p}_z\) defined as the intersection \(W^{1,p}_0\cap W^{k+1,p}\). Hence, we can not use this map \(S_0\) in the setting of Theorem \ref{th-universal-property}. In the framework of higher order Sobolev spaces and H\"older spaces, we will rely instead on the following construction \cite[Theorem 9.2, Remark 9.3]{Csato-Dacorogna-Kneuss}, which requires that \(\Omega\) be at least \(C^2\). \begin{theorem} \label{theorem:a is zero higher regularity} Assume that $\Omega$ is a bounded open set of class \(C^2\). Then there exists a linear continuous operator $\overline{S_0}$ satisfying the same conclusion as in Theorem \ref{th-Bogovski} with the following additional properties: \begin{enumerate} \item We assume that \(\Omega\) is of class \(C^{m+2}\) for some \(m\in \N\). Then for every \(p\in (1,\infty)\), \(\overline{S_0}\) maps continuously \((W^{m,p}\cap L^{1}_{\sharp})(\Omega)\) into \(W^{m+1,p}_z(\Omega;\R^n)\). \item We assume that \(\Omega\) is of class \(C^{m+2, \alpha}\) for some \(m\in \N\) and \(\alpha \in (0,1)\). Then \(\overline{S_0}\) maps continuously \((C^{m,\alpha}\cap L^{1}_{\sharp})(\overline{\Omega})\) into \(C^{m+1,\alpha}_z(\overline{\Omega};\R^n)\). \end{enumerate} \end{theorem} In the proof of Theorem \ref{th-universal-property}, the above map \(\overline{S_0}\) plays a crucial role, in a similar way as the operator \(S_0\) is used in the proof of Theorem \ref{th-main}. We also rely on standard properties of the pointwise multiplication in higher order Sobolev spaces and in H\"older spaces. \subsection{Plan of the paper} The next section is dedicated to the proof of Theorem \ref{th-main} while Theorem \ref{th-universal-property} is proven in Section 3. In the last section, we discuss the non-existence in $L^1$ and $L^{\infty}.$ In contrast to the previous existence and regularity results, the proof is exactly the same as for the case $a=0.$ Finally, for the convenience of the reader, we have gathered in the Appendix some technical tools. \section{Construction of \(S_a\) on \(L^{p}\)} The construction of $S_a$ is inspired from the proof of Lemma \ref{lemma_fa} in \cite{Bourgain-Brezis}. In our setting, we take \(E=W^{1,p}_0(\Omega;\R^n)\), \(F=L^{p}(\Omega)\) for some \(p\in (1,\infty)\) and \begin{equation} \label{eq:T definition} T_a(u)=\diw u + \langle a , u \rangle. \end{equation} Throughout this section, we assume that \(\Omega\) is Lipschitz. We first observe that $T_a$ is continuous: \begin{lemma} \label{lemma:T continuous} Let $T_a$ be given by \eqref{eq:T definition}. If $q> n,$ $a\in L^q(\Omega),$ then for every $1\leq p\leq q$, $T_a$ maps continously \(W^{1,p}(\Omega)\) into $L^p(\Omega)$. \end{lemma} \begin{proof} We recall that for two functions $f\in L^r$ and $g\in L^s$, we have \begin{equation} \label{eq:multipliciation in Lp} fg\in L^t\quad\text{ for all $t$ satisfying }\quad \frac{1}{r}+\frac{1}{s}\leq \frac{1}{t}. \end{equation} We only need to show that $\langle a,u\rangle\in L^p$ with the appropriate estimate. If $p<n$ then by the Sobolev embedding, $u\in L^{p^*}$ with $p^*=\frac{np}{n-p}$. Using \eqref{eq:multipliciation in Lp} and the assumption $q\geq n$, it follows that $\langle a,u\rangle\in L^p$. For the case $p=n$, one uses that $W^{1,n}$ embeds into any $L^s$ as long as $s<\infty$ and one applies again \eqref{eq:multipliciation in Lp} together with the fact that $q>n=p$. In the third case $n<p\leq q$, one relies on the Morrey embedding \(W^{1,p}\subset L^{\infty}\) to conclude. \end{proof} \smallskip The next lemma is the key tool to prove that $N(T_{a}^{\ast})=\{0 \}.$ This is the step where the assumption that \emph{\(a\) is not a gradient} plays a crucial role. \begin{lemma}\label{lemma_kernel} Let $\Omega\subset\re^n$ be a bounded open Lipschitz set and \(a\in L^{q}(\Omega;\R^n)\) for some \(q\in [1,\infty]\). We assume that \(a\) is not a gradient: there exists no $A\in W^{1,q}(\Omega)$ such that $\nabla A=a.$ If \(g\in L^{q'}(\Omega)\) is such that \(\nabla g =g a\) in the sense of distributions, i.e. $$ \intomega g\diw u =-\intomega \langle g a,u\rangle\quad \text{for all }u\in C_c^{\infty}(\Omega,\re^n), $$ then $$ g\equiv 0. $$ \end{lemma} \begin{remark} The proof of this lemma is much simpler if one assumes $a\in C^0(\Omega;\re^n).$ Applying repeatedly the Sobolev and Morrey embeddings, one obtains that $g\in C^1$ and $\nabla g=ga$ in a classical sense. Let $x,y\in\Omega$ and assume that $\gamma$ is a $C^1$ curve connecting $x$ and $y$ with $\gamma(0)=x$ and $\gamma(1)=y.$ Thus $g\circ\gamma$ satisfies the differential equation $ (g\circ\gamma)'=\langle \nabla g\circ\gamma,\gamma'\rangle=(g\circ\gamma)\langle a\circ\gamma,\gamma'\rangle $ and it follows that \begin{equation*} g(y)=g(x)e^{\int_0^1\langle a(\gamma(t)),\gamma'(t)\rangle dt}. \end{equation*} This implies that either $g\equiv 0$ or $g(x)\neq 0$ for all $x\in\Omega$. But the second case cannot occur, because if $g$ never vanishes, then $A=\ln |g| \in C^1(\Omega)$ satisfies $\nabla A=\nabla g/g=(ga)/g=a,$ which is a contradiction to the hypothesis on $a.$ \end{remark} \begin{proof} Without loss of generality, one can assume that \(\Omega\) is connected. The equality \(\nabla g =g a\) implies that \(\nabla g \in L^{1}(\Omega)\) and thus \(g\in W^{1,1}(\Omega)\). Let \(Q\) be a cube contained in \(\Omega\). Up to a dilation and an isometry, we can assume that \(Q=(0,1)^n \subset \Omega\). For almost every \(x'\in (0,1)^{n-1}\), the map \(g_{x'}:x_n\mapsto g(x',x_n)\) belongs to \(W^{1,1}(0,1)\) while \(x_n\mapsto a_n(x',x_n)\) belongs to \(L^{1}(0,1)\) and moreover \[ g_{x'}'(x_n) = a_n(x',x_n)g_{x'}(x_n). \] It thus follows that for such \(x'\), for every \(x_n, y_n \in [0,1]\), \begin{equation}\label{eq590} g(x',y_n)=g(x',x_n) e^{\int_{x_n}^{y_n}a_n(x',t)\,dt}. \end{equation} Repeating the above argument in every direction parallel to the coordinate axes between \[ (x_{1}, \dots, x_{i-1}, x_{i}, y_{i+1}, \dots, y_n) \textrm{ and } (x_{1}, \dots, x_{i-1}, y_{i}, y_{i+1}, \dots, y_n), \qquad 1\leq i \leq n, \] we deduce that for a.e. \(x,y\in Q\), \(g(y)=g(x)e^{D(y,x)}\) with \begin{equation}\label{eq589} D(y,x)=\sum_{i=1}^n \int_{x_i}^{y_i}a_i(x_1,\dots, x_{i-1}, t, y_{i+1}, \dots , y_n)\,dt. \end{equation} In particular, either \(g\equiv 0\) on \(Q\) or \(g>0\) a.e. on \(Q\) or \(g<0\) a.e. on \(Q\). Since for every two cubes \(Q_1, Q_2 \subset \Omega\) such that \(Q_1\cap Q_2 \not=\emptyset\), the \emph{same} conclusion among the three above alternatives must hold true, the connectedness of \(\Omega\) implies that either \(g>0\) a.e. on \(\Omega\), or \(g<0\) a.e. on \(\Omega\) or \(g\equiv 0\) on \(\Omega\). Assume by contradiction that \(g>0\) a.e. on \(\Omega\) and consider again the cube \(Q=(0,1)^n \subset \Omega\). By the Fubini theorem and the fact that the function \(D\) defined in \eqref{eq589} belongs to \(L^{1}(Q\times Q)\), for a.e. \(x\in Q\), the function \(y\mapsto D(y,x)\) belongs to \(L^{1}(Q)\). We fix such an \(x\) for which we further require that \(g(x)>0\). Then the identity \(g(y)=g(x)e^{D(y,x)}\) shows that \(\ln g \in L^{1}(Q)\). We claim that for every \(\varphi \in C^{\infty}_c(Q)\) and \(1\leq i \leq n\), \begin{equation}\label{eq607} \int_{Q} \ln g \, \partial_i \varphi = -\int_{Q} a_i \varphi. \end{equation} Let us prove the claim for \(i=n\). By the Fubini theorem, \begin{align*} \int_{Q} \ln g \, \partial_n \varphi &= \int_{(0,1)^{n-1}}\,dx' \int_{0}^1 \ln g (x',t_n) \, \partial_n \varphi(x',t_n)\,dt_n\\ & = \int_{(0,1)^{n-1}}\,dx'\int_{0}^1 \ln g (x',0)\, \partial_n \varphi(x',t_n)\,dt_n \\ &\qquad + \int_{(0,1)^{n-1}}\,dx'\int_{0}^1 \left(\int_{0}^{t_n}a_n(x',s)\,ds\right) \partial_n \varphi(x',t_n)\,dt_n \end{align*} where the last line follows from \eqref{eq590} with \(x_n=0\) and \(y_n=t_n\). Now, by the Fubini theorem, \[ \int_{0}^1 \left(\int_{0}^{t_n}a_n(x',s)\,ds\right) \partial_n \varphi(x',t_n)\,dt_n = -\int_{0}^{1} \varphi(x',s)a_n(x',s)\,ds. \] Since \[ \int_{0}^1 \ln g (x',0)\, \partial_n \varphi(x',t_n)\,dt_n = \ln g(x',0) \int_{0}^1 \partial_n \varphi(x',t_n)\,dt_n=0, \] we get \[ \int_{Q} \ln g \, \partial_n \varphi = -\int_{Q}a_n \varphi. \] We can repeat this calculation in every direction \(i=1, \dots n\) by using the identity corresponding to \eqref{eq590} where \(n\) is replaced by \(i\). This proves claim \eqref{eq607}. We deduce therefrom that \(\ln g \in W^{1,q}(Q)\) with \(\nabla \ln g = a\). Since this is true for every cube \(Q \subset \Omega\), this implies that \(a=\nabla (\ln g)\) on \(\Omega\), which contradicts the fact that \(a\) is not a gradient on \(\Omega\). Hence, we cannot have \(g>0\) a.e. The case \(g<0\) a.e. can be treated similarly. This proves that \(g\equiv 0\) as desired. \end{proof} We proceed to explain how the above lemma implies that \(N(T_{a}^*)=\{0\}\). \begin{lemma}\label{lm-nta-t} Let $q>n$ and suppose that $a\in L^q(\Omega;\re^n)$ is not the gradient of a $W^{1,q}$ function. Then for every \(1<p\leq q\), the operator \(T_a : u\in W^{1,p}(\Omega;\R^n) \mapsto \diw u + \langle a , u \rangle \in L^{p}(\Omega)\) satisfies: \[ N(T_{a}^*)=\{0\}. \] \end{lemma} \begin{proof} Let \(g\in (L^{p})^*\) such that \(T_{a}^*(g)=0\). Identifying \(g\) with an element of \(L^{p'}\), \(p'=\frac{p}{p-1}\), this means that \[ \forall u \in W^{1,p}_0, \qquad \int_{\Omega}g(\diw u + \langle a , u \rangle)=0. \] Note that $a\in L^q\subset L^p$. Moreover, \(a\) is not the gradient of a $W^{1,p}$ function for otherwise, there would exist \(A\in W^{1,p}\) such that \(a=\nabla A\). This would imply that \(\nabla A \in L^{q}\) and thus by the Poincar\'e-Wirtinger inequality, \(A\in W^{1,q}\), a contradiction to the assumption on \(a\). In view of Lemma \ref{lemma_kernel} applied with \(p\) instead of \(q\), we deduce that \(g=0\). This proves that \(N(T_{a}^*)=0\). \end{proof} By the Hahn-Banach theorem, see e.g. \cite[Corollary 1.8]{Brezis functional a}, this implies that the range \(R(T_a)\) of \(T_a\) is dense in \(L^{p}(\Omega)\). We shall see later that in fact \(R(T_a)=L^{p}(\Omega)\). We now introduce two operators that will play a crucial role in the sequel. \begin{definition} \label{definition of S and K} Let $\Omega\subset\re^n$ be a bounded open Lipschitz set. Let $S_0$ be the map given by Theorem \ref{th-Bogovski}. We then define for \(f\in \bigcup_{1<p<\infty}L^{p}(\Omega)\): \begin{equation}\label{eq_def_S} S(f) = S_0\left(f-\fint_{\Omega}f\right), \end{equation} \begin{equation}\label{eq_def_K} K(f)=- \fint_{\Omega}f +\langle a ; S(f) \rangle. \end{equation} \end{definition} For every \(1<p<\infty\), \(S\) defines a continuous linear map from \(L^{p}(\Omega)\) into \(W^{1,p}_0(\Omega;\re^n)\). The map $K$ is continuous as well, but also compact under an appropriate assumption on $a.$ \begin{lemma}\label{lemma_compacity_K} Assume that \(a\in L^{q}(\Omega;\re^n)\) for some \(q>n\). For every \(1<p\leq q\), the map \(K\) is a compact linear map from \(L^{p}(\Omega)\) into \(L^{p}(\Omega)\). \end{lemma} \begin{proof} We define the exponent \(r\in (1,\infty]\) by \(\frac{1}{r}+\frac{1}{q}=\frac{1}{p}\). When \(p<n\), the fact that \(q>n\) implies that \(r< p^*=\frac{np}{n-p}\). It follows that the embedding \(W^{1,p}_0\subset L^{r}\) is compact and this remains true when \(p\geq n\). By the H\"older inequality, the map \[ u\in L^r(\Omega;\R^n) \mapsto \langle a ; u \rangle \in L^p(\Omega) \] is continuous. It thus follows that the map \[ u\in W^{1,p}_0(\Omega;\R^n) \mapsto \langle a ; u \rangle \in L^p(\Omega) \] is compact as the composition of a continuous operator with a compact one. Composing again by the continuous operator \(S\), we infer that the map \(f\in L^p \mapsto \langle a , S(f)\rangle \in L^p\) is compact. The map $f\mapsto \fint_{\Omega} f$ is clearly compact from $L^p$ to $L^p.$ It follows that \(K\) is compact as well. \end{proof} \smallskip We observe that for every \(f\in \bigcup_{1<p\leq q}L^{p}(\Omega)\), \begin{equation} \label{eq:TS is I plus K} \begin{split} T_a\circ S (f) &= \diw S_0\left( f-\fint_{\Omega}f\right) +\langle a ; S_0\left( f- \fint_{\Omega}f\right)\rangle\\ &=f-\fint_{\Omega}f +\langle a ; S_0\left( f- \fint_{\Omega}f\right)\rangle=(I+K)(f). \end{split} \end{equation} \begin{remark}\label{remark-498} In view of Lemma \ref{lm-nta-t} and Lemma \ref{lemma_compacity_K}, one can apply Lemma \ref{lemma_fa} to \(\widetilde{S}=S\), where \(S\) is defined in \eqref{eq_def_S}. We thus obtain that \(T_a:W^{1,p}_0(\Omega;\R^n) \to L^{p}(\Omega)\) has a right inverse, provided that \(a\) is not a gradient. However, we shall not use Lemma \ref{lemma_fa} directly, but slightly modify its original proof from \cite{Bourgain-Brezis} to obtain a \emph{universal} construction, first in the whole scale of Lebesgue spaces (this will imply Theorem \ref{th-main}), and next, in higher order Sobolev and H\"older spaces (to get Theorem \ref{th-universal-property}). \end{remark} The kernel of the operator \(I+K\) only contains \(L^q\) functions provided that \(a\in L^q\). More generally, \begin{lemma}\label{lm_inva_p} Let \(q>n\) and \(a\in L^{q}(\Omega;\re^n)\). Then for every \(1<p\leq q\) and every \(f\in L^{p}(\Omega)\), if \((I+K)(f)\in L^{q}(\Omega)\), then \(f\in L^{q}(\Omega)\). \end{lemma} \begin{proof} Let \(f\in L^{p}\) such that \[ \widetilde{f} +\langle a , S_0(\widetilde{f})\rangle \in L^{q}(\Omega), \qquad \widetilde{f}=f-\fint_{\Omega}f. \] We now distinguish three cases: \smallskip \textit{Case 1.} Assume first that \(p>n\). Since \(S_0\) maps \(L^{p}_\sharp\) into \(W^{1,p}\), it follows from the Morrey embedding that \(S_0(\widetilde{f})\in L^{\infty}\). Since \(a\in L^{q}(\Omega)\), we have \(\langle a , S_0(\widetilde{f})\rangle \in L^{q}(\Omega)\), which completes the proof in that case. \smallskip \textit{Case 2.} When \(p=n\), we use that \(S_0(\widetilde{f})\in W^{1,n}\). The latter space being contained in \( \bigcap_{1\leq r <\infty} L^{r}\), we deduce that \[ \langle a , S_0(\widetilde{f})\rangle \in \bigcap_{1\leq r <q}L^r. \] Since $q>n,$ we can choose some $r\in (n,q)$ such that $\langle a , S_0(\widetilde{f})\rangle \in L^r$. Hence $\widetilde{f}\in L^r$ and we are thus reduced to the first case. \smallskip \textit{Case 3.} Finally, if \(p<n\), we rely on the fact that \(S_0\) maps \(L^{r}_\sharp\) into \(W^{1,r}\) for every \(1<r<\infty\) and the Sobolev embedding \(W^{1,r}\subset L^{r^*}\), with \(\frac{1}{r^*}=\frac{1}{r}-\frac{1}{n}\). For \(r=p\), this first implies that \(S_0(\widetilde{f})\in L^{p^*}\). Since \(a\in L^{q}\) with \(q>n\), if follows that \(\langle a , S_0(\widetilde{f})\rangle \in L^{p_1}\) with \[ \frac{1}{p_1}=\frac{1}{p^*} + \frac{1}{q} = \frac{1}{p} + \frac{1}{q}-\frac{1}{n}. \] Since \(p_1<q\), one also has \begin{equation}\label{eq540} \widetilde{f}\in L^{p_1}. \end{equation} Since \(\frac{1}{q}-\frac{1}{n} <0\), one gets that \(p_1>p\). If \(p_1<n\), one can repeat this argument with \(r=p_1\). This proves that \(\widetilde{f}\in L^{p_2}\) with \[ \frac{1}{p_2}= \frac{1}{p_1} + \frac{1}{q}-\frac{1}{n}=\frac{1}{p} + 2\left(\frac{1}{q}-\frac{1}{n}\right). \] Let us define the sequence \((p_k)_{k\in \N}\) as $$ \frac{1}{p_k}=\frac{1}{q}-\frac{1}{n}+\frac{1}{p_{k-1}}=\frac{1}{p}+k\left(\frac{1}{q}-\frac{1}{n}\right). $$ We observe that if \(1<p_{k-1}<n\), then \(\frac{1}{p_k}>\frac{1}{q}\) and thus \(p_k>0\). Since \(\frac{1}{p_k}<\frac{1}{p_{k-1}}\), it follows that \(p_k>1\). Moreover, the sequence \((\frac{1}{p_k})_{k\in \N}\) tends to \(-\infty\). Hence, there exists \(k_0\in \N\) such that \(1<p_{k}<n\) for every \(k\leq k_0\) and \(p_{k_{0}+1}\geq n\). Bootstrapping the argument leading to \eqref{eq540}, we deduce that \(\widetilde{f}\) belongs to \(L^{p_{k_0+1}}\). The two first parts of the proof then apply to yield \(\widetilde{f}\in L^{q}\). \end{proof} \smallskip According to Remark \ref{remark-498}, we need to check that the construction described in the original proof of Lemma \ref{lemma_fa}, see \cite{Bourgain-Brezis}, can be slightly adapted to yield a \emph{universal} right inverse to \(T_a\). A first tool is provided by the following: \begin{lemma} \label{lemma:decomp of Lp in Xp plus N} Let $q>n$ and suppose that $a\in L^q(\Omega;\re^n).$ \smallskip (i) Then there exists a closed subspace $X=X_q$ of $L^1(\Omega)$ (which is independent of $p$) such that for every $1<p\leq q$ $$ L^p(\Omega)=(X\cap L^p(\Omega))\oplus N, $$ where $N=\{f\in L^q(\Omega):\, (I+K)(f)=0\}.$ \smallskip (ii) If in addition $a\in L^r(\Omega;\re^n)$ for some $r\geq q,$ then \(N\subset L^{r}(\Omega)\) and for the same $X=X_q$ of (i), one has $$ L^r(\Omega)=(X\cap L^r(\Omega))\oplus N. $$ \end{lemma} \begin{proof} (i) By Lemma \ref{lemma_compacity_K}, the map \(K\) is a compact operator from \(L^{q}\) into \(L^{q}\). This implies that \(N\) is a finite dimensional subspace of \(L^{q}\), and thus, a finite dimensional subspace of \(L^{1}\). Hence, there exists a closed subspace \(X\) of \(L^{1}\) such that \[ L^{1}=X \oplus N. \] For every \(1<p\leq q\), the fact that \(N\subset L^p\) implies that \[ L^{p}=(X\cap L^p) \oplus N. \] Note that $X\cap L^p$ is a closed subspace of $L^p.$ \smallskip (ii) First observe that if $a\in L^r,$ then by Lemma \ref{lm_inva_p} with \(r\) instead of \(q\), one has \(N\subset L^{r}\). We can thus conclude as in (i) that \(L^{r}=(X\cap L^r) \oplus N\). \end{proof} \begin{lemma} \label{lemma:decomp of Lp in range plus Z} Let $q>n$ and suppose that $a\in L^q(\Omega;\re^n)$ is not the gradient of a $W^{1,q}$ function. Then there exists a finite dimensional space $Z=Z_q\subset T_a(C_c^{\infty}(\Omega;\re^n))$ such that for all $1<p\leq q$ \begin{equation} \label{eq369} (I+K)(L^p(\Omega))\oplus Z=L^p(\Omega). \end{equation} \end{lemma} \begin{proof} By Lemma \ref{lm-nta-t}, \(\{f\in L^{q'} : T_{a}^*(f)=0\}=\{0\}\), which implies that \(T_{a}(W^{1,q}_0)\) is dense in \(L^{q}\) by the Hahn-Banach theorem. Since \(C^{\infty}_c\) is dense in \(W^{1,q}_0\), it follows that \(T_{a}(C^{\infty}_c)\) is dense in \(L^q\). Moreover, \((I+K)(L^q)\) has finite codimension in \(L^{q}\), because $K:L^q\to L^q$ is compact, see \cite[Theorem 6.6]{Brezis functional a}. In view of Lemma \ref{lemma_complementing_density} in the Appendix, there exists a finite dimensional space \(Z\subset T_{a}(C^{\infty}_c)\) such that \(Z\oplus (I+K)(L^{q})=L^{q}\). We claim that $Z$ has the desired property. We first prove that $(I+K)(L^p)\cap Z=\{0\}.$ Indeed, let \(f\in L^{p}\) such that \((I+K)(f)\in Z\). Since \(Z\subset L^q\), Lemma \ref{lm_inva_p} implies that \(f\in L^{q}\). Hence, \((I+K)(f)\in Z\cap (I+K)(L^q)=\{0\}\). This proves that $(I+K)(L^p)\cap Z=\{0\}.$ Since $K:L^p\to L^p$ is compact, $(I+K)(L^p)$ is closed in $L^p,$ see \cite[Theorem 6.6]{Brezis functional a}. Using that \(Z\) is finite dimensional, it follows that \((I+K)(L^p)\oplus Z\) is closed in \(L^p\). Since \((I+K)(L^p)\oplus Z\) contains \((I+K)(L^q)\oplus Z = L^q\) which is dense in \(L^{p}\), we deduce that \(Z\oplus (I+K)(L^p)=L^{p}\), as desired. \end{proof} \smallskip Having prepared all the necessary ingredients, we can now complete the \begin{proof}[Proof of Theorem \ref{th-main}] \textit{Step 1.} Fix \(p\in (1,q]\). By Lemma \ref{lm_inva_p}, \(\{f\in L^{p} : (I+K)(f)=0\}=\{f\in L^{q} : (I+K)(f)=0\}\). By construction of \(X\) and Lemma \ref{lemma:decomp of Lp in Xp plus N} (i), the map \(I+K\) defines an isomorphism from \(X\cap L^p\) onto \((I+K)(L^p)\) which are two Banach spaces, as two closed subsets of Banach spaces. We denote by \(V_p : (I+K)(L^p) \to X\cap L^p\) its inverse (continuous by the inverse mapping theorem) and we have the following diagram $$ \begin{array}{cccccc} L^p & = & X\cap L^p & \oplus & N &\phantom{aaaaaaa}\text{ where }N=N(I+K)\\ & & \;\vert \quad\quad\; \uparrow & & & \\ & & I+K \quad V_p & & & \phantom{aaaaaa} V_p=\left((I+K)\big|_{X\cap L^p}\right)^{-1} \\ & & \,\downarrow \quad\quad\; \vert & & & \\ L^p & = & (I+K)(L^p) & \oplus& Z . & \end{array} $$ We claim that for every \(1<p_1\leq p_2 \leq q\) $$ V_{p_1}(f)=V_{p_2}(f) \quad\text{ for every }f\in (I+K)(L^{p_2}). $$ In fact, for every \(f\in (I+K)(L^{p_2})\), let \(g\) be the unique element in \(X\cap L^{p_2}\) such that \((I+K)(g)=f\). Then \(V_{p_2}(f)=g\). Moreover, \(g\in X\cap L^{p_1}\) so that \(V_{p_1}(f)=g\). In view of Lemma \ref{lemma:decomp of Lp in range plus Z}, for every \(1<p\leq q\), \(L^{p}= (I+K)(L^p)\oplus Z\). Hence, there exist two continuous projections \[ Q_p : L^{p} \mapsto (I+K)(L^p), \] \[ \zeta_p : L^{p} \mapsto Z, \] such that for every \(f\in L^{p}\), \(f= Q_p(f) + \zeta_p(f)\). Let \((e_\alpha)_{\alpha}\) be a finite basis of \(Z\). Since \(Z\subset T(C^{\infty}_c)\), there exists \((\bar{e}_\alpha)\subset C^{\infty}_c\) such that \(T(\bar{e}_\alpha)=e_\alpha\) for every \(\alpha\). Let \((e_{\alpha}^*)_{\alpha}\subset Z^*\) be the dual basis of \((e_\alpha)_\alpha\), which has the property $$ \sum_{\alpha}e_{\alpha}^*(g)e_{\alpha}=g\quad\text{ for all }g\in Z. $$ At last we define for every \(1<p\leq q\), for every \(f\in L^p\), \[ S_{a}^p(f) = S\circ V_p \circ Q_p (f)+ \sum_{\alpha} e_{\alpha}^* \circ \zeta_p(f)\bar{e}_\alpha. \] Remember that \(S\) is given by \eqref{eq_def_S}. Then, by composition, \(S_{a}^p\) is a linear continuous map from \(L^{p}\) into \(W^{1,p}_0\). Moreover, by \eqref{eq:TS is I plus K}, \(T_a\circ S = I+K\), and thus \begin{align*} T_a\circ S_{a}^p(f) &= T_a\circ \left(S\circ V_p \circ Q_p(f) + \sum_{\alpha} e_{\alpha}^* \circ \zeta_p(f)\bar{e}_\alpha\right)\\ &=T_a\circ S\circ V_p \circ Q_p(f) + \sum_{\alpha} e_{\alpha}^*\circ \zeta_p(f)T(\bar{e}_\alpha)\\ &=(I+K) \circ V_p \circ Q_p(f) + \sum_{\alpha} e_{\alpha}^*\circ \zeta_p(f) e_\alpha\\ &=Q_p(f) + \zeta_p(f)=f. \end{align*} \textit{Step 2.} The above construction is universal: if \(1<p_1\leq p_2\leq q\), then for every \(f\in L^{p_2}\), we claim that \(S_{a}^{p_1}(f)=S_{a}^{p_2}(f)\). Indeed, one can write \(f= Q_{p_2}(f)+\zeta_{p_2}(f)\) according to the decomposition \(L^{p_2}=(I+K)(L^{p_2}) \oplus Z\). Since \((I+K)(L^{p_2})\subset (I+K)(L^{p_1})\), this is also the decomposition of \(f\) in \(L^{p_1}=(I+K)(L^{p_1}) \oplus Z\): \[ Q_{p_2}(f)=Q_{p_1}(f) \qquad \textrm{ and } \qquad \zeta_{p_2}(f)=\zeta_{p_1}(f). \] We have already checked that \(V_{p_1}\) and \(V_{p_2}\) agree on \((I+K)(L^{p_2})\). Hence, \begin{align*} S_{a}^{p_1}(f) &= S\circ V_{p_1} \circ Q_{p_1}(f) + \sum_{\alpha} e_{\alpha}^{*}\circ \zeta_{p_1}(f)\bar{e}_{\alpha}\\ &=S\circ V_{p_2} \circ Q_{p_2}(f) + \sum_{\alpha} e_{\alpha}^{*}\circ \zeta_{p_2}(f)\bar{e}_\alpha\\ &=S_{a}^{p_2}(f). \end{align*} It follows that we can define the map \(S_a : \bigcup_{1<p\leq q}L^{p} \to \bigcup_{1<p\leq q}W^{1,p}_0\) by setting \(S_{a}|_{L^{p}}=S_{a}^p\). \end{proof} \smallskip We now prove that the above construction of \(S_{a}\) does not depend on \(q\), see Remark \ref{remark-185}: \begin{lemma} \label{lemma:q replace by r} If \(a\in L^{r}(\Omega)\) for some \(r\geq q\), then \(S_a\) maps continuously \(L^{r}(\Omega)\) into \(W^{1,r}_{0}(\Omega)\). \end{lemma} \begin{remark} \label{remark:unicity of Sa} Given some $a\in L^r(\Omega;\re^n),$ the map $S_a$ is uniquely determined by the choice of $S_0,$ by the choice of some $q\leq r$ and $X=X_q$ (in Lemma \ref{lemma:decomp of Lp in Xp plus N}), and finally by the choice of the $\overline{e}_{\alpha}$ (which in turn depends on the choice of $Z=Z_q$). \end{remark} Lemma \ref{lemma:q replace by r} is an easy consequence of the following more general result which will be needed in the next section: \begin{lemma}\label{lm_univ_prop} Let $\Omega\subset\re^n$ be a bounded open Lipschitz set. Let $q>n$ and $a\in L^q(\Omega;\re^n)$. Let $1<p\leq q$ and \(E\), \(F\) be two Banach spaces continuously embedded in \(W^{1,p}_0(\Omega;\R^n)\) and \(L^{p}(\Omega)\) respectively. We assume that \begin{enumerate} \item The set \(E\) contains \(C^{\infty}_c(\Omega;\R^n)\). \item For every \(\varphi\in C^{\infty}_c(\Omega;\R^n)\), \(\diw \varphi+\langle a, \varphi \rangle\) belongs to \(F\). \item For every \((u,f)\in L^{p}(\Omega)\times F\), \begin{equation}\label{eq_univ_prop} (I+K)(u)=f \quad\Rightarrow\quad u\in F. \end{equation} \item The map \(S\) defined in \eqref{eq_def_S} maps continuously \(F\) into \(E\). \end{enumerate} Then the map \(S_a\) constructed in Theorem \ref{th-main} maps continuously \(F\) into \(E\). \end{lemma} We obtain Lemma \ref{lemma:q replace by r} from Lemma \ref{lm_univ_prop} by taking \(p=q\), \(E=W^{1,r}_0\) and \(F= L^{r}\). The assumption \eqref{eq_univ_prop} is satisfied thanks to Lemma \ref{lm_inva_p}. \begin{proof}[Proof of Lemma \ref{lm_univ_prop}] By the closed graph theorem, we only need to prove that \(S_a(F)\subset E\). By construction, for every \(f\in L^{p}\), \[ S_a(f) = S\circ V_p \circ Q_p(f) + \sum_{\alpha}e_{\alpha}^* \circ \zeta_p(f)\bar{e}_\alpha. \] Each \(\bar{e}_\alpha\) belongs to \(C^{\infty}_c\) and thus the second term belongs to \(E\). We proceed to prove that if \(f\) belongs to \(F\), then \(S\circ V_p \circ Q_p(f)\) belongs to \(E\). Since \(S(F)\subset E\), this amounts to proving that \(V_p\circ Q_p(F) \subset F\). Let \(f\in F\). It follows from the definition of \(T_a\) and the second assumption that \(T_a(C^{\infty}_c)\subset F\). Since by construction, \(Z\) is contained in \(T_a(C^{\infty}_c)\), we deduce that \(Z\) is a subset of \(F\) and thus \(\zeta_p(f)\in F\). Using the decomposition \(f=Q_p(f)+ \zeta_p(f)\) and the fact that \(f\in F\), one gets that \(Q_p(f)\in F\). Let \(u= V_p \circ Q_p(f)\). Then \((I+K)(u)=Q_p(f)\in F\). We now rely on \eqref{eq_univ_prop} to get that \(u\in F\); that is \(V_p\circ Q_p(f)\in F\). The proof is complete. \end{proof} \section{Universal property for \(S_a\)} \label{section:higher regularity} In this section we deal with higher order Sobolev and H\"older spaces. As explained in the introduction, a right inverse \(\overline{S_0}\) to the divergence operator \(T_0=\diw\) which is universal in the scales of these spaces, is only available in the literature when \(\Omega\) is at least \(C^2\), see Theorem \ref{theorem:a is zero higher regularity}. Hence, we assume this regularity property on \(\Omega\) throughout this section. We can repeat the same construction as in Section 2 with \(\overline{S_0}\) instead \(S_0\). More precisely, in the definition \eqref{eq_def_S} of \(S\), the operator \(S_0\) has to be replaced by \(\overline{S_0}\). We thus obtain a linear map \(\overline{S_a}\) which is a right inverse to \(T_a\) and is universal in the scale of Lebesgue spaces. We proceed to prove that \(\overline{S_a}\) is also universal in the scale of higher order Sobolev spaces and H\"older spaces. We rely on Lemma \ref{lm_univ_prop} and we first check that the assumption \ref{eq_univ_prop} is satisfied in our framework. \begin{lemma}\label{lm-univ-sobolev} Let \(m\geq 0\), \(r> \frac{n}{m+1}\) and \(a\in W^{m,r}(\Omega ; \R^n)\). Assume that \(\Omega\) is of class \(C^{m+2}\). Then for every \(1<p\leq r\) and \(f\in L^{p}(\Omega)\), \[ (I+K)(f)\in W^{m,p}(\Omega) \;\Longrightarrow\; f\in W^{m,p}(\Omega). \] \end{lemma} \begin{proof} We prove by induction on \(m\in \N\) the following more general result: If \(\Omega\) is of class \(C^{m+2}\) and \(a\in W^{m,r}(\Omega;\re^n)\) with \(r>\frac{n}{m+1}\), then for every \(1<p\leq s\leq r\) and every \(f\in L^{p}(\Omega)\), \[ (I+K)(f)\in W^{m,s}(\Omega) \;\Longrightarrow\; f\in W^{m,s}(\Omega). \] Remember that $$ (I+K)(f)=\widetilde{f} + \langle a, \overline{S_0}(\widetilde{f})\rangle \qquad \textrm{ where } \widetilde{f}=f-\fint_{\Omega}f. $$ We first consider the case \(m=0\) which is a slight generalization of Lemma \ref{lm_inva_p}. If \(p=s\), there is nothing to prove. Hence, we assume that \(p<s\). If \(s>n\), then we can apply Lemma \ref{lm_inva_p} with \(q=s\), which gives that \(f\in L^s\). If \(s\leq n\), we rely on the fact that \(\overline{S_0}(\widetilde{f})\in W^{1,p}\) and the Sobolev embedding \(W^{1,p}\subset L^{p^*}\) (observe that we are in the case where \(p<s\leq n\)) to get that \(\overline{S_0}(\widetilde{f})\in L^{p^*}\). This implies that \begin{equation}\label{eq794} \langle a, \overline{S_0}(\widetilde{f}) \rangle \in L^{p_1}, \qquad \textrm{with } \frac{1}{p_1}=\frac{1}{p}-\frac{1}{n}+\frac{1}{r}. \end{equation} If \(p_1\geq s\), then \(\widetilde{f}\in L^{s}\), as desired. Otherwise, \(\widetilde{f}\in L^{p_1}\). We next introduce as in the proof of Lemma \ref{lm_inva_p} the sequence \[ \frac{1}{p_k} =\frac{1}{p_{k-1}} +\frac{1}{r} -\frac{1}{n}, \qquad k\in \N. \] We have already seen that there exists \(k_0\in \N\) such that \(1<p_k<n\) for \(k\leq k_0\) and \(p_{k_0+1}\geq n\). This implies that there exists \(k_1\in \N\) such that \(1<p_k<s\) for every \(k\leq k_1\) and \(p_{k_1+1}\geq s\). Repeating the argument leading to \eqref{eq794} for \(p_1, \dots, p_{k_1}\), we finally obtain that \(\langle a, \overline{S_0}(\widetilde{f}) \rangle \in L^{p_{k_1+1}}\). This gives \(\widetilde{f}\in L^{s}\), and completes the proof in the case \(m=0\). We now assume that the property is true for some \(m\geq 0\) and let us prove it for \(m+1\). Let $$ a\in W^{m+1,r}(\Omega ; \R^n)\quad\text{ with }\quad r>\frac{n}{m+2}. $$ Let \(1<p\leq s\leq r\) and \(f\in L^{p}(\Omega;\R^n)\) such that \[ \widetilde{f} + \langle a , \overline{S_0}(\widetilde{f}) \rangle \in W^{m+1,s} \] When \(r<n\), we set \(\overline{r}=r^*\) and observe that \(r>\frac{n}{m+2}\) implies \(\overline{r}>\frac{n}{m+1}\). When \(r\geq n\), we take for \(\overline{r}\) any number \(>\frac{n}{m+1}\). In any case, by the Sobolev embeddings, \(a\in W^{m,\overline{r}}\). Let \[ \overline{s}:=\begin{cases} s^* & \textrm{ if } s<n,\\ \overline{r} & \textrm{ if } s\geq n. \end{cases} \] Then \(\widetilde{f} + \langle a , \overline{S_0}(\widetilde{f}) \rangle \in W^{m,\overline{s}}\). Since \(1<p\leq \overline{s}\leq \overline{r}\), the induction assumption yields \(\widetilde{f}\in W^{m, \overline{s}}\). Hence, \(\overline{S_0}(\widetilde{f}) \in W^{m+1, \overline{s}}\). When \(s\geq n\), we use that \(a\in W^{m+1, s}\) and rely on Corollary \ref{corollary:Sob multiply} (i) below (applied with $s$ instead of \(p\), $q=\overline{s}$ and with \(m+1\) instead of \(m\)). This gives $$ \langle a, \overline{S_0}(\widetilde{f})\rangle\in W^{m+1,s}. $$ Hence, \(\widetilde{f}\in W^{m+1,s}\), which proves the induction assumption for \(m+1\). When \(s<n\), we use that \(\overline{S_0}(\widetilde{f}) \in W^{m+1, s^*}\) and apply Corollary \ref{corollary:Sob multiply} (ii) (with $s$ instead of \(p\), \(q=r\) and \(m+1\) instead of \(m\)). We deduce that $$ \langle a, \overline{S_0}(\widetilde{f})\rangle\in W^{m+1,s} $$ and conclude as before. The proof is complete. \end{proof} The corresponding statement in the scale of H\"older spaces reads: \begin{lemma}\label{lm-univ-holder} Let \(m\geq 0\), \(\alpha \in (0,1)\) and \(a\in C^{m,\alpha}(\Omegabar)\). Assume that \(\Omega\) is of class \(C^{m+2, \alpha}\). Then for every \(1< p <\infty\) and for every \(f\in L^{p}(\Omega)\), \[ (I+K)(f)\in C^{m,\alpha}(\Omegabar) \Longrightarrow f\in C^{m,\alpha}(\Omegabar). \] \end{lemma} \begin{proof} Let \(f\in L^{p}\), \(1<p<\infty\), such that for some \(m\geq 0\) and \(\alpha\in (0,1)\), \begin{equation} \label{eq:bootstrap in Hoelder} \widetilde{f} + \langle a , \overline{S_0}(\widetilde{f}) \rangle \in C^{m,\alpha}, \end{equation} where as usual $\widetilde{f}=f-\frac{1}{|\Omega|}\int_{\Omega}f$. \smallskip \textit{Step 1.} We shall first show that $\widetilde{f}\in C^{0,\alpha}.$ We can assume that $p<n,$ because if the lemma holds in this case, then it certainly also holds for $p\geq n.$ Using that \(C^{m,\alpha}\subset L^{\infty}\), \( \overline{S_0}(\widetilde{f})\in W^{1,p}\subset L^{p_*}\) and \(a\in L^{\infty}\), we deduce that \begin{equation}\label{eq847} \tilde{f}\in L^{p^*},\quad\text{ with }\quad p^*=\frac{np}{n-p}. \end{equation} By a similar argument as in Case 3 in the proof of Lemma \ref{lm_inva_p}, there exists \(k_0\in \N\) such that the sequence \((p_k)_{k\in \N}\) defined by \[ p_0=p, \qquad \qquad \forall k\geq 1, \qquad \frac{1}{p_k}=\frac{1}{p_{k-1}}-\frac{1}{n}=\frac{1}{p}-\frac{k}{n} \] satisfies \(1<p_k<n\) for \(k\leq k_0\) and \(p_{k_0+1}\geq n\). By bootstrapping the argument leading to \eqref{eq847}, we obtain that $\widetilde{f}\in L^{p_{k_0+1}},$ Thus we get $ \langle a, \overline{S_0}(\widetilde{f})\rangle\in L^{p_{k_0+1}}$ which implies that $\widetilde{f}\in L^{p_{k_0+1}}$. Hence, $ \overline{S_0}(\widetilde{f})\in W^{1,p_{k_0+1}}\subset W^{1,n}$ which in turn implies by \eqref{eq:bootstrap in Hoelder} that $\tilde{f}\in L^r$ for any $r<\infty.$ It follows that \( \overline{S_0}(\widetilde{f})\) belongs to \(\cap_{1<r<\infty}W^{1,r}\). By the Morrey embedding, we deduce therefrom that \( \overline{S_0}(\widetilde{f})\) belongs to \(\cap_{0<\beta<1}C^{0,\beta}\). Since \(a\in C^{0,\alpha}\), this implies that \(\langle a, \overline{S_0}(\widetilde{f})\rangle\) belongs to \(C^{0,\alpha}\). Therefore, by \eqref{eq:bootstrap in Hoelder}, \(\widetilde{f}\in C^{0,\alpha}\). \smallskip \textit{Step 2.} If $m=0$ we are done by Step 1. Otherwise, for every \(0\leq k \leq m-1\), the operator \(u\mapsto \langle a , u \rangle\) maps continuously \(C^{k+1,\alpha}\) into \(C^{k+1,\alpha}\), see \cite[Theorem 16.28]{Csato-Dacorogna-Kneuss}. Together with the facts that \( \overline{S_0} (C^{k,\alpha}) \subset C^{k+1,\alpha}\) and \(C^{m,\alpha} \subset C^{k+1,\alpha}\), it follows by induction on \(k=0, \dots, m-1\) that \(\widetilde{f}\in C^{m,\alpha}(\Omegabar)\). \end{proof} \smallskip \begin{proof}[Proof of Theorem \ref{th-universal-property}] In the scale of Sobolev spaces, we apply Lemma \ref{lm_univ_prop} with $\overline{S_a}$ instead of \(S_a\), to the sets \(E=W^{m+1,p}_z(\Omega;\R^n)\) and \(F=W^{m,p}(\Omega)\). In view of Lemma \ref{lm-univ-sobolev}, the assumption \eqref{eq_univ_prop} is satisfied. Theorem \ref{theorem:a is zero higher regularity} implies that \(\overline{S} : f\mapsto \overline{S_0}\left(f-\frac{1}{|\Omega|}\int_{\Omega}f\right)\) maps continuously \(W^{m,p}\) into \(W^{m+1,p}_{z}\). The conclusion follows in that case. In the scale of H\"older spaces, we rely on Lemma \ref{lm_univ_prop} with \(E=C^{m+1,\alpha}_z(\Omegabar;\R^n)\) and \(F=C^{m,\alpha}(\Omegabar)\). Here, assumption \eqref{eq_univ_prop} follows from Lemma \ref{lm-univ-holder}. Theorem \ref{theorem:a is zero higher regularity} now implies that \(\overline{S}\) maps continuously \(C^{m,\alpha}\) into \(C^{m+1,\alpha}_{z}\). The proof is complete. \end{proof} \section{Non-existence results in $L^1$ and $L^{\infty}$} First we discuss the existence in $L^1,$ which is easier. These results have nothing to do with the boundary conditions, and the proofs are almost identical to the case $a=0,$ see for instance \cite{Bourgain-Brezis}. \begin{theorem} \label{theorem:non reg in L1} Let $\Omega\subset\re^n$ be a bounded Lipschitz set. Let $a\in L^{\infty}(\Omega;\re^n).$ Then there exists $f\in L^1(\Omega)$ such that there is no $u\in W^{1,1}(\Omega;\re^n)$ with $\diw u+\langle a,u\rangle =f.$ \end{theorem} \begin{proof} Suppose by contradiction that for all $f\in L^1$ there exists a $u\in W^{1,1}$ such that $$ T_a(u)=\diw u+\langle a;u\rangle=f. $$ Hence the map $T_a:W^{1,1}\to L^1$ is onto. Moreover it is bounded and linear. Hence by the open mapping theorem there is a constant $C>0,$ such that for every $f\in L^1$ there exists a $u$ with $T_a(u)=f$ and $$ \|u\|_{W^{1,1}}\leq C_1 \|f\|_{L^1}. $$ We will show that for all $\varphi\in W_0^{1,n}$ and all $f\in L^1$ we have \begin{equation} \label{eq:proof:first contradiction f and u L1} \left|\intomega \varphi f\right|\leq C(\Omega,\|a\|_{L^{\infty}},C_1) \|\varphi\|_{W^{1,n}}\|f\|_{L^1}. \end{equation} Indeed we have that \begin{align*} \left|\intomega \varphi f\right|=&\left|\intomega \varphi(\diw u+\langle a;u\rangle)\right|=\left|-\intomega (\nabla \varphi)\, u+\intomega \varphi\langle a;u\rangle\right| \smallskip \\ \leq& \|\nabla \varphi\|_{L^n}\|u\|_{L^{\frac{n}{n-1}}}+\|a\|_{L^{\infty}}\|\varphi\|_{L^n}\|u\|_{L^{\frac{n}{n-1}}}. \end{align*} We now use the continuous embedding $W^{1,1}\subset L^{\frac{n}{n-1}}$ and obtain that \begin{align*} \left|\intomega \varphi f\right|\leq C \|\varphi\|_{W^{1,n}}\|u\|_{W^{1,1}}\leq C \|\varphi\|_{W^{1,n}}\|f\|_{L^1}. \end{align*} This proves \eqref{eq:proof:first contradiction f and u L1}. We are now able to conclude. Consider the map $M_{\varphi}:L^1\to \re$, for every fixed $\varphi\in W^{1,n}_0$, given by $M_{\varphi}(f)=\int \varphi f.$ Then by \eqref{eq:proof:first contradiction f and u L1} we have that $$ M_\varphi\in (L^1)^{\ast}=L^{\infty} \quad \textrm{ and } \quad \|\varphi\|_{L^{\infty}}\leq C\|\varphi\|_{W^{1,n}}. $$ This is a contradiction to the non embedding \(W^{1,n}\not\subset L^{\infty}\). \end{proof} \smallskip We now deal with the $L^{\infty}$ non-existence. \begin{theorem} Let $\Omega\subset\re^n$ be a bounded open set. Let $a\in W^{1,\infty}(\Omega;\re^n).$ Then there exists $f\in L^{\infty}(\Omega)$ such that there is no $u\in W^{1,\infty}(\Omega)$ with $\diw u+\langle a;u\rangle =f.$ \end{theorem} \begin{proof} For simplicity we present the proof for $n=2.$ We assume by contradiction that for every $f\in L^{\infty}$ there exists $u$ satisfying (as in the proof of Theorem \ref{theorem:non reg in L1}) $$ T_a(u)=\diw u+\langle a;u\rangle =f,\quad \|u\|_{W^{1,\infty}}\leq C\|f\|_{L^{\infty}}. $$ Let $\psi\in C_c^{\infty}(\Omega)$ and define $f\in L^{\infty}(\Omega)$ as ($\psi_{x_i},\psi_{x_i x_j}$ denote the partial derivatives of \(\psi\)) $$ f=\operatorname{sign}\psi_{x_1x_2}. $$ Let $u$ be a solution of $T_a(u)=f$ with $\|u\|_{W^{1,\infty}}\leq C\|f\|_{L^{\infty}}=C.$ We therefore obtain that \begin{align*} \|\psi_{x_1x_2}\|_{L^1}=&\left|\int\psi_{x_1x_2}f\right|=\left|\int\psi_{x_1x_2}\left(\diw u+\langle a,u\rangle\right) \right| \smallskip \\ =&\left|\int\left(\psi_{x_1x_1}\frac{\partial u_1}{\partial x_2}+\psi_{x_2x_2}\frac{\partial u_2}{\partial x_1}-\psi_{x_1}\frac{\partial }{\partial x_2}\langle a,u\rangle\right)\right| \smallskip \\ \leq & \left(\|\psi_{x_1x_1}\|_{L^1}+\|\psi_{x_2x_2}\|_{L^1}\right)\|u\|_{W^{1,\infty}}+\|\psi_{x_1}\|_{L^1}\|a\|_{W^{1,\infty}}\|u\|_{W^{1,\infty}}. \end{align*} We now use that $\|\psi_{x_1}\|_{L^1}\leq C \|\psi_{x_1x_1}\|_{L^1}$ for some constant $C>0$ depending only on $\Omega.$ This gives $$ \|\psi_{x_1x_2}\|_{L^1}\leq C(1+\|a\|_{W^{1,\infty}})\left(\|\psi_{x_1x_1}\|_{L^1}+\|\psi_{x_2x_2}\|_{L^1}\right), $$ which is a contradiction to the non-inequality of Ornstein \cite{Ornstein}. \end{proof} \section{Appendix} \subsection{Multiplication in Sobolev spaces} \begin{lemma}\label{lemma_multiplier} Let \(1\leq p\leq q \leq \infty\) and let $l\in \N$ be an integer such that $n-pl>0$ and let \(m\in \N^*\). Define $$ p(l)=\frac{np}{n-pl} $$ We assume that \begin{equation} \label{eq:assumption on pml} q> \frac{n}{m+l}. \end{equation} Then for every \(f\in W^{m,p(l)}(\Omega)\) and \(g\in W^{m,q}(\Omega)\), the function \(f g \) belongs to \(W^{m,p}(\Omega)\) and \[ \|f g\|_{W^{m,p}} \leq C \|f\|_{W^{m,p(l)}} \|g\|_{W^{m,q}} \] for some constant \(C>0\) which depends only on \(n,m,p,q\) and \(l\). \end{lemma} We will only use the cases $l=0$ and $l=1,$ which we emphasize as a Corollary. \begin{corollary} \label{corollary:Sob multiply} Let \(1\leq p\leq q \leq \infty\) and \(m\in \N^*\). (i) We assume that \(q> \frac{n}{m}\). Then for every \(f\in W^{m,p}(\Omega)\) and \(g\in W^{m,q}(\Omega)\), the function \(f g \) belongs to \(W^{m,p}(\Omega)\) and \[ \|f g\|_{W^{m,p}} \leq C \|f\|_{W^{m,p}} \|g\|_{W^{m,q}}. \] (ii) We assume that \(q> \frac{n}{m+1}\), $p<n$ and $p^*$ denotes the Sobolev conjugate of $p.$ Then for every \(f\in W^{m,p^*}(\Omega)\) and \(g\in W^{m,q}(\Omega)\), the function \(f g \) belongs to \(W^{m,p}(\Omega)\) and \[ \|f g\|_{W^{m,p}} \leq C \|f\|_{W^{m,p^*}} \|g\|_{W^{m,q}}. \] \end{corollary} \begin{proof}[Proof of Lemma \ref{lemma_multiplier}.] Let \(f\in W^{m,p(l)}\) and \(g\in W^{m,q}\). By the Sobolev and the Morrey embeddings, for every \(k=0, \dots, m\), \(W^{k,p(l)} \subset L^{p_{k}(l)}\) with \begin{equation} \label{eq:qk} \frac{1}{p_k(l)}:=\begin{cases} \frac{1}{p(l)} - \frac{k}{n}=\frac{1}{p} - \frac{k+l}{n} & \textrm{ if } k<\frac{n}{p(l)}\quad\Leftrightarrow\quad k+l<\frac{n}{p},\\ \text{as small as we wish} & \textrm{ if } k=\frac{n}{p(l)},\\ 0 & \textrm{ if } k>\frac{n}{p(l)}, \end{cases} \end{equation} where the third case has to be understood as $p_k(l)=\infty.$ The same holds true for \(W^{k,q}\subset L^{q_k}(\Omega)\) with \begin{equation} \label{eq:pk} \frac{1}{q_k}:=\begin{cases} \frac{1}{q} - \frac{k}{n} & \textrm{ if } k<\frac{n}{q},\\ \text{as small as we wish}& \textrm{ if } k=\frac{n}{q},\\ 0 & \textrm{ if } k>\frac{n}{q}. \end{cases} \end{equation} By considering each case successively, one can check that for every \(k=0,\dots,m\), \begin{equation}\label{eq-1056} \frac{1}{p_k(l)} + \frac{1}{q_{m-k}} \leq \frac{1}{p}. \end{equation} Indeed, if we are in the first cases of both \eqref{eq:qk} and \eqref{eq:pk}, then using that \(q\geq \frac{n}{m+l}\), one gets \begin{align*} \frac{1}{p_k(l)}+\frac{1}{q_{m-k}}=\frac{1}{p}+\frac{1}{q}-\frac{m+l}{n}\leq \frac{1}{p}. \end{align*} Assume next that \(k=\frac{n}{p(l)}\) and \(m-k<\frac{n}{q}\). Then \eqref{eq-1056} is satisfied provided that \(\frac{1}{q_{m-k}}<\frac{1}{p}\), or equivalently, \begin{equation}\label{eq-1064} \frac{1}{q}-\frac{m-k}{n}<\frac{1}{p}. \end{equation} This holds true when \(q>p\). When \(q=p\), we have \[ k=\frac{n}{p(l)}=\frac{n}{p}-l = \frac{n}{q}-l. \] Since by assumption \(q>\frac{n}{m+l}\), one has \(k<m\), which implies \eqref{eq-1064}. In the case \(k<\frac{n}{p(l)}\) and \(m-k=\frac{n}{q}\), the proof is essentially the same (when \(l=0\), we use that \(q>\frac{n}{m}\) which implies that \(k>0\)). The remaining cases are obvious and we omit them. This completes the proof of \eqref{eq-1056}. One also verifies that $$ \quad \frac{1}{p_m(l)} + \frac{1}{q_m}\leq \frac{1}{p}. $$ In particular, since \(f\in L^{p_m(l)}\), \(g\in L^{q_m}\), the H\"older inequality implies \(fg\in L^{p}\). We now prove that \(D^m(fg)\in L^{p}\), where \(D^m\) denotes the tensor of all partial derivatives of order \(m\). By the Leibniz rule, \begin{equation*} \|D^m(fg)\|_{L^{p}}\leq C \sum_{k=0}^{m}\||D^{m-k}f||D^k g|\|_{L^{p}}. \end{equation*} Since \(D^{m-k}f\in W^{k,p(l)}\subset L^{p_k(l)}\) and \(D^k g \in W^{m-k,q}\subset L^{q_{m-k}}\), the H\"older inequality and \eqref{eq-1056} imply that each term of the above sum belongs to \(L^{p}\) and \[ \|D^m(fg)\|_{L^{p}}\leq C' \sum_{k=0}^{m}\|D^{m-k}f\|_{L^{p_k(l)}}\|D^k g\|_{L^{q_{m-k}}}\leq C''\|f\|_{W^{m,p(l)}}\|g\|_{W^{m,q}}. \] This completes the proof. \end{proof} \subsection{A density lemma} \begin{lemma}\label{lemma_complementing_density} Let \(X\) be a Banach space, \(F\) a closed subspace of \(X\) of codimension \(k\geq 1\) and \(D\) a dense subset of \(X\). Then there exists a complemented subspace \(G\) for \(F\) such that \(G\) has a basis contained in \(D\). \end{lemma} \begin{proof} Let \(\widetilde{G}\) be a complemented subspace of \(F\) and \((e_1, \dots, e_k)\) be a basis of \(\widetilde{G}\). We consider an approximation \((e_{1}^{n},\dots, e_{k}^{n})_{n\in\N}\subset D^{k}\) of this basis: \begin{equation}\label{eq413} \forall i\in \{1, \dots, k\}, \qquad \lim_{n\to +\infty}e_{i}^{n}=e_i \end{equation} and we define \[ G^n:=\textrm{Vect}\,(e_{1}^{n},\dots, e_{k}^{n}). \] We claim that there exists an \(n_0\in \N\) such that (actually this is true for all $n$ sufficiently large), \begin{equation}\label{eq:lambda zero} \forall \lambda=(\lambda_1, \dots, \lambda_k) \in \R^k,\qquad \sum_{j=1}^k\lambda_j e_{j}^{n_0} \in F \Longrightarrow \lambda=0. \end{equation} Indeed, assume by contradiction that for every \(n\in \N\), there exists a \(\lambda^{n}=(\lambda_{1}^{n}, \dots, \lambda_{k}^{n})\in \R^k \setminus \{0\}\) such that \[ \sum_{j=1}^{k}\lambda_{j}^{n}e_{j}^{n}\in F. \] We can assume that \(\|\lambda^{n}\|=1\) for some fixed norm \(\|\cdot\|\) in \(\R^k\). Hence, up to a subsequence (we do not relabel), \((\lambda^{n})_{n\in \N}\) converges to some \(\lambda \in \R^k\) such that \(\|\lambda\|=1\). Then \[ \lim_{n\to +\infty}\sum_{j=1}^{k}\lambda_{j}^{n}e_{j}^{n} = \sum_{j=1}^{k}\lambda_{j}e_{j}. \] Since \(F\) is closed, this implies that \(\sum_{j=1}^{k}\lambda_{j}e_{j}\in F\). Using that \(\widetilde{G}\cap F =\{0\}\), we deduce that \(\sum_{j=1}^{k}\lambda_{j}e_{j}=0\), which implies that \(\lambda=0\) in view of the fact that \((e_j)_{1\leq j \leq k}\) is a basis of $\widetilde{G}.$ This contradicts the fact that \(\|\lambda\|=1\) and the claim \eqref{eq:lambda zero} is proved. \smallskip Set $G=G^{n_0}$ which satisfies by \eqref{eq:lambda zero}, $G\cap F=\{0\}$. Since \(F\) has codimension \(k\), this proves that \(X=F\oplus G\), as desired. \end{proof} \smallskip \noindent\textbf{Acknowledgements} The second author was supported by Chilean Fondecyt Iniciaci\'on grant nr. 11150017 and is member of the Barcelona Graduate School of Mathematics. He also acknowledges financial support from the Spanish Ministry of Economy and Competitiveness, through the ``Mar\'ia de Maeztu'' Programme for Units of Excellence in RD (MDM-2014-0445).

Microgravity condition leads to endothelium dysfunction which in turn causes the individuals to experience cardiovascular deconditioning as well as physiologic changes[@b1]. Gravitational alterations influence endothelial cell (EC) proliferation, differentiation, signaling, gene expression, surface adhesion molecules, extracellular matrix proteins expression, and cause significant changes in cytoskeletal polymers[@b2][@b3][@b4][@b5][@b6][@b7][@b8][@b9][@b10]. Endothelium plays a crucial role in local blood flow, regulation of coagulation, permeability, leukocyte adhesion, and vascular smooth muscle cell growth. The structural and functional integrity of the ECs are vital to regulate vascular homeostasis and prevent atherosclerosis[@b11][@b12]. Chemical regulators, extracellular matrix, and environmental conditions adjust the cells' functions. Mechanical loadings such as shear stresses resulted from blood flow, hydrostatic pressure, and cyclic stretch are some examples of environmental stimuli able to alter the functionality and mechanical properties of ECs[@b13][@b14]. In addition to regulatory effects of such stresses, in critical conditions, the extreme values may initiate pathologic problems such as atherosclerosis and intima hyperplasia[@b15]. DNA microarray analysis revealed that about 600 genes (3% of all genes) of ECs responded to shear stress[@b16]. On the other hand, altered mechanical properties of cells may contribute to cell remodeling and regulation of stresses within the cell body[@b17]. Actin filaments align to the direction of shear stress, which consequently lead the cells spindle-shaped. Cells experiencing turbulent flow found a rounder shape with non-uniform orientations[@b18][@b19]. It has been demonstrated that mechanical unloading (MU) remodeled actin cytoskeleton polymer, reduced the total amount of this biopolymer through transcriptional mechanism[@b4][@b20], and had a significant effect on the arrangement and dynamics of microtubules[@b21]. The detailed mechanisms involved in conversion of MU to intracellular biochemical reactions are still elusive. Therefore, *in vitro* studying of the endothelial behavior in response to microgravity conditions in term of mechanical properties assists in understanding of endothelium alterations, mechanisms involved in vascular problems, and pathogenesis of related diseases in space mission. Several studies have been done to evaluate the mechanical properties of cells exposed to a variety of mechanical stresses by using either quantitative and qualitative techniques[@b22]. Micropipette aspiration (MA) is the most feasible method, which enables measurement of the whole cell mechanical properties. It eliminates disadvantageous effects of cell-matrix interaction and more importantly the stiffness of substrate[@b23][@b24]. The goal of this study was to investigate the impact of s-μg on viscoelastic parameters of ECs and the content of main cytoskeleton polymers. A developed RPM was utilized to simulate weightless conditions by means of continuous random change of orientation, relative to the gravity vector ([Fig. 1](#f1){ref-type="fig"})[@b25]. Concerning difficulties with space experiments, RPM is an appropriate candidate for ground-based evaluations with a reliable performance comparing to real microgravity[@b26][@b27][@b28]. In this report, primary ECs were placed under s-μg conditions by which viscoelastic properties, proliferation, morphology, and cytoskeleton content were investigated. To the best of our knowledge, this work is the very first study on the effects of microgravity on the cell mechanical properties and the linkage with the main cytoskeletal polymers contents. Results ======= Mechanical properties of ECs ---------------------------- All ECs had typical viscoelastic solid creep behavior responding to a 600 Pa step pull-in pressure. Deformation (*L*) and time (*t*) had a nonlinear regression with a mean of R^2^ = 0.9781 ± 0.017 for all cells. The average diameters of EC cells were measured in the range of 12--16.8 μm. In [Fig. 2](#f2){ref-type="fig"} a real image of MA experiment for an endothelial cell after 72 hrs of being subjected to unloading is illustrated. Estimated viscoelastic parameters are shown in [Fig. 3](#f3){ref-type="fig"}. There was a considerable reduction in the measured elasticity and the viscosity for cells under microgravity. The results indicated statistically significant differences in *K*~*1*~ within 24 hrs and later (*p* \< *0.01*). The mean Young's modulus of elasticity () for normal ECs and the cells under microgravity at 12, 24, 48, and 72 hrs were obtained to be 129.62 ± 19, 122.88 ± 12, 101.46 ± 13, 62.15 ± 9, and 32.21 ± 6 Pa, respectively ([Fig. 4](#f4){ref-type="fig"}). A similar declined trend was observed for the viscosity (*μ*) with no statistically significant drop at 12 hrs (*p* \> *0.22*). The time constant (*τ*) reduced at the selected time points; however, the reduction was statistically meaningful at 48 hrs and 72 hrs (*p* \< *0.04*). The instantaneous Young's modulus () decreased significantly (*p* \< *0.001*), where the values at 24 hrs and 72 hrs were approximately 30% and 60% of the control cells, respectively. No significant correlations were observed between the diameters and the viscoelastic parameters of cells (*p* \> *0.08*). Circularity ----------- Circularity was defined as the ratio of width to length of each cell. Endothelial morphology changed from a normal spindle shape ([Fig. 5a](#f5){ref-type="fig"}) to a cobblestone phenotype ([Fig. 5b](#f5){ref-type="fig"}). It is obvious that the value increased under s-μg. After 24 hrs of being unloaded, circularity increased about 3.3 times in comparison with control cells ([Fig. 5c](#f5){ref-type="fig"}). The average surface area of ECs was also measured using ImageJ which showed an increase of 43% under s-μg. 4 different petri dishes with minimum cell number of 200 cells were evaluated for circularity measurement (*p* \< *0.05*). Cell proliferation assay ------------------------ The MTT (3-(4, 5-dimethylthiazol-2-yl) 2, 5-diphenyl tetrazolium bromide) assay was used to evaluate the effect of microgravity on ECs viability at 24 hrs. Results confirmed slight changes in cell populations ([Fig. 5d](#f5){ref-type="fig"}). Being exposed under s-μg caused an average drop of 7% in cells' viability. The reduction for cells under s-μg is not statistically insignificant (n = 6, p \> 0.05) comparing to control cells. Cytoskeleton Content and staining --------------------------------- Western Blot (WB) analysis was performed to estimate the content of F-actins and MTs in ECs. The results helped to quantify the alterations of two main cytoskeletal biopolymers. [Figure 6](#f6){ref-type="fig"} shows the densitometric quantization of the relative band intensities for cells under 24 hrs of being subjected to microgravity and control cells. It was clear that the expression of both cytoskeleton filaments is reduced. The content of F-actins protein reduced around 65% and the reduction for β-tubulin expression was 26% under the microgravity simulation. Similarly, fluorescent staining revealed a considerable disruption in structural filamentous polymers. [Figure 7](#f7){ref-type="fig"} shows the MTs and the actin filaments of the cells for control and s-μg condition. The images show that the more cell being under simulated microgravity, the more the cytoskeleton disrupts. The actin rim underneath the plasma membrane was no more continuous. Microtubules network obviously disorganized and relatively concentrated around the nucleus. Discussion ========== Mechanical properties of EC provide the cell ability to resist widespread hemodynamic forces and at the same time, to respond differentially to variations in different surrounding stresses[@b29]. ECs constantly experience shear stress, cyclic stretch, and hydrostatic pressure. The developed mechanical stresses within the cell body have shown an evident role in the endothelium function. Furthermore, altered mechanical properties of cells by microenvironmental loadings have influenced cell remodeling and regulation of the stresses[@b17]. Vascular endothelium dysfunction undoubtedly influences vascular smooth muscle proliferation, angiogenesis, and wound healing. It is generally believed that the endothelial mechanical properties, gravisensation, and gravitropism are mostly defined by the cytoskeleton. The adaptation of cell mechanical tension to any changes in external mechanical stresses depends on the cell mechanical characteristics and the sensitivity of mechanosensors[@b30]. Similar to other eukaryotic cells, the endothelial cytoskeleton strongly depend on three major entangled filamentous biopolymers: actin filaments, microtubules, and to a less extent intermediate filaments. This highly dynamic network defines cell shape, withstand external forces, and effectively respond to mechanical stimuli[@b29]. Actin filaments, besides structural responsibility, play key functions in cell morphology, migration, phagocytosis, vesicular movement, cytokinesis, and molecular transport between the plasma membrane and the nucleus[@b31][@b32]. Cell locomotion is mediated by actin networks in cell cortex. While F-actins resist against deformations, they fluidize under high shear stresses. They respond to external forces and they have substantial effect in formation of leading-edge protrusions during cell motility[@b33]. Therefore, any disorganization of actin filaments may interfere with the transmission of forces across the cell membrane[@b34]. Microtubules are essential skeletal elements in cell division, polarization, and migration. Microtubules as the second major constituent of the cytoskeleton act in concert with the other filament biopolymers to stabilize the cell structure under compression loadings. According to the tensegrity theory, MTs as load-bearing compression elements and actin filaments as tension elements support integrity of cells[@b35]. ECs showed sensitivity to short term microgravity (22 s) in rearrangement of their β-tubulin[@b5]. Moreover, being subjected to s-μg more than 96 hrs caused a decrease in the total amount of actin as well as disorganization of actin cytoskeleton which could lead to apoptosis. In one word, the responses of cells to microgravity are cell-type and exposure-time dependent[@b4]. In this study, we demonstrated that microgravity could change mechanical properties, content of the main cytoskeleton polymers, and morphology of ECs. Previous studies confirmed viscoelastic behavior of ECs[@b14][@b36]. While deformation of cell body is mostly determined by elasticity, viscoelasticity characterizes time dependent alteration in deformation. Viscoelastic parameters of ECs were evaluated quantitatively by MA technique over 72 hrs. Young's modulus and viscosity of the ECs decreased during microgravity experiments. As described by results of the current study, microgravity caused a significant increase in deformability of the cultured ECs. This effect was intensified by s-μg exposure time. Although mechanical parameters of ECs did not experience statistically meaningful reduction at 12 hrs, *K*~*1*~ and *μ* decreased substantially about 20% (p \< 0.01) and 40% (p \< 0.001) at 24 hrs and it continued reduction until 72 hrs. As the viscoelastic properties of the cells showed significant alterations at 24 hrs, this time point was selected to compare the cytoskeleton reorganization with control cells. Accordingly, MTT assay, western blotting were performed at that time point. There are a few studies that addressed a short period viability for ECs under microgravity and any possible consequences on the cell structure[@b4][@b11]. MTT assay was performed to make sure the least side effect of s-μg on cell survival at the time of cytoskeleton evaluation. Viability assay depicted a trivial effect of microgravity on ECs proliferation at 24 hrs. Subsequently, WB analysis and fluorescence staining were done at this time point. Although trypsin used for cell detachment may affect the cytoskeleton, the observed changes in mechanical properties are mainly generated by s-μg. All mechanically investigated cells were treated in similar conditions with similar detaching approach and MA. Trypsin's concentration, volume, exposing time were maintained similar in all experiments, thereby the differences in mechanical properties were mostly associated with the microgravity exposure time. Disruption of stress and tensile fibers and orientation of filamentous bundles can be influential in both elastic and viscous parameters of cellular mechanical model. Although polymerization and depolymerization of the main cytoskeletal fibers are much faster comparing to continuous random change of orientation by RPM, it is shown in line with other reports that the total amount of actin filaments and microtubules decreased[@b8][@b19][@b37][@b38][@b39][@b40]. Western blot analysis, which was performed very immediately after simulated microgravity, confirmed a considerable reduction in content of both actin filaments and microtubules. Obviously, actin filaments were more susceptible to unloading by experiencing a dramatic drop of 65% comparing to 26% reduction in microtubules. It is believed that a drop in actin filaments is an adaptive mechanism to avoid the accumulation of redundant actin fibers[@b5]. This phenomena is likely due to transcriptional regulation of the cytoskeletal polymers[@b4][@b38][@b41], where the down-regulation can lead to a lasting reduction in the synthesis of the cytoskeletal polymers. Cytoskeleton staining revealed disorganization of both actin filaments and microtubules. The more cells being under microgravity the more the cytoskeleton disrupts. Cell staining requires fixation using materials such as paraformaldehyde which helps to preserve cellular architecture and composition of cells. As the staining procedure started vey immediately after each s-μg experiment, the visualized cytoskeleton of the cells obviously proved the disorganization of actin filaments and microtubules caused by simulated microgravity. The amount of actin filaments particularly around the membrane was reduced but the reduction for microtubules was less considerable. In fact, the altered mechanical characteristics of ECs can be mostly attributed to the rearrangement and reduction in the content of cytoskeletal actin filaments. There was some time interval between sample dismounting and cells' preparation for MA which might affect cytoskeleton structure however, it could not have a substantial contribution in alteration of mechanical properties among samples. The evaluated ECs were treated carefully in a same manner therefore, the differences between viscoelastic properties of the cells are stemmed from the effects of s-μg. More importantly, the values emphasized in this research are the differences in mechanical properties rather than absolute values. This means that even if there are some effects of delays or trypsin on cytoskeletal properties, we are still able to observe the effect of simulated microgravity on mechanical properties and cytoskeleton of cells, although this effect may not be linearly associated. Shear stresses are transmitted to the cortical actin rim immediately underneath the plasma membrane of EC. Actin polymers form a cortical rim in quiescent ECs which provides a link between extracellular events and intercellular organelles[@b42]. Densely polymerized array of actin filaments underneath of membrane has the major role to reinforce membrane against intense cortical deformation[@b43]. The rigidity of the cortex as a fundamental property is vital in regulation of cell shape, cytokinesis, proliferation, differentiation, and the response to a change in external environment. In addition, formation of protrusive organelles such as filopodia, lamellipodia, and microvilli is regulated by very local changes in membrane rigidity[@b44]. Disruption of actin bundles suggested evaluating mechanical behaviors of EC cortex. MA experiment begins with the investigation of cell cortex and its mechanical properties can be estimated by the instantaneous Young modulus (*E*~*0*~). This parameter decreased approximately 20% at 12 hrs and reached 67% of its initial value at 24 hrs. On the other hand, E lessened insignificantly 5% at 12 hrs and the reduction was approximately 20% at 24 hrs of being in s-μg condition. At 48 hrs, the reduction of both E~0~ and E were similar and around 50% of their initial values. Finally at 72 hrs, reduction of Young's modulus exceed instantaneous Young's modulus. Indeed, E and E~0~ decreased approximately 75% and 60%, respectively, at 72 hrs. In other words, cell membrane stiffness was more susceptible to disruption than whole cell body stiffness at early hours of experiencing microgravity condition, whereas long term s-μg had more influence on whole cell rigidity than cortical rigidity. It is evident that disruption of cortical actin boosts endothelial permeability and stops barrier function of the cells[@b12][@b45]. Fluorescence staining suggested that the mechanical alterations of the cortical actin rim occurred because of extensive disruption of actin filaments. This has a significant impact on endothelial dysfunction and vascular damage. Microgravity had effects on both cell shape and cytoskeleton. Image processing algorithms showed that ECs roundness increased under s-μg significantly. The length and width of the cells decreased in a manner that the circularity of the cells raised. This morphological alteration was a consequence of cytoskeletal reorganization at the subcellular level. The circularity parameter increased about 3 times after 24 hrs of s-μg, which is consistent with previous reports[@b8][@b46]. The imbalance of forces on a cell seemingly caused an extracellular matrix remodeling and consequently changes in morphology[@b46][@b47]. It is believed that cell rounding and, consequently, barrier integrity is mainly determined by the actin filaments organization. It was suggested that a reduction in cell-matrix attachment comparing to cytoskeletal arrangement and susceptibility of PKC-mediated signal transduction to microgravity caused the cells to be rounded in shape and not in spread[@b3][@b46][@b48]. In conclusions, it is clearly demonstrated that there is a close relationship between mechanical properties, the main cytoskeletal polymer contents, and endothelial cell shape. The findings of this paper indicate an adaptive response of ECs to microgravity conditions through an alteration in cytoskeleton structure. Microgravity disrupts the polymerization and depolymerization processes of the main cytoskeleton polymers and consequently causes improper function by reducing cytoskeleton-generated tensions. Although s-μg and its effect on mechanical properties of ECs have been studied, the related mechanotransduction should be fully understood by further studies. These changes certainly have a direct influence on signal transduction, synthesis and secretion of cytokines, and gene expression, which can ultimately lead to apoptosis. Hence, an exact evaluation of the proteome and utilizing the gene array techniques may expand knowledge about the involved signaling pathways in endothelial dysfunction and mechanism involved in vascular disorders in microgravity conditions. This may eventually help to find novel approaches to prevent vascular disorders. Possibility of creating 3D differentiated tissue assemblies has attracted several researchers to culture cells in microgravity conditions[@b49][@b50][@b51]. Accordingly, alterations of the mechanical features under s-μg need to be known. In addition, the results from s-μg conditions on mechanical behaviors of ECs should be compared with real microgravity experience in space. Methods ======= Cell culture ------------ Human umbilical vein endothelial cells (HUVECs) were obtained from a commercial source (ATCC CRL-1730). Cells were maintained at 37 °C (5% CO~2~, 95% air) in Ham's F12 + DMEM (1/1V), supplemented with 10% fetal bovine serum (Gibco), and 1% penicillin/streptomycin (Gibco). ECs were cultured on standard 35-mm diameter petri dishes at a density of 2 × 10^5^ cells per dish. The fresh medium was replaced every other day. The ECs were used up to passage 5. To detach cells, 0.25% Trypsin-EDTA solution (Invitrogen) was used. To investigate the effect of microgravity, each petri dish with subconfluent ECs was mounted on a specific holder to insure nonexistence of any bubbles which possibly exert undesired shear on the cells. Total of 36 experiments were performed. Each experiment contains two independent samples; the test sample was under s-μg whereas the control one was maintained in the same cell culture incubator. A detail statistical analysis was performed and P \< 0.05 was considered as a maximum to represent a statistically significant difference. ImageJ software was used to calculate the amount of circularity for the cells. Random Positioning Machine -------------------------- An RPM was developed to provide the s-μg by means of continuous random change of samples' orientation relative to the gravity vector[@b19][@b52]. The machine contains two perpendicular frames controlled by a programmable logic controller (PLC) using two independent servomotors. The speed and rotation direction of each frame are determined based on random walk scenario[@b53]. To lessen the effect of cell-cell contact on mechanical properties of cells, non-confluent dishes were chosen for MA analysis. The cavity between petri dish and holder were completely filled with cell culture medium. Thereafter, a 22 μm filter was inserted in the cap to enable air/medium exchange and to maintain 5% CO~2~. Finally, the samples were mounted as close as possible to the center of the RPM which was initially located within the incubator ([Fig. 1](#f1){ref-type="fig"}). The RPM functional parameters, random angular walk speed and the maximum distance of ECs to the center of rotation[@b52], were adjusted to be 0.3--0.7 rad/s and 2 mm respectively. Before suspending the cells, ECs beyond the maximum distance were mechanically detached by cell scraper. This approach guaranteed the evaluated ECs to experience 10^−4 ^g as the maximum gravitational acceleration and 5 × 10^−4^ as the maximum tangential acceleration. The detailed calculation of the developed microgravity together with the data obained from a 3D acceleration sensor embedded in the test plate confirmed that the samples experience a gravity of less than 10^−4 ^g. Cell proliferation assay ------------------------ MTT cell proliferation assay was used to evaluate the effect of microgravity on ECs proliferation. Briefly, after incubating for 24 hrs on normal conditions, one petri dish was mounted on the RPM stage utilizing the holder for another 24 hrs. Then, tetrazolium salt (M2128, Sigma) was added for an additional 4 hrs. Thereafter, formazan crystals were solubilized using Dimethyl sulfoxide (Sigma) while shaking for 20 min. Absorbance at 570 nm was measured by a standard spectrophotometer (CECIL, UK). The assay was performed six times for cells exposed to the s-μg condition for 24 hrs. Micropipette aspiration experiment ---------------------------------- MA is a reproducible technique to quantify mechanical properties of cells. Considering appropriate mechanical models and subsequent theories, MA has been broadly utilized to estimate the whole body mechanical behavior of various cell types such as cancerous cells[@b24][@b54], chondrocytes[@b55], fibroblast[@b56], and endothelial cells[@b14]. MA has also been used to investigate the alteration in skeletal and mechanical parameters of ECs as a consequence of external loadings[@b36]. Our developed MA arrangement for the present study is depicted in [Fig. 2](#f2){ref-type="fig"}. The following procedure was performed according to the previous works[@b36][@b57]. In brief, cells were suspended in the culture medium by 0.25% Trypsin-EDTA in less than 4 min to minimize the effect of trypsinization[@b58]. Furthermore, the measurement time for all samples was limited to less than 400 s to minimize the possible cytoskeletal reorganization after demounting from RPM machine. A movable water reservoir was implemented to apply the desired pressure at the end of the glass micropipette. Displacing the reservoir up or down leads to a suitable pressure to pull in or force away the cell, respectively. This is also applicable to find zero pressure. Actual pressure was related to the displacement of the reservoir provided by a precise motor. The step constant negative pressure imposed to each cell maintained to be between 500--600 Pa in all experiments. Micropipettes, with internal diameters within the range of 4.5--5 μm were coated with Sigmacote chemical agent (Sigma) to minimize the possible adhesion of the cells to the inner wall of the micropipettes[@b59]. The cell surface was aspirated into the pipette by exerting a controlled suction pressure on the cell body at room temperature around 20--22 °C. An inverted microscope equipped with a digital camera (Nikon Eclipse, DXM1200) provided monitoring of the leading edge of cell surface. Each micropipette was under control by a micromanipulator (Transfer Man Nk2, Eppendorf). At least three cells were investigated from each petri dish for MA study. Analysis of images was performed by Axiovision LE Software (Zeiss). Theoretical Model of Micropipette aspiration -------------------------------------------- Mechanical properties of cells in MA technique are extracted with a generalized Maxwell model. Cells are assumed to be a homogeneous, incompressible, and viscoelastic material subjected to a uniform axisymmetric aspiration pressure as depicted in [Fig. 2](#f2){ref-type="fig"}. The model includes three constants: a spring (k~1~) that provides the restoring force necessary to recover the initial shape after the release of the stress in parallel with serially arranging of a damper (μ) and a spring (k~2~)[@b36]. The relation between elastic and viscous parameters is described by Equ. 1 in which the boundary condition of no axial displacement of the cell is assumed at the micropipette end. where *∆P* is the applied pressure, *L*(*t*) is the aspirated length, *h*(*t*) is the unit step function, and *a* is considered as the inner radius of the micropipette. The time constant (*τ*) of viscoelastic model is calculated by Equ. 2: Given the following equations, viscoelastic parameters of cells are obtained by curve fitting of experimental data (*L/a*) with time using the least square method implemented in Matlab: Fluorescence labeling for microscopy ------------------------------------ Cells were seeded to proliferate for 24 hrs on sterile cover slips. The cover slips were then mounted on the simulating microgravity machine for another 24 hrs. Fluorescein isothiocyanate labeled Phalloidin (Sigma, P5282) was used to visualize actin filaments. The staining procedure started immediately after each s-μg experiment. After several times rinsing with phosphate buffered saline (PBS), the cells were fixed with 3.7% paraformaldehyde (dissolved in PBS buffer) for 5 min. Upon permeabilizing cells with 0.1% Triton-X100 in PBS and washing again in PBS, they were stained with a 50 mg/ml fluorescent phalloidin conjugate solution in PBS for 40 min at room temperature. Before staining for microtubules, samples were washed several times with PBS to remove unbound phalloidin conjugate. Detection and localization of microtubules were performed using monoclonal Anti-β-Tubulin-Cy3 (Sigma, C4585). Diluted antibody conjugate in PBS containing 1% BSA was added to cover cell layer and incubated for 60 min. Upon 3 times washing of the cells with PBS, 5 minutes each, samples were left to dry. Images were captured with an invert fluorescent microscope (Olympus, BX51 equipped with DP72 camera) and processed with ImageJ software. Western Blot Analysis --------------------- Western Blot analysis was performed immediately after simulated microgravity experiments to quantify the content of the major cytoskeleton polymers[@b60]. Whole cells were lysed by RIPA buffer added with 1x complete protease inhibitor cocktail tablet (Roche). Total protein concentration was determined by the Bradford method (Bio-Rad protein assay). After electrophoresis, proteins were transferred to Methanol pre-wetted polyvinylidene difluoride (PVDF) membranes (Millipore). Membranes were blocked with 2% ECL Advance blocking agent (Amersham) in PBS-Tween. Mouse monoclonal antibodies targeting total actin and β-tubulin (Sigma) were diluted in the blocking solution and incubated over night at 4 °C. Detection of bound antibodies was performed with peroxidase coupled secondary antibodies using the ECL Advance Western blotting detection system (Amersham). The bands obtained on X-ray films were quantified using the NIH ImageJ program. Expression of actin and β-tubulin content are presented in the average of three biological replicates. To ensure equal loading, ribosomal protein L19 or γ--tubulin (Sigma--Aldrich) was used. Statistical analysis -------------------- The analysis of variance (ANOVA) followed by Dunnet post-test was done to compare the results. The Dunnet post-test decreases considerably the number of comparisons and increases the power for detecting the differences. This method compares each test group with the control instead of involving nonrelated groups. All data are presented as mean ± standard error of the mean. All statistical analysis and the related graphs were performed in GraphPad Prism 6.0. (CA, USA). P \< 0.05 was considered as a maximum to represent a statistically significant difference. Additional Information ====================== **How to cite this article**: Janmaleki, M. *et al.* Impact of Simulated Microgravity on Cytoskeleton and Viscoelastic Properties of Endothelial Cell. *Sci. Rep.* **6**, 32418; doi: 10.1038/srep32418 (2016). The authors would like to thank Dr. H. Peirovi, Head of the Medical Nanotechnology and Tissue Engineering Research Centre for his invaluable advice and Mrs. N. Pachenari and Associate Professor M. Javan at Molecular Neurophysiology Laboratory of Department of Physiology School of Medicine, Tarbiat Modares University for their help in obtaining microscopy images and Mr. Morteza FataliBeigi for the support for simulating microgravity machine. Financially, this work was partially supported by Natural Sciences and Engineering Research Council of Canada. **Author Contributions** M.J., M.P., S.M.S., R.S. and A.S.N. were fully involved in the study and preparation of the manuscript. M.J., M.P. and S.M.S. performed the experiments. {#f1} {#f2} {#f3} {#f4} {#f5} {#f6} {#f7} [^1]: These authors contributed equally to this work.

\section{Introduction} \label{sec:intro} Soon after the discovery of the binary pulsar B1913+16 \citep{ht74} it was pointed out that if the pulsar's spin axis and orbital angular momentum vector were misaligned, they should precess around their common sum on a time scale of around 300 years due to General relativistic spin-orbit coupling \citep{dr74,hdr75,bo75,eh75}. This phenomenon is often referred to as ``geodetic precession''. Because radio pulsars are thought to possess lighthouse-like beams of emission that only beam to a small fraction of the sky, such precession should lead to observable pulse shape changes once a significant fraction of the precession period has passed. Secular changes in the mean pulse profile of PSR B1913+16 were first reported by \citet{wrt89}. The changes became more pronounced in later years, allowing limited modeling of the emission cone \citep{kra98,wt02}. The variations seen in PSR B1913+16 suggest a small misalignment angle between the spin and orbital angular momentum vectors, consistent with a natal kick being imparted to the most recently formed neutron star \citep{bai88}. Large misalignment angles raise the possibility that we might be able to, for the first time, map out the entire emission cone of a pulsar until it disappears completely from view \citep{kra98}. For many years, PSR B1913+16 was the only binary pulsar with the right combination of relativistic parameters and an evolutionary history that would make detection of geodetic precession possible. However, shortly after geodetic precession was first hinted at, new relativistic pulsars were uncovered in pulsar surveys that would become suitable for measurements in the future. \citet{agk+90} discovered a near clone of the original binary pulsar in the globular cluster M15C, which although weak, might one day exhibit the phenomenon. The much closer PSR B1534+12 \citep{wol90a} is a relativistic pulsar in a 10 hour binary orbit that has recently been shown to exhibit pulse shape changes. These changes have been combined with polarimetric models to make the first reliable estimate of the rate at which the spin axis of the pulsar is tilting away from us \citep{sta04}. Thus, measurements of geodetic precession are allowing new tests of General relativity and constraining evolutionary models \citep{wkk00,kws03}. These new tests complement the earlier pioneering work of authors such as \cite{tw82}, that examined other aspects of the theory. In principle, it should be possible to use the relativistic pulsars to map both shape and intensity changes across a pulsar emission cone. However, this is complicated by the fact that many of the pulsars have random variations in their total intensity because of refractive and diffractive interstellar scintillation. Pulsars with small dispersion measures (DM) often have their fluxes change by factors of several on time scales that vary between minutes and days \citep{ssh+00}. However, large dispersion measure pulsars ($DM > 100$ pc cm$^{-3}$) have relatively stable fluxes when observed at high frequencies ($\nu > $ 1.4 GHz) if long integration times are used and the observing system samples a large ($> 100$ MHz) bandwidth. The discovery of a relativistic pulsar at a large DM with a rapid geodetic precession time scale might then offer the first hope of determining how rapidly pulsar emission varies in both intensity and shape across the emission cone. This discovery would have important ramifications for pulsar statistics, where one of the great uncertainties is the pulsar ``beaming fraction'', often guessed at by observing the duty cycles of radio pulsars. Various arguments have been made in favour of both meridional compression \citep{big90b} and the elongation of pulsar beams \citep{nv83}, making the actual measurement very important. In the future it seems as though we will have a much larger number of relativistic systems with which to both test General relativity and map pulsar emission cones. Recent surveys have been finding relativistic pulsars at a welcome rate \citep{fsk+04} and the total pulsar population now exceeds 1500 objects. In particular, the Parkes multibeam surveys have discovered a large number of pulsars in relativistic orbits. The spectacular ``double pulsar'' is a 2.4 hour binary with two active pulsars, and a geodetic precession period of just 80 years \citep{lbk+04}. \cite{fkl+05} report the recent discovery of the 7.7-hour binary pulsar J1756--2251 with an eccentricity of 0.18 that almost certainly consists of two neutron stars. The first relativistic binary pulsar discovered by the multibeam surveys was however the 4.8 hour binary pulsar J1141--6545 \citep{klm+00}. This pulsar orbits what is most likely a white dwarf companion, but the system is unique in that it still possesses a significant orbital eccentricity ($e=0.17$). An eccentricity of this magnitude suggests that the system was put in its final configuration in an explosive event that may have given the pulsar a significant kick. Recent timing of the system is consistent with a 1.3 M$_\odot$ pulsar orbiting a 1.0 M$_\odot$ white dwarf companion \citep{bok03}. It is likely that the system originated as a binary containing two main sequence stars, both below the critical mass required for a supernova. The initially more massive star begins to transfer matter onto its companion as it evolves, causing it to exceed the critical mass. If the system remains bound after the resulting supernova, a young neutron star is left orbiting a white dwarf companion. In the case of symmetric supernovae, the eccentricity of the orbit is induced by the sudden mass loss, and allows us to determine the pre-supernova mass uniquely \citep{rs85}. This ejected mass can be related to the expected runaway velocity of the system, which in the case of PSR J1141--6545, is less than 50 km s$^{-1}$. Significant progress has been made in understanding the geometry and location of PSR J1141--6545 through a range of timing and spectroscopic studies. \citet{obv02a} demonstrated via HI absorption analysis that the pulsar is at least 3.7 kpc distant. In addition, PSR J1141--6545 was the first pulsar to exhibit dramatic changes in its scintillation time scale as a function of orbital phase, which have enabled an independent estimate of both its orientation and velocity. \citet{obv02} used the orbital modulation of the scintillation time scale to calibrate the usually unknown scale factor that relates the scintillation time scale, scattering screen distance and intrinsic motion to pulsar velocity. Their subsequently determined space motion was greater than that expected from a symmetric supernova \citep{obv02}. Several years ago we commenced an observing campaign of PSR J1141--6545 to study its HI, scintillation and timing properties over long baselines. It was clear that such observations could also be used to search for relativistic effects such as orbital decay and precession. The DM of the pulsar is 116 pc cm$^{-3}$, and its scintillation properties are well understood. \citet{obv02} showed that the scintillation bandwidth is much smaller than the 256 MHz observing band used by the analogue filter bank at the Parkes radio telescope, resulting in fairly stable fluxes when integrated over the 4.8 hr orbit. In this paper we demonstrate that PSR J1141--6545 is undergoing rapid secular evolution of both its total intensity and polarimetric profiles in a manner consistent with geodetic precession. In section \ref{sec:observations} we describe the many instrumental systems used to observe this pulsar since its discovery in 1999, along with the associated data reduction methods. Section \ref{sec:1141} describes the parameters of the binary system in greater detail and includes a calculation of the expected precession rate. It also introduces two polarimetric profiles that are considerably different from each other, providing the first evidence of profile evolution. Section \ref{sec:tip_change} describes in detail the observed secular changes in the total intensity profile. Polarimetric evolution is considered in section \ref{sec:pol_change}, which shows that there has been a convergence of the linear and circular components of the pulsar profile in the last twelve months and that the slope of the position angle swing is steepening, presumably as we approach the emission pole of the pulsar. Finally, in section \ref{sec:discussion} we discuss the implication of our observations for pulsar surveys and pulsar emission models. \section{Observations} \label{sec:observations} Observations were made at the Parkes radio telescope between July 1999 and May 2004, at centre frequencies ranging from 1318.25 MHz to 1413.50 MHz. Two different receiver packages were used to record data during this period; the central beam of the Parkes multibeam receiver and the wide band H-OH receiver. The multibeam receiver \citep{swb+96} has a system temperature of approximately 21K at 20 cm, which was about 5 degrees cooler than the H-OH receiver before it was upgraded near the end of 2003. Flux calibration of both systems using the radio galaxy 3C218 (Hydra A) suggests that the post-upgrade difference is only one or two degrees Kelvin. Our data were recorded with three different instruments, each designed for high time resolution observations across the widest possible bandwidths. Due to the rapid development of digital electronics within the past decade, each new observing system differed significantly from its predecessor. The Caltech Fast Pulsar Timing Machine (FPTM), described by \citet{nav94}, was a hardware-based auto-correlation spectrometer that performed incoherent dedispersion of dual orthogonal polarisations across two bands, each 128 MHz wide. Although the sampling rate was high enough to observe millisecond pulsars (MSPs) with only a few microseconds of smearing (at low DM), this instrument suffered from a number of artifacts induced by radio frequency interference and some deterioration in the correlator boards themselves. In some pulsars this led to oscillations in the passband that contaminated the pulsar profile. Nevertheless, many observations with this instrument were not affected by these problems and it successfully timed many MSPs to high accuracy \citep{tsb+99}. The FPTM 2-bit sampled the raw data and formed auto-correlation functions that were binned at the apparent spin period of the pulsar. We were able to apply routine 2-bit corrections to enable accurate polarimetry. Being an incoherent detector with a finite number of lags, the FPTM could divide the passband into 4$\times$128$\times$1 MHz channels, leaving a residual dispersion smear ($t_{\rm{smear}}$) given by Eq \ref{eq:smear}. \begin{equation} t_{\rm{smear}} = {8.3}\frac{B DM}{\nu^3} \mu s \label{eq:smear} \end{equation} \noindent Here, the channel bandwidth $B$ is in units of MHz, the sky frequency $\nu$ is in units of GHz and the dispersion measure $DM$ is in units of pc cm$^{-3}$. For the configuration used to observe PSR J1141--6545, this corresponds to 350 $\mu$s of smearing in the detected pulse profile. Given that the FPTM uses 1024 phase bins across a single pulse period and that PSR J1141--6545 rotates once every 394 ms, each phase bin represents 384 $\mu$s of time. The detrimental effects of dispersion smearing are therefore confined to within a single phase bin. However, if the pulsar spin period is two orders of magnitude shorter (as is typical of the millisecond pulsar population), dispersion smearing can significantly reduce the resolution of an incoherent detector. Motivated by a desire to overcome this problem for MSPs, the Caltech Parkes Swinburne Recorder Mk I (CPSR1) was commissioned in 1998 August. This system implemented a technique called coherent dedispersion \citep{hr75}, which requires Nyquist sampling of the observed band, followed by deconvolution with a response function characteristic of the interstellar medium (ISM). While this approach effectively eliminates dispersion smearing in the detected profiles, it is highly computationally intensive both in terms of the initial data rate and subsequent analysis. CPSR1 streamed digital samples to a striped set of four DLT drives (analogous to the method implemented for the S2 VLBI recorder) whose tapes were shipped to the Swinburne Centre for Astrophysics and Supercomputing for processing. Even with four striped tape drives, CPSR1 was limited to a bandwidth of 20 MHz. Rapid advances in consumer digital electronics soon made it feasible to upgrade the capabilities of the system and in 2002 August, CPSR Mk II was installed at Parkes. CPSR2 performs coherent dedispersion in near real-time, using a cluster of 30 server-class computers located at the telescope. It is capable of recording 2$\times$64 MHz dual-polarisation bands simultaneously, providing a total bandwidth comparable to that of the previous generation of incoherent detectors, like the FPTM. The coherent dedispersion method employed by both CPSR machines allows essentially arbitrary spectral resolution and reduces the dispersion smearing in each channel to a minute fraction of PSR J1141--6545's period, giving an effective sampling time of a few microseconds. Individual observations of PSR J1141--6545 ranged in duration from a few minutes to several hours. In recent years, our strategy has been to maximise orbital phase coverage by observing in concentrated sessions during which the pulsar is tracked continuously for two whole orbits ($\sim$9.6 hr). To calibrate the data we point the telescope one degree south of the pulsar and drive the in-built receiver noise source with a square wave at a frequency of 11.122976 Hz at least once per orbit, to characterise the polarimetric response of the system. In addition, at least once per month we observe the flux calibration source 3C218 (Hydra A). Our highest density of observations were taken with CPSR2 during 2003 and 2004 (MJD 52845 - 53134), during which time we have a full record of polarimetry, flux and profile morphology. The CPSR1 recorder was designed primarily to observe the bright southern millisecond pulsar J0437--4715 \citep{vbb+01}, however in 2001 January it took data on PSR J1141--6545 for a total of 30 hours beginning on MJD 51922. The resulting calibrated profile provides important, early epoch information. We have selected three profiles at the widely spaced epochs of MJD 51381, 51781 and 52087, representing high quality FPTM data to further supplement our temporal coverage. It is fortuitous that our earliest FPTM pointing (MJD 51381) dates all the way back to 1999, extending our time baseline by almost two full years. For this reason, we include a 1999 profile despite the fact that the observation was only 12 minutes in duration. Fortunately, PSR J1141--6545 is a bright pulsar with an average flux density of approximately 4 mJy, so the signal-to-noise (S/N) ratio of this 12 minute observations is 125, quite sufficient for our analysis. All data were processed using the tools included with the PSRCHIVE \citep{hvm04} scheme, with the addition of several extra routines specific to pulse variability analysis\footnote{All PSRCHIVE code is freely available for academic use, see http://astronomy.swin.edu.au/pulsar}. \section{PSR J1141--6545} \label{sec:1141} PSR J1141--6545 was discovered in the first Parkes multibeam survey \citep{klm+00a}. It resides in an unusual relativistic binary system, orbiting what is most likely a heavy white dwarf companion once every 4.8 hours. The pulsar does not appear to be recycled and is assumed to be the most recently evolved member of the system. Positioned close to the galactic plane, it is one of the few pulsars whose distance can be estimated by analysis of neutral hydrogen absorption features in its frequency spectrum. \citet{obv02a} obtain a lower limit of 3.7 kpc using this method. In addition, the signal from this pulsar exhibits diffractive scintillation over small bandwidths ($\sim$1 MHz) and time scales of a few minutes \citep{obv02}, which can be used to place timing-independent constraints on the binary parameters. \citet{obv02} report a significant detection of orbital modulation in the observed scintillation velocity (due to the motion of the pulsar in its orbit) and infer both a relative velocity of $\sim$115 km s$^{-1}$ and an orbital inclination angle $i$ = 76 $\pm$ 2.5$^\circ$. This velocity is large compared to the value of $<$ 50 km s$^{-1}$ expected to result from a symmetric supernova. Uncertainties in our knowledge of the true distance to the pulsar, and hence the relative velocity of the Earth's standard of rest, combined with uncertainties introduced by the Earth's orbital motion and bulk flows or anisotropies in the ISM, mean that we cannot convincingly state that the pulsar has an anomalous velocity due to an asymmetric supernova. However, if the profile evolution reported here is due to spin and orbital angular momentum misalignment and geodetic precession, we would expect the pulsar to have received a kick at birth, which would increase its runaway velocity over that of a symmetric explosion. The rotation period of PSR J1141--6545 is 394 ms, so the precision obtainable through pulse timing experiments is somewhat limited when compared to results \citep{vbb+01} obtained by timing millisecond pulsars, whose spin periods are of order 100 times shorter. Despite this, several post-Keplerian parameters are measurable. \citet{bok03} describe a timing solution that includes significant detections of periastron advance ($\dot \omega \sim$ 5.3$^\circ$ yr$^{-1}$), combined transverse Doppler and gravitational redshift ($\gamma$) and a marginal detection of orbital period derivative ($\dot P_{\rm b}$). Despite the lack of any Shapiro delay measurement, we can still derive a good estimate of the component masses. The post-Keplerian parameters $\dot \omega$ and $\gamma$ are related to the pulsar mass ($m_{\rm p}$) and companion mass ($m_{\rm c}$) by Eq \ref{eq:omdot} and Eq \ref{eq:gamma} respectively. In addition, pulse timing accurately determines the quantity, derived from Newtonian gravitation, known as the mass function (Eq \ref{eq:massfn}). These equations allow determination of the inclination angle and component masses. \begin{equation} \dot \omega = 3 \left({{2 \pi}\over{P_{\rm b}}}\right)^{5/3} \left({{G (m_{\rm p} + m_{\rm c})}\over{c^3}}\right)^{2/3} (1-e^2)^{-1} \label{eq:omdot} \end{equation} \begin{equation} \gamma = e \left({\frac{P_{\rm b}}{2 \pi}}\right)^{1/3} \frac{G^{2/3}}{c^{2}} m_{\rm c} (m_{\rm p} + 2 m_{\rm c}) {(m_{\rm p} + m_{\rm c})}^{4/3} \label{eq:gamma} \end{equation} \begin{equation} f(m_{\rm p},m_{\rm c}) = {{m_{\rm c}^3} \sin^3 i\over{(m_{\rm p} + m_{\rm c})^2}} = {{4 \pi^2}\over{G}}{{a^3 \sin^3 i}\over{P_{\rm b}^2}} \label{eq:massfn} \end{equation} \noindent Here, $G$ is Newton's gravitational constant, $P_{\rm b}$ is the pulsar orbital period and $a \sin i$ is the projected semi-major axis. According to \citet{bok03}, $m_{\rm p}$ = 1.30 $\pm$ 0.02 M$_\odot$ and $m_{\rm c}$ = 0.986 $\pm$ 0.02 M$_\odot$. The timing-derived inclination angle limit ($i > 75^\circ$) compares well with the value obtained from scintillation experiments \citep{obv02}. Assuming that General relativity is the correct theory of gravity, \citet{bo75b} present an expression (Eq \ref{eq:prate}) for the expected, time-averaged precession rate of the pulsar spin axis, $\Omega_{\rm p}$. \begin{equation} \Omega_{\rm p} = {\frac{1}{2}} {\left(\frac{G}{c^3}\right)}^{2/3} {\left(\frac{P_{\rm b}}{2\pi}\right)}^{-5/3} {\frac{m_{\rm c}(4m_{\rm p}+3m_{\rm c})}{(1-e^2)(m_{\rm p}+m_{\rm c})^{4/3}}} \label{eq:prate} \end{equation} \noindent Here, $c$ is the speed of light and $e$ is the eccentricity of the system. For PSR J1141--6545, this evaluates to an intrinsic precession rate of 1.36$^\circ$ yr$^{-1}$, which implies a precession period of 265 years. \citet{bai88} showed graphically that the maximum observable rate of precession may be significantly less than the intrinsic value. The geometry of the system and our viewing angle have a significant impact on our ability to detect geodetic precession. The observable quantity is the rate at which the angle $\delta$ between the observer's line of sight and the pulsar spin axis changes, as this will manifest itself as a changing cut through the emission cone. \citet{bai88} and \citet{cwb90} present expressions for the rate of change of $\delta$, of the form reproduced in Eq \ref{eq:delta}. The most important parameters in the expression are the misalignment angle between the spin axis and the orbital angular momentum vector, and the precessional phase at the current epoch, neither of which are known. \begin{equation} d\delta/dt = \Omega_{p} {\bf n} \cdot ({\bf s} \times {\bf j}) (1 - [{\bf n} \cdot {\bf s}]^{2})^{-1/2} \label{eq:delta} \end{equation} \noindent Here, ${\bf n}$ is a unit vector along the line of sight to the observer, ${\bf s}$ is a unit vector along the pulsar spin axis and ${\bf j}$ is a unit vector in the direction of the orbital angular momentum. To evaluate this equation, we must know the misalignment angle, the current precessional phase and the orbital inclination angle $i$. The first two parameters are unknown for the PSR J1141--6545 system, but we can assume the value of $i$ derived from scintillation studies and plot one precessional period of $d\delta/dt$ for various misalignment angles (see Fig \ref{fig:delta}). \begin{centering} \begin{figure} \includegraphics[angle=270,scale=0.65]{f1.eps} \caption{Rate of change of the angle between our line of sight and the spin axis of the pulsar ($d\delta/dt$) as a function of precessional phase, for various misalignment angles (shown inset top left, in degrees) for PSR J1141--6545. We have assumed an inclination angle of 76 degrees, consistent with scintillation measurements. The amplitude of the observable precession signature is highly dependent on both the misalignment angle and our current position in the precession cycle, but any derived value of $d\delta/dt$ gives a minimum misalignment angle for the system.} \label{fig:delta} \end{figure} \end{centering} Unfortunately, it is difficult to translate an observed difference in mean pulse profile morphology into a quantitative measure of angular shift. This requires knowledge of the intrinsic beam shape, which of course we do not possess, making the result model dependent. It is possible that polarimetric studies may offer a key alternative method, however it is still necessary to assume some model of pulsar polarisation as a function of impact angle. In this paper we restrict ourselves, where possible, to a quantitative description of the observed evolution of the mean pulse profile. \subsection{Selected Observations} Fig \ref{fig:cpsr1} shows a coherently dedispersed, polarimetrically calibrated mean profile, observed with the multibeam receiver and CPSR1 at a frequency of 1413 MHz in 2001 January (MJD 51922). The pulse profile is morphologically quite simple, consisting of a single component flanked on the left by a shoulder of emission. Note the small ``bump'' high on the leading edge of the profile, which is also present in the FPTM and analogue filter bank data. The peak fractional polarisation is of order 20 percent in both linear and circular. The position angle swing does not seem to fit the predictions of the rotating vector model and is similar to that seen by \citet{klm+00}, although lacking the orthogonal mode change that is present in the leading shoulder of the Kaspi et al. profile. \begin{centering} \begin{figure} \includegraphics[angle=270,scale=0.65]{f2.eps} \caption{PSR J1141--6545 mean profile, obtained from 28 hours of data taken in 2001 January (MJD 51922), at a centre frequency of 1413 MHz. The solid line represents total intensity, the dashed line total linear and the dotted line total circularly polarised emission. There are 1024 phase bins across the profile, which has been polarimetrically calibrated using a simple model of relative gain and phase in the orthogonal linear receiver probes. Note the slight ``bump'' on the leading edge of the profile and the absence of any steep position angle evolution across the phase range shown.} \label{fig:cpsr1} \end{figure} \end{centering} Fig \ref{fig:cpsr2} shows our most recent fully calibrated mean profile, observed in 2004 May (MJD 53134) with the H-OH receiver and CPSR2, at a centre frequency of 1341 MHz. There are a number of striking differences when compared to Fig \ref{fig:cpsr1}, most notably an overall broadening of the profile, which now has an extended trailing component; loss of the leading ``bump'' and general steepening of the position angle swing, which now has an identifiable slope. \begin{centering} \begin{figure} \includegraphics[angle=270,scale=0.65]{f3.eps} \caption{2004 May (MJD 53134) PSR J1141--6545 profile (as in Fig \ref{fig:cpsr1}) obtained from 2 hours of data taken with CPSR2 at a centre frequency of 1341 MHz. Note the smooth leading edge and extended trailing component, as well as the more pronounced position angle sweep.} \label{fig:cpsr2} \end{figure} \end{centering} Pulsar profiles can be corrupted by systematic errors associated with instrumentation. Typical effects include baseline artifacts due to improper binning and radio frequency interference, insufficient quantisation capabilities in the samplers (leading to the removal of power near the pulse as the samplers attempt to maintain an optimum mean level), and smearing due to insufficient time and frequency resolution. Fortunately, in the case of PSR J1141--6545 we have a variety of instruments, each of which has sufficient time resolution to over-sample the profile, and that give self-consistent results at similar epochs. We would be concerned if observed changes only coincided with equipment upgrades, but this has not been the case. The mean pulse profile has changed significantly in the space of three years. This is confirmed in the next two sections where we present an analysis of data taken during (and before) the epochs presented in Figs \ref{fig:cpsr1} \& \ref{fig:cpsr2}. We observe a smooth secular change in the characteristics of the mean pulse. \section{Evolution of the Total Intensity Profile} \label{sec:tip_change} To examine the evolution of PSR J1141--6545's mean pulse profile in greater detail, we sum all the data within each observing session to produce a set of 13 well spaced, high S/N profiles. These mean profiles typically span one or two days, with total integrated times of a few hours. Table 1 summarises the most important parameters of each profile including observing system, start date, observing frequency and S/N. \begin{table} \begin{center} \begin{tabular}{c|c|c|c} \hline Instrument & Date (MJD) & Freq (MHz) & S/N \\ \hline FPTM & 51381 & 1318.25 & 125 \\ FPTM & 51781 & 1413.50 & 791 \\ CPSR1 & 51922 & 1413.00 & 778 \\ FPTM & 52087 & 1413.50 & 743 \\ CPSR2 & 52614 & 1405.00 & 1051 \\ CPSR2 & 52845 & 1341.00 & 427 \\ CPSR2 & 52902 & 1341.00 & 822 \\ CPSR2 & 52902 & 1341.00 & 1354 \\ CPSR2 & 52920 & 1341.00 & 390 \\ CPSR2 & 53109 & 1341.00 & 346 \\ CPSR2 & 53130 & 1341.00 & 728 \\ CPSR2 & 53133 & 1341.00 & 537 \\ CPSR2 & 53134 & 1341.00 & 553 \\ \hline \end{tabular} \caption{List of parameters associated with the 13 observations used to characterise the secular evolution of the PSR J1141--6545 mean pulse profile.} \end{center} \end{table} We demonstrate that even though our points are not evenly spaced in time, the data describe a clear trend in profile evolution (Figs \ref{fig:mean} \& \ref{fig:diff}). The changes are so great that visual inspection of the profiles can reveal much qualitative information including an overall broadening, extension of the trailing shoulder and smoothing of the leading edge (Fig \ref{fig:mean}). In addition, we compute a quantitative measure of the width of each mean profile using an algorithm that defines thresholds in pulse phase based on where the flux under the pulse exceeds 10 percent of the peak value for the first time, and drops below this value for the last time. We choose the 10 percent threshold because it includes the majority of the on-pulse region, thus incorporating the important leading and trailing components of the pulse, which are seen to evolve significantly. The secular trend remains if the level is set to 50 percent, however the $\chi^{2}$ of the linear fit worsens marginally as we would expect from this narrower region of the profile. \begin{centering} \begin{figure} \includegraphics[angle=270,scale=0.65]{f4.eps} \caption{The upper panel shows superposed mean total intensity profiles from observations taken at various epochs during the past five years. Each profile was observed at a wavelength near 20cm. The MJD of each observation is marked on the plot, in line with the peak of the corresponding profile. For example, at pulse phase 0.495, the MJD of each observation increases monotonically with decreasing amplitude. The flux under each profile in the plotted region has been normalised to the flux under the earliest profile to allow direct comparison, regardless of the amplitude scaling schemes used by individual instruments. Each mean profile was aligned using an ephemeris obtained from a global timing solution across the entire data set, effectively maximising the cross-correlations between each profile. The lower panel shows the evolution of 10 percent width (see text) as a function of time, with one point for each profile in the upper panel. The profile clearly broadens over the span of our observations. Error bars are derived from consideration of the root-mean-square (RMS) noise level in each profile and represent 1-$\sigma$ uncertainties. The line of best fit, obtained using a linear least-squares method, is shown (dashed).} \label{fig:mean} \end{figure} \end{centering} A simple linear least-squares fit to the width data shows that the rate of profile broadening is well approximated by a straight line with slope 1.3 $\pm$ 0.06 ms yr$^{-1}$. Unlike PSR B1913+16, PSR J1141--6545 has only a single pulse profile component. It is therefore difficult to constrain the angular extent of the beam, or the angle between the magnetic axis and the line of sight, using total intensity information alone. A similar total intensity profile can be produced by intersecting the centre of a narrow beam or the edge of a wider beam, which introduces a degeneracy in the interpretation. \citet{sta04} describe a method for determining the full geometry of the system through the detection of a secondary precession effect caused by orbitally modulated aberration. Unfortunately we have been unable to detect this in PSR J1141--6545. Given a sufficient time baseline, the profile variations should eventually deviate from the currently observed linear trend. This will also provide an opportunity to constrain the three dimensional geometry of the pulsar system. In order to characterise the rate of change in more detail, we perform a difference profile analysis on our data set. This involves using (arbitrarily) the first profile in the series as a standard template whose amplitudes are subtracted from the remaining profiles after their flux and alignment have been normalised to the standard. Measurement of the remaining flux in the difference profile gives an indication of how much the profiles vary across a particular epoch (Fig \ref{fig:diff}). This method is similar to the technique of principle component analysis performed on PSR B1534+12 by \citet{sta04}, where only the mean profile and one orthogonal component are taken into account. We use this numerical measure of profile difference to examine quantitatively the rate at which evolution is occurring. \begin{centering} \begin{figure} \includegraphics[angle=270,scale=0.65]{f5.eps} \caption{The upper panel shows superposed difference profiles, constructed from the data shown in Fig \ref{fig:mean}, using the earliest profile as a standard template (shown as a horizontal line through zero). The vertical scale is in units of the standard profile flux, indicating that some individual components of the profile have changed by up to 20 percent. There is a monotonic increase in difference amplitude with time, at pulse phase 0.515 for example, the amplitude increases with the MJD of the observation. The lower panel shows the fractional difference between each difference profile and the standard template, as a function of time. Each point represents a single difference profile. The fractional difference is found by summing the absolute values of the amplitude in each pulse phase bin in the difference profile and dividing by the flux in the standard template. Errors are based on measurements of the off-pulse RMS. Whilst this technique is sensitive to changes in S/N as well as morphology, we are confident that our choice of high S/N observations and a narrow phase window allows the morphological information to dominate. The line of best fit is shown (dashed), obtained using a linear least-squares method. Note that the earliest profile has a fractional difference (and error) of zero, providing a reference point.} \label{fig:diff} \end{figure} \end{centering} Fig \ref{fig:diff} shows that the fractional difference trend is well approximated by a straight line with slope $2.4 \pm 0.07 \times10^{-4}$ fractional difference units per day. The mean pulse profile of PSR J1141--6545 is therefore changing at a steady rate of approximately 9 percent per year. The profile evolution seen in Figs \ref{fig:mean} \& \ref{fig:diff} is unlikely to be instrumental in origin because it occurs smoothly over the entire time span, instead of jumping discontinuously at the points when new hardware was introduced. In addition, the instrumental upgrades always decreased systematic smearing of the observed profile, whereas we observe the profile width increasing with time. Given that our observations span a frequency range of almost 100 MHz, it is possible that intrinsic evolution of the profile shape with frequency might contaminate the result. Such contamination is however unlikely to be responsible for the observed secular trend in pulse width, because as Table 1 shows, the changes in observing frequency have not been linear in time. To demonstrate that PSR J1141--6545 does not exhibit significant profile evolution over the range of observing frequencies in our data, we compare two CPSR2 profiles observed on the same day (MJD 53204) in July 2004. These two profiles were observed at 1405 MHz and 1341 MHz respectively and we analyse them in a similar fashion to Fig \ref{fig:diff}, using the 1405 MHz profile as a standard template and constructing a difference profile. Fig \ref{fig:freqs} shows that the profile does not evolve significantly across a bandwidth of 64 MHz, therefore we can be confident that frequency evolution does not contaminate our result. \begin{centering} \begin{figure} \includegraphics[angle=270,scale=0.65]{f8.eps} \caption{Difference profile constructed from two CPSR2 observations taken on MJD 53204 at two frequencies separated by 64 MHz. There is no systematic morphological difference above the level of the noise that might be expected if the pulsar profile was rapidly evolving as a function of radio frequency.} \label{fig:freqs} \end{figure} \end{centering} A number of different authors have reported observing the mean profile of a pulsar changing in some way. Some slower pulsars have been observed to emit in two or more ``modes'' of pulse shape, that each remain stable for a time before switching rapidly to another mode \citep{lyn71a}. In more recent years, various authors have reported observing random profile variations in some of the millisecond pulsars, most notably PSR B1821--24 \citep{bs97} and PSR J1022+1001 \citep{kxc+99}. None of the reported variations, or the mode changing phenomenon, match the steady secular change we have observed in PSR J1141--6545, which is itself a slow pulsar and therefore may not suffer from the erratic variations seen in a small number of the millisecond pulsars. It is still unclear whether or not random variations in MSP profiles are intrinsic to the pulsar as a recent analysis of PSR J1022+1001 observations made by \cite{hbo04} finds no instabilities that might induce strange timing behaviour. Can we hope to distinguish between geodetic precession and, say, free precession of the pulsar? Pulse shape variations are attributed to free precession in observations of PSR B1828--11 \citep{sls00} and PSR B1642--03 \citep{slu01}. This evidence manifests in the form of profile shape changes correlated to variations in the pulse arrival times. These changes are cyclic, but with no clear {\it a priori} time scale. The geodetic precession time scale is already well determined to be near 265 years. From Fig 1. we would hope to see the emission cone tilt at a rate $< 1.36$ deg yr$^{-1}$ and continue any secular trend for decades unless we are at one of the two crossing points encountered every geodetic precession period. The overwhelming majority of slow pulsars exhibit no evidence for free precession whatsoever. On the other hand, we expect any binary pulsar that has received a misaligning kick to precess to some degree. We therefore assert that geodetic precession is responsible for the observed secular variation in the mean profile of PSR J1141--6545 and we discuss this interpretation in section \ref{sec:discussion}. Long-term monitoring of the precession will ultimately determine a time scale and hence the true mechanism. We now present an analysis of the polarimetry of this pulsar, providing further evidence that our line of sight to the emission cone is changing steadily with time. \section{Evolution of Polarised Emission} \label{sec:pol_change} The standard interpretation of pulsar polarimetry is the rotating vector model (RVM), put forward by \citet{rc69a}. This assumes a dipolar magnetic field whose central axis is offset from the neutron star rotation axis. As the emission cone sweeps past our line of sight, the changing orientation between the observer and the magnetic field is expected to produce a characteristic ``S'' shaped curve in the measured position angle of any linearly polarised components. In addition, the rate at which the linear polarisation vector rotates as the beam crosses the observer is dependent on whether or not the line of sight cuts close to the centre of the beam. Polarimetric observations may therefore offer a sensitive indicator of both the rate at which the beam is precessing past the observer and where in the emission cone we are at the present time. First we consider the morphology of the polarised component of the mean pulse. The polarimetric changes observed between Figs \ref{fig:cpsr1} \& \ref{fig:cpsr2} are extreme. To convince ourselves that rapid polarimetric evolution is taking place, we focus on the most recent data with a high time density of observations and the same pulsar back-end (CPSR2). The polarimetric capabilities of CPSR2 and the PSRCHIVE scheme \citep{hvm04} have recently been verified by comparison with published results with good agreement \citep{osh+04}. Fig \ref{fig:pr} shows profiles of the polarised emission of PSR J1141--6545, recorded using CPSR2 and ordered consecutively in time. \begin{centering} \begin{figure} \includegraphics[angle=270,scale=0.7]{f6.eps} \caption{Consecutive profiles of the polarised emission from PSR J1141--6545, observed with CPSR2 over a period of 289 days beginning in 2003 July (MJD 52845). The dashed line represents total linear polarisation and the dashed-doted line represents total circular. The vertical axis (unlabeled for simplicity) is in mJy, the vertical range is kept constant across all sub-panels. All profiles have been polarimetrically calibrated using a simple model of relative gain and phase for two orthogonal linear receiver probes and flux calibrated against Hydra A. The peaks of the linear and circular components can be seen to move closer together as time progresses.} \label{fig:pr} \end{figure} \end{centering} Fig \ref{fig:pr} shows that the polarimetric profile is certainly changing, even over a period of less than 10 months. The linear and circular emission appears to converge during this time. Significant morphological differences in polarised emission appear between two widely separated observations that also correspond to a change in receiver, when the multibeam system was replaced by the refurbished H-OH. Therefore it is still possible that the differences in polarised emission could be instrumental in origin, although comparison with Fig \ref{fig:cpsr1} suggests that position angle evolution is also occurring. Observations over a longer time span are required to make more conclusive (and perhaps quantitative) statements. Regular polarimetric observations of this pulsar will be a high priority in future years. Next we attempt a more quantitative analysis of the position angle of the linearly polarised component of the pulsar beam. In the formalism of the RVM, measured position angle (PA, $\psi$), is presented as a function of the angle between the spin and magnetic axes ($\alpha$), the minimum angle between the magnetic axis and the line of sight ($\beta$) and pulse phase ($\phi$), as shown in Eq \ref{eq:rvm}. \begin{equation} \label{eq:rvm} \tan(\psi(\phi) - \psi_{0})\, =\, \frac{\sin \alpha \sin(\phi - \phi_0)} {\cos\alpha\sin\delta - \sin\alpha\cos\delta\cos(\phi - \phi_0)}, \end{equation} Here $\delta = \alpha + \beta$ is the angle between the spin axis and the line of sight, as in section \ref{sec:1141}; $\phi_{0}$ is the pulse phase of steepest PA swing and $\psi_{0}$ is a constant position angle offset. Unfortunately, application of this method to pulsars with narrow duty cycles and shallow or complicated PA swings does not well constrain $\alpha$ or $\beta$ independently. In common with the analysis presented by \citet{klm+00} we find it impossible to fit the RVM model to the early observations of PSR J1141--6545 with any degree of confidence. More recent observations are better suited to a partial RVM analysis, in that the recent PA behaviour includes a sweep in the central region of the pulse profile. The narrow duty cycle still restricts the applicability of the RVM and it is therefore impossible to constrain $\alpha$. Under the assumption that the steepest PA evolution is still providing an indication of the orientation of the magnetic field lines with respect to both the spin axis and the line of sight we have applied a simplification of the RVM model with the sole purpose of determining the general evolution of $\beta$. Within the RVM formalism the rate of change of PA as a function of pulse phase has a maximum value, given in Eq \ref{eq:rvm2}: \begin{equation} \label{eq:rvm2} \left(\frac{d\psi}{d\phi}\right)_{\rm max}\, = \, \frac{\sin\alpha}{\sin\beta} \end{equation} We have evaluated the gradient of the steepest PA swing in those observations for which a straight line fit can be obtained from the same central region of the pulse profile. This procedure could only be applied to epochs between which the linear behaviour of the PA swing was considerably wider in pulse phase than any possible translation, allowing the profiles to be aligned by a suitable ephemeris. The applicable observations are those obtained with CPSR2 since mid 2003. Earlier observations cannot be subject to this analysis as they display PA behaviour which is too complicated. This fact alone indicates that the detected emission represents a different cut through the emission region than was evident in earlier observations. The data have been grouped into two epochs (2003 and 2004) separated by approximately 0.7~years. Average position angle profiles were formed from the constituent observations at each epoch. These average PA profiles include observations from both 64 MHz observing bands and within each epoch are separated by several days. Although there were very slight variations between the PA profiles within each epoch, there was no systematic trend. A linear least squares minimisation of a straight line fit to the PA profiles demonstrates a significant deviation in the gradient obtained at the two epochs. Epoch one (2003) displays a gradient of --15.1 $\pm$ 0.3 degrees of PA swing per degree of phase. Epoch two (2004) displays a gradient of --17.1 $\pm$ 0.3 degrees of PA swing per degree of phase. A straight line fit to a difference PA profile, formed by the subtraction of the epoch two profile from the epoch one profile, is presented in Fig \ref{fig:diffpa}. The gradient of this difference fit is 2.3 $\pm$ 0.4, which is consistent with the simple subtraction of the best fit for epoch two from that of epoch one. Eq \ref{eq:rvm2} was then evaluated for all $\alpha$ and a rate of change of $\beta$ between the two epochs was determined. The calculated value of $d\beta/dt$ is a strong function of $\alpha$ but suggests that $\beta$ is increasing. The peak rate of change is 0.8 degrees yr $^{-1}$ (68 \% confidence), but this interpretation requires the pulsar spin axis to be 90 deg from the magnetic axis, which is unlikely if we assume the orientation is random. It is therefore probable that $d\beta/dt$ is less than 0.8 degrees yr$^{-1}$. The gradient of PA also indicates that $\beta$ is currently negative for all possible $\alpha$. We therefore infer that the beam is precessing into the line of sight at a rate less than 0.8 degrees yr$^{-1}$. It should be noted that $d\beta/dt \equiv d\delta/dt$ as the two angles are related by a constant offset. Thus we can use this result, coupled with Fig \ref{fig:delta}, to make some general statements about the unknowns in Eq \ref{eq:delta}. It is clear that either the misalignment angle is smaller than approximately 30$^\circ$, or we are currently at a special precessional phase where $d\delta/dt$ is changing rapidly and has not attained its maximum value in the span of our observations. Conversely, unless $\alpha$ is near 0$^\circ$ or 180$^\circ$, we expect that the misalignment angle is greater than approximately 15$^\circ$. \begin{centering} \begin{figure} \includegraphics[angle=270,scale=0.65]{f7.eps} \caption{The difference position angle (PA) as a function of phase. The average PA from the mid-2004 epoch was subtracted from that of the mid-2003 epoch. A straight line fit has been applied to the residual. The steepening of the PA swing as a function of time is an indication that the pulsar emission beam may be precessing into, and not away from our line of sight.} \label{fig:diffpa} \end{figure} \end{centering} The mean flux of the pulsar in these observations is 4.0 $\pm$ 0.5 mJy. There does not appear to be a significant trend in mean flux within the past year, but this may change as the time baseline of flux calibrated data grows. \section{Discussion \& Conclusion} \label{sec:discussion} Our observations indicate that the line of sight to the emission cone of PSR J1141--6545 is precessing deeper into the core, producing a steeper position angle swing. Depending on the chosen beam model, this could also explain why the pulse profile is seen to increase in width. Our result has implications both for the study of pulsar emission beams and for the detection rates of relativistic pulsars in large-scale surveys. PSR J1141--6545 is the slowest pulsar (by an order of magnitude) for which geodetic precession is observable, providing a unique means of examining the emission cone of a normal (un-recycled) pulsar. Further polarimetric observations, extended time baselines or the detection of orbitally modulated aberration may allow determination of our present location in the emission cone and the geometry of the beam as a whole. Given the expected precession period of order 265 years and the fact that we seem to still be moving towards the central axis of the beam, it is possible that this pulsar may only have precessed into view within the past few decades. This might explain its non-detection in early pulsar surveys with flux limits well below the required threshold \citep{jlm+92, mld+96}. If this is the case, it could be argued that surveys of the sky for relativistic pulsars should continue on a regular basis. By their very nature, the most interesting objects are likely to be visible for the least amount of time. \acknowledgments The Parkes radio telescope is operated by the Australia Telescope National Facility on behalf of the Commonwealth Scientific and Industrial Research Organisation (CSIRO). We are grateful to Ruth Musgrave for assistance with data reduction and Willem van Straten and Haydon Knight for assistance with observations and software development. AWH is the recipient of an Australian Post-graduate Award (APA) and CSIRO top-up allowance and thanks Claire Trenham for support and encouragement. \bibliographystyle{apj}

\section{Introduction} \Label{d1} Let $\Om$ be a pseudoconvex domain in $\C^n$ and $z_o$ be a boundary point. For a smooth monontonic increasing function $f :[1+\infty)\to[1,+\infty)$ with $f (t)\le t^{\frac{1}{2}}$, we say that $\Om$ has the $f$-Property at $z_o$ if there exist a neigborhood $U$ of $z_o$ and a family of functions $\{\phi_\delta\}$ such that \begin{enumerate} \item [(i)] $\phi_\delta$ are plurisubharmonic and $C^2$ on $U$ and $-1\le \phi_\delta \le0$; \item[(ii)] $\di\dib \phi_\delta\simge f(\delta^{-1})^2Id$ and $|D\phi_\delta|\lesssim \delta^{-1}$ for any $z\in U\cap \{z\in \Om:-\delta<r(z)<0\}$, where $r$ is a defining function of $\Om$. \end{enumerate} Here and in what follows, $\lesssim$ and $\simge$ denote inequality up to a positive constant. Morever, we will use $\approx$ for the combination of $\lesssim$ and $\simge$. \\ In the joint work with G. Zampieri \cite{KZ10}, we show that the $f$-Property implies an $f$-estimate for the $\dib$-Neumann problem. In another paper \cite{KZ12}, we prove that an $f$-estimate with $\dfrac{f}{\log }\to \infty$ at $\infty$ implies that the Bergman metric has a lower bound with the rate $g(t)=\dfrac{f}{\log}(t^{1-\eta})$ for $\eta>0$. The ideas leading to these results follow by Kohn \cite{Koh02}, Catlin \cite{Cat83, Cat87} and McNeal \cite{McN92a}. Combining the two results above, we obtain \begin{theorem}\Label{B}Let $\Om$ be a pseudoconvex domain in $\C^n$ with $C^\infty$-smooth boundary and $z_o$ a point in the boundary $b\Om$. Assume that the $f$-Property holds at $z_o$ with $\dfrac{f}{\log }\nearrow \infty$ for $t\to\infty$. Then for any $\eta>0$ there is a neigborhood $U_\eta$ of $z_o$ and a constant $C_\eta$ such that the Bergman metric $B$ of $\Om$ satisfies \begin{eqnarray}\Label{b1} B(z,X)\ge C_\eta \dfrac{f}{\log}(\delta^{-1+\eta}_\Om(z))|X| \end{eqnarray} for any $z\in U_\eta\cap\Om$ and $X\in T_z^{1,0}\C^n$. \end{theorem} The purpose of this paper is to prove a result similar to Theorem \ref{B} for the Kobayashi metric. Let us recall the definition of the Kobayashi metric. \\ Let $\Om$ be a pseudoconvex domain in $\C^n$; the function $K: T^{1,0}\Om\to \R$ on the holomorphic tangent bundle, given by \begin{eqnarray} \begin{split} K(z,X)=&\inf\{ \alpha>0| \exists g: \Delta\to \Om \T{~holomorphic with~} g(0)=z, g'(0)=\alpha^{-1}X\}\\ =& \inf\{r^{-1}| \exists g: \Delta_r\to \Om \T{~holomorphic with~} g(0)=z, g'(0)=X\}, \end{split} \end{eqnarray} is called the Kobayashi metric of $\Om$. Here $\Delta$ denotes the unit disc and $\Delta_r$ the disc in $\C$ centered at $0$ with radius $r$. \\ Our main result is the following \begin{theorem}\Label{t1}Let $\Om$ be a pseudoconvex domain in $\C^n$ with $C^2$-smooth boundary $b\Om$ and $z_o$ be a boundary point. Assume that $\Om$ has the $f$-Property at $z_o$ with $f$ satisfying $\displaystyle\int_t^\infty \dfrac{da}{a f(a)}<\infty$ for some $t>1$, and denote by $(g(t))^{-1}$ the above, finite, integral. Then, there is a neighborhood $V$ of $z_o$ such that \begin{eqnarray}\Label{K1} K(z,X)\simge g(\delta^{-1}_{\Om}(z))|X| \end{eqnarray} for any $z\in V\cap\Om$ and $X\in T^{1,0}_z\C^n$. \end{theorem} We remark that Theorem~\ref{t1} may apply to domains of both finite and infinite type; in the first case we take $f(t)=t^\epsilon,\, g(t)=t^\epsilon$, and in the second $f(t) = \log^{1+\epsilon}$, $g(t)=\log^\epsilon t$. \\ Comparing with Theorem~\ref{B}, we reduce the $C^\infty$-smoothness of the boundary and slightly strengthen the hypothesis of $f$ in the $f$-Property since $\displaystyle\int_t^\infty \dfrac{da}{a f(a)}<\infty$ is stronger than $\underset{a\to \infty}{\lim}\dfrac{f(a)}{\log a}=\infty$. Morever, we obtain a larger size of the lower bound of the Koybayashi metric; for example, in the case $f=t^\epsilon$ we have $g=t^\epsilon$ instead of $g=t^{\epsilon-\eta}$ and, for $f=\log^{1+\epsilon}t$ we have $g=log^\epsilon t$ instead of $\log^\epsilon(t^{1-\eta})$. \\ Using the $f$-Property constructed by Catlin in \cite{Cat87, Cat89}, McNeal \cite{McN91b, McN92b}, Khanh-Zampieri \cite{KZ10}, Khanh \cite{Kha10}, we have the following \begin{corollary}1) Let $\Om$ be a pseudoconvex domain of finite type $m$ in $\C^n$. Then \eqref{K1} holds for $g(t)=t^{\frac{1}{m}}$ if $\Om$ satisfies at least one of the following conditions: $\Om$ is strongly pseudoconvex, or $\Om$ is convex, or $n=2$, or $\Om$ is decoupled. In any case, we have $g(t)=t^\epsilon$ with $\epsilon=m^{-n^2m^{n^2}}$.\\ 2) Let $\Om$ be defined by $\Om=\{z\in \C^n:\T{Im}z_n+\sum_{j=1}^{n-1}P_j(z_j)<0\},$ where $\Delta P_j(z_j)\simge \dfrac{\exp(-1/|x_j|^\alpha)}{x_j^2}$ or $\dfrac{\exp(-1/|y_j|^\alpha)}{y_j^2}$ wih $\alpha<1$. Then \eqref{K1} holds for $g(t)=\log^{\frac1\alpha-1}t$. \end{corollary} The lower bound of the Kobayashi metric is an important tool in the function theory of several complex variables and has been studied by many authors. In the following, we briefly review some significant, classical results.\\ When $\Om$ is strongly pseudoconvex or else it is pseudoconvex of finite type in $\C^2$ and decoupled or convex in $\C^n$, then the size of the Kobayashi metric has been described by I. Graham \cite{Gra75a}, D. Catlin \cite{Cat89}, G. Herbort \cite{Her92} and L. Lee \cite{Lee08}. In these classes of domains, there exists a quantity $M(z,X)$ which satisfies the asymptotic formula $$\underset{z\to b\Om}{\lim}M(z,X)=\delta_\Om^{-1/m}(z)|X^\tau|+\delta_\Om^{-1}(z)|X^\nu|,$$ (where, $X^\tau$ and $X^\nu$ are the tangential and normal components of $X$ and $m$ is the type of the boundary), such that $$K(z, X)\approx M(z,X).$$ For a general pseudoconvex domain in $\C^n$, K. Diederich and J. E. Fornaess \cite{DF79} proved, by using Kohn's algorithm \cite{Koh79}, that there is a $\epsilon>0$ such that $K(z,X)\simge \delta(z)^{-\epsilon}|X|$ if $b\Om$ is real analytic of finite type. By using the method of Catlin in \cite{Cat87, Cat89}, S. Cho \cite{Cho92} improved the result of \cite{DF79} for domains which are not necessarily real analytic. However, in the case of infinite type we know very little except from the recent results by S. Lee \cite{Lee01} for the exponentially-flat infinite type. \\ Among other uses of the lower bound of the Kobayashi metric, we mention the continuous extendibility of proper holomorphic maps to the boundary of a domain of general type. We refer readers to \cite{Hen73, BF78e, DF79, Ran78} for this problem on domains of finite type. \begin{theorem}\Label{fHolder}Let $\Om$ and $\Om'$ be pseudoconvex domains. Let $\eta, 0<\eta\le 1$ be such that there is a $C^2$ defining function $r$ of $\Om$ with the property that $-(-r)^\eta$ is strictly plurisubharmonic on $\Om$. Assume that $\Om'$ has the $f$-Property with $f$ satisfying $\displaystyle\int_t^\infty \dfrac{(\ln a-\ln t) da}{a f(a)}<\infty$ for some $t>1$, and denote by $(\tilde f(t))^{-1}$ this finite integral. Then any proper holomorphic map $\Psi: \Om\to \Om'$ can be extended as a general H\"older continuous map $\hat\Psi:\bar\Om\to \bar\Om'$ with a rate $\tilde f(t^{\eta})$, that is, $$|\hat\Psi(z)-\hat\Psi(w)|\lesssim \tilde{f}(|z-w|^{-\eta})^{-1}$$ for any $z,w\in \bar\Om$. \end{theorem} The paper is organized as follows. In section 2, using the $f$-Property, we construct the bumping functions. By the existence of suitable exhaustion functions, we obtain the plurisubharmonic peak functions having the good estimates. The lower bound of the Kobayashi metric follows from the estimates of the plurisubharmonic peak functions (cf. Section 3). In Section 4, we prove Theorem~\ref{fHolder}. \\ \section{The bumping function} In this section, we construct the bumping functions, which might also be useful for other purposes. We will prove that, for any boundary point $z_o$ on $b\Om$ which satisfies the $f$-Property, we can find a pseudoconvex hypersurface touching $\bar\Om$ exactly at $w$ from the outside such that the distance from $z\in\Om$ to the new hypersurface is exactly controlled by the rate in$ |z-w|^{-1}$ of the reciprocal of the inverse of $g$. \begin{theorem} \Label{bumpingfunct} Let $\Om$ be pseudoconvex and $z_o$ be a boundary point. Assume that $\Om$ has the $f$-Property at $z_o$ with $f$ satisfying $\displaystyle\int_t^\infty \dfrac{da}{a f(a)}<\infty$ for some $t>1$, and denote by $(g(t))^{-1}$ this finite integral. Then there is a neigborhood $U$ of $z_o$ and a real $C^2$ function $\rho$ on $U\times(U\cap b\Om)$ with the following properties: \begin{enumerate} \item $\rho(w,w)=0$. \item $\rho(z,w)\le -G(|z-w|)$ for any $(z,w)\in (U\cap \Om)\times (U\cap b\Om)$ where $G(\delta)=\left(g^*(\gamma\delta^{-1}\right)^{-1}$. Here, the supercript $^*$ denotes the inverse function and $\gamma>0$ sufficiently small. \item $\rho(z,\pi(z))\gtrsim -\delta_\Om(z)$ for any $z\in U\cap \Om$ where $\pi(z)$ is the projection of $z$ to the boundary. \item For each fixed $w\in U\cap b\Om$, denote $S_w=\{z\in U: \rho(z,w)=0 \}$. One has: \begin{enumerate} \item $|D_z\rho(z,w)| \approx1$ everywhere on $S_w$. \item $S_w$ is pseudoconvex. In fact, one can choose $\rho$ such that $S_w$ is strongly pseudoconvex outside of $w$. \item $S_w$ touches $\bar\Om$ exactly at $w$ from outside. \end{enumerate} \end{enumerate} \end{theorem} The proof is divided in four steps. In step 1, we show the equivalence of the $f$-Property between the pseudoconvex and the pseudoconcave side of a hypersurface. In step 2, we prove that there exists a single function with self-bounded gradient which has a lower bound $f(r^{-1}(z))$ for the Levi form. In step 3, we estimate the function $G$. The properties of bumping function is checked on step 4. \\ {\it Proof of Theorem~\ref{bumpingfunct}.}\\ {\bf Step 1.} Since the hypersurface defined by each bumping function lies outside the original domain except from one point and the $f$-Property takes place inside the domain, we first show hat the $f$-Property still holds outside the domain. \\ Without loss of generality, we can assume that the original point $z_o$ belongs to $U\cap b\Om$. We choose special coordinates $z=(x,r)\in \R^{2n-1}\times \R$ at $z_o$. Assume that there is a family of functions $\phi_\delta$ which have properties (i) and (ii) in the first paragraph of Section 1. Define $\tilde\phi_\delta(x,r):=\phi_\delta(x,r-\delta)$ for each $\delta>0$ and still call $\phi_\delta$ for $\tilde\phi_\delta$. Then, for each $\delta$, $\phi_\delta$ is $C^2$, plurisubharmonic, and satisfies on $U$ $-1\le\phi_\delta\le0$, $\di\dib \phi_\delta\simge f^2(\delta^{-1})Id$, and $|D\phi_\delta|\lesssim \delta^{-1}$ on $-\delta<r-\delta<0$ or $0<r<\delta$.\\ {\bf Step 2.} In this step we build a single function which has self-bounded gradient and has lower bound $f(r^{-1}(z))$ for the Levi form. \begin{lemma}\Label{A3} Assume that $\Om$ enjoys the $f$-Property at $z_0$. Then there is a single function $\Phi$ and constants $c,C>0$ such that \begin{enumerate} \item $-1\lesssim \Phi\le 0$ \item $\di\dib\Phi(X,X)\ge-C(\dfrac{1}{r}|\di\dib r(X,\bar X)|+\dfrac{1}{r^2}|Xr|^2)+\dfrac{1}{8}|X\Phi|^2+ cf^2(\dfrac{1}{r})|X|^2$ \item $|D\Phi|\lesssim \dfrac{1}{r}$ \end{enumerate} for any $z\in U\setminus\bar\Om$. \end{lemma} {\it Proof. } Let $\chi$ be a cut-off function such that $\chi(t)= \begin{cases} 0 &\T{~if~} t\le \dfrac{1}{4} ~~\T{or}~~t\ge 2.\\ 1&\T{~if~} \dfrac{1}{2}\le t\le 1. \end{cases}$ We also suppose that $|\dot\chi|$, $|\ddot\chi|$ and $\dfrac{\dot\chi^2}{\chi}$ are bounded. Define \begin{eqnarray}\Label{a1} \Phi(z):=\sum_{j=1}^\infty\left(\exp(\phi_{2^{-j}}(z))-1\right)\chi(2^jr(z)). \end{eqnarray} Denote $S_\delta:=\{z\in U| 0<r(z)<\delta\}$. Let $z\in U\setminus\bar\Om=\underset{j=1}{\overset{\infty}{\cup}} S_{2^{-k}}\setminus S_{2^{-(k+1)}}$, then there is an integer $k$ such that \begin{eqnarray}\Label{2.1a} z\in S_{2^{-k}}\setminus S_{2^{-(k+1)}}=\{z\in U: 2^{-k-1}\le r(z)<2^{-k}\}. \end{eqnarray} We notice that $\chi(2^{j}r(z))=0$ if $j<k-1$ or $j>k+1$, and $\chi(2^kr(z))=1$ for any $z\in S_{2^{-k}}\setminus S_{2^{-(k+1)}}$. Hence, \eqref{a1} can be rewritten that $$\Phi(z)=\sum_{j=k-1}^{k+1}\left(\exp(\phi_{2^{-j}}(z))-1\right)\chi(2^jr(z)).$$ This proves (i). We observe that \begin{eqnarray}\Label{2.2} \begin{split} \di\dib\left(\left(e^{\phi_{2^{-j}}}-1\right)\chi(2^jr)\right)(X,\bar X)= &\left(\di\dib\phi_{2^{-j}}(X,\bar X)+|X\phi_{2^{-j}}|^2\right ) e^{\phi_{2^{-j}}}\chi\\ &+2^{j+1}\T{Re}\la X\phi_{2^{-j}}, \overline{Xr}\ra e^{\phi_{2^{-j}}}\dot\chi\\ &+\left(2^j\di\dib r(X,\bar X)\dot\chi+2^{2j}|Xr|^2\ddot\chi\right)(e^{\phi_{2^{-j}}}-1)\\ \ge&\left( \di\dib\phi_{2^{-j}}(X,\bar X)+\dfrac{1}{2}|X\phi_{2^{-j}}|^2\right)e^{\phi_{2^{-j}}}\chi\\ &-2^j\dot\chi(1-e^{\phi_{2^{-j}}})|\di\dib r(X,\bar X)|\\ &-2^{2j}(|\ddot\chi|(1-e^{\phi_{2^{-j}}})+2\frac{\dot\chi^2}{\chi}e^{\phi_{2^{-j}}})|Xr|^2. \end{split} \end{eqnarray} Here, we use the Cauchy-Schwartz inequality for the second line of \eqref{2.2}, that is, $$\left|2^{j+1}\T{Re}\la X\phi_{2^{-j}} \overline{Xr}\ra e^{\phi_{2^{-j}}}\dot\chi\right| \le \frac{1}{2}|X\phi_{2^{-j}}|^2e^{\phi_{2^{-j}}}\chi+2^{2j+1}|Xr|^2\frac{\dot\chi^2}{\chi}.$$ Moreover, we also observe that \begin{eqnarray}\Label{2.3} \begin{split} |X\Phi(z)|^2=&|\sum_{j=k-1}^{k+1}X(\phi_{2^{-j}})e^{\phi_{2^{-j}}}\chi+2^j(e^{\phi_{2^{-j}}}-1)X(r)\dot\chi(2^jr)|^2\\ \le & 4\sum_{j=k-1}^{k+1}|X\phi_{2^{-j}}|^2e^{\phi_{2^{-j}}}\chi(2^jr)+2^{2k+2}(1-e^{-1})^2|Xr|^2\left(\frac{1}{4}\dot\chi^2(2^{k-1})+4\dot\chi^2(2^{k+1})\right) \end{split} \end{eqnarray} Combining \eqref{2.2} and \eqref{2.3}, we obtain \begin{eqnarray}\Label{2.4} \begin{split} \di\dib\Phi(X,\bar X)\ge&e^{-1}\di\dib\phi_{2^{-k}}(X,\bar X)+\dfrac{1}{8}|X\Phi|^2-C(2^k|\di\dib r(X,\bar X)|+2^{2k}|Xr|^2)\\ \ge&cf^2(2^k)|X|^2+\dfrac{1}{8}|X\Phi|^2-C(2^k|\di\dib r(X,\bar X)|+2^{2k}|Xr|^2)\\ \end{split} \end{eqnarray} for any $z\in S_{2^{-k}}\setminus S_{2^{-(k+1)}}$. From \eqref{2.3}, we also obtain $|D\Phi|\lesssim 2^k$ $z\in S_{2^{-k}}\setminus S_{2^{-(k+1)}}$ since $|D\phi_\delta|\lesssim \delta^{-1}$ and $|Dr|\lesssim 1$. This completes the proof of (ii) and (iii).\\ $\hfill\Box$ {\bf Step 3.} We recall that $g(t)=\left(\displaystyle\int_t^\infty \dfrac{da}{a f(a)}\right)^{-1}=\left(\displaystyle\int_0^{t^{-1}}\dfrac{da}{a f(a^{-1})}\right)^{-1}$ for any $t>1$. Then, it is easy to check that $g$ is increasing, $g\to \infty$ at $\infty$, and $g\le f$ on $(1,+\infty)$. We define $$G(\delta):=g^*\left(((\gamma\delta)^{-1}\right)^{-1}$$ where $\gamma>0$ is a constant to be chosen later. We also notice that $G$ is an increasing function, and $G(0)=0$.\\ {\it Claim: For $\delta>0$, we have \begin{enumerate} \item $\dfrac{\dot G(\delta)}{G(\delta)}=\gamma f(G^{-1}(\delta))$; \item $G(\delta)\ddot G(\delta)\le \dot G^2(\delta)$; \item $\dfrac{G(\delta)}{\delta}\le \dot G(\delta)$. \end{enumerate}} {\it Proof of the Claim.} By the definition of $G$ and $g$, we have \begin{eqnarray}\Label{2.6} g(G(\delta)^{-1}=(\gamma\delta)^{-1}\quad \T{ or }\quad \displaystyle\int_0^{G(\delta)}\dfrac{da}{a f(a^{-1})}=\gamma \delta. \end{eqnarray} Taking the derivative with respect to $\delta$ in the second equation of \eqref{2.6}, we prove the first claim, that is, \begin{eqnarray}\Label{2.7} \dfrac{\dot G(\delta)}{G(\delta)}=\gamma f(G^{-1}(\delta)). \end{eqnarray} Taking again the derivative with respect to $\delta$ in \eqref{2.7}, and observing that \begin{eqnarray}\Label{2.8} -\frac{\gamma\dot G(\delta)\dot f(G^{-1}(\delta))}{G^2(\delta)}=\dfrac{G(\delta)\ddot G(\delta)-\dot G^2(\delta)}{G^2(\delta)}, \end{eqnarray} (since $G$ and $f$ are increasing functions) we get the proof of the second claim. Moreover, since $f\ge g$, then $G^{-1}(\delta)=g^*\left((\gamma\delta)^{-1}\right)\ge f^*\left((\gamma\delta^{-1})\right)$. From \eqref{2.7} we then get $\dfrac{\dot G(\delta)}{G(\delta)}\ge \gamma f(f^*((\gamma\delta)^{-1}))=\delta^{-1}$. The proof of the Claim is complete.\\ $\hfill\Box$ {\bf Step 4.} We define $$\rho(z,w)=r(z)+G(|z-w|)\left(-1+\epsilon\Phi(z)\right)$$ where $\epsilon>0$ will be chosen later. \\ Let $S_w= \{z\in U| \rho(z,w)=0\}$ be a hypersurface defined by $\rho(z,w)=0$ where $w$ is fixed. We will prove that $\rho$ satisfies the following properties: \begin{enumerate} \item[(i)] $\rho(w,w) = 0$ for any $w\in b\Om$. \item[(ii)] $\rho(z,w)\le -G(|z-w|)$ for $z\in U\cap \Om$ and $w\in U\cap b\Om$. \item[(iii)] $\rho(z,\pi(z)) \gtrsim - r(z)$ for $z\in U\cap \Om$, where $\pi(z)$ is the projection of $z$ to the boundary. \item[(v)] $S_w$ is pseudoconvex. \item[(iv)] $|D_z \rho(z,w)|\approx 1$ on $S_w$. \end{enumerate} Now, (i) is obvious. Since $\Phi$ is negative and bounded, we first choose $\epsilon$ so small that $-2\le -1+\epsilon\Phi\le -1$. For $z\in U\cap \Om$, we have $r(z)<0$, and $|r(z)|\ge G(|r(z)|)$, hence (ii) and (iii) follow. \\ We estimate the Levi form of $\rho$ with respect to $z$, \begin{eqnarray}\Label{phi} \begin{split} \di_z\dib_z\rho(z,w)(X,X)=&\di\dib r(z)(X,X)+\big(\dfrac{\dot{G}(|z-w|)}{|z-w|}+\ddot{G}(|z-w|)\big)(-1+\epsilon\Phi(z))|X|^2\\ &+2\epsilon\T{Re}\la XG(|z-w|),X\Phi(z)\ra+\epsilon G(|z-w|)\di\dib \Phi(z)(X,X)\\ \ge&\di\dib r(z)(X,X)-2\big(\dfrac{\dot{G}(|z-w|)}{|z-w|}+\ddot{G}(|z-w|)+8\epsilon\dfrac{\dot{G}^2(|z-w|)}{G(|z-w|)} \big)|X|^2\\ &+\epsilon G(|z-w|)\big(\di\dib \Phi(z)(X,X)-\frac{1}{16}|X\Phi(z)|^2\big)\\ \ge&\di\dib r(z)(X,X)-\epsilon C G(|z-w|)\left(\dfrac{1}{r}|\di\dib r(X,\bar X)|+\dfrac{1}{r^2}|Xr|^2\right)\\ &+\epsilon c G(|z-w|)f^2(\dfrac{1}{r})|X|^2+\dfrac{\epsilon }{16}G(|z-w|)|X\Phi|^2 \\ &-2\big(\dfrac{\dot{G}(|z-w|)}{|z-w|}+\ddot{G}(|z-w|)+8\epsilon\dfrac{\dot{G}^2(|z-w|)}{G(|z-w|)} \big)|X|^2\\ \end{split} \end{eqnarray} Here, the first inequality follows from Cauchy-Schwartz inequality as for the second line; the last inequality follows from Lemma~\ref{A3}(ii).\\ Now we consider $z\in (S_w\cap U)\setminus{w} \subset U\setminus \bar\Om$, that is, $r(z)=G(|z-w|)\big(1-\epsilon\Phi(z)\big)$. By the choice of $\epsilon$, we obtain \begin{eqnarray}\Label{c2.10} G(|z-w|)\le| r(z)|\le 2G(|z-w). \end{eqnarray} Thus the inequality of \eqref{phi} continues as \begin{eqnarray}\Label{b2} \begin{split} \di_z\dib_z\rho(z,w)(X,X)\ge&\di\dib r(z)(X,X)-2\epsilon C|\di\dib r(X,X)|-\frac{4C\epsilon }{G(|z-w|)}|Xr|^2\\ &+\frac{\epsilon}{16}G(|z-w|)|X\Phi|^2\\ &+\left( c\epsilon G(|z-w|)f^2\left(\frac{1}{G(|z-w|}\right)-(4+16\epsilon)\dfrac{\dot{G}^2(|z-w|)}{G^2(|z-w|)} \right)|X|^2\\ \end{split} \end{eqnarray} Here the last line follows from Claim (2) and (3).\\ Choose $\epsilon$ small such that $2\epsilon C\le 1$; the first line of \eqref{b2} can be estimated as follows \begin{eqnarray}\Label{2.11} \begin{split} &\di\dib r(z)(X,X)-2\epsilon C |\di\dib r(X,X)|-\frac{4\epsilon C}{G(|z-w|)}|Xr|^2\\ &\ge \di\dib r(z)(X,X)-|\di\dib r(X,X)| -\frac{2}{G(|z-w|)}|Xr|^2\\ &\ge C_1|X| |Xr|-\frac{2}{G(|z-w|)}|Xr|^2\\ &\ge-\frac{C_1^2}{4}G(|z-w|)|X|^2-\frac{3}{G(|z-w|)}|Xr|^2\\ \end{split} \end{eqnarray} For $X\in T^{1,0}S_w$, that is, $X\rho=0$. This implies $$Xr=(1-\epsilon \Phi)XG(|z-w|)-\epsilon G(|z-w|)X\Phi,$$ and hence, $$ |Xr|^2\le 8\dot G^2(|z-w|)|X|^2+2\epsilon^2 G^2(|z-w|)|X\Phi|^2. $$ The inequality \eqref{2.11} continues as \begin{eqnarray}\Label{2.12} \begin{split} &\ge-C_2 \frac{\dot G^2(|z-w|)}{G(|z-w|)}|X|^2-6\epsilon^2 G(|z-w|)|X\Phi|^2. \end{split} \end{eqnarray} Combining \eqref{b2} and \eqref{2.12}, we obtain \begin{eqnarray}\Label{2.13} \begin{split} \di\dib\rho(X,\bar X)\ge& \left(\dfrac{\epsilon}{16}-6\epsilon^2\right)G(|z-w|)|X\Phi|^2\\ &+c\epsilon G(|z-w|)\left(f^2\left(\frac{1}{G(|z-w|)}\right)-C_3 \frac{\dot G^2(|z-w|)}{G^2(|z-w|)}\right)|X|^2 \end{split} \end{eqnarray} Again, choose $\epsilon$ such that $\dfrac{\epsilon}{16}-6\epsilon^2\ge 0$; then the term in left hand side of the first line of \eqref{2.12} can be disregarded. Using Claim (1) with $\gamma>0$ small enough, we obtain that the term in the second line is positive. We conclude that $\di\dib\rho(X,X)\ge 0$ on $S_w$ for any $X\in T^{1,0}S_w$. The proof of property (iv) is complete. \\ For any $z\in (S_w\cap U)\setminus {w}$, we have \begin{eqnarray} \begin{split} \left|D\left(G(|z-w|)(-1+\epsilon\Phi(z))\right)\right|\le& 2\dot G(|z-w|)+\epsilon G(|z-w|)|D\Phi(z)|\\ \le& 2\eta G(|z-w|)f(G^{-1}(|z-w|))+\epsilon \frac{G(|z-w|)}{r(z)}\\ \le& 2\eta+\epsilon\\ \end{split} \end{eqnarray} where, the second inequality follows from Claim (1) and Lemma~\ref{A3}.(3), the third inequality follows from the hypothesis that $f(t)\le t$ and \eqref{c2.10}. Since $|Dr|\approx 1$, then for $\epsilon$ and $\eta$ small enough, we obtain $|D\rho|\approx 1$. That is the proof of property (iv). The proof of Theorem \ref{bumpingfunct} is complete. $\hfill\Box$ \section{Proof of Theorem~\ref{t1}} The proof of Theorem~\ref{t1} follows immediately from Theorem~\ref{pshpeak} and Theorem~\ref{lowerbound} below. Theorem~\ref{pshpeak} consists in the construction of plurisubharmonic peak functions with good estimates. This is a consequence of the construction of bumping functions in last section. More precisely, we obtain the following \begin{theorem}\Label{pshpeak}Assume that there exists a family of bumping functions on a local path $V$ of the boundary as in the conclusion of Theorem \ref{bumpingfunct}. Fix $0< \eta<1$; then for any $w\in V\cap b\Om$ there is a plurisuhharmonic function $\psi_w$ on $V\setminus \{w\}$ verifying \begin{enumerate} \item $|\psi_w(z)-\psi_w(z')|\lesssim |z-z'|^\eta$ \item $\psi_w(z)\le -G^\eta(|z-w|)$ \item $\psi_{\pi(z)}(z)\gtrsim-\delta_\Om(z)^{\eta}$ \end{enumerate} for any $z$ and $z'$ in $V\cap\bar\Om$. \end{theorem} \begin{remark} The construction of the plurisubharmonic peak functions on a pseudoconvex domain of finite type in $\C^2$ and a convex domain of finite type in $\C^n$ has been obtained by J. E. Fornaess and N. Sibony in \cite{FS89}. \end{remark} {\it Proof of Theorem~\ref{pshpeak}.} Using the argument in Section 3 and 4 of Diederich-Fornaess \cite{DF79}, we obtain that for any $\eta>0$, there exist an open neighborhood $V\subset U$, and a constant $L>0$, such that $$\psi_w(z)=-\left(-\rho(z,w) e^{L|z-w|^2}\right)^{\eta},$$ is a plurisubharmonic function on $V\cap \{z\in U:\rho(z,w)<0\}$. By the properties of $\rho$, we can check that $\psi_w$ satisfies (1), (2) and (3). That is the proof of Theorem~\ref{pshpeak}.\\ $\hfill\Box$ Now, we prove the lower bound for the Kobayashi metric by using the plurisubharmonic peak function. We state the theorem in a more general setting \begin{theorem}\Label{lowerbound} Let $\Om$ be a pseudoconvex domain in $\C^n$, $z_o$ be a given boundary point, $F_1$ and $F_2$ are postive functions such that $F_1$ is increasing and convex. Assume that there is a neighborhood $V$ of $z_o$ such that for each $w\in V\cap b\Om$, there is a plurisubharmonic function $\psi_w$ such that \begin{enumerate} \item[i)] $\psi_w(z)\le -F_1(|z-w|)$ \item[ii)] $\psi_{\pi(z)}(z)\ge -F_2(\delta_\Om(z))$ \end{enumerate} for $z\in U\cap\Om$. \\ Then $$K_\Om(z,X)\ge (F_1^*(F_2(\delta_\Om(z)))^{-1}|X|$$ for all $z\in V\cap \Om$, $X\in T_z^{1,0}{\C^n}$. \end{theorem} \begin{proof} We fix now a point $z\in V\cap \Om$, put $w=\pi(z)$ and assume that $g=(g_1,\dots,g_n):\overline\Delta\to \Om$ is a holomorphic map of the closed unit disc into $\Om$ with $g(0)=z$.\\ By applying the mean value inequality to the subharmonic function $\psi_w(g(t))$ on $\overline\Delta$ we get $$\psi_w(z)=\psi_w(g(0))\le \int_{0}^{1}\psi_w\circ g(e^{i2\pi\theta})d\theta$$ The hypothesis (ii) gives \begin{eqnarray}\Label{ineq1} F_2(\delta(z)))\ge \int_{0}^1-\psi_w\circ g(e^{i2\pi\theta})d\theta \end{eqnarray} We now use the hypothesis (i) of $\psi_w$, \begin{eqnarray}\Label{ineq2} \begin{split} \int^1_0-\psi_w\circ f(e^{i2\pi \theta})d\theta=&\int^1_0\big(-\psi_w\circ g(e^{i2\pi \theta})-F_1(|g_j(e^{i2\pi\theta})-w_j|)\big)d\theta\\ &+\int^1_0F_1\big(|g_j(e^{i2\pi\theta})-w_j|\big)d\theta\\ \ge &\int^1_0F_1\big(|g_j(e^{i2\pi\theta})-w_j|\big)d\theta. \end{split} \end{eqnarray} Using the Jensen inequality for the increasing, convex function $F_1$, we get $$F_1( |g_j'(0)|)\le F_1 \Big(\int^1_0|g_j(e^{i2\pi\theta})-w_j|d\theta\Big)\le \int^1_0 F_1\big(|g_j(e^{i2\pi\theta})-w_j|\big)d\theta .$$ Combining the above inequality with \eqref{ineq1} and \eqref{ineq2}, we obtain $$F_1(|g'_j(0)|)\le F_2(\delta_\Om(z)).$$ An immediate consequence of this is $$|g'_j(0)|\le F_1^*(F_2(\delta_\Om(z)).$$ By the definition of $K(z,X)$, we must have for all $X \in T^{1,0}\C^n$ $$K(z,X)\ge (F_1^*(F_2(\delta_\Om(z)))^{-1}|X|.$$ \end{proof} \section{Application to proper holomorphic maps } Let $\Om_1$ and $\Om_2$ be bounded domains in $\C^n$ with smooth boundary. Assume that $\Om_2$ is pseudoconvex of finite type at every boundary point. It is well-known that there is $\alpha>0$, such that every proper holomorphic map $\Psi: \Om_1\to \Om_2$ is H\"{o}lder continuous of order $\alpha$, in particular, $\Psi$ extends continuosly to $\bar\Om_1$. In this section, we prove a similar result for domains of infinite type. For this purpose we give a suitable estimate for generalized H\"older regularity.\\ Let $f$ be an increasing function such that $\underset{t\to+\infty}{\lim} f(t)=+\infty$. For $\Om\subset \C^n$, define the $f$-H\"older space on $\bar\Om$ by $$\Lambda^{f}(\bar\Om)=\{u : \no{u}_{\infty}+\sup_{z,z+h\in \overline\Om}f(|h|^{-1}) \cdot |u(z+h)-u(z)|<\infty \}$$ and set $$\no{u}_{f}= \no{u}_{\infty}+\sup_{z,z+h\in \overline\Om}f(|h|^{-1})\cdot |u(z+h)-u(z)|. $$ Note that the $f$-H\"older space include the standard H\"older space $\Lambda_\alpha(\bar\Om)$ by taking $f(t) = t^{\alpha}$ (so that $f(|h|^{-1}) = |h|^{-\alpha}$) with $0<\alpha<1$. \\ Before proving Theorem~\ref{fHolder}, we need a generalization of the Hardy-Littlewood Lemma. \begin{lemma} \Label{HL}Let $\Om$ be a bounded Lipschitz domain in $\R^N$ and let $\delta_{\Om}(x)$ denote the distance function from $x$ to the boundary of $\Om$. Let $G:\R^+\to\R^+$ be an increasing function such that $\dfrac{G(\delta)}{\delta}$ is decreasing and $\displaystyle\int_0^d\frac{G(\delta)}{\delta}d\delta<\infty$ for $d>0$ small enough. Let $u\in C^1(\Om)$ satisfy \begin{eqnarray}\Label{5.1} \quad |\nabla u(x)|\lesssim \frac{G(\delta_{\Om}(x))}{\delta_{\Om}(x)}~~\T{~~for every ~~} x\in \Om. \end{eqnarray} Then $u\in \Lambda^f(\bar\Om)$ where $f(d^{-1})=\Big(\displaystyle\int_0^d\frac{G(\delta)}{\delta}d\delta\Big)^{-1}.$ \end{lemma} The proof of this theorem can be found in \cite{Kha12}. \begin{remark} If $G(t)=t^\alpha$, Lemma \ref{HL} is the classical Hardy-Littlewood Lemma for a domain of finite type. \end{remark} {\it Proof of Theorem~\ref{fHolder}} Using Theorem~\ref{t1} for $\Om'$, the Schwarz-Pick lemma for the Kobayashi metric, and the upper bound of the Kobayashi metric, we obtain the following estimate \begin{eqnarray} g\left(\delta_{\Om'}^{-1}(\Psi(z))\right)|\Psi'(z)X|\lesssim K_{\Om'}(\Psi(z), \Psi'(z)X)\le K_\Om(z,X)\lesssim \delta^{-1}_\Om(z)|X| \end{eqnarray} for any $z\in\Om$ and $X\in T^{1,0}\C^n$. Moreover, by the fact that $-(-r)^\eta$ is strictly plurisubharmonic on $\Om$, one has $\delta_{\Om'}(\Psi(z))\lesssim \delta^\eta_\Om(z)$ for any $z\in\Om$ (Lemma 8 in \cite{DF79}). Therefore, $$|\Psi'(z)X|\lesssim \delta^{-1}_\Om(z) g^{-1}(\delta_\Om^{-\eta}(z))|X|$$ for any $z\in\Om$ and $X\in T^{1,0}\C^n$. Using Lemma~\ref{HL}, $\Psi$ can be extended to a $h$-H\"older continuous map $\hat\Psi:\bar\Om\to\bar\Om'$ with the rate $h(t)$ defined by \begin{eqnarray} \begin{split} (h(t))^{-1}:=&\int_0^{t^{-1}}\frac{d\delta}{\delta g(\delta^{-\eta})}=\frac{1}{\eta}\int_{t^\eta}^{\infty}\frac{db}{b g(b)}\\ =&\frac{1}{\eta}\int_{t\eta}^\infty\frac{1}{b} \left(\int_{b}^\infty \frac{da}{a f(a)}\right)db=\frac{1}{\eta}\int_{t^\eta}^\infty\frac{1}{af(a)} \left(\int_{t^\eta}^a \frac{db}{b}\right)da\\ =&\frac{1}{\eta}\int_{t^\eta}^\infty\frac{\ln a-\ln t^{\eta}}{af(a)}da=\frac{1}{\eta}(\tilde f(t^\eta))^{-1}.\\ \end{split} \end{eqnarray} The proof of Theorem~\ref{fHolder} is complete. $\hfill\Box$ \bibliographystyle{alpha}

\section*{Introduction} In this paper we introduce a framework of model theoretic forcing for metric structures, i.e., structures based on metric spaces. We use the language of infinitary continuous logic, which we define below. This is a variant of finitary continuous logic which is exposed in \cite{BenYaacov-Usvyatsov:localstability} or \cite{BenYaacov-Berenstein-Henson-Usvyatsov:metricstructures}. The model theoretic forcing framework introduced here is analogous to that developed by Keisler~\cite{Keisler:1973} for structures of the form considered in first-order model theory. The paper concludes with an application to separable quotients of Banach spaces. The long standing Separable Quotient Problem is whether for every nonseparable Banach space $X$ there exists a operator $T\colon X\to Y$ such that $T(X)$ is a separable, infinite dimensional Banach space. We prove the following result (Theorem~\ref{T:separable quotient}): If $X$ is an infinite dimensional Banach space and $T\colon X \to Y$ is a surjective operator with infinite dimensional kernel, then there exist Banach spaces $\Hat{X},\Hat{Y}$ and a surjective operator $\Hat{T}\colon\Hat{X}\to\Hat{Y}$ such that \begin{enumerate} \item $\Hat{X}$ has density character $\omega_1$, \item The range of $\Hat{T}$ is separable, \item $(X,Y,T)$ and $(\Hat{X},\Hat{Y},\Hat{T})$ are elementarily equivalent as metric structures. \end{enumerate} The paper is organized as follows. In \fref{sec:preliminaries} we introduce the syntax that will be used in the paper. In \fref{sec:forcing}, we introduce model theoretic forcing for metric structures. \fref{sec:special forcing properties} we focus our attention on two particular forcing properties. These properties are used in \fref{sec:omitting types} to prove the general Omitting Types Theorem. The last section, \fref{sec:separable quotients}, is devoted to the aforementioned application to separable quotients. For the exposition of the material we focus on one-sorted languages. However, as the reader will notice, the results presented here hold true, mutatis mutandi, for multi-sorted contexts. In fact, the structures used in the last section are multi-sorted. The authors are grateful to Yi Zhang for his encouragement and patience. \section{Preliminaries} \label{sec:preliminaries} Recall that if $f\colon (X,d) \to (X',d')$ is a mapping between two metric spaces, then $f$ is uniformly continuous if and only if there exists a mapping $\delta\colon (0,\infty)\to (0,\infty]$ such that for all $x,y \in X$ and $\varepsilon > 0$, \begin{gather} \label{eq:UnifCont} d(x,y) < \delta(\varepsilon) \Longrightarrow d'(f(x),f(y)) \leq \varepsilon. \end{gather} If \eqref{eq:UnifCont} holds, we say that $\delta$ is a \emph{uniform continuity modulus} and that $f$ \emph{respects} $\delta$. The choice of strict and weak inequalities here is so that the property of respecting $\delta$ be preserved under certain important constructions (e.g., completions and ultraproducts). Let $\delta'\colon (0,\infty) \to (0,\infty]$ be any mapping, and define: \begin{gather} \label{eq:GoodContMod} \delta(\varepsilon) = \sup \{\delta'(\varepsilon')\mid 0<\varepsilon'<\varepsilon\}. \end{gather} Then $\delta$ and $\delta'$ are equivalent as uniform continuity moduli, in the sense that a function $f$ respects $\delta$ if and only if it respects $\delta'$. In addition we have \begin{gather} \label{eq:GoodContMod2} \delta(\varepsilon) = \sup \{\delta(\varepsilon')\mid 0<\varepsilon'<\varepsilon\}, \end{gather} i.e., $\delta$ is increasing and continuous on the left. As a consequence, \fref{eq:UnifCont} is equivalent to the apparently stronger version: \begin{gather} \label{eq:UnifContStrict} d(x,y) < \delta(\varepsilon) \Longrightarrow d'(f(x),f(y)) < \varepsilon. \end{gather} From this point on, when referring to a uniform continuity modulus $\delta$, we mean one that satisfies \fref{eq:GoodContMod2}. In this section we introduce infinitary continuous formulas. For a general text regarding continuous structures and finitary continuous first order formulas we refer the reader to Sections 2 and 3 of~\cite{BenYaacov-Usvyatsov:localstability} or Sections~2--6 of~\cite{BenYaacov-Berenstein-Henson-Usvyatsov:metricstructures}. Recall that a continuous signature $\mathcal{L}$ consists of the following data: \begin{itemize} \item For each $n$, a set of $n$-ary function and predicate symbols. \item A distinguished binary predicate symbol $d$. \item For each $n$-ary symbol $s$ and $i<n$, a uniform continuity modulus for the $i$th argument denoted $\delta_{s,i}$. \end{itemize} A continuous $\mathcal{L}$-structure is a set $M$ equipped with interpretations of the symbols of the language: \begin{itemize} \item Each $n$-ary function symbol is interpreted by an $n$-ary function: $$f^M\colon M^n \to M.$$ \item Each $n$-ary predicate symbol is interpreted by a continuous $n$-ary predicate: $$P^M\colon M^n \to [0,1].$$ \item The interpretation $d^M$ of the distinguished symbol $d$ is a complete metric. \item For each $n$-ary symbol $s$ and $i < n$, the interpretation $s^M$, viewed as a function of its $i$th argument, respects the uniform continuity modulus $\delta_{s,i}$. \end{itemize} It is proved in \cite{BenYaacov-Usvyatsov:localstability} that the following system of connectives is \emph{full}: \[ x\mapsto \lnot x, \qquad x\mapsto \frac{x}{2}, \qquad (x,y)\mapsto x\mathbin{\mathpalette\dotminussym{}} y := \max(x-y,0) \] This means that for every $n \geq 1$, the family of functions from $[0,1]^n \to [0,1]$ which can be written using these three operations is dense in the class of all continuous functions $[0,1]^n \to [0,1]$. For the purposes of this paper (namely, to simplify the treatment of forcing, in \fref{sec:forcing}), it is convenient to use the connective $\dotplus$ instead of $\mathbin{\mathpalette\dotminussym{}}$. Note that this causes no loss in expressive power, since $x\mathbin{\mathpalette\dotminussym{}} y =\lnot(\lnot x\dotplus y)$. In this paper we extend the class of first-order continuous formulas by considering formulas that may contain the infinitary connectives $\bigwedge$ and $\bigvee$, where for a set of formulas $\Phi$, $\bigwedge_{\phi\in\Phi} \phi$ and $\bigvee_{\phi\in\Phi} \phi$ stand for $\sup\{\phi\mid\phi\in\Phi\} \phi$ and $\inf\{\phi\mid\phi\in\Phi\}$, respectively. Because of the infinitary nature of this language, in order to form formulas with these connectives, one needs to be particularly careful about the uniform continuity moduli of the terms and formulas with respect to each variable, denoted $\delta_{\tau,x}$ and $\delta_{\varphi,x}$, respectively; thus, we have the following definition. \begin{definition} Let $\mathcal{L}$ be a continuous signature. We define the formulas of $\mathcal{L}_{\omega_1,\omega}$. Simultaneously, for each variable $x$, each term $\tau$ and each formula $\phi$ of $\mathcal{L}_{\omega_1,\omega}$ we define uniform continuity moduli $\delta_{\tau,x}$ and $\delta_{\varphi,x}$. Both definitions are inductive. \begin{itemize} \item A variable is a term, with $\delta_{x,x} = \id$ and $\delta_{x,y} = \infty$ for $y\neq x$. \item If $f$ is an $n$-ary function symbol and $\tau_0,\dots,\tau_{n-1}$ are terms, then $f\tau_0\ldots\tau_{n-1}$ is a term. If $\tau$ is a term of this form, \begin{gather*} \delta_{\tau,x}(\varepsilon) = \sup_{\varepsilon_0+\ldots+\varepsilon_{n-1} < \varepsilon} \min \{\delta_{\tau_i,x}\circ\delta_{f,i}(\varepsilon_i)\mid i < n\}. \end{gather*} Here we follow the convention that $\delta(\infty) = \infty$. \item If $P$ is an $n$-ary predicate symbol and $\tau_0,\dots\tau_{n-1}$ are terms, then $P\tau_0\ldots\tau_{n-1}$ is a formula (called an \emph{atomic formula}). The definition of $\delta_{P\tau_0\ldots\tau_{n-1},x}$ is formally identical that of $\delta_{f\tau_0\ldots\tau_{n-1},x}$. \item If $\varphi$ and $\psi$ are formulas then so are $\lnot\varphi$, $\half \varphi$ and $\varphi \dotplus \psi$. We have: \begin{align*} \delta_{\lnot\varphi,x}(\varepsilon) & = \delta_{\varphi,x}(\varepsilon) \\ \delta_{\half \varphi,x}(\varepsilon) & = \delta_{\varphi,x}(2\varepsilon) \\ \delta_{\varphi\dotplus \psi,x}(\varepsilon) & = \sup_{\varepsilon_0+\varepsilon_1 < \varepsilon} \min \{\delta_{\varphi,x}(\varepsilon_0),\delta_{\psi,x}(\varepsilon_1)\}. \end{align*} \item Let $\Phi$ be a countable set of formulas in a finite tuple of free variables $\bar x$. For each variable $x$, let $\delta'_{\bigwedge\Phi,x} = \inf_{\varphi\in\Phi} \delta_{\varphi,x}\colon (0,\infty) \to [0,\infty]$. If $\delta'_{\bigwedge\Phi,x}(\varepsilon) > 0$ for all $\varepsilon > 0$ and $x \in \bar x$, then $\bigwedge\Phi$ is a formula, also denoted $\bigwedge_{\varphi\in\Phi} \varphi$. Its uniform continuity moduli are given by \[ \delta_{\bigwedge\Phi,x}(\varepsilon) = \sup\{\delta'_{\bigwedge\Phi,x}(\varepsilon')\colon 0<\varepsilon'<\varepsilon\}, \] so that \fref{eq:GoodContMod2} is satisfied. \item If $\varphi$ is a formula and $x$ a variable, then $\inf_x \varphi$ is a formula. For $y\neq x$ we have $\delta_{\inf_x\varphi,y} = \delta_{\varphi,y}$, while $\delta_{\inf_x\varphi,x} = \infty$. \end{itemize} \end{definition} \begin{notation} Rather than putting $\bigvee$ and $\sup$ in our language we define them as abbreviations: \begin{align*} \bigvee\Phi & := \lnot\bigwedge_{\varphi\in\Phi}\lnot\varphi \\ \sup_x \varphi & := \lnot\inf_x\lnot\varphi. \end{align*} \end{notation} If $M$ is an $\mathcal{L}$-structure and $\varphi(x_0,\dots,x_{n-1}) \in \mathcal{L}_{\omega_1,\omega}$, one constructs the interpretation $\varphi^M\colon M^n \to [0,1]$ in the obvious manner. By induction on the structure of $\varphi$ one also shows that for each variable $x$, $\varphi^M$ is uniformly continuous in $x$ respecting $\delta_{\varphi,x}$. Finitary continuous first order formulas, as defined in \cite{BenYaacov-Usvyatsov:localstability} and \cite{BenYaacov-Berenstein-Henson-Usvyatsov:metricstructures}, are constructed in the same manner, with the exclusion of the infinitary connectives $\bigwedge$ and $\bigvee$ (i.e., only using the connectives $\lnot,\half,\dotplus$, or equivalently $\lnot,\half,\mathbin{\mathpalette\dotminussym{}}$). We observe that $\varphi\land\psi$ is equivalent to $\varphi \mathbin{\mathpalette\dotminussym{}} (\varphi \mathbin{\mathpalette\dotminussym{}} \psi)$, so finitary instances of $\bigwedge$ and $\bigvee$ are allowed there as well. The set of all such formulas is denoted $\mathcal{L}_{\omega,\omega}$. \begin{definition} Let $\mathcal{L}$ be a continuous signature and let $\phi$ be an $\mathcal{L}_{\omega_1,\omega}$-formula. The set of \emph{subformulas} of $\phi$ denoted $\operatorname{sub}(\phi)$, is defined inductively as follows. \begin{itemize} \item If $P$ is a predicate symbol and $\tau_0,\dots\tau_{n-1}$ are terms, then $\operatorname{sub}(P\tau_0\ldots\tau_{n-1})=\{P\tau_0\ldots\tau_{n-1}\}$. \item $\operatorname{sub}(\lnot \phi)=\{\lnot\phi\}\cup\operatorname{sub}(\phi)$ and $\operatorname{sub}(\half \phi)=\{\half\phi\}\cup\operatorname{sub}(\phi)$. \item $\operatorname{sub}(\phi \dotplus \psi)=\{\,\phi \dotplus \psi\,\}\cup\operatorname{sub}(\phi)\cup\operatorname{sub}(\psi)$. \item $\operatorname{sub}(\bigwedge_{\phi\in\Phi} \phi) = \{\,\bigwedge_{\phi\in\Phi} \phi\,\}\cup \bigcup_{\phi\in \Phi}\operatorname{sub}(\phi)$. \item $\operatorname{sub}(\inf_x\phi) = \{\,\inf_x\phi\,\}\cup \operatorname{sub}(\phi)$. \end{itemize} \end{definition} $\mathcal{L}_{\omega_1,\omega}$ need not be countable if $\mathcal{L}$ is countable. Nevertheless, it is often sufficient to work with countable fragments of $\mathcal{L}_{\omega_1,\omega}$: \begin{definition} A \emph{fragment} of $\mathcal{L}_{\omega_1,\omega}$ is subset of $\mathcal{L}_{\omega_1,\omega}$ which contains all atomic formulas and is closed under subformulas and substitution of terms for free variables. \end{definition} \begin{remark} Every countable subset of $\mathcal{L}_{\omega_1,\omega}$ is contained in a countable fragment of $\mathcal{L}_{\omega_1,\omega}$. \end{remark} For the next three sections (that is, the rest of the paper minus the last section), $\mathcal{L}$ will denote a fixed countable continuous signature, and $\mathcal{L}_A$ will denote a fixed countable fragment of $\mathcal{L}_{\omega_1,\omega}$. We will let $C = \{c_i \mid i < \omega\}$ be a set of new constant symbols, and $\mathcal{L}(C) = \mathcal{L}\cup C$. An $\mathcal{L}(C)$-structure $M$ will be called \emph{canonical} if the set $\{c_i^M \mid i < \omega\}$ is dense in $M$. By $\mathcal{L}_A(C)$ we will denote the smallest countable fragment of $\mathcal{L}_{\omega_1,\omega}(C)$ that contains $\mathcal{L}_A$; notice that $\mathcal{L}_A(C)$ is obtained allowing closing $\mathcal{L}_A$ under substitution of constant symbols from $C$ for free variables. We will also use the following notation: \begin{itemize} \item The set of all sentences in $\mathcal{L}_A(C)$ will be denoted $\mathcal{L}_A^s(C)$. \item The set of all atomic sentences in $\mathcal{L}_A(C)$ will be denoted $\mathcal{L}_A^{as}(C)$. \item The set of variable-free terms in $\mathcal{L}(C)$ will be denoted $\mathcal{T}(C)$. \end{itemize} \section{Forcing} \label{sec:forcing} \begin{definition} A \emph{forcing property} for $\mathcal{L}_A$ is a triplet $(\mathbb{P},\leq,f)$ where $(\mathbb{P},\leq)$ is a partially ordered set. The elements of $\mathbb{P}$ are called \emph{conditions}. For each condition $p$, $f$ assigns a mapping $f_p \colon \mathcal{L}_A^{as}(C) \to [0,1]$ satisfying the following conditions. \begin{enumerate} \item $p \leq q$ implies $f_p \leq f_q$ i.e., $f_p(\varphi) \leq f_q(\varphi)$ for all $\varphi \in \mathcal{L}_A^{as}(C)$. \item Given $p\in \mathbb{P}$, $\varepsilon>0$, $\tau,\sigma \in \mathcal{T}(C)$, and an atomic $\mathcal{L}(C)$-formula $\varphi(x)$ there are $q \leq p$ and $c \in C$ such that: \begin{gather*} f_q(d(\tau,c)) < \varepsilon,\\ f_q(d(\tau,\sigma)) < f_p(d(\sigma,\tau)) + \varepsilon,\\ \intertext{and if $f_p(d(\tau,\sigma)) < \delta_{\varphi,x}(\varepsilon)$,} f_q(\varphi(\sigma)) < f_p(\varphi(\tau)) + \varepsilon. \end{gather*} \end{enumerate} \end{definition} For the rest of this section, $(\mathbb{P},\leq,f)$ will denote a fixed forcing property. \begin{definition} \label{D:forcing} Let $p \in \mathbb{P}$ be a condition and $\varphi \in \mathcal{L}_A^s(C)$ a sentence. We define $F_p(\varphi) \in [0,1]$ by induction on $\varphi$. For $\varphi$ atomic, \[ F_p(\varphi) = f_p(\varphi).\\ \] Otherwise, \[ \begin{array}{lcl} F_p(\lnot\varphi) & = & \lnot \inf_{q\leq p} F_q(\varphi) \\ F_p(\half \varphi) & = & \half F_p(\varphi) \\ F_p(\varphi \dotplus \psi) & = & F_p(\varphi) \dotplus F_p(\psi) \\ F_p(\bigwedge\Phi) & = & \inf_{\varphi\in\Phi} F_p(\varphi) \\ F_p(\inf_x \varphi(x)) & = & \inf_{c\in C} F_p(\varphi(c)). \end{array} \] If $r \in \mathbb{R}$ and $F_p(\varphi) < r$ we say that $p$ \emph{forces} that $\varphi < r$, in symbols $p \Vdash \varphi<r$. \end{definition} \begin{remark} \label{R:forcing inequality properties} Let $p \in \mathbb{P}$ be a condition, $\varphi \in \mathcal{L}_A^s(C)$ a sentence, and $r \in \mathbb{R}$. Then, \[ \begin{array}{lcl} p \Vdash \varphi<r &\Longleftrightarrow & f_p(\varphi) < r, \text{ if $\varphi$ is atomic}\\ p \Vdash \half \varphi < r & \Longleftrightarrow & p \Vdash \varphi < 2r\\ p \Vdash \lnot\varphi < r & \Longleftrightarrow & (\exists s > 1-r)(\forall q \leq p)(q \nVdash \varphi < s)\\ p \Vdash (\varphi \dotplus \psi) < r & \Longleftrightarrow & (\exists s)(p \Vdash \varphi < s \text{ and } p \Vdash \psi < r-s)\\ p \Vdash \bigwedge\Phi < r & \Longleftrightarrow & (\exists\varphi\in\Phi)(p \Vdash \varphi < r)\\ p \Vdash \inf_x\varphi(x) < r & \Longleftrightarrow & (\exists c\in C)(p \Vdash \varphi(c) <r). \end{array} \] \end{remark} \begin{remark} The forcing relation $\Vdash$ can be defined inductively, without reference to the function $F_p(\phi)$, by the list of equivalences in the preceding remark. One can then define $F_p(\varphi)$ as $\inf\{\,r\in\mathbb{R}\mid p \Vdash \varphi<r \,\}$. \end{remark} The following basic properties will be used many times. \begin{lemma} \label{lem:ForcingProps} For all $p,\varphi$, \begin{enumerate} \item $F_p(\varphi) \in [0,1]$. \item $q \leq p \Longrightarrow F_q(\varphi) \leq F_p(\varphi)$. \item $F_p(\varphi) + F_p(\lnot\varphi) \geq 1$. \end{enumerate} \end{lemma} \begin{proof} The first two items are by induction on the structure of $\varphi$. The last one follows directly from the definition. \end{proof} \begin{definition} We also define $F^w_p$ by: \begin{gather*} F^w_p(\varphi) = \sup_{q\leq p}\inf_{q'\leq q} F_{q'}(\varphi). \end{gather*} If $r \in \mathbb{R}$ and $F^w_p(\varphi) < r$ we say that $p$ \emph{weakly forces} that $\varphi < r$, in symbols $p \Vdash^w \varphi<r$. \end{definition} By \fref{lem:ForcingProps}, $F^w_p(\varphi) \leq F_p(\varphi)$. \begin{remark} The weak forcing relation $\Vdash^w$ can be defined without reference to the function $F^w_p$ as follows: $p \Vdash^w \varphi<r$ if and only if $(\exists s < r)(\forall q\leq p)(\exists q'\leq q)(q' \Vdash \varphi<s)$. We can then define $F^w_p(\varphi)$ as $\inf\{\,r\mid p \Vdash^w \varphi<r \,\}$. \end{remark} \begin{lemma} \label{lem:WForceRobust} Let $p \in \mathbb{P}$, $\varphi \in \mathcal{L}_A^s(C)$ and $r \in \mathbb{R}$. Then \begin{gather*} F^w_p(\varphi) = \sup_{q\leq p} F^w_q(\varphi) = \sup_{q\leq p}\inf_{q'\leq q} F^w_{q'}(\varphi). \end{gather*} \end{lemma} \begin{proof} That $F^w_p(\varphi) = \sup_{q\leq p} F^w_q(\varphi)$ follows easily from the definitions, and $\sup_{q\leq p} F^w_q(\varphi) \geq \sup_{q\leq p}\inf_{q'\leq q} F^w_{q'}(\varphi)$ is immediate. Finally: \begin{align*} \sup_{q\leq p}\inf_{q'\leq q} F^w_{q'}(\varphi) & = \sup_{q\leq p}\inf_{q'\leq q} \sup_{q''\leq q'} \inf_{q'''\leq q''} F_{q'''}(\varphi) \geq \sup_{q\leq p}\inf_{q'\leq q} \inf_{q'''\leq q'} F_{q'''}(\varphi) \\ & = \sup_{q\leq p}\inf_{q'\leq q} F_{q'}(\varphi) = F^w_p(\varphi). \qedhere \end{align*} \end{proof} \begin{proposition} \label{prp:WForceInd} The weak forcing function $F^w$ obeys the following inductive rules: \[ \begin{array}{lcl} F^w_p(\lnot\varphi) & = & \lnot \inf_{q\leq p} F^w_q(\varphi) \\ F^w_p(\half \varphi) & = & \half F^w_p(\varphi) \\ F^w_p(\varphi \dotplus \psi) & = & \sup_{q\leq p} \inf_{q'\leq q} F^w_{q'}(\varphi) \dotplus F^w_{q'}(\psi) \\ F^w_p(\bigwedge\Phi) & = & \sup_{q\leq p} \inf_{q'\leq q} \inf_{\varphi\in\Phi} F^w_{q'}(\varphi) \\ F^w_p(\inf_x \varphi(x)) & = & \sup_{q\leq p} \inf_{q'\leq q} \inf_{c\in C} F^w_{q'}(\varphi(c)). \end{array} \] \end{proposition} \begin{proof} For $\lnot\varphi$ and $\half\varphi$ this follows from a straightforward calculation. For example: \begin{align*} F^w_p(\lnot\varphi) & = \sup_{q\leq p}\inf_{q'\leq q} F_{q'}(\lnot\varphi) = \sup_{q\leq p}\inf_{q'\leq q} \lnot \inf_{q''\leq q'} F_{q'}(\varphi) \\ & = \lnot \inf_{q\leq p}\sup_{q'\leq q} \inf_{q''\leq q'} F_{q'}(\varphi) = \lnot \inf_{q\leq p} F^w_q(\varphi), \end{align*} For the other three, the inequality $\geq$ is obtained substituting the definition of $F^w_p$ on the left hand side and using the fact that $F_p \geq F^w_p$. For $\leq$, we first use \fref{lem:WForceRobust} to replace each occurrence of $F^w_p$ on the left hand side with $\sup_{q<p}\inf_{q'\leq q} F^w_{q'}$. Thus, it will suffice to show that: \[ \begin{array}{lcl} F^w_p(\varphi \dotplus \psi) & \leq & F^w_p(\varphi) \dotplus F^w_p(\psi) \\ F^w_p(\bigwedge\Phi) & \leq & \inf_{\varphi\in\Phi} F^w_p(\varphi) \\ F^w_p(\inf_x \varphi(x)) & \leq & \inf_{c\in C} F^w_p(\varphi(c)). \end{array} \] For $\dotplus$ assume $F^w_p(\varphi) = r$ and $F^w_p(\psi) = s$. Then for all $\varepsilon>0$ and for all $q \leq p$ there is $q'_0\leq q$ such that $F_{q'_0}(\varphi) < r+\varepsilon$, and as $q'_0 \leq p$ there is $q' \leq q'_0$ such that $F_{q'}(\psi) < s+\varepsilon$. Then $F_{q'}(\varphi \dotplus \psi) < r+s + 2\varepsilon$, yielding $F^w_{q'}(\varphi \dotplus \psi) \leq r+s$. For $\bigwedge\Phi$ and $\inf_x \varphi(x)$ it's a straightforward quantifier exchange argument, e.g.: \begin{align*} F^w_p(\bigwedge\Phi) & = \sup_{q\leq p}\inf_{q'\leq q} \inf_{\varphi\in\Phi} F_p(\varphi) \leq \inf_{\varphi\in\Phi} \sup_{q\leq p}\inf_{q'\leq q} F_p(\varphi) = \inf_{\varphi\in\Phi} F^w(\varphi). \qedhere \end{align*} \end{proof} \begin{lemma} \label{lem:WForceDense} For all $p \in \mathbb{P}$ and $\tau$: $F^w_p(\inf_x d(\tau,x)) = 0$. \end{lemma} \begin{proof} If not then $F^w_p(\inf_x d(\tau,x)) = \sup_{q\leq p} \inf_{q'\leq q} \inf_{c\in C} f_p(d(\tau,x)) > 0$. But this contradicts the definition of forcing property. \end{proof} \begin{definition} \label{dfn:Generic} A nonempty $G \subseteq \mathbb{P}$ is \emph{generic} if: \begin{enumerate} \item It is directed downwards, i.e., for all $p,q \in G$ there is $p' \in G$ such that $p' \leq p,q$. \item It is closed upwards, i.e., if $p \in G$ and $q \geq p$ then $q \in G$. \item For every $\varphi \in \mathcal{L}_A^s(C)$ and $r > 1$ there is $p \in G$ such that $F_p(\varphi) + F_p(\lnot\varphi) < r$. \end{enumerate} If $G$ is a generic set and $\varphi \in \mathcal{L}_A^s(C)$ we define $$\varphi^G = \inf_{p\in G} F_p(\varphi).$$ \end{definition} \begin{proposition} \label{P:generic set} Every condition belongs to a generic set. \end{proposition} \begin{proof} Fix $p\in\mathbb{P}$. Let $(\,(r_n,\varphi_n)\colon n < \omega\,)$ enumerate all pairs $(r,\varphi)$, where $r \in \mathbb{Q}$, $r>1$, and $\varphi \in \mathcal{L}_A^s(C)$. Construct a sequence $p_0 \geq p_1 \geq \ldots \geq p_n \geq \ldots$ in $\mathbb{P}$ as follows. We start with $p_0 = p$. Assume $p_n$ has already been chosen. By definition $F_{p_n}(\lnot\varphi_n) + \inf_{q\leq p_n} F_{q}(\varphi_n) = 1 < r_n$, so we can choose $p_{n+1}\leq p_n$ such that $F_{p_n}(\lnot\varphi_n) + F_{p_{n+1}}(\varphi_n) < r_n$, whereby $F_{p_{n+1}}(\lnot\varphi_n) + F_{p_{n+1}}(\varphi_n) < r_n$. Define $$G = \{\,q \in \mathbb{P}\mid q \geq p_n\text{ for some }n\,\}.$$ Then $G$ is generic, and $p \in G$. \end{proof} \begin{lemma} \label{lem:GenSetWForce} Let $G$ be generic and $\varphi \in \mathcal{L}_A^s(C)$. Then $\varphi^G = \inf_{p\in G} F^w_p(\varphi)$. \end{lemma} \begin{proof} The inequality $\geq$ is immediate since $F^w_p(\phi)\leq F_p(\phi)$. For the other, assume $\varphi^G > \inf_{p\in G} F^w_p(\varphi)$, so there are $\varepsilon > 0$ and $p \in G$ such that $\varphi^G - \varepsilon > F^w_p(\varphi)$. As $G$ is generic there is $q \in G$ such that $F_q(\varphi) + F_q(\lnot\varphi) < 1+\varepsilon$, and as $p \in G$ we may assume $q \leq p$. We obtain $$F^w_p(\varphi) \geq \inf_{q'\leq q}F_{q'}(\varphi) = 1-F_q(\lnot\varphi) > F_q(\varphi) - \varepsilon \geq \varphi^G - \varepsilon > F^w_p(\varphi),$$ a contradiction. \end{proof} \begin{lemma} \label{lem:GenSetNeg} If $G$ is generic and $\varphi \in \mathcal{L}_A^s(C)$, then $(\lnot\varphi)^G = 1-\varphi^G$. \end{lemma} \begin{proof} From \fref{lem:ForcingProps} we have $\varphi^G+(\lnot\varphi)^G \geq 1$, while $\varphi^G+(\lnot\varphi)^G \leq 1$ follows from \fref{dfn:Generic}. \end{proof} \begin{lemma} \label{lem:GenSetMetric} Let $G$ be a generic set and $\tau,\sigma \in \mathcal{T}(C)$. Then: \begin{enumerate} \item For every $\varepsilon > 0$ there is $c_{\tau,\varepsilon,G} \in C$ such that $d(\tau,c_{\tau,\varepsilon,G})^G < \varepsilon$. \item $d(\tau,\sigma)^G = d(\sigma,\tau)^G$. \item For every atomic $\mathcal{L}(C)$-formula $\varphi(x)$, if $d(\tau,\sigma)^G < \delta_{\varphi,x}(\varepsilon)$ then $|\varphi(\tau)^G-\varphi(\sigma)^G| < \varepsilon$. \end{enumerate} \end{lemma} \begin{proof} For (i), observe that $(\inf_x d(\tau,x))^G =0$ by \fref{lem:WForceDense} and \fref{lem:GenSetWForce}, so there is $p \in G$ such that $F_p(\inf_x d(\tau,x)) < \varepsilon$, and thus there exists $c \in C$ such that $d(\tau,c)^G \leq F_p(d(\tau,c)) < \varepsilon$. The other two statements follow directly from \fref{lem:GenSetWForce} and the definition of forcing property. \end{proof} \begin{lemma} \label{L:generic model} Let $M_0^G$ be the term algebra $\mathcal{T}(C)$ equipped with the natural interpretation of the function symbols, and interpreting the predicate symbols by: $P^{M_0^G}(\bar \tau) = P(\bar \tau)^G$. Then $M_0^G$ is a pre-$\mathcal{L}(C)$-structure, and its completion $M^G$ is a canonical structure. \end{lemma} \begin{proof} First we use \fref{lem:GenSetMetric} to show that $d^{M_0^G}$ is a pseudometric. Symmetry is \fref{lem:GenSetMetric}(ii). The triangle inequality follows from \fref{lem:GenSetMetric}(iii), keeping in mind that $\delta_{d(x,\sigma),x} = \id$. That $d^{M_0^G}(\tau,\tau) = 0$ follows from the triangle inequality and \fref{lem:GenSetMetric}(i). Finally, by \fref{lem:GenSetMetric}(iii), every symbol respects its uniform continuity modulus. Thus $M_0^G$ is a pre-structure, and we can define $M^G$ to be its completion. That $C^{M^G}$ is dense in $M^G$ now follows from \fref{lem:GenSetMetric}(i). \end{proof} \begin{theorem} \label{T:generic model} For all $\varphi \in \mathcal{L}_A^s(C)$ we have $\varphi^{M^G} = \varphi^G$. \end{theorem} \begin{proof} By induction on $\varphi$: \begin{enumerate} \item For $\varphi$ atomic, this is immediate from the construction of $M^G$. \item For $\half \varphi$, $\varphi \dotplus \psi$ and $\bigwedge\Phi$, this is immediate from the definition of forcing and the induction hypothesis. \item For $\lnot\varphi$, this is immediate from \fref{lem:GenSetNeg} and the induction hypothesis. \item For $\inf_x \varphi(x)$, it follows from the definition of forcing and the induction hypothesis that $(\inf_x \varphi)^G = \inf\{\varphi(c)^{M^G}\mid c \in C\}$. Since $C^{M^G}$ is dense in $M^G$ and $\varphi(x)^{M^G}$ is uniformly continuous in $x$, the latter is equal to $(\inf_x \varphi)^{M^G}$. \qedhere \end{enumerate} \end{proof} \section{The forcing Properties $\mathcal{P}(\mathcal{M})$ and $\mathcal{P}(\mathcal{M},\Sigma)$} \label{sec:special forcing properties} If $\mathcal{M}$ is class of $\mathcal{L}$-structures, we denote by $\mathcal{M}(C)$ the class of all structures of the form $(M, a_c)_{c\in C_0}$, where $M$ is in $\mathcal{M}$ and $C_0$ is a finite subset of $C$; such a structure is regarded naturally as an $\mathcal{L}(C_0)$-structure by letting $a_c$ be the interpretation of $c$ in $M$, for each $c\in C_0$. Let $\Sigma$ be a class of formulas of $\mathcal{L}_A$ that contains all the atomic formulas and is closed under subformulas, and let $\Sigma(C)$ denote the subset of $\mathcal{L}(C)$ obtained from formulas $\phi$ in $\Sigma$ by replacing finitely many free variables of $\phi$ with constant symbols from $C$. The forcing property $\mathcal{P}(\mathcal{M},\Sigma)$ is defined as follows. The conditions of $\mathcal{P}(\mathcal{M},\Sigma)$ are the finite sets of the form \[ \{\,\phi_1<r_1,\dots,\phi_n<r_n\,\}, \] where $\phi_1,\dots,\phi_n\in\Sigma(C)$ and there exist $M\in \mathcal{M}(C)$ such that $\phi_i^M<r_i$, for $i=1,\dots, n$. The partial order $\leq$ on conditions is reverse inclusion, i.e., if $p,q$ are conditions of $\mathcal{P}(\mathcal{M},\Sigma)$, then $p\leq q$ if and only $p\supseteq q$. If $p$ is a condition of $\mathcal{P}(\mathcal{M}_\Delta,\Sigma)$ and $\phi$ is an atomic sentence of $\mathcal{L}(C)$, we define \[ f_p(\phi)= \begin{cases} \min \{r\leq 1 \mid \phi<r\in p\} ,\quad &\text{if $\{r\leq 1 \mid \phi<r\in p\}\neq \emptyset$,}\\ 1,\quad&\text{otherwise}. \end{cases} \] When $\Sigma$ is the set of all atomic $\mathcal{L}$-formulas, the forcing property $\mathcal{P}(\mathcal{M},\Sigma)$ is denoted simply $\mathcal{P}(\mathcal{M})$. The main result of this section is Proposition~\ref{P:w-forcing in P(M,Sigma)}, below, which characterizes weak forcing for the forcing property $\mathcal{P}(\mathcal{M},\Sigma)$; for the proof, we need two lemmas. \begin{definition} We extend the definition of $f_p$ above to all sentences of $\Sigma(C)$: \[ H_p(\varphi) = \begin{cases} \min \{r\leq 1 \mid \phi<r\in p\} ,\quad &\text{if $\{r\leq 1 \mid \phi<r\in p\}\neq \emptyset$,}\\ 1,\quad&\text{otherwise}. \end{cases} \] We define $H^w_p$ accordingly: $H^w_p(\varphi) = \sup_{q\leq p} \inf_{p\leq q} H_p(\varphi)$. \end{definition} Clearly if $q \leq p$ then $H_q(\varphi) \leq H_p(\varphi)$ and $H^w_q(\varphi) \leq H^w_p(\varphi)$, whereby for all $p$: $H^w_p(\varphi) \leq H_p(\varphi)$. \begin{lemma} \label{lem:WPMSChar} For all $p \in \cP(\cM,\Sigma)$ and $\varphi \in \Sigma(C)$: \begin{align*} H^w_p(\varphi) & = \inf \{r \in [0,1] \mid (\forall q \leq p)(q \cup \{\varphi<r\} \in \cP(\cM,\Sigma))\} \\ & = \sup \{r \in [0,1]\mid p\cup\{\lnot\varphi < 1-r\} \in \cP(\cM,\Sigma)\} \end{align*} (Here $\inf \emptyset = 1$, $\sup \emptyset = 0$.) \end{lemma} \begin{proof} The first equality is a mere rephrasing: $H^w_p(\varphi) \leq r$ if and only if $\inf_{q'\leq q} H_p(\varphi) \leq r$ for all $q \leq p$, i.e., if and only if $q \cup \{\varphi<r\} \in \cP(\cM,\Sigma)$ for all $q \leq p$. For the second equality: Assume first that $q = p\cup\{\lnot\varphi < 1-r\} \in \cP(\cM,\Sigma)$. Then $q \leq p$ but $q \cup \{\varphi<r\} \notin \cP(\cM,\Sigma)$. This gives $\geq$. Now assume $p\cup\{\lnot\varphi < 1-r\} \notin \cP(\cM,\Sigma)$. Then $p\cup\{\lnot\varphi < 1-r\}$ cannot be realized in the given class. Thus, for every $q \leq p$, as $q$ can be realized, it is realized in a model where $\varphi \leq r$. Thus $q \cup \{\varphi<s\} \in \cP(\cM,\Sigma)$ for all $q \leq p$ and $s > r$. This gives $\leq$. \end{proof} \begin{proposition} \label{prp:WPMSInd} The functions $H^w_p$ satisfy the properties stated for $F^w_p$ in \fref{lem:WForceRobust} and \fref{prp:WForceInd}, i.e.: \[ \begin{array}{lcl} H^w_p(\varphi) & = & \sup_{q\leq p} H^w_q(\varphi) = \sup_{q\leq p}\inf_{q'\leq q} H^w_{q'}(\varphi) \\ H^w_p(\lnot\varphi) & = & \lnot \inf_{q\leq p} H^w_q(\varphi) \\ H^w_p(\half \varphi) & = & \half H^w_p(\varphi) \\ H^w_p(\varphi \dotplus \psi) & = & \sup_{q\leq p} \inf_{q'\leq q} H^w_{q'}(\varphi) \dotplus H^w_{q'}(\psi) \\ H^w_p(\bigwedge\Phi) & = & \sup_{q\leq p} \inf_{q'\leq q} \inf_{\varphi\in\Phi} H^w_{q'}(\varphi) \\ H^w_p(\inf_x \varphi(x)) & = & \sup_{q\leq p} \inf_{q'\leq q} \inf_{c\in C} H^w_{q'}(\varphi(c)). \end{array} \] \end{proposition} \begin{proof} The first property is proved precisely as in \fref{lem:WForceRobust}. For $\lnot$: it follows from \fref{lem:WPMSChar} that $H^w_p(\lnot\varphi) = \lnot\inf_{q\leq p}H_q(\varphi)$, and we conclude as in the proof of \fref{prp:WForceInd}. For $\half$: observe that $q\cup\{\varphi<r\} \in \cP(\cM,\Sigma)$ if and only if $q\cup\{\half \varphi < \half r\} \in \cP(\cM,\Sigma)$ and apply \fref{lem:WPMSChar}. For the last three we reduce as in the proof of \fref{prp:WForceInd} to showing that: \[ \begin{array}{lcl} H^w_p(\varphi \dotplus \psi) & \leq & H^w_p(\varphi) \dotplus H^w_p(\psi) \\ H^w_p(\bigwedge\Phi) & \leq & \inf_{\varphi\in\Phi} H^w_p(\varphi) \\ H^w_p(\inf_x \varphi(x)) & \leq & \inf_{c\in C} H^w_p(\varphi(c)). \end{array} \] For $\dotplus$ this follows from \fref{lem:WPMSChar}. For $\bigwedge$ and $\inf$ the quantifier exchange argument from proof of the corresponding items in \fref{prp:WForceInd} works here too. \end{proof} \begin{proposition} \label{P:w-forcing in P(M,Sigma)} Suppose that $p$ is a condition in the forcing property $\mathcal{P}(\mathcal{M},\Sigma)$ and $\sigma$ is a sentence of $\Sigma(C)$. Then $F^w_p(\sigma) = H^w_p(\sigma)$. \end{proposition} \begin{proof} For atomic $\sigma$ the equality is immediate. We then proceed by induction on $\sigma$, noting that \fref{prp:WForceInd} on the one hand and \fref{prp:WPMSInd} on the other tell us that $F^w_p$ and $H^w_p$ obey the same inductive definitions. \end{proof} \section{Generic Models and \ensuremath{\sup\bigwedge\inf}-Formulas} \label{sec:omitting types} Recall from \fref{sec:preliminaries} that the expressions $\bigvee\Phi$ and $\sup_x \varphi$ are regarded abbreviations of $\lnot\bigwedge_{\varphi\in\Phi}\lnot\varphi$ and $\lnot\inf_x\lnot\varphi$ respectively. \begin{proposition} \label{P:disjunctions} Let $(\mathbb{P},\leq,f)$ be a forcing property for $\mathcal{L}_A(C)$ and let $p\in \mathbb{P}$. Then \begin{enumerate} \item $F_p(\bigvee\Phi)=\sup_{\phi\in\Phi}F_p^w(\phi)$. \item $F_p(\sup_x\phi(x))=\sup_{c\in C} F_p^w(\phi(c))$. \end{enumerate} \end{proposition} \begin{proof} The proofs are straightforward applications of the definitions: for~(i), \begin{align*} F_p(\bigvee \Phi)= F_p(\lnot\bigwedge_{\phi\in\Phi} \lnot\phi) &=\lnot \inf_{q\leq p}F_q(\bigwedge_{\phi\in\Phi}\lnot\phi)\\ &=\lnot \inf_{q\leq p}\inf_{\phi\in\Phi}F_q(\lnot\phi)\\ &=\lnot \inf_{q\leq p}\inf_{\phi\in\Phi}\lnot\inf_{q'\leq q} F_{q'}(\phi)\\ &=\sup_{q\leq p}\sup_{\phi\in\Phi}\inf_{q'\leq q} F_{q'}(\phi)\\ &=\sup_{\phi\in\Phi}\sup_{q\leq p}\inf_{q'\leq q} F_{q'}(\phi)\\ &=\sup_{\phi\in\Phi}F_p^w(\phi), \end{align*} and for~(ii), \begin{align*} F_p(\sup_x\phi(x))= F_p(\lnot\inf_x \lnot\phi(x))&=\lnot\inf_{q\leq p}F_q(\inf_x \lnot\phi(x))\\ &=\lnot\inf_{q\leq p}\inf_{c\in C}F_q(\lnot\phi(c))\\ &=\lnot\inf_{q\leq p}\inf_{c\in C}\lnot\inf_{q'\leq q}F_{q'}(\phi(c))\\ &=\sup_{q\leq p}\sup_{c\in C}\inf_{q'\leq q}F_{q'}(\phi(c))\\ &=\sup_{c\in C}\sup_{q\leq p}\inf_{q'\leq q}F_{q'}(\phi(c))\\ &=\sup_{c\in C}F_{q'}^w(\phi(c)). \qedhere \end{align*} \end{proof} \begin{notation} If $\Phi$ is a finite set of formulas, say $\Phi=\{\phi_1,\dots,\phi_n\}$, we write \[ \phi_1\land\dots\land\phi_n\quad\text{and}\quad \phi_1\lor\dots\lor\phi_n \] as abbreviations of $\bigwedge\Phi$ and $\bigvee\Phi$, respectively. \end{notation} \begin{proposition} \label{P:finite disjunctions} If $(\mathbb{P},\leq,f)$ is a forcing property for $\mathcal{L}_A(C)$ and $p\in \mathbb{P}$, then \[ F_p(\phi_1\lor\dots\lor\phi_n)=\max_{i}F_p^w(\phi_i). \] \end{proposition} \begin{proof} By Proposition~\ref{P:disjunctions}. \end{proof} \begin{definition} Let $\Sigma$ be a class of formulas of $\mathcal{L}_A$ which contains all atomic formulas and is closed under subformulas. A \emph{$\sup\bigwedge\inf$-formula over $\Sigma$} is an $\mathcal{L}_A$-formula of the form \[ \sup_{x_1}\dots\sup_{x_m}\bigwedge_{n<\omega}\inf_{y_1}\dots \inf_{y_{i(n)}} (\sigma_{n,1}(\bar x,\bar y_n)\lor \dots \lor \sigma_{n,j(n)}(\bar x,\bar y_n)), \] where $\sigma_{n,\nu}$ belongs to $\Sigma$ for $n<\omega$ and $\nu=1,\dots,j(n)$, $\bar x=x_1,\dots,x_m$, and $\bar y_n=y_1,\dots, y_{i(n)}$. \end{definition} \begin{proposition} \label{P:forcing special formulas} Let $\Sigma$ be a class of formulas of $\mathcal{L}_A$ which contains all atomic formulas and is closed under subformulas. Suppose that $\phi$ is a $\sup\bigwedge\inf$-formula over $\Sigma$, of the form \[ \sup_{x_1}\dots\sup_{x_m}\bigwedge_{n<\omega}\inf_{y_1}\dots \inf_{y_{i(n)}} (\sigma_{n,1}(\bar x,\bar y_n)\lor \dots \lor \sigma_{n,j(n)}(\bar x,\bar y_n)), \] where $\sigma_{n,\nu}$ belongs to $\Sigma$ for $n<\omega$ and $\nu=1,\dots,j(n)$, $\bar x=x_1,\dots,x_m$, and $\bar y_n=y_1,\dots, y_{i(n)}$. Then, if $(\mathbb{P},\leq,f)$ is a forcing property for $\mathcal{L}_A(C)$ and $p\in \mathbb{P}$, \[ F_p(\phi)= \sup_{\substack{\bar c\in C^m\\q\leq p}} \inf_{\substack{q'\leq q\\ \bar d\in C^{i(n)}\\ n<\omega}} \max_{1\leq\nu\leq j(n)} F_{q'}^w(\, \sigma_{n,\nu}(\bar c,\bar d)\,). \] \end{proposition} \begin{proof} We use Propositions~\ref{P:disjunctions} and \ref{P:finite disjunctions} to compute $F_p(\phi)$: \begin{align*} &F_p(\,\sup_{x_1}\dots\sup_{x_m}\bigwedge_{n<\omega}\inf_{y_1}\dots \inf_{y_{i(n)}} (\sigma_{n,1}(\bar x,\bar y_n)\lor \dots \lor \sigma_{n,j(n)}(\bar x,\bar y_n)\, )) & \\ =&\sup_{\bar c\in C^m} F_p^w(\, \bigwedge_{n<\omega}\inf_{y_1}\dots \inf_{y_{i(n)}} (\sigma_{n,1}(\bar c,\bar y_n)\lor \dots \lor \sigma_{n,j(n)}(\bar c,\bar y_n))\,) &\text{(by \ref{P:disjunctions})}\\ =&\sup_{\bar c\in C^m} \sup_{q\leq p}\inf_{q'\leq q} F_{q'}(\, \bigwedge_{n<\omega}\inf_{y_1}\dots \inf_{y_{i(n)}} (\sigma_{n,1}(\bar c,\bar y_n)\lor \dots \lor \sigma_{n,j(n)}(\bar c,\bar y_n))\,) & \\ =&\sup_{\bar c\in C^m} \sup_{q\leq p}\inf_{q'\leq q}\inf_{n<\omega}\inf_{\bar d\in C^{i(n)}} F_{q'}(\,\sigma_{n,1}(\bar c,\bar d)\lor \dots \lor \sigma_{n,j(n)}(\bar c,\bar d)\,) & \\ =&\sup_{\bar c\in C^m} \sup_{q\leq p}\inf_{q'\leq q}\inf_{n<\omega}\inf_{\bar d\in C^{i(n)}}\max_{1\leq\nu\leq j(n)} F_{q'}^w(\, \sigma_{n,\nu}(\bar c,\bar d)\,)&\text{(by \ref{P:finite disjunctions})}. \end{align*} \end{proof} \begin{remark} \label{R:forcing special formulas} If $p$ is a condition in the forcing property $\mathcal{P}(\mathcal{M},\Sigma)$, then $F_p^w(\phi)=H_p^w(\phi)$, by Proposition~\ref{P:w-forcing in P(M,Sigma)}. Hence, if $\phi$ is as in Proposition~\ref{P:forcing special formulas}, \[ F_p(\phi)= \sup_{\substack{\bar c\in C^m\\q\leq p}} \inf_{\substack{q'\leq q\\ \bar d\in C^{i(n)}\\ n<\omega}} \max_{1\leq\nu\leq j(n)} H_{q'}^w(\, \sigma_{n,\nu}(\bar c,\bar d)\,). \] \end{remark} Recall (\fref{sec:preliminaries}) that $C$ denotes countable set of constants not in $\mathcal{L}$ and that $\mathcal{L}(C)=\mathcal{L}\cup C$. As in Section~\ref{sec:special forcing properties}, if $\mathcal{M}$ is a class of $\mathcal{L}$-structures, $\mathcal{M}(C)$ denotes the class of structures of the form $(M, a_c)_{c\in C_0}$, where $M$ is in $\mathcal{M}$ and $C_0$ is a finite subset of $C$. If $\Gamma$ is a set of inequalities of the form $\phi<r$, where $\phi$ is an $\mathcal{L}_A(C)$-formula and $r$ is a real number, we will say that $\Gamma$ is \emph{satisfiable} in $\mathcal{M}$ if there exists a structure $M$ in $\mathcal{M}(C)$ such that $\phi^M<r$ for every inequality $\phi<r$ in $\Gamma$. Let $\Sigma$ be class of formulas of $\mathcal{L}_A$ that contains all the atomic formulas and is closed under subformulas. A \emph{finite $\Sigma$-piece of $\mathcal{M}$} is a finite set $p$ of inequalities of the form $\phi<r$, where $\phi\in\Sigma$, such that $p$ that is satisfiable in $\mathcal{M}$. \begin{corollary}[Omitting Types Theorem] \label{cor:omitting types} Let $(\,\phi_n\mid n<\omega\,)$ be a sequence of $\mathcal{L}$-formulas such that for each $n<\omega$ $\phi_n$ is a $\sup\bigwedge\inf$-formula over $\Sigma$, of the form \begin{gather*} \sup_{x_1}\dots\sup_{x_{m(n)}}\psi_n(x_1,\dots,x_{m(n)}), \end{gather*} where for each $n<\omega$ $\psi_n$ is of the form \[ \bigwedge_{k<\omega}\inf_{y_1}\dots \inf_{y_{i(n,k)}} (\sigma_{n,k,1}(\bar x_n,\bar y_{n,k})\lor \dots \lor \sigma_{n,k,j(n,k)}(\bar x_n,\bar y_{n,k})), \] with $\bar x_n=x_1,\dots,x_{m(n)}$, and $\bar y_{n,k}=y_1,\dots,y_{i(n,k)}, $ and let $(\,r_n\mid n<\omega\,)$ be a sequence of real numbers such that for every finite $\Sigma$-piece $p$ of $\mathcal{M}$ and every $\bar c_n\in C^{m(n)}$, the set $p\cup\{\psi_n(\bar c_n)<r_n\}$ is satisfiable in $\mathcal{M}$. Then there exists a canonical $\mathcal{L}(C)$-structure $M$ such that $\phi_n^M\leq r_n$ for every $n<\omega$. \end{corollary} \begin{proof} Let $p$ be a condition in the forcing property $\mathcal{P}(\mathcal{M},\Sigma)$. Fix a condition $q \leq p$, $n<\omega$ and $\bar c_n\in C^{m(n)}$. Since $q\cup\{\psi_n(\bar c_n)<r_n\}$ is satisfiable in $\mathcal{M}$, there exist $k<\omega$ and $\bar d_{n,k}\in C^{i(n,k)}$ such that \[ q\cup\{\, \sigma_{n,k,1}(\bar c_n,\bar d_{n,k})\lor \dots \lor \sigma_{n,k,j(n,k)}(\bar c_n,\bar d_{n,k})<r_n \,\} \] is satisfiable in $\mathcal{M}$. Let \[ q'=q\cup\{\, \sigma_{n,k,1}(\bar c_n,\bar d_{n,k})<r_n, \dots, \sigma_{n,k,j(n,k)}(\bar c_n,\bar d_{n,k})<r_n \,\}. \] Then, $q$ is a condition in $\mathcal{P}(\mathcal{M},\Sigma)$, and by \fref{lem:WPMSChar}, \[ \max_{1\leq\nu\leq j(n,k)}H_{q'}^w(\, \sigma_{n,k,\nu}(\bar c_n,\bar d_{n,k})\,)< r_n, \] so \[ \inf_{\substack{q'\leq q\\ \bar d\in C^{i(n,k)}\\ k<\omega}} \max_{1\leq\nu\leq j(n,k)} H_{q'}^w(\, \sigma_{n,k,\nu}(\bar c_n,\bar d_{n,k}) \,)< r_n. \] Thus, by Remark~\ref{R:forcing special formulas}, $F_p(\phi_n)\leq r_n$. Let $G$ be a generic set for $\mathcal{P}(\mathcal{M},\Sigma)$ (the existence of $G$ is guaranteed by Proposition~\ref{P:generic set}). For every $n<\omega$, $\phi_n^G=\inf_{p\in G} F_p(\phi_n)\leq r_n$ (see Definition~\ref{dfn:Generic}). Let now $M^G$ be as in Lemma~\ref{L:generic model}. Then, by Theorem~\ref{T:generic model}, $\phi_n^{M^G}=\phi_n^G\leq r_n$. \end{proof} \begin{remark} The reader may worry about the fact that the assumptions of \fref{cor:omitting types} involve strict inequalities while the conclusion only yields a weak inequality. In fact it would be enough to assume a weak inequality, i.e., that $p \cup \{ \psi_n(\bar c_n) \leq r_n\}$ is satisfiable in $\mathcal{M}$ for every $p$ and $\psi_n$ as in the statement of \fref{cor:omitting types}, or more precisely, that $p \cup \{ \psi_n(\bar c_n) < r_n + \varepsilon\}$ is satisfiable for every $\varepsilon > 0$. Indeed, in this case we would be able to find $M$ in which $\varphi_n^M \leq r_n + 2^{-m}$ for all $n,m$, i.e., such that $\varphi_n^M \leq r_n$ for all $n$. \end{remark} \section{Application: Separable Quotients} \label{sec:separable quotients} If $\phi$ is a formula of $\mathcal{L}_{\omega_1,\omega}$, we will say that $\phi$ is \emph{finitary} if all the occurrences of $\bigwedge$ in $\phi$ are finitary, i.e., if whenever $\bigwedge_{\psi\in\Phi} \psi$ is a subformula of $\phi$, the set $\Phi$ is finite. We recall from \fref{sec:preliminaries} that the set of all finitary formulas is denoted $\mathcal{L}_{\omega,\omega}$. If $M$ and $N$ are $\mathcal{L}$-structures, $M$ and $N$ are said to be \emph{elementary equivalent}, written $M\equiv N$, if $\phi^M=\phi^N$ for every finitary $\mathcal{L}$-sentence $\phi$. Thus $M\equiv N$ if and only if $\phi^M<r$ implies $\phi^N<r$ for every finitary $\mathcal{L}$-sentence $\phi$ and every rational number $r$. If $M$ is a substructure of $N$, $M$ is said to be an \emph{elementary substructure} of $N$ if $(M,a \mid a\in M)\equiv(N,a\mid a\in M)$. A Banach space $(X,\|\cdot\|)$ can be regarded as a metric structure in a number of ways. A natural approach is to introduce for each nonnegative rational $r$ a distinct sort for the closed ball $B_X(r)$ of radius $r$ around $0$; the metric on $B_X(r)$ is given by the norm $\|\cdot\|$; in the structure we also include: \begin{itemize} \item the inclusion maps $I_{r,s}\colon B_X(r)\to B_X(s)$ for $r<s$, \item the vector addition, which maps $B_X(r)\times B_X(s)$ onto $B_X(r+s)$, \item for each $\lambda\in\mathbb{Q}$, the scalar multiplication by $\lambda$, which maps $B_X(r)$ onto $B_X(|\lambda|r)$, \item the normalized norm predicate $\|\cdot\|/r$, which maps $B_X(r)$ onto the interval $[0,1]$, \item the normalized distance predicate on $B_X(r)$ defined by $d(x,y) = \|x-y\|/(2r)$ (as $x-y \in B_X(2r)$). \end{itemize} Notice that with the normalized norm and distance, all symbols are $1$-Lipschitz, meaning that the identity function $\delta(\epsilon)=\epsilon$ is a modulus of uniform continuity for each and every one of them. Other ways of regarding Banach space as metric structures are discussed in Section~3 of~\cite{BenYaacov-Usvyatsov:localstability} and in Section~4 \cite{BenYaacov:Perturbations}. If $X$ and $Y$ are Banach spaces and $T\colon X\to Y$ is a Banach space operator, we denote by $(X,Y,T)$ the structure that includes, in addition to the Banach space structure of $X$ and the Banach space structure on $Y$, in separate sorts, the operator $T$ as a family of functions between the appropriate sorts, i.e., from $B_X(r)$ to $B_Y(s)$ if $s \geq \|T\|r$. If $T$ is nonzero, the function $\delta(\epsilon)=\epsilon/\|T\|$ is a modulus of uniform continuity for $T$. \begin{proposition} \label{P:surjectivity} Let $X,Y,\Hat{X},\Hat{Y}$ be Banach spaces and let $T\colon X\to Y$ and $\Hat{T}\colon\Hat{X}\to\Hat{Y}$ be bounded linear operators such that $(X,Y,T)\equiv(\Hat{X},\Hat{Y},\Hat{T})$. Then $T$ is surjective if and only if $\Hat{T}$ is surjective. \end{proposition} \begin{proof} For a real number $r\geq 0$, let $B_X(r)$ and $B_Y(r)$ denote the closed balls of radius $r$ around $0$ in $X$ and $Y$, respectively. The proof of the Open Mapping Theorem shows that $T\colon X\to Y$ is surjective if and only if the following holds: for every $\epsilon>0$ there exists $\delta(\epsilon)>0$ such that \[ \forall y\in B_Y(\delta(\epsilon)) \ \exists x\in B_X(1) \ (\,\|T(x)-y\|\leq \epsilon \,). \] We can assume that $\delta(\epsilon)<1$ for $\epsilon<1$. Thus, $T$ is surjective if and only if for every $\epsilon$ with $0<\epsilon<1$ we have $\phi_\epsilon^{(X,Y,T)}=0$, where $\phi_\epsilon$ is the following finitary sentence (the variables $x$ and $y$ are of sort $B_X(1)$ and $B_Y(1)$, respectively): \[ \sup_y \ \inf_x \ \min\big(\ \delta(\epsilon)\mathbin{\mathpalette\dotminussym{}} \|y\|,\|T(x)- y\|\mathbin{\mathpalette\dotminussym{}} \epsilon \ \big). \] (Or, if one wishes to be pedantic, replace $\|T(x)- y\|\mathbin{\mathpalette\dotminussym{}} \epsilon$ with $\frac{1}{2}\|T(x)-y\| \mathbin{\mathpalette\dotminussym{}} \frac{\epsilon}{2}$.) Hence, if $(X,Y,T)\equiv(\Hat{X},\Hat{Y},\Hat{T})$ and $T$ is surjective, $\Hat{T}$ is surjective too. \end{proof} The authors are grateful to William B.~Johnson for pointing out that Proposition~\ref{P:surjectivity} is given by the proof of the Open Mapping Theorem. All the Banach spaces mentioned henceforth will be infinite dimensional. The Separable Quotient Problem is perhaps the most prominent open problem in nonseparable Banach space theory. The question is whether for every nonseparable Banach space $X$ there exist a separable Banach space $Y$ and a surjective operator $T\colon X\to Y$. Let $T\colon X\to Y$ be a Banach space operator and consider the structure $(X, Y, T)$ (the sorts of this structure are $X$ and $Y$). In this section we use Corollary~\ref{cor:omitting types} to prove that if $T\colon X\to Y$ is surjective and has infinite dimensional kernel, then there exists an operator $\Hat{T}\colon \Hat{X}\to \Hat{Y}$ such that \begin{enumerate} \item $(X,Y,T)\equiv(\Hat{X},\Hat{Y},\Hat{T})$, \item $\Hat{X}$ has density character $\omega_1$, \item $\Hat{Y}$ is separable. \end{enumerate} It follows from (i) and Proposition~\ref{P:surjectivity} that $\Hat{T}$ is surjective. \begin{lemma} \label{L:Hahn-Banach} Suppose that $X$ is a Banach space and $Y$ is a closed proper subspace of $X$. Then there exists a non-zero linear functional $f\colon X\to\mathbb{R}$ whose restriction to $Y$ is zero. Up to multiplication by a scalar we may further assume that $\|f\| = 1$. \end{lemma} \begin{proof} This is a well-known application of the Hahn-Banach theorem; the proof can be found in a textbook, e.g.,~\cite{Fabian-Habala-Hajek-Montesinos-Pelant-Zizler:2002}. \end{proof} \begin{lemma} \label{L:sequence} If $X$ is a Banach space of density character $\kappa$, there exists a family $(x_i)_{i<\kappa}$ in $X$ such that $\|x_i\|=1$ for every $i<\kappa$ and $\|x_i-x_j\|\geq 1$ for $i<j<\kappa$. \end{lemma} \begin{proof} The construction of $(x_i)_{i<\kappa}$ is inductive. Fix $j<\kappa$ and suppose that constructed $x_i$ is defined for $i<j$. Let $Y$ be the closed linear span of $\{ x_i \mid i<j\}$, which is a proper closed subspace of $X$. By Lemma~\ref{L:Hahn-Banach}, take $f\colon X\to\mathbb{R}$ such that $f(x)=0$ for every $x\in Y$ and $\|f\|=1$. Let now $x_j$ be an element of the unit sphere of $X$ such that $|f(x)|=\|f\|=1$. Then, if $y\in Y$, we have $\|x_j-y\|\geq |f(x_j)-f(y)|=1$; in particular, $\|x_j-x_i\|\geq 1$ for $i<j$. \end{proof} \begin{theorem} \label{T:separable quotient} For every surjective operator $T\colon X\to Y$ with infinite dimensional kernel there exists an operator $\Hat{T}\colon \Hat{X}\to \Hat{Y}$ such that \begin{itemize} \item $(X, Y,T)\equiv(\Hat{X},\Hat{Y},\Hat{T})$, \item $\Hat{X}$ has density character $\omega_1$, \item $\Hat{Y}$ is separable. \end{itemize} Furthermore, if $D$ is a given countable subset of $X$, the structure $(\Hat{X},\Hat{Y},\Hat{T})$ can be chosen with the following property: there exists a separable subspace $X_0$ of $X$ such that $D\subseteq X_0$ and if $T_0$ denotes the restriction of $T$ to $X_0$, \begin{itemize} \item $(X_0,T_0(X_0),T_0)\prec (X,Y,T)$, \item $(X_0,T_0(X_0),T_0)\prec (\Hat{X},\Hat{Y},\Hat{T})$. \end{itemize} \end{theorem} \begin{proof} By the L\"owenheim-Skolem Theorem~\cite[page 47]{Henson-Iovino:2002}, there exists a separable subspace $X_0$ of $X$ such that if $T_0$ denotes the restriction of $T$ to $X_0$ and $Y_0=T_0(X_0)$, \[ (X_0,Y_0,T_0)\prec (X,Y,T). \] (Note that $X_0$ can be taken so that it contains any given countable subset of $X$.) Let $A$ be a countable dense subset of $X_0$ and consider the structure \[ (\,X_0,Y_0,T,a \mid a\in A\,). \] Let $\mathcal{L}$ be the signature that results from expanding the signature of $(\,X,Y,T,a \mid a\in A\,)$ with new constant symbols $c_0,c_1,\dots$ and $c_0^*$ of sort $B_X(1)$ as well as new constant symbols $d_0,d_1,\dots$ and of sort $B_Y(1)$. Let us introduce some temporary terminology. If $\phi$ is an $\mathcal{L}$-sentence, an $\mathcal{L}$-structure $M$ \emph{satisfies} the inequality $\phi<r$ if $\phi^M<r$. An inequality $\phi<r$ will be called \emph{finitary} if the formula $\phi$ is finitary. Let $\Gamma$ consist of the following inequalities (the variable $x$ in~\ref{I:Surjectivity} is of sort $B_Y(1)$): \begin{enumerate} \item \label{item:ElemIneq} All the finitary inequalities satisfied by the structure $(\,X_0,Y_0,T,a \mid a\in A\,)$ \item $\lnot\|c_n\|<\epsilon$ (i.e., $\|c_n\| > 1-\varepsilon$) and $\lnot\|d_n\|<\epsilon$, for every $n<\omega$ and every rational $\epsilon>0$. \item $\lnot\half\|c_m-c_n\| < \half + \epsilon$ (i.e., $\|c_n-c_m\| > 1-2\varepsilon$) and $\lnot\half\|d_m-d_n\|<\half+\epsilon$, for every pair $m,n$ with $m<n<\omega$ and every rational $\epsilon>0$. \item $\lnot \|c_0^*\|<\epsilon$, for every rational $\epsilon>0$. \item $\lnot\half\|c_0^*-c_n\|<\half+\epsilon$, for every $n<\omega$ and every rational $\epsilon>0$. \item $\|T(c_\omega)\|<\epsilon$, for every rational $\epsilon>0$. \item \label{I:Surjectivity} For every rational $\epsilon>0$, the inequality \[ \sup_x \left( \bigwedge_{\substack{r_0,\dots, r_{n-1}\in\mathbb{Q}\cap[-1,1]\\ n<\omega}}(n+1)\cdot\frac{1}{n+1}\| T(x)-\sum_{i<n} r_id_i\|\right)<\epsilon. \] Here $\frac{1}{n+1}\|\cdot\|$ is just the normalized norm predicate on the sort of $ T(x)-\sum_{i<n} r_id_i $, and $(n+1)\cdot\varphi$ is defined in general as $\varphi \dotplus \cdots \dotplus \varphi$ $n+1$ times. \end{enumerate} Lemma~\ref{L:sequence} ensures that the hypotheses of \fref{cor:omitting types} are satisfied with $\Sigma = \mathcal{L}_{\omega,\omega}$. Thus by \fref{cor:omitting types}, $\Gamma$ has a separable model \[ (\,X_1,Y_1,T_1, a, a_n, b_n, a_0^*\mid a\in A, n<\omega\,), \] where for each $n<\omega$, $a_n$ is the interpretation of $c_n$, $b_n$ is the interpretation of $d_n$, and $a_0^*$ is the interpretation of $c_0^*$. By \fref{item:ElemIneq}, we have \[ (\,X, Y,T)\prec (\,X_1, Y_1,T_1), \] so, in particular, by Proposition~\ref{P:surjectivity}, $T_1$ is surjective. Now we iterate the preceding process to find for each ordinal $\alpha$ with $0<\alpha<\omega_1$ a separable structure \[ (\,X_\alpha, Y_\alpha,T_\alpha, a_n, b_n, a_i^* \mid n<\omega, i<\alpha\,) \] such that if $0<\alpha<\beta<\omega_1$, \begin{itemize} \item $(\,X_\alpha, Y_\alpha,T_\alpha, a_n, b_n, a_i^* \mid n<\omega, i<\alpha\,)\prec (\,X_\beta, Y_\beta,T_\beta, a_n, b_n, a_i^* \mid n<\omega, i<\alpha\,)$ \item $ a_i^*\in X_\alpha$ for $i<\alpha$ \item $\| a_i^*\|=1$ and $\| a_i^*- a_j^*\|=1$ for $i<j<\alpha$ \item The linear span of $\{b_n\mid n<\omega\}$ is a dense subset of $T_\alpha(X_\alpha)$. \end{itemize} The theorem then follows by taking $\Hat{X}=\bigcup_{\alpha<\omega_1} X_\alpha$, $\Hat{Y}=\bigcup_{\alpha<\omega_1} Y_\alpha$, and $\Hat{T}=\bigcup_{\alpha<\omega_1} T_\alpha$. \end{proof} \newcommand{\etalchar}[1]{$^{#1}$} \providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace} \providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR } \providecommand{\MRhref}[2]{% \href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2} } \providecommand{\href}[2]{#2}

1. Introduction {#s0005} =============== Porcine epidemic diarrhea virus (PEDV) is an enteric virus that causes severe diarrhea, dehydration, and depression [@b0005], [@b0010]. PEDV is a species within the *Alphacoronavirus* genus (previously the type I *Coronaviridae*) in the *Coronaviridae* family of the *Nidovirales* order [@b0015], [@b0020]. Instances of PEDV began in the early 1970s, but outbreaks had been relatively tame, small, and sporadic until a panzootic occurred beginning in 2010. A mutation in the virus led to an increased virulence in some strains. This most recent outbreak caused severe economic losses worldwide as well as the introduction of the disease into North America and the United States in 2013 [@b0025], [@b0030], [@b0035], [@b0040], [@b0045]. Transmission of PEDV occurs primarily through the oral-nasal route. Incubation time for PEDV varies with age and strain, and the original strain, CV777, can take up to 8 days before viral shedding is detected [@b0050]. Newborn piglets (\<7 days) develop symptoms to PEDV including fecal shedding within 24 h of infection and die at a high rate within 3--4 days, causing significant economic loss to the swine industry [@b0055], [@b0060]. The successful replication of PEDV in intestinal epithelial cells of young piglets causes mal-absorptive diarrhea induced by decreased efficacy of brush border membrane-bound digestive enzymes and leaky junctions in the small intestine [@b0065], [@b0070]. The symptoms are severe enough in these cases to result in a death rate of nearly 100% upon exposure due to dehydration. This is not true for adult pigs, however, which may be due to their delayed infection and more mature microvilli. The typical genome for all *Coronaviridae* contains a 5′ and a 3′ untranslated region (UTR), and contains a range of open reading frames (ORFs) [@b0075]. PEDV has an approximately 28 kb genome with ORFs arranged 5′ to 3′ in the order of: ORF1a, ORF1b, spike (S) protein, envelope (E) protein, membrane (M) glycoprotein, ORF3, and nucleocapsid (N) protein. For most coronaviruses, the S protein undergoes post-translational modifications including cleavage into an S1 and an S2 domain [@b0080]. This is not true for PEDV, but there is a conventional separation of the protein into these domains around amino acids (aa) 781/782. Recent research in recombinant vaccination against PEDV has predominantly focused on the S protein, with particular attention paid to the S1 domain [@b0085], [@b0090], [@b0095], [@b0100], [@b0105]. There are two known, small, and linear B-cell epitopes within the S1 domain corresponding to the aa sequences ^748^YSNIGVCK^755^, and ^764^SQYGQVKI^771^ from the original CV777 strain [@b0110], [@b0115], [@b0120]. An additional small B-cell epitope was discovered within the S2 domain corresponding to the aa sequence ^1368^GPRLQPY^1374^ [@b0110]. A fourth short, linear B-cell epitope is located in the M protein corresponding to the sequence ^195^WAFYVR^200^ [@b0125]. Hepatitis B virus core antigen (HBcAg) is a popular protein expressed in *E. coli* for use in vaccine development. The resultant proteins are able to assemble into virus-like particles (VLPs) *in vitro* after purification [@b0130] *.* Used in a chimeric form, foreign antigens from other viruses can be inserted into the protein genetically for efficient immune system presentation from the VLP platform. HBcAg is typically divided into two separate domains for chimeric protein design. The change from the N-terminal to the C-terminal domain occurs around Isoleucine-80. Insertion of foreign epitopes near this residue will have a higher immunogenic response compared to an N-terminal or C-terminal insertion, and this site was identified as the major immunodominant region (MIR) [@b0135]. Previous work by this lab used HBcAg as a vector protein in a novel PEDV vaccination strategy. During the original experiment, five treatment vaccines were designed including four candidates with each of the four (^748^YSNIGVCK^755^, ^764^SQYGQVKI^771^, ^1368^GPRLQPY^1374^, ^195^WAFYVR^200^) linear B-cell epitopes inserted individually into the MIR. A fifth candidate had all four epitopes together in a linear peptide sequence inserted into the MIR of the HBcAg "backbone" protein [@b0140]. The resultant epitope antibody (Ab) titers as measured by ELISA were not strongly correlated with virus neutralization with the exception of the ^748^YSNIGVCK^755^ motif. It was therefore concluded that an immune response to this particular antigen from PEDV might be more critical than others for virus neutralization. The inclusion of the YSNIGVCK epitope in future vaccine designs could therefore be critical to their efficacy. This report focuses on several design changes to previous HBcAg-based PEDV vaccines in an effort to improve the exposure of the inserted PEDV epitopes to the immune system and increase the overall immune response. The elimination of three residues from the MIR of the HBcAg protein, ^79^PIS^81^, was reported to increase the immune response to the inserted epitope [@b0145]. This modification was performed for 5 vaccine candidates ([Fig. 1](#f0005){ref-type="fig"} B--D and [Fig. 1](#f0005){ref-type="fig"}F--G). The addition of flexible, water-soluble residues between foreign epitope insertions at the MIR was also tested in two vaccine candidates to increase the solubility and presentation of the included antigens ([Fig. 1](#f0005){ref-type="fig"}C and D) [@b0150]. The substitution of two residues, D29C and R127C, in the HBcAg vector protein of two vaccine candidates ([Fig. 1](#f0005){ref-type="fig"}D and G) was tested for increased efficacy. Assembly of HBcAg-based VLPs with these substitutions was previously shown by Lu et al to add a network of disulfide covalent bonds between pentamers and hexamers leading to an increased structural stability (Native VLP shown in [Fig. S1](#s0130){ref-type="sec"}A and C, additional disulfide bonds shown in red in [Fig. S1](#s0130){ref-type="sec"}B and D) [@b0155]. Finally, the locations of the epitopes ^1368^GPRLQPY^1374^ and ^195^WAFYVR^200^ were changed so as to only include the ^748^YSNIGVCK^755^ motif in the MIR of the assembled VLPs. All vaccine design changes were performed and tested *in vivo* in a first animal trial ([Fig. 1](#f0005){ref-type="fig"}A--D), and a second animal trial ([Fig. 1](#f0005){ref-type="fig"}A, E--G).Fig. 1Schematic diagrams of each vaccine tested. Foreign epitope sequences are indicated for the MIR, the N-terminus, and the C-terminus of the HBcAg protein as well as their location within the vector protein. Additional mutations are reported at their approximate location. (A) The truncated HBcAg protein with an N-terminal Histidine tag; (B--D) treatment vaccines used during the first animal trial; (E--G) treatment vaccines used during the second animal trial. 2. Materials and methods {#s0010} ======================== 2.1. Plasmid construction {#s0015} ------------------------- Plasmid construction was performed using the pET-28a (+) plasmid (Novagen, Madison, WI). Briefly, genes coding for the optimized sequences (IDT, Coralville, IA) for each protein were amplified to contain an overlapping sequence corresponding to their placement in the vector plasmid. The genes were then used to amplify the pET-28a plasmid, which then contained the gene of interest. Each circular plasmid was transformed into BL21 (DE3) *E. Coli* T7 express® (NEB) cells using the heat shock method. Each construct was verified by Sanger sequencing after transformation for the correct sequence and codon usage. 2.2. Expression {#s0020} --------------- Overnight cultures of BL21 (DE3) *E. coli* cells that were grown in LB media supplemented with 30 μg/mL kanamycin were used to inoculate 1 L batches of 2 × YT media supplemented with the same antibiotic at a rate of 0.2% v/v. The *E. coli* cultures were incubated at 28 °C until the OD~600~ reached 0.7 AU. Protein production was induced with 1 mM Isopropyl β-D-1-thiogalactopyranoside (IPTG), and cells were incubated overnight. The resulting cells were centrifuged and stored at −20 °C until purification. 2.3. Purification {#s0025} ----------------- Protein purification was performed as previously described [@b0140]. Briefly, cell pellets were re-suspended and sonicated in PBS. All proteins aside from ΔHBcAg were located in the inclusion bodies. Proteins were denatured in 6 M urea and loaded onto an anion exchange chromatography (AEX) column utilizing DEAE Sepharose FF resin (GE Healthcare, Marlborough, MA). Eluents were then pH adjusted and loaded onto an immobilized metal affinity chromatography (IMAC) column. Following IMAC purification, step-wise dialysis was performed in PBS at pH 7.8 with additives. These additives were, in order: 0.9% Sarkosyl, 0.45% Sarkosyl, 0.01% Sarkosyl, 0% Sarkosyl (PBS only). Each vaccine candidate was then diafiltered further into PBS buffer using Spin X UF tubes with a molecular weight cut off of 100 kDa (Corning, Corning, NY). 2.4. Formulation {#s0030} ---------------- Concentrated vaccines in PBS were sterile filtered through 30 mm, 0.10 μm syringe filters (Celltreat, Pepperell, MA) followed by adjuvant addition and dilution with sterile PBS such that each dosage of 0.5 mL contained 20 μg of VLPs. Final vaccines for the first animal trial contained 20 μg protein in PBS at pH 7.8 with 100 μg per dose of chitosan (Invivogen) [@b0160]. Candidates formulated for the second experiment had a total of 20 μg protein in PBS at pH 7.8 with 100 μg Alhydrogel (Invivogen). Vaccine candidates were then held for \< 8 h at 4 °C until ready for use [@b0165]. Vaccine dosages were chosen to be at or below previous studies [@b0145], [@b0170], [@b0175]. 2.5. Confirmation of assembly {#s0035} ----------------------------- A zetasizer nano (Malvern Instruments, Malvern UK) instrument was used according to manufacturers instructions, where particle size distributions were measured in PBS for initial confirmation of VLP assembly. Formvar coated copper grids were treated with protein samples for a few seconds until the grid was blotted with filter paper and allowed to dry. The samples were then stained with 2% phosphotungstic acid at pH 7, again for a few seconds before the stain was blotted with filter paper and allowed to dry [@b0180]. Images were taken using a JEM-1400 (JEOL, Peabody, MA) microscope at a maximum of 300,000x. 2.6. Protein visualization {#s0040} -------------------------- Protein samples were prepared for analysis by adding water, NuPage Reducing Agent (10×) (Invitrogen, Carlsbad, CA), and NuPage LDS Buffer (4×) (Invitrogen) for a total of 10 μL per well. Non-reduced samples were prepared similarly, substituting 1 μL of water for the reducing agent. Each sample was put into a NuPage™ 4--12% Bis-Tris PAGE gel (Invitrogen) along with a PrecisionPlus™ protein ladder (Biorad, Hercules, CA). Gels were run in MES buffer (Invitrogen). After electrophoresis, gels were rinsed three times with 18 MΩ water for 5 min each. Gels were then treated with SimplyBlue™ Safestain (Invitrogen) for at least one hour, and then rinsed with water overnight. Images of stained gels were taken using a Chemidoc Touch Imager (BioRad). Densitometry measurements were performed using ImageLab software (BioRad). Densitometry readings of \>96% were considered to be pure enough to proceed with animal trials. 2.7. Western blot {#s0045} ----------------- Western blots were performed using the same protocol as reduced PAGE samples, but with a Precision Plus WesternC protein ladder (Biorad). Gels were loaded with 2 μL of cell lysate per well. Each Gel was blotted onto a nitrocellulose membrane using a Transblot® Turbo™ instrument (Biorad) according to manufacturers instructions. Membranes were blocked for at least one hour using 5% non-fat dry milk in TBST. A 3 × wash was performed between each step consisting of three five-minute rinses in TBST. Each membrane was treated with a primary mouse anti-histag antibody (Thermo) followed by a secondary HRP conjugated goat anti-mouse antibody (Millipore, Billerica, MA). Detection was done using a Chemidoc Touch instrument (Biorad) with Clarity® ECL substrate (Biorad). 2.8. Endotoxin testing {#s0050} ---------------------- Endotoxin levels were tested after removal of detergent and chaotropic agents. Tests were performed using a Pierce® LAL Chromogenic Endotoxin Quantitation Kit (Thermo). The kit was used according to manufacturer's instructions, with a endotoxin cut off level of \<5 EU/mL. 2.9. Animals {#s0055} ------------ All animal experiments were performed in accordance with guidelines and approved by the Virginia Tech Institutional Animal Care and Use Committee (IACUC). Groups of 8 germ free balb/c mice from Charles River were used for *in vivo* experiments. 7--8 week old mice were allowed to acclimate for 13 days prior to a Day -1 blood draw via submandibular bleed. Vaccine candidates and two negative controls (ΔHBcAg and PBS Only) were administered intraperitoneally on day 0, day 14, and day 28. Blood samples were collected on days 13 and 27 via the cheek punch method, and on day 42 by exsanguination. Collected blood samples were incubated at 4 °C for 2 h and centrifuged. Aliquots of supernatants were made, if required, and frozen at −20 °C until tested. 2.10. Enzyme linked immunosorbent assay (ELISA) {#s0060} ----------------------------------------------- Ab titers to the inserted foreign epitopes were measured by a previously described process utilizing custom designed and purified ELISA fusion proteins (EFPs) [@b0140]. All proteins used in coating were diluted to 10 μg/mL in coating buffer (50 mM Na~2~CO~3~, 50 mM NaHCO~3~, pH 9.6). Plates were coated at 37 °C for \> 2 h, washed, and blocked overnight at 4 °C. Serum samples from Day 42, serially diluted in blocking buffer with a starting dilution of 1:25 were applied for 1 h at room temperature (RT). Plates were next incubated with goat anti-mouse IgG HRP (Millipore, Billerica, MA). Visualization was done with TMB One Component Microwell Substrate (SouthernBiotech) and read at 450 nm using a Biotek Synergy HTX plate reader (Winooski, VT). Titers were calculated as the reciprocal of the dilution at which test sera measurements were statistically similar to PBS negative control samples (n = 3) located on each plate. The *Acinetobacter* phage-derived VLP, AP205, was expressed and purified from the same plasmid (pET-28a) as the PEDV vaccine candidates using the same process. Different fractions were collected during AEX and the proteins were not assembled. ELISA plates coated with VLP proteins were performed at the same dilutions as the EFP coated plates, with the same process. 2.11. Whole virus (WV) antibody ELISA {#s0065} ------------------------------------- The WV ELISA process used was adapted from Thomas et al. [@b0185]. Microlon 600 flat bottom plates were coated with isolated CO strain (USDA APHIS, Ames, IA) of PEDV at a concentration of 10^4.5^ TCID~50~ diluted 1:50 in sterile PBS at 4 °C for 6 h. Plates were then washed 5 × in sterile TBS. Plates were blocked in 1% BSA at 4 °C overnight, followed by a wash. Murine serum samples were diluted 1:25 in TBS and each well was filled with 100 μL of diluted serum. Serum samples were allowed to incubate for 1 h at RT, and were then washed 5 × with sterile TBS. Each well was next treated with a 100 μL volume of 1:10,000 diluted goat anti-mouse IgG HRP conjugate (Millipore) for 1 h at RT. After 5 × washes in TBS, the wells were detected with 100 μL of TMB One Component Microwell Substrate (SouthernBiotech). The reaction was stopped using 0.05% sulfuric acid. A Biotek Synergy HTX plate reader was used to read plates at 450 nm. 2.12. PEDV virus neutralization (VN) test {#s0070} ----------------------------------------- PEDV VN titers were measured using a focus forming assay (FFA) as described previously and executed by the Animal Disease Research and Diagnostic Laboratory in the Veterinary and Biomedical Sciences Department at South Dakota State University [@b0190]. Briefly, serum samples were heat-treated and diluted 1:10 in minimal essential medium (MEM) with 1.5 μg/mL tosyl phenylalanyl chloromethyl ketone (TPCK) trypsin. A serial, 2-fold dilution was then performed in 96-well plates with the MEM/TPCK-trypsin diluent. Cell culture adapted PEDV was then added at 100 foci forming units (FFU)/100 μL and incubated for 1 h at 37 °C. The resulting mixture was then added to confluent monolayers of Vero-76 cells and incubated for 2 h at 37 °C. These new wells were then washed with MEM/TPCK-trypsin and allowed to incubate for 20--24 h. Plates were fixed with acetone and stained with fluorescein isothiocyanate (FITC) conjugated mAb SD6-29 for visualization. End point titers were determined as the lowest concentration resulting in a 90% or greater reduction in fluorescence relative to controls. 2.13. Statistics {#s0075} ---------------- JMP® Pro 12 software (SAS, Cary, NC) was used in all statistical analysis. All samples were run in a nested experimental design using administration, peptide/protein, and group, respectively. Individual ANOVA contrasts were performed to determine statistical significance between individual groups. 3. Results {#s0080} ========== 3.1. Epitope antibody generation {#s0085} -------------------------------- All vaccines used for *in vivo* testing were successfully purified ([Fig. S2](#s0130){ref-type="sec"}, further discussion in supplemental material) and were able to form VLPs as indicated by dynamic light scattering (data not shown) and by transmission electron microscope (TEM) images ([Fig. 2](#f0010){ref-type="fig"} A--H). After the completion of the first animal trial ([Fig. 3](#f0015){ref-type="fig"} ), there was an immune response seen to the SQSGQVKI, GPRLQPY, and WAFYVR epitopes as seen in [Fig. 3](#f0015){ref-type="fig"}A--D for two of the three test vaccine groups compared to the ΔHBcAg control (HBcAg + s4E and mHBcAg + ms4E). In each case, the epitope-specific Ab titer increased in a dose-dependent manner over the course of the experiment (data not shown). It is important to note that the responses seen to the YSNIGVCK motif were minimal or non-existent ([Fig. 3](#f0015){ref-type="fig"}A). The lack of response to the YSNIGVCK motif may indicate that there is an interaction taking place between residues in the MIR of all test candidates that reduces or eliminates the presentation of the YSNIGVCK epitope [@b0195].Fig. 2TEM verification of VLP assembly. TEM images were taken of each vaccine candidate prior to each animal trial. Concentration and buffer vary (18 MΩ H~2~O, or PBS) with each sample for visual confirmation. Larger images were taken at the maximum magnification. Smaller images are close-ups manipulated after the images were taken for greater detail.Fig. 3Epitope antibody titers from the first animal trial. Ab titers from mouse sera after completion of the first animal trial as measured by ELISA corresponding to plates coated with EFPs are shown. Ab titers specific to (A) EFP-YSNIGVCK, (B) EFP-SQSGQVKI, (C), EFP-GPRLQPY and (D) EFP-WAFYVR were measured. Each treatment group's mean titer is compared to the ΔHBcAg negative control by ANOVA analysis. Standard error bars are shown for each treatment group (^\*^ = P ≤ 0.05, ^\*\*^ = P ≤ 0.01, ^\*\*\*^ = P ≤ 0.001). During the second animal trial ([Fig. 4](#f0020){ref-type="fig"} ), alum was included in the formulation of three of the vaccine candidates in an effort to test its efficacy as an adjuvant. [Fig. 4](#f0020){ref-type="fig"}A indicates that the addition of Alum to the formulation was detrimental to vaccine efficacy as measured by the immune response to the YSNIGVCK epitope (comparison of the HBcAg + SE candidate vaccine with and without alum added). This phenomenon appears to hold true for the HBcAg + SSM candidate as well. Interestingly, the addition of alum did not appear to disrupt the response to the mHBcAg + SSM vaccine, which was identical to the HBcAg + SSM vaccine with the exception of the D29C and R127C mutations. These results indicate the addition of Cysteine residues enabled the mHBcAg + SSM vaccine candidate to invoke an immune response even in the presence of alum, possibly due to the added network of disulfide bonds. However, inefficient antigen presentation in the MIR of the HBcAg + SSM candidate due to peptide interactions cannot be ruled out as a contributing factor to the low Ab response level. The mHBcAg + SSM candidate was also able to elicit an immune response to the N-terminal GPRLQPY and C-terminal WAFYVR epitopes as seen in [Fig. 4](#f0020){ref-type="fig"}B and C. Similar to the first animal trial, Ab titers for each treatment increased in a dose dependent manner throughout the experiment (data not shown).Fig. 4Epitope antibody titers from the second animal trial. Ab titers from mouse sera after completion of the second animal trial as measured by ELISA corresponding to plates coated with EFPs are shown. Ab titers specific to (A) EFP-YSNIGVCK, (B) EFP-GPRLQPY, and (C) EFP-WAFYVR were measured. Each treatment group's mean titer is compared to the ΔHBcAg negative control by ANOVA analysis. Standard error bars are shown for each treatment group (^\*^ = P ≤ 0.05, ^\*\*^ = P ≤ 0.01, ^\*\*\*^ = P ≤ 0.001). 3.2. Vector protein antibody generation {#s0090} --------------------------------------- There is a great concern in the recombinant production of chimeric protein-based vaccines that pre-existing Abs to the vector protein may cause the elimination of the vaccine prior to antigen presentation. The Ab titers specific to the ΔHBcAg protein itself were therefore tested in conjunction with the individual epitope inserts. Results from [Fig. 5](#f0025){ref-type="fig"} A indicate that the Ab titers to ΔHBcAg were statistically similar across all treatments during the first animal trial, indicating that each vaccine induced an immune response, though not necessarily to the inserted epitopes ([Fig. 3](#f0015){ref-type="fig"}). During the second animal trial ([Fig. 5](#f0025){ref-type="fig"}B), however, the data suggest that the two vaccine candidates with an epitope inserted N-terminal to the Histidine tag (HBcAg + SSM and mHBcAg + SSM) had significantly lower Ab responses to ΔHBcAg compared to the other treatments. To further elucidate the immune response to the vector protein, Ab titers to an unrelated vector protein containing the same N-terminal Histidine tag and cleavage site were tested (AP205, as described in the materials and methods section). [Fig. 5](#f0025){ref-type="fig"}C and D indicate that all vaccines elicited a significantly higher Ab titer to the Histidine tag than those vaccines with a foreign epitope insertion N-terminal to the Histidine tag (HBcAg + SSM and mHBcAg + SSM), following the same pattern as the Ab titers to the whole ΔHBcAg protein. Together these data demonstrate that the histidine tag located at the N-terminus of the vaccine candidates may play a large role in Ab recognition to each protein, as has been seen previously by others, and the placement of an epitope N-terminal to the tag may reduce its immune interaction [@b0200].Fig. 5Antibody titers to the vector protein and Histidine tag. Ab titers from mouse sera after completion of both animal trials as measured by ELISA corresponding to plates coated with His6xΔHBcAg (A and B) or His6xAP205 (C and D) are shown. Sera were tested for the presence of Abs after the first animal trial (A and C) as well as the second animal trial (B and D). Each treatment group's mean titer is compared to the ΔHBcAg negative control by ANOVA analysis. Standard error bars are shown for each treatment group (^\*^ = P ≤ 0.05, ^\*\*^ = P ≤ 0.01, ^\*\*\*^ = P ≤ 0.001). 3.3. Virus neutralization (VN) {#s0095} ------------------------------ To further test the efficacy of the vaccine candidates, the response of the murine serum to the virus itself was determined by foci forming assay (FFA) utilizing the CO strain of PEDV. [Fig. 6](#f0030){ref-type="fig"} A shows the VN titers from the first animal trial. Although there is statistical significance between the three test groups (HBcAg + 4E, HBcAg + ms4E, and mHBcAg + ms4E) and the negative controls (ΔHBcAg, and PBS Only), the numbers remain relatively low. Serum samples from the second animal trial ([Fig. 6](#f0030){ref-type="fig"}B) demonstrated a stronger ability to neutralize virus for two of the treatment groups only (HBcAg + SE and mHBcAg + SSM). When combined with peptide Ab titers, the data demonstrate a strong correlation between high VN titer and high Ab titer to the YSNIGVCK epitope ([Figs. 6](#f0030){ref-type="fig"}B and [4](#f0020){ref-type="fig"}A) compared to the GPRLQPY epitope ([Figs. 6](#f0030){ref-type="fig"}A and [3](#f0015){ref-type="fig"}C).Fig. 6VN and WV ELISA. A foci forming assay (FFA) and a whole virus ELISA were used to determine the serum Ab titers specific to the CO strain of PEDV. Sera collected at Day 42 for each treatment group from (A) the first animal trial and (B) the second animal trial were tested for PEDV neutralization *in vitro.* These same serum samples were subsequently tested for their ability to bind to PEDV without neutralization measurement (C). Each treatment group's mean titer is compared to both the ΔHBcAg and PBS Only negative controls by ANOVA analysis. Standard error bars are shown for each treatment group (^\*^ = P ≤ 0.05, ^\*\*^ = P ≤ 0.01, ^\*\*\*^ = P ≤ 0.001). 3.4. Whole virus (WV) ELISA {#s0100} --------------------------- To determine if a low binding affinity of Abs generated through recombinant vaccination to the native proteins from virions could contribute to VN discrepancies, a WV ELISA assay was performed on all serum samples ([Fig. 6](#f0030){ref-type="fig"}C). For each treatment, the WV ELISA results appear to correlate well with individual peptide ELISA results independent of the specificity of the Abs generated. The candidates with high Ab titers to B-cell epitopes from PEDV (HBcAg + 4E, HBcAg + ms4E, and mHBcAg + ms4E from [Fig. 3](#f0015){ref-type="fig"}, HBcAg + SE, and mHBcAg + SSM + Alum from [Fig. 4](#f0020){ref-type="fig"}) had significantly higher absorption signals than the ΔHBcAg, and PBS Only treatments. These data indicate that the peptides presented by a vector protein did not prevent them from eliciting Abs that could bind to the virus during the virus neutralization assay, and that there may be a true difference in the ability of the YSNIGVCK to neutralize PEDV compared to GPRLQPY. 4. Discussion {#s0105} ============= The original goal of this project was to improve upon vaccine designs made and tested previously for PEDV by design modification to increase antigen presentation and virus neutralization on a per dose basis [@b0140]. The elimination of the ^79^PIS^81^ residues and the addition of Serine residues between foreign epitopes tested during the first animal trial revealed a major flaw where either or both modifications prevented the presentation of the YSNIGVCK epitope. It is hypothesized that this phenomenon is due to residue interactions at the MIR of the VLP, as each modification is predicted to increase the flexibility of the region [@b0145]. This newly added flexibility might therefore have removed a small α-helix native to the vector protein, allowing the new interactions to take place. These treatments were able to generate Abs to the other B-cell epitopes and to the vector protein, indicating that assembly of the VLPs with some antigen presentation did indeed take place. VN titers to these vaccine candidates were significant but low, indicating that high Ab titers to the SQSGQVKI, GPRLQPY, and WAFYVR epitopes are not necessarily associated with virus neutralization. With the preliminary data in mind, the focus of the second experiment was two-fold: to test the use of alum as an adjuvant, and attempt two redesigns to increase the immune response to the YSNIGVCK epitope while maintaining the immune response to the GPRLQPY and WAFYVR B-cell epitopes. Although neutralization is a key measurement, there may be a benefit to including multiple conserved antigens in a vaccine for future mutations in some strains. The HBcAg + SE protein was used as a positive control during the second animal trial in two different formulations. The addition of alum to the formulation of the HBcAg + SE treatment group appeared to be detrimental to its immune response as indicated by a lack of YSNIGVCK-specific Abs to the adjuvanted vaccine ([Fig. 4](#f0020){ref-type="fig"}A). Interestingly, the HBcAg + SE + Alum treatment group *was* capable of inducing a response to ΔHBcAg ([Fig. 5](#f0025){ref-type="fig"}B), possibly through the recognition of the Histidine tag ([Fig. 5](#f0025){ref-type="fig"}D). Together, the data indicate that the HBcAg + SE VLPs were assembled with successful YSNIGVCK antigen presentation prior to formulation with alum, but the antigen presentation was reduced upon formulation leading to the Ab recognition of vector protein alone. The use of alum in an IP vaccine delivery may indeed reduce the immune response as seen by Chang et al. [@b0205]. Cubas et al have also indicated that IP vaccine delivery of VLPs may not be the best route for drainage into the lymph nodes, which might reduce the exposure of the foreign antigens included in the VLP designs [@b0210]. However, the complete elimination of an immune response to the foreign antigen with a similar immune response to the vector protein indicates that there may be additional factors present within this study. Future *in vivo* work may therefore require a more detailed analysis of the administration route as a possible variable in vaccine improvement. There was also a significant immune response difference seen between the two redesigned vaccine candidates (HBcAg + SSM and mHBcAg + SSM) during the second animal trial. The mHBcAg + SSM vaccine, with the Cysteine substitutions, generated Abs specific to all three PEDV epitopes included ([Figs. 4](#f0020){ref-type="fig"}A--[3](#f0015){ref-type="fig"}C) and was able to neutralize virus ([Fig. 6](#f0030){ref-type="fig"}B), while the non-substituted version elicited an immune response to the vector protein alone ([Fig. 5](#f0025){ref-type="fig"}B). The sole difference between the HBcAg + SSM and mHBcAg + SSM candidate designs was in the vector protein located away from the MIR. Non-reduced PAGE analysis ([Fig. S2](#s0130){ref-type="sec"}B) along with TEM images ([Fig. 2](#f0010){ref-type="fig"}G) have indicated that there were no issues with assembly of the HBcAg + SSM VLPs. The difference in immune response therefore seems to indicate that the Cysteine residues have improved antigen presentation in the mHBcAg + SSM vaccine, possibly due to an increase in VLP stabilization as reported by Lu et al. [@b0155]. We hypothesize this increased stabilization gives the mHBcAg + SSM vaccine a higher half-life *in vivo*, leading to a prolonged antigen presentation. However, this hypothesis was not investigated further. The placement of the GPRLQPY foreign epitope N-terminal to the Histidine tag in two vaccine candidates (HBcAg + SSM and mHBcAg + SSM) is correlated with a reduced immune reaction to ΔHBcAg when compared to all of the other treatments tested ([Fig. 5](#f0025){ref-type="fig"}A and B). Ab response to the histidine tag cannot be separated from that of ΔHBcAg in this case, as the Histidine tag was used for purification of the ΔHBcAg coat proteins used in the assay. However, ELISA values for an unrelated protein containing the same N-terminal sequence showed a similar pattern amongst all treatment groups ([Fig. 5](#f0025){ref-type="fig"}C and D), indicating that the placement of the foreign epitope upstream may reduce the Ab recognition of the Histidine tag. These results may make physiological sense. If an extended N-terminus on the ΔHBcAg protein forms spikes upon VLP assembly, then it is possible that antigens located within that region (including Histidine tags) would have an opportunity to interact with immune cells [@b0215]. The N-terminal placement of a more desirable epitope within this region may therefore cause a preferential interaction between the inserted foreign epitope and the immune system compared to that of the tag. It seems unlikely that the other changes (removal of ^79^PIS^81^, C-terminal addition of the WAFYVR epitope) would influence the detection of the Histidine tag, but they cannot be ruled out as possibilities from the presented results. [Fig. 6](#f0030){ref-type="fig"}A and B indicates that a high Ab response to the YSNIGVCK epitope is of higher significance in vaccine design for virus neutralization compared to the other B-cell epitopes used in this study. Although several design changes were made between the two animal trials outlined in this report, the individual epitope Ab concentrations were measured using the same assay. Furthermore, the significance of the YSNIGVCK motif has been implicated in previous reports by this lab as well as others [@b0120], [@b0140]. The inclusion of B-cell epitopes other than YSNIGVCK in vaccines for PEDV elicited Abs to multiple peptide sequences without any increase in VN titers. This result would make sense if virus neutralization were more dependent upon the concentration of Abs specific to the YSNIGVCK peptide sequence, meaning the addition of other epitope specific Abs have diminished returns. Since apoptosis of piglet intestinal epithelial cells is of primary concern, virus neutralization should be of the utmost importance. One possible explanation for the discrepancy in VN titers could be a fundamental difference in the GPRLQPY amino acid presentation in the context of the recombinant VLP compared to that of the native PEDV virus. If a difference were to exist, then it would be expected that Abs generated by the *E. coli* expressed proteins would have reduced or negligible binding to the PEDV virions. However, WV ELISA results ([Fig. 6](#f0030){ref-type="fig"}C) demonstrate a strong correlation between epitope Ab titers and WV titers independent of the antigen sequence. Thus it seems as though the preferential neutralization of the YSNIGVCK epitope is a true phenomenon. VN results in this report are based on systemic IgG levels in a murine model. The results may therefore not correlate directly with prevention of viral infection in the mesenteric epithelia of suckling piglets without further development. However, the results herein help demonstrate an improved efficacy of a vaccine in generating neutralizing Abs toward PEDV. Future studies building on these results should therefore include strategies to enhance lactogenic immunity from the pregnant sow in the form of secretory IgAs specific to the YSNIGVCK antigen. Appendix A. Supplementary material {#s0130} ================================== The following are the Supplementary data to this article:Supplementary data 1 Supplementary data 2 Acknowledgement {#s0110} =============== We thank the generous financial support by Smithfield Foods and Murphy-Brown LLC and the insightful discussion with their scientists, especially Dr. Terry Coffey, throughout the project period. We also wish to thank Aaron Singrey, Eric Nelson, and the rest of the team at the Animal Disease Research and Diagnostic Laboratory in the Veterinary and Biomedical Sciences Department at South Dakota State University. Conflict of interest {#s0115} ==================== None. Funding {#s0120} ======= This work was supported by Smithfield Foods and Murphy-Brown LLC. Supplementary data associated with this article can be found, in the online version, at <https://doi.org/10.1016/j.vaccine.2018.06.015>.

\section{Introduction} \label{sec:intro} Direction of arrival (DOA) estimation is the task of identifying the relative position of the sound sources with respect to the microphone. DOA estimation is a fundamental operation in microphone array processing and forms an integral part of speech enhancement~\cite{Woelfel2009}, multichannel sound source separation \cite{nikunen2014direction} and spatial audio coding~\cite{politis2015sector}. Popular approaches to DOA estimation are based on time-delay-of-arrival (TDOA)~\cite{Huang2001}, the steered-response-power (SRP) ~\cite{Brandstein1997}, or on subspace methods such as multiple signal classification (MUSIC)~\cite{Schmidt1986} and the estimation of signal parameters via rotational invariance technique (ESPRIT)~\cite{Roy1989}. The aforementioned methods differ from each other in terms of algorithmic complexity, and their suitability to various arrays and sound scenarios. MUSIC specifically is very generic with regards to array geometry, directional properties and can handle multiple simultaneously active narrowband sources. On the other hand, MUSIC and subspace methods in general, require a good estimate of the number of active sources, which are often unavailable or difficult to obtain. Furthermore, MUSIC can suffer at low signal to noise ratio (SNR) and in reverberant scenarios~\cite{dibiase2001robust}. In this paper, we propose to overcome the above shortcomings with a deep neural network (DNN) method, referred to as DOAnet, that learns the number of sources from the input data, generates high precision DOA estimates and is robust to reverberation. The proposed DOAnet also generates a spatial acoustic activity map similar to the MUSIC pseudo-spectrum (SPS) as an intermediate output. The SPS has numerous applications that rely on a directional map of acoustic activity such as soundfield visualizations~\cite{o2008imaging}, and room acoustics analysis~\cite{khaykin2012acoustic}. In comparison, the proposed DOAnet outputs the SPS and DOA's of multiple overlapping sources similar to any popular DOA estimators like MUSIC, ESPRIT or SRP without requiring the critical information of the number of active sound sources. A successful implementation of this will enable the integration of such DNN methods to higher-level learning based end-to-end sound analysis and detection systems. Recently, several DNN-based approaches have been proposed for DOA estimation~\cite{Roden2015, Xiao2015, Takeda2016,Zermini2016, Vesperini2016, Chakrabarty2017}. There are six significant differences between them and the proposed method: a) All the aforementioned works focused on azimuth estimation, with the exception of ~\cite{Vesperini2016} where the 2-D Cartesian coordinates of sound sources in a room were predicted, and~\cite{Roden2015} trained separate networks for azimuth and elevation estimation. In contrast, we demonstrate the estimation of both azimuth and elevation for the DOA by sampling the unit sphere uniformly and predicting the probability of sound source at each direction. b) The past works focused on the estimation of a single DOA at every time frame, with the exception of~\cite{Takeda2016} where localization of azimuth for up to two sources simultaneously was proposed. On the other hand, the proposed DOAnet does not algorithmically limit the number of directions to be estimated, i.e., with a higher number of audio channels input, the DOAnet can potentially estimate a larger number of sound events. c) Past works were evaluated with different array geometries making comparison difficult. Although the DOAnet can be applied to any array geometry, we evaluate the method using real spherical harmonic input signals, which is an emerging popular spatial audio format under the name Ambisonics. Microphone signals from various arrays, such as spherical, circular, planar or volumetric, can be transformed to Ambisonic signals by an appropriate transform~\cite{teutsch2007modal}, resulting in a common representation of the 3-D sound recording. Although the DOAnet is scalable to higher-order Ambisonics, in this paper we evaluate it using the compact four-channel first-order Ambisonics (FOA). d) Regarding classifiers, earlier methods have used fully connected (FC) neural networks~\cite{Roden2015, Xiao2015, Takeda2016,Zermini2016, Vesperini2016} and convolutional neural networks (CNN)~\cite{Chakrabarty2017}. In this work, along with the CNNs we use recurrent neural network (RNN) layers. The usage of RNN allows the network to learn long-term temporal information. Such an architecture is referred to as a convolutional recurrent neural network (CRNN) in literature and is the state-of-the-art method in many single-~\cite{sainath2015,Malik2017} and multichannel~\cite{Sainath2017, Adavanne2017} audio tasks. e) Previous methods used inter-channel features such as generalized cross-correlation with phase transform (GCC-PHAT)~\cite{Vesperini2016,Xiao2015}, eigen-decomposition of the spatial covariance matrix~\cite{Takeda2016}, inter-channel time delay (ITD) and inter-channel level differences (ILD)~\cite{Roden2015,Zermini2016}. More recently, Chakrabarty et al.~\cite{Chakrabarty2017} proposed to use only the phase component of the spectrogram, avoiding explicit feature extraction. In the proposed method, we use both the magnitude and the phase component. Contrary to~\cite{Chakrabarty2017}, which employed omnidirectional sensors only, general arrays with directional microphones additionally encode the DOA information in magnitude differences, while Ambisonics format especially encode directional information mainly in the magnitude component. f) All previous methods were evaluated on speech recordings that were synthetically spatialized and spatially static. We continue to use the static sound sources in the present work and extend them to a larger variety of sound events, such as impulsive and transient sounds. \section{Method} \label{sec:method} The block diagram of the proposed DOAnet is presented in Figure~\ref{fig:crnn}. The DOAnet takes multichannel audio as the input and first extracts the spectrograms of all the channels. The phases and the magnitudes of the spectrograms are mapped using a CRNN to two outputs sequentially. The first output, spatial pseudo-spectrum (SPS) is generated as a regression task, followed by the DOA estimates as a classification task. The DOA is defined by the azimuth $\phi$ and elevation $\lambda$ with respect to the microphone and the SPS is the intensity of sound along the DOA given by $S(\phi, \lambda)$. In this paper, we use discrete $\phi$ and $\lambda$ by uniformly sampling the 2-D polar coordinate space, with a resolution of 10 degrees in both azimuth and elevation, resulting in 614 sampled directions. The SPS is computed at each sampled direction, whereas, a subset of 432 directions is used for DOA, where the elevations are limited between -60 and 60 degrees. \subsection{Feature extraction} The spectrogram is calculated for each of the audio channels whose sampling frequencies are 44100 Hz. A 2048-point discrete Fourier transform (DFT) is calculated on Hamming windows of 40 ms with 50~\% overlap. We keep 1024 values of the DFT corresponding to the positive frequencies, without the zeroth bin. $L$ frames of features, each containing 1024 magnitude and phase values of the DFT extracted in all the $C$ channels, are stacked in a $L\times 1024 \times 2C$ 3-D tensor and used as the input to the proposed neural network. The $2C$ dimension results from ordering the magnitude component of all channels first, followed by the phase. We use a sequence length $L$ of 100 (= 2 s) in this work. \begin{figure}[!tb] \centering \centerline{\includegraphics[width=\columnwidth, keepaspectratio]{CRNN_DOA_SPLIT.pdf}} \caption{DOAnet - neural network architecture for direction of arrival estimation of multiple sound sources.} \label{fig:crnn} \vspace{-15pt} \end{figure} \subsection{Direction of arrival estimation network (DOAnet)} Local shift-invariant features are extracted from the input spectrogram tensor ($L\times 1024 \times 2C$ dimension) using CNN layers. In every CNN layer, the intra-channel time-frequency features are processed using a receptive field of $3\times3$, rectified linear unit (ReLU) activation and pad zeros to the resulting activation map to keep the output dimension equal to input. Batch normalization and max-pooling operation along frequency axis are performed after every CNN layer to reduce the final dimension to $L\times 2 \times N_{C}$, where $N_{C}$ is the number of CNN filters in the last CNN layer. The CNN activations are reshaped to $L\times2N_{C}$ keeping the time axis length unchanged and fed to RNN layers in order to learn temporal structure. Specifically, the bi-directional gated recurrent units (GRU) with tanh activation are used. Further, the RNN output is mapped to the first output, the SPS, in regression manner using FC layers with linear activation. The SPS is further mapped to DOA estimates--the final output of the proposed method--using a similar CRNN network as above with two minor architectural changes. An FC layer is introduced between the CNN and RNN layers to reduce the dimension of the RNN output. Additionally, the output layer which predicts the DOA uses sigmoid activation in order to estimate more than one DOA for a given time frame. Each node in this output layer represents a direction in 2-D polar space. During testing, the probabilities at these nodes are thresholded with a value of 0.5, so that anything greater suggests the presence of a source in the direction or otherwise absence of source. We refer to the combined architecture of SPS and DOA estimation in this work as DOAnet. The DOAnet is trained using the target SPS computed at each sampled direction, and for every time frame applying MUSIC (see Section~\ref{ssec:baseline}), and is represented using nonnegative real numbers. For the DOA output, the DOAnet aims to make a discrete decision about the presence of a source in a certain direction; and during training, the DOAnet uses the ground truth DOAs utilized to synthesize the audio (see Section~\ref{ssec:dataset}). The DOAnet was trained for 1000 epochs using Adam optimizer, mean squared error loss for SPS output and binary cross entropy loss for DOA output. The sum of the two losses was used for back propagation. Dropout was used after every layer and early stopping was used if the DOA metric (Section~\ref{ssec:metric}) did not improve for 100 epochs. The DOAnet was implemented using Keras framework with Theano backend. \section{Evaluation} \label{sec:eval} \subsection{Dataset} \label{ssec:dataset} In order to evaluate the proposed DOAnet, there are no publicly available real or synthetic datasets which consist of general sound events each associated with a 2D spatial coordinate. Since DNN-based methods need sufficiently large datasets to train on, most DNN-based methods proposed~\cite{Roden2015, Xiao2015, Zermini2016, Vesperini2016, Chakrabarty2017} have studied the performance on synthetic datasets. In similar fashion, we evaluate the proposed DOAnet on synthetic datasets about the same size as in the previous works. We synthesize datasets consisting of static point sources associated with a spatial coordinate in the space in two contexts - anechoic and reverberant. For each context, three datasets are generated with no temporally overlapping sources ($O1$), maximum two overlapping sources ($O2$), and maximum three overlapping sound sources ($O3$). We refer to the anechoic context dataset as $OxA$ and reverberant as $OxR$, where $x$ denotes the number of overlapping sources. Each of these datasets has three cross-validation (CV) splits with 240 recordings for training and 60 for testing. Recordings are sampled at 44.1 kHz and 30 s long. In order to generate these datasets, we use the isolated real-life sound event recordings from the DCASE 2016 task 2~\cite{dcase2016Task2}. This dataset consists of 11 sound event classes, each with 20 examples. The classes in this dataset included speech, coughing, door slam, page-turning, phone ringing and keyboard sounds. During CV, for each of the splits, we randomly chose disjoint sets of 16 and 4 examples for training and testing, amounting to 176 examples for training and 44 for testing. In order to synthesize a recording, a random subset of the 176 or 44 sound examples was chosen from the respective split. The subset size varied for each recording based on the chosen sound examples. We start synthesizing a recording by randomly choosing the beginning time of the first randomly chosen sound example within the first second of the recording. The next randomly chosen sound example is placed 250-500 ms after the end of the first sound example. On reaching the maximum recording length of 30 s, the process is repeated as many times as the number of required overlapping sound events. Each of the sound examples were assigned a DOA randomly using the following conditions. All sound events were placed in a spatial grid of ten degrees resolution along both azimuth and elevation. Two temporally overlapping sound events have at least ten degrees of spatial separation to avoid spatial overlapping. The elevation was constrained within the range of [-60, 60] degrees, as most natural sound events occur in this range. Finally, for the anechoic dataset, the sound sources were randomly placed at a distance $d$ in the range 1-10 m. For the reverberant dataset, the sound events were randomly placed inside a room of dimensions $10\times8\times4$ m with the microphone in the center of the room. Spatialization for the anechoic case was done as following. Each point source signal $s_i$ with DOA $(\phi_i,\lambda_i)$, was converted to Ambisonics format by multiplying the signal with the vector $\mathbf{y}(\phi_i,\lambda_i) = [Y_{00}(\phi_i,\lambda_i), Y_{1(-1)}(\phi_i,\lambda_i), Y_{10}(\phi_i,\lambda_i), Y_{11}(\phi_i,\lambda_i)]^\mathrm{T}$ of real orthonormalized spherical harmonics $Y_{nm}(\phi,\lambda)$. The complete anechoic sound scene multichannel recording $\mathbf{x}_\mathrm{A}$ was generated as $\mathbf{x}_\mathrm{A} = \sum_{i}g_i s_i \mathbf{y}(\phi_i,\lambda_i)$, with the gains $g_i<1$ modeling the distance attenuation. Each entry of $\mathbf{x}_\mathrm{A}$ corresponds to one channel and $g_i = \sqrt{1 / 10^{d/d_{max}}}$, where $d_{max} = 10$ m is the maximum distance. In the reverberant case, a fast geometrical acoustics simulator was used to model natural reverberation based on the rectangular room image-source model \cite{Allen1979}. For each point source $s_i$ with DOA in the dataset, $K$ image sources were generated modeling reflections up to a predefined time-limit. Based on the room and its propagation properties, each image source was associated with a propagation filter $h_{ik}$ and DOA $(\phi_k,\lambda_k)$ resulting in the spatial impulse response $\mathbf{h}_i = \sum_{k=1}^K h_{ik} \mathbf{y}(\phi_k,\lambda_k).$ The reverberant scene signal was finally generated by $\mathbf{x}_\mathrm{R} = \sum_{i} s_i*\mathbf{h}_i,$ where $(*)$ denotes convolution of the source signal with the spatial impulse responses. The room absorption properties were adjusted to match reverberation times of typical office spaces. Three sets of testing data were generated with similar room size as training data (Room 1), $80\%$ of room size ($8\times8\times4$ m) and reverberation time (Room 2), and $60\%$ of room size ($8\times6\times4$ m) and reverberation time (Room 3). \subsection{Baseline} \label{ssec:baseline} The proposed method to our knowledge is the first DNN-based implementation for 2D DOA estimation of multiple overlapping sound events. Thus in order to evaluate the complete features of the proposed DOAnet, we compare the performance with the conventional, high-resolution DOA estimator based on MUSIC. Similar to the SPS and DOA outputs estimated by the DOAnet, the MUSIC method also estimates SPS and DOA, thus allowing a direct one-to-one comparison. The MUSIC SPS is based on a measure of orthogonality between the signal subspace (dominated by the source signals) of the spatial covariance matrix $\mathbf{C_s}$ and the noise subspace (dominated by diffuse and ambient sounds, late reverberation, and microphone noise). The spatial covariance matrix is calculated as $\mathbf{C_s} = \mathbb{E}_{f,t}\left[\mathbf{X}(f,t)\mathbf{X}(f,t)^\mathrm{H}\right]$, where spectrogram $\mathbf{X}(f,t)$ is a frequency $f$ and time $t$ dependent $C$-dimensional vector, where $C$ is the number of channels, $^\mathrm{H}$ is the conjugate transpose and $\mathbb{E}_{f,t}$ denotes the expectation over $f$ and $t$. For a sound scene with $O$ number of sources, the MUSIC SPS $S_{GT}$ is obtained from $\mathbf{C_s}$ by first performing an eigenvalue decomposition on $\mathbf{C_s} = \mathbf{E}\mathbf{\Lambda}\mathbf{E}^\mathbf{H}$. The sorted eigenvectors $\mathbf{E}$ (according to eigenvalues with decreasing magnitude) are further partitioned into the two aforementioned subspaces $\mathbf{E} = [\mathbf{U}_\mathrm{s}\; \mathbf{U}_n]$, where $\mathbf{U}_\mathrm{s}$ denotes the signal subspace and will be composed of $O$ eigenvectors corresponding to the higher eigenvalues and the rest will form the noise subspace $\mathbf{U}_\mathrm{n}$. The $S_{GT}$ along the direction $(\phi_i,\lambda_i)$ is now given by $S_{GT}(\phi_i,\lambda_i) = 1/(\mathbf{y}^\mathrm{T}(\phi_i,\lambda_i)\mathbf{U}_n \mathbf{U}_n^\mathrm{H}\mathbf{y}(\phi_i,\lambda_i))$. Finally, the source DOAs are found by selecting the directions $(\phi_i,\lambda_i)$ corresponding to the $O$ largest peaks from $S_{GT}$. \subsection{Metric} \label{ssec:metric} The DOAnet estimated SPS ($S_{E}(\phi, \lambda)$) is evaluated with respect to the baseline MUSIC estimated ground truth ($S_{GT}(\phi, \lambda)$) using the SNR metric calculated as $SNR = 10 \log_{10} (\sum_{\phi}\sum_{\lambda}S_{GT}(\phi, \lambda)^2 / \sum_{\phi}\sum_{\lambda} (S_{E}(\phi, \lambda) - S_{GT}(\phi, \lambda))^2)$. As the DOA metric we use the angle between the estimate DOA (defined by azimuth $\phi_E$ and elevation $\lambda_E$) and the ground truth DOA ($\phi_{GT}$, $\lambda_{GT}$) used to synthesize the dataset in degrees. This is calculated as $\sigma = \arccos (\sin\phi_E\sin\phi_{GT} + \cos\phi_E\cos\phi_{GT} \cos(\lambda_{GT}-\lambda_E)) \cdot 180.0 / \pi$. Further, to accommodate the scenario of unequal number of estimated and ground truth DOAs we calculate and report the minimum distance between them using the Hungarian algorithm~\cite{Hungarian} along with the percentage of frames in which the number of DOAs estimated were correct. The final metric for the entire dataset, referred as DOA error, is calculated by normalizing the minimum distance with the total number of estimated DOA's. \begin{figure}[!b] \vspace{-15pt} \begin{minipage}[b]{0.48\linewidth} \centering \centerline{\includegraphics[height=4cm,width=\columnwidth,trim={1.5cm 3cm 2cm 3cm},clip,keepaspectratio]{gt_heatmap}} \centerline{(a) MUSIC estimated }\medskip \end{minipage} \begin{minipage}[b]{0.48\linewidth} \centering \centerline{\includegraphics[height=4cm,width=\columnwidth, trim={1.5cm 3cm 2cm 3cm},clip,keepaspectratio]{pred_heatmap}} \centerline{(b) DOAnet estimated}\medskip \end{minipage} \vspace{-15pt} \caption{SPS for two closely located sound sources. The black-cross markers represent the ground truth DOA. The horizontal axis is azimuth and vertical axis is elevation angle (in degrees)} \label{fig:spm} \end{figure} \subsection{Evaluation procedure} The parameter tuning for DOAnet was performed on the $O1A$ test data, and the best configuration is as shown in Figure~\ref{fig:crnn}. This configuration has 677 K weights, and the same configuration is used in all of the following studies. At test time, the SNR metric for SPS output of the DOAnet ($S_{E}$) is calculated with respect to SPS of baseline MUSIC ($S_{GT}$). The DOA metric for the DOAs predicted by DOAnet and baseline MUSIC are calculated with respect to the ground truth DOA used to synthesize the dataset. In the above experiment, the baseline MUSIC algorithm uses the knowledge of the number of active sources. In order to have a fair evaluation, we test the DOAnet in a similar scenario where the number of sources is known. We use this knowledge to choose the top probabilities in prediction layer of the DOAnet instead of thresholding it with a value of 0.5. \begin{table}[!tb] \caption{Evaluation metric scores for the spatial power map and DOAs estimated by the DOAnet for different datasets.} \label{T2:results} \resizebox{\columnwidth}{!}{ \centering \begin{tabular}{l|ccc|ccc} & \multicolumn{3}{c|}{Anechoic} & \multicolumn{3}{c}{Reverberant (Room 1)} \\\hline \begin{tabular}[c]{@{}l@{}}Max. no. of \\ overlapping sources\end{tabular} & 1 & 2 & 3 & 1 & 2 & 3 \\ \hline SPS SNR (in dB) & 9.90 & 3.35 & -0.26 & 3.11 & 1.24 & 0.13 \\ \multicolumn{7}{l}{\multirow{2}{*}{}} \\ \multicolumn{7}{l}{DOA error with unknown number of active sources (threshold of 0.5)} \\\hline DOAnet & 0.57 & 8.03 & 18.34 & 6.31 & 11.46 & 38.41 \\ \begin{tabular}[c]{@{}l@{}}Correctly predicted \\ frames (in \%)\end{tabular} & 95.4 & 42.7 & 1.8 & 59.3 & 15.8 & 1.2 \\ \multicolumn{7}{l}{\multirow{2}{*}{}} \\ \multicolumn{7}{l}{DOA error with known number of active sources} \\\hline DOAnet & 1.14 & 27.52 & 49.30 & 12.61 & 38.98 & 67.07 \\ MUSIC & 2.29 & 8.60 & 28.66 & 25.80 & 57.33 & 91.72 \end{tabular} } \vspace{-15pt} \end{table} \begin{figure}[!hb] \vspace{-15pt} \hspace{11pt} \begin{minipage}[b]{0.10\linewidth} \centering \centerline{\includegraphics[height=2cm,trim={8cm 2.5cm 8cm 3.5cm},clip,keepaspectratio]{o1ae}} \centerline{(a) $O1A$}\medskip \end{minipage} \hspace{5pt} \begin{minipage}[b]{0.35\linewidth} \centering \centerline{\includegraphics[height=2cm, trim={6cm 2.5cm 8cm 3.5cm},clip,keepaspectratio]{o2ae}} \centerline{(b) $O2A$}\medskip \end{minipage} \hspace{-10pt} \begin{minipage}[b]{0.25\linewidth} \centering \centerline{\includegraphics[height=2cm, trim={8.5cm 2.5cm 8cm 3.5cm},clip,keepaspectratio]{o1e}} \centerline{(c) $O1R$}\medskip \end{minipage} \hspace{-10pt} \begin{minipage}[b]{0.25\linewidth} \centering \centerline{\includegraphics[height=2cm, trim={8.5cm 2.5cm 8cm 3.5cm},clip,keepaspectratio]{o2e}} \centerline{(d) $O2R$}\medskip \end{minipage} \vspace{-25pt} \caption{Confusion matrix for the number of DOA estimated per frame by the DOAnet. The horizontal axis is the DOAnet estimate, and the vertical axis is the ground truth.} \label{fig:conf_mat} \end{figure} \section{Results and Discussion} \label{sec:results} The results of the evaluations are presented in Table~\ref{T2:results}. The high SNRs for SPS in both the contexts, with up to one and two overlapping sound events show that the SPS generated by DOAnet ($S_{E}$) is comparable with the baseline MUSIC SPS ($S_{GT}$). Figure~\ref{fig:spm} shows the $S_{E}$ and the respective $S_{GT}$ when two active sources are closely located. In the case of up to three overlapping sound events, the baseline MUSIC is already at its theoretical limit of estimating $N-1$ sources from $N$-dimensional signal space~\cite{Ottersten1993}. In practice, for $N-1$ sources only one noise subspace vector $\mathbf{U}_n$ is used to generate SPS, which for real signals is too weak for stable estimation. In the present evaluation of DOAnet which is trained with four-channel audio features and MUSIC SPS, for the case of three overlapping sound sources the SPS used is an unstable estimate resulting in poor training and consequently the results. With more than four-channels input, which the proposed DOAnet can easily extend to, it can potentially localize more than two sound sources simultaneously. The DOA error for the proposed DOAnet when the number of active sources are unknown is presented in Table~\ref{T2:results}. The DOAnet error is considerably better in comparison to the baseline MUSIC that uses the active sources knowledge for all datasets. However, the number of frames in which DOAnet produced the correct number of active sources were few. For example, in the case of anechoic recordings with up to two overlapping sound events, only 42.7\% of the estimated frames had the correct number of DOA predictions. This prediction drops even drastically when the number of sources is three, due to the theoretical limit of MUSIC as explained previously, and consequently for the DOAnet as MUSIC SPS is used for training. Finally, the confusion matrix for the number of DOA estimates per frame for $O1$ and $O2$ datasets are visualized in Figure~\ref{fig:conf_mat}. We skipped the confusion matrices for the $O3$ datasets as they were not meaningful for similar reasons as explained above. With the knowledge of the number of active sources (Table~\ref{T2:results}), the DOAnet performs considerably better than baseline MUSIC for all datasets other than the $O2A$ and $O3A$. The MUSIC DOA's were chosen using a 2D peak finder on the MUSIC SPS, whereas the DOA's in DOAnet were chosen by simply picking the top probabilities in the final DOA prediction layer. A smarter peak picking method from the DOAnet, or using the number of sources as an additional input can potentially result in better scores across all datasets. Further, the DOAnet error on unmatched reverberant data is presented in Table~\ref{T3:reverb}. The performance of DOAnet is seen to be consistent in comparison to the matched reverberant data in Table~\ref{T2:results}, and significantly better than the performance of MUSIC. In this paper, since the baseline was chosen to be MUSIC, for a fair comparison the DOAnet was also trained using MUSIC SPS. In an ideal scenario, considering the DOAnet is trained using datasets for which the ground truth DOAs are known, we can generate accurate high-resolution SPS from the ground truth DOA's as per the required application and use them for training. Alternatively, the DOAnet can be trained without the SPS to directly generate the DOAs, it was only used in this paper to present the complete potential of the method in the limited paper space. In general, the above results show that the proposed DOAnet has the potential to learn the 2D direction information of multiple overlapping sound sources directly from the spectrogram of the input audio without the knowledge of the number of active sound sources. An exhaustive study with more detailed experiments including both synthetic and real datasets are planned for future work. \begin{table}[!t] \caption{Evaluation scores for unmatched reverberant room.} \label{T3:reverb} \centering \resizebox{0.9\columnwidth}{!}{ \begin{tabular}{l|cc|cc} & \multicolumn{2}{c|}{Room 2} & \multicolumn{2}{c}{Room 3} \\\hline Max. no. of overlapping sources & 1 & 2 & 1 & 2 \\ \hline SPS SNR (in dB) & 3.53 & 1.49 & 3.49 & 1.46 \\ \multicolumn{5}{l}{\multirow{2}{*}{}} \\ \multicolumn{5}{l}{DOAnet error (Unknown number of sources)} \\\hline DOAnet & 3.44 & 6.88 & 4.59 & 10.89\\ Correctly predicted frames (in \%) & 46.2 & 14.3 & 49.7 & 14.1 \\ \multicolumn{5}{l}{\multirow{2}{*}{}} \\ \multicolumn{5}{l}{DOA error (Known number of sources)} \\\hline DOAnet & 8.60 & 32.10 & 9.17 & 33.82\\ MUSIC & 31.52 & 58.47 & 33.25 & 60.76 \end{tabular} } \vspace{-15pt} \end{table} \section{Conclusion} \label{sec:conclusion} A convolutional recurrent neural network (DOAnet) was proposed for multiple source localization. The DOAnet was shown to learn the number of active sources directly from the input spectrogram, and estimate precise DOA in 2-D polar space. The method was evaluated on anechoic, matched and unmatched reverberant dataset. The proposed DOAnet performed considerably better than baseline MUSIC in most scenarios. Thereby showing the potential of DOAnet in learning highly computational algorithm without prior knowledge of the number of sources. \bibliographystyle{IEEEtran}

\section{Introduction} Exactly-solvable systems play an important role for the understanding of the physics of quantum many-body systems. They offer an insight into the behaviour of strongly correlated systems in ways that would otherwise be impossible. One class of these systems is the class of Richardson-Gaudin (RG) integrable systems, which can be derived from a generalized Gaudin algebra \cite{dukelsky_colloquium:_2004,ortiz_exactly-solvable_2005}. The pairing model in the reduced Bardeen-Cooper-Schrieffer (BCS) approximation, used to describe superconductivity, has been shown to be RG integrable \cite{cambiaggio_integrability_1997}, as has the $p_x + i p_y$ pairing Hamiltonian \cite{ibanez_exactly_2009,rombouts_quantum_2010}, the central spin model \cite{gaudin_diagonalisation_1976}, factorisable pairing models in heavy nuclei \cite{dukelsky_exactly_2011}, an extended $d+id$ pairing Hamiltonian \cite{marquette_integrability_2013} and several atom-molecule Hamiltonians such as the inhomogeneous Dicke model \cite{dukelsky_exactly_2004}. For these models, diagonalizing the Hamiltonian in an exponentially scaling Hilbert space can be reduced to solving a set of non-linear equations scaling linearly with system size. For most of these systems, computationally inexpensive expressions are also known for several form factors and overlaps, which can be used to investigate observables in systems where the Hilbert space is too large for traditional exact methods. Unfortunately, the Bethe Ansatz or Richardson-Gaudin equations \cite{richardson_exact_1964,gaudin_diagonalisation_1976,dukelsky_class_2001} that need to be solved are highly non-linear and give rise to singularities, making a straightforward numerical solution challenging \cite{richardson_numerical_1966,dominguez_solving_2006}. Several methods have been introduced as a way to resolve this difficulty, such as a change of variables \cite{rombouts_solving_2004,dominguez_solving_2006}, a (pseudo-)deformation of the algebra \cite{de_baerdemacker_richardson-gaudin_2012,van_raemdonck_exact_2014}, or a Heine-Stieltjes connection, reducing the problem to a differential equation \cite{guan_heine-stieltjes_2012}. The ground state energy in the thermodynamic limit has also been obtained by treating the interaction as an effective temperature \cite{pogosov_probabilistic_2012}. Our method consists of a generalization of the numerical method first proposed by Faribault et al. \cite{faribault_gaudin_2011} for a set of non-degenerate rational RG models, and later extended to degenerate rational models by El Araby, Gritsev and Faribault \cite{el_araby_bethe_2012}. The equations for the Dicke model were independently presented by Babelon and Talalaev \cite{babelon_bethe_2007}. In this method, an alternative set of equations is derived in terms of variables related to the eigenvalues of the constants of motion. One of the advantages of this method is that it offers not only the solutions of the RG equations, but also efficient numerical expressions for form factors and overlaps \cite{faribault_determinant_2012,tschirhart_algebraic_2014}. Such expressions for the rational model have already been used to study quantum quenches \cite{faribault_quantum_2009} and decoherence in quantum dots \cite{van_den_berg_competing_2014}, and should facilitate the use of the XXZ model as a variational or projected approach for non-integrable systems, similar to the use of the XXX model in quantum chemistry \cite{johnson_size-consistent_2013,limacher_new_2013,tecmer_assessing_2014}. In the following, the results for the rational XXX model are generalized towards the full class of XXZ RG integrable systems and possible applications are discussed. \section{Richardson-Gaudin models} \subsection{Definitions} Unlike its classical counterpart, no single unique definition is known for quantum integrability \cite{caux_remarks_2011}. One class of systems usually denoted integrable is the set of RG integrable systems, since these support as many (nontrivial) conserved operators commuting with the Hamiltonian as there are degrees of freedom in the system \cite{dukelsky_colloquium:_2004,links_algebraic_2003}. These systems can be diagonalized exactly by means of a Bethe Ansatz wave function, where diagonalizing the full Hamiltonian can be reduced to solving a set of nonlinear coupled equations for the variables in the ansatz \cite{richardson_exact_1964,gaudin_diagonalisation_1976}. The families of RG integrable systems have their roots in a generalized Gaudin algebra \cite{gaudin_diagonalisation_1976,ortiz_exactly-solvable_2005} based on the $su(2)$ algebra of (quasi-)spin operators \cite{talmi_simple_1993}. Depending on the physical systems, they can either represent spin degrees of freedom, fermion pairs or bosonic/fermionic Schwinger representations. The relevant (quasi-)spin operators can be defined as satisfying the following commutation relations \begin{eqnarray} [S_i^0,S^{\dagger}_j]&=&\delta_{ij}S^{\dagger}_i, \qquad [S_i^0,S_j]=-\delta_{ij}S_i, \nonumber\\ {}[S^{\dagger}_i,S_j]&=&2\delta_{ij}S_i^0, \end{eqnarray} where each separate algebra spans a $su(2)_i$ algebra associated with a level $i$ and irreducible representations (irrep) $\ket{d_i,\mu_i}$. For a set of $n$ levels, the RG integrable models are then defined by a set of $n$ mutually commuting operators \begin{equation} R_i=S_i^0+g \sum_{k \neq i}^n\left[\frac{1}{2}X_{ik}(S^{\dagger}_i S_k+S_iS^{\dagger}_k)+Z_{ik}S_i^0 S_k^0\right] \end{equation} Following Gaudin \cite{gaudin_diagonalisation_1976} and Dukelsky et al. \cite{dukelsky_class_2001}, it is possible to obtain a set of conditions for which the set of operators $R_i$ commute mutually. These are given by \begin{eqnarray}\label{ga:gaeq} X_{ij}=-X_{ji}, \qquad Z_{ij}=-Z_{ji}, \\ X_{ij}X_{jk}-X_{ik}(Z_{ij}+Z_{jk})=0, \end{eqnarray} the so-called Gaudin equations. They were originally discovered by Gaudin as defining a general class of quadratic Hamiltonians in the spin variables, among which the Gaudin magnet. The derivation by Dukelsky et al. differs from the derivation by Gaudin in the presence of the generator of the Cartan subalgebra $S_i^0$, which however does not influence these conditions. In fact, the Gaudin models are given by Dukelsky's constants of motion in the limit $g \to \infty$ . Gaudin mentioned three classes of solutions of these equations, where all classes consider $X_{ij}$ and $Z_{ij}$ as odd functions of some real arbitrary parameters $X_{ij}=X(\epsilon_i,\epsilon_j)$. The physical interpretation of these parameters follows from the expression for the Hamiltonian constructed with these parameters. \begin{enumerate} \item The rational model \begin{equation}\label{ga:rat} X_{ij} = Z_{ij} = \frac{1}{\epsilon_i-\epsilon_j} \end{equation} \item The trigonometric model \begin{equation}\label{ga:trig} X_{ij} = \frac{1}{\sin(\epsilon_i-\epsilon_j)}, \qquad Z_{ij}=\cot(\epsilon_i-\epsilon_j) \end{equation} \item The hyperbolic model \begin{equation} X_{ij} = \frac{1}{\sinh(\epsilon_i-\epsilon_j)}, \qquad Z_{ij}=\coth(\epsilon_i-\epsilon_j) \end{equation} \end{enumerate} Here the rational model is also referred to as the XXX-model \cite{dukelsky_class_2001}, indicating that the coefficients in the expression for the constants of motion are identical for all 3 components of the spin. In the same vein, the trigonometric and hyperbolic models are called XXZ-models. Alternative solutions are given by \begin{eqnarray}\label{ga:richalg} X_{ij}&=&\frac{\sqrt{1+2\alpha\epsilon_i+\beta \epsilon_i^2}\sqrt{1+2\alpha\epsilon_j+\beta \epsilon_j^2}}{\epsilon_i-\epsilon_j},\nonumber\\ Z_{ij}&=&\frac{1+\alpha(\epsilon_i+\epsilon_j)+\beta \epsilon_i\epsilon_j}{\epsilon_i-\epsilon_j}, \end{eqnarray} which was proposed by Richardson \cite{richardson_new_2002}, and \begin{equation}\label{ga:hyp} X_{ij}=\frac{2\sqrt{\epsilon_i\epsilon_j}}{\epsilon_i-\epsilon_j}, \qquad Z_{ij}=\frac{\epsilon_i+\epsilon_j}{\epsilon_i-\epsilon_j}, \end{equation} which is a reparametrization of the hyperbolic model \cite{ortiz_exactly-solvable_2005}. \subsection{Diagonalizing integrable Hamiltonians} Any general RG integrable Hamiltonian can be written as a linear combination of the constants of motion \begin{equation}\label{RG:hamiltonian} \hat{H}=\sum_{i=1}^n \eta_i R_i, \end{equation} resulting in a wide variety of systems. The reduced BCS Hamiltonian can be found from the rational model \cite{cambiaggio_integrability_1997}, the $p_x+ip_y$ pairing Hamiltonian can be derived from the hyperbolic model \cite{ibanez_exactly_2009,rombouts_quantum_2010}, and the central spin model Hamiltonian is identical to one of the constants of motion of the rational model \cite{gaudin_diagonalisation_1976}. By introducing a bosonic degree of freedom, the inhomogeneous Dicke model can be found as a limiting case of the XXZ model \cite{dukelsky_exactly_2004}. Since all constants of motion commute mutually, they also commute with any Hamiltonian that can be written as Eq. (\ref{RG:hamiltonian}), so this reduces the diagonalization of the Hamiltonian to the diagonalization of one of the constants of motion. It can be shown that the eigenstates are given by a Bethe Ansatz \begin{equation}\label{rg:bas} \ket{\psi_N}=\prod_{\alpha=1}^N\left(\sum_{i=1}^n X_{i\alpha}S^{\dagger}_i\right)\ket{\theta}, \end{equation} with $N$ the number of excitations \footnote{With excitations we denote spin-up flips for magnetic (spin) systems, and the number of pairs for pairing (quasispin) systems.} and the vacuum state $\ket{\theta}=\otimes_{i=1}^n \ket{d_i,-d_i}$ the tensor product of the lowest-weight irreps of each $su(2)_i$ copy, provided the RG equations \begin{equation}\label{ga:RGeq} 1+g\sum_{i=1}^n Z_{i\alpha}d_i-g\sum_{\beta \neq \alpha}^N Z_{\beta\alpha}=0, \ \ \forall \alpha=1\dots N, \end{equation} are satisfied, where the algebra has been extended to $X_{i\alpha}=X(\epsilon_i,x_{\alpha})$ and $Z_{i\alpha}=Z(\epsilon_i,x_{\alpha})$ by introducing a set of $N$ parameters $\{x_{\alpha}\}$. These parameters, also known as the RG variables or rapidities, fix the wave function and need to be determined from the RG equations (\ref{ga:RGeq}). There are multiple strategies to derive this result. Richardson obtained these equations for the reduced level-independent BCS model starting from a variational approach \cite{richardson_restricted_1963,richardson_exact_1964}, Gaudin started from integrability constraints \cite{gaudin_diagonalisation_1976}, Zhou et al. derived these results starting from the Algebraic Bethe Ansatz \cite{zhou_superconducting_2002} and Ortiz et al. used a commutator scheme based on the Gaudin algebra \cite{ortiz_exactly-solvable_2005}. Once the RG equations have been solved and the rapidities determined, the eigenvalues of the constants of motion $R_i$ can be evaluated as \begin{equation} r_i=d_i\left(-1+g\sum_{k\neq i}^nZ_{ik}d_k+g\sum_{\beta=1}^N Z_{\beta i}\right), \ \ \forall i=1 \dots n. \end{equation} Although the set of RG equations seems to be linear in the elements $Z_{i\alpha}$ and $Z_{\alpha\beta}$, these variables are still coupled through the Gaudin equations, leading to a set of coupled and nonlinear equations. As an example, the RG equations for the doubly-degenerate ($d_i=1/2, \forall i$) XXX model are given by \begin{equation}\label{rg:bcs} 1+\frac{g}{2}\sum_{i=1}^n \frac{1}{\epsilon_i-x_{\alpha}} - g \sum_{\beta \neq \alpha}^N \frac{1}{x_{\beta}-x_{\alpha}}=0, \ \ \forall \alpha=1 \dots N. \end{equation} These equations need to be solved for the set of rapidities $\{x_{\alpha}\}$ and are coupled and highly nonlinear. It has been shown that singular points arise in these equations at certain values of the coupling constant, where multiple rapidities $x_{\alpha}$ equal one of the single-particle levels $\epsilon_i$ \cite{richardson_numerical_1966,rombouts_solving_2004,dominguez_solving_2006}. It can be readily seen that both the second term and the third term in the RG equations contain diverging poles, but it can also be shown that these singularities cancel exactly. Unfortunately, the behaviour arising at these so-called singular points hampers straightforward solutions of these equations, so involved numerical methods have to be found to circumvent these points. \section{An eigenvalue-based numerical method} \subsection{Doubly degenerate models} For the XXX spin-$1/2$ model ($d_i=1/2, \forall i$), it is possible to introduce a new set of variables \begin{equation} \Lambda_i \equiv \Lambda(\epsilon_i)=\sum_{\alpha=1}^N\frac{1}{\epsilon_i-x_{\alpha}}, \end{equation} circumventing the singular points in the Richardson-Gaudin equations \cite{faribault_gaudin_2011}. A set of equations equivalent to the RG equations can be found for these variables, however void of singular behaviour. As an illustration, the set of RG equations for the doubly-degenerate XXX model (Eq. \ref{rg:bcs}) are isomorphic to the set of quadratic equations\cite{faribault_gaudin_2011} \begin{equation} \Lambda_i^2=-\frac{2}{g}\Lambda_i+\sum_{j \neq i}^n \frac{\Lambda_j-\Lambda_i}{\epsilon_j-\epsilon_i}, \ \ \forall i=1 \dots n. \end{equation} A straightforward numerical solution of these equations can easily be implemented and no singular behaviour will arise since no variables occur in the denominator. We wish to extend these results to the class of more general XXZ models, relying only on the Gaudin algebra and not on an explicit rational expression of $X$ and $Z$. Define \begin{equation} \Lambda_i \equiv \sum_{\alpha=1}^NZ_{i\alpha}=\sum_{\alpha=1}^NZ(\epsilon_i,x_{\alpha}), \end{equation} which reduces to the previously-introduced variables in the XXX model. It is interesting to note that the eigenvalue $r_i$ of the constant of motion $R_i$ is related to the variable $\Lambda_i$ as \begin{equation} r_i=d_i\left(-1-g\Lambda_i+g\sum_{k\neq i}^n Z_{ik}d_k\right), \end{equation} which has led to the denomination eigenvalue-based variables. In the following, we will start from the Richardson-Gaudin equations for spin-1/2 systems \begin{equation} 1+\frac{g}{2}\sum_{i=1}^nZ_{i\alpha} = g \sum_{\beta \neq \alpha}^NZ_{\beta\alpha}, \qquad \forall \alpha=1 \dots N, \end{equation} and several properties that can be derived from the Gaudin algebra, as previously obtained by Ortiz et al. \cite{ortiz_exactly-solvable_2005}. Firstly, it has been shown that \begin{equation} (X_{ij})^2-(Z_{ij})^2=\Gamma, \qquad \forall i \neq j, \end{equation} where $\Gamma$ is a constant. The rational model corresponds to $\Gamma=0$, while positive and negative $\Gamma$ result in the trigonometric and hyperbolic models respectively. It also follows from the Gaudin algebra that \begin{equation}\label{ga:Zeq} Z_{ij}Z_{jk} - Z_{ik}(Z_{ij}+Z_{jk})=\Gamma, \qquad \forall i \neq j \neq k \neq i, \end{equation} which is a generalization of the Gaudin equations for the rational model, where $X_{ij}=Z_{ij}$ and $\Gamma=0$. Using only these equations, it can straightforwardly be shown that \begin{equation}\label{ebv:eq} \Lambda_i^2=N(n-N)\Gamma-\frac{2}{g}\Lambda_i+\sum_{j \neq i}^n Z_{ji}(\Lambda_j-\Lambda_i), \ \ \ \forall i=1 \dots n \end{equation} with $N$ the total number of excitations and $n$ the number of single-particle levels. The full derivation can be found in Appendix \ref{appendixA}. Unlike the case of the rational model, these equations depend explicitly on the number of excitations $N$. When solving these equations numerically, it was found that the total number of solutions exceeds the dimension of the Hilbert space for $N$ excitations distributed over $n$ levels. Therefore, these equations necessarily support unphysical solutions, not corresponding to any eigenstate, implying that this new set of equations is not equivalent to the original set of RG equations (Eq. \ref{ga:RGeq}). In order to obtain a set of equations equivalent to the original equations, additional constraints for the $\Lambda_i$ are needed. It is clear from the $\Gamma=0$ case (XXX) that the new set of equations can not distinguish between the different excitation sectors $N$. This can be imposed by noting that the sum of all constants of motion is given by the operator counting the number of excitations \begin{equation} \sum_{i=1}^nR_i=\sum_{i=1}^n S_i^0. \end{equation} Writing out the eigenvalues of the constants of motion in the new variables results in \begin{equation}\label{ebv:sum} -\frac{g}{2}\sum_{i=1}^n \Lambda_i=N. \end{equation} In the following section it will be shown that the full set of equations (\ref{ebv:eq}) and (\ref{ebv:sum}) has as many solutions as the dimension of the Hilbert space, so introducing this additional equation leads to a system of equations fully equivalent to the original set of RG equations. These eigenvalue-based equations do not show singular behaviour and can easily be solved numerically, as will be shown in the following. \subsection{The weak-coupling limit} In order to obtain some insight in the behaviour of the solutions of these equations, an approximate solution can be found in the weak-coupling limit (small $|g|$). This can be done by proposing a series expansion in $g$ for the solutions of these equations and solving the equations at each order. For small $g$, a series expansion of $\Lambda_i$ in $g$ can be proposed up to $\mathcal{O}(g)$, keeping only the two dominant terms \begin{equation} \Lambda_i = \frac{\lambda_i^{(-1)}}{g}+\lambda_i^{(0)}+\mathcal{O}(g). \end{equation} By plugging this expansion in the equations, we obtain \begin{eqnarray} &\frac{1}{g^2}\lambda_i^{(-1)}\left(\lambda_i^{(-1)}+2\right)\nonumber\\ &+\frac{1}{g}\left(2\lambda_i^{(0)}(1+\lambda_i^{(-1)})-\sum_{j \neq i}^nZ_{ji}\left[\lambda_j^{(-1)}-\lambda_i^{(-1)}\right]\right) \nonumber\\ &+\mathcal{O}(g^0)=0 \end{eqnarray} This results in \begin{equation} \lambda_i^{(-1)}=0 \ \ \textrm{or} \ -2 \end{equation} and \begin{equation} \lambda_i^{(0)}=\frac{1}{2\lambda_i^{(-1)}+2}\sum_{j \neq i}^n Z_{ji}\left(\lambda_j^{(-1)}-\lambda_i^{(-1)}\right). \end{equation} In order to satisfy Eq. (\ref{ebv:sum}), the number of dominant terms different from $0$ has to equal the number of excitations $N$, resulting in $\binom{n}{N}$ solutions. The total number of solutions then equals the dimension of the Hilbert space for $N$ excitations distributed over $n$ doubly degenerate levels. Any solution in the weak-coupling limit can be adiabatically connected to a solution for arbitrary coupling, indicating that all possible solutions are always found and no unphysical solutions are present. This series expansion can also be connected to the series expansion obtained for the rapidities \cite{ortiz_exactly-solvable_2005}. In the limit $g \to 0$ we obtain a noninteracting model, where the rapidities $x_{\alpha}$ converge to the parameters $\epsilon_i$, depending on the corresponding distribution of excitations over energy levels. A rapidity converging to $\epsilon_i$ then corresponds to an excited level $i$ in the noninteracting limit. For finite but small $g$, dominant corrections of $\mathcal{O}(g)$ are present, which are proportional to the roots of orthogonal polynomials via a Heine-Stieltjes connection \cite{stieltjes_theoreme_1885,sriram_shastry_solution_2001}. For $x_{\alpha}$ converging to $\epsilon_i$, this results in $Z_{i\alpha}=Z(\epsilon_i,x_{\alpha})$ diverging as $1/g$ in the weak-coupling limit, where the proportionality factor can be found to be $-2$ from the Heine-Stieltjes connection for $d_i=1/2$. For all other levels $j \neq i$, $Z_{j\alpha}$ remains finite and the dominant order is $\mathcal{O}(g^0)$. A dominant order of $\mathcal{O}(1/g)$ in $\Lambda_i$ then corresponds to an excited state $i$ in the noninteracting limit, while a dominant order of $\mathcal{O}(g^0)$ results in an non-excited state $i$. This behaviour can be generalized through a connection of $\Lambda_i$ to occupation numbers, as will be shown in the section about form factors. Note that the divergence in $g \to 0$ can pose numerical problems, so this suggests the use of $g\Lambda_i$ as variables instead of $\Lambda_i$. \subsection{Solving the equations} Due to the similarity of Eq. (\ref{ebv:eq}) to the set of equations found for the rational model \cite{faribault_gaudin_2011}, it is straightforward to extend the solution method for the rational model to our equations. General sets of nonlinear equations have to be solved by an iterative approach starting from an initial guess, such as the Newton-Raphson method. This method converges quadratically to the solution if the initial guess lies in the basin of attraction. An efficient numerical approach can be implemented based on this method once we have access to a sufficiently good initial guess for the solution. As in [\onlinecite{faribault_gaudin_2011}], we start from an approximation of the solution in the weak-coupling limit ($|g|\ll$). A solution at any value of the coupling constant can then be found by adiabatically varying $g$ starting from the weak-coupling limit and using the solution at the previous step as the starting point for an iterative solution at the current step. This initial guess can be improved by using a Taylor expansion of the solutions at the previous step, \begin{equation} \Lambda_i(g+\delta g) \approx \Lambda_i(g)+\delta g \frac{\partial \Lambda_i}{\partial g} \end{equation} since the derivatives of the $\Lambda_i$ can be found by solving a linear system once the set of $\Lambda_i$ is known, similar to the procedure followed in [\onlinecite{faribault_gaudin_2011}]. The equations for the derivatives can be found by deriving the set of equations to $g$, leading to \begin{equation} \Lambda_i \frac{\partial \Lambda_i}{\partial g}=\frac{\Lambda_i}{g^2}-\frac{1}{g}\frac{\partial \Lambda_i}{\partial g}+\frac{1}{2}\sum_{j \neq i}^n Z_{ji}\left(\frac{\partial \Lambda_j}{\partial g}-\frac{\partial \Lambda_i}{\partial g}\right) \end{equation} By taking the higher derivatives of the original set of equations, linear equations can be found for higher order derivatives of $\Lambda_i$, which allows a Taylor expansion up to arbitrary order. For an efficient numerical implementation, combining the Newton-Raphson method with a Taylor approximation up to first order already offers a remarkable increase in speed. \subsection{Inverting the transformation} Although many properties of the XXZ systems can be written in terms of the eigenvalue-based variables $\Lambda_i$, as will be shown in the following section, it is still necessary to obtain the rapidities from these variables for the evaluation of e.g. certain form factors. For the rational model, Slavnov's determinant formula \cite{slavnov_calculation_1989} is the building block of many such expressions, and this formula relies on the construction of a matrix expressed in the rapidities. If we wish to obtain a general formalism, it should be possible to obtain the rapidities from the eigenvalue-based variables for any XXZ model. In order to be as general as possible, we propose a method to obtain the rapidities that does not rely on the explicit expression for $X$ and $Z$ except in the very last step. We introduce an auxiliary index $r$, corresponding to an auxiliary single-particle level or an additional uncoupled rapidity $\epsilon_r$, such that the Gaudin algebra can be extended. Following Ortiz et al. \cite{ortiz_exactly-solvable_2005} and Eq. (\ref{ga:Zeq}), each $Z_{i\alpha}$ can be written as \begin{equation}\label{var:parZ} Z_{i\alpha}=\frac{\Gamma+Z_{ri}Z_{r \alpha}}{Z_{ri}-Z_{r\alpha}}. \end{equation} Once an explicit expression for $Z_{ij}$ is known, $Z_{ri}=Z(\epsilon_r,\epsilon_i)$ can easily be calculated after choosing $\epsilon_r$. Instead of solving for the rapidities $\{x_{\alpha}\}$, it is now possible to find a transformation that allows us to determine the set of Gaudin algebra elements $\{Z_{r\alpha}\}$. If these are known, the rapidities can always be found by solving each separate equation $Z_{r\alpha}=Z(\epsilon_r,x_{\alpha})$ for $x_{\alpha}$. This expression can be used to show that \begin{eqnarray}\label{inv:lam} \Lambda_i&=&\sum_{\alpha=1}^NZ_{i\alpha}=\sum_{\alpha=1}^N\frac{\Gamma+Z_{ri}Z_{r \alpha}}{Z_{ri}-Z_{r\alpha}}\nonumber\\ &=&-NZ_{ri}+(\Gamma+Z_{ri}^2)\sum_{\alpha=1}^N\frac{1}{Z_{ri}-Z_{r\alpha}}. \end{eqnarray} We can now define a polynomial with the full set ($\alpha=1 \dots N$) of $Z_{r\alpha}$ as roots \begin{equation} P(z)=\prod_{\alpha=1}^N(z-Z_{r\alpha})=\sum_{m=0}^N P_{N-m}z^m, \end{equation} with $P_0=1$. Once the coefficients $P_{N-m}$ are known, the roots of this polynomial can be determined to find the variables. This method arises naturally for different problems in the theory of integrable systems, such as the Heine-Stieltjes connection \cite{stieltjes_theoreme_1885,sriram_shastry_solution_2001}, the numerical methods by Guan et al. \cite{guan_heine-stieltjes_2012} and Rombouts et al. \cite{rombouts_solving_2004}, and the weak-coupling limit in RG models \cite{ortiz_exactly-solvable_2005}. The definition of $P(z)$ can now be used to consider \begin{equation} \frac{P'(z)}{P(z)}=\frac{\sum_{m=0}^N m P_{N-m}z^{m-1}}{\sum_{m=0}^N P_{N-m}z^m}=\sum_{\alpha=1}^N\frac{1}{z-Z_{r\alpha}}. \end{equation} Evaluating this equality in $z=Z_{ri}$, we find that \begin{equation} \Lambda_i=-N Z_{ri}+(\Gamma+Z_{ri}^2)\frac{P'(Z_{ri})}{P(Z_{ri})} \end{equation} resulting in \begin{equation} \frac{P'(Z_{ri})}{P(Z_{ri})}=\sum_{\alpha=1}^N\frac{1}{Z_{ri}-Z_{r\alpha}}=\frac{\Lambda_i+N Z_{ri}}{\Gamma+Z_{ri}^2}. \end{equation} The coefficients of the polynomial $P(z)$ can then be found by solving a linear problem \begin{eqnarray} \sum_{m=0}^{N-1}P_{N-m}&&\left[m\Gamma Z_{ri}^{m-1}-\Lambda_iZ_{ri}^m+(m-N)Z_{ri}^{m+1}\right]\nonumber\\ &&=\Lambda_i Z_{ri}^N-N\Gamma Z_{ri}^{N-1}. \end{eqnarray} We obtain $n$ equations for $N<n$ variables, but if the set of $\Lambda_i$ variables can be written as Eq. (\ref{inv:lam}), these are linearly dependent and we are free to choose $N$ equations from the full set and solve these. Once these coefficients are known, the variables $Z_{r\alpha}$ can be determined using a root-finding algorithm such as Laguerre's method. Although easy to implement, this method has the disadvantage of being prone to numerical errors. Indeed, it is well-known that the roots of a polynomial are highly sensitive to changes in the coefficients, sometimes even changing the solutions at the qualitative level (a set of complex conjugate roots can be found numerically instead of two separate real roots). These considerations are not pressing for a limited number of excitations, but become more and more important for an increasing number of excitations. This problem was also encountered for the rational model \cite{faribault_gaudin_2011}, after which an improved method was proposed by decomposing $P(z)$ in the basis of Lagrange polynomials \cite{el_araby_bethe_2012}. This led to an increased accuracy and allowed the method to tackle problems with a large number of excitations (up to a few hundreds). It is expected that these results can be generalized towards our problem because of the similarity of all necessary equations. We refer the reader to [\onlinecite{el_araby_bethe_2012}] for a detailed analysis of the method. As an illustration, the eigenvalue-based variables and the related rapidities have been plotted for the different models in Figure \ref{fig:models} as a function of the coupling constant. For each model, the levels $\epsilon_i$ are given by a picket-fence model \cite{hirsch_fully_2002} with equal level spacing: $\epsilon_i=i, \forall i$. For each system singular points occur, where multiple real rapidities combine and continue as a pair of complex conjugate variables. As is clear from Figure \ref{fig:models}, the RG equations (Eq. \ref{ga:RGeq}) become highly singular, whereas the eigenvalue-based variables vary very smoothly. \begin{figure*}[htb!] \begin{center} \includegraphics[width=\textwidth]{models_neg} \caption{Evolution of the variables for the ground state of a 'picket-fence' model with 6 excitations in 12 doubly-degenerate levels $[1.,2.,3., \dots ,12.]$. The rational, the hyperbolic and the trigonometric models are organized in columns, while the rows depict the set of eigenvalue-based variables $g\Lambda_i$ and the real and imaginary part of the rapidities as a function of the coupling constant $g=0 \dots -1$. The used parametrizations are given by Eqs. (\ref{ga:rat}), (\ref{ga:hyp}) and (\ref{ga:trig}) respectively. Note the Moore-Read point at $g=-2/n=-0.167$ for the hyperbolic model \cite{rombouts_quantum_2010}, where all rapidities condense to zero, and the eigenvalue-based variables all become equal to $-1$. Due to the periodicity of the trigonometric model, the real parts of the rapidities lie within the interval $[-\pi/2,+\pi/2]$, as marked in the figure. \label{fig:models}} \end{center} \end{figure*} \subsection{Degenerate models} The discussion has been limited to $d_i=1/2$ models only thus far. This is sufficient for the majority of interacting (quasi-)spin systems, however in situations with higher symmetries arbitrary degeneracies $d_i > 1/2$ may occur. For the rational model, the set of equations for the eigenvalue-based variables were derived starting from a differential equation, where each equation for $\Lambda_i$ is obtained by evaluating this equation at a different level $\epsilon_i$. For degenerate levels, additional equations could be obtained by taking derivatives of the differential equation and evaluating these at the different values $\epsilon_i$ \cite{faribault_gaudin_2011,el_araby_bethe_2012}. This necessitated the introduction of new variables \begin{equation} \Lambda_i^{(n)} \sim \sum_{\alpha=1}^N \frac{1}{(\epsilon_i-x_{\alpha})^n}, \end{equation} which are highly reminiscent of the set of variables introduced by Rombouts et al. \cite{rombouts_solving_2004}. The total number of variables is determined by the number of single-particle levels weighted by their degeneracies. For each single particle level $i$ with degeneracy $2d_i+1$, taking the first $2d_i-1$ derivatives of the differential equation and evaluating each equation at $\epsilon_i$ then results in a closed set of equations for the set of variables $\Lambda_i,\Lambda_{i}^{(2)},\dots,\Lambda_{i}^{(2d_i)}, \forall i$. The outlined procedure is slightly more involved for the class of XXZ RG models, however it is analogous to the rational case, and remains therefore tractable. The main idea is identical: starting from a continuous representation of the equations it is possible to obtain any number of equations for the total set of variables. The continuous representation of the variables is given by \begin{equation} \Lambda(z)=\sum_{\alpha=1}^NZ(z,x_{\alpha}), \end{equation} where $\Lambda(\epsilon_i)\equiv\Lambda_i$. A similar derivation as for the $d_i=1/2$ model in Appendix \ref{appendixA} results in a continuous equation \begin{eqnarray}\label{deg:conteq} [\Lambda(z)]^2&=&\Gamma N\left(1-N+2\sum_{j \neq i}d_j\right)-\frac{2}{g}\Lambda(z)\nonumber\\ &+&2\sum_{j \neq i}d_jZ(z,\epsilon_j)\left(\Lambda(z)-\Lambda(\epsilon_j)\right) \nonumber\\ &+&\sum_{\alpha}Z(z,x_{\alpha})\left[Z(z,x_{\alpha})-2d_i Z(\epsilon_i,x_{\alpha})\right], \end{eqnarray} which holds if $z \neq \epsilon_j$ for all $j\neq i$. The evaluation of these equations at $z=\epsilon_i$ results in the known set of equations for $\Lambda_i$ if $d_i=1/2$, but for larger degeneracies the set of equations also depends on \begin{equation} \Lambda_{i}^{(2)}\equiv\sum_{\alpha=1}^N Z_{i\alpha}^2. \end{equation} By taking the derivative of Eq. (\ref{deg:conteq}), additional equations can be obtained linking these higher-order variables to the original variables. The derivation of this method is given in Appendix \ref{appendixB}. It is remarkable that the derivative of $\Lambda(z)$, a summation over $Z(z,x_{\alpha})$, can still be related to $\Lambda^{(2)}(z)$, a summation over $Z(z,x_{\alpha})^2$. As an example, the evaluation of the first derivative results in \begin{eqnarray}\label{deg:firstder} \Lambda_i\left(N\Gamma+\Lambda_{i}^{(2)}\right)&=&-\frac{1}{g}\left(N\Gamma+\Lambda_{i}^{(2)}\right)\nonumber\\&&+\sum_{j \neq i}d_j(\Gamma+Z_{ij}^2)\left(\Lambda_i-\Lambda_j\right) \nonumber\\ &&+\sum_{j \neq i}d_j Z_{ij}\left(\Lambda_{i}^{(2)}+\Gamma\right)\nonumber\\ && + \Gamma(1-d_i)\Lambda_i +(1-d_i)\Lambda_{i}^{(3)}. \end{eqnarray} For $d_i=1$, we have obtained a closed set of equations in $\{\Lambda_i,\Lambda_i^{(2)}\}$ for each level. For arbitrary $d_i$, the first $2d_i-1$ derivatives results in a closed set of equations. An additional equation for the total number of excitations can easily be determined as \begin{equation} N=-\sum_{i=1}^n g d_i \Lambda_i. \end{equation} As an illustration, the results for a level-independent reduced BCS model describing neutron superfluidity\cite{rombouts_solving_2004} in $^{56}$Fe are given in Figure \ref{fig:Stefan}. The energy levels are determined as the levels of a spherical symmetrical Woods-Saxon potential and the $d_i$, dependent on the angular momentum quantum numbers, vary from $1/2$ to $5/2$. \begin{figure}[tb!] \begin{center} \includegraphics[width=\columnwidth]{figureDegen} \caption{Evolution of the variables for the ground state containing 11 excitations in 10 levels for a varying (negative) coupling constant $g$, describing the pairing of neutrons in $^{56}$Fe by means of a level-independent reduced BCS model. The parameters of the system are taken from [\onlinecite{rombouts_solving_2004}], where the set of $d_i$ are given by $[3/2,1,1/2,2,1,3/2,1/2,1/2,1,5/2]$ with a maximal degeneracy of $2d_i+1=6$. The first figure shows the evolution of the eigenvalue-based variables, while the second and third figure correspond to the real respectively the imaginary part of the rapidities. \label{fig:Stefan}} \end{center} \end{figure} \section{Form factors} \subsection{Overlap with Slater determinants} When performing calculations with Bethe Ansatz states, it is customary to expand them in a basis set of uncorrelated wavefunctions. These expansion coefficients are given by permanents of matrices in the general case, following from the theory of multivariate generating functions \cite{percus_combinatorial_1971}. However, these expressions are not practical for computational purposes since the evaluation of a permanent scales exponentially with the matrix size. Compared to the determinant, which can be evaluated in a polynomial time, permanents require a factorial scaling computational time and therefore severely limit the size of the systems that can be considered. Luckily, multiple formulae exist for the rational model linking the permanents to determinants, allowing numerically efficient expansions of the Bethe Ansatz state. Here, we present several such results for $d_i=1/2$ XXZ systems as a generalization of the results previously obtained for the XXX model \cite{faribault_determinant_2012} and the inhomogeneous Dicke model \cite{tschirhart_algebraic_2014}. Starting from these expressions, determinant expressions for systems with arbitrary $d_i$ can also be obtained as a limiting case where several rapidities coincide, but this will not be discussed here. Multiple determinant formulae exist for the overlap of a Bethe Ansatz state in the rational model with a Slater determinant \begin{equation} \ket{\{i_a\}}=\ket{\{i_1\dots i_N\}}=\prod_{a=1}^N S^{\dagger}_{i_a}\ket{\downarrow \dots \downarrow}, \end{equation} which is an uncorrelated wavefunction and an eigenstate in the uncoupled limit ($g=0$). The Bethe Ansatz state (\ref{rg:bas}) can be expanded in this basis as \begin{eqnarray} \ket{\psi_N}&=&\prod_{\alpha=1}^N\left(\sum_{i=1}^n X_{i\alpha} S^{\dagger}_i\right)\ket{\theta} \nonumber \\ &=&\sum_{\{i_a\}}\ket{\{i_a\}}\braket{\{i_a\}|\psi_N} \\ &=&\sum_{\{i_a\}}\ket{\{i_a\}} \perm \left(X_{i_a \alpha}\right), \end{eqnarray} where $\{i_a\}$ runs over the partitioning of $n$ levels over $N$ different excitations modes and \begin{equation} \braket{\{i_a\}|\psi_N}= \perm \left(\left[X_{i_a \alpha}\right]\right), \end{equation} which is the permanent of an $(N \times N)$ matrix with matrix elements given by $X_{i_a \alpha}$. This is a formula independent of the integrability or the explicit expression for $X$ and is simply a result of the structure of the wavefunction. However, by imposing the Gaudin algebra on the matrix elements, it is always possible to rewrite this permanent as a determinant. \begin{eqnarray} &&\perm \left(\left[X_{i_a \alpha}\right]\right)\nonumber\\ &&\ \ =\frac{\left(\prod_{a>b}X_{i_ai_b}\right)\left(\prod_{\beta>\alpha}X_{\alpha\beta}\right)}{\prod_{a,\alpha}X_{i_a\alpha}}\det\left[X_{i_a\alpha}*X_{i_a\alpha}\right] \end{eqnarray} where the Hadamard product of two matrices is introduced, defined as $(A*B)_{ij}=A_{ij}B_{ij}$. This is a generalization of the Borchardt or Izergin determinant representation for the rational model \cite{borchardt_bestimmung_1857,izergin_determinant_1992,singer_bijective_2004} and has a similar structure to some results presented previously for the trigonometric model\cite{hao_determinant_2012}. In ref. [\onlinecite{faribault_determinant_2012}], a case was made for a RG theory that would require only the eigenvalue-based variables instead of the rapidities. This would be desirable from a numerical point of view as the eigenvalue-based variables are free of singularities, opposed to the singularity-prone rapidities. In addition, this would enable us to skip the inversion step, which is the main bottleneck of this method. For the rational model, it was shown by Faribault and Schuricht \cite{faribault_determinant_2012} that \begin{equation} \perm\left(\left[\frac{1}{\epsilon_{i_a}-x_{\alpha}}\right]\right)=\det J \end{equation} where the left-hand size is the permanent of an $N \times N$ Cauchy matrix and the right hand side is the determinant of a matrix $J$ defined as \begin{equation}\label{rat:overlap} J_{ab}=\begin{cases}\sum_{\alpha=1}^N\frac{1}{\epsilon_{i_a}-x_{\alpha}}-\sum_{c \neq a}^N\frac{1}{\epsilon_{i_a}-\epsilon_{i_c}} &\text{if } a=b\\ \frac{1}{\epsilon_{i_a}-\epsilon_{i_b}} &\text{if } a \neq b \end{cases} \end{equation} where the only dependence on the rapidities is through the diagonal elements, which can be expressed in terms of the eigenvalue-based variables. The ingenious proof of this expression is based on a recursion relation which was found to hold for both sides of Eq. (\ref{rat:overlap}). We will show that these results can be easily generalized to XXZ models. The Gaudin equations can be used to rewrite the permanent in the XXZ model in a structure similar to a Cauchy matrix. We once again introduce an auxiliary level $r$ and use the Gaudin equations to write \begin{equation} X_{i\alpha}=\frac{X_{ri}X_{r\alpha}}{Z_{ri}-Z_{r\alpha}}, \qquad Z_{i\alpha}=\frac{\Gamma+Z_{ri}Z_{r \alpha}}{Z_{ri}-Z_{r\alpha}}. \end{equation} Rewriting all matrix elements in the permanent $\phi$ allows us to take \begin{eqnarray} &&\perm\left(\left[X_{i_a \alpha}\right]\right)=\perm\left(\left[\frac{X_{ri_a}X_{r\alpha}}{Z_{ri_a}-Z_{r\alpha}}\right]\right)\nonumber\\ &&\ =\left(\prod_a X_{ri_a}\right)\left( \prod_{\alpha} X_{r\alpha}\right) \perm\left(\left[\frac{1}{Z_{ri_a}-Z_{r\alpha}}\right]\right), \end{eqnarray} resulting in a Cauchy matrix, which we can rewrite as a determinant \begin{equation} \perm\left(\left[X_{i_a \alpha}\right]\right)= \left(\prod_a X_{ri_a}\right)\left( \prod_{\alpha} X_{r\alpha}\right) \det J \end{equation} with $J$ redefined as \begin{equation} J_{ab}= \begin{cases} \sum_{\alpha=1}^N\frac{1}{Z_{ri_a}-Z_{r\alpha}}-\sum_{c \neq a}^N\frac{1}{Z_{ri_a}-Z_{ri_c}} &\text{if } a=b\\ \frac{1}{Z_{ri_a}-Z_{ri_b}} &\text{if } a \neq b\\ \end{cases} \end{equation} By multiplying each row and column $c$ with $X_{ri_c}$ and compensating for these factors in the prefactor, this can be written as \begin{equation} \perm\left(\left[X_{i_a \alpha}\right]\right)=\frac{\prod_{\alpha}X_{r\alpha}}{\prod_{a}X_{ri_a}} \det J \end{equation} with \begin{equation} J_{ab}=\begin{cases} \sum_{\alpha=1}^N\frac{X_{ri_a}^2}{Z_{ri_a}-Z_{r\alpha}}-\sum_{c \neq a}^N\frac{X_{ri_a}^2}{Z_{ri_a}-Z_{ri_c}} &\text{if } a=b\\ \frac{X_{ri_a}X_{ri_b}}{Z_{ri_a}-Z_{ri_b}} &\text{if } a \neq b, \end{cases} \end{equation} where we recognize $X_{i_ai_b}$ in the off-diagonal elements and the diagonal elements can be rewritten by using the Gaudin algebra until \begin{equation} J_{ab}=\begin{cases}\sum_{\alpha=1}^NZ_{i_a\alpha}-\sum_{c \neq a}^NZ_{i_ai_c}+Z_{ri_a} &\text{if } a=b\\ X_{i_ai_b} &\text{if } a \neq b \end{cases}, \end{equation} The only explicit dependency on the rapidities is now found in the prefactor and in the diagonal elements, which only depend on the eigenvalue-based variables. Fortunately, the prefactor can be absorbed in the definition of the state without loss of generality, which will still allow the calculation of all necessary normalizations and prefactors. Note the appearance of $Z_{ri}$ in the diagonal elements, which can be easily evaluated and links the matrix to the prefactor for the Bethe Ansatz state through our choice of the additional level $r$. To recapitulate, the overlap of a state \begin{equation} \ket{\{x_{\alpha}\}}=\prod_{\alpha=1}^N\left(\sum_{i=1}^n \frac{X_{i\alpha}}{X_{r\alpha}} S^{\dagger}_i\right)\ket{\downarrow \dots \downarrow} \end{equation} with a Slater determinant \begin{equation} \ket{\{i_a\}}=\ket{\{i_1\dots i_N\}}=\prod_{j=1}^N S^{\dagger}_{i_j}\ket{\downarrow \dots \downarrow} \end{equation} is given by \begin{equation}\label{ov:XXZ} \braket{\{i_1\dots i_N\}|\{x_{\alpha}\}}=\frac{\det J}{\prod_{a}X_{r i_a}} \end{equation} with \begin{equation} J_{ab}=\begin{cases} \Lambda_{i_a}-\sum_{c \neq a}^NZ_{i_ai_c}+Z_{ri_a} &\text{if } a=b\\ X_{i_ai_b} &\text{if } a \neq b \end{cases}. \end{equation} \subsection{Dual representations} So far, all Bethe Ansatz states have been created by acting with creation operators on the vacuum state, destroyed by all annihilation operators. Due to the particle-hole symmetry of the RG models, every Bethe Ansatz state can also be constructed by acting on the fully-filled state, annihilated by all creation operators, with a set of generalized annihilation operators, the so-called dual representation. A renormalized Bethe Ansatz state is given in its normal and dual representation as \begin{eqnarray} \ket{\{x_{\alpha}\}}&=&\prod_{\alpha=1}^N\left(\sum_{i=1}^n \frac{X_{i\alpha}}{X_{r\alpha}} S^{\dagger}_i\right)\ket{\downarrow \dots \downarrow}, \\ \ket{\{x_{\alpha'}\}}&=&\prod_{\alpha'=1}^{n-N}\left(\sum_{i=1}^n \frac{X_{i\alpha'}}{X_{r\alpha'}} S_i\right)\ket{\uparrow \dots \uparrow}. \end{eqnarray} The RG equations for the dual state are given by \begin{equation} -1+\frac{g}{2}\sum_{i}Z_{j\alpha'} = g \sum_{\beta' \neq \alpha'}^{n-N}Z_{\beta'\alpha'}, \qquad \forall \alpha'=1 \dots n-N \end{equation} or, written in the eigenvalue-based variables, \begin{equation} \left[\Lambda_i'\right]^2=N(n-N)\Gamma+\frac{2}{g}\Lambda_i'+\sum_{j \neq i} Z_{ji}(\Lambda_j'-\Lambda_i'), \ \ \forall i=1\dots n. \end{equation} Both representations describe, up to the normalization, the same state. So by comparing the eigenvalues of the constants of motion \begin{eqnarray} r_i&=&\frac{1}{2}\left(-1-g\Lambda_i +\frac{g}{2}\sum_{k\neq i}^nZ_{ik}\right) \nonumber\\ &=&\frac{1}{2}\left(1-g\Lambda_i'+\frac{g}{2}\sum_{k\neq i}^n Z_{ik}\right), \end{eqnarray} it can be seen that the dual representation of an eigenstate is determined by a set of dual eigenvalue-based variables given by \begin{equation} \Lambda_i'=\Lambda_i+\frac{2}{g}, \qquad \forall i=1 \dots n, \end{equation} which can be verified by simply substituting these variables in the dual eigenvalue-based equations. The existence of the dual state allows one to write the overlap of a normal state and a dual state as the overlap of a Bethe Ansatz state with the fully-filled state. This will be exploited in the following section to obtain the normalization of the Bethe Ansatz states and several form factors. \subsection{Normalization and form factors} Due to the dual representation, we have two different representations of each Bethe Ansatz state. These two states have different normalizations and can be written as \begin{eqnarray} \ket{\{x_\alpha\}}&=&\prod_{\alpha=1}^N \left(\sum_{i=1}^n \frac{X_{i\alpha}}{X_{r\alpha}}S^{\dagger}_i\right)\ket{\downarrow \dots \downarrow} = N_{\alpha} \ket{\{x_\alpha\}}_n \nonumber \\ \ket{\{x_\alpha'\}}&=&\prod_{\alpha'=1}^{n-N} \left(\sum_{i=1}^n \frac{X_{i\alpha'}}{X_{r\alpha'}}S_i\right)\ket{\uparrow \dots \uparrow} = N_{\alpha'} \ket{\{x_\alpha\}}_n \nonumber \end{eqnarray} where $\ket{\{x_{\alpha}\}}_n$ is the normalized eigenstate. This can be used to calculate \begin{eqnarray} &&\braket{\{x_{\alpha'}\}|\{x_{\alpha}\}}=N_{\alpha}N_{\alpha'}\nonumber\\ &&\ \ =\braket{\uparrow \dots \uparrow |\prod_{\alpha'=1}^{n-N} \left(\sum_{i=1}^n \frac{X_{i\alpha'}}{X_{r\alpha'}}S^{\dagger}_i\right)\prod_{\alpha=1}^N \left(\sum_{i=1}^n \frac{X_{i\alpha}}{X_{r\alpha}}S^{\dagger}_i\right) |\downarrow \dots \downarrow} \nonumber\\ &&\ \ =\frac{\det J}{\prod_{i=1}^n X_{ri}} \end{eqnarray} with \begin{equation} J_{ij}=\begin{cases} 2\Lambda_{i}+\frac{2}{g}-\sum_{k \neq i}^nZ_{ik}+Z_{ri} &\text{if } i=j\\ X_{ij} &\text{if } i \neq j \end{cases}, \end{equation} since the overlap between the two states is just the overlap of a Bethe Ansatz state determined by $\{x\}=\{x_{\alpha}\}\cup\{x_{\alpha'}\}$ or by $\Lambda_i=\Lambda_i+\Lambda_i'=2\Lambda_i+2/g$ with $\ket{\uparrow \dots \uparrow}$. The ratio $N_{\alpha}/N_{\alpha'}$ can easily be found by taking the overlap of both representations with a single reference state and taking this ratio. Knowing both the ratio and the product of the normalizations, the normalization of both representations is uniquely defined In a similar manner as for the XXX model \cite{faribault_determinant_2012,tschirhart_algebraic_2014}, these results can be used to calculate form factors. Since this is a generalization of the results presented in these papers, we will only summarize the results for the XXZ model here. The form factor for the raising operator between a state $\{x_\alpha\}$ with $N$ excitations and a dual state $\{x_\mu'\}$ with $N+1$ excitations is given by \begin{equation} \braket{\{x_\mu'\}|S^{\dagger}_k|\{x_\alpha\}}=\frac{\det J^k}{\prod_{i \neq k} X_{ri}} \end{equation} with \begin{equation} J^k_{ab}=\begin{cases} \Lambda_{i}^{\alpha}+\Lambda_{i}^{\mu}+\frac{2}{g}-\sum_{l \neq i,l \neq k}^nZ_{il}+Z_{ri} &\text{if } a=b\\ X_{ab} &\text{if } a \neq b \end{cases} \end{equation} and $a,b \neq k$. The form factor for a lowering operator is then given by the Hermitian conjugate of this expression. The form factor for local operators $S_k^0$ can similarly be found, both for diagonal and off-diagonal expectation values. Here we will make the dependency on the coupling constant explicit and define $\{x_\alpha(g)\}$ as an eigenstate at coupling constant $g$. From the Hellmann-Feynman theorem we obtain for the diagonal expectation values \begin{eqnarray} &&\braket{\{x_\alpha'(g)\}|S^0_k|\{x_\alpha(g)\}}=\nonumber\\ &&\qquad\frac{1}{2}\left(-1+g^2\frac{\partial \Lambda^{\alpha}_k}{\partial g}\right)\braket{\{x_\alpha'(g)\}|\{x_\alpha(g)\}} \end{eqnarray} while the off-diagonal expectation values can be found as \begin{eqnarray} &&\braket{\{x_\mu'(g)\}|S^0_k|\{x_\alpha(g)\}} \nonumber\\ &&\qquad=\frac{1}{2}\left(g\Lambda^{\mu}_k-g\Lambda^{\alpha}_k+2\right)\frac{\braket{\{x_\mu'(g)\}|\{x_\alpha(g+dg)\}}}{dg} \nonumber\\ &&\qquad=\frac{1}{2}\left(g\Lambda^{\mu}_k-g\Lambda^{\alpha}_k+2\right)\sum_{k=1}^n\frac{\partial \Lambda^{\alpha}_k}{\partial g} \det \tilde{J}^k \end{eqnarray} with \begin{equation} \tilde{J}^k_{ab}= \begin{cases} \Lambda_{i}^{\alpha}+\Lambda_{i}^{\mu}+\frac{2}{g}-\sum_{l \neq i}^nZ_{il}+Z_{ri} &\text{if } a=b\\ X_{ab} &\text{if } a \neq b \end{cases} \end{equation} and $a,b \neq k$. These form factors can be written in an expression only dependent on the eigenvalue-based variables, so the rapidities do not need to be determined explicitly. However, if these have been determined, other form factors can also be calculated by commuting the operator through the creation operators in the Bethe Ansatz state, as has been shown in [\onlinecite{zhou_superconducting_2002}]. This allows the computation of form factors such as e.g. $S^{\dagger}_k S_l$ as a sum over determinants. It is interesting to note that the expectation value for $S_i^0$ offers a physical interpretation for $\Lambda_i$, since the derivative of $\Lambda_i$ to $g$ fully determines the filling of the level $i$ as described by $\braket{S_i^0}$. Knowledge of the evolution of $\Lambda_i$ with a changing coupling constant is then equivalent to knowing how the $N$ excitations are distributed over the $n$ levels. \section{Conclusions and outlook} In the present paper, we presented a numerical solution method for the RG equations for XXZ integrable models, while also proposing determinant expressions for the normalization and form factors of the Bethe Ansatz states that are only dependent on the new set of eigenvalue-based variables. The availability of an efficient solution method and explicit expressions for the overlaps opens up possibilities for the investigation of integrable and non-integrable quantum systems. An efficient numerical expression for overlaps between two Bethe Ansatz states allows for a numerically nearly exact investigation of the dynamics resulting from a quantum quench in the reduced BCS pairing model \cite{faribault_quantum_2009}, while for models close to integrability the eigenstates of the integrable model are being used to study decoherence \cite{van_den_berg_competing_2014}. These results were all obtained for the rational XXX model, and it would be interesting to investigate similar dynamics for XXZ models, where a rich phase diagram is known \cite{rombouts_quantum_2010}. Due to the favourable computational scaling of this method it is also possible to numerically approach the thermodynamic limit and investigate the limiting behaviour of the system. Such an investigation has been carried out for the reduced BCS Hamiltonian by El Araby and Baeriswyl \cite{el_araby_order_2014}, where exact results for large system sizes were compared with the BCS Ansatz. Again, it would be worthwhile to compare the results rigourously with the results from BCS mean-field theory \cite{ibanez_exactly_2009}. The overlaps between Bethe Ansatz states and Slater determinants have also been used in quantum chemistry for strongly correlated systems. These can not be accurately described by means of a single Slater determinant, and several wavefunctions have been proposed to capture the correlation present in the system. One example is the class of geminal-based wavefunctions, which have a structure highly reminiscent of the Bethe Ansatz states found in the Gaudin models. Unfortunately, these wavefunctions are only computationally feasible if a numerical efficient expression is known for the overlaps with Slater determinants. One such class (APr2G) is based on the rational XXX model \cite{johnson_size-consistent_2013,limacher_new_2013,tecmer_assessing_2014}, and the presented results should allow for an extension of these geminal-based wavefunctions with a class based on the XXZ models. The results presented here should also allow for a variational approach based on the Bethe Ansatz states in the XXZ model. By starting from Richardson's general solution for the Gaudin algebra (Eq. \ref{ga:richalg}) and varying over the free parameters in the wavefunction, it is possible to scan over all possible Gaudin models (XXX and XXZ), allowing more freedom and as such a better approximation to the energy compared to a variational approach based solely on the XXX model. It was already noted by Babelon and Talalaev that similar equations could be found for the XXX Heisenberg spin chain \cite{babelon_bethe_2007}, so the question naturally arises if this method can be generalized to other Bethe Ansatz equations obtained from the quantum inverse scattering method \cite{korepin_quantum_1993}. \section*{Acknowledgements} PWC and SDB acknowledge financial support from FWO-Vlaanderen as pre-doctoral and post-doctoral fellows respectively.

Introduction {#s0005} ============ Construction delay means a time overrun either beyond the contract date or beyond the date that the parties have agreed upon for the delivery of the project. In both cases, a delay is usually a costly situation [@b0005]. Delay was also defined as an act or event which extends required time to perform or complete work of the contract manifests itself as additional days of work [@b0010]. Poor site management can cause project delay and affect productivity [@b0015]. A lot of research efforts have been made to study delay causes in different countries [@b0020; @b0025; @b0030; @b0035; @b0040; @b0045; @b0050; @b0055; @b0060; @b0065; @b0070; @b0075]. For example, material-related delay is the main cause of project delays in Saudi Arabia [@b0080]. Bordoli and Baldwin [@b0085] examined the causes of delays in building projects in the United States. Weather, labor supply, and sub-contractors were found to be the major causes of delays. Poor risk management, poor supervision, unforeseen site conditions, slow decision making involving variation, and necessary variation works are the principle delay factors in Hong Kong [@b0090]. Unforeseen soil condition, poor site supervision, low speed of decision making involving all project teams, client initiated variations, necessary variations of work, and inadequate contractor experience are the six significant factors contributing to delays in building and civil engineering works [@b0095]. Materials-, equipment-, and labor-related delays were identified as major causes of contractors' performance delays [@b0100]. Design changes, poor labor productivity, and inadequate planning and resources were found to be responsible for delays in Indonesia [@b0105]. In Saudi Arabia, contractor performance, owner's administration, early planning and design, government regulation, site and environment conditions, and site supervision were found to be the important causes of delay [@b0110]. Whereas, the financing and payment for completed works, poor contract management, change in site conditions, and shortages of materials were found the most important items of delay causes in Nigeria [@b0115]. Odeyinka and Yusif [@b0120] studied client-, contractor-, and consultant-caused delays in housing projects in Nigeria. Variation orders, slow decision making, and cash flow problems were found as client-caused delays. Financial difficulties, material management problems, planning and scheduling problems, inadequate site inspection, equipment management problems and shortage of manpower were found as contractor-caused delays. Incomplete drawing, slow response by consultant, variation orders, late issuance of instruction, and poor communications were classified as consultant-caused delays. Inclement weather, act of God, labor dispute, and strikes were found to be extraneous factors responsible for delays. Bramble and Callahan [@b0125] studied owner-, designed-, contractor-, and others-related delays in U.S.A. Late release of site to the contractor, late approval, financial difficulties, contract administration responsibilities, change orders, and interference were found to be owner-caused delays. Design defects, slow correction of design errors, tardy shop drawings review, and delays due to test and inspection were considered to be designed-caused delay. Failure to evaluate the site and design, construction defects, contractor management problems, and inadequate resources were found to be contractor-related delays. Weather, act of God, strikes, and labor disputes were found to be delays not caused by the design and construction parties. In Egypt, Amer [@b0130] studied the major delay causes for construction projects which they are: poor contract management, unrealistic scheduling, lack of owner's financing/payment for completed work, design modifications during construction, and shortages in materials such as cement and steel. Abd El-Razek et al. [@b0135] considered several delay causes in construction projects in Egypt which are: financing by contractor during construction, delays in contractor's payment by owner, design changes by owner or his agent during construction, partial payments during construction, and non-utilization of professional construction/contractual management. Marzouk et. al. [@b0140] studied delays that are related to engineering factors which arise due to design development, workshop drawings, and change then he developed a knowledge based expert system for assessing the engineering related delay claims. Kazaz et al. [@b0145] conducted a study on the causes oof time extensions in the Turkish construction industry and levels of their importance, considering 34 factors. A questionnaire survey was conducted with 71 construction companies in Turkey, and the outcomes were evaluated by means of statistical analyses. Methodology {#s0010} =========== Delay causes survey {#s0015} ------------------- The questionnaire designed for use in the survey comprised demographic information about respondents and 43 delay causes which were grouped to seven categories: owner related, consultant related, contractor related, material related, labor and equipment related, project related, and external related (see [Table 1](#t0005){ref-type="table"}). The respondents were requested to choose one degree of frequency for each delay cause which is rarely, sometimes, often, or always. Also they were requested to choose one degree of severity which is low, moderate, high, or extreme. The questionnaire was distributed only to owners who are representatives of large investment projects and affiliated to first class of consultants who are classified as house of expertise in the Egyptian Syndicate of Engineers as well as to contractors who are representatives to high class companies registered in the Egyptian Federation for Construction and Building Contractors (EFCBC). Total of 33 experts responded. The experts were divided into three groups each group consists of eleven experts the first group represents owners and the second represents consultants while the third represents contractors. All respondents hold senior positions with related working experience and the majority of them had practiced in the field for 20--30 years. The size of the sample required from the targeted population, i.e. 33 respondents was determined statistically [@b0150]. The results suggested that the minimum sample size required was:$$n_{0} = \frac{p \ast q}{v^{2}}$$$$n = \frac{n_{0}}{1 + \frac{n_{0}}{N}}$$where *n*~0~ is the first estimate of sample size. *p* the proportion of the characteristic being measured in the target population. *q* the complement of *p* or 1-*p*. *v* the maximum standard error allowed. *N* the population size. *n* is the sample size. The total number of contractors registered the Egyptian Federation for Construction and Building Contractor (EFCBC) in 2011 are 58991 and the first class contractors are 1716. Then, *N* is 58991 and *p* is 1716/58991 = 0.0291. To account for the possible error in the qualitative answers from the questionnaire, the maximum standard error *V* was set as 10%. Substituting in Eqs. [(1) and (2)](#e0005 e0010){ref-type="disp-formula"}, the minimum sample required is 2.8 ≈ 3. It is obvious that this number of required sample is less than the number of respondents who provided their feedback (i.e., 33 respondents). Since the number of contractor companies in Egypt is more than the number of consultant companies and owner representatives, therefore, it is sufficient to utilize the same sample size for owner and consultant representatives as for contractor. A detailed questionnaire comprises 43 delay causes were prepared and presented to construction experts. The respondents were asked to determine the frequency of occurrence of each cause as follows: *Rarely* (*R*) = 1, *Sometimes* (*S*) = 2, *Often* (*O*) = 3, *Always* (*A*) = 4. To determine the degree of severity of that cause, the following levels are considered: *Low* (*L*) = 1, *Moderate* (*M*) = 2, *High* (*H*) = 3, *Extreme* (*E*) = 4. The Frequency Index (F.I), Severity Index (S.I), and Importance Index (IMP.I) are calculated using Eqs. [(3)--(5)](#e0015 e0020 e0025){ref-type="disp-formula"} respectively as stated in Assaf and Al-Hejji [@b0155].$$\text{Frequency\ Index}\ (\text{F}.\text{I})(\%) = \sum\limits_{i = 1}^{4}\frac{a_{\mathit{if}} \ast n_{\mathit{if}}}{4 \ast N} \times 100$$$$\text{Severity\ Index}\ (\text{S}.\text{I})(\%) = \sum\limits_{i = 1}^{4}\frac{a_{\mathit{is}} \ast n_{\mathit{is}}}{4 \ast N} \times 100$$$$\text{Importance\ Index}\ (\text{IMP}.\text{I})(\%) = \frac{\text{F}.\text{I} \ast \text{S}.\text{I}}{100}$$ where *a~if~* and *a~is~* are numbers of respondents who choose certain frequency and severity degree respectively, *n~if~* and *n~is~* are degrees of frequency and severity respectively (1 or 2 or 3 or 4), *N* is total number of respondents. Average values for Frequency and Importance Indexes for all respondents are shown in [Fig. 1](#f0005){ref-type="fig"}. The top ten delay causes that are ranked according to Frequency Index (F.I), Severity Index (S.I), and Importance Index (IMP.I) are listed in [Tables 2--4](#t0010 t0015 t0020){ref-type="table"}, respectively. Case study {#s0020} ========== A contract was signed between the claimant (contractor) and the defendant (owner) to construct 16 residential buildings including utilities and landscape in 6 October city -- Egypt. The duration of the contract was 24 months and the contract amount was 62.25 million pounds. The project was delayed for the following reasons:•Strike of drivers and tractors sub-contractors (Reason A) which delays reaching required material to site. The contractor notified the owner and the consultant that this event delayed the project one month.•Revolution of 25 January and its effects, security chaos, curfew, closure of banks and material factories, departure of labor to their towns to secure their families (Reason B). The contractor notified the owner and the consultant that this event delayed the project from 25 January 2011 until that event was finished.•Execution of upper semells due to increase height of the ground floor (Reason C). The contractor notified the owner and the consultant that this event delayed the project for four months.•Official holidays (Reason D). The contractor notified the owner and the consultant that this event delayed the project for 180 days. After negotiation, the consultant recommended time extension of 132 days as follows: 21 days for Reason A, 75 days for Reason B, 36 days for Reason C, and no time extension for Reason D. By comparing the above delay causes with delay causes listed in questionnaire, these causes are equivalent respectively to the following causes:•Force Majeure as war, revolution, riot, strike, and earthquake, etc. (Cause 7.8).•Shortage of construction materials in market (Cause 4.1).•Shortage of labors (Cause 5.1).•Variation orders/changes of scope by owner during construction (Cause 6.1). Statistical analysis {#s0025} -------------------- Data are analyzed using Statistical Package for Social Sciences (SPSSs). Numerical data were expressed as median and range to compare between the three groups of respondents (owners, consultants, and contractors) using Kruskal--Wallis test (non-parametric ANOVA), and post-Hoc "Schefe test" on ranks by performing pair-wise comparison. A *P*-value \< 0.05 was considered significant. According to subjective meaning of gathered data from questionnaire which is ordinal data and express of experts' opinions, therefore, median values were utilized to indicate the frequency of delay causes and the major difference between respondents' opinions for values of *P* \< 0.05. [Table 5](#t0025){ref-type="table"} illustrates median and range for delay causes as well as *P* value (probability of error) is presented in the table for each cause which indicate to major difference between respondents opinions for values of *P* \< 0.05. It is clear from the table that there are no differences in opinions among opinions of experts for the three groups of owners, consultants, and contractors in total and for each delay causes group. This means there is good correlation between respondents with regard to delay causes groups in general. [Table 6](#t0030){ref-type="table"} lists the median and range for each delay cause as well as *P* value which indicate to major difference between respondents' opinions considering *P* \< 0.05. The results reveal that there are no differences in opinions among opinions of experts for the three groups of owners, consultants, and contractors in total and for each delay causes group which indicates that there is good correlation between respondents with regard to delay causes groups in general. [Table 7](#t0035){ref-type="table"} lists the median values of Frequency Index and Severity Index for each delay causes group. They are classified according to its percentage as *Very Low* (*V.L*), *Low* (*L*), *Medium* (*M*), *High* (*H*), and *Very High* (*V.H*). The table indicates that the frequency scores are in compliance with severity where owner related delay causes has degree *Very High* and contractor related delay causes has degree *High* while external related delay causes has degree *Medium* but delay causes related to consultant, labor and equipment, and project have degree *Low* while material related delay causes has degree *Very Low*. [Fig. 2](#f0010){ref-type="fig"} depicts the findings regarding the most important delay causes groups of construction projects in Egypt. These causes are prioritized regarding significance to five levels *Very High*, *High*, *Medium*, *Low*, and *Very Low*. Results and discussion {#s0030} ====================== The results of questionnaire are analyzed to obtain the important causes of delays for owners, contractors, and consultants respondents. These causes of delays are listed in below sub-sections. Owners respondents {#s0035} ------------------ The important causes are; ineffective planning and scheduling of project, difficulties in financing project by contractor, variation orders/changes of scope by owner during construction, poor site management and supervision, type of project bidding and award (negotiation, lowest bidder), low productivity level of labors, effects of subsurface conditions (e.g., soil, high water table, etc.), unqualified workforce, shortage of construction materials in market, and delays in sub-contractors work. Contractors respondents {#s0040} ----------------------- The important causes are; finance and payments of completed work by owner, late in revising and approving design documents by owner, owner interference, suspension of work, mistakes and discrepancies in design documents, shortage of construction materials in market, variation orders/changes of scope by owner during construction, slow decision making, effects of subsurface conditions (e.g., soil, high water table, etc.), and type of project bidding and award (negotiation, lowest bidder). Consultants respondents {#s0045} ----------------------- The important causes are; effects of subsurface conditions (e.g., soil, high water table, etc.), low productivity level of labors, difficulties in financing project by contractor, ineffective planning and scheduling of project, poor site management and supervision, finance and payments of completed work by owner, type of project bidding and award (negotiation, lowest bidder), unqualified workforce, late in revising and approving design documents by owner, and variation orders/changes of scope by owner during construction. According to the considered case study, "variation orders/changes of scope by owner during construction" is the most frequent delay cause which is one of the top ten delay causes considered in the questionnaire. According to the findings of the questionnaire and case studies, it is recommended for all project parties to avoid occurrence of delay causes which lead to disputes. The following recommendations might help accomplishing delay reduction in construction projects: ### For owners {#s0050} -Specification of a realistic duration in the contract for the contractor to execute the project.-Having sufficient time to prepare feasibility study for the project, as well as the preparation of a comprehensive financial plan and cash flow.-Obtaining the required approvals for the project from the relevant authorities and ensure the availability of the necessary funding.-Choosing a consultant to the project with sufficient experience in the field of work and has a good reputation.-Making sure tender documents are complete, clear and free of errors and/or contradiction.-Payment of the dues to the contractor for the work being carried as well as the payments of finished items according to terms of the contract.-Hiring an experienced contractor in the field of work who has a good reputation. ### For consultants {#s0055} -Avoiding delaying the response to contractor's queries as well as the approval the submitted submittals and shop drawings.-Establishment of a control system to handle, control, and evaluate variation orders, initiated by the owner. ### For contractors {#s0060} -Development of a comprehensive financial plan and cash flow.-Development of a monitoring and periodical reporting of critical and long lead items and periodically providing a narrative explanation of causes of any experienced delay.-Choosing experienced subcontractors with good reputation.-Development of a good system for site management and supervision also develops effective planning and scheduling for the project. ### For all parties {#s0065} -Project parties should preview the site Due-Diligence Reviews and execution of necessary borings during the tender stage to make sure that the need for adjustments in design or make amendments if necessary before the issuance of notice to proceed.-Formal relationships among project parties should be clearly identified, as well as roles and responsibilities. Conclusion {#s0070} ========== This paper analyzed causes of construction delays in Egypt. The feedback of construction experts was obtained through interviews and questionnaire surveys. Frequency Index, Severity Index, and Importance Index are calculated and according to the highest values of them the top ten delay causes of construction projects in Egypt are determined. Statistical analysis is carried out using analysis of variance ANOVA method to test delay causes, obtained from the survey. The most important delay causes groups of construction projects in Egypt are prioritized according to their significance to five levels *Very High*, *High*, *Medium*, *Low*, and *Very Low*. The survey results were discussed. Finally, recommendations have been made to construction projects' parties to accomplish delay reduction in construction projects. Conflict of interest {#s0075} ==================== The authors have declared no conflict of interest. Peer review under responsibility of Cairo University. {#f0005} {#f0010} ###### Delay causes of construction projects. Delay group Causes ------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------- 1\. Owner related 1.1 Slow decision making 1.2 Suspension of work 1.3 Late in revising and approving design documents by owner 1.4 Delay to furnish and deliver the site to the contractor 1.5 Delay in finance and payments of completed work by owner 1.6 Variation orders/changes of scope by owner during construction 1.7 Type of project bidding and award (negotiation, lowest bidder) 1.8 Unrealistic contract duration 1.9 Ineffective delay penalties 1.10 Owner interference 

 2\. Consultant related 2.1 Inadequate experience of consultant 2.2 Delay in approving shop drawings and sample materials 2.3 Mistakes and discrepancies in design documents 2.4 Unclear and inadequate details in drawings 2.5 Quality assurance/control 

 3\. Contractor related 3.1 Difficulties in financing project by contractor 3.2 Poor site management and supervision 3.3 Ineffective planning and scheduling of project 3.4 Rework due to errors during construction 3.5 Delays in sub-contractors work 3.6 Inadequate contractor experience 3.7 Delay in site mobilization 3.8 Delay in preparation of shop drawings and material samples 

 4\. Material related 4.1 Shortage of construction materials in market 4.2 Delay in material delivery 4.3 Changes in material types and specifications during construction 

 5\. Labor & equipment related 5.1 Shortage of labors 5.2 Unqualified workforce 5.3 Low productivity level of labors 5.4 Equipment availability and failure 

 6\. Project related 6.1 Effects of subsurface conditions (e.g., soil, high water table, etc.) 6.2 Traffic control and restriction at job site 6.3 Unavailability of utilities in site or Delay in providing services from utilities such as (water, etc.) 6.4 Accident during construction 6.5 Problem with neighbors 

 7\. External related 7.1 Weather effect (hot, rain, etc.) 7.2 Environmental restrictions 7.3 Changes in government regulations and laws 7.4 Slow permit by government/municipality 7.5 Delay in performing final inspection and certification by a third party 7.6 Lack of communication between the parties 7.7 Fluctuations in cost/ currency 7.8 Force Majeure as war, revolution, riot, strike, and earthquake, etc. ###### Top ten delay causes ranked according to Frequency Index. Delay group Cause F.I -------------------- --------------------------------------------------------------------------- ------- Owner related 1.7 Type of project bidding and award (negotiation, lowest bidder) 68.94 Contractor related 3.3 Ineffective planning and scheduling of project 68.94 Owner related 1.6 Variation orders/changes of scope by owner during construction 66.67 Owner related 1.3 Late in revising and approving design documents by owner 65.15 Owner related 1.5 Finance and payments of completed work by owner 64.39 Contractor related 3.2 Poor site management and supervision 64.39 Labor & equipment 5.3 Low productivity level of labors 64.39 Contractor related 3.1 Difficulties in financing project by contractor 63.64 Project related 6.1 Effects of subsurface conditions (e.g., soil, high water table, etc.) 62.12 Owner related 1.1 Slow decision making 57.58 ###### Top ten delay causes ranked according to Severity Index. Delay group Cause S.I -------------------- --------------------------------------------------------------------------- ------ Material related 4.1 Shortage of construction materials in market 93.9 External related 7.7 Fluctuations in cost/currency 91.7 Owner related 1.5 Finance and payments of completed work by owner. 90.9 Project related 6.1 Effects of subsurface conditions (e.g., soil, high water table, etc.) 90.9 Labor & equipment 5.1 Shortage of labors 88.6 Consultant related 2.1 Inadequate experience of consultant 86.4 Contractor related 3.1 Difficulties in financing project by contractor 86.4 Labor & equipment 5.3 Low productivity level of labors 86.4 Labor & equipment 5.2 Unqualified workforce 85.6 Owner related 1.6 Variation orders/changes of scope by owner during construction 84.8 ###### Top ten delay causes ranked according to Importance Index. Delay group Cause Imp.I -------------------- --------------------------------------------------------------------------- ------- Owner related 1.5 Finance and payments of completed work by owner 58.54 Owner related 1.6 Variation orders/changes of scope by owner during construction 56.57 Project related 6.1 Effects of subsurface conditions (e.g., soil, high water table, etc.) 56.47 Labor & equipment 5.3 Low productivity level of labors 55.61 Contractor related 3.3 Ineffective planning and scheduling of project 55.36 Contractor related 3.1 Difficulties in financing project by contractor 54.96 Owner related 1.7 Type of project bidding and award (negotiation, lowest bidder) 53.27 Material related 4.1 Shortage of construction materials in market 51.24 Owner related 1.3 Late in revising and approving design documents by owner 50.84 Labor & equipment 5.2 Unqualified workforce 48.64 ###### Median and range of frequency for delay causes groups. Delay causes group Owner respondents Consultant respondents Contractor respondents *P* value --------------------------- ------------------- ------------------------ ------------------------ ----------- Owner related 21(12--31) 24(11--28) 24(15--31) 0.369 Consultant related 6(5--11)^a^ 7(5--10)^a^ 10(6--13)^b^ 0.024 Contractor related 18(12--24)^a^ 17(12--23)^a^ 15(8--19)^b^ 0.059 Material related 5(3--7) 5(3--7) 4(3--7) 0.925 Labor & equipment related 8(5--10) 8(5--11) 7(4--11) 0.368 Project related 9(5--11) 10(6--12) 9(5--14) 0.891 External related 13(9--18) 12(8--17) 12(8--20) 0.764 Total 80(51--112) 83(50--108) 81(49--115) 0.966 *Notes*: The respondents group which has a different superscript b is significantly different from the other two respondents groups which have superscript a (e.g. Causes group 2). ###### Median and range of severity for delay causes groups. Delay causes group Owner respondents Consultant respondents Contractor respondents *P* value --------------------------- ------------------- ------------------------ ------------------------ ----------- Owner related 28(18--36)^a^ 28(18--37)^a^ 36(23--38)^b^ 0.065 Consultant related 13(7--17) 12(9--17) 15(11--17) 0.502 Contractor related 22(15--27) 25(17--30) 23(16--29) 0.670 Material related 7(5--10)^b^ 10(7--12)^a^ 9(6--11)^a^ 0.005 Labor & equipment related 14(9--16) 15(10--16) 11(9--15) 0.307 Project related 14(8--17) 17(11--19) 16(10--18) 0.237 External related 20(14--26)^a^ 19(14--24)^a^ 24(16--28)^b^ 0.022 Total 118(76--149) 126(86--155) 134(91--156) 0.401 *Notes*: The respondents group which has a different superscript b is significantly different from the other two respondents groups which have superscript a (e.g. Causes group 1). ###### Median and degrees of Frequency Index and Severity Index. Delay causes group Frequency Severity --------------------------- ----------- ---------- ----- ------- ------- ----- Owner related 23 28.28 V.H 30.67 24.34 V.H Consultant related 7.67 9.43 L 13.33 10.58 L Contractor related 16.67 20.49 H 23.33 18.51 H Material related 4.67 5.74 V.L 8.67 6.88 V.L Labor & equipment related 7.67 9.43 L 13.33 10.58 L Project related 9.33 11.47 L 15.67 12.44 L External related 12.33 15.16 M 21 16.67 M Total 81.34 100 --- 126 100 ---

1. Introduction {#sec1} =============== The introduction of silicone oil (polydimethylsiloxane, PDMS) to retinal detachment surgery in the early 1980s was one of the main steps in the effective treatment of this pathology \[[@B1]--[@B4]\]. In the last three decades, vitreoretinal surgery combined with PDMS tamponade has become the widespread treatment for complicated cases of retinal detachment caused by a proliferative process. Silicone application mainly serves two functions. The first is the displacement of the retina toward the eye-wall by its surface tension effect and volume displacement, and the second, to a lesser degree, is the tamponade of the superior retina by its flotation force. More than 30 years of clinical use has demonstrated that the tamponade effect of PDMS is usually sufficient, provided that the retina is completely mobile and provided that no new membranes develop. Moreover, the stability and immunological tolerability of PDMS make it relatively safe as a long-term internal tamponade. Histological examination of the human retina after more than 3 years of PDMS endotamponade did not show significant morphological alterations \[[@B5]\]. Intraretinal or intracellular deposits suggestive of silicone have been observed in attached retinas only if subretinal silicone deposition occurred in accidental situations \[[@B5]\]. However, PDMS and long-acting gases provide good support only for the superior retina and lack efficacy for the inferior retina, especially when the fill is subtotal. This makes these tamponade agents less useful for closing inferior retinal breaks and for defending them from the proliferative vitreoretinopathy (PVR) that usually begins in the inferior quadrants. Placing an agent that is heavier than water in contact with the retina should reduce the redetachment rate and the rate of inferior PVR. In the past two decades, clinicians and researchers have attempted to identify internal tamponades that are heavier than water and have good tolerability. The first heavy tamponade used was fluorinated silicone oil or fluorosilicone (FSiO), but its high rate of complications such as early emulsification and development of intraocular inflammation and PVR limited its use \[[@B6]\]. A second group of heavy internal tamponades, the perfluorocarbon liquids (PFCLs), was studied for prolonged postoperative endotamponade at the end of the 1980s. These are fully fluorinated alkane compounds with a high specific gravity. However, these compounds turned out to be unsuitable long-term internal tamponade because of the mechanical damage on the retina and the tendency for droplet dispersion \[[@B7]--[@B9]\]. Presently, these compounds are widely used as intraoperative tools, but not as vitreous substitutes. A third group of substances, liquid semifluorinated alkanes (SFAs), appeared to have the potential to act as heavy internal tamponade agents \[[@B10]\]; in particular, perfluorohexyloctane (F6H8) seemed to be well tolerated in long-term animal studies \[[@B11]\]. In clinical practice, the use of F6H8 provided adequate reattachment rates and few signs of retinal damage; however, it was associated with a high rate of postsurgical inflammation and an early rate of emulsification of droplets into the entire eye \[[@B12]\]. SFAs have the ability to bring the silicone oil into solution, creating a fourth category of heavy tamponades, the heavy silicone oils (HSOs). HSOs are admixtures of different concentrations of highly viscous PDMS and SFAs, combining the advantages of increased gravity and high viscosity. Some of these mixtures were more tolerated by ocular tissues compared to SFAs, and these mixtures have been successfully investigated as long-term endotamponades. However, in some clinical situations, the combination of two or more tamponade agents is suspected to increase postsurgical inflammation. Two compounds belonging to the HSO group are available for clinical use: Densiron 68 (Fluoron: a combination of F8H8 and silicone oil) and Oxane HD (Bausch and Lomb: a combination of olefins RMN3 and silicone oil). A third compound, HWS 46-3000, appeared to be well tolerated, but it is not yet available in common practice. In this review, we describe the current knowledge on HSOs and heavy tamponades and discuss the fundamental clinical and biochemical/molecular events involved in the ocular inflammation induced by these compounds. 1.1. Physical Properties of an Optimal Heavy Tamponade {#sec1.1} ------------------------------------------------------ The essential attribute of the PDMS is its ability to keep the retina in contact with the pigment epithelium by the hydraulic force of its volume displacement, thereby alleviating the tractions. The efficacy of an internal tamponade depends on its ability to make contact with the internal surface of the retina. The PDMS cannot flatten the retina because it has a weak flotation force; rather, it exerts a tamponade effect mainly by immobilizing the retina and reducing fluid circulation. Thus, the tamponade effect of the injected PDMS is modest and is not comparable to that of air or gas. A layer of fluid between the retina and the silicone bubble is always present, and closed contact between the oil and retina is not possible. However, the stabilization of the eye for a long period after surgery is the main advantage of silicone oil compared to gas. PVR is an exaggerated wound-healing phenomenon in which inflammation, proliferation, and remodeling lead to a retinal scar \[[@B13]\]. At the end of surgery, the meniscus of fluid that remains between the endotamponade and the retina is a milieu of rich proinflammatory cytokines and growth factors that promote PVR development, and this is the main cause of failure after retinal detachment surgery. With conventional "light" endotamponades (either gas or PDMS), the PVR is located in the inferior quadrants where the remnant fluid is displaced in almost all cases. In cases of inferior breaks, the contact between an agent that is heavier than water and the inferior retina can prevent the passage of aqueous through the hole and displace water upwards. An ideal heavy tamponade agent should possess the following qualities: optical clarity, no effects on the eye\'s refractive state, no toxic effects on eye structures, no effects on eye pressure, no cataractogenic effects, and the ability to inhibit inflammation, cellular migration, and glial proliferation \[[@B14]\]. Moreover, the following physical properties regulate endotamponade effectiveness: the difference in the specific gravity of the agent and the aqueous (buoyancy), the interfacial tension, and the viscosity \[[@B15]\]. Unfortunately, all of the presently used agents have both advantages and disadvantages related to their different properties. The specific gravity (the difference between the specific gravity of the agent and water) determines whether the tamponade will sink or float in water and the shape of the intraocular bubble. The specific gravity and the interfacial tension determine the effectiveness of an internal tamponade in the short term. The viscosity of the material is crucial for maintaining its integrity, thus reducing dispersion in the long term. In contrast to PDMS (specific gravity, 0.97 g/cm^3^), the high specific gravities of perfluorodecalin (1.93 g/cm^3^) and F6H8 (1.35 g/cm^3^) allow these substances to stay perfectly in contact with the lower retina. These compounds are able to flatten the retina because of their strong sinking force; they fit perfectly over all of the irregularities of the posterior pole and the recesses of the indents, and no fluid remains between the inferior retina and the tamponade agent. However, the specific gravity of these agents is probably too high, and the absence of water between the agent and the inferior retina produces a mechanical or metabolic negative effect that impairs retinal function \[[@B16]\]. The lower specific gravities of "lighter" heavy tamponades, such as Oxane HD (specific gravity, 1.02 g/cm^3^) and Densiron 68 (1.06 g/cm^3^) minimize these effects. Thus, these compounds should be less toxic, although this reduces their tamponade effects, especially in the presence of a retinal indents \[[@B17]\]. Further, an effective tamponade must have a high interfacial tension against water in order to push the retina toward the eye-wall. Gas or air has the highest interfacial tension against water (approximately 80 mN/m), whereas PFCLs and silicone oil derivatives (PDMS or HSO) have a lower tamponade capability because of their lower interfacial tension against water (around 40--45 mN/m or 35 mN/m, resp.). According to Archimedes\' principles, the tamponade force that presses against the retina depends on the gravity of a submerged bubble, namely, buoyancy. When in contact with water, the bubble of a light or heavy silicone oil is rounded because of its small "pressing" force. In an eye that is almost completely filled with a tamponade, this substance is in contact with only a portion of the retina (superior or inferior, depending on the gravity), while it forms a convex meniscus on the opposite side that is not in contact with the retina. The shape of this meniscus is more or less convex, depending on the physical characteristics (gravity, buoyancy, and superficial tension) of the substance; in general, a flat meniscus is a characteristic of a good endotamponade agent. For example, gas or air has a flat meniscus, while PDMS and HSO have a convex meniscus, and it seems that the Densiron 68 meniscus is less convex than the Oxane HD meniscus. In clinical practice, it remains unclear whether the differences between Densiron 68 and Oxane HD are significant. The essential role of any vitreous substitute is presumably its ability to fill the eye and maintain the retina in contact with the pigment epithelium rather than to flatten it \[[@B18]\]. Emulsification is a frequent complication associated with the use of heavy internal tamponade \[[@B19]\]. This phenomenon is influenced by many factors including the interfacial tension, the viscosity of the oil, and the presence of impurities such as low-molecular-weight siloxanes and catalytic remnants \[[@B15]\]. The viscosity rate is the main factor influencing emulsification; a reduction in viscosity reduces the mechanical energy needed to disperse a large bubble in small droplets. In theory, an intraocular tamponade should be highly viscous, thus reducing the tendency to emulsify and to disperse into small bubbles that can cross retinal breaks or the zonula to the anterior segment, causing inflammation or glaucoma \[[@B20]\]. Silicone oil, which has high viscosity (5000 mPas), is more stable and tends to have less dispersion; therefore, it is associated with a lower rate of complications related to emulsification compared to the less viscous PDMS (1000 mPas) \[[@B21]\]. In the clinical practice, however, the PDMS is usually removed after 3-4 months, and such a difference in dispersion may be not significant in this time interval. The high viscosity of 5000 mPas PDMS increases the difficulties associated with handling the substance. A PDMS of 1000 Cs can be introduced and removed much more easily than a PDMS of 5000 Cs; thus, the former is largely utilized by most vitreoretinal ophthalmologists. Moreover, with the advent of minimally invasive surgery (23--25 gauge), the use of a less viscous silicone oil is preferable in order to save time during its introduction and its passage through the small gauge system. Therefore, obtaining an HSO of low viscosity that does not generate the phenomena of emulsification would be desirable. Heavy tamponades have lower viscosity than PDMS: F6H8 and the other SFAs have a viscosity of 2.5--3 mPas, close to that of water (1 mPas). These compounds are easy to handle, but they tend to emulsify very early after surgery. Dispersion was described in 30% to 100% of cases treated after a few weeks with F6H8, depending on the time to removal. The mixture of SFA with a PDMS that has a viscosity of 1000 mPas can inhibit dispersion by F6H8, but this mixture was found to be unstable, depending on the temperature and movement of the eyes \[[@B22], [@B23]\]. The mixtures of an SFA and a PDMS with a viscosity greater than 5000 mPas, forming the HSO compounds, seem to be more stable. HSO compounds have higher viscosity than pure FSA: approximately 1400 mPas for Densiron 68 and approximately 3800 mPas for Oxane HD. Although this quality slows the emulsification rate, it influences their handiness during removal \[[@B24]\]. The amount of emulsification of heavy tamponades is, among other factors, time-dependent. Thus, the tendency to emulsify is the main factor that influences the time to removal of these tamponades. This factor is crucial for stabilizing the retina for the time that is necessary for PVR to develop (usually 4--6 weeks). The better tolerance of the new HSOs allows these substances to remain for up to 3-4 months without detrimental effects \[[@B25]\]. 1.2. Immune Response and the Proinflammatory Nature of HSOs {#sec1.2} ----------------------------------------------------------- The inflammatory response after prolonged retinal detachment and after vitreoretinal surgery peaks in the development of PVR, which occurs when the retinal cells are exposed to the inflammatory milieu in the vitreous humor \[[@B26]\]. The "PVR soup" consists of the aqueous humor containing growth factors and cytokines \[[@B27]\]; it tends to settle at the level of the inferior retina and posterior pole because of gravity \[[@B28]\]. This situation is common in complicated retinal detachment, but it is amplified after invasive surgery and by the use of intraocular tamponades that float over a subtle film of liquid where the inflammatory cytokines and growth factors reach the critical concentration over the inferior retina. The accumulation of the PVR soup beneath the inferior meniscus of the PDMS or gas exposes the inferior retina (in the orthostatic position) and the posterior pole (in the supine position) to factors that may generate epiretinal membranes. Heavy tamponades theoretically possess the quality to displace this inflammatory environment away from the inferior retina and the posterior pole \[[@B29]\]. With a heavy tamponade, the head movements during common daily postures should frequently displace the liquid meniscus from the upper retina to the posterior capsule of the lens. In contrast, with PDMS, head movements frequently displace the liquid from the inferior retina to the posterior pole, increasing the risk of damaging the macula. However, PDMS has been used for more than three decades and is appreciated for its stability and immunological tolerability, which make it safe for use as a long-term internal tamponade. The same level of safety has not yet been achieved by any of the heavy tamponades used up to now, especially for their physical and immunological interaction with ocular tissues. Many authors have noted that heavy tamponades are more prone to causing intraocular inflammation compared to standard silicone if they remain for several months in the eye. It is difficult and often impossible to distinguish between inflammation caused by the tamponade and the inflammatory reaction that is associated with the underlying complicated retinal disease. High inflammation can be commonly expected after a complicated retinal detachment surgery, and this is not related to the tamponade used. Fibrin formation, corneal edema, and cataract progression are frequent complications related to surgical trauma or to the ocular disease itself (i.e., in cases of retinal detachment after an ocular injury). Moreover, severe reproliferation is the major reason for anatomical and functional failure, and it can be seen with or without the use of heavy tamponades. However, detecting any possible adverse inflammatory event related to the physical characteristics of any endotamponade agent is crucial because it could modify or amplify the wound-healing response and stimulate PVR, which is the primary reason for visual loss and poor visual outcome. Four mechanisms are involved in the genesis of the inflammatory response: toxicity due to impurities or the instability of the agent, direct toxicity and immunogenicity, oil emulsification, and mechanical injury due to gravity \[[@B30]\]. 1.3. Toxicity due to Impurities or the Instability of the Agents {#sec1.3} ---------------------------------------------------------------- PDMS and FSiO contain impurities like linear and cyclic low-molecular-weight components (LMWCs), ionic compounds, and compounds with cleavable fluoride that are thought to cause ocular toxicity \[[@B31]\]. LMWCs (less than 2,500 Da) have high volatility and may diffuse as vaporized molecules into the surrounding tissues, where they can produce toxic effects. The vaporized siloxanes can also condense and become silicone oil droplets in areas of temperature change, such as near the iris or in the anterior chamber, or in presence of polarized molecules in the anterior chamber fluid. Further, the inactivated catalysts remaining in the silicone oil may be toxic. Severe inflammation and corneal edema can be induced when small species of linear and cyclic LMWCs of endotamponades are injected into the anterior chambers of animals. The ocular responses to the single species of the LMWCs increase as the molecular weights decrease. However, unpurified PDMS and FSiO, as well as purified oils (via solvent fractionation), usually do not cause significant adverse ocular responses, presumably because the amounts of LMWCs (especially the smallest species) in the oils are relatively small. Using gas chromatography, several authors analyzed the PDMS and FSiO recovered from rabbits and human vitreous cavities up to 2 years after injection and discovered that LMWCs may diffuse from the oils into the ocular tissues \[[@B32]\]. Although the long-term effect of LMWCs in the intraocular PDMS and FSiO has not been determined, the diffusion of LMWCs into ocular tissues may be related to the chronic ocular toxicity of the oils. In addition, postoperative emulsification of silicone is related to the number of low-molecular-weight polymer chains \[[@B32]\]. In HSOs, the semifluorinated alkanes are embedded in silicone oil molecules that may theoretically contain LMWCs. However, the companies that produce Densiron 68 and Oxane HD stated that these agents are 100% pure preparations and that they do not contain low-molecular siloxanes and other impurities. The biocompatibility of the SFAs and their admixtures with PDMS (the HSO) depends on the lipophilic behavior and on the molecular dimension of the semifluorinated alkanes. Because cell membranes and other physiological borders are composed of lipophilic substances, it is possible that they could be damaged or solubilized into the silicone bubble at certain temperatures \[[@B10]--[@B12]\]. The composition of the HSO may vary with time and temperature and from contact with other chemical agents. For example, the higher temperature of the anterior chamber might separate the F6H8 and PDMS in some situations. Further, the stability of the combination of two different agents may cause unexpected ocular toxicity. The interaction between F6H8 and other substances like PFCLs, PDMS remnants, or the cortical humor vitreous and humor acqueous may alter the stability between the two compounds and the properties of HSOs. The decomposition of these substances was shown to cause intraocular inflammation or phenomena like "sticky silicone oil" \[[@B33], [@B34]\]. It was shown that F6H8 might react with remnants of the humor vitreous and humor acqueous either in the vitreous base or in the posterior pole, creating whitish epiretinal membranes \[[@B35]\]. Even if an apparently complete exchange of the PFCL with air is assumed, a thin layer of the PFCL may remain on the retinal surface and in ciliary bodies; these remnants can be found in droplets in many patients months or years after the surgery at the follow-up visits \[[@B36]\]. The interaction between the HSO and the volatile remnants of the PFCL or vitreous remnants may generate drops of sticky silicone, a sort of "glued oil," attached on the retinal surface and, in the worst cases, on the macula. PFCL remnants were found in high concentrations in the sticky samples of several patients \[[@B33]\]. Contamination of the tamponade with the heavy liquids used during intraoperative manipulations is also suspected to cause granulomatous uveitis with the use of Oxane HD \[[@B37]\]. For these reasons, it is recommended that a PFCL-air exchange be performed before injecting any HSO in order to avoid direct contact between the PFCL and HSO, thus preventing unpredictable side effects. The biocompatibility of the SFA and their admixtures with PDMS (the HSO) is dependent on the lipophilic behavior and on the molecular dimension of the semifluorinated alkanes. Because cell membranes and other physiological borders are composed of lipophilic substances, it is possible that they could be damaged or solubilized into the silicone bubble at certain temperatures. 1.4. Direct Immunogenicity and Toxicity of the Compounds {#sec1.4} -------------------------------------------------------- The early clinical reports of some heavy tamponades showed a relatively high rate of intraocular inflammation. A fibrinoid reaction and even retinal necrosis associated with the use of high-density fluorosilicone oils as well as semifluorinated alkanes such as F6H8 and their oligomers have been reported \[[@B12], [@B35], [@B38]--[@B40]\]. In particular, F6H8 is suspected to increase the wound-healing reaction and to cause granulomatous reactions, fibrinoid reactions, and retropupillary membrane formation. The direct immunogenicity of this compound has been demonstrated by the finding of a granulomatous reaction with epithelioid cells containing minute drops of F6H8 \[[@B37]\]. The introduction of HSO reduced the rate of intraocular inflammation compared to previous reports. However, several cases of fibrin formation and unusual anterior chamber inflammation were reported either with Oxane HD or with Densiron 68 \[[@B37], [@B41]\]. An abnormal inflammatory reaction was not found in any patients treated with HWS 46-3000 \[[@B42]\]. The chronic presence of an intraocular endotamponade may also indirectly cause some form of toxicity. An endotamponade that remains in the vitreous cavity for several months may absorb endogenous substances from the ocular tissues or exogenous substances via the blood stream. The analysis of PDMS and FSiO extracted after several months of intraocular placement demonstrated the presence of cholesterol, retinol, and lipophilic acids that were extracted from the retinal cells or from the blood. Further, depending on their molecular dimensions and temperatures, SFAs may extract cholesterol from ocular plasma membranes that are damaged or are solubilized into the silicone bubble \[[@B10]\]. These findings suggest that intravitreal endotamponades containing PDMS or SFAs are not completely inert and may extract cellular components or accumulate substances not normally present in the vitreous cavity, and these substances may have a cytotoxic effect over time \[[@B43], [@B44]\]. 1.5. Emulsification {#sec1.5} ------------------- Heavy tamponades with a viscosity that is lower than that of silicone oil are more prone to emulsification compared to standard silicone oil, which in turn gives rise to inflammation. The dispersion and diffusion of a tamponade agent in the aqueous are responsible for the subsequent formation of an emulsion of droplets or "fish eggs" \[[@B45]\]. Emulsification is probably either the effect or the cause of intraocular inflammation, quite apart from the fact that individual agents might be a stimulant for inflammatory reaction. Intraocular inflammation promotes early emulsification of the endotamponade, while the diffusion of foreign molecules from the endotamponade promotes further inflammation. Minute bubbles of oil are suspected to trigger inflammatory cell chemotaxis and phagocytosis, which stimulate a foreign body-type reaction \[[@B46]\]. However, it is not clear whether the size of the bubble or the combination of the vesicle shape with a specific stabilizing surfactant activates neutrophils or stimulates phagocytosis by monocytes \[[@B47]\]. Dispersion also depends on the underfilling of the tamponade after surgery especially in large-volume eyes and if severe postoperative inflammation coexists. Silicone oils are composed of polymers and hence show the characteristics of non-Newtonian fluids, which means that the viscosity changes along with the share rate. Saccadic and pursuit movements of the eyes and of the head may cause intraocular fluid currents that exert shear stress on the silicone bubble surface. Therefore, the shear force or the lateral attrition, created by rotatory movements, exceeds the surface tension of the bubble, creating a dispersion of small fractions of the tamponade in small bubbles. Because the viscosity of silicone oil is determined by its molecular weight, low viscosity silicone emulsifies more easily. Differences in the rates of emulsification are not due to differences in surface tension because surface tension changes minimally with increasing viscosity. Different samples of silicone oil with the same viscosity may be composed of a narrow band of different molecular weight chains containing only a few short chains, whereas another sample of the same viscosity may be composed of a wider range of molecular weight chains with more short-chain molecules capable of emulsification. The homogeneity of the silicone components and the low concentration of the LMWCs are important factors for avoiding toxicity and emulsification. While emulsification is transitory in the first phase, it becomes permanent in the presence of blood components and inflammatory proteins that act as surfactants \[[@B47]\]. Red blood cell membranes, plasma lipoproteins, and HDL-apolipoproteins support silicone oil emulsification \[[@B48]\]. Further, vigorous physical activity with the tamponade in situ is reported as a possible cause of dispersion, opacification of the endotamponade, and intraocular inflammation \[[@B35]\]. Finally, the contact of silicone oil with any type of substance during a direct exchange may increase emulsification \[[@B49]\]. The first agents used as heavy tamponades (FSi, PFCL, and F6H8) have low viscosity and fast intraocular emulsification; however, the resistance to extensional deformation and therefore the extensional viscosity of F6H8 may be increased by mixing a certain amount of very long-chain silicone molecules into the heavy tamponade. This maintains the specific weight at a value greater than 1 and increases the resistance to emulsification. Rachel et al. studied a combination of high-molecular-weight (423 kDa) PDMS and silicone oil 1000 at 5% and 10% w/w concentrations in order to increase the emulsification resistance of the tamponade agents while maintaining ease of injection and removal \[[@B50]\]. HSOs are derived from a mixture of a highly viscous PDMS (more than 5000 mPas) and different semifluorinated alkanes (F6H8, F4H5, and F4H6) or a similar substance (RMN-3), and these have a lower tendency to create dispersion and emulsion. However, the concentration of the two components may vary with time and temperature, and the possible chemical decomposition of HSO has been reported, where the heavier component tends to settle over time in the inferior part of the bubble, separating it from PDMS. Thus, the specific gravity of HSO in the eye may become heterogeneous over time with the oil because the SFA dissociates from the silicone oil, thereby producing droplets of PDMS and droplets of SFA. This dissociation may result in an anterior uveal reaction \[[@B51], [@B52]\]. The iris pigment could be absorbed by HSO droplets in some cases, leading to iris depigmentation \[[@B52]\]. In an in vitro model, Caramoy et al. demonstrated that increasing the extensional viscosity by the addition of small amounts of very long-chain silicone molecules significantly influenced the reduction of the emulsification for 1000 cSt silicone oil (Siluron 2000) and for 1000 cSt silicone oil with an admixture of F6H8 (Densiron 68 HV) \[[@B53]\]. These findings are expected to be investigated further in an in vivo model. 1.6. Effect of Gravity in Long-Term Vitreous Tamponade {#sec1.6} ------------------------------------------------------ Previous reports showed that PFCL agents (perfluorodecalin, perfluoroperhydrophenanthrene, and perfluorooctane) are clinically tolerated in the eyes for only a few days (5--7 days) \[[@B54]--[@B57]\]. Mechanical pressure on the retina may be partly responsible for the changes observed in the retina when PFCL agents are used. These considerations are mainly dependent on experiments and histological evaluations conducted in animal models. A few weeks of endotamponade with PFCL may cause the following ultrastructural changes in the inferior retina of rabbits: narrowing of the outer plexiform layer, ultrastructural distortions of the photoreceptor outer segments, and migration of the receptor cell nuclei to the photoreceptor layer \[[@B8], [@B9], [@B58]\]. These changes may represent a mechanical rather than toxic effect; in fact, similar changes have been reported in the superior retina in silicone-filled eyes. The specific gravity of PFCL ranges from between 1.7 g/cm^3^ and more than 2.0 g/cm^3^. The histologic changes in the retina may be partly attributed to the dystrophic effect of the "heavy" liquids that press the inferior retina. However, it was noted that the retinal damage was more evident in the external layers rather than in the inner retinal layers that are in direct contact with the heavy substance. A mechanism of damage different from a simple mechanical interaction was assumed. Recent observations indicate that PFCL toxicity is not primarily due to the high specific gravity or possible chemical impurities but rather due to their inability to dissolve ions. Gravity might not be causally linked to retinal damage that may rather depend on a metabolic disturbance. OCT measurements indicate that PFCLs, including the semifluorocarbon PFH with low specific gravity, replace most of the aqueous sink volume available for potassium (K^+^) siphoning. Thus, impairment of retinal K^+^ clearance may be an important mechanism of PFCL-induced retinal injury. These observations explained the morphological alteration reported regarding Müller cells. Müller cells have been shown to develop features of reactive gliosis including hypertrophy, expression of glial fibrillary acidic protein, and drop-like protrusions between the inner segments of the photoreceptors. Müller cells may be directly injured by the elevated \[K^+^\], thus causing subsequent atrophy of the photoreceptors that occupy the external retinal layers. HSO is less efficient compared to PFCLs and SFAs at remaining in contact with the retina and is unable to fit into small recesses; however, this relatively poor contact allows a thin film of aqueous to remain in contact with the retinal surface, and this is important for retinal cell survival and for potassium siphoning by retinal Müller cells \[[@B16]\]. 2. Internal Tamponade Agents {#sec2} ============================ 2.1. Fluorosilicone {#sec2.1} ------------------- Fluorinated silicone oil (trifluoropropylmethylsiloxane or fluorosilicone-FsiO), which has a density of 1.30 g/cm^3^, was the first heavy tamponade used. Clinically, it was marked by immediate, albeit transient, iritis. The ocular toxicity of fluorinated silicone oils was attributed to their low-molecular-weight components and to their high dispersion rate \[[@B31], [@B32]\]. In animal models, FsiO fluorosilicone caused inflammatory responses that exceeded those observed with PDMS \[[@B6]\]. This agent is thought to promote PVR in the longer term, with an epiretinal membrane forming around the oil bubble. Histologically, these membranes showed foreign body reactions \[[@B59]\]. A copolymer of PDMS and FsiO was evaluated in order to avoid the anatomical damages caused by PFCLs with the aim of decreasing the specific gravity (density, 1.16 g/cm^3^) of the tamponade \[[@B60]\]. The atrophic retinal changes were much less than those observed with the heavier perfluorotetradecahydrophenanthrene (density, 2.03 g/cm^3^). However, thinning of the outer plexiform layer in rabbit retina was still observed after 6--8 weeks and small droplets ingested by mononuclear cells were found in the vitreous cavity or preretina after 4--6 months \[[@B61], [@B62]\]. 2.2. Perfluorohexyloctane (F6H8) {#sec2.2} -------------------------------- Perfluorohexyloctane (F6H8) is the most extensively investigated agent belonging to a group of fluorinated hydrocarbons: the semifluorinated alkanes (SFA) \[[@B10]\]. These agents have specific gravities greater than those of water, but slightly lower than those of perfluorooctane (1.35 g/cm^3^), and their surface tension and interfacial tension against water are equal to those of perfluorocarbon liquids (45.3 mN/m). F6H8 is chemically and physically inert because of the strength of its hydrocarbon (C--H) and fluorocarbon (C--F) bonds. The fluorocarbon moiety is lipophobic, while the hydrocarbon moiety is lipophilic; thus, the SFAs are amphiphilic molecules that are soluble in both silicone oils and perfluorocarbon liquids but are insoluble in water. F6H8 is a biocompatible compound that was investigated as a candidate for blood substitutes \[[@B63]\]. F6H8 was well tolerated for three months in rabbit eyes \[[@B11]\], and it was introduced initially as a solvent for silicone oil to remove silicone oil remnants from intraocular surfaces \[[@B64], [@B65]\]. Further, F6H8 was investigated as an intraoperative tool and as a long-term tamponade in several small case series \[[@B12], [@B66]\]. Its low density and viscosity (2.5 mPas) reduced the risk of mechanical retinal damage, but it promoted dispersion and the phenomena of emulsification in the eye in up to 100% of all treated cases \[[@B67]\]. According to some authors, the ability of this compound to generate inflammatory responses is mostly due to its propensity to disperse and to form small, emulsified droplets. Minute bubbles of oil subsequently trigger chemotaxis of inflammatory cells and phagocytosis \[[@B46], [@B47]\]. Despite good results with the use of F6H8 in animal models and in some small case series \[[@B11]\], studies conducted in vitro and in vivo showed evidence that F6H8 had proinflammatory activity. In preclinical studies, blood-retinal barrier breakdown associated with local vasoconstriction, hypertrophy of Müller cells, and vacuolization of the inner retinal surface were observed in rabbit retinas after 6 weeks of tamponade \[[@B68]\]. An evaluation using the live/dead assay on cultured ocular cells that were incubated with F6H8 for up to 5 days showed a significant reduction of vital EPR cells. Due to its lipophilicity, F6H8 seemed to be able to interact with cell membranes, causing a change in the adherence of the cells to extracellular matrix \[[@B69]\]. Some evidence for an irritating effect has been observed in clinical pilot studies using F6H8 as a retinal tamponade. In some cases, retrolental, epiretinal, and simil-PVR membranes were associated with its use as a prolonged tamponade. These membranes were similar to the classical PVR membrane histologically, but they also exhibited dense macrophagic infiltration and foreign body reactions. Further, they contained vacuolated and pigmented CD68-positive cells, exhibiting a macrophagic and EPR phenotype. These observations supported possible differentiation of the EPR cells in response to the proinflammatory stimulus induced by F6H8 \[[@B35]\]. The presence of intracellular droplets of F6H8 in the vacuolated cells suggested that the contact with the oil in the form of microemulsion causes activation of the monocyte-macrophage population. This finding indicated that the inflammatory reaction was enhanced by droplets of a certain vesicle size. In an in vitro study, however, the inflammatory response appeared only when the vesicles interacted with specific artificial, but not natural surfactants \[[@B46]\]. From the clinic-pathological point of view, the inflammatory reaction leads to the formation of epiretinal membranes that sometimes extend to the posterior surface of the lens. The difference between these membranes and the classical ones encountered in PVR is greater infiltration of leukocytes, which appear to be mostly CD68-macrophages, or rather RPE cells, which have "transdifferentiated" to a macrophage-like phenotype \[[@B47]\]. Epithelioid cells, which are typical of a granulomatous reaction, were found in some specimens, suggesting that emulsified F6H8 could result in the release of growth-promoting factors for macrophages. Regarding the development of retrolental membranes, it is known that silicone oil usually causes cataract formation because it interferes with the metabolism of posterior capsule epithelial cells \[[@B70]\]. The microscopic examination of lens capsule in eyes after F6H8 tamponade demonstrated the presence of macrophages adhering to the lens capsule with epithelioid cells and with fibroblastic differentiation, thus adding a probable inflammatory genesis to cataract formation \[[@B47]\]. 2.3. Other Perfluoroalkanes Oligomers: Perfluorobutylpentane (F4H5), Perfluorobutylhexane (F4H6), and Perfluorobutyloctane (F4H8) {#sec2.3} --------------------------------------------------------------------------------------------------------------------------------- A recent study by Mackiewicz et al. conducted on rabbits showed that the use of different semifluorinated alkanes leads to quite different immunologic reactions. Whereas F6H8 (perfluoroexyloctane) and F4H5 (perfluorobutylpentane) were well tolerated, F4H6 (perfluorobutylhexane) and F4H8 (perfluorobutyloctane) resulted in a severe inflammatory response, which appeared to be more pronounced when these substances were used in pure form rather than in an admixture with silicone oil. Microscopic investigation showed that the vitreous was replete with immune cells, mostly neutrophils. Chemically, these tamponades are amphiphilic (either hydrophilic or lipophilic). The capacity to penetrate the cellular membranes depends on the lipophilic property, and this is directly proportional to the length of the alkylic chain. A minimal increase in the lipophilic properties of some semifluorinated alkanes may lead to their penetration into the cell membranes, causing cellular damage and complete disorganization of the retinal layers and lens structure \[[@B30]\]. However, experimental studies have produced a new biocompatible perfluoroalkane, F4H5 (perfluorobutylpentane). The combination of F4H5 with PDMS 100.000 mPas gave rise to a new HSO, HWS 46-3000. This oil is very viscous (3109 mPas); it did not show a tendency to emulsify in clinical trials, and it is well tolerated. However, its high viscosity limits its use because the removal of this oil is reportedly difficult and time-consuming \[[@B42]\]. 2.4. Heavy Silicone Oils {#sec2.4} ------------------------ ### 2.4.1. Oxane HD {#sec2.4.1} Oxane HD (Bausch and Lomb, Toulouse, France) is a mixture of 5700 mPas PDMS and RMN-3 (perfluorooctyl-5-methyl-hex-2-ene), a mixed fluorinated and hydrocarbonated olefin. The surface tension and interfacial tension of this agent against water are similar to those of perfluorocarbon liquids (41 mN/m), and its specific gravity is only slightly greater than that of water (1.02 g/cm^3^). Its high viscosity (3800 mPas) reduces the risk of early emulsification. The rate of inflammatory reactions related to the use of Oxane HD was reported to be from 3% to 37% of treated patients ([Table 1](#tab1){ref-type="table"}). The immunogenicity of Oxane HD was investigated in a recent study in which immunohistochemistry was performed on epiretinal membranes formed in redetached retinas under this HSO \[[@B76]\]. Using monoclonal antibodies against retinal pigment epithelium cells, glia, macrophages, and T-lymphocytes, the inflammatory cell population was found to be similar to that obtained with conventional silicone oils; however, several aspects emerged that were attributed to a reaction against a foreign body. CD68-positive macrophages and epithelioid cells containing phagocytosed silicone oil were found in the area adjacent to the fibrocellular component of the membrane. Another study that investigated intraocular inflammation following endotamponade with Oxane HD showed that 37% of treated patients presented with a severe inflammatory reaction that assumed the characteristics of a granulomatous anterior uveitis \[[@B37]\]. Seven patients in this series developed pigmented endothelial precipitates, flare, and cellularity of the aqueous humor. In contrast to what has been shown in other studies, the uveitic reaction did not regress after the administration of topical corticosteroids and was reversible only after tamponade removal. The immune reaction was attributed to a granulomatous type IV reaction, in which an immune complex of insoluble antigens can cause T-lymphocyte-mediated reaction. The high percentage of intraocular inflammation in this series was probably due to the intraoperative contact between the Oxane HD and the PFCL. In fact, other authors did not report this phenomenon. Thus, a direct exchange between PFCL and Oxane HD has been contraindicated, and a PFCL-air exchange is recommended before injecting the HSO. Several case series performing this maneuver did not report uncommon posterior chamber reactions; thereby it was concluded that Oxane HD is well tolerated by the eye for up to 3 months of the endotamponade period \[[@B71], [@B72], [@B73]\]. ### 2.4.2. Densiron 68 {#sec2.4.2} Densiron 68 (Fluoron, Neu Ulm, Germany) is an admixture of F6H8 (30.5%) and PDMS 5000 mPas (69.5%); thereby, the viscosity was increased to 1387 mPas. This translates into a reduced ability for dispersion and emulsification, consequently reducing irritability to ocular structures \[[@B41]\]. Hence, compared with F6H8 alone, Densiron 68 is associated with significantly less inflammatory side effects \[[@B24]\] ([Table 1](#tab1){ref-type="table"}). A comparison of Densiron 68 with 1000 mPas PDMS demonstrated that Densiron 68 does not have a higher rate of postoperative inflammation in the middle period \[[@B79]\]. Moreover, in cases likely to develop PVR, Densiron 68 was demonstrated to be useful for avoiding repeated surgeries with scleral buckle usage \[[@B80], [@B79]\]. A common finding was a mild-to-moderate anterior chamber reaction \[[@B25], [@B81], [@B84]\]. This inflammatory reaction was sometimes associated with the development of fibrous membranes, the appearance of keratic precipitates, and cataract formation with inflammatory precipitation on the lens. Posterior capsular opacification could be caused by an increased cellular infiltration as a reaction to emulsified tamponade \[[@B81], [@B84]\]. The percentage of patients who developed significant postoperative inflammation varies greatly in different studies, depending mostly on the tamponade period. The probability of having complications increases if prolonged retention of this agent is required. A high rate of inflammatory reactions (40.7%) was recorded in a study in which Densiron 68 remained for more than 6 months \[[@B83]\]. In this retrospective study, an inflammatory reaction that was sometimes associated with fibrin exudation or with the appearance of a sterile hypopyon was detected in 11 patients out of 29 affected by complicated inferior retinal detachment. Due to its low viscosity, Densiron 68 also appears to be correlated with a high rate of dispersion and emulsification in droplets, which in turn precipitates inflammation if a long tamponade period is required \[[@B84], [@B86], [@B87]\]. ### 2.4.3. HeavySil (HSIL) {#sec2.4.3} HeavySil (ALCHIMIA srl, Padua, Italy) is made from the combination of high purity 75% silicone oil 5000 cSt (polydimethylsiloxane) and 25% perfluoroalkyloxyoctane (C11H11F13O); it has a density of 1032 and a viscosity of 1500 cSt. Its stability and high affinity for silicone oil are due to the presence of a partially fluorinated ether instead of an alkane. In a prospective, noncomparative interventional study on 31 consecutive eyes, Romano et al. investigated the anatomic and functional results and complications of this ocular tamponade. They found that HSIL is a safe and effective tamponade agent for the treatment of complicated RD; the main complications were cataract formation (71%), emulsification (19%), sticky oil formation (9.6%), and severe intraocular inflammation (3.2%) ([Table 1](#tab1){ref-type="table"}). One of the coauthors (B. Parolini) reviewed retrospectively 13 eyes of 13 patients with retinal detachment complicated with inferior PVR, treated using HeavySil 1500 as tamponade. All surgeries were performed with standard three-port 20-gauge pars plana vitrectomy. Additional surgical procedures such as membrane peeling and relaxing retinotomy were performed when necessary to allow retinal reattachment. Retinal breaks were treated by endophotocoagulation. In patients with preexisting endotamponade, the silicone oil was removed first. All patients were pseudophakic and underwent already at least one previous vitreoretinal surgery. Three patients were lost at 16-month follow-up. After tamponade with HeavySil, retina appeared to be attached in 9 cases over 10 (90%). Only one patient developed an IOP increase that was successfully treated with topical therapy. Another patient presented with emulsification in anterior chamber. Persistent subretinal fluid was never detected after surgery. Mean best corrected visual acuity was 2.1 ± 0.2 logMar preoperatively and 0.9 ± 0.1 logMar postoperatively. Three cases developed severe retinal inflammation 2 weeks after Heavysil 1500 tamponade. All three patients presented with optic disc swelling and retinal edema with diffuse narrowing of arteries and veins ([Figure 1](#fig1){ref-type="fig"}). One patient developed pain and the other two developed significant discomfort. Another case showed retinal inflammation with features resembling herpes retinitis, although virology was negative. In all cases, oil removal was performed within 1 week after the occurrence of retinal inflammation. The appearance of the fundus slightly improved within 2 weeks after oil removal. Silicone oil was analysed in these three cases with cytology and only in one case inflammatory cells were found. During oil-removal surgery, the sticky oil phenomenon appeared in one case. A retinal tissue sample was collected for histology examination; however, the result showed nonspecific signs of inflammation (Figures [2](#fig2){ref-type="fig"}, [3](#fig3){ref-type="fig"}, and [4](#fig4){ref-type="fig"}). In this particular group visual acuity remained very low after surgery even if retinal reattachment was reached in all patients. The cause for final low vision was cystoid edema in one patient and persistent macular hole and retinal thinning in the other two. It is difficult and often impossible to distinguish between problems caused by the tamponade and those that are associated with the underlying complicated retinal disease. In these three cases the timing of acute appearance of inflammation and retinitis was considered significant and differed from other more chronic and subtle signs of silicone related inflammatory reactions. Our case series, with all the limitations due to the retrospective examination of data, shows a significant rate of severe acute retinal inflammation when using HeavySil (70%). Larger prospective clinical trials will be needed in order to define the safety of this new heavy tamponade. 2.5. HWS 46-3000 and HWS 45-3000 {#sec2.5} -------------------------------- HWS 46-3000 and HWS 45-3000 are admixtures of 45% silicone oil 100,000 and 55% perfluorobutylhexane (F4H6) and perfluorobutylpentane (F5H6), respectively. Of the three new generation tamponades, HWS 46-3000 is the heaviest and has the greatest viscosity (3.109 mPas). In Rizzo\'s pilot study of a case series of 32 patients published in 2003, the major side effect detected was the development of early posterior subcapsular cataract (100%); intraocular inflammation and emulsification were not observed ([Table 1](#tab1){ref-type="table"}). Rizzo et al. postulated that the low rate of postsurgical reproliferation and epiretinal membranes formation (9%) was due to adequate contact with the buffering of the retina, reducing the infiltration of the PVR soup. HWS 45-3000 has a density of 1.118 and a viscosity of 2.903 mPas. In 2010, Rizzo et al. did not observe significant emulsification or a significant inflammatory reaction with this agent. 3. Conclusions {#sec3} ============== The treatment of complex retinal detachments using internal tamponade agents produces successful restoration of vision in many cases. However, the recurrence rates for complicated retinal detachment are as high as 20--25%, and this rate increases in the presence of PVR \[[@B4]--[@B13]\]. Although vitreoretinal techniques have been improved over the past years, the rate of PVR has not decreased considerably \[[@B90]\]. PDMS or gas exposes the inferior retina (in the orthostatic position) and the posterior pole (in the supine position) to proinflammatory growth factors and cytokines that may generate epiretinal membranes. Compared to PDMS, heavy tamponades theoretically possess the quality to provide better protection to the posterior pole from PVR \[[@B29]\]. With a heavy tamponade, the head movements during common daily postures are expected to displace the liquid meniscus from the upper retina to the posterior capsule of the lens frequently. In contrast, with PDMS, head movements frequently displace the liquid from the inferior retina to the posterior pole, increasing the risk of damaging the macula. Further, when postoperative posturing is more important, such as in cases of posterior breaks, macular hole in highly myopic eyes, or inferior retinectomies, heavy tamponades are advantageous, especially for patients with orthopedic disability or mental retardation and for children \[[@B91], [@B92]\]. The hypothetical advantage of using a HSO is that the physical separation of the "PVR soup" from the effector cells (retinal pigment epithelial cells, Müller cells, and fibroblasts of the inferior breaks) inhibits or mitigates fibroplasia. The most important presumed advantage for HSO compared to PDMS is a lower redetachment rate after endotamponade removal and a lower rate of macular redetachments. However, the preliminary results of a recent multicentric randomized trial failed to demonstrate the real superiority of HSO in comparison with standard PDMS in eyes with proliferative PVR of the lower retina \[[@B93]\]. Regarding final acuity, HSO was neither inferior nor superior to PDMS in almost all clinical series. Further, the rate of PVR in HSO-treated patients was not inferior to that registered for PDMS-treated patients; rather, HSO caused a shift of the PVR to the upper retina above the horizontal meridian \[[@B93]\]. To prevent this complication, several authors proposed performing a prophylactic superior laser photocoagulation, while others suggested a shorter endotamponade period with subsequent silicone oil use \[[@B25], [@B81]\]. The presence of a subtle meniscus of fluid around bubbles with a specific gravity very close to that of water is probably the main reason for the diffusion of growth factors and cytokines from the inferior breaks to the upper retina, which generates epiretinal proliferation. On the other hand, this subtle meniscus of fluid is essential for the correct K^+^ siphoning of the Müller cells and is necessary for avoiding the functional damage due to the excessive drying of the retinal surface that has been reported for heavier agents or for gas \[[@B94]--[@B96]\]. The rates of complications, such an inflammatory reaction, macular epiretinal membranes, IOP rise, cataract, and emulsification of HSOs, seemed to be similar to those in patients treated with PDMS in the middle period. This indicates that the intraocular behavior and tolerance of HSOs and PDMS would be similar if they remained in the eye for 3-4 months. This is an important safety result obtained for HSOs in comparison with all of the previously used heavy agents, because none of these agents could be utilized for such a long period without severe complications. The clear advantages of using HSOs rather than PDMS are shortening of the surgical time, easy handling, and a reduction in the necessity for utilizing external buckles or macular indents. However, when using HSOs, a strict follow-up period is required and the timing of the endotamponade removal should be respected more strictly in comparison with PDMS. Intraocular inflammation is common if it remains for more than 6 months \[[@B83]\]. The real utility of the use of HSOs depends on the correct selection of the patients for treatment. In a number of situations, such as myopic macular holes with or without retinal detachment, myopic foveoschisis, penetrating ocular injuries with retinal detachment, and inferior giant retinal tears, treatment with a heavy substance is easier and should therefore be the first choice. Moreover, HSOs offer new strategies for treating very complicated cases of retinal detachment caused by a proliferative process, such as alternating the tamponade agent in two different surgeries (i.e., first using an HSO and using PDMS or gas after a few months), or by combining in a single step the HSO with PDMS to reach a tamponade effect on both the upper and lower retina \[[@B97]\]. A mixed bubble of 70% Densiron 68 −30% PDMS has been recently used to obtain a "filling effect," suggesting that this strategy could minimize the stress produced by the tractional forces originating from eye movements. However, the results of using HSOs in many of these clinical situations have not yet been evaluated in extensive multicentric clinical trials. In conclusion, the introduction of HSOs represents an improvement in vitreoretinal techniques because the intraocular tolerance of these agents is good for 3-4 months. Even if the goal to prevent PVR formation is not reached and the visual results obtained with HSOs are comparable and not superior to those obtained with the "old PDMS," these new agents represent a useful new surgical tool. In the same way that the small gauge vitrectomy represents an improvement over the "old" 20-gauge vitrectomy, the HSOs are better than the "old PDMS" in some clinical situations. Although both agents obtain comparable visual results, the new agent gives similar results while the procedures are performed more easily, with appreciable advantages for both the surgeon and the patient. Conflict of Interests ===================== The authors declare that there is no conflict of interests regarding the publication of this paper. {#fig1} {#fig2} {#fig3} {#fig4} ###### Ocular complications after HSOs use in CRT. -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Author, year (reference) Tamponade *N* Pathology Time to removal Follow-up Complications Conclusions -------------------------------- ----------------- ----- ---------------------------------------------------- ----------------- ----------------------------------------- --------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Wolf et al.\ Oxane HD 33 Complicated RRD of inferior quadrants Within 3 months 12--16 months Rise in IOP (18%)\ Complications are similar to those reported with conventional silicone oil 2003 \[[@B71]\] Pupillary block (6%)\ Marked AC inflammation (3%)\ Retinal hemorrhages (6%) Theelen et al.\ Oxane HD 19 Complicated RRD of inferior quadrants 1--4 months 2--4 months after tamponade removal Keratic precipitates, pigmented clumps, and anterior chamber cellular reaction (37%)\ Inflammatory response resembling granulomatous uveitis; it is likely that Oxane HD is an immunogenic agent 2004 \[[@B37]\] Emulsification (11%) Wong et al.\ Densiron 68 42 Complicated RRD of inferior or posterior quadrants 10--16 weeks \>3 months after tamponade removal Cataract progression (100%)\ We neither observed clinically significant dispersion nor found any inflammation more than we would except from routine vitreoretinal surgery 2005 \[[@B41]\] Rise in IOP (8%)\ Moderate AC inflammation (8%) Rizzo et al.\ Oxane HD 28 Complicated RRD 45--96 days 6 months after tamponade removal Cataract progression (38%)\ Good intraocular tolerance with few minor complications 2006 \[[@B72]\] Rise in IOP (14%)\ Tamponade in AC (4%)\ Membrane formation (54%) Sandner and Engelmann\ Densiron 68 48 Complicated RRD 27--4000 days 103 days after tamponade removal Dispersion (16%)\ Compared with conventional silicone oil, a temporary inflammation and early emulsification developed more frequently with Densiron 68 2006 \[[@B81]\] Emulsification (15%)\ Hypotony (2%) Ocular hypotension (13%)\ Glaucoma (10%)\ Cataract progression (50%)\ Moderate AC inflammation (21%)\ Sterile hypopyon (4%) Cheung et al.\ Oxane HD 12 RD secondary to myopic macular hole 3-4 months 9--15 months Transient rise in IOP (42%)\   2007 \[[@B73]\] Mild oil emulsification and transient peripheral choroidal detachment (8%) Rizzo et al.\ HWS 46-3000 32 Complicated RRD of inferior or posterior quadrants 3 months 6 months after tamponade removal Early posterior subcapsular cataract (100%)\ No evidence of emulsification and intraocular inflammation 2007 \[[@B42]\] Epiretinal membranes (9%)\ Rise in IOP (0.3%) Sandner et al.\ Densiron 68 12 Complicated primary RRD 33--126 days 400 days after tamponade removal Emulsification (17%)\   2007 \[[@B25]\] Ocular hypotension (8%)\ Glaucoma (17%)\ Cataract progression (100%)\ Moderate AC inflammation (33%)\ Suspected intraretinal gliosis (25%) Berker et al.\ Oxane HD 21 Complicated primary RRD 3 months 11.5 months Rise in IOP (14%)\ Its complications were acceptable, and mostly due to its physical properties 2007 \[[@B74]\] Cataract progression and PVR (19%)\ Dispersion (9.5%)\ Rubeosis iridis (14%)\ Vitreous hemorrhage and optic atrophy (5%)\ Ocular pain and photophobia (100%) Boscia et al.\ Oxane HD 10       Posterior capsular opacification (22%)\   2008 \[[@B75]\] ERM (30%) Romano et al.\ Densiron 68 41 Complicated RRD of inferior quadrants 3 months 6 months AC shallowing (5%)\ Analyzing the observed side effects in our series, we found no presence of clinically significant emulsified Densiron 68 or intraocular inflammation 2008 \[[@B80]\] Emulsification (5%)\ Posterior synechiae (5%)\ Cataract progression (54%)\ Rise in IOP (2.4%)\ PVR (2%) Majid et al.\ Densiron 68 40 Complicated or primary RRD 9--14 weeks 6--12 months after tamponade removal Emulsification (20%)\ Emulsified Densiron may have contributed to significant intraocular inflammation, ERM formation, and CMO. This has potentially significant implications on the indications for Densiron-68 use. 2008 \[[@B82]\] Fibrinous uveitis (5%)\ ERM (12%)\ Cataract progression\ CMO (8%) Auriol et al.\ Densiron 68 27 Complicated primary RRD 14 weeks 6 months AC inflammation with fibrin accumulation (41%)\ Special attention must be paid to unusual adverse effects like inflammatory reactions and fibrin accumulation in the anterior chamber. 2008 \[[@B83]\] Sterile hypopyon (1%)\ PC inflammation with preretinal membranes without traction (0.3%)\ Hyphema and endothelial corneal dystrophy (1%)\ Rise in IOP (25%)\ Emulsification with pseudohypopyon (1%) Wickham et al. 2010 \[[@B76]\] Oxane HD 18 Complicated RRD of inferior quadrants 3 months 6 months Postoperative PVR (28%)\ Histopathological analysis showed that the structure and associated inflammatory response of membranes were similar to those observed following the use of conventional oils Hypotony (17%)\ Uveitis (11%)\ Rise in IOP (22.2%)\ Glaucoma (6%) Meng et al.\ Oxane HD 40 Complicated RRD Mean 87.9 days Mean 438.1 days after tamponade removal Mild-to-moderate AC inflammation (45%)\ Oxane HD showed an encouraging anatomical and functional success rate and good intraocular tolerance, with a few complications in complicated RD patients 2010 \[[@B77]\] Fibrin accumulation (40%)\ Temporary pupillary synechiae (15%)\ Rise in IOP (17.5%)\ Emulsification and dispersion (22.5%)\ Cataract progression (77.8%)\ ERM (30%) Li et al.\ Densiron 68 21 Complicated primary RRD 1--3 months 15 months Mild-to-moderate AC inflammation (24%)\ Vitreoretinal surgery with temporary Densiron 68 intraocular tamponade appears to increase anatomical success, while giving rise to minimal complications, in selected cases of complicated RD and PVR 2010 \[[@B84]\] Severe intraocular inflammatory reaction (10%)\ Emulsification (24%) and pseudohypopyon (10%)\ Early posterior capsular cataract (24%)\ Choroidal detachment/ocular hypotension (5%)\ Rise in IOP (20%)\ Pupillary block (5%)\ Transient corneal edema (5%) Li et al.\ Densiron 68 27 Complicated RRD of inferior or posterior quadrants 3 months 15 months Cataract progression (25%) or early posterior capsular opacification (26%)\ Postoperative complications did\ 2010 \[[@B85]\] Intraocular inflammation (22%)\ not increase significantly in the vitreoretinal surgery with temporary Densiron 68\ Corneal edema and rubeosis iridis (7%)\ intraocular tamponade Emulsification and dispersion (19%)\ Rise in IOP (19%)\ Intraocular hemorrhage (4%)\ Choroidal detachment/ocular hypotension (11%)\ Pupillary block (4%)\ PVR (15%) and ERM (7%) Ang et al.\ Oxane HD 18 Complicated RRD of inferior quadrants Mean 27 weeks Mean 66 weeks Posterior capsular opacification (22%)\ The relatively high rate\ 2010 \[[@B78]\] Severe intraocular inflammation (6%)\ of emulsification and increased\ Pseudohypopyon (6%)\ risk of intraocular complication when compared to other reported series is less promising and\ Emulsification (33%)\ warrants further evaluation Rise in IOP (11%)\ ERM (28%) Rizzo et al. 2011 \[[@B88]\] HWS-45 3000 10 Complicated RRD of inferior quadrants 2 months 6 months after tamponade removal Tiny droplets in CA (30%)\ HWS-45 3000 appears to be a well-tolerated heavy oil suitable\ Mild inflammatory reaction (40%) for the treatment of complicated inferior retinal detachment Romano\ HeavySil (HSIL) 31 Complicated RRD of inferior quadrants 1 month 2 months after tamponade removal Emulsification in AC (19%)\ HeavySil is a safe and effective tamponade\ et al. 2013 \[[@B89]\] Mild/severe intraocular inflammation (9.6%)\ agent for the treatment of complicated RD; no major complication in terms of corneal damage, hypotension or hypertension, or postoperative proliferation was observed Cataract formation (71%)\ Sticky oil formation (9.6%)\ Posterior synechiae (3.2%)\ Kocak and Koc 2013 \[[@B79]\] Densiron 68 31 Complicated RRD of inferior quadrants 3 months 6 months after tamponade removal Rise in IOP (25%)\ Densiron 68 does not have higher complication rates than conventional silicone oil Cataract progression (80%)\ AC inflammation with fibrin accumulation (3%)\ PVR (17%) and ERM (6%) -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- [^1]: Academic Editor: Kenneth Li

1. Introduction {#sec1} =============== Epidemiological studies have suggested a protective effect of obesity on postmenopausal bone loss \[[@B1]\]. In general, obesity in women is associated with lower risk of osteoporosis, which may be due to the beneficial effect of hyperestrogenemia \[[@B2]\]. The obese are characterized by calciotropic hormones disturbances, including lower 25-hydroxyvitamin-D~3~ (25-OH-D~3~) and higher parathyroid hormone (PTH) serum concentrations \[[@B3], [@B4]\], which increases along with the degree of obesity \[[@B5], [@B6]\]. Lower 25-OH-D~3~ levels in obese may be explained by decreased exposure to sunlight (limited mobility of some obese) \[[@B7]\], inhibited vitamin D hydroxylation in the liver \[[@B8]\], and probably excessive storage of vitamin D in adipose tissue \[[@B6]\]. The regulation of bone metabolism is complex including hormones and a long list of cytokines and growth factors. Recently, a new hormone involved in calcium-phosphate homeostasis was found---fibroblast growth factor-23 (FGF-23). FGF-23 is a 251-aminoacid residues protein that belongs to FGF family \[[@B9]\]. Fibroblast growth factors regulate the development of subcutaneous white adipose tissue (scWAT); however only five of them (FGF-1, 2, 7, 9 and 18) and not FGF-23 were identified in scWAT \[[@B10]\]. Osteocytes and osteoblast turned out to be the major physiological sources of FGF-23 \[[@B11]--[@B13]\]. FGF-23 regulates serum phosphate and 1,25(OH)~2~D~3~ levels by acting on the kidney and intestines. FGF-23 reduces serum phosphorus level by suppressing proximal tubular phosphorus reabsorption (similarly to PTH) and, in indirect way, intestinal phosphorus absorption \[[@B14]\]. FGF-23 also suppresses the expression of 1*α*-hydroxylase and enhances the expression of 24*α*-hydroxylase in the kidney, causing the reduction of serum 1,25(OH)~2~D~3~ level \[[@B14]\]. An association of plasma FGF-23 concentration with metabolic syndrome has already been proved \[[@B15]\]; however there are no data evaluating its association with bone metabolism in the obese. The aim of this study was to evaluate plasma FGF-23 concentrations and its relationship with calcium-phosphate homeostasis in obese perimenopausal women. 2. Materials and Methods {#sec2} ======================== Forty-eight obese perimenopausal women were enrolled into this study ([Table 1](#tab1){ref-type="table"}). Each woman was diagnosed with simple obesity without concomitant diseases. None of them underwent ovariectomy. All subjects were informed about the nature of the study and signed their informed consent forms. The exclusion criteria included current or recent (preceding 3 months) infections, pharmacotherapy affecting bone metabolism, cigarette smoking, and drinking of more than 2 alcohol beverages a week. The control group consisted of 29 nonobese healthy women ([Table 1](#tab1){ref-type="table"}). Body mass index (BMI) was calculated based on height and weight measurements. Blood sample was obtained from each subject between 8.00 and 9.00 am, after an overnight fast. Serum creatinine, total calcium, and phosphorus concentrations were measured using the commercially available test kits (Roche, Switzerland). Glomerular filtration rate (eGFR) was estimated using the CKD-EPI formula \[[@B16]\]. Plasma and serum samples for intact fibroblast growth factor 23 (iFGF-23), intact parathyroid hormone (iPTH), 25-OH-D~3~, CTX~1~, and osteocalcin estimations were stored at −80°C until the time of the assay. 2.1. Laboratory Measurements {#sec2.1} ---------------------------- Intact FGF-23 was measured by ELISA (Immutopics, San Clemente, USA). The Electrochemiluminiscent ImmunoAssay (Elecsys, Roche Diagnostics GmbH, Germany) was used for iPTH, osteocalcin, and C-terminal telopeptide of type I collagen (CTX1) assays. 25-(OH)-D~3~ was measured using RIA method (Bio Source-EUROPE S.A, Nivelles, Belgium). The concentrations of total calcium and inorganic phosphorus were assessed using spectrophotometer (Point Scientific Inc. MI, USA). Body composition was analyzed using bioimpedance method (Bodystat 1500, Bodystat Ltd., Great Britain). Dual energy X-ray absorptiometry (DXA method) of lumbar spine and femoral neck for BMD and BMC determinations was performed using Lunar Prodigy Advance apparatus. 2.2. Statistical Analysis {#sec2.2} ------------------------- All values are expressed as means ± 95% confidence intervals. Statistical analysis was performed using the STATISTICA 8.0 PL software. The results were analyzed using the Kolmogorov-Smirnoff test. As iFGF-23 values show marked squvenes, logarithmic transformation was performed for calculation of univariate correlations and multiple regression analyses. Mann-Whitney *U* test was used to compare obese to control group. Correlation coefficients were calculated according to Spearman. Separate multiple regression analyses were performed to explain variability of iFGF-23 concentrations, BMD of proximal femur, and lumbar spine. *P* values \<0.05 were considered to be statistically significant. 3. Results {#sec3} ========== Obese subjects were characterized by higher BMC and BMD both in the lumbar spine and femoral neck. There were higher serum levels of iPTH and iFGF-23 (by 42%), but lower of 25-OH-vitamin D3, phosphorus, and osteocalcin in obese women than in controls ([Table 1](#tab1){ref-type="table"}). In the combined group of obese and nonobese women iFGF-23 concentration correlated significantly with BMI (*R* = 0.292; *P* = 0.01) ([Figure 1](#fig1){ref-type="fig"}), body mass (*R* = 0.241; *P* = 0.035), and fat content (*R* = 0.259; *P* = 0.024). The relationship with waist circumference did not reach the level of statistical significance (*R* = 0.224; *P* = 0.073). There was a correlation between iFGF-23 concentration and proximal femur BMD (*R* = 0.321; *P* = 0.010) but not lumbar spine BMD (*R* = 0.223; *P* = 0.099). Moreover, a significant correlation between iFGF-23 and iPTH (*R* = 0.254; *P* = 0.026) was found. No correlation was demonstrated between serum phosphorus or eGFR and plasma iFGF-23 and between eGFR and serum phosphorus ([Figure 2](#fig2){ref-type="fig"}). In the group of obese women iFGF-23 concentration correlated significantly only with age (*R* = −0.346; *P* = 0.016) and proximal femur BMD (*R* = 0.285; *P* = 0.05). There was also a correlation with eGFR of borderline significance (*R* = 0.266; *P* = 0.07). No significant correlation was found between iFGF-23 plasma concentration and other study parameters in controls. Multiple regression analysis did not reveal any independent contribution of iFGF-23 concentrations to both proximal femur and lumbar spine BMD. The variability of proximal femur BMD was in 36.6% explained by age (*β* = −0.391; *P* \< 0.001) and BMI (*β* = 0.329; *P* = 0.006). Similarly the variability of lumbar spine BMD was in 17.9% explained by age *β* = −0.354; *P* = 0.007) and serum level of 25(OH)D~3~ (*β* = 0.271; *P* = 0.046). 4. Discussion {#sec4} ============= To the best of our knowledge, this is the first study that evaluated plasma iFGF-23 concentration with reference to bone metabolism in obesity. Our study revealed higher (by 42%) serum level of iFGF-23 in obese individuals compared to nonobese ones. A positive correlation between serum level of FGF-23 and BMI, waist circumference, waist-to-hip ratio, and triglycerides has been previously described by Mirza et al. \[[@B15]\]. The authors concluded that higher serum level of iFGF-23 may indicate an increased cardiovascular risk in elderly population; however its direct toxicity has never been shown. The authors did not explain the potential reason of FGF-23 increase in obese individuals. Moreover, the same authors in another study reported an association between elevated serum iFGF-23 concentration and the overall fracture risk in elderly men. This relation remained unaltered after adjustment for BMI, BMD, glomerular filtration rate, 25(OH)~2~D~3~, PTH, and other fracture risk factors \[[@B17]\]. To find out the explanation for our finding we focused on mechanisms potentially influencing serum FGF-23 levels. It has been previously shown that FGF-23 release is regulated by vitamin D~3~, phosphorus, and possibly PTH. Administration of 1,25(OH)~2~D~3~ and dietary phosphorus load was associated with increased FGF-23 levels \[[@B18], [@B19]\]. It has been demonstrated that PTH stimulates FGF-23 expression, whereas FGF-23 suppresses PTH secretion both *in vitro* and *in vivo*\[[@B20]--[@B22]\]. It was suggested that osteoblast stimulation by PTH directly increases skeletal FGF-23 release \[[@B19]\]. However the precise mechanism of this action remains unknown. We may suspect that higher level of PTH, previously reported in obese subjects, as well as higher dietary phosphorus may both account for serum FGF-23 increase. However, lower serum phosphorus concentrations in the obese do not support our hypothesis on a compensatory release of FGF-23 in the response to phosphorus accumulation. Perhaps more important than serum phosphorus level is dietary phorphorus intake. As it has been previously shown by Burnett et al., sustained increase in dietary phosphorus was associated with increasing iFGF-23 levels and declining calcitriol levels, while dietary phosphorus restriction reversed these trends \[[@B23]\]. The magnitude of change in serum FGF-23 level with dietary phosphorus depletion was smaller than with dietary phosphorus loading. It suggests that FGF-23 is a hormone that primarily promotes phosphorus wasting \[[@B22]\]. In our study we did not quantify phosphorus content in the diet or its urinary excretion; therefore we cannot analyze the influence of phosphorus consumption on iFGF-23 serum levels. We may suppose that elevated serum FGF-23, observed in obese subjects, may decrease phosphorus levels *via* a direct effect on proximal tubular phosphate reabsorption and/or intestinal phosphorus absorption in addition to glomerular hyperfiltration. The lack of correlation between serum phosphorus and iFGF-23 concentrations and simultaneously an almost significant inverse correlation between eGFR and serum phosphorus seem to support our hypothesis. It is well known that obesity is associated with both functional and structural changes within the kidney \[[@B24], [@B25]\]. Increased GFR in the obese may influence serum phosphorus and even secondarily suppress iFGF-23 secretion. In the study group of obese women, serum iFGF-23 concentration correlated significantly with age and proximal femur BMD only. Surprisingly, no correlation was found between iFGF-23 and studied markers of bone turnover, and the more so, because the study of Wang et al. showed that FGF-23 appeared to directly inhibit osteoblasts maturation and matrix mineralization \[[@B26]\]. In our study, the osteocalcin serum level was decreased in obese but no correlation with FGF-23 was found. We believe that this finding might be explained by decreased bone turnover in obese individuals, as already described in our previous study \[[@B27]\]. We suspect that the lack of significant correlations may have resulted from a limited number of study subjects. In line with this explanation we have demonstrated a significant, though weak, correlation between iPTH and iFGF-23 concentrations only in the combined group analysis. Recently Marsell et al. have found a positive, yet very weak (*R* = 0.04--0.07), correlation between serum FGF-23 and BMD in elderly men, which disappeared in multivariate correlation adjusting for body weight \[[@B28]\]. This correlation is quite surprising; however, it might be explained by the positive association between FGF-23 and body weight. The authors speculated, that osteocytes have a mechanosensory role, thus increased mechanical load (due to a larger body weight) may stimulate FGF-23 expression. Finally, the authors concluded that FGF-23 is a poor marker of BMD, due to the weak strength of correlation and its dependence on body weight. Our study has several limitations. The main one, previously mentioned, is the lack of measurement of daily phosphorus consumption and its urinary excretion. Moreover, metabolic and hormonal disturbances found in obese individuals, such as higher serum PTH, eGFR, and lower phosphorus, 25-OH-vitamin D~3~, and osteocalcin levels, may influence iFGF-23 secretion in obese individuals. Given these limitations, our study shows that obese individuals have higher level of iFGF-23 that correlates with proximal femur BMD. However, the mechanism by which iFGF-23 may affect bone metabolism in obese subjects remains unknown. Further studies are necessary to elucidate the cause and role of increased serum iFGF-23 level in the obese. 5. Conclusion {#sec5} ============= Serum levels of iFGF-23 are increased in obese perimenopausal women. This may partially explain lower phosphorus levels in obese subjects and seems not to reflect bone turnover. The study was funded by a Grant from the Ministry of Science and Higher Education (no. 404 207 135). {#fig1} {#fig2} ###### Characteristics of obese women and controls. Obese *N* = 48 Controls *N* = 29 ------------------------------ ---------------------- ---------------------- -------------- Age (yrs) 55.4 (53.6--57.1) 50.4 (46.0--54.7) *P* = 0.22 Body mass (kg) 94.5 (90.6--98.3) 62.4 (59.2--65.6) *P* \< 0.001 BMI (kg/m^2^) 36.5 (35.1--37.9) 23.5 (22.4--24.6) *P* \< 0.001 Waist circumference (cm) 106.7 (103.3--110.1) 82.3 (78.3--86.4) *P* \< 0.001 Fat tissue content (%) 49.6 (47.8--51.3) 35.2 (32.3--38.2) *P* \< 0.001 BMD L~2~--L~4~ (g/cm^2^) 1.196 (1.133--1.259) 1.035 (0.930--1.140) *P* = 0.023 BMD proximal femur (g/cm^2^) 1.026 (0.976--1.076) 0.858 (0.808--0.909) *P* \< 0.001 BMC L~2~--L~4~ (g/cm) 65.0 (60.7--69.3) 56.4 (51.1--61.7) *P* = 0.036 BMC proximal femur (g/cm) 4.87 (4.58--5.16) 4.23 (3.81--4.64) *P* = 0.030 Creatinine (mg/dL) 0.64 (0.61--0.67) 0.66 (0.61--0.71) *P* = 0.77 eGFR (mL/min) 97.4 (94.4--100.3) 99.5 (92.7--106.3) *P* = 0.46 Total calcium (mmol/L) 2.22 (2.18--2.25) 2.19 (2.13--2.25) *P* = 0.52 Phosphorus (mmol/L) 1.04 (1.00--1.08) 1.15 (1.09--1.21) *P* = 0.003 25(OH)D~3~ (ng/mL) 30.2 (26.2--34.2) 43.5 (34.8--52.1) *P* = 0.016 Osteocalcin (ng/mL) 19.1 (17.4--20.9) 24.0 (20.4--27.6) *P* = 0.026 iPTH (pg/mL) 47.3 (42.2--52.3) 35.4 (30.4--40.4) *P* = 0.002 CTX~1~ (ng/mL) 0.24 (0.21--0.27) 0.28 (0.21--0.35) *P* = 0.73 iFGF-23 (ng/mL) 9.14 (7.12--11.15) 6.44 (5.70--7.18) *P* = 0.005 PTH: parathormone; CTX~1~: C-terminal telopeptide of collagen type I; BMI: body mass, BMD: bone mineral density, BMC: bone mineral content, iFGF-23: intact fibroblast growth factor 23, eGFR: estimated glomerular filtration rate. [^1]: Academic Editor: Vin Tangpricha

Background {#Sec1} ========== Chronic obstructive pulmonary disease (COPD) is characterized by airflow limitation that is progressive and not fully reversible. Cigarette smoking is the main risk factor for COPD contributing to structural changes in the airways during COPD progression \[[@CR1]\]. Airway remodeling events in COPD include goblet cell hyperplasia and mucus hypersecretion, increased smooth muscle mass, and airway wall fibrosis. Thickening of the airway wall and peribronchiolar fibrosis play an important role in the development of symptoms associated with loss of lung function in COPD. An important characteristic of the fibrotic response is the increased deposition of extracellular matrix (ECM) proteins, particularly collagen types I and III, fibronectin and proteoglycans. One of the mechanisms now implicated in remodeling events associated with COPD pathogenesis is that fibroblastic cells arise from local conversion of epithelial cells by epithelial-mesenchymal transition (EMT) induced by cigarette smoking \[[@CR2]\]. During EMT, epithelial cells lose their typical cell-cell junctions and cell polarity and acquire a more motile, mesenchymal phenotype \[[@CR3]\]. The most common transformation occurs from epithelial cells to fibroblasts, which subsequently make the transition to myofibroblasts. EMT is mainly characterized by the loss of epithelial markers such as cytokeratins, tight junction proteins and E-cadherin, the acquisition of mesenchymal markers such as vimentin and N-cadherin, and increased expression of the transcription factors Snail, Twist and Zeb \[[@CR3]\]. EMT has been recently shown to be important in COPD \[[@CR2]\]. Sohal et al. reported that, in smokers and COPD patients, cells in the fragmented reticular basement membrane (Rbm) are double-positive for vimentin and cytokeratins, suggestive of EMT \[[@CR4]\]. The presence of cells expressing S100A4 and vimentin in the basal layer of the epithelium in smokers and COPD patients also indicates the contribution of smoke exposure to EMT process \[[@CR5]\]. Human bronchial epithelial cells (HBECs) and A549 cells undergo EMT *in vitro* after treatment with cigarette smoke condensate \[[@CR6]--[@CR8]\] further strengthening the rationale that EMT is a contributing factor in remodeling events of COPD. Interaction between lung structural cells, particularly epithelial cells and fibroblasts, may be key in driving the EMT process in COPD. Bronchial epithelial cells are the first anatomical barrier to noxious cigarette smoke particles and are involved in the initiation of airway remodeling through the production of proinflammatory mediators, ECM protein, growth factors and matrix metalloproteinases \[[@CR9]\]. Supernatants from bronchial epithelial cell cultures contain factors which both stimulate and inhibit fibroblast proliferation \[[@CR10]\]. Fibroblasts are also important in regulating ECM turnover and epithelial cell differentiation via growth factor secretion and mesenchymal-epithelial cell interactions \[[@CR11]\]. However, the interactions of fibroblasts and epithelial cells and the participation of fibroblasts in the EMT process remain poorly understood in COPD. In this study, we hypothesized that EMT is active in bronchial epithelial cells of patients with COPD, and that mediators secreted by COPD lung fibroblasts could induce EMT. We therefore investigated the EMT process in bronchial epithelial cells of COPD patients, together with the effect of mediators secreted by human lung fibroblasts (HLF) from normal and COPD subjects on the expression of epithelial and mesenchymal markers in human bronchial epithelial (HBE) cells. Methods {#Sec2} ======= Epithelial cell culture {#Sec3} ----------------------- Primary human bronchial epithelial cells from normal subjects (NHBE) and COPD patients (DHBE) were purchased from Lonza (Walkersville, MD) and were maintained in serum-free bronchial epithelial cell growth medium (BEGM, Lonza) supplemented with a bullet kit containing bovine pituitary extract, insulin, hydrocortisone, gentamicin/amphotericin, retinoic acid, transferrin, epinephrine and human epithelial growth factor (hEGF) (Lonza). NHBE and DHBE cells were used before passage 6. Fibroblast cell culture and collection of conditioned media (CM) {#Sec4} ---------------------------------------------------------------- Lung tissue was obtained from individuals undergoing lung resection surgery for suspected lung cancer at McMaster University. Recruited individuals included those with COPD as well as never-smokers without COPD (controls). This study was approved by the Research Ethics Board of St Joseph's Healthcare Hamilton and all patients gave written informed consent. Primary lung fibroblasts were cultured as previously described \[[@CR12], [@CR13]\] from parenchymal lung tissue. Only tissue from cancer-free regions was used for the derivation of fibroblasts. Prior to experimentation, fibroblasts were characterized based on morphology, vimentin expression and absence of cytokeratin (epithelial cell marker), desmin (muscle cell marker) and α-smooth muscle actin (α-SMA; myofibroblast marker) \[[@CR12], [@CR13]\]. All fibroblasts used in this study had a typical fibroblast morphology (flat, elongate with oval nuclei) and expressed vimentin; no staining was observed for cytokeratin or desmin, Following characterization, cells were expanded and either frozen in liquid nitrogen or maintained in culture. For experimentation, primary human lung fibroblasts from normal subjects (NHLF) and COPD patients (DHLF) were maintained in minimum essential medium (MEM) supplemented with 2 mM L-glutamine (Life Technologies, Burlington, ON), 10 % fetal bovine serum, and antibiotics (50 μg/ml streptomycin and 50 U/ml penicillin). Cells were maintained at 37 °C and incubated in humidified 5 % CO~2~ atmosphere. Fibroblasts were used before passage 14. For experimentation, fibroblasts (3 × 10^5^ cells per flask) were plated in T-75 flasks in 10 % FBS MEM and allowed to reach 80 % confluence. The cultures were then rinsed twice and cultured in serum-free growth medium consisted of BEGM and Dulbecco's Modified Eagle's Medium (DMEM) and CM was collected 48 h later. Treatment of epithelial cells with fibroblast-derived CM {#Sec5} -------------------------------------------------------- NHBE and DHBE were grown at 37 °C in a humidified 5 % CO~2~ atmosphere. Prior to the experiments, the cells were starved in 1 % BEGM for 24 h and then stimulated with CM derived from NHLF or DHLF for a period of 3 (RNA analysis) or 5 (protein analysis) days. The cells were then harvested for further analysis. Protein quantification and immunoblotting {#Sec6} ----------------------------------------- Primary bronchial epithelial cells were lysed in 80 μL of lysis buffer (50 mM Tris--HCl pH 7.5, 1 mM EGTA, 1 mM EDTA, 1 % (v/v) Triton X-100, 1 mM sodium orthovanadate, 5 mM sodium pyrophosphate, 50 mM sodium fluoride, 0.27 M sucrose, 5 mM sodium pyrophosphate decahydrate and protease inhibitors). Protein concentrations were quantified using the BCA Protein Assay Kit (ThermoScientific, Rockford, IL) according to the manufacturer's instructions. Forty micrograms of protein were boiled and separated on a 10 % Pro-Pure Next Gel with Pro-Pure Running Buffer (Amresco, Solon, OH). After transferring proteins to nitrocellulose, membranes were blocked for 1 h at room temperature in Odyssey Blocking Buffer (Li-Cor Biosciences, Lincoln, NE). Blots were then incubated with a mouse anti-human vimentin antibody (1:200, sc-6260, Santa Cruz Biotechnology, Santa Cruz, CA), a mouse anti-human E-cadherin antibody (1:600, ab1416, Abcam, Cambridge, MA), a rabbit anti-human N-cadherin antibody (1:1000, ab76057, Abcam) or a mouse anti-human GAPDH antibody (1:1500, MAB374, Millipore, Billerica, MA) overnight at 4 °C. Goat anti-mouse IgG (1:15,000, \#35518, DyLight™680, Thermo Scientific) antibody or goat anti-rabbit IgG (1:15,000, \#35571, DyLight™800, Thermo Scientific) antibody was applied for 1 h in the dark at room temperature (1:15,000). The signal was detected and quantified using a LI-COR Odyssey imaging system (LI-COR Biosciences). All samples were normalized to GAPDH and expressed as a ratio relative to the control sample. Measurement of hepatocyte growth factor (HGF) {#Sec7} --------------------------------------------- HGF concentration in cell cultured conditioned medium was measured with human HGF ELISA kit (R&D Systems, Inc. Minneapolis, MN, USA) according to the manufacturer's protocol. HGF secretion by fibroblasts was expressed in picogram/ml. Semi-quantitative RT-PCR {#Sec8} ------------------------ Total RNA was isolated from cultured primary bronchial epithelial cells and purified using the RNeasy Mini Kit (Qiagen, Toronto, Canada), supplemented with the RNase-Free DNase Set (Qiagen). cDNA was obtained using the SuperScript^TM^ III Reverse Transcriptase (Life Technologies), and the absence of DNA contamination was verified by excluding the reverse transcriptase from subsequent PCR reactions. cDNA was subjected to PCR using the Power SYBR Green PCR Master Mix (Applied Biosystems, Foster City, CA) to amplify human transcripts of E-cadherin, N-cadherin, vimentin, Snail1, Snail2, Twist1, Twist2, Zeb1, Zeb2 and GAPDH using primers from Life Technologies. Each set of primers are shown in Table [1](#Tab1){ref-type="table"}. Each PCR reaction was carried out as follows: 15 min at 95 °C, 15 s at 94 °C, 30 s at 60 °C, and 30 s at 72 °C. Each cycle was repeated 40 times following the manufacturer's recommendations using a 7500 Fast Real-Time PCR System (Applied Biosystems) thermal cycler. Based on the comparative Ct method, gene expression levels were calculated and GAPDH was used as the housekeeping gene. Untreated control samples for each cell line were set to 100 % and the fold change in expression following treatment is represented as mean ± standard error of the mean. Each condition was assessed based on duplicates with *n* = 3 cell lines per group.Table 1Primers for quantitative real-time RT-PCR analyses of gene transcript expressionGeneForward primerReverse primerSize (bp)E-cadherin5'-gccgagagctacacgttca-3'5'-gaccggtgcaatcttcaaa-3'88N-cadherin5'-ctccatgtgccggatagc-3'5'-cgatttcaccagaagcctctac-3'92Vimentin5'-gtttcccctaaaccgctagg-3'5'-agcgagagtggcagagga-3'68Snail5'-gctgcaggactctaatccaga-3'5'-atctccggaggtgggatg-3'84Slug5'-tggttgcttcaaggacacat-3'5'-gttgcagtgagggcaagaa-3'66Twist15'-aaggcatcactatggactttctct-3'5'-gccagtttgatcccagtatttt-3'96Twist25'-tctgaaacctgaacaacctcag-3'5'-ctgctgtcccttctctcgac70Zeb15'-gctaagaactgctgggaggat-3'5'-atcctgcttcatctgcctga-3'79-82Zeb25'-aagccagggacagatcagc-3'5'-ccacactctgtgcatttgaact-3'74HGF5'-cagagggacaaaggaaaagaa-3'5'-gcaagtgaatggaagtccttta-3'167GAPDH5'-agtcaacggatttggtcgtatt-3'5'-atgggtggaatcatattggaac-3'139 Immunohistochemistry {#Sec9} -------------------- Paraffin-embedded lung sections were stained with specific antibodies for E-cadherin (a mouse anti-human E-cadherin antibody, 1:300 dilution, ab1416, Abcam, Cambridge, MA) or snail1 (1:250 dilution, a rabbit anti-Snail + Slug, ab85936, Abcam, Cambridge, MA) or N-cadherin (a rabbit polyclonal anti-human N-cadherin antibody, 1:300 dilution, ab18203, Abcam, Cambridge, MA). For negative controls, isotype-matched control antibody was substituted for primary antibody (data not shown). Lung sections were incubated with peroxidase-antiperoxidase complex and developed with diaminobenzidine (DAB) after blocking endogenous peroxidase activity. All sections were counterstained with Gill II hematoxylin, inspected using an Olympus light microscope and digital images were saved. Dark brown color indicates positive immunostaining for a particular antigen expression. Statistical analysis {#Sec10} -------------------- For statistical analyses between two groups, a *t*-test was used. Comparisons among more than two groups were performed by ANOVA, followed by a Tukey post hoc test. A p-value of \< 0 .05 was considered to be statistically significant. Data are expressed as mean ± standard error of the mean. Results {#Sec11} ======= Basal expression of EMT markers and EMT-related transcription factors are increased in COPD bronchial epithelial cells and COPD lung tissues {#Sec12} -------------------------------------------------------------------------------------------------------------------------------------------- To determine the basal expression of EMT-related markers in HBE cells, we examined the mRNA expression of epithelial cells derived from normal subjects (NHBE) and COPD patients (DHBE). While basal expression of E-cadherin in DHBE cells was not different from that of NHBE cells (Fig. [1a](#Fig1){ref-type="fig"}), DHBE cells showed a significant increase in mRNA expression of typical mesenchymal markers including N-cadherin and vimentin (Fig. [1b,c](#Fig1){ref-type="fig"}). The ratio of E-cadherin to N-cadherin, an index of the epithelial phenotype, (Fig. [1d](#Fig1){ref-type="fig"}) was lower (but not statistically significant) in DHBE cells than in NHBE cells, suggesting the potential for EMT in COPD epithelial cells.Fig. 1Basal mRNA expression of mesenchymal markers was increased in DHBE cells. Basal mRNA expression of EMT-related markers in non-COPD (NHBE, open bar, *n* = 3) and COPD (DHBE, closed bar, *n* = 3) human bronchial epithelial cells were analyzed by real-time RT-PCR. Results were normalized to NHBE and expressed as mean (SE) of three independent experiments. Data were analysed by Student *t* test, \**p* \< 0.05, compared with NHBE. (**a**) E-cadherin (**b**) N-cadherin (**c**) vimentin (**d**) Ratio of E-/N-cadherin. ns: not significant We next measured changes in the expression of EMT-related transcription factors in NHBE and DHBE cells, including Snail, Snai2 (also called Slug), Twist1, Twist2, and Zeb1. There were significant increases in the expression of these transcription factors in DHBE cells compared to NHBE cells (Fig. [2](#Fig2){ref-type="fig"}), findings which are consistent with the hypothesis that the EMT process is more evident in DHBE cells. In contrast, Zeb2 expression was decreased in DHBE cells compared to NHBE cells (Fig. [2f](#Fig2){ref-type="fig"}).Fig. 2Basal mRNA expression of EMT-related transcription factors, except Zeb2, was higher in DHBE cells. Basal mRNA expression of EMT-related transcription factors in non-COPD (NHBE, open bar, *n* = 3) and COPD (DHBE, closed bar, *n* = 3) human bronchial epithelial cells were analyzed by real-time RT-PCR. Results were normalized to NHBE and expressed as mean (SE) of three independent experiments. Data were analysed by Student *t* test, \**p* \< 0.05, \*\**p* \< 0.01, \*\*\**p* \< 0.0001, compared with NHBE. (**a**) Snail (**b**) Slug (**c**) Twist1 (**d**) Twist2 (**e**) Zeb1 and (**f**) Zeb2 To determine if the results presented above in bronchial epithelial cells of control (non-COPD) subjects and COPD subjects is also reflective of changes within the lung, we evaluated the expression of these transcription factors in RNA derived from resected lung tissue. As shown in Fig. [3](#Fig3){ref-type="fig"}, there was a significant increase in the mRNA expression of the EMT-related transcription factors Snai1 (Fig. [3a](#Fig3){ref-type="fig"}), Twist1 (Fig. [3c](#Fig3){ref-type="fig"}) and Twist2 (Fig. [3d](#Fig3){ref-type="fig"}) in tissue from COPD patients compared to tissue from normal subjects. However, there was no difference in the expression level of Snai2 mRNA (Fig. [3b](#Fig3){ref-type="fig"}) between non-COPD and COPD lung.Fig. 3mRNA expression analysis of EMT-related transcription factors in non-COPD and COPD lungs. Basal mRNA expression of EMT-related transcription factors in non-COPD (white bars, *n* = 4) and COPD (black bars, *n* = 4) subjects. Lung tissues from non-COPD and COPD subjects were analyzed by real-time RT-PCR. Results were normalized to housekeeping gene (GAPDH mRNA) and expressed as mean (SE) of four independent experiments. Data were analysed by unpaired Student's *t* test, (**a**) Snail (**b**) Snai2 (**c**) Twist1 (**d**) Twist2. Non-COPD subjects are non smokers We also evaluated EMT markers in histological sections of non-COPD (Control) (Fig. [4a,c](#Fig4){ref-type="fig"}) and COPD (Fig. [4b,d](#Fig4){ref-type="fig"}) lungs. In agreement with mRNA expression, we observed enhanced Snail1 (Fig. [4d](#Fig4){ref-type="fig"}) and N-cadherin (Fig. [4f](#Fig4){ref-type="fig"}) protein--positive cells in COPD lungs whereas E-cadherin was markedly decreased in COPD (Fig. [4b](#Fig4){ref-type="fig"}) compared to non-COPD lungs. Taken together these results corroborate the data obtained from the cell culture model, and support a role for EMT in COPD pathogenesis.Fig. 4Snail1 and E-cadherin protein expression in non-COPD and COPD lungs. Immunohistochemical localization of EMT markers in the airway epithelium of non-COPD (**a**, **c**, **e**) and COPD (**b**, **d**, **f**) lungs (*n* = 3). Immunostaining for E-cadherin (**a**, **b**) and Snail1 transcription factor (**c**, **d**) and N-cadherin (**e**, **f**). The dark brown color is indicative of strong expression (denoted by red arrows) of E-cadherin in non-COPD (**a**), Snail1 in COPD (**d**) and N-cadherin in COPD (**f**) lung. Negative control was stained with isotype-matched control antibody (data not shown). Non-COPD subjects are non smokers. aw: airway lumen Basal expression of HGF in non-COPD and COPD lung fibroblasts {#Sec13} ------------------------------------------------------------- Using RT-PCR, the mRNA expression level of HGF was measured in normal and COPD lung fibroblasts. No significant difference in mRNA expression was found (Fig. [5a](#Fig5){ref-type="fig"}). Nor did we find a significant differences between normal and COPD lung fibroblasts in HGF protein secretion in the culture supernatants, although there was a tendency to an increase in COPD fibroblasts (Fig. [5b](#Fig5){ref-type="fig"}).Fig. 5Basal expression of HGF in non-COPD (NHPF) and COPD (CHPF) lung fibroblasts. mRNA expression (**a**) and the protein secretion (**b**) of hepatocyte growth factor in conditioned medium of normal (NHPF) and COPD (CHPF) lung fibroblasts. Results were expressed as mean (SE) of six independent experiments. Data were analysed by unpaired Student's *t* test, Non-COPD subjects are non smokers CM derived from NHLF increased mRNA expression of vimentin in NHBE and DHBE, and that of N-cadherin in DHBE {#Sec14} ----------------------------------------------------------------------------------------------------------- Although the above findings indicate the evidence of EMT in COPD epithelial cells and lungs, it is important to further dissect the contribution of fibroblast-epithelial cell interactions in the EMT process. To this end, we investigated whether soluble mediators secreted by lung fibroblasts would modulate the phenotypic properties of bronchial epithelial cells derived from normal (non-COPD) subjects and COPD patients. NHBE cells and DHBE cells were treated with or without CM derived from NHLF or DHLF for 3 days. In both NHBE and DHBE cells following the 3-day stimulation with CM derived from NHLF, the expression of vimentin mRNA was significantly increased compared to control, and the induction was higher in NHBE cells than in DHBE cells (Fig. [6a](#Fig6){ref-type="fig"}). Only in DHBE cells did stimulation with CM derived from NHLF increase N-cadherin mRNA expression (Fig. [6b](#Fig6){ref-type="fig"}). Stimulation of HBE cells with CM derived from DHLF did not affect the mRNA expression of N-cadherin or vimentin. Also, there was no significant difference in the expression level of E-cadherin mRNA (Fig. [6c](#Fig6){ref-type="fig"}) or E-cadherin/N-cadherin ratio (Fig. [6d](#Fig6){ref-type="fig"}) in epithelial cells stimulated with CM derived from NHLF or DHLF. These results suggest that mediators secreted from lung fibroblasts of non-COPD subjects influence EMT events in epithelial cells.Fig. 6mRNA expression of EMT markers in HBE cells after 3-day treatment with CM from HLF. HBE cells from non-COPD (N) or COPD patients (D) were stimulated with 100 % CM (black bars), or without CM (white bars) collected from human lung fibroblasts (HLF) derived from non-smokers (N) or COPD subjects (D) for 3 days and total RNA was extracted. The mRNA expression of EMT markers were determined by real-time RT-PCR and normalized to GAPDH mRNA expression in the same origin of HBE cells. (**a**) E-cadherin, (**b**) N-cadherin, (**c**) vimentin, and (**d**) the ratio of E-/N-cadherin were calculated. Data were analysed and expressed as mean (SE) of from six to nine independent experiments by one-way ANOVA with Tukey's multiple comparison test, \**p* \< 0.05, \*\**p* \< 0.01, compared between the indicated groups CM from DHLF increased protein expression of vimentin as well as E-cadherin in NHBE cells {#Sec15} ----------------------------------------------------------------------------------------- Based on the mRNA results above, we next investigated the effect of CM on protein expression of EMT markers. Treatment with CM from DHLF significantly increased the protein expression of E-cadherin and N-cadherin in NHBE cells compared to NHLF (Fig. [7a,b](#Fig7){ref-type="fig"}). Vimentin protein expression in NHBE cells was increased significantly by CM from DHLF (Fig. [6c](#Fig6){ref-type="fig"}), whereas CM from NHLF had no effect. The above data (Figs. [6](#Fig6){ref-type="fig"} and [7](#Fig7){ref-type="fig"}) suggest a divergence between protein expression of EMT markers and the corresponding mRNA expression stimulated by NHLF and DHLF CM in NHBE and DHBE cells.Fig. 7Effects of CM from HLF on protein expression of EMT markers in HBE cells. HBE cells from normal subjects (N) or COPD patients (D) were lysed after stimulation with or without 100 % CM collected from human lung fibroblasts of N or D for 5 days. Total protein was prepared analysed by western blotting, using 40 μg of protein per lane. The membranes were probed with anti-E-cadherin, anti-N-cadherin, anti-vimentin, or anti-GAPDH antibodies. The intensities of the bands for target protein were normalized to the corresponding GAPDH bands for the each treated cells. (**a**) E-cadherin, (**b**) N-cadherin, (**c**) vimentin, and (**d**) the ratio of E-/N-cadherin were calculated. Data were analysed and expressed as mean (SE) of from seven to ten independent experiments by one-way ANOVA with Tukey's multiple comparison test, \**p* \< 0.05, \*\**p* \< 0.01, compared between the indicated groups CM from NHLF increased the mRNA expression of some EMT-related transcription factors in HBE cells {#Sec16} ------------------------------------------------------------------------------------------------- Given the effect of CM on EMT markers in epithelial cells (Figs. [6](#Fig6){ref-type="fig"} and [7](#Fig7){ref-type="fig"}), we next determined the changes in the expression of EMT-related transcription factors. CM from NHLF or DHLF did not alter the mRNA expression of Snai1 transcription factor in epithelial cells (Fig. [8a](#Fig8){ref-type="fig"}). While CM obtained from NHLF increased the mRNA expression of Snai2 in DHBE cells (Fig. [8b](#Fig8){ref-type="fig"}), CM derived from DHLF had no effect. CM from NHLF, but not from DHLF, significantly increased the mRNA expression of Twist1 and Twist2 in NHBE cells (Fig. [8c,d](#Fig8){ref-type="fig"}). Collectively, these data may reflect differences in functional phenotypes related to repair and remodeling of fibroblasts derived from non-COPD and COPD subjects in driving the EMT process.Fig. 8mRNA expression of EMT-related transcription factors in HBE cells after 3-day treatment with CM from HLF. HBE cells from normal subjects (N) or COPD patients (D) were stimulated with or without 100 % CM collected from human lung fibroblasts of N or D for 3 days and total RNA was extracted. The mRNA expressions of EMT-related transcription factors were determined by real-time RT-PCR and normalized to GAPDH mRNA expression in the same origin of HBE cells. (**a**) Snail, (**b**) Slug, (**c**) Twist1, (**d**) Twist2, and (**e**) Zeb1 were calculated. Data were analysed and expressed as mean (SE) of from six to nine independent experiments by one-way ANOVA with Tukey's multiple comparison test, \**p* \< 0.05, \*\**p* \< 0.01, compared between the indicated groups Discussion {#Sec17} ========== COPD is a complex disease caused primarily by cigarette smoke exposure. The pathogenic processes associated with COPD include α~1~-anti-trypsin deficiency, oxidative stress, chronic inflammation, aberrant repair processes and small airway fibrosis. Recently, EMT has been considered as another causative mechanism implicated in the small airway fibrosis in COPD \[[@CR4], [@CR5], [@CR14], [@CR15]\]. In this regard, EMT may contribute to the myofibroblast accumulation associated with fibrotic lesions. Here we demonstrate that under basal culture conditions mRNA expression of mesenchymal markers and EMT-related transcriptional factors are upregulated in DHBE cells compared to NHBE cells. Additionally, our studies demonstrate increased expression of EMT markers and EMT-related transcription factors in COPD lungs compared to non-COPD tissue. Milara et al. \[[@CR14]\] recently reported that mesenchymal markers such as vimentin and α-smooth muscle actin (α-SMA) were upregulated in HBEC from smokers and patients with COPD compared to those from non-smokers while epithelial markers such as E-cadherin and ZO-1 were downregulated, findings consistent with our study. To the best of our knowledge, we are the first to show that mRNA expression of EMT-related transcription factors are upregulated in DHBE cells and COPD lungs compared to NHBE cells and non-COPD tissue. These results suggest that DHBE cells have already acquired the phenotype consistent with the EMT process. Furthermore, elevation of EMT markers and transcription factors in DHBE cells relative to normal epithelial cells suggest that COPD epithelial cells are capable of maintaining the EMT phenotype in culture conditions. These data indicate that acquired changes of epithelial phenotypes can contribute to the pathogenesis of COPD. A defining feature of EMT is a reduction in E-cadherin levels and a concomitant production of N-cadherin. While our immunostaining data revealed evidence of a decrease in E-cadherin and an increase in N-cadherin expression in COPD patients, we did not detect a significant change in E-cadherin gene expression in COPD epithelial cells or in the conditioned medium treated epithelial cells. Based on these observations it could be suggested that epithelial cells in these COPD subjects are in the process of undergoing an EMT program or have attained 'hybrid states', with features of both epithelial and mesenchymal cells. Our findings are in agreement with the studies of Sohal et al. \[[@CR4]\], who reported that cells in the reticular basement membrane were positive for mesenchymal markers and an epithelial marker in COPD subjects. It has been claimed that cells expressing mesenchymal and epithelial features are migratory, invade the basement membrane, accumulate in the interstitium, eventually lose their epithelial properties and attain a complete fibroblastic phenotype \[[@CR3]\]. Mechanisms which drive the EMT process in COPD are largely unknown. Epithelial cells are at the interface with the external environment, and interact extensively with both immune and structural cells such as fibroblasts. Fibroblasts, in addition to being the key cell involved in structure and repair, have enormous synthetic capabilities. COPD fibroblasts exhibit alterations in proliferation, repair and inflammatory cytokine secretion, making it possible that factors released by fibroblasts can influence overlying epithelial cells, thereby altering the EMT process \[[@CR16]\]. To address this possibility, we utilized cell culture media (CM) collected from normal and COPD fibroblasts and exposed normal and COPD epithelial cells to this media. Our results demonstrate that CM from NHLF increases mesenchymal markers at mRNA levels in NHBE cells. While the identity of the factor secreted by the fibroblasts is not yet known, a candidate protein is hepatocyte growth factor (HGF). HGF secretion by pulmonary fibroblasts is crucial to an effective epithelial repair after lung injury and can drive the EMT process. HGF secretion by human fibroblasts from emphysematous lung is lower even after stimulation with interleukin-1β or prostaglandin E~2~ compared to control fibroblasts \[[@CR17]\]. Thus, low HGF production in COPD fibroblasts, as reported earlier by Plantier et al. \[[@CR17]\], may explain why these cells are less capable of inducing EMT compared to normal fibroblasts. Our results showed no significant change in the expression of HGF between normal (NHLF) and diseased lung fibroblasts (DHLF), although the protein expression data showed a tendency to an increase. The differences in HGF expression between our study and that of Plantier et al. \[[@CR17]\] might be due to differences in disease severity. Whereas fibroblasts used in our study were obtained from patients with mild to moderate COPD, cells used in the study of Plantier et al. \[[@CR17]\] were from patients with more severe disease who were undergoing lung volume reduction surgery. As fibroblasts secrete soluble factors which act in a paracrine manner targeting epithelial repair and maintenance, alterations in HGF production due to cigarette smoking may well have implications for wound repair functions of fibroblasts in COPD. Multiple transcription factors, including Twist, Snail, Slug, and Zeb orchestrate the EMT and the migratory processes during embryogenesis \[[@CR18]--[@CR22]\]. Of these, Twist is particularly important, as it can bind to the E-cadherin promoter and represses gene transcription of E-cadherin, a protein required in cell-cell contact and one that is typically down-modulated during EMT. In the lung, Pozharskay et al. first demonstrated the upregulation of Twist protein expression in a murine model of virus-induced lung fibrosis and in idiopathic pulmonary fibrosis (IPF) lung tissue \[[@CR23]\]. However, there is no information on the expression of Twist in COPD or influence by epithelial cells. Therefore, we evaluated whether CM could alter the expression of EMT-associated transcription factors in COPD cells. Our data show that only CM from NHLF significantly affected Twist1 and Twist2 expression in NHBE cells. While the functional significance of this observation in the context of COPD is not clear, it is possible that Twist2 contributes to the maintenance of epithelial cell characteristics \[[@CR21]\] or participates as an anti-apoptotic factor \[[@CR22]\]. Further studies are needed to characterize the functional significance of Twist in COPD. Regardless, our studies suggest that CM from NHLF-induced Twist activation might represent an important mechanism for EMT in the normal airway epithelial repair. Interestingly, NHLF-derived CM was also more effective than DHLF-derived CM at inducing Slug mRNA in DHBE cells. Concomitant with the induction of EMT, Slug has a critical role in re-epithelialization of cutaneous wounds in mice, a process that involves a transient loss of epithelial polarity without full acquisition of mesenchymal characteristics. Such partial EMT processes occur during some types of tubulogenesis (e.g., in mammary gland development). Slug is also required for survival during partial EMT \[[@CR24]\]. In our study, CM from normal HLF increased the expression of Slug in DHBE cells, which may mean that normal lung fibroblasts are capable of promoting cell survival and thus contribute to the normal repair of the injured DHBE cells. We also examined the expression of Zeb1 and Zeb2 mRNA in DHBE cells. While Zeb1 expression was upregulated in DHBE cells, Zeb2 mRNA expression was decreased in DHBE compared to NHBE. It is known that E-cadherin expression is regulated by a family of transcriptional factors including Snail, Slug, Twist, Zeb1, and Zeb2, and these proteins are, in turn, regulated by growth factors such as PDGF, TGF-β, and Wnt proteins. Both Zebs suppress E-cadherin expression. However, it is not clear if Zeb1 has a predominant role over Zeb2 or *vice versa* in suppressing E-cadherin expression during EMT. It is likely that both Zebs contribute to EMT, along with involvement of other EMT-related transcription factors. Further studies are needed to address their respective roles using molecular strategies designed to target simultaneously both Zebs and/or other transcription factors. Our data also raise the question whether different repressors participate in silencing of E-cadherin at defined stages of EMT program in COPD. Together, our results illustrate the complexity of E-cadherin expression and demonstrate that a mechanism that fully defines its expression in COPD has yet to be described. We also found that protein expression of both epithelial and mesenchymal markers were increased only in NHBE cells treated with CM from DHLF. A possible explanation is that because the basal expression of EMT markers is higher in COPD epithelial cells, these cells are refractory to subsequent treatment with CM and cannot further upregulate the expression of EMT markers. The increased levels of EMT markers in diseased epithelial cells derived from COPD subjects may be due to cigarette smoking, which has been shown to promote EMT in epithelial cells and emphasizes the importance of environmental factor in the pathogenesis of chronic inflammatory diseases, including COPD \[[@CR5], [@CR6]\]. The above changes in EMT protein markers seem to be contradictory with data obtained for mRNA expression where NHLF-derived CM upregulated EMT-associated transcription factors in DHBE cells. The molecular mechanisms responsible for the differential actions of CM from normal vs. COPD lung fibroblasts on protein and mRNA expression are not clearly understood. In summary, our work reveals increases in EMT-associated transcription factors and mesenchymal markers in bronchial epithelial cells derived from COPD subjects, demonstrating that these cells are undergoing EMT process. The fact that only normal bronchial epithelial cells responded to cell culture conditioned-medium derived from COPD lung fibroblasts suggest that environmental factors (e.g., cigarette smoke) may play an important role in the development of EMT. Although recent reports have indicated that human bronchial epithelial cells from COPD patients compared to non-smokers, undergo EMT, the hypothesis that epithelial trans-differentiation during tissue repair in COPD may be induced through the communication with fibroblasts, independent of inflammatory cells, was not considered in previous studies \[[@CR4], [@CR5], [@CR14]\]. The epithelium-fibroblast interaction is important in normal lung development, where a fibroblast-secreted mediator affected epithelial maturation and function \[[@CR25]\]. During an injury-repair response secreted mediators are required for epithelial regeneration. In the case of anomalous repair the balance is tipped in the direction of a fibrotic growth, where interruption of the normal intercellular relationships, principally between epithelial cells and fibroblasts, occurs. Whether these intercellular contacts transfer inductive messages associated with switching on EMT is unknown. Here, we investigated whether fibroblasts in COPD subjects contribute to the abnormal function of epithelial cells, and the effect of mediators secreted by human lung fibroblasts from normal and COPD subjects on phenotypic changes in normal and diseased bronchial epithelial cells. We demonstrate that cell cultured conditioned medium from COPD fibroblasts (DHLF-CM) promoted EMT in normal human bronchial epithelial cells (NHBE), suggesting that interactions between fibroblast and epithelial cells may play an important role in the EMT process in COPD. Whereas mediators secreted by infiltrating inflammatory cells, such as lymphocytes, platelets or macrophages, upregulate EMT-related transcription factors \[[@CR25]--[@CR29]\], and may be critical to establishment of repair response by epithelial cells and fibroblasts, our data are novel as they demonstrate that the interrelationship between epithelial and fibroblast cells is sufficient to induce EMT process in the absence of blood-borne cells or their secreted mediators. Our findings also indicate the possibility that fibroblasts are involved in the control of EMT via soluble mediators, a finding which merits further investigation. The finding that structural cells may be involved directly or indirectly in promoting EMT further establishes the importance of these cells in COPD pathogenesis. CM : Conditioned medium COPD : Chronic obstructive pulmonary disease ECM : Extracellular matrix EMT : Epithelial-mesenchymal transition HBE : Human bronchial epithelial cell HLF : Human lung fibroblast NHBE : Normal human bronchial epithelial cells DHBE : COPD bronchial epithelial cells NHLF : Normal hulan lung fibroblasts DHLF : COPD lung fibroblasts HGF : Hepatocyte growth factor. **Competing interests** The authors declare that they have no competing interests. **Authors' contributions** MN, NV, ML, QH: Concept and design of experiments. NV, ML, CB, DE and QH helped with study design and draft the manuscript. MN, NV, KD, ML and CB: Data analysis and manuscript preparation. AM performed RT-PCR studies. PN supplied and characterized patient specimens. SH and TL helped with study design and statistical analysis. All authors read and approved the final manuscript. This study was supported by the Canadian Institutes of Health Research, the Costello Memorial Fund, the McGill University Health Centre--Research Institute and the Richard and Edith Strauss foundation. We would like to thank Fazila Chouiali for her help in immunostaining.

\section{Introduction} Drinfeld - Kohno (DK) theorem \cite{Dr1} - \cite{Ko} states braided tensor equivalence between seemingly different categories of modules: on the one hand, quasitriangular quasi-Hopf universal enveloping algebra modules associated to a simple Lie algebra $\mathfrak{g}$ with associator and braiding defined through Knizhnik-Zamolodchikov (KZ) equation with quantum deformation parameter $h$, and on the other hand -- of modules of quasitriangular Hopf $h$-quantized universal enveloping algebra associated to $\mathfrak{g}$. By this equivalence the quantization parameter $h$ of the latter algebra corresponds in the former to deformation parameter of associator which arises as a monodromy of KZ equation solutions associated with the chosen representations category. This theorem was proved by Drinfeld using series expansion in $h$ around zero, and is valid for generic values of this parameter, with excluded specific rational values. Later on, in the seminal series of papers \cite{KL1} - \cite{KL4}, this equivalence were addressed more generally by Kazhdan and Lusztig. In \ \cite{KL1}, \cit {KL2} they showed that when $l\notin \mathbb{Q} $ or when $l\in \mathbb{Q} $ but $l<-h^{\vee }$, where $h^{\vee }$ is the dual Coxeter number of \mathfrak{g}$, a certain category of affine level $l$ Lie algebra $\widehat \mathfrak{g}}$ modules has a natural braided tensor category structure, and they proved rigidity for most of these tensor categories in \cite{KL4}. Kazhdan - Lusztig construction was then extended by Finkelberg to rational positive level categories of affine modules in \cite{Fin}. In this affine algebraic context the KZ equation appears naturally, and Kazhdan and Lusztig have proved \cite{KL3}, \cite{KL4} braided tensor equivalence of their category at level $l$ to the category of finite-dimensional modules over the quantum group $U_{q}(\mathfrak{g})$ where $q=e^{\frac{i\pi }{m((l+h^{\vee }) }$, $m$ is the ratio of the squared length of the long roots of $\mathfrak{g} $ to the squared length of the short roots. The interest to this equivalence of representation categories was renewed in the context of attempts to understand representation theory of logarithmic conformal field theories \cite{FSem1}, \cite{FSem2} or of logarithmic vertex operator algebras (VOA) - their mathematically rigorous incarnation (see e.g. \cite{HuRev} and references therein in for mathematically oriented, and \cite{CRbeyond} for physically oriented reviews). One of the main ingredients which differ logarithmic VOA from rational ones is essential role played by reducible but indecomposable modules. The current understanding of representation theory of logarithmic VOA is far from being complete. Since set of intertwiner operators of VOA satisfy KZ equations, analogs of DK theorem, and especially its extension to all the values of deformation parameter, can add to understanding of the representation theory of logarithmic VOAs. The VOA related to the affine Lie superalgebra $\hat{\mathfrak{gl}}(1|1)$ is one of the archetypical examples of logarithmic VOAs \cite{CRarhy}. This motivates to start from DK theorem for this algebra for suitable category of representations, for generic values of the affine level, with a hope to extend analysis of this example beyond the scope of generic values, with further extension to logarithmic VOAs. The description of the category of modules we consider and restrictions on their parameters corresponding to situation of generic level (deformation parameter) will be given below. Of course, the question about DK theorem for superalgebras was addressed before. It turns out that direct copy of Drinfeld's proof of DK theorem for Lie superalgebras is impossible because of the obstacles explained in particular in \cite{Gee}. Nevertheless the author succeeded to prove DK theorem for the classical superalgebras applying Etingof-Kazhdan approach to quantization \cite{EK1} - \cite{EK3} as a bridge for tensor equivalence. We refer to \cite{Gee} and references therein for details, which will be reviewed below. The main object which makes the equivalence of categories explicit is the twist $\mathcal{F}$. Its explicit, non perturbative in $h$ construction in the case of simple Lie algebras is difficult. Some attempt of such explicit construction for simple Lie algebras known to us, without proofs that the constructions indeed implement full braided tensor equivalence, is \cit {Engel}. It is based on basis dependent fundamental representations projectors of simple Lie algebras. Our way of rigorous proof of tensor equivalence is a repeat of the proof of Geer with a trivial argumentation why it works for the case of non-semisimple Lie superalgebra $\mathfrak{gl}(1|1)$, which formally not in the list of superalgebras he considered. The main result of the paper is the Theorem \ref{thm7}. It claims that for the superalgebra $\mathfrak{gl}(1|1)$ two non-semisimple categories of modules are braided tensor equivalent. The first category is the Drinfeld category $\mathcal{D}$ with (equivalence classes of) the typical $\mathcal{T _{e,n}$, atypical $\mathcal{A}_{n}$, and projective $\mathcal{P}_{n}$ modules as objects, such that the parameters $e_{i}$ satisfy $e_{i}/\kappa \notin \mathbb{Z} $ and $(e_{i}+e_{j})/\kappa \notin \mathbb{Z} \backslash \{0\}$ for any pair of typical modules. The second category is the tensor category $\mathcal{C}_{\kappa }$ of corresponding modules \mathcal{T}_{e,n}^{\kappa }$, $\mathcal{A}_{n}^{\kappa }$, $\mathcal{P _{n}^{\kappa }$ of quantum group $U_{\kappa }(\mathfrak{gl}(1|1))$ The paper is organized as follows. In the next Section \ref{section2} we review the main steps of derivation of KZ equations, first in operator form for intertwining operators, then -- for correlation functions of intertwiners. There is almost no difference in it compared to Lie algebra case when non-semisimple finite dimensional modules are included. In the Section \ref{section3} we define Drinfeld category $\mathcal{D}$ for any Lie (super)algebra, and its tensor ring structure in the $\mathfrak{gl}(1|1)$ case for three types of $\mathfrak{gl}(1|1)$-modules. The main part of this section is the proof of existence of associator in the $\mathfrak{gl}(1|1)$ case with its standard properties, as well as the braiding. The Section \ref{section4} defines the category \mathcal{C}_{\kappa }$ of corresponding $U_h(\mathfrak{gl}(1|1))$ quantum group modules with its tensor product ring and other braided tensor category structures. The Section \ref{section5} reviews different aspects of proof of equivalence of the two categories of modules. Some perspectives of continuation of this research is summarized in the Section \ref{section6}. Many technical details, such as bases of the representations, solutions of KZ equations, their asymptotic needed for the proof of associator existence are collected in the Appendix A \ref{appA}. Similar technical information about the quantum group side, including the proof of the tensor product ring structure of the modules in specified bases one can find in the Appendix B \ref{appB}. For the rest of the paper we make an important remark: \textit{The proofs of statements and theorems cited below as known do not use the fact of algebra semisimplicity or semisimplicity of the category of its modules under consideration. The cases where it requires different proofs or leads to different results (like as in analysis of asymptotic solutions of KZ equations) are considered in details. Modifications of proofs related to the fact that we deal with superalgebra are trivial and do not change the cited statements of known theorems. The only needed modifications is in definition of } \mathbb{Z} _{2}$ \textit{graded commutator \begin{equation*} \lbrack A,B]=AB-(-1)^{p(A)p(B)}BA \end{equation* \textit{and the manipulations with tensor products \begin{equation*} (A\otimes B)(a\otimes b)=(-1)^{p(B)p(a)}Aa\otimes Bb \end{equation* \textit{where }$p(x)$ \textit{is the parity of the object }$x$. \textit{An exception from this general rule appears in tensor product decomposition of \mathbb{Z} _{2}$\textit{\ graded modules which sometimes involve parity reverse operator. (It will be explained in the proper cases below)} \textbf{Acknowledgements. }The author is thankful to I.Scherbak for clarifying explanations related to asymptotic solutions of KZ equations, and especially grateful to P.Etingof for many valuable stimulating discussions. \section{Generic $\protect\kappa $ KZ equation} \label{section2} Below we recall standard derivation of operator KZ equation for intertwiners of any affine algebra $\widehat{\mathfrak{g}}$ with some remarks specifying the super case, for affinization of any category of finite dimensional \mathfrak{g}$-modules (possibly indecomposable) at generic $\kappa $. By \kappa $ we denote the inverse quantization parameter discussed above \kappa =h^{-1}=h^{\vee }+k$, $h^{\vee }$ is dual Coxeter number and $\ k$ is the level of affine (super)algebra $\mathfrak{g}$. We also recall standard derivation of KZ equations for correlation functions. The fact that some of modules are indecomposable doesn't hamper to repeat the standard steps of derivation \textit{for generic }$\kappa $ (see \cite{EFK} Lecture 3 for a review). In the case of $\mathfrak{gl}(1|1)$ \ we have $h^{\vee }=0$ and generic means generic values of $k$ which will be specified below. \subsection{Intertwining operators} \label{subIntertwin} We start from\ formulation of affine intertwiners of Tsuchia and Kanie \cit {TK}, summarized in \cite{EFK}, Lecture 3, generalizing it to the non-semisimple $\mathfrak{g}$-modules. Let $\mathfrak{g}$ be a simple Lie (super)algebra over \mathbb{C} $. Let $M_{p}$ be a \textit{finite dimensional} indecomposable (possibly reducible) $\mathfrak{g}$-module, $p$ - some set of parameters which characterise the module. The module is weight: for any homogeneous vector u\in M_{p}$, $\mathfrak{h}u=\lambda _{u}u$ for some $\lambda _{u}\in \mathbb{C} $. We assume that Casimir element $\Omega $ of $U(\mathfrak{g)}$ can act non diagonally and we decompose $\Omega =C_{d}+C_{nil}$ where $C_{d}$ acts diagonally with the same eigenvalue $\lambda _{p}$ on all the vectors of the module, and $C_{nil}$ is a nilpotent part of non-diagonal action: (C_{nil})^{n}=0$ for some non negative integer $n$. In the case of non-super Lie algebras $M_{p}$ is assumed to be a highest weight module and $C_{nil}=0$. We relax this requirement. In what follows all commutations and tensor products are understood as \mathbb{Z} _{2}$ graded for the case of superalgebras. We recall some notations and definitions related to affine Lie (super)algebraic modules. We consider triangular decomposition with respect to \mathbb{Z} $-grading $\widehat{\mathfrak{g}}=\widehat{\mathfrak{g}}_{<0}\mathfrak \oplus }\widehat{\mathfrak{g}}_{0}\mathfrak{\oplus }\widehat{\mathfrak{g} _{>0}$, and induced $\widehat{\mathfrak{g}}$-modules $M_{p,k}=Ind_{\widehat \mathfrak{g}}_{\geq 0}}^{\widehat{\mathfrak{g}}}M_{p}$, \textit{for generic $k$ , where the action of $\widehat{\mathfrak{g}}_{>0}=\mathfrak{g\otimes \mathbb{C} \lbrack t]}$ is trivial, and the action of $\widehat{\mathfrak{g}}_{0} \mathfrak{g\oplus }k\mathfrak \mathbb{C} }$ is such that it is isomorphic to the action of $\mathfrak{g}$ for the first summand, and is multiplication by $k$ - for the second. The modules M_{p,k}$ admit \mathbb{Z} $-grading \begin{equation*} M_{p,k}=\bigoplus\limits_{n\geq 0}M_{p,k}[-n] \end{equation* and $M_{p,k}[-n]$ is the eigenspace of affine Lie (super)algebra derivation d$ with the eigenvalue $-n-\Delta _{p,k}$, $\Delta _{p,k}$ is the conformal dimension. The $M_{p,k}[0]$ is naturally a $\mathfrak{g}$-module isomorphic to $M_{p}$. Recall that for simple Lie algebras we can restrict and define as generic \kappa =k+h^{\vee }$ such that $k\notin \mathbb{Q} $. In our superalgebra case we mean by generic $k\in \mathbb{C} $ restricted by suitable constraints dependent on modules parameters, such that they guarantee that if $M_{p}$ is irreducible then $M_{p,k}$ is also irreducible, and if $M_{p}$ is of finite length with composition factors L_{p_{i}}$ then $M_{p,k}$ is also of finite length with the composition factors $L_{p_{i},\kappa }=Ind_{\widehat{\mathfrak{g}}_{\geq 0}}^{\widehat \mathfrak{g}}}L_{p_{i}}$. We will see below what restrictions on the level k $ for the affine $\mathfrak{gl}(1|1)_k$ case it implies. Another important kind of $\widehat{\mathfrak{g}}$-modules we need in order to define affine intertwiner is \textit{evaluation module. }Let $M_{p}$ be a $\mathfrak{g}$-module which admits weight decomposition with finite dimensional weight spaces. For a non zero complex number $z$ and for any element $x\otimes P(t)$ of $\mathfrak{g}\otimes \mathfrak \mathbb{C} }[t,t^{-1}]$ where $P$ is a polynom, we define its action $x\otimes P(t)\cdot u=P(z)xu$, where $u\in M_{p}$, and the central element of \widehat{\mathfrak{g}}$ acts trivially. After that one can extend such action on the whole $\widehat{\mathfrak{g}}$ by replacing the space of action by a bigger one $M_{p}\otimes z^{-\Delta }\mathfrak \mathbb{C} }[z,z^{-1}]\footnote Strictly speaking it requires to consider $z$ as a formal variable $z$ and restoration of status of complex variable requires subtle procedure worked out in vertex operator algebras formalism by Huang and Lepowsky, and later by Huang, Lepowsky and Zhang for logarithmic vertex operator algebras (see the list of references in \cite{HuRev})}$, With a standard extension of \widehat{\mathfrak{g}}$ to $\widetilde{\mathfrak{g}}$ by derivation $d$, we define the action of the affine derivation as $d=z\frac{d}{dz}$. \ Here \Delta \in \mathbb{C} $ is specified as module dependent parameter (its conformal dimension). Now we have the generating function for \mathbb{Z} $-graded modes $v[n]=v\otimes t^{n}\in M_{p,k}$ : $v(z)=\sum_{n\in \mathbb{Z} }v[n]z^{-\Delta -n}$. We denote the vector space of such objects as M_{p}(z) $. We are equipped now with all necessary ingredients for construction of affine intertwiner. Assume we can classify all $\mathfrak{g}$-homomorphisms of the form $\varphi :M_{p_{1}}\rightarrow M_{p_{0}}\otimes M_{p}$, $g\in \mathfrak{g}$. We call them $\mathfrak{g}$-intertwiners and we want to lift them to $\widehat{\mathfrak{g}}$-intertwiners $\Phi $. Our consideration will be restricted to the special class of intertwiners which preserve the superalgebra \mathbb{Z} _{2}$ grading. We define a $\widehat{\mathfrak{g}}$-intertwiner as a homomorphism $\Phi :M_{p_{1},k}\rightarrow M_{p_{0},k}\widehat{\otimes M_{p}(z)$ which satisfies \begin{equation*} \Phi (z)x[n]=(x[n]\otimes 1+z^{n}\cdot 1\otimes x)\Phi (z) \end{equation* where $z$ is a non zero complex number. Here $\widehat{\otimes }$ is understood as a completed tensor product consisting of infinite sums of tensor products of homogeneous vectors. Recall that in the case of simple non-super algebras, for the picture of complete braided tensor category (BTC) structure, it is enough to consider affinization of finite dimensional highest weight $\mathfrak{g}$-modules (sometimes called Weyl modules), and evaluation modules. We will do the same for superalgebras relaxing the condition that we affinize and build evaluation modules over highest weight irreducible modules: the $\mathfrak{g} $-modules are not necessarily highest weight, and may be reducible but indecomposable. Almost all the steps of intertwiners construction can be copied from the non-super case. In particular the following assertion can be proved as a slight generalization to the superalgebra case of the Theorem 3.1.1 of \cite{EFK}, which was originally proved in \cite{TK} in $\widehat \mathfrak{sl}}(2)$ case (Theorem 1, followed from Proposition 2.1 and Theorem 2.2) \begin{proposition} \label{pr0}Let $\varphi :M_{p_{1}}\rightarrow M_{p_{0}}\otimes M_{p}$ be a $\mathfrak{g}$-homomorphism. \ Then for generic $k$ there exists a unique $\widehat{\mathfrak{g}}$-intertwiner $\Phi (z):M_{p_{1},k}\rightarrow M_{p_{0},k}\widehat{\otimes }M_{p}(z)$ such that for every vector $v\in M_{p_{1},k}[0]\simeq $ $M_{p_{1}}$the zero degree component of $\Phi (z)v$ is equal to $\varphi v$. \end{proposition} We recall here the sketch of the proof\textit{.} Because of the annihilation condition for $w\in M_{p_{1},k}[0]\simeq M_{p_{1}}$ by $\mathfrak{g\otimes \mathbb{C} \lbrack t]}$ we have \begin{equation} \Phi (z)w\in \left( M_{p_{0},k}\widehat{\otimes }M_{p}(z)\right) ^{\mathfrak g}\otimes \mathbb{C} \lbrack t]}=Hom_{\mathfrak{g}\otimes \mathbb{C} \lbrack t]}\left( M_{p_{0},k}^{\ast },M_{p}(z)\right) \label{n1} \end{equation The contragredient module $M_{p_{0},k}^{\ast }$ is freely generated over \mathfrak{g\otimes \mathbb{C} \lbrack t]}$ \textit{for generic }$k$, therefore the restriction map \begin{equation*} Hom_{\mathfrak{g}\otimes \mathbb{C} \lbrack t]}\left( M_{p_{0},k}^{\ast },M_{p}(z)\right) \rightarrow Hom_ \mathbb{C} }\left( M_{p_{0}}^{\ast },M_{p}\right) \simeq (\Pi )M_{p_{0}}\otimes M_{p} \end{equation* is an isomorphism. ($\Pi $ means the parity inversion which sometimes is necessary in superalgebra case, see below.) Therefore $\Phi (z)w$ is uniquely defined by its zero grade component. Then we define the homomorphic action of $\Phi (z)$ on any $Xw=u\in M_{p_{1},k}$, where $X\in U(\widehat{\mathfrak{g}})$, $X=\pro \limits_{i}x_{n_{i}}^{(i)}$, is written in PBW basis, where $x^{(i)}\in \mathfrak{g}$ and $n_{i}\in \mathbb{Z} $. We define it inductively over each factor $x_{n_{i}}^{(i)}$ of $X$ by \begin{equation} \Phi (z)x_{n_{i}}^{(i)}=[x_{n_{i}}^{(i)}\otimes 1+z^{n}(1\otimes x_{n_{i}}^{(i)})]\Phi (z),\;n<0 \label{n2} \end{equation It defines an $\widehat{\mathfrak{g}}$-intertwiner uniquely with an obvious property that it is a lifting of the $\mathfrak{g}$-intertwiner $\varphi $. A subtle point of the above definition is that some more general and rigorous construction is needed in order to treat $z$ as a genuine complex variable and not just formal variable. This construction was elaborated in the seminal series of papers by Huang and Lepowsky - see the footnote above - in the framework of vertex operator algebras (VOA). In particular more general intertwiners of the form \begin{equation*} \mathcal{Y}(\_,z):W_{1}\rightarrow Hom(W_{2},W_{3})\{z\}[\log z] \end{equation* are usually needed in logarithmic vertex operator superalgebra (VOSA) $V$ case, where $W_{i}$ are some $V$-modules. It is precisely relevant for our \mathfrak{gl}(1|1)$ case, but we will continue to use the definition of intertwiners described above, without use of VOSA language. Despite the lack of proper rigorously we will continue to treat $z$ in our approach as complex variable, defining, where it is needed a branch of multivalued functions. In particular for what follows we chose $\ln z=\ln |z|+iArg(z),-\pi <Arg(z)\leq \pi $. An important remark is in order here. Recently, when a preliminary version of this paper was finished, an important progress was achieved in understanding of braided tensor category structure of the $\mathfrak{gl (1|1) $ VOSA \cite{CMY}. The next standard step is to extend this $\widehat{\mathfrak{g}} -homomorphism to $\widetilde{\mathfrak{g}}$-homomorphism, where $\widetilde \mathfrak{g}}$ is the standard extension of $\widehat{\mathfrak{g}}$ by affine derivation $d=-L_{0}$, with $L_{m}$ defined by Sugawara construction. \begin{equation} L_{m}=\frac{1}{2\left( k+h^{\vee}\right) }\sum_{a,b}\sum_{n\i \mathbb{Z} }B_{ab}^{-1}:J_{n}^{a}J_{m-n}^{b}: \label{n4} \end{equation} where $h^{\vee}$ is a dual Coxeter number of (super)algebra\footnote This construction can be modified in the case of non semisimple (super)algebra. It acts as a scalar on simple modules, but sometimes acts non diagonally on indecomposables, as for example in the case of $\widehat{g }(1|1)$.}, and $B$ - $\mathfrak{g}$-invariant (super)symmetric non-degenerated bilinear form. If we want to extend the intertwining homomorphism $\Phi(z)$ defined in Proposition \ref{pr0} to $\widetilde{\mathfrak{g}}$-homomorphism we have to twist it. We define two twisted intertwiners: for $w\in M_{p_{1},k}$ \begin{align} \widehat{\Phi }^{g}(z)w& =(z^{L_{0}}\otimes z^{L_{0}})\left( z^{-L_{0}}\Phi (z)wz^{L_{0}}\right) (z^{-L_{0}}\otimes z^{-L_{0}}), \label{n3} \\ \widetilde{\Phi }^{g}(z)w& =(z^{L_{0}}\otimes 1)\left( z^{-L_{0}}\Phi (z)wz^{L_{0}}\right) (z^{-L_{0}}\otimes 1) \label{n31} \end{align They remain intertwiners with image in \begin{equation} z^{-L_{0}}M_{p_{0},k}z^{L_{0}}\widehat{\otimes z^{-L_{0}}M_{p}z^{L_{0}}[z,z^{-1}] \notag \end{equation and \begin{equation} z^{-L_{0}}M_{p_{0},k}z^{L_{0}}\widehat{\otimes }M_{p}[z,z^{-1}] \notag \end{equation respectively. In the case of irreducible highest weight modules $M_{p_{i}}$ with highest weight $p_{i}$ these twists reduce to the standard scalar factors twists $\widehat{\Phi }^{g}(z)=\sum_{n}\Phi(n)z^{-n-\Delta }$, \Delta =\Delta (p_{1})-\Delta (p_{0})-\Delta (p)$, and the same for \widetilde{\Phi }^{g}$ with $\Delta (p_{1})-\Delta (p_{0})$, where $\Delta _{i}=\frac{\langle p_{i},p_{i}+2\rho \rangle }{2(k+h^{\vee })}$. (The factor $z^{\Delta (p_{0})+\Delta (p)}$ is moved to the definition of $\widehat{\Phi }^{g}(z)$ by the first and the last parenthesis factors.) For the restricted dual $M_{p}^{\ast }$ and its evaluation module M_{p}^{\ast }(z)\cong (M_{p}(z))^{\ast }$ which are assumed to be well defined, we can take any vector $u\in M_{p}^{\ast }$, define $\widehat{\Phi _{u}^{g}(z)w=\langle 1\otimes u,\widehat{\Phi }^{g}(z)w\rangle ,w\in M_{p_{1},k}$ and regard it as an operator $\widehat{\Phi _{u}^{g}(z):M_{p_{1},k}\rightarrow M_{p_{0},k}$. Then the proof of the theorem \cite{KZ}, \cite{FR} Theorem 2.1, about the operator form of KZ equation which says that \begin{equation} (k+h^{\vee })\frac{d}{dz}\widehat{\Phi }_{u}^{g}(z)=\sum_{a\in B}:J_{a}(z \widehat{\Phi }_{au}^{g}(z): \label{n8} \end{equation (summation is over the basis $B$ of $\mathfrak{g}$) generalizes to the case of indecomposable modules $M_{p_{i}}$ actually without changes. Recall the proof. Obviously the intertwining relation (\ref{n2}) is satisfied for $\widehat \Phi }^{g}(z)$ as well. Applying contravariant bilinear form in the space M_{p}$ this relation can be written as \begin{equation*} \lbrack \widehat{\Phi }_{u}^{g}(z),x[n]]=z^{n}\widehat{\Phi }_{xu}^{g}(z) \end{equation* If we introduce currents $J_{x}^{\pm }(z)$ for any algebra element $x$ \begin{equation*} J_{x}(z)=J_{x}^{+}(z)-J_{x}^{-}(z),\ J_{x}^{+}(z)=\sum_{n<0}x[n]z^{-n-1},\ J_{x}^{-}(z)=-\sum_{n\geq 0}x[n]z^{-n-1} \end{equation* then in terms of these currents the last intertwining property takes the for \begin{equation} \lbrack J_{x}^{\pm }(\zeta ),\widetilde{\Phi }_{u}^{g}(z)]=\frac{1}{z-\zeta \widetilde{\Phi }_{xu}^{g}(z) \label{crcur} \end{equation (plus sign corresponds to $|\zeta |<|z|$, and minus sign -- to $|\zeta |>|z| ). Now we write the $d$-invariance property of $\widetilde{\Phi }_{u}^{g}(z) \begin{equation*} z\frac{d}{dz}\widetilde{\Phi }_{u}^{g}(z)=-[d,\widetilde{\Phi }_{u}^{g}(z)] \end{equation* which is the same as \begin{equation*} z\frac{d}{dz}\widehat{\Phi }_{u}^{g}(z)=-[d,\widehat{\Phi }_{u}^{g}(z)]+ \frac{d}{dz}(1\otimes z^{-L_{0}})\widehat{\Phi }_{u}^{g}(z)(1\otimes z^{L_{0}}) \end{equation* We can continue by Sugawara construction \begin{align*} & \frac{B_{a,b}^{-1}}{2(k+h^{\vee })}\left( \sum_{n\leq 0}[J^{a}[n]J^{b}[-n] \widehat{\Phi }_{u}^{g}(z)]+\sum_{n>0}[J^{a}[-n]J^{b}[n],\widehat{\Phi _{u}^{g}(z)]\right) + \\ &z\frac{d}{dz}(1\otimes z^{-L_{0}})\widehat{\Phi _{u}^{g}(z)(1\otimes z^{L_{0}})= \\ & \frac{B_{a,b}^{-1}}{2(k+h^{\vee })}\{2zJ_{b}^{+}(z)\widehat{\Phi _{au}^{g}(z)-2z\widehat{\Phi }_{bu}^{g}(z)J_{a}^{-}(z)+J_{b}^{+}[0]\widehat \Phi }_{au}^{g}(z)-\widehat{\Phi }_{bu}^{g}(z)J_{a}^{-}[0]\} \\ & +z\frac{d}{dz}(1\otimes z^{-L_{0}})\widehat{\Phi }_{u}^{g}(z)(1\otimes z^{L_{0}})= \\ & \frac{B_{a,b}^{-1}}{k+h^{\vee }}:J_{a}(z)\widehat{\Phi }_{bu}^{g}(z): \frac{B_{a,b}^{-1}}{2(k+h^{\vee })}(J_{b}^{+}[0]\widehat{\Phi }_{au}^{g}(z) \widehat{\Phi }_{bu}^{g}(z)J_{a}^{-}[0])+ \\ &z\frac{d}{dz}(1\otimes z^{-L_{0}} \widehat{\Phi }_{u}^{g}(z)(1\otimes z^{L_{0}}) \end{align* The last two terms cancel because they can be written as \begin{equation*} \frac{1}{2(k+h^{\vee })}\widehat{\Phi }_{Cu}^{g}(z)-\Delta (p)\widehat{\Phi _{u}^{g}(z) \end{equation* where $C=B_{a,b}^{-1}J_{a}J_{b}$ is a Casimir element of the algebra \mathfrak{g}$, (not to be confused with tensor Casimir defined in (\ref{n15 ). This completes the proof. Concluding this section about systematic definition of intertwining operators for affine Lie (super)algebra we can illustrate an important difference of a non-semisimple case from a semisimple one. Suppose we have a finite dimensional $\mathfrak{g}$-module with non-semisimple action of the Casimir element $C$ which we can represent as $C=\lambda I+C_{nil}$, where I $ acts as identity and $C_{nil}$ acts nilpotently: $C_{nil}^{n}=0$ on each vector of the module. Then the action of $z^{aC}$ where $z,\lambda \in \mathbb{C} $, on any vector $w$ of the module can be written a \begin{equation} z^{C}w=z^{\lambda }\sum_{m=1}^{n-1}\frac{(\ln z)^{m}}{m!}C_{nil}^{m}w \label{n6} \end{equation leading to a presence of logarithms (with a choice of a branch), which necessarily arises with non-semisimplicity in logarithmic vertex operator algebras and in logarithmic conformal field theories. A modification of this example for the operator like $z^{\Omega _{ij}}$ will appear below where the Casimir $C$ is replaced by quantum Casimir $\Omega _{ij}$, see (\ref{n15}) below. \subsection{KZ equation for correlation functions} \label{subCorFun} The way of derivation of correlation functions KZ equation from operator KZ equation (\ref{n8}) first appeared in the seminal paper of Knizhnik and Zamolodchikov \cite{KZ}. Later it was derived more rigorously in \cite{FR} and in textbooks like e.g. \cite{EFK}, Sect. 3.4. Possible non-semisimplicity of Lie (super)algebra $\mathfrak{g}$-modules doesn't lead to serious modifications in the derivation. We again recall the main steps of it. In order to define correlation function consider the modules M_{q_{i},k},i=1,...,N$, and $M_{p_{i}}$, $i=1,...,N+1$. Let $\widehat{\Phi ^{g_{i}}(z_{i}):M_{p_{i},k}\rightarrow M_{p_{i-1},k}\widehat{\otimes M_{q_{i}}[z_{i}^{\pm 1}]$ be an intertwiner as explained above, where M_{q_{i}}(z_{i})$ is evaluation module. We consider the homomorphis \begin{eqnarray} \Psi (z_{1},...,z_{N})=\left( \widehat{\Phi }^{g_{1}}(z_{1})\otimes 1...\otimes 1\right) ...\left( 1...\otimes \widehat{\Phi ^{g_{N-1}}(z_{N-1})\otimes 1\right) \times \nonumber\\ \left( 1\otimes ...\otimes \widehat{\Phi}^{g_{N}}(z_{N})\right) \label{n10} \end{eqnarray that maps $M_{p_{N},k}\rightarrow M_{p_{0},k}\widehat{\otimes }M_{q_{1} \widehat{\otimes }...\widehat{\otimes }M_{q_{N}}$. This formula for homomorphism makes sense at least being understood as formal power series in $z_{1},z_{2},...,z_{N}$ and their logarithms. Consider a subspace of weight $\lambda_{N}$ of $M_{p_{N},k}[0]$, and subspace of weight $-\lambda_{0}$ of $M_{p_{0},k}^{\ast}[0]$. The object \Psi (z_{1},...,z_{N})|\lambda_{N}\rangle$ takes values in the space M_{q_{1}}\otimes M_{q_{2}}...\otimes M_{q_{N}}\otimes M_{p_{0}}$. We can take a projection of it onto finite dimensional invariant subspace of the weight $\lambda_{N}-\lambda_{0}$ in the $M_{p_{0}}$ component of it V=\left( M_{q_{1}}\otimes M_{q_{2}}...\otimes M_{q_{N}}\right) ^{\lambda_{N}-\lambda_{0}}$. If we take $\lambda_{N}=\lambda_{0}$ then we get the $\mathfrak{g}$ invariant subspace $V^{\mathfrak{g}}$. This sort of projection of $\Psi$ on such a subspace, with some chosen $u_{N+1}\in M_{p_{N},k}[0]$, $u_{0}\in$ $M_{p_{0},k}[0]$, is called a \textit correlation function} \begin{align} \psi(z_{1},...z_{N}) & =\langle u_{0},\Psi(z_{1},...z_{N})u_{N+1}\rangle \label{n111} \\ \psi(z_{1},...z_{N}) & \in\left( M_{q_{1}}\otimes M_{q_{2}}...\otimes M_{q_{N}}\right) ^{\lambda_{N}-\lambda_{0}} \notag \end{align} (the vector $\langle u_{0}|\in M_{p_{0},k}^{\ast}[0]$ ) Taking into account the remark (\ref{n6}) we can say that $\psi$ here is defined as a formal power series: it belongs to \\ $\prod \limits_{i}z_{i}^{-\Delta(p_{i})+\Delta(p_{i-1})+\Delta(q_{i})}(\ln\frac z_{i}}{z_{i-1}})^{n_{i} \mathbb{C} \lbrack\lbrack\frac{z_{2}}{z_{1}},...\frac{z_{N}}{z_{N-1}}]]$. Equivalently one can define correlation function as \mathbb{C} $-valued if choosing $u_{i}\in M_{q_{i}},i=1,...,N$, we define \begin{equation} \psi_{u_{1},...,u_{N+1}}(z_{1},...z_{N})=\langle u_{0},\widehat{\Phi _{u_{1}}^{g_{1}}(z_{1})...\widehat{\Phi}_{u_{N}}^{g_{N}}(z_{N})u_{N+1 \rangle\i \mathbb{C} \label{n13} \end{equation} In particular one can take $M_{p_{0},k}=M_{p_{N},k}$ to be the scalar representation $M_{0}$, i.e. $M_{p_{0},k},M_{p_{N},k}$ -- induced vacuum modules with the zero grade vector $u_{0}$, and define $V$-valued correlation function. \begin{equation} \phi(z_{1},...z_{N})=\langle u_{0},\Psi(z_{1},...,z_{N})u_{0}\rangle \label{n12} \end{equation} Then $\phi(z_{1},...z_{N})\in V^{\mathfrak{g}}$.\footnote In the super algebras case it sometimes happens that a scalar representation appears only as a (part of) atypical module. By general tensor category "ideology" atypical modules should be replaced by their projective covers. But even then there is a "bottom" vector $u_{N+1}$ in it satisfying \mathfrak{g}u_{N+1}=0$.} The main theorem proved in \cite{KZ} for simple highest weight modules of (non super) algebra, claims the KZ equation on (\ref{n111}) in the form \begin{equation} (k+h^{\vee })\partial _{i}\psi =\left( \sum_{j\neq i=1}^{N}\frac{\Omega _{ij }{z_{i}-z_{j}}+\frac{\Omega _{i,N+1}}{z_{N}}\right) \psi ,\;i=1,...,N+1 \label{n141} \end{equation Equivalent form of KZ equation can be obtained by adding one more formal variable $z_{N+1}$ to the function \ $\psi (z_{1},...z_{N})=\psi (z_{1}-z_{N+1},...z_{N}-z_{N+1})$, giving \begin{equation} (k+h^{\vee })\partial _{i}\psi =\left( \sum_{j\neq i=1}^{N+1}\frac{\Omega _{ij}}{z_{i}-z_{j}}\right) \psi ,\;i=1,...,N+1 \label{n14} \end{equation Here we denote tensor Casimir \begin{equation} \Omega _{ij}=B_{ab}^{-1}(x^{a})_{i}\otimes ^{s}(x^{b})_{j} \label{n15} \end{equation (the lower indices $i,j$ indicate the spaces of the tensor product in $V$ where the generators $x^{a}$ act.) and $z_{N+1}=0$. Recall that the vectors u_{0}\in M_{p_{0},k}[0]$ and $u_{N+1}\in M_{p_{N},k}[0]$ have grade 0 . Here we use the super tensor product which for two matrices $A_{\alpha \gamma }$ and $B_{\beta \delta }$ is defined as $(A\otimes ^{s}B)_{\alpha \beta }^{\gamma \delta }=(-1)^{\beta (\alpha +\gamma )}A_{\alpha \gamma }B_{\beta \delta }$, where the indices lifted to exponential of $(-1)$ are parities of corresponding indices in \mathbb{Z} _{2}$ graded vector spaces. The main difference compared to the usual non superalgebras and irreducible finite dimensional highest weight modules is that $\Omega _{ij}$ can act now non diagonally on the modules. In this sense they are not eigenvalue numbers but operators. With the assumption that u_{N+1}$ is the vector of scalar representation (at least in the sense described in the footnote) the last term in (\ref{n141}) disappears, and the equation we will deal with in what follows \begin{equation} (k+h^{\vee })\partial _{i}\psi =\sum_{j\neq i=1}^{N}\frac{\Omega _{ij}} z_{i}-z_{j}}\psi ,\;i=1,...,N \label{n16} \end{equation} The proof of the theorem claiming (\ref{n16}) for correlation functions for superalgebras with non-semisimple modules is a copy of the proof in the case of simple modules over usual Lie algebras. The proof uses commutation relations (\ref{crcur}) and the fact that $u_{0},u_{N+1}$ are zero grade states. Looking for solutions for $\psi \in V^{\mathfrak{g}}$ is not the only option. One can get a set of solutions when $\psi $ is projected onto some weight subspace $\psi \in V^{\lambda }$ of weight $\lambda $. Usually, when the spaces $M_{p_{i}}$ are highest weight ones $\mu _{i}$, the solutions with values in the space $(V^{\mathfrak{n}^{+}})^{\lambda }$ are considered. If $\lambda =\sum \mu _{i}-\mu $, $\mu =\sum n_{i}\alpha _{i}$, $\alpha _{i}\in Q^{+}$, the value $|\mu |=\sum n_{i}$ is called level of the equatio \footnote It will be interesting to find a direct way to obtain non zero level solution from the zero level solutions ones, as it was done in non-super case \cite{Var}}. Usually level one solutions for $N=3$ already give solutions with a basis of hypergeometric functions. But in order to see such hypergeometric solutions in $V^{\mathfrak{g}}$, one has to take at least N=4 $ correlation functions. Important particular case of KZ equation when it becomes an ordinary differential equation, is the $N=3$ case. As one can show (see e.g. \cit {EFK}), in this case any solution of KZ equation can be written as \begin{equation*} \psi(z_{1},z_{2},z_{3})=(z_{1}-z_{3})^{(\Omega_{12}+\Omega_{13}+\Omega _{23})/\kappa}f\left( \frac{z_{1}-z_{2}}{z_{1}-z_{3}}\right) \end{equation*} where $f(z)\in V$ satisfies the differential equatio \begin{equation} \kappa\partial_{z}f=\left( \frac{\Omega_{12}}{z}+\frac{\Omega_{23}}{z-1 \right) f \label{3kz} \end{equation} For the irreducible modules $M_{q_{1}},...,M_{q_{N}}$ of highest/lowest weight there is a classification and explicit form of solutions of KZ equation for specified level of weights in root lattice grading. Level zero solution is always of the for \begin{align*} \Psi _{0}(z_{1},...z_{N})& =\psi _{0}(z_{1},...z_{N})v,\;v=\mu _{1}\otimes \mu _{2}...\otimes \mu _{N}, \\ \psi _{0}(z_{1},...z_{N})& =\prod\limits_{i<j}(z_{i}-z_{j})^{\mu _{i}\mu _{j}/2\kappa } \end{align*} Solutions of higher levels of KZ equations in the case of highest or lowest weight modules $M_{\lambda _{i}}$ at generic $\kappa $ one can obtain by the following procedure. (We consider highest weight modules). Define multi-valued function \begin{equation*} \phi _{1}(z_{1},...z_{N},t)=\prod\limits_{i=1}^{N}(t-z_{i})^{\mu _{i}/\kappa } \end{equation* and fix a closed contour $C$ in $t$ complex plane not containing any of z_{i}$, and having a continuous branch along $C$. Example of such contour is Pochhammer contour for two $z_{a},z_{b}.$ Existence and classification of such contours is known for semisimple case, but is a non trivial question for non semisimple case. Then a general level one solution $\Psi _{1}(z_{1},...z_{N})$ can be obtained as \begin{equation} \Psi _{1}(z_{1},...z_{N})=\psi _{0}(z_{1},...z_{N})\sum_{r=1}^{N}\left( \int_{C}dt\phi _{1}(z_{1},...z_{N},t)\frac{1}{t-z_{r}}\right) f_{r}v \label{integr} \end{equation where $v=v_{1}\otimes ...\otimes v_{N}$ is the highest weights tensor product, and the step operator $f_{r}$ acts on the $r$th component of tensor product. The proof is by direct calculations. Explicit realization of this solution gives rise to integral representations of hypergeometric functions _{2}F_{1}$. Level $l$ solution can be similarly generated by integration of operator valued differential $l$-forms. The answer in this case is much more involved \cite{Var}. For the case of semisimple categories of finite dimensional $\mathfrak{g} -modules at generic level $\kappa $ the most important statement says that the monodromy of KZ equations gives rise to braided tensor categories, and that they equivalent to the categories of specific quantum group representation. One of the ways to see it for generic level case was worked out by Schechtman and Varchenko \cite{ScheVar} using the integral formulas of the KZ solutions by analysis of geometry of integration cycles. Can the same be done in the case of non-semisimple categories of $\mathfrak{g} -modules when solutions involve logarithms? We are going to address this question elsewhere. All the construction above treats $z_{i}$ as formal variables. There is a theorem proved for KZ equations in semisimple case that $\psi $ is an analytic function of $z_{i}$ in the region $|z_{1}|>|z_{2}|>...>0$. This analyticity should be modified in the non semisimple case because of presence of logarithms in intertwiners mode expansions. Consistency and $\mathfrak{g}$-invariance of KZ equation, as in semisimple case, follows from $\mathfrak{g}$-invariance of Casimir operator. It has an important practical application: in order to find the full set of independent KZ equations for a given correlation function one should find the basis of invariants of the space $V$ -- the set of tensor product vectors annihilated by all the generators of $\mathfrak{g}$, and then project the equations on these vectors. One can find some examples of such calculations in Appendix \ref{subCorf}. Explicit construction of tensor category structures of solutions of KZ equations requires calculations up to $N=4$ -- four point correlation functions. The final goal is investigation of monodromy properties of solutions of KZ equation. By this we mean the following. The system of KZ equations being consistent can be interpreted as a flat connection in the trivial vector bundle with the fiber $V$ over the configuration space $X_{N} \{(z_{1},z_{2},...,z_{N})\in \mathbb{C} ^{N}\;|\;z_{i}\neq z_{j}\}$. For any path $\gamma :[0,1]\rightarrow X_{N}$ we denote by $M_{\gamma }$ the operator of holonomy along $\gamma $. It can be considered as an operator in $V$ and it depends only on homotopy class of $\gamma $, or as operator of analytic continuation along $\gamma $. From \mathfrak{g}$-invariance of $\Omega $ follows that for any $\gamma $ M_{\gamma }:V\rightarrow V$ is a $\mathfrak{g}$-homomorphism. If $V$ is completely reducible, then it means that $M_{\gamma }$ preserves subspace of singular vectors in $V$ and is uniquely defined by its action on this subspace. \section{Drinfeld category of $\mathfrak{gl}(1|1)$ modules} \label{section3} In this section we consider the KZ equation as an equation on functions \begin{equation*} \psi (z_{1},...,z_{N}) \mathbb{C} ^{N}\setminus \{Diag\}\rightarrow V[[\kappa ^{-1}]] \end{equation* where the set of points $\{Diag\}:z_i=z_j, i\neq j$ are removed from the domain $\mathbb{C}^N$. The functions are valued in $V[[\kappa ^{-1}]]$, where $V=V_{1}\otimes ...\otimes V_{N}$ is a tensor product of moduli representation spaces $V_{i}$ of the superalgebra \mathfrak{gl}(1|1)$. We define abelian tensor Drinfeld supercategory \mathcal{D}$ of subset of finite dimensional non semisimple $\mathfrak{gl (1|1)$-moduli with all moduli homomorphisms as the category morphisms, and construct its braided tensor category structure. The superalgebra $\mathfrak{gl}(1|1)$ is the algebra of endomorphisms of the vector superspace \mathbb{C} ^{1|1}$ $\mathfrak{gl}(1|1)=span\{E,N,\psi ^{+},\psi ^{-}\}$ with two dimensional even $\mathfrak{gl}(1|1)_{\overline{0}}=span\{E,N\}$ and two dimensional odd $\mathfrak{gl}(1|1)_{\overline{1}}=span\{\psi ^{+},\psi ^{-}\}$ subspaces written in the superalgebra basis. The commutation relations of the algebra, explicit form of the basis in the representations we consider are collected in the Appendix \ref{subgl11}. The objects of $\mathcal{D}$ include (isomorphism classes of) three types of modules - typical $\mathcal{T}_{e,n}$, and atypical $\mathcal{P}_{n} \mathcal{A}_{n}$ four- and one-dimensional modules (see the Appendix \re {subgl11} for description of the meaning of moduli parameters $e,n$). We impose the restrictions on the parameters $e_{i}$ of the set of typical modules $\mathcal{T}_{e_{i},n_{i}}$ to be such that three conditions are satisfied. First, $e_{i}\kappa ^{-1}\notin \mathbb{Z} $, just by definition of the typical module. Second, $\kappa ^{-1}\sum_{k}e_{i_{k}}\notin \mathbb{Z} \backslash \{0\}$ for any subset of the set of typical modules $\mathcal{T _{e_{i},n_{i}}$ in the category (the reason for this restrictions will be explained below). And the third, for any subset of typical modules \sum_{k}e_{i_{k}}$ is equal either to some $e_{n}$ or to 0. The reason for the latter restriction is that such choice guarantee the closure of the set of objects under the tensor product decomposition. This decomposition is well known (see for example \cite{CRarhy}, eq. (2.11)-(2.16)) \begin{align} \mathcal{A}_{n}\otimes \mathcal{A}_{n^{\prime }}& =\mathcal{A}_{n+n^{\prime }},\;\mathcal{A}_{n}\otimes \mathcal{T}_{e,n^{\prime }}=\mathcal{T _{e,n+n^{\prime }} \label{gl1} \\ \mathcal{T}_{e,n}\otimes \mathcal{T}_{e^{\prime },n^{\prime }}& =\mathcal{T _{e+e^{\prime },n+n^{\prime }+1/2}\oplus \Pi \mathcal{T}_{e+e^{\prime },n+n^{\prime }-1/2}, \notag \\ \mathcal{T}_{e,n}\otimes \mathcal{T}_{-e,n^{\prime }}& =\mathcal{P _{n+n^{\prime }},\;\mathcal{A}_{n}\otimes \mathcal{P}_{n^{\prime }}=\mathcal P}_{n+n^{\prime }}, \notag \\ \mathcal{T}_{e,n}\otimes \mathcal{P}_{n^{\prime }}& =\Pi \mathcal{T _{e,n+n^{\prime }+1}\oplus 2\mathcal{T}_{e,n+n^{\prime }}\oplus \Pi \mathcal T}_{e,n+n^{\prime }-1}, \notag \\ \mathcal{P}_{n}\otimes \mathcal{P}_{n^{\prime }}& =\mathcal{P}_{n+n^{\prime }+1}\oplus 2\Pi \mathcal{P}_{n+n^{\prime }}\oplus \mathcal{P}_{n+n^{\prime }-1}. \notag \end{align The functor $\Pi $ for some modules on the right hand side denotes parity reversion of the \mathbb{Z} _{2}$ grading of even and odd module subspaces. The obvious requirements on the set of parameters $n_{i}$ of the modules in the category similarly follows by closure of the tensor product decomposition. No other restrictions on the moduli parameters $n_{i}$ are imposed. Of course there are infinitely many other finite dimensional indecomposable \mathfrak{gl}(1|1)$-modules, but our choice seems to be the minimal set of (isomorphism classes of) modules closed under the tensor product decomposition with a non trivial braiding structure described below. The indecomposable modules $\mathcal{P}_{n}$ are called projective, because they are projective covers for $\mathcal{A}_{n}$. The typical modules \mathcal{T}_{e,n}$ are their own projective covers. Indecomposable structure of the modules can be found in the same reference \cite{CRarhy}, Section 2.2. The modules of our category are finitely generated and are semisimple under the action of the even part of the superalgebra. Some properties of such categories of $\mathfrak{gl}(m|n)$-modules were reviewed in \cite{Br}. We see that one should include in the category the modules obtained by the parity change functor $\Pi $. It means the above tensor rules should be completed by the copy of them with the obvious action of $\Pi $, which we omit for brevity. All the statements below will be proved for the part of tensor ring (\ref{gl1}), and is identical for its parity change analog. The standard parity for the modules are chosen in the following way. We assume the highest weight of the two dimensional typical module $\mathcal{T}_{e,n}$ ($e\neq 0$) to be grassmann even, as well as the one dimensional atypical module $\mathcal{A}_{n}$, and the top vector of the projective module \mathcal{P}_{n}$ (see Appendix \ref{subgl11}) to be also even. The structure of braided tensor category $(\mathcal{D},\times ,\mathbf{1 ,\lambda ,\rho ,\sigma )$ is defined as follows. The \ bifunctor $\mathcal D\times D\rightarrow D}$ is the tensor product of the modules that was described above.The unit object of $\mathcal{D}$ is $\mathbf{1}=\mathcal{A _{0}$ is simple, and as follows from (\ref{gl1}) the functorial isomorphisms $\lambda :\mathbf{1}\otimes U\tilde{\longrightarrow}U$, $\rho :U\otimes \mathbf{1}\tilde{\longrightarrow}U$ are trivial. Below we will define and prove the existence of invertible associator - functorial isomorphism \alpha _{X,Y,Z}:(X\otimes Y)\otimes Z\tilde{\longrightarrow}X\otimes (Y\otimes Z)$ for any triple of objects $X,Y,Z\in Obj(\mathcal{D})$. This isomorphism is defined using asymptotic solutions of KZ equations. The braiding $\sigma :X\otimes Y\rightarrow Y\otimes X$ of any two objects is defined by $\sigma =Pe^{i\pi \Omega _{12}/\kappa }$ where $P$ is graded permutation. The prove of coherence theorem for associator, i.e. pentagon and triangle relations for monoidal structure becomes standard after the explicit construction of associator, as well as the proof of hexagon relation for braiding. First we briefly recall the monodromy structure and asymptotic solutions of KZ equations for semisimple category of modules. We follow and recapitulate the main steps presented in \cite{Var2}, Section 2. The system of KZ equations can be interpreted as a flat connection in a trivial vector bundle with a fiber $V=V_{1}\otimes ...\otimes V_{N}$, $V_{i}$ are objects of \mathcal{D}$, over the configuration space $X_{N}=\{(z_{1},...,z_{N})\in \mathbb{C} ^{N}\;|\;z_{i}\neq z_{j}\}$. For any path $\gamma :[0,1]\rightarrow X_{N}$ one denotes by $M_{\gamma }:V\rightarrow V$ the operator of holonomy along \gamma $, which can be considered as analytic continuation of KZ equation solutions $\psi (z_{1},...,z_{N})$ along $\gamma $. $M_{\gamma }$ is \mathfrak{g}$-homomorphism since the tensor Casimir operator $\Omega $ of KZ equation is $\mathfrak{g}$-invariant. Operator $M_{\gamma }$ with $\gamma (0)=\gamma (1)=z^{0}=(z_{1}^{0},...,z_{N}^{0})$ is called the monodromy operator. We have such $M_{\gamma }$ as a monodromy representation of the fundamental group $\pi _{1}(X_{N},z^{0})$ in $V$. The dependence on the base point $z^{0}$ can be eliminated by conjugation, because $X_{N}$ is connected. But the fundamental group $\pi _{1}(X_{N})$ is well known -- it is $PB_{N}$ -- pure braid group. Moreover, one can construct the homomorphism of braid group $B_{N}\rightarrow \pi _{1}(X_{N}/S_{N})$ where S_{N}$ is the symmetric group: if we choose the $z^{0}$ such that z_{i}^{0}\in \mathbb{R} $ and $z_{1}^{0}>z_{2}^{0}>...>z_{N}^{0}$ then the action of $b_{i}$ generator of $B_{N}$ on $z^{0}$ corresponds to transposition of $z_{i}^{0}$ and $z_{i+1}^{0}$ (say, $z_{i+1}^{0}$ and $z_{i}^{0}$ exchange their locations such that $z_{i}^{0}$ passes above $z_{i+1}^{0}$). For a fixed base point $z^{0}$ a loop $\gamma $ in $X_{N}/S_{N}$ can be considered as an element of $B_{N}$. Then we can lift it to a path in $X_{N}$ defining the operator $\check{M}_{\gamma }=\sigma M_{\gamma }:V\rightarrow V^{\sigma }$, where $\sigma \in S_{N}$ is the image of $\gamma $ under the map B_{N}\rightarrow S_{N}$ and $V^{\sigma }=V_{\sigma ^{-1}(1)}\otimes ...\otimes V_{\sigma ^{-1}(N)}$. For example, for the $\gamma $ which exchanges $z_{i}^{0}$ and $z_{i+1}^{0}$ we will have $\check{M}_{i}^{\pm }(z^{0})=\check{M}_{\gamma _{i}}^{\pm 1}$. The fact that the operators \check{M}_{i}^{\pm }$ called \textit{half monodromy operators} satisfy the equations \begin{eqnarray*} \check{M}_{i}^{\pm }\check{M}_{i}^{\mp } &=&I, \\ \check{M}_{i}^{\pm }\check{M}_{i+1}^{\pm }\check{M}_{i}^{\pm } &=&\check{M _{i+1}^{\pm }\check{M}_{i}^{\pm }\check{M}_{i+1}^{\pm } \end{eqnarray* follows from the relation $\gamma _{i}\gamma _{i+1}\gamma _{i}=\gamma _{i+1}\gamma _{i}\gamma _{i+1}$ in the fundamental group of $X_{N}/S_{N}$. The (half)monodromy operators being independent on the choice of base point, can be calculated with a specific choice of it. One of the convenient choices of the base point is $z^{0}:z_{1}^{0}\gg z_{2}^{0}\gg ...\gg z_{N}^{0}$. We will need also another choice of the base point for $N=3$ correlation function below. We fix the region $D\subset X_{N}$, D=\{z=(z_{1},...,z_{N})\in \mathbb{R} ^{N}\;|\;z_{1}>...>z_{N}\}$. There is an isomorphism between the space of $V -valued solutions $\Gamma _{f}(D,V_{KZ})$ of the KZ equation in the region D $ and $V$ : for any $z\in D$ the solution $\psi (z)$ is this isomorphism. It is useful to make the following change of variables \begin{align} u_{i}& =\frac{z_{i}-z_{i+1}}{z_{i-1}-z_{i}},\;i=2,...,N-1 \label{35} \\ u_{1}& =z_{1}-z_{2},\;u_{N}=z_{1}+...+z_{N} \notag \end{align All $u_{i}$ are positive on $D$. One can see that $(z_{1},...,z_{N} \rightarrow (u_{1},...,u_{N})$ is one to one map with inverse polynomial mapping, therefore any analytic function $f(z)$ on $D$ can be considered as analytic function of $u$ on some subset $D_{u}\subset \mathbb{C} ^{N}$ - the image of the mapping, and closure of $D_{u}$ contains the origin. The change of variables (\ref{35}) is chosen so that if we have a curve $z(t)$ such that $z(t)\rightarrow 0$ when $t\rightarrow 0$, then the condition $z_{i}(t)/z_{i+1}(t)\rightarrow \infty $ for $i=1,...,N-1$, implies $u_{i}(t)\rightarrow 0$ for $i=1,...,N$. We can define now the limit $\lim_{z_{1}\gg ...\gg z_{N}}f(z)=v$ as a vector which satisfies $\lim_{u_{i}\rightarrow 0}f(u)=v$ for $i=1,...,N$ with $f$ being written in terms of new variables $u_{i}$. We define the asymptotic $f\sim \phi _{1}(z)v$ of a smooth vector valued function $f(z)$ as the $z_{1}\gg ...\gg z_{N}$ limit of $f$ in \ $D$, for some scalar function $\phi _{1}(z)$ and a vector $v\in V$, if \begin{equation} f(z)=\phi _{1}(z)(v+o(z)) \label{36} \end{equation where $o(z)$ considered as a $V$-valued function of $u$ in some neighborhood of the origin is regular and $o(u=0)=0$. We will sometimes put $z_{N}=0$. If $f$ is translation invariant then $\lim_{z_{1}\gg ...\gg z_{N}}f(z)=\lim_{z_{1}\gg ...\gg 0}f(z)$. Another region we need is $D_{0}(z):z_{1}-z_{2}\ll z_{2}-z_{3}\ll ...\ll z_{N-1}-z_{N}$ and as above we define the asymptotic of a function $f(z)$ in the region $D_{0}(z)$ as $f\sim \phi _{0}(z)v$ if $f(z)=\phi _{0}(z)(v+o(z))$ where $o(z)$ considered as a $V$-valued function of $u$ in some neighborhood of the point $u_{i}\rightarrow \infty $. The special case important for the proof of associator existence is $N=3$. The KZ equation takes the form of ordinary differential equation in one variable. In terms of the variables (\ref{35}) $u_{1}=z_{1}-z_{2}$, $u_{2} \frac{z_{2}-z_{3}}{z_{1}-z_{2}}$, $u_{3}=z_{1}+z_{2}+z_{3}$ the KZ equations look like \begin{eqnarray} \kappa \partial _{u_{1}}\psi &=&\frac{\Omega _{12}+\Omega _{13}+\Omega _{23 }{u_{1}}\psi \label{KZ31} \\ \kappa \partial _{u_{2}}\psi &=&\left( \frac{\Omega _{12}}{u_{2}+1}+\frac \Omega _{23}}{u_{2}}\right) \psi \notag \\ \kappa \partial _{u_{3}}\psi &=&0 \notag \end{eqnarray We introduce the function $f$ defined by\footnote This function is well defined with the choice of the branch of logarithm fixed above because the operators $\Omega _{ij}$ acting in the space $V$ have nilpotent non diagonalizable part.} \begin{equation*} \psi (z_{1},z_{2},z_{3})=(z_{1}-z_{3})^{(\Omega _{12}+\Omega _{13}+\Omega _{23})/\kappa }f\left( \frac{z_{1}-z_{2}}{z_{1}-z_{3}}\right) \end{equation* Using the fact that all $\Omega _{ij}$ commute with $\Omega _{12}+\Omega _{13}+\Omega _{23}$ one can see by direct calculation that f=u_{1}^{-(\Omega _{12}+\Omega _{13}+\Omega _{23})/\kappa }\psi $ depends only on $x=\frac{1}{u_{2}+1}$ and is $u_{1},u_{3}$ independent. Thus we get one ODE for the $V$-valued function $f(x)$ \begin{equation} \kappa \partial _{x}f(x)=\left( \frac{\Omega _{12}}{x}+\frac{\Omega _{23}} x-1}\right) f(x) \label{KZf} \end{equation The asymptotic regions $D_{0}(z),D_{1}(z)$ correspond to $x\rightarrow 0$ and $x\rightarrow 1$ respectively. The existence of asymptotic solutions of KZ equation as they are defined above is the main tool for the proof of existence of associator. \begin{theorem} \label{thm1} Let $V=V_{1}\otimes V_{2}\otimes V_{3}$ where $\{V_{i}\}$ -- any combination from the set $\{\mathcal{A},\mathcal{P},\mathcal{T}\}$. If e_{i}\notin \mathbb{Z} $\textbf{\ }and $e_{1}+e_{2}\notin \mathbb{Z} \backslash \{0\}$ in the case $V_{i}=\mathcal{T}$ $,i=1,2$\ in $V$, then for every eigenvector $v\in V$ of $\Omega _{12}$ there exists unique asymptotic solution of (\ref{KZf}) around 0 corresponding to $v$ and this correspondence gives isomorphism $\phi _{0}:\Gamma _{f}(D,V_{KZ})\rightarrow V$. \end{theorem} \textit{Proof. }The proof is based on straightforward linear algebra manipulations which we moved to Appendix A.\textit{\ }We apply Lemma \re {lemma: A1} or \ref{lemma: A2} (see Appendix A), considering all possible 6 combinations (up to a permutation) of $V_{1},V_{2}$: $\mathcal{T _{e_{1},n_{1}}\otimes \mathcal{T}_{e_{2},n_{2}}$, $\mathcal{T _{e_{1},n_{1}}\otimes \mathcal{P}_{n_{2}}$, $\mathcal{P}_{n_{1}}\otimes \mathcal{P}_{n_{2}}$, $\mathcal{T}_{e_{1},n_{1}}\otimes \mathcal{A}_{n_{2}} , $\mathcal{P}_{e_{1},n_{1}}\otimes \mathcal{A}_{n_{2}}$, $\mathcal{A _{n_{1}}\otimes \mathcal{A}_{n_{2}}$. The explicit form of the function solution $\phi (x)$ is not important at this point, but one can find it in the Appendix A. All we have to do is to check, case by case, the applicability of Lemmas \ref{lemma: A1},\ref{lemma: A2}. Isomorphism to the space $\Gamma _{f}(D,V_{KZ})$ of KZ solution follows by linearity. The following data is obtained by direct diagonalization of $\Omega _{12}$ on the basis of $V_{1}\otimes V_{2}$. 1. $\mathcal{T}_{e_{1},n_{1}}\otimes \mathcal{T}_{e_{2},n_{2}}$. When $e_{2}+e_{1}\notin \mathbb{Z} $ there are no Jordan blocks and the eigenvalues are $\lambda _{1}=\delta _{12}^{++}$, $\lambda _{2}=\delta _{12}^{--}$, with two eigenvectors for each of them. Here and below $\delta _{ij}^{\alpha \beta }=e_{i}e_{j}+e_{i}(n_{j}+\beta /2)+e_{j}(n_{i}+\alpha /2)$. The difference \lambda _{1}-\lambda _{2}=e_{1}+e_{2}\notin \mathbb{N} $ and by the Lemma \ref{lemma: A1} there are four asymptotic solutions for four different eigenvectors. When $e_{2}+e_{1}=0$ there is one eigenvalue $e_{1}(n_{2}-n_{1})-e_{1}^{2}$ with two eigenvectors without Jordan block and two other ones with Jordan block of size 2. By the Lemma \ref{lemma: A2} there are four asymptotic solutions. We cannot prove existence of asymptotic solutions using Lemma \ref{lemma: A1} in the case $e_{2}+e_{1}\in \mathbb{Z} \backslash \{0\}$, but this case, from the perspective of affine Lie superalgebra, exactly corresponds to what we call non generic case of representations \cite{CRbeyond}. 2. $\mathcal{T}_{e_{1},n_{1}}\otimes \mathcal{P}_{n_{2}}$ The set of eigenvalues are $\lambda _{1}=e_{1}(n_{2}-1)$ and $\lambda _{2}=e_{1}(n_{2}+1)$ with the difference $2e_{1}\notin \mathbb{N} $. Each of them \ correspond to two eigenvectors without Jordan block and one Jordan block of size 2. By the Lemmas \ref{lemma: A1},\ref{lemma: A2} there are asymptotic solutions for each eigenvector. 3. $\mathcal{P}_{n_{1}}\otimes \mathcal{P}_{n_{2}}$ There is one eigenvalue $\lambda =0$ with the following structure of eigenvectors: there are 3 Jordan blocks of rank 2, one Jordan block of rank 3 and 7 eigenvectors without Jordan block structure. Again the condition \lambda \mathbb{N} $ is not an eigenvalue is satisfied, therefore by Lemmas \ref{lemma: A1},\re {lemma: A2} there are asymptotic solutions corresponding to each eigenvector. 4. $\mathcal{T}_{e_{1},n_{1}}\otimes \mathcal{A}_{n_{2}}$ There is one eigenvalue $\lambda =e_{1}n_{2}$ with two different eigenvectors without a Jordan block. Lemma \ref{lemma: A1} is applicable. 5. $\mathcal{P}_{n_{1}}\otimes \mathcal{A}_{n_{2}}$ There is one eigenvalue $\lambda =0$ with four different eigenvectors without a Jordan block. Lemma \ref{lemma: A1} is applicable. 6. $\mathcal{A}_{n_{1}}\otimes \mathcal{A}_{n_{2}}$ There is one eigenvalue $\lambda =0$ with one eigenvector. Lemma \ref{lemma: A1} is applicable. $\square $ \begin{theorem} \label{thm2} The same claim as in the Theorem \ref{thm1}, with the same restrictions on the parameters of typical modules $\mathcal{T}$ appearing as $V_{i},i=2,3$ in $V$, is valid for existence and uniqueness of asymptotic solutions of KZ equation (\ref{KZf}) around $x=1$. \end{theorem} \textit{Proof.} The proof is based on the Lemma \ref{lemma: A3} (see Appendix A) that replaces the Lemmas \ref{lemma: A1},\ref{lemma: A2} in the proof of Theorem \ref{thm1}. $\square $ As we see, there are specific cases $2e_{i}\in \mathbb{Z} $ and $e_{1}+e_{2}\in \mathbb{Z} \backslash \{0\}$ for parameters of typical representations when we are not able to guarantee the existence and uniqueness of asymptotic solutions by Lemmas \ref{lemma: A1},\ref{lemma: A2},\ref{lemma: A3}. We notice that for affine $\widehat{\mathfrak{gl}}(1|1)$ (where the we always can put $\kappa =k=1$) these cases correspond to reducibility of the induced affine modules, and as we said above, we exclude these cases in the process of derivation of KZ equation. \begin{proposition} With the restrictions on the parameters of typical modules as in the Theorem \ref{thm1} there is an isomorphisms of the spaces \begin{equation} \alpha _{1,2,3}:(V_{1}\otimes V_{2})\otimes V_{3}\tilde{\longrightarrow \Gamma _{f}(D,V_{KZ})\tilde{\longrightarrow}V_{1}\otimes (V_{2}\otimes V_{3}) \label{asoc} \end{equation which will serve the associator in the Drinfeld tensor category. \end{proposition} \textit{Proof. }The first isomorphism $\phi _{0}:(V_{1}\otimes V_{2})\otimes V_{3}\tilde{\longrightarrow}\Gamma _{f}(D,V_{KZ})$ is defined by the correspondence between the eigenvectors of $\Omega _{12}$ in $V$ and asymptotic solutions of KZ equation (\ref{KZf}) around $x=0$ established by the Theorem \ref{thm1}. The second isomorphism $\phi _{1}^{-1}:\Gamma _{f}(D,V_{KZ})\tilde{\longrightarrow}V_{1}\otimes (V_{2}\otimes V_{3})$ is the inverse of the isomorphism $\phi _{1}$ established by the Theorem \re {thm2}. $\square $ \textbf{Remark 1. }One can easily see that the associator (\ref{asoc}) is trivial (equal to 1) when one of the spaces $V_{i},i=1,2,3$ is one dimensional, as for example in the cases 4,5,6 of the proof of the Theorem \ref{thm1}. \begin{theorem} \label{thm3} For any quadruple of objects $V_{i},i=1,...,4$ in the \mathfrak{gl}(1|1)$ Drinfeld category $\mathcal{D}$, with the restrictions on the parameters of typical modules $e_i\notin\mathbb{Z}$, $e_{i}+e_{j}\notin \mathbb{Z} \backslash \{0\}$ for any pair $\mathcal{T}_{e_{i},n_{i}},\mathcal{T _{e_{j},n_{j}}$, the isomorphism $\alpha _{1,2,3}$ (\ref{asoc}) satisfies pentagon equation $((V_{1}\otimes V_{2})\otimes V_{3})\otimes V_{4}\longrightarrow V_{1}\otimes (V_{2}(\otimes V_{3}\otimes V_{4})) \begin{equation} \alpha _{id_{1}\otimes 2,3,4}\circ \alpha _{1,2\otimes 3,4}\circ \alpha _{1,2,3\otimes Id_{4}}=\alpha _{1,2,3\otimes 4}\circ \alpha _{1\otimes 2,3,4} \label{pen} \end{equation} \end{theorem} The proof is based on decomposition of pentagon diagram into triangle ones, and each triangle is a commutative diagram which includes as a part the isomorphism (\ref{asoc}). The proof uses only the fact of existence and uniqueness of invertible associator irrespectively of details of its construction from asymptotic solutions. We refer to the books \cite{BK}, p.25, or \cite{ChP}, p.545 for details of the proof, which is independent on concrete form of asymptotic solutions but only on the fact of their existence. $\square $ Recall the standard derivation of braiding $\sigma _{X,Y}$ from half-monodromy of KZ solutions (See \cite{ChP} Section 16.2 and original references therein.) Since the solution of KZ equations for $N=2$ is a function of difference $z_{2}-z_{1}$, one can represent the braid group B_{2}$ generator $\sigma _{12}$ which swaps $z_{1}$ and $z_{2}$, z_{1},z_{2}\in D\subset \mathbb{C} ^{2}$ by the loop contour $\overrightarrow{z}(s)=(z_{1}(s),z_{2}(s))$, z_{1,2}(s)=a+be^{i\pi s}$, $a=(z_{1}+z_{2})/2$, $b=(z_{1}-z_{2})/2$, parametrized by $s\in \lbrack 0,1]$. It satisfies $\overrightarrow{z (0)=z_{1}$, $\overrightarrow{z}(1)=z_{2}$. A pull back of the KZ $N=2$ equation written for a one form $dw$ along this contour leads to the equation \begin{equation} \frac{dw}{ds}=\frac{\Omega _{12}}{\kappa }w(s) \label{br1} \end{equation with the solution \begin{equation} w(s)=e^{\frac{\Omega _{12}}{\kappa }s}w(0) \label{brsol} \end{equation As before the exponent is understood here as classical series $\sum \left( \frac{\Omega _{12}}{\kappa }s\right) ^{n}\frac{1}{n!}$, which converges on Aut(V_{1}\otimes V_{2})$ because of the nilpotency of non diagonal part of \Omega _{12}$ acting on any tensor product of vectors. Therefore if we put s=1$ in the last equation we get the monodromy representation of braid group \begin{equation} \rho _{N=2}(\sigma _{12})(v_{1}\otimes v_{2})=Pe^{\frac{\Omega _{12}}{\kappa }}(v_{1}\otimes v_{2}) \label{rosol} \end{equation It is straight forward now to generalize this representation of braiding through half-monodromy of KZ solution to $N>2$. \begin{equation} \rho _{N}(\sigma _{i,i+1})(v_{1}\otimes ...\otimes v_{N})=P_{i,i+1}e^{\frac \Omega _{i,i+1}}{\kappa }}(v_{1}\otimes ...\otimes v_{N}) \label{rosol2} \end{equation} \begin{theorem} \label{thm4} For any triple of objects $V_{1},V_{2},V_{3}$ in the Drinfeld category $\mathcal{D}$ with the restrictions on parameters of $V_{i}=$ \mathcal{T}_{e_{i},n_{i}}$ as above, associator $\alpha _{1,2,3}$ and braiding $\sigma _{1,2}:V_{i}\otimes V_{j}\longrightarrow V_{j}\otimes V_{i} , $\sigma _{1,2}=P\exp (i\pi \Omega _{12}/\kappa )$ where $P$ is super permutation of spaces, satisfy the hexagon relation $(V_{1}\otimes V_{2})\otimes V_{3}\longrightarrow V_{2}\otimes (V_{3}\otimes V_{1}) \begin{equation} \alpha _{2,3,1}\circ \sigma _{1,2\otimes 3}^{\pm 1}\circ \alpha _{1,2,3}=(Id_{2}\otimes \sigma _{1,3}^{\pm 1})\circ \alpha _{2,1,3}\circ (\sigma _{1,2}^{\pm 1}\otimes Id_{3}) \label{hex} \end{equation Moreover the half monodromy operators $\check{M}_{1}$ acting on V_{1}\otimes (V_{2}\otimes V_{3})$ defined above coincide with $\alpha _{1,2,3}^{-1}\sigma _{12}\alpha _{1,2,3}$. \end{theorem} The existence of the universal form of the representation of braiding (\re {rosol}), (\ref{rosol2}) allows to apply the same proof as in the case of semisimple categories. We refer to \cite{ChP}, p.547 for details of the proof, which follows \cite{Dr0}, \cite{Ko}. There is an interesting explicit representation of the associator written in terms of P-exponential. It was suggested by Drinfeld and a proof that this is indeed an associator can be found in \cite{Saf \begin{equation} \alpha _{1,2,3}=\lim_{t\rightarrow 0}\left[ t^{-\Omega _{23}/\kappa }P\exp \left( \frac{1}{\kappa }\int\limits_{t}^{1-t}\left( \frac{\Omega _{12}}{z} \frac{\Omega _{23}}{z-1}\right) dz\right) t^{\Omega _{12}/\kappa }\right] \label{asocPexp} \end{equation Unfortunately even in the case of $\mathfrak{gl}(1|1)$ superalgebra an explicit calculation of this expression is hard and leads to a complicated series and interesting algebraic structure \cite{Fur} which we will not discuss here. Braided tensor structure of this category is standard for modules category of quasitriangular Hopf algebra: trivial unit object, trivial associator and unit morphisms, and braiding morphisms $\sigma _{V,W}=PR_{V,W}$ where $P$ is super permutation. The proof is standard, and doesn't refer to any particular data and we refer to textbooks, for example to \cite{ChP}. For the correspondence with the Drinfeld category we mention the functorial isomorphism $\beta _{X,Y,Z}^{\pm }:X\otimes (Y\otimes Z)\rightarrow Y\otimes (X\otimes Z)$ defined by \begin{equation} \beta _{X,Y,Z}^{\pm }=\alpha (\sigma _{XY}^{\pm 1}\otimes Id_{Z})\alpha ^{-1} \label{qmonodr} \end{equation It satisfies \begin{equation} \beta _{X,Y,Z}^{\pm }\beta _{Y,X,Z}^{\mp }=Id \label{qmon1} \end{equation Then the functorial isomorphisms \begin{eqnarray} \beta _{12}^{\pm } &=&\beta _{X,Y,Z\otimes U}^{\pm }:X\otimes (Y\otimes (Z\otimes U))\rightarrow Y\otimes (X\otimes (Z\otimes U)), \label{qmon2} \\ \beta _{23}^{\pm } &=&Id_{X}\otimes \beta _{Y,Z,U}^{\pm }:X\otimes (Y\otimes (Z\otimes U))\rightarrow X\otimes (Z\otimes (Y\otimes U)) \notag \end{eqnarray satisfy the relation \begin{equation} \beta _{12}^{\pm }\beta _{23}^{\pm }\beta _{12}^{\pm }=\beta _{23}^{\pm }\beta _{12}^{\pm }\beta _{23}^{\pm } \label{qmon3} \end{equation} We can summarise the construction of Drinfeld category by the following proposition based on the Theorems \ref{thm1},\ref{thm2},\ref{thm3},\ref{thm4 . \begin{proposition} \label{prop1} The category $\mathcal{D}$ of typical, atypical and projective $\mathfrak{gl}(1|1)$-modules with the restrictions on typicals with $e_i\notin\mathbb{Z}$, $(e_{i}+e_{j})/\kappa \notin \mathbb{Z} \backslash \{0\}$ is braided tensor category with the structures as described above. \end{proposition} With these structures category $\mathcal{D}$ of $\mathfrak{gl}(1|1)$-modules will be considered as category of modules of the algebra denoted by \mathcal{A}_{\mathfrak{g},\Omega }$, ($\mathfrak{g=gl}(1|1)$). \section{Category $\mathcal{C}_{\protect\kappa }$ of $U_{h}(\mathfrak{gl (1|1))$-modules} \label{section4} We denote $i\pi \kappa ^{-1}=h$. The structure of quasitriangular $h$-adic Hopf superalgebra $A=U_{h}(\mathfrak{gl}(1|1))$, $\kappa \in \mathbb{R} ^{\times }$, is defined by the following commutation relations of its generators $\psi ^{\pm },N,E$ \begin{equation*} \{\psi ^{+},\psi ^{-}\}=2\sinh (hE) \end{equation* \begin{equation} \lbrack N,\psi ^{\pm }]=\pm \psi ^{\pm },\;(\psi ^{+})^{2}=(\psi ^{-})^{2}=0,\;[E,X]=0\;\forall X\in U_{h}(\mathfrak{gl}(1|1)) \notag \end{equation where $\exp (\pm Eh)$ is understood as its Taylor series around $h=0$ ( \kappa =\infty $). The Hopf algebra structure is defined as follows. Coproduc \begin{eqnarray} \overline{\Delta }(E) &=&E\otimes I+I\otimes E,\;\overline{\Delta (N)=N\otimes I+I\otimes N, \label{qcop} \\ \overline{\Delta }(\psi ^{+}) &=&\psi ^{+}\otimes e^{Eh/2}+e^{-Eh/2}\otimes \psi ^{+},\;\overline{\Delta }(\psi ^{-})=\psi ^{-}\otimes e^{Eh/2}+e^{-Eh/2}\otimes \psi ^{-}, \notag \end{eqnarray couni \begin{equation} \epsilon (E)=\epsilon (N)=\epsilon (\psi ^{\pm })=0, \label{qco1} \end{equation and antipod \begin{eqnarray} \gamma (E) &=&-E,\;\gamma (N)=-N, \label{qanti} \\ \gamma (\psi ^{+}) &=&-e^{Eh/2}\psi ^{+},\;\gamma (\psi ^{-})=-\psi ^{-}e^{-Eh/2}, \notag \end{eqnarray The algebra $U_{h}(\mathfrak{gl}(1|1))$ is quasitriangular. One can choose the universal R-matrix $\overline{R}:A\otimes A\rightarrow A\otimes A$ in the for \begin{equation} \overline{R}=\exp [h(E\otimes E+E\otimes N+N\otimes E)](1-e^{Eh/2}\psi ^{+}\otimes e^{-Eh/2}\psi ^{-}) \label{qRuni} \end{equation It satisfies the standard quasitriangular Hopf algebra relation \begin{eqnarray} \overline{R}\overline{\Delta }(X) &=&\overline{\Delta }^{op}(X)\overline{R ,\;\forall X\in A \label{qRrel} \\ (\overline{\Delta }\otimes Id)\overline{R} &=&\overline{R}_{13}\overline{R _{23}, \notag \\ (Id\otimes \overline{\Delta })R &=&\overline{R}_{13}\overline{R}_{12}, \notag \end{eqnarray} As any quasitriangular Hopf superalgebra $U_{h}(\mathfrak{gl}(1|1))$ induces braided tensor category structure on the category of finite dimensional modules provided the latter is closed under the tensor product functor. \begin{proposition} Restrictions on $\kappa $ and parameters $e$ of typical modules $e_i\notin\mathbb{Z}$, $e_{i}+e_{j}\notin \mathbb{Z} \backslash \{0\}$ are enough for the category $\mathcal{C}_{\kappa }$ of (equivalence classes of) the modules $\mathcal{T}_{e,n}^{\kappa },\mathcal{P _{n}^{\kappa },\mathcal{A}_{n}^{\kappa }$ to form a tensor product ring isomorphic to the tensor product ring (\ref{gl1}) of the modules $\mathcal{T _{e,n},\mathcal{P}_{n},\mathcal{A}_{n}$. (See Appendix B \ref{appB} for definition of the tensor category $\mathcal{C}_{\kappa }$ in a specified basis.) \end{proposition} We check this by direct calculation in Appendix B \ref{appB} using explicit basis of three types of representations. It is shown that with the restrictions on parameters mentioned in the theorem the same tensor product decomposition works in the quantum case, and the tensor rings are isomorphic. \section{Proof of braided tensor equivalence} \label{section5} The main result of this paper is the following theorem. \begin{theorem} \label{thm7} The categories of modules $\mathcal{D}$ and $\mathcal{C _{\kappa }$ with the restrictions on objects of typical modules $e_i/\kappa\notin\mathbb{Z}$, $e_i/\kappa + e_j/\kappa\notin\mathbb{Z}\backslash \{0\}$ are braided tensor equivalent categories. \end{theorem} Since our proof of this theorem follows \cite{Gee} , we have change the approach to KZ equation to a more general one used in \cite{Gee}. In stead of the KZ equation (\ref{n16}) for correlation functions $\psi $ built on intertwiners of $\mathfrak{g}$-modules consider the equation - \ we will call it KZ$_{\mathfrak{g}}$- one can consider KZ equation for superalgebra valued element $\omega \in (U(\mathfrak{g}))^{\otimes N}[[h]]$ of the form \begin{equation} \frac{1}{h}\partial _{i}\omega =\sum_{j\neq i=1}^{N}\frac{\Omega _{ij}} z_{i}-z_{j}}\omega ,\;i=1,...,N \label{n16a} \end{equation} This gives rise to the topologically free quasitriangular quasi-Hopf superalgebra $\mathcal{A}_{\mathfrak{g},\Omega }$ with the braiding defined as $e^{h\Omega }$ and the coassociator defined by the monodromy of solutions of the equation (\ref{n16a}). \ We refer to the standard description of this algebra in \cite{Dr1} for non-super case, and to it straight forward generalization for the super case \cite{Gee} \ Section 4. The Drinfeld category $\mathcal{D}$ is a category of topologically free modules over \mathcal{A}_{\mathfrak{g},\Omega }$. The equivalence partner for the algebra $\mathcal{A}_{\mathfrak{g},\Omega }$ is the Drinfeld-Jimbo $h$-adic quantum superalgebra $U_{h}(\mathfrak{g})$. Its structure in our specific case was described in the previous section. The theorem proved in \cite{Dr2}, which can be modified to the superalgebra case at hand, claims that if two topological algebras \mathcal{A}_{\mathfrak{g},\Omega }$, $U_{h}(\mathfrak{g})$ are gauge equivalent (we will explain what it means below), then the categories of their topologically free modules of finite rank are braided tensor equivalent. Therefore it is enough for us to show gauge equivalence of the two superalgebras. The plan of this section is the following. We start with recalling the standard proof of the gauge equivalence in the case of simple Lie algebras which one can find in Drinfeld's paper \cite{Dr3} and explain why it is in general not applicable in the case of superalgebras. After that we explain the details of Geer's proof \cite{Gee} of gauge equivalence which avoids the points of Drinfeld's proof problematical for superalgebras, but applicable for superalgebras of types $A-G$. At the end we argue why a proof found by Geer for classical superalgebras of types $A-G$ works also for $\mathfrak{gl (1|1)$ case. \textbf{The Drinfeld's proof.} We recall a proof of braided tensor equivalence of $U_{ih}(\mathfrak{g})$ and $\mathcal{A}_{\mathfrak{g},\Omega } $ for $\mathfrak{g}$ -- \textit{non-super} Lie algebra \cite{Dr3}, (see also the Section 16 of \cite{ChP}). This proof is based on the proof of existence of the invertible element $\mathcal{F}_{h}\in (U(\mathfrak{g )\otimes U(\mathfrak{g}))[[h]]$ which implements the twist of the structures of the algebra $U(\mathfrak{g})[[h]]$ to the structures of $\mathcal{A}_ \mathfrak{g},\Omega }$. The algebra $U_{ih}(\mathfrak{g})$ is isomorphic \textit{as } \mathbb{C} \lbrack \lbrack h]]$ \textit{algebra }to $U(\mathfrak{g})[[h]]$. First, one obtains the algebra $(U(\mathfrak{g})\otimes U(\mathfrak{g}))[[h]]$ from $U \mathfrak{g})[[h]]$ by application of the composite homomorphism \begin{equation*} \widetilde{\Delta }_{h}:U(\mathfrak{g})[[h]]\tilde{\rightarrow}U_{ih} \mathfrak{g})\rightarrow \overline{\Delta }\rightarrow U_{ih}(\mathfrak{g )\otimes U_{ih}(\mathfrak{g})\tilde{\rightarrow}(U(\mathfrak{g})\otimes U \mathfrak{g}))[[h]] \end{equation* If one requires that $\widetilde{\Delta }_{h}=\Delta (\bmod h)$ where \Delta $ is the usual comultiplication in $U(\mathfrak{g})$, then using the fact that $H^{1}(\mathfrak{g},U(\mathfrak{g})\otimes U(\mathfrak{g}))=0$ for simple Lie algebras, one gets that there must exist $\mathcal{F}_{h}\in (U \mathfrak{g})\otimes U(\mathfrak{g}))[[h]]$ such that \begin{equation*} \mathcal{F}_{h}\equiv 1\otimes 1(\bmod h) \end{equation* and \begin{equation} \mathcal{F}_{h}^{-1}\overline{\Delta }(x)\mathcal{F}_{h}=\widetilde{\Delta _{h}(x),\text{ }\forall x\in U(\mathfrak{g}) \label{tw0} \end{equation Let the image of the universal R-matrix $\overline{R}$ of $U_{ih}(\mathfrak{ })\approxeq U(\mathfrak{g})[[h]]$ in $(U(\mathfrak{g})\otimes U(\mathfrak{g ))[[h]]$ under $\widetilde{\Delta }_{h}$ be $\widetilde{R}$. The quasitriangular Hopf algebra $U(\mathfrak{g})[[h]]$ with trivial coassociator, the coproduct $\widetilde{\Delta }_{h}$ and the R-matrix \widetilde{R}$ can now be twisted by the element $\mathcal{F}_{h}$, giving quasitriangular quasi-Hopf algebra $U(\mathfrak{g})[[h]]$ with different comultiplication, different R-matrix and non trivial coassociator. We would like them to be the same as of the algebra $\mathcal{A}_{\mathfrak{g},\Omega }$, i.e $\Delta $ - the trivial coproduct of $U(\mathfrak{g})$. The standard properties of quasitriangular quasi-Hopf algebras are used to prove that all three structures can fit to the required ones of $\mathcal{A}_{\mathfrak{g ,\Omega }$ using the existing twist element $\mathcal{F}_{h}$. This element implements what we called above the gauge equivalence. Explicitly the twist equations ar \begin{equation} (\epsilon \otimes id)\mathcal{F}_{h}\mathcal{=}(id\otimes \epsilon )\mathcal F}_{h}=1 \label{tw00} \end{equation \begin{equation} \mathcal{F}_{h}^{-1}\overline{\Delta }(x)\mathcal{F}_{h}=\Delta (x) \label{tw1} \end{equation \begin{equation} (\mathcal{F}_{h})_{21}^{-1}\overline{R}_{12}(\mathcal{F}_{h})_{12}=R_{12}, \label{tw2} \end{equation \begin{equation} (\mathcal{F}_{h})_{23}(1\otimes \Delta )(\mathcal{F}_{h}).\alpha .[(\mathcal F}_{h})_{12}(\Delta \otimes 1)(\mathcal{F}_{h})]^{-1}=1\otimes 1\otimes 1 \label{tw3} \end{equation} Therefore the braided equivalence prove is reduced to a proof of existence of invertible $\mathcal{F}_{h}$ which satisfies the equations (\ref{tw1}) - \ref{tw3}). The equation (\ref{tw3}) is the most important one. However explicit solution of the equations (\ref{tw1}) - (\ref{tw3}) is a very hard problem, which requires an explicit form of associator. All we are able to do in this context is to prove its existence, in a way described above. One of the problems to repeat these arguments of twist $\mathcal{F}_{h}$ existence for a superalgebra case, is that the vanishing of the first cohomology $H^{1}(\mathfrak{g},U(\mathfrak{g})\otimes U(\mathfrak{g}))=0$ used above doesn't not hold in general for superalgebras, in particular for \mathfrak{g}=\mathfrak{gl}(1|1)$, (see for example \cite{Moe}). One should look for a way which avoids the cohomology vanishing arguments. One of such ways was suggested by N. Geer \cite{Gee} by superalgebra modification of quantization procedure suggested by Etingof and Kazhdan (EK) \cite{EK1} - \cite{EK3}\footnote The complete list of relevant sequel of their papers is longer, but the others will not be used in our discussion below.}. \textbf{The Geer's proof.} We start from an concise outline and the main steps of the proof in \cit {Gee} and then provide some details. The EK construction \cite{EK1} includes two algebras -- the algebra $H=U_{h}(D(\mathfrak{g}))$, the quantisation of quantum double, and the algebra $\overline{U}_{h}(\mathfrak{g})$. The latter is a quantum double which in general admits a non trivial $h$-adic topolog \footnote In \ \cite{EK1} the $\overline{U}_{h}(\mathfrak{g})$ quantisation was built to be applied to infinite dimensional Lie algebras, when the quantization $H$ doesn't cannot be applied, but both quantizations work for finite dimensional Lie algebras.}. Both quantizations are generalized in the Sections 5,6,7 for the types $A-G$ superalgebras in \cite{Gee} and shown to be equivalent. The main features which make this way of quantization effective is \textit{commutativity with quantum double}, by construction, of the first quantisation, and \textit{functoriality} of the second (Section 8 of \cite{Gee}). The next step of the proof (\cite{Gee}, Section 9) is isomorphism of the two equivalent EK quantizations $U_{h}(\mathfrak{g})$ to the standard Drinfeld-Jimbo quantization $U_{h}^{DJ}(\mathfrak{g})$. The proof follows \cite{EK3} where the assertion was proved for non-super Kac-Moody algebras case but works for finite Lie algebras as well. The superalgebra case requires explicit check of additional Serre relations typical for the most of the quantum superalgebras of types $A-G$. The final steps of the proof are in the Section 10. (All the references below are to the sections and equations of the paper \cite{Gee}.) If there is a gauge isomorphism $\alpha $ between two quasitriangular quasi-superbialgebras then it induces braided tensor equivalence between their modules (Theorem 47). Using the previous results on quantization of double with explicit form of the twist (eq. (32)) one leads to the conclusion that $U_{h}(\mathfrak{g})$ is a gauge twist $(\mathcal{A}_ \mathfrak{g},\Omega })_{\mathcal{F}}$ of $\mathcal{A}_{\mathfrak{g},\Omega } . The collection of these assertions finally leads to the required conclusion that the categories of topologically free $\mathcal{A}_{\mathfrak g},\Omega }$ and $U_{h}(\mathfrak{g})$ finite dimensional modules are braided tensor equivalent (Theorem 48). Now we explain some details of the proof steps described above and point out specific features of these steps in our $\mathfrak{gl}(1|1)$ case at the end. The superalgebra $\mathcal{A}_{\Omega ,\kappa }$ is topologically free quasitriangular quasi-Hopf superalgebra built from $\mathfrak{g}$, and the Drinfeld category of modules $\mathcal{D}_{\mathfrak{g}}$ is braided tensor category of its modules with the structures based on the KZ equation, as described above. Let $\mathfrak{g}_{+}$ be a finite dimensional superbialgebra and $\mathfrak g}=D(\mathfrak{g}_{+})$ be its double. In Section 5, following \cite{EK1}, Verma modules $M_{\pm }=U(\mathfrak{g})\otimes _{U(\mathfrak{g}_{\pm })}c_{\pm }$ over $\mathfrak{g}$ are used in order to construct forgetful functor $F$ from the Drinfeld category $\mathcal{D}_{\mathfrak{g}}$ to the category of topologically free \mathbb{C} \lbrack \lbrack h]]$-modules $\mathcal{A}$ \begin{equation} F(V)=Hom_{\mathcal{D}_{\mathfrak{g}}}(M_{+}\otimes M_{-},V) \label{Fn} \end{equation The Theorem 12 asserts that it is a tensor functor. More precisely, there exists a family of isomorphisms $\mathcal{J}_{V,W}$, $V,W\in \mathcal{D}_ \mathfrak{g}}$ such tha \begin{equation} \mathcal{J}_{U\otimes V,W}\circ (\mathcal{J}_{U,V}\otimes 1)=\mathcal{J _{U,V\otimes W}\circ (1\otimes \mathcal{J}_{V,W}) \label{FcPr} \end{equation namely \begin{equation} \mathcal{J}_{V,W}(v\otimes w)=(v\otimes w)\circ \alpha _{1,2,34}^{-1}(1\otimes \alpha _{2,3,4})\circ \beta _{23}\circ (1\otimes \alpha _{2,3,4}^{-1})\circ \alpha _{1,2,34}\circ (i_{+}\otimes i_{-}) \label{FcDef} \end{equation Here $i_{\pm }$ is a coproduct defined on the highest (lowest) weights of the Verma modules as $i_{\pm }(v_{\pm })=v_{\pm }\otimes v_{\pm }$ and \beta $ is the morphism given by $\tau e^{\Omega \kappa /2}$. Theorem 12 with a proof copied from \cite{EK1} asserts that $\mathcal{J}_{V,W}$ together with $F$ is a tensor functor. The functor $F$ can be thought of as a forgetful functor $F(V):V\rightarrow Hom_{\mathcal{D}_{\mathfrak{g}}}(U \mathfrak{g}),V)$. Being a tensor functor it induces a bialgebra structure on the target. Moreover, it induces superbialgebra structure on $U(\mathfrak g})[[h]]$ and give rise to a Hopf algebra $H$ with structure isomorphic to a twist $\mathcal{F}\in U(\mathfrak{g})^{\otimes 2}[[h]]$ determined by \mathcal{J}_{V,W}$ (eq. 32) of the usual structure of $U(\mathfrak{g})[[h]] . Its R-matrix $R=(\mathcal{F}^{op})^{-1}e^{\kappa \Omega /2}\mathcal{F}$. This $R$ is polarized, i.e. $R\in U_{h}(\mathfrak{g}_{+})\otimes U_{h} \mathfrak{g}_{-})$. The final assertion of this part (Theorem 17) is that $H$ is a quantization of superbialgebra $\mathfrak{g}$. Two important features of this construction is that $U_{h}(\mathfrak{g}_{\pm })$ are closed under coproduct, and that this quantization commutes with taking the double: D(U_{h}(\mathfrak{g}_{+}))\cong U_{h}(\mathfrak{g}_{+})\otimes U_{h} \mathfrak{g}_{-})=H$ (Corollary 23). \ We refer to the Section 5 of \cit {Gee} for details of this part of the proof steps. The construction of this first EK quantisation can be preserved in our \mathfrak{gl}(1|1)$ case. The Verma modules in our notations isomorphic to the typical modules $\mathcal{T}_{e,n}$. The atypical modules are the quotients of $\mathcal{T}_{0,n}$, and the projectives $\mathcal{P}_{n}$ can be identified in this construction with $M_{+}\otimes M_{-}\cong \mathcal{T _{e,n_{1}}\otimes \mathcal{T}_{-e,n_{2}}$ , \ $e\neq 0$, $n_{1}+n_{2}=n$. The second EK quantization is in a sense "dual" to the first EK quantization. For a finite dimensional superbialgebra with a discrete topology (given by inverse limit of finite dimensional topological superspaces) its modules are topological superspaces. For such topological modules of Drinfeld double one defines the dual Drinfeld category $\mathcal{ }_{\mathfrak{g}}^{t}$ with these \ modules as objects and morphisms $Hom_ \mathcal{D}_{\mathfrak{g}}^{t}}(V,W)=$ $Hom_{\mathfrak{g}}(V,W)[[h]]$. \ One can also dualize $i_{\pm }^{\ast }$ the maps $i_{\pm }$. Similarly functor \overline{F}:\mathcal{D}_{\mathfrak{g}}^{t}\rightarrow \mathcal{A}^{t}$ from dual Drinfeld category to a symmetric tensor category $\mathcal{A}^{t}$ of \mathbb{C} \lbrack \lbrack h]]$ modules with continuous maps as morphisms, can be defined now by $\overline{F}(V)=Hom_{\mathcal{D}_{\mathfrak{g }}(M_{-},M_{+}^{\ast }\otimes V)$. \ Similarly to the first EK quantization the Theorem 26 asserts that together with the isomorphism \begin{equation} \overline{\mathcal{J}}_{V,W}(v\otimes w)=(i_{+}^{\ast }\otimes 1\otimes 1)\circ \alpha _{1,2,34}^{-1}(1\otimes \alpha _{2,3,4})\circ \beta _{23}^{-1}\circ (1\otimes \alpha _{2,3,4}^{-1})\circ \alpha _{1,2,34}\circ (v\otimes w)\circ i_{-} \label{FcDefd} \end{equation it defines a tensor structure on $\overline{F}$. Further steps of the second EK quantization are parallel to the first one, similarly leading to the Hopf algebra $\overline{H}$ which is a quantization of $\mathfrak{g}$ (Theorem 28). Using Proposition 9.7 of \cite{EK1} it is proved that there is an isomorphism of Hopf superbialgebras $H$ and $\overline{H}$ The following theorems summarize the previous constructions of this step. Theorem 33: There exists a functor from the category of finite dimensional superbialgebra $\mathfrak{g}$ over \mathbb{C} $ and the category of quantum universal enveloping superalgebra over \mathbb{C} \lbrack \lbrack h]]$ such that $\mathfrak{g}$ is mapped to $U_{h}(\mathfrak{ })$ which is the second EK quantization. Theorem 34: There exist a functor from the category of quasitriangular finite dimensional superbialgebra $(\mathfrak{g},r)$ over \mathbb{C} $, where $r$ is classical R-matrix, to the category of quasitriangular quantum universal enveloping superalgebra $(U_{h}(\mathfrak{g}),R)$ which is the first EK quantization. Functoriality is the main feature in the proof of Theorem 35: There ia an isomorphism of the first and the second EK quantizations of quasitriangular superbialgebra as Hopf algebras. And at last the Theorem 37: The quantization of a finite dimensional superbialgebra commutes with taking the double $D(U_{h}(\mathfrak{g}))\cong U_{h}(D(\mathfrak{g}))$ Next step (Section 9) is to show that (both of) EK quantizations are isomorphic to the Drinfeld-Jimbo quantization. Following \cite{EK3} it requires to show that the EK quantization is given by the desired generators and relations. The main effort in this part is to prove the additional quantum Serre-type relations which appear in the type $A-G$ superalgebras. Since there are no such additional relations for $\mathfrak{gl}(1|1)$ superalgebra we omit this details. The conclusive steps which lead to the main Theorem 48 were already described above and do not require details. We remark that the Geer's proof has no any restriction explicitly related to semisimplicity of superalgebra and to semisimplicity of the category of its finite dimensional modules therefore can be applied to non-semisimple superalgebra $\mathfrak{gl}(1|1)$. The only feature one should be careful about is that the tensor functors $F,\overline{F}$ between categories are well defined for $\mathfrak{gl}(1|1)$ and their construction covers the set of objects of our category of $\mathfrak{gl}(1|1)$ modules. We argued above why it is the case. This concludes the proof of our Theorem \ref{thm7} claiming braided tensor equivalence of the Drinfeld category $\mathcal{D}$ and the category $\mathcal{C}_{\kappa }$ of Drinfeld-Jimbo quantized superalgebra $\mathfrak{gl}(1|1)$. Summarizing, we have checked that all the steps of the proof of braided tensor equivalence in \cite{Gee} can be applied to the superalgebra \mathfrak{gl}(1|1)$. It is based on the twist (\ref{FcDef}), which exists and is unique, at least on the categories of the solutions of KZ equations we consider. Unfortunately the formula (\ref{FcDef}) for twist is not practically useful in explicit calculations because it requires to know the explicit form of associator. \section{Outlook} \label{section6} The proved braided tensor equivalence of non semisimple categories of \mathcal{A}_{\Omega ,\kappa }$ and $U_{h}(\mathfrak{g})$ modules at generic values of $\kappa $ is a preliminary step towards an understanding of relation between corresponding modules for non generic values of $\kappa $. In this case the problem actually becomes about a correspondence between the categories of modules of logarithmic vertex operator superalgebra $V \mathfrak{gl}(1|1),\kappa )$ and quantum group $U_{q}(\mathfrak{gl}(1|1))$. Despite a\ big progress done in understanding of this correspondence in the last years for non-superalgebra case, the situation with superalgebras remains, to our knowledge, unclear. Recall that in the known cases of such correspondence for non superalgebras the relevant second partner of the correspondence is restricted quantum group, or in the case of logarithmic VOA, unrolled restricted quantum group \cite{CMR}, \cite{Ru}. As we mentioned in Subsection 2.1 an essential progress has been achieved recently in \cite{CMY} in understanding of vertex tensor category structure of $V \mathfrak{gl}(1|1),\kappa )$ for any $\kappa $ including non-generic $\kappa $. It would be interesting to understand what is the quantum group partner for $V(\mathfrak{gl}(1|1),\kappa )$ - modules category for non-generic values of $\kappa $. On a VOA part of the correspondence a rigorous construction of intertwining operators for vertex operator superalgebras at non-generic $\kappa $ is an important first step (for non-superalgebras it was recently done in \cite{McR}). Another hard problem is to understand practical applicability of vertex tensor categories structures (see \cit {HuRev} and references therein) in concrete cases of superalgebras \cit {CKMcR}. Another interesting problem is a logarithmic generalization of the way to construct all the solutions of KZ equations for corrtelation function including non-semisimple finitely generated modules, by an integration operator as in (\ref{integr}) from some minimal set of basic solutions. It is natural to expect as a result a sort of logarithmic deformations of hypergeometric functions structures discovered in \cite{ScheVar}. \section{Appendix A} \label{appA} In this Appendix we collect some data about $\mathfrak{gl}(1|1)$ and details of solutions of its KZ equations. \subsection{Asymptotic solutions of KZ equation} \label{section:subAsymp} \begin{lemma} \label{lemma: A1} If there is an eigenvector (not generalized) $v$ of \Omega _{12}$ with eigenvalue $\lambda $, and there are no eigenvalues of \Omega _{12}$ such that $\lambda +n\kappa ,n\in \mathbb{N} $, then there exists unique asymptotic solution around $x=0$ \begin{equation*} f(x)=x^{\lambda /\kappa }(v+o(x)),\;\;\lim_{x\rightarrow 0}o(x)=0 \end{equation*} \end{lemma} \textit{Proof. }By not generalized eigenvector we mean that $v$ is not a member of a Jordan block. We check existence and uniqueness of asymptotic solution of the form \begin{equation} f(x)=x^{\lambda /\kappa }(v+xv_{1}+x^{2}v_{2}+...),\;\;o(v)=\sum_{n=1}x^{n}v_{n} \label{as} \end{equation with some perhaps infinite set of vectors $v_{n}$. After the substitution of it into the left hand side of the equation (\ref{KZf}) we ge \begin{equation} lhs=x^{\lambda /\kappa }[\lambda x^{-1}v+(\lambda +\kappa )v_{1}+x(\lambda +2\kappa )v_{2}+x^{2}(\lambda +3\kappa )v_{3}+...] \label{lhs} \end{equation The right hand side we rewrite in the vicinity of $x=0$ as \begin{equation*} \frac{\Omega _{12}}{x}-\Omega _{23}(1+x+x^{2}+...) \end{equation* and now we act by it onto (\ref{as}) \begin{eqnarray} rhs &=&x^{\lambda /\kappa }\left( \frac{\Omega _{12}}{x}-\Omega _{23}(1+x+x^{2}+...)\right) (v+xv_{1}+x^{2}v_{2}+...) \label{rhs} \\ &=&x^{\lambda /\kappa }[\lambda vx^{-1}+(\Omega _{12}v_{1}-\Omega _{23}v)+(\Omega _{12}v_{2}-\Omega _{23}v_{1}-\Omega _{23}v)x \notag \\ &&+(\Omega _{12}v_{3}-\Omega _{23}v_{2}-\Omega _{23}v_{1})x^{2}+...] \notag \end{eqnarray Now we compare the multipliers of the same powers of $x$ in (\ref{lhs}) and \ref{rhs}) and get the infinite set of equations \begin{eqnarray} x^{0} &:&(\Omega _{12}-(\lambda +\kappa )Id)v_{1}=\Omega _{23}v, \label{rhs1} \\ x^{1} &:&(\Omega _{12}-(\lambda +2\kappa )Id)v_{2}=\Omega _{23}v_{1}+\Omega _{23}v, \notag \\ x^{2} &:&(\Omega _{12}-(\lambda +3\kappa )Id)v_{3}=\Omega _{23}v_{2}+\Omega _{23}v_{1}, \notag \\ &&....... \notag \end{eqnarray They can be solved one after another. Indeed, the right hand side of the first equation is a known vector. $\det [\Omega _{12}-(\lambda +\kappa )Id]\neq 0$ because $\lambda +\kappa $ is not an eigenvalue of $\Omega _{12} . Therefore the first equation has a unique solution $v_{1}$. The same arguments can now be applied to the second equation : $\Omega _{23}v_{1}$ is now a known vector. We can solve the second equation for $v_{2}$, which is possible because $\det [\Omega _{12}-(\lambda +2\kappa )Id]\neq 0$, for \lambda +2\kappa $ is not an eigenvalue of $\Omega _{12}$. And so on. Thus we find uniquely each vector $v_{i}$ by this recurrent procedure, which proves the statement. We don't discuss the convergency question of the infinite sum of vectors in $o(v)$ because we prove only the existence of asymptotic expansion.\ \ $\square $ The case of a Jordan block requires more general ansatz. The operator x^{\Omega _{12}/\kappa }$ is a well defined operator on any finite dimensional representation space on which $\Omega _{12}$ acts nilpotently. In this case the operator \begin{equation} x^{\Omega _{12}/\kappa }=\sum_{i=0}^{n}\frac{(\ln x)^{i}}{i!}\frac{\Omega _{12}^{i}}{\kappa ^{i}} \label{logex} \end{equation where $n$ is the degree of nilpotency of $\Omega _{12}$. Then we can reformulate the lemma in the following way. \begin{lemma} \label{lemma: A2} If there is a Jordan block of $\Omega _{12}$ with eigenvalue $\lambda $ with the set of eigenvectors $v^{(i)},i=0,...,n-1$, \Omega _{12}v^{(i)}=\lambda v^{(i)}+v^{(i-1)}$, ($v^{(-1)}=0$) and there are no eigenvalues of $\Omega _{12}$ such that $\lambda +n\kappa ,n\in \mathbb{N} $, then there exist $n$ asymptotic solutions around $x=0$ of the form \begin{equation} f_{i}(x)=x^{\lambda /\kappa }(v^{(i)}(\ln x)^{i}+\kappa ^{-1}v^{(i-1)}(\ln x)^{i-1}+o^{(i)}(x)),\;\;\lim_{x\rightarrow 0^{+}}o^{(i)}(x)=0,\;\;i=0,...,n-1 \label{asvj} \end{equation} \end{lemma} \textit{Proof. }To make the presentation more clear we put $\kappa =1$ and prove the statement for the case of rank $n=2$ Jordan block. With a more lengthy formulas the same proof can be repeated for $n>2$. The claim of the lemma for $f_{0}(x)$ becomes identical to the claim of the Lemma \ref{lemma: A1}, with the same proof and the same form of the vector o^{(0)}(x)=xv_{1}+x^{2}v_{2}+...$. Now we prove the lemma for $f_{1}(x).$ We show existence and uniqueness of $v_{j},u_{j},j=1,2,...$ such tha \begin{eqnarray} f_{1}(x) &=&x^{\Omega _{12}}(v^{(1)}\ln x+v^{(0)}+o^{(1)}(x)), \label{as1} \\ o^{(1)}(x) &=&\sum_{j=1}^{\infty }v_{j}x^{j}\ln x+\sum_{j=1}^{\infty }u_{j}x^{j} \notag \end{eqnarray First we prove existence of the vectors $v_{j}$. We substitute this ansatz for $o^{(1)}(x)$ into the KZ equation (\ref{KZf}). We see that the terms proportional to $\ln x/x$ and $1/x$ cancel. Using the same expansion in powers of $x$ of the term $\Omega _{23}/(x-1)$ in before and extracting the terms containing $\ln x$ we get the equations \begin{eqnarray} \ln x &:&(\Omega _{12}-(\lambda +1)Id)v_{1}=\Omega _{23}v^{(0)}, \label{as2} \\ x\ln x &:&(\Omega _{12}-(\lambda +2)Id)v_{2}=\Omega _{23}(v^{(0)}+v_{1}) \notag \\ &&........ \notag \end{eqnarray As before we can solve these equations for $v_{1},v_{2},...$ sequentially because $\lambda +n,n\geq 1$ is not an eigenvalue of $\Omega _{12}$ and the right hand side of these equations are known vectors. After we found $v_{i} s we do the same extracting on both hand side of KZ equation the terms which are not proportional to $\ln x$. We ge \begin{eqnarray} x &:&(\Omega _{12}-(\lambda +1)Id)u_{1}=\Omega _{23}v^{(1)}+v_{1}, \label{as3} \\ x^{2} &:&(\Omega _{12}-(\lambda +2)Id)u_{2}=\Omega _{23}v^{(1)}+v_{2}+\Omega _{23}u_{1}, \notag \\ &&........ \notag \end{eqnarray By the same reasons as before the equations can be uniquely solved sequentially for $u_{i}$. This completes the proof. $\square $ In the same way we can prove similar statements about existence of unique asymptotic solutions of the \ref{KZf} equation around $x=1,x<1$. \begin{lemma} \label{lemma: A3} If there is an eigenvector $v$ of $\Omega _{23}$ with eigenvalue $\lambda $, and there are no eigenvalues of $\Omega _{23}$ such that $\lambda +n\kappa ,n\in \mathbb{N} $, then there exists unique asymptotic solution around $x=1$ of the form \begin{equation} f(x)=(1-x)^{-\lambda /\kappa }(v+o(x)),\;\;\lim_{x\rightarrow 1^{-}}o(x)=0 \label{x1} \end{equation in the case this eigenvector is not a member of a Jordan block. For the case of Jordan block of the size $n$ the $n$ asymptotic solutions are of the form \begin{eqnarray*} f_{i}(x) &=&(1-x)^{-\lambda /\kappa }(v^{(i)}(\ln (1-x))^{i}+\kappa ^{-1}v^{(i-1)}(\ln (1-x))^{i-1}+o^{(i)}(x)),\;\; \\ \lim_{x\rightarrow 1^{-}}o^{(i)}(x) &=&0,\;\;i=0,...,n-1 \end{eqnarray*} \end{lemma} Proof is the same as for Lemmas \ref{lemma: A1},\ref{lemma: A2}. \begin{corol} \label{corol: A4} If the above restriction conditions on the parameters of typical modules are satisfied an equivalent form of asymptotic solutions of \ref{KZf}) around $x=0$ is \begin{equation} f(x)=x^{\Omega _{12}/\kappa }(v_{b}+o(v)) \label{op} \end{equation where $v_{t}$ is the same as $v$ in the case when there are no Jordan block structure for the action of $\Omega _{12}$, and $v_{b}$ is the bottom vector $v^{(n-1)}$ when there is a Jordan block of size $n$ for the action of \Omega _{12}$. \end{corol} \textit{Proof.} In the case without Jordan block this is just change of notations. In the case when there is Jordan block of size $n$ we split \Omega _{12}=$ $\Omega _{12}^{d}+\Omega _{12}^{nil}$ into diagonal and nilpotent parts and write $x^{\Omega _{12}/\kappa }=x^{\Omega _{12}^{d}/\kappa }\sum_{i}\frac{1}{i!}\left( \frac{\Omega _{12}^{nil}} \kappa }\ln x\right) ^{i}$. The action of it on the bottom vector of the set of generalized eigenvectors of $\Omega _{12}$ will generate the sum of vectors proportional to $(\ln x)^{i}v^{(i)}$ where $v^{(i)}$ are the same as in (\ref{asvj}). Therefore the representation (\ref{asvj}) is related to the expansion (\ref{op}) by a change of basis of solutions of KZ equation. $\square $ This corollary enables to use without changes the standard proofs of BTC structure of category of $\mathfrak{gl}(1|1)$-modules with associator and braiding defined through the KZ solutions and their monodromies. \subsection{Basis for $\mathfrak{gl}(1|1)$ and its modules} \label{subgl11} The $\mathfrak{gl}(1|1)$ generators are $E,N,$ $\psi ^{\pm }$ with commutation relations $[N,\psi ^{\pm }]=\pm \psi ^{\pm }$, $\{\psi ^{+},\psi ^{-}\}=E$ and $E$ is central. (Maybe some other choice of basis will be more convenient?) Chevalley involution can be chosen as $\omega (E)=-E,$ $\omega (N)=-N,$ $\omega (\psi ^{\pm })=\pm \psi ^{\mp }$ and produces the dual representation. The basis for typical representation $\mathcal{T}_{e,n}$ of gl(1|1)$ can be chosen as \begin{equation} N=\left( \begin{array}{cc} n+1/2 & 0 \\ 0 & n-1/ \end{array \right) ,\;E=\left( \begin{array}{cc} e & 0 \\ 0 & \end{array \right) ,\;\psi ^{+}=\left( \begin{array}{cc} 0 & e \\ 0 & \end{array \right) ,\;\psi ^{-}=\left( \begin{array}{cc} 0 & 0 \\ 1 & \end{array \right) \label{glbt} \end{equation The basis for weights of module $\mathcal{T}_{e,n}$ is $u=\uparrow =\binom{ }{0}$ (even highest weight), and $v=\downarrow =\binom{0}{1}$ (odd), and for dual module $\mathcal{T}_{e,n}^{\ast }$ -- $u^{\ast }=\binom{0}{1}$ (odd lowest weight), and $v^{\ast }=\binom{-1}{0}$ (even). For one dimensional atypical representation $\mathcal{A}_{n}$ there is one vector $v_{0}$ with the action of the algebra generators $\psi ^{+}v_{0}=\psi ^{-}v_{0}=Ev_{0}=0 , $Nv_{0}=nv_{0}$. The algebra action on it explicitly \begin{equation} N\cdot \uparrow =(n+1/2)\uparrow ,\;N\cdot \downarrow =(n-1/2)\uparrow ,\;\psi ^{+}\cdot \uparrow =\psi ^{-}\cdot \downarrow =0,\;\psi ^{-}\cdot \uparrow =\downarrow ,\;\psi ^{+}\cdot \downarrow =e\uparrow \label{acttyp} \end{equation For four dimensional atypical representation $\mathcal{P}_{n}$ one can choose \begin{eqnarray} N &=&\left( \begin{array}{cccc} n+1 & 0 & 0 & 0 \\ 0 & n & 0 & 0 \\ 0 & 0 & n & 0 \\ 0 & 0 & 0 & n- \end{array \right) ,\;\psi ^{+}=\frac{1}{2}\left( \begin{array}{cccc} 0 & 1 & 1 & 0 \\ 0 & 0 & 0 & 1 \\ 0 & 0 & 0 & -1 \\ 0 & 0 & 0 & \end{array \right) ,\;\psi ^{-}=\frac{1}{2}\left( \begin{array}{cccc} 0 & 0 & 0 & 0 \\ -1 & 0 & 0 & 0 \\ 1 & 0 & 0 & 0 \\ 0 & 1 & 1 & \end{array \right) , \label{glbp} \\ E &=&0\times Id_{4} \notag \end{eqnarray And the weights of the module \begin{equation} u_{1}=t=\left( \begin{array}{c} 0 \\ 1 \\ 1 \\ \end{array \right) ,\;v_{1}=r=\left( \begin{array}{c} 1 \\ 0 \\ 0 \\ \end{array \right) ,\;v_{2}=l=\left( \begin{array}{c} 0 \\ 0 \\ 0 \\ \end{array \right) ,\;u_{2}=b=\frac{1}{2}\left( \begin{array}{c} 0 \\ 1 \\ -1 \\ \end{array \right) \label{glpv} \end{equation the even vectors are $u_{1,2}$, the odd \ $v_{1,2}$. This module is self dual. The algebra action on i \begin{align} N\cdot t& =nt,\;N\cdot r=(n+1)r,\;N\cdot l=(n-1)l,\;N\cdot b=nb, \label{actproj} \\ \psi ^{+}\cdot t& =r,\;\psi ^{+}\cdot l=b,\;\psi ^{+}\cdot r=\psi ^{+}\cdot b=0,\; \notag \\ \psi ^{-}\cdot t& =l,\;\psi ^{-}\cdot r=-b,\;\psi ^{-}\cdot l=\psi ^{-}\cdot b=0,\; \notag \end{align} We will use the following choice of Casimir element \begin{equation} \Omega=NE+EN+\psi^{-}\psi^{+}-\psi^{+}\psi^{-}+E^{2} \label{cas} \end{equation} and its tensor analog \begin{equation} \Omega_{ij}=N_{i}\otimes E_{j}+E_{i}\otimes N_{j}+\psi_{i}^{-}\otimes\psi _{j}^{+}-\psi_{i}^{+}\otimes\psi_{j}^{-}+E_{i}\otimes E_{j} \label{cast} \end{equation} where the lower indices denote the spaces where the generator acts. $\widehat{\mathfrak{gl}}(1|1)$ commutation relations \begin{equation} \lbrack N_{r},E_{s}]=rk\delta _{r+s},\;[N_{r},\psi _{s}^{\pm }]=\pm \psi _{r+s}^{\pm },\;\{\psi _{r}^{+},\psi _{s}^{-}\}=E_{r+s}+rk\delta _{r+s} \label{acm} \end{equation One can rescale generators in such a way that $k$ will become 1 (if it is not 0), but we will keep it. The generic $k$ will mean $e/k\notin \mathbb{Z} $ for all the modules involved into correlation function, as well as for all the modules appearing in tensor product decomposition. A remark: the structure of all modules for non generic $k$ for $\widehat{\mathfrak{gl} (1|1)$ and their tensor product decomposition is of course well known, but the KZ for this case and its solutions is another (next...) problem. Conformal dimension of Virasoro primary field $h=e\left( n+\frac{e}{2 \right) $. We are going to find basis for invariants of level zero KZ equations for N=2,3,4$. Recall that level zero equations in the case of $\mathfrak{gl (1|1) $ means that $\sum e_{i}=0$, if typical representations are involved in correlation function. In addition the invariants can be classified according to the $N -grading of the space of states $V$ of correlation function. \subsection{Examples of solutions of KZ equation for correlation functions} \label{subCorf} In this section we collect examples of explicit form of KZ $N=2,3$ solutions on the space of $\mathfrak{gl}(1|1)$ invariant functions. This class of solutions is the most interesting in the context of KZ equations for correlation functions of intertwining operators of affine Lie superalgebra \mathfrak{gl}(1|1)^{\vee }$. Similar calculations has been done in the paper \cite{Troo}. \textbf{1. }$N=2$ There is one invariant for $\mathcal{TT}$ correlation function in the basis described above $I_{0}^{\mathcal{TT}}=\uparrow \downarrow +\downarrow \uparrow ,$and the list of invariants for $\mathcal{PP}$ correlation function is \begin{align} I_{-1}^{\mathcal{PP}}& =rb-br \label{PPinv} \\ I_{0,1}^{\mathcal{PP}}& =tb+rl-lr+bt \notag \\ I_{0,2}^{\mathcal{PP}}& =bb \notag \\ I_{1}^{\mathcal{PP}}& =lb-bl \notag \end{align The first subindex denotes the value of $n_{1}+n_{2}$. (Recall that it is not an eigenvalue of $N$ acting on the tensor product state. The latter is 0 for $\mathfrak{g}$-invariant correlation function.) Projection of KZ $N=2$ equation onto this basis gives an ODE with solution \begin{equation} f(z_{1},z_{2})=[A(z_{1}-z_{2})^{\delta _{12}/k}]I_{0}^{\mathcal{TT },\;\delta _{ij}=n_{i}e_{j}+n_{j}e_{i}+e_{i}e_{j} \label{sol1} \end{equation for $\mathcal{T}_{e_{1},n_{1}}\mathcal{T}_{-e_{1},n_{2}}$ correlation function ($A$ is a constant), and solutions \begin{eqnarray} f(z_{1},z_{2}) &=&const\times I_{\pm 1}^{\mathcal{PP}},\text{ for n_{1}+n_{2}=\pm 1 \label{sol2} \\ f(z_{1},z_{2}) &=&AI_{0,2}^{\mathcal{PP}}+(2A\kappa ^{-1}\ln (z_{1}-z_{2})+B)I_{0,1}^{\mathcal{PP}}\text{ \ for }n_{1}+n_{2}=0 \notag \end{eqnarray where $A,B$ are constants. This is an example of logarithms in correlation functions of logarithmic vertex operator algebras. \textbf{2. }$N=3$ There are two invariants for $\mathcal{TTT}$ correlation in the same notations as above \begin{align} I_{-1/2}^{\mathcal{TTT}}& =(\uparrow \uparrow \downarrow +\uparrow \downarrow \uparrow +\downarrow \uparrow \uparrow ) \label{TTTinv} \\ I_{+1/2}^{\mathcal{TTT}}& =(e_{1}\uparrow \downarrow \downarrow -e_{2}\downarrow \uparrow \downarrow +e_{3}\downarrow \downarrow \uparrow ), \notag \end{align (Of course $e_{1}+e_{2}+e_{3}=0$.) Invariants of $\mathcal{TTP}$ correlations ar \begin{align} I_{-1}^{\mathcal{TTP}}& =\uparrow \uparrow b-\uparrow \downarrow r-\downarrow \uparrow r \label{TTPinv} \\ I_{0,1}^{\mathcal{TTP}}& =e_{1}(\uparrow \uparrow l+\uparrow \downarrow t+\downarrow \uparrow t)+\uparrow \downarrow b+\downarrow \downarrow r \notag \\ I_{0,2}^{\mathcal{TTP}}& =\uparrow \downarrow b+\downarrow \uparrow b \notag \\ I_{1}^{\mathcal{TTP}}& =e_{1}(\uparrow \downarrow l+\downarrow \uparrow l)+\downarrow \downarrow b \notag \end{align and the list of invariants of $\mathcal{PPP}$ correlations are \begin{align} I_{-2}^{\mathcal{PPP}}& =rrb-rbr+brr \label{PPPinv} \\ I_{-1,1}^{\mathcal{PPP}}& =trb-tbr-rrl-rbt+lrr+brt \notag \\ I_{-1,2}^{\mathcal{PPP}}& =rtb+rrl-rlr+rbt-btr-brt \notag \\ I_{-1,3}^{\mathcal{PPP}}& =rbb-brb \notag \\ I_{-1,4}^{\mathcal{PPP}}& =rbb-bbr \notag \\ I_{0,1}^{\mathcal{PPP}}& =btb+brl-blr+bbt \notag \\ I_{0,2}^{\mathcal{PPP}}& =bbb \notag \\ I_{1,1}^{\mathcal{PPP}}& =tlb-tbl-rll+llr-lbt+blt \notag \\ I_{1,2}^{\mathcal{PPP}}& =ltb+lrl-llr+lbt-btl-blt \notag \\ I_{1,3}^{\mathcal{PPP}}& =lbb-bbl \notag \\ I_{1,4}^{\mathcal{PPP}}& =blb-bbl \notag \\ I_{2}^{\mathcal{PPP}}& =llb-lbl+bll \notag \end{align Projection of KZ equation in the form (\ref{KZf}) onto these bases gives systems of ODEs with the following solutions. If the space of invariants with fixed first subindex, i.e. fixed sum of $n_{1}+n_{2}+n_{3}$ is one dimensional equal to $I$ then the solution for correlation function in all three cases can be written as \begin{equation} f(x)=Ax^{\alpha /\kappa }(1-x)^{\beta /\kappa }I \label{sol31} \end{equation where $A\in \mathbb{C} $ is a constant, and $\alpha ,\beta $ are eigenvalues of $\Omega _{12},\Omega _{23}$ acting on $I$ respectively. In the $\mathcal{TTP}$ case with $n_{1}+n_{2}+n_{3}=0$ solution contains logarithms \begin{equation} f(x)=Ax^{\delta _{12}/\kappa }(1-x)^{\delta _{23}/\kappa }[I_{0,1}^{\mathcal TTP}}+(B+\frac{e_{1}}{\kappa }(\ln (1-x)-\ln x))I_{0,2}^{\mathcal{TTP}}] \label{TTP1} \end{equation} In the $\mathcal{PPP}$ case with $n_{1}+n_{2}+n_{3}=0$ the solution is trivial \begin{equation} f(x)=AI_{0,1}^{\mathcal{PPP}}+BI_{0,2}^{\mathcal{PPP}},\;A,B\in \mathbb{C} \label{PPP0} \end{equation But in the case $n_{1}+n_{2}+n_{3}=\pm 1$ there are logarithms in the solutions \begin{eqnarray} f^{\pm }(x) &=&A^{\pm }I_{\pm 1,1}^{\mathcal{PPP}}+B^{\pm }I_{\pm 1,2}^ \mathcal{PPP}}+\left( C_{3}^{\pm }+\frac{A^{\pm }-B^{\pm }}{\kappa }\ln x \frac{B^{\pm }-2A^{\pm }}{\kappa }\ln (1-x)\right) I_{\pm 1,3}^{\mathcal{PPP } \notag \\ &&+\left( C_{4}^{\pm }+\frac{B^{\pm }}{\kappa }\ln x+\frac{A^{\pm }-B^{\pm }{\kappa }\ln (1-x)\right) I_{\pm 1,3}^{\mathcal{PPP}} \label{PPP1} \end{eqnarray where $A^{\pm },B^{\pm },C_{3,4}^{\pm }$ are constants. Another interesting problem is structure of solutions of KZ equations on a wider $N$-graded spaces, not necessarily invariants of $\mathfrak{gl}(1|1)$. We will address this problem elsewhere. \section{Appendix B} \label{appB} Here we will describe the basis and tensor product decomposition of $U_{h} \mathfrak{gl}(1|1))$-modules and will prove the Proposition 3. We will choose $i\pi \kappa ^{-1}=h$ and consider real $\kappa $. We use the following matrix basis for the three types of $U_{h}(\mathfrak{gl}(1|1)) -modules $\mathcal{T}_{e,n}^{\kappa },\mathcal{A}_{n}^{\kappa },\mathcal{P _{n}^{\kappa }$ included into $\mathcal{C}_{\kappa }$, as the basis for construction of tensor ring. For $\mathcal{T}_{e,n}^{\kappa } \begin{equation*} E=\left( \begin{array}{cc} e & 0 \\ 0 & \end{array \right) ,\;N=\left( \begin{array}{cc} n+1/2 & 0 \\ 0 & n-1/ \end{array \right) ,\;\psi ^{+}=\left( \begin{array}{cc} 0 & 2\sinh (eh) \\ 0 & \end{array \right) ,\;\;\psi ^{-}=\left( \begin{array}{cc} 0 & 0 \\ 1 & \end{array \right) \end{equation* with the vectors of the module \begin{equation*} |e,n\rangle =\left( \begin{array}{c} 1 \\ \end{array \right) \text{ \ (even)},\;|e,n-1\rangle =\psi ^{-}|e,n\rangle =\left( \begin{array}{c} 0 \\ \end{array \right) \text{ \ (odd)} \end{equation*} and for four dimensional module we choose \begin{eqnarray*} N &=&\left( \begin{array}{cccc} n+1 & 0 & 0 & 0 \\ 0 & n & 0 & 0 \\ 0 & 0 & n & 0 \\ 0 & 0 & 0 & n- \end{array \right) ,\;\psi ^{+}=\left( \begin{array}{cccc} 0 & 1 & -e^{h} & 0 \\ 0 & 0 & 0 & e^{h} \\ 0 & 0 & 0 & 1 \\ 0 & 0 & 0 & \end{array \right) , \\ \psi ^{-} &=&\left( \allowbreak \begin{array}{cccc} 0 & 0 & 0 & 0 \\ -1 & 0 & 0 & 0 \\ -e^{-h} & 0 & 0 & 0 \\ 0 & e^{-h} & -1 & \end{array \right) ,\;E=0\times Id_{4}, \end{eqnarray* The coordinates of the vectors of the four dimensional vector space of this representation are graded as in (\ref{glpv}). Let us note that there are many other matrix presentations of $\mathcal{P}_{n}^{\kappa }$ module which can contain some more free numerical parameters. \textit{Proof of Proposition 3}. With these basis we can consider decomposition of tensor product of this set of three types of modules using the coproduct (\ref{qcop}) and show that under some suitable assumptions on parameters of modules they form a ring. The cases \begin{equation*} \mathcal{A}_{n}^{\kappa }\otimes \mathcal{A}_{n^{\prime }}^{\kappa } \mathcal{A}_{n+n^{\prime }}^{\kappa },\;\mathcal{A}_{n}^{\kappa }\otimes \mathcal{T}_{e,n^{\prime }}^{\kappa }=\mathcal{T}_{e,n+n^{\prime }}^{\kappa },\;\mathcal{A}_{n}^{\kappa }\otimes \mathcal{P}_{e,n^{\prime }}^{\kappa } \mathcal{P}_{e,n+n^{\prime }}^{\kappa } \end{equation* are obvious. More interesting are the remaining three cases. Consider $\mathcal{T}_{e,n}^{\kappa }\otimes \mathcal{T}_{e^{\prime },n^{\prime }}^{\kappa }$. The calculations of tensor product decomposition of two $U_{ih}(\mathfrak{gl}(1|1))$-modules $\mathcal{T}_{e_{1},n_{1}}^ \kappa }\otimes \mathcal{T}_{e_{2},n_{2}}^{\kappa }$ is completely parallel to the same calculations for $\mathfrak{gl}(1|1)$-modules. $\mathcal{T _{e,n}^{\kappa }$ has two states - the highest weight $v_{1}=|\uparrow \rangle $ Grassmann even and $v_{2}=\psi ^{-}v_{1}=|\downarrow \rangle $ - Grassmann odd. We can start from two vectors $w_{2}=\alpha _{2}|\uparrow \rangle \otimes |\downarrow \rangle +\beta _{2}|\downarrow \rangle \otimes |\uparrow \rangle $ and $u_{1}=\alpha _{1}|\uparrow \rangle \otimes |\downarrow \rangle +\beta _{1}|\downarrow \rangle \otimes |\uparrow \rangle $ with constraint $\alpha _{2}=-\beta _{1}\beta _{2}/\alpha _{1}$ which guarantees their orthogonality. We consider $w_{2}$ as highest weight of a grading reversed module, i.e. $\Delta (\psi ^{+})w_{2}=0$. It gives $\beta _{2}=-2\alpha _{2}e^{-he_{1}}\sinh (e_{2}h)$. And we consider $u_{1}$ \ as lowest weight module, with Grassmann even highest weight. It means $\Delta (\psi ^{-})u_{1}=0$, which gives $\beta _{1}=\alpha _{1}e^{he_{2}}$. Then one can easily check that corresponding lowest weight module of the first (grading reversed) module is $2\alpha _{1}\sinh ((e_{1}+e_{2})h)|\downarrow \rangle \otimes |\downarrow \rangle $, and highest weight of the second module is $\alpha _{2}\sinh ((e_{1}+e_{2})h)/\sinh (e_{1}h)|\uparrow \rangle \otimes |\uparrow \rangle $. We see that conditions $\sinh ((e_{1}+e_{2})h)\neq 0$, $\sinh (e_{1}h)\neq 0$, which mean $e_{1}/\kappa \notin \mathbb{Z} $, $(e_{1}+e_{2})/\kappa \notin \mathbb{Z} \backslash \{0\}$ are sufficient for decomposition \begin{equation*} \mathcal{T}_{e_{1},n_{1}}^{\kappa }\otimes \mathcal{T}_{e_{2},n_{2}}^{\kappa }=\mathcal{T}_{e_{1}+e_{2},,n_{1}+n_{2}+1/2}^{\kappa }\oplus \mathcal{T _{e_{1}+e_{2},,n_{1}+n_{2}-1/2}^{\kappa \prime } \end{equation* In the case $e_{1}+e_{2}=0$ one can check that any vector of the form |t\rangle =\alpha |\uparrow \rangle \otimes $ $|\downarrow \rangle +\beta |\downarrow \rangle \otimes $ $|\uparrow \rangle $ with $\alpha \neq e^{he_{1}}\beta $ serves as the $|t\rangle $-vector in the basis of the \mathcal{P}_{n_{1}+n_{2}}^{\kappa }$ module of four vectors of the tensor product $\mathcal{T}_{e_{1},n_{1}}^{\kappa }\otimes \mathcal{T _{-e_{1},n_{2}}^{\kappa }$. We see that the tensor product ring composed of the $U_{h}(\mathfrak{gl}(1|1))$-modules $\mathcal{A}_{n}^{\kappa },\mathcal{ }_{e,n}^{\kappa },\mathcal{P}_{n}^{\kappa }$ is the same as the tensor product ring of the category $\mathcal{C}_{\kappa }$ composed of $\mathcal{A _{n},\mathcal{T}_{e,n},\mathcal{P}_{n}$ for restriction on parameters the same as in the Proposition 3. $\square $

{.203} {.204}

\section{Introduction}\label{Sec:intro} In scalable quantum computation \cite{Nielsen2000}, the quantum gates in a quantum circuit have to be implemented with very high fidelity, with the admissible error in the range of $10^{-3}$ to $10^{-4}$, depending on the quantum error correction protocol \cite{Lidar2013}. The current state of the art in trapped-ions experiments features errors of $10^{-5}$ \cite{Brown2011,Gaebler2016} and even $10^{-6}$ \cite{Harty2014} for single-qubit gates, and errors of the order of $10^{-3}$ \cite{Ballance2016,Gaebler2016} for two-qubit gates. The detrimental cross-talk to neighboring ion qubits has been suppressed to values of $10^{-5}$ \cite{Piltz2014} and $10^{-6}$ \cite{Craik2017}, and the errors in ion transport have been reduced below $10^{-5}$ too \cite{Kaufmann2018}. In superconducting qubits, fidelities of 99.9\% for single-qubit and 99.4\% for two-qubit gates have been reported \cite{Barends2014}. Various steps in qubit readout and quantum gate tomography, e.g. population shelving \cite{Stevens1998,Sorensen2006,Moller2007,Schafer2020}, have to be implemented with very high fidelity too. Such tiny errors require very good control of the driving field as well as the environment, e.g. compensation of ambient electric and magnetic fields to a very high degree. This is crucial in a quantum circuit where the entire propagator of the particular gate must be very well controlled, i.e. both the probability and the propagator phases have to be very stable. The phases, in particular, are very sensitive to a detuning shift, which can emerge due to both fluctuating frequency of the laser, microwave or radiofrequency generator, and energy level shifts caused by uncompensated magnetic or electric fields. In order to characterize such unwanted detuning shifts in the course of the measurement it is very useful to have a simple, fast and reliable method to detect and measure such shifts. Measuring the populations after the application of a single pulse, which generates the quantum gate, is not reliable because the populations are fairly insensitive to detuning shifts, whereas the gate phases, which are prone to these shifts, are invisible to such a measurement. Performing full quantum tomography would reveal the changes in the gate phases but such a tomography could be very time-consuming. To this end, I propose in this paper a conceptually very simple technique for sensing and measuring small detuning shifts. It is based on the repeated application of the same gate-generating pulse, which greatly amplifies the effect of the detuning shift due to quantum interference and maps it onto the populations. In this manner, one does not need any change in the experimental apparatus (i.e. the pulse amplitude, duration, shape, frequency and possibly chirp), and avoids the introduction of additional experimental parameters beyond the gate-generating pulse. In this manner, the proposed technique is simpler than other, more sophisticated quantum sensing methods \cite{Degen2016} and quantum gate tomography \cite{Merkel2013}. I consider the implementation of this sensing method by applying the pulses used in two major quantum control techniques for population inversion: a resonant $\pi$ pulse and an adiabatic chirped pulse. The paper is organized as follows. The concept of the sensing method is introduced in Sec.~\ref{Sec:concept} and the general features of the transition probability generated the two types of sensing sequences are presented in Sec.~\ref{Sec:multi-pass}. Three exactly soluble analytic models are presented in Secs.~\ref{Sec:RZ}, \ref{Sec:Rabi}, and \ref{Sec:DK}. The last section \ref{Sec:conclusions} wraps up the results and presents some discussion on the limitations of the method and an outlook of its possible extensions. \section{Concept \label{Sec:concept}} \begin{figure}[t] \begin{tabular}{cc} \begin{tabular}{c} \includegraphics[width=0.22\columnwidth]{fig1a.eps} \end{tabular} & \begin{tabular}{c} \includegraphics[width=0.75\columnwidth]{fig1c.eps} \end{tabular} \\ \\ \begin{tabular}{c} \includegraphics[width=0.22\columnwidth]{fig1b.eps} \end{tabular} & \begin{tabular}{c} \includegraphics[width=0.75\columnwidth]{fig1d.eps} \end{tabular} \\ & \begin{tabular}{c} \includegraphics[width=0.52\columnwidth]{fig1e.eps} \end{tabular} \end{tabular} \caption{ Pulse sequences used to measure a small detuning shift $\Delta_0$ from resonance by repeating the single-pulse interaction, while keeping its phase the same (top), or flipping it by $\pi$ from pulse to pulse (bottom). } \label{fig:2ss} \end{figure} I consider two types of pulse sequences applied to a two-state quantum system, see Fig.~\ref{fig:2ss}. In one of them, the same pulse is applied $N$ times, see Fig.~\ref{fig:2ss} (top), and then the populations are measured. In the other, the same pulse is again applied $N$ times but the phase of every other pulse is shifted by $\pi$, i.e. a minus sign is applied to the Rabi frequency of every even-numbered pulse in the sequence, see Fig.~\ref{fig:2ss} (bottom). The sequences may contain an odd or even number of pulses. On exact resonance it is assumed that each pulse produces complete (for resonant pulse) or almost complete (for adiabatic chirped pulse) population inversion. This is achieved by a $\pi$ pulse in the former case and a chirped pulse with a pulse area of a few $\pi$ in the latter. Because the excitation profile has its maximum value (and hence a vanishing first-order derivative versus the detuning) at resonance, a small detuning shift has a very little effect on the populations. However, such a detuning shift changes the dynamic phases of the propagator much more significantly. If one is restricted to measuring populations, as it is assumed here, these phase shifts are invisible in a single-pulse interaction; however, they are mapped onto the populations by the interference generated by a train of pulses. These pulses are chosen to be identical to the single $\pi$ pulse, or the single chirped adiabatic pulse, mentioned above, with the only difference being the sign flip of the Rabi frequency in the sign-alternating sequence. In this manner, no new parameters are introduced in addition to the dynamic parameters of the single-pulse propagator (the transition probability and the two dynamic phases). This fact is important because, in the absence of other uncertainties, the population changes are directly linked to the detuning shift. The rationale for the two pulse sequences considered here is the following. It might appear at first sight that repeating the same pulse, without phase shifts, is the most natural approach to coherent error amplification. Indeed, this case is carefully analyzed here. However, it turns out that a sequence of pulses with alternating Rabi frequency signs is a far better approach as far as detuning shift sensing is concerned. When considered more carefully, this is readily understood: a pair of two resonant pulses with a $\pi$ phase shift cancel each other's effect exactly because the propagator for the second pulse is the Hermitean conjugate of the propagator of the first, thereby resulting in the identity operation. In other words, a pair of resonant pulses with Rabi frequencies $\Omega$ and $-\Omega$ will produce no overall change and exactly zero transition probability. A small detuning will break this symmetry and the cancellation will not occur, giving rise to a nonzero transition probability. When such a pair is repeated $N$ times, the nonzero probability is coherently amplified very quickly. Furthermore, sequences of both even and odd number of pulses $N$ are used. With the transition probability for zero static detuning being equal to one, or almost one, one should keep in mind that the multi-pass transition probability for an odd $N$ is equal to one, or almost one, while it is zero, or almost zero, for an even $N$. Hence the sensing feature around the zero static detuning appears as a \emph{spike} for odd $N$ and a \emph{dip} for even $N$. It is the \emph{width} of this feature and its \emph{slope}, which are important for the frequency shift sensing rather than whether it is a spike or a dip. In order to not only sense an energy level shift but accurately measure it using the pulse sequence scenario, an analytic relation between the single-pulse and $N$-pulse probabilities is needed. Such a relation for a SU(2) propagator is available \cite{Vitanov1995} and it has been used recently for quantum gate tomography \cite{Vitanov2018,Vitanov2020}. Here it is used extensively, in combination with explicit analytic formulas for the propagator for three popular exactly soluble two-state models, the non-crossing Rosen-Zener \cite{Rosen1932,Vitanov1994} and Rabi \cite{Shore1990} models and the level-crossing Demkov-Kunike model \cite{Demkov1969,Hioe1985,Zakrzewski1985,Suominen1992}. They allow one to explicitly link the $N$-pulse transition probability to the detuning shift. \section{Multi-pass transition probability}\label{Sec:multi-pass} The Hamiltonian of a coherently driven two-state quantum system, in the rotating-wave approximation \cite{Shore1990}, reads \begin{equation} \label{H-2} \mathbf{H}(t) = \tfrac12 \left[\begin{array}{cc} -\Delta(t) & \Omega(t) \\ \Omega(t) & \Delta(t) \end{array}\right], \end{equation} where $\Delta(t)$ is the system-field frequency detuning and $\Omega(t)$ is the Rabi frequency of the coupling between the two states. The Rabi frequency is supposed to be pulse-shaped and the detuning may contain a static part $\Delta_0$ and a chirp, \begin{equation} \Delta (t) = \Delta_0 + \beta f(t). \end{equation} In this paper, the objective is to measure a static error in the detuning around zero, i.e. the static detuning $\Delta_0$ will be the (unknown) quantity to be determined. For arbitrary $\Omega(t)$ and $\Delta(t)$, the propagator corresponding to the traceless Hamiltonian \eqref{H-2} is a SU(2) matrix, which is expressed in terms of the complex-valued Cayley-Klein parameters $a$ and $b$ ($|a|^2 +|b|^2=1$) as \begin{equation} \label{U} \mathbf{U} = \left[\begin{array}{cc} a & -b^* \\ b & a^* \end{array}\right]. \end{equation} If the system is initially in state $\ket{1}$, the probabilities for remaining in state $\ket{1}$ and for transition to state $\ket{2}$ are \begin{equation} \label{2ss-pq} q = |a|^2,\quad p = |b|^2. \end{equation} with $p+q = 1$. In the following sections, the connections between $\Omega(t)$ and $\Delta(t)$ and the Cayley-Klein parameters will be explicitly presented for three specific analytically soluble models. Instead of the two complex Cayley-Klein parameters $a$ and $b$, the propagator \eqref{U} can be expressed in terms of three real parameters: the transition probability $p$ and the dynamical phases $\xi$ and $\eta$ (sometimes called St\"uckelberg phases), \begin{equation} \label{propagator} \mathbf{U} = \left[\begin{array}{cc} e^{i \xi} \sqrt{1-p} & - e^{-i\eta} \sqrt{p} \\ e^{i\eta} \sqrt{p} & e^{-i \xi} \sqrt{1-p} \end{array}\right] , \end{equation} where $a = e^{i \xi} \sqrt{1-p}$ and $b = e^{i\eta} \sqrt{p}$. \subsection{Sequence of pulses with the same phases\label{Sec:plus}} In order to determine the populations after $N$ passes we need to find the $N$-pass propagator $\mathbfcal{U}_N = \mathbf{U}^{N}$. It has been proved \cite{Vitanov1995} that the $N$-th power of any SU(2) propagator, parameterized as in Eq.~\eqref{U}, reads \begin{equation} \label{U-N} \mathbfcal{U}_N = \left[\begin{array}{cc} \cos N\theta + ia_i \dfrac{\sin N\theta}{\sin \theta} & -b^* \dfrac{\sin N\theta}{\sin \theta} \\ b \dfrac{\sin N\theta}{\sin \theta} & \cos N\theta - ia_i \dfrac{\sin N\theta}{\sin \theta} \end{array}\right], \end{equation} where $a = a_r + i a_i$ and \begin{equation} \label{theta} \theta = \arccos (a_r) \quad (0 \leqq \theta \leqq \pi). \end{equation} Therefore, the transition probability after $N$ passes is \begin{equation} \label{P-N} \mathcal{P}_N = p\, \frac{\sin^2 (N\theta)}{\sin^2 (\theta)}. \end{equation} For $N=2$, after simple algebra one finds \begin{equation} \mathbf{U}_{\Omega,\Delta} \mathbf{U}_{\Omega,\Delta} = \left[\begin{array}{cc} a^2 - |b|^2 & -2 b^* a_r \\ 2 b a_r & (a^*)^2 - |b|^2 \end{array}\right]. \label{2ss-UppU} \end{equation} The double-pass transition probability is equal to $\mathcal{P}_2 = 4 p (1-p) \cos^2 \xi$. \subsection{Sequence of pulses with alternating phases\label{Sec:minus}} For the pulse sequence with alternating phases we need a modification of the above result as follows. Consider a second interaction with the same magnitudes of $\Omega(t)$ and $\Delta(t)$, but with the opposite sign of $\Omega(t)$, see Fig.~\ref{fig:2ss}. The respective propagator can be obtained from Eq.~\eqref{U} by simple algebraic operations \cite{Vitanov1999nonlinear,Vitanov2018,Vitanov2020} and, very importantly, can be expressed with the \emph{same} Cayley-Klein parameters $a$ and $b$, \begin{equation} \label{U-} \mathbf{U}_{-\Omega,\Delta} = \left[\begin{array}{cc} a & b^* \\ -b & a^* \end{array}\right]. \end{equation} The respective double-pass propagator reads \begin{equation} \mathbf{U}_{-\Omega,\Delta} \mathbf{U}_{\Omega,\Delta} = \left[\begin{array}{cc} a^2 + |b|^2 & -2 i b^* a_i \\ -2 i b a_i & (a^*)^2 + |b|^2 \end{array}\right], \label{2ss-UmpU} \end{equation} where $\mathbf{U}_{\Omega,\Delta}$ is the same as $\mathbf{U}$ of Eq.~\eqref{U}. The double-pass transition probability is equal to $\mathcal{P}_2^\pm = 4 p (1-p) \sin^2 \xi$, which is different from the double-pass transition probability of $\mathcal{P}_2 = 4 p (1-p) \cos^2 \xi$ in the case of the same phases due to the St\"uckelberg phase $\xi$. This difference will be greatly amplified in the $N$-pass probabilities. For the sign-alternating sequence, we can use Eq.~\eqref{2ss-UmpU} as the basic building block and derive the $2n$-pass propagator using the connection between Eqs.~\eqref{U} and \eqref{U-N} above. Obviously, the parameter $a$ is now replaced by $a^2 + |b|^2 = 1 -2 a_i^2 + 2i a_r a_i $, the parameter $b$ is replaced by $-2 i b a_i$, and the parameter $\theta$ is redefined as \begin{equation} \label{Theta} \Theta = \arccos (1-2a_i^2). \end{equation} Then the propagator for the sign-alternating sequence of $2N$ pulses reads \begin{equation} \label{U2-N} \mathbfcal{U}_{2n}^\pm = \left[\begin{array}{cc} \cos n\Theta + 2i a_r a_i \dfrac{\sin n\Theta}{\sin \Theta} & -2 i b^* a_i \dfrac{\sin n\Theta}{\sin \Theta} \\ -2 i b a_i \dfrac{\sin n\Theta}{\sin \Theta} & \cos n\Theta - 2i a_r a_i \dfrac{\sin n\Theta}{\sin \Theta} \end{array}\right]. \end{equation} The propagator for $2n+1$ pulses can be obtained by multiplying $\mathbfcal{U}_{2n}^\pm$ by $\mathbf{U}$ of Eq.~\eqref{U}, \begin{equation} \label{U2-Nodd} \mathbfcal{U}_{2n+1}^\pm = \mathbf{U} \mathbfcal{U}_{2n}^\pm = \left[\begin{array}{cc} a C + 2i a_i S & - b^* C \\ b C & a^* C - 2i a_i S \end{array}\right], \end{equation} with \begin{equation} C= \dfrac{\cos (n+\frac12)\Theta}{\cos \frac12\Theta},\quad S = \dfrac{\sin n\Theta}{\sin \Theta}. \end{equation} Therefore, the transition probability after $2n$ and $2n+1$ pulses reads \begin{subequations} \label{P-N-pm} \begin{align} \mathcal{P}_{2n} &= p\, \frac{\sin^2 n\Theta}{\cos^2 \frac12\Theta}, \label{Ppme} \\ \mathcal{P}_{2n+1} &= p\, \frac{\cos^2 (n+\frac12)\Theta}{\cos^2 \frac12\Theta} . \label{Ppmo} \end{align} \end{subequations} In the derivation of Eq.~\eqref{Ppme}, it is used that $a_i^2 = \sin^2 \frac12\Theta$, which follows from Eq.~\eqref{Theta}. In the limit of nearly complete population transfer by a single pulse, which is concerned here, we have $|b| \approx 1$ and $|a| \ll 1$. Then $|a_r| \ll 1$ and $|a_i| \ll 1$ and hence $\theta \approx \pi/2$ and $\Theta \ll 1$. More accurately, we have \begin{subequations} \begin{align} \theta &\approx \tfrac12\pi - a_r - \tfrac16 a_r^3 + \cdots, \\ \Theta &\approx 2|a_i| + \tfrac13 |a_i|^3 + \cdots \end{align} \end{subequations} Therefore, for sequences of pulses of the same phase, \begin{subequations} \label{PN++} \begin{align} \mathcal{P}_{2n} &\approx (2n)^2 a_r^2, \label{PN++even} \\ \mathcal{P}_{2n+1} &\approx 1 - a_i^2 - (2n+1)^2 a_r^2, \label{PN++odd} \end{align} \end{subequations} while for sequences of pulses of alternating phases, \begin{subequations} \label{PN+-} \begin{align} \mathcal{P}_{2n}^\pm &\approx (2n)^2 a_i^2, \label{PN+-even} \\ \mathcal{P}_{2n+1}^\pm &\approx 1 - (2n+1)^2 a_i^2, \label{PN+-odd} \end{align} \end{subequations} where the relation $p = 1 - a_r^2 -a_i^2$ has been accounted for. While retaining the dominant asymptotic terms, especially in Eq.~\eqref{PN++odd}, it has been also used that $a_i$ is of order $O(\delta)$ and $a_r$ is of order $O(\delta^2)$. Obviously, the difference between the two types of sequences is that for a sequence of pulses with the same phases, the $N$-pulse transition probability \eqref{P-N} is controlled by the real part $a_r$ of the diagonal Cayley-Klein parameter $a$, while for a sequence of pulses with alternating phases, the $N$-pulse transition probability is controlled by the imaginary part $a_i$ of $a$. As we will see, this is an important difference and these observations show the significance of the properties of the parameter $a$. \subsection{Case studies \label{Sec:models}} In order to examine the performance of the general formulas \eqref{PN++} and \eqref{PN+-} in regard to sensing, we consider three analytically soluble models, which provide explicit analytic expressions for the parameters $a_r$ and $a_i$, from which one can find out the dependence of the populations on the static detuning shift $\Delta_0$. The first model is the Rosen-Zener model, which assumes a bell-shaped hyperbolic-secant pulse shape and a static detuning. It is suitable for studying the sensitivity of a population inverting resonant $\pi$ pulse to a small detuning shift. The second model is the Rabi model, which assumes a rectangular pulse shape and a constant detuning. It is of the same type as, and simpler than the Rosen-Zener model. The comparison of the two models reveals the effects of the sharp pulse edges in the Rabi model. The third model is the Demkov-Kunike model, which assumes the same hyperbolic-secant pulse shape as the Rosen-Zener model but the detuning is a sum of a hyperbolic-tangent-shaped detuning and a static detuning. It is suitable for modelling adiabatic passage via a level crossing in the presence of a small detuning shift. \begin{figure}[t] \includegraphics[width=0.96\columnwidth]{fig2.eps} \caption{ Transition probability vs the static detuning shift $\Delta_0$ for sequences of $N=2$, 4, 8, and 16 identical pulses indicated by the boxed numbers. The pulse shape is hyperbolic secant, with a pulse area of $\pi$ and width $T$, and the detuning is constant (Rosen-Zener model). In each frame, the transition probability is plotted for a sequence of pulses with the same phase (thin blue solid curve), and a sequence of pulses with alternating phases (thick red solid curve). The dashed curve shows the no-transition probability for a single pulse and serves as a reference. } \label{fig:rz-even} \end{figure} \section{Resonant $\pi$ pulses: Rosen-Zener model}\label{Sec:RZ} The non-crossing \emph{Rosen-Zener model} is defined as \cite{Rosen1932} \begin{equation} \label{RZ model} \Omega(t) = \Omega_0\, \text{sech}\, (t/T),\quad \Delta(t) = \Delta_0. \end{equation} The Cayley-Klein parameters of the Rosen-Zener propagator are \cite{Rosen1932,Vitanov1994,Vitanov2012} \begin{subequations} \begin{align} a &= \frac{\Gamma(\nu) \Gamma(\nu-\lambda-\mu)} { \Gamma(\nu-\lambda) \Gamma(\nu-\mu)} , \\ b &= -i \frac{\sin(\pi \alpha / 2)}{\cosh(\pi \delta / 2)}, \end{align} \end{subequations} where $\Gamma(z)$ is Euler's gamma function \cite{wolfram} and \begin{equation} \lambda = \alpha/2, \quad \mu = -\alpha/2, \quad \nu = (1+i\delta)/2, \end{equation} with $\alpha = \Omega_0 T$ and $\delta = \Delta_0 T$. For $\Delta_0 = 0$, the Rosen-Zener model describes a resonant pulse of pulse area $\pi \Omega_0 T = \pi \alpha$. Therefore, a resonant $\pi$ pulse is realized with $\alpha = 1$. The presence of the static detuning $\Delta_0$ allows to simulate the effect of a detuning shift on resonant excitation. Of particular interest in the present context is the behavior of the parameter $a$ for small detuning. For a $\pi$ pulse ($\alpha=1$) we have \begin{equation} \label{a-rz} a \approx i \frac{\pi\delta}{2} + \pi\delta^2\ln2 - i \pi \delta^3 \left[ \frac{\pi^2}{24} + (\ln2)^2 \right] + O(\delta^4). \end{equation} From here, it follows that \begin{subequations} \begin{align} \theta &\approx \frac \pi 2 - \pi\delta^2\ln2 + O(\delta^4), \\ \Theta &\approx \pi \delta + O(\delta^3). \end{align} \end{subequations} \begin{figure}[t] \includegraphics[width=0.96\columnwidth]{fig3.eps} \caption{ The same as Fig.~\ref{fig:rz-even} but for an odd number of pulses, $N=3$, 5, 9, and 15. Here the dashed curve shows the transition probability for a single pulse and serves as a reference. } \label{fig:rz-odd} \end{figure} Figure \ref{fig:rz-even} shows the transition probability vs the static detuning shift $\Delta_0$ for sequences of an even number of hyperbolic-secant pulses, each with a pulse area of $\pi$, and Fig.~\ref{fig:rz-odd} shows the transition probability for sequences of an odd number of pulses. The probabilities are calculated from Eqs.~\eqref{P-N} and \eqref{P-N-pm}. Around resonance, a narrow feature forms which gets more narrow as the number of pulses $N$ increases. For an even number of pulses, the feature shows up as a dip (Fig.~\ref{fig:rz-even}), and for an odd number of pulses, it appears as a spike (Fig.~\ref{fig:rz-odd}). This feature is much more narrow for sequences of alternating phases (thick curves) than for sequences of the same phases (thin curves). Using Eqs.~\eqref{PN++}, \eqref{PN+-} and \eqref{a-rz}, it is easy to derive the behavior of the transition probability for $|\delta|\ll 1$. For a sequence of pulses with the same phases, the approximation is \begin{subequations} \begin{align}\label{Peven-p-a} P_{2n} &\approx (2n \pi \ln2)^2 \delta^4 , \\ \label{Podd-p-a} P_{2n+1} &\approx 1 - \tfrac14 \pi^2 \delta^2 + \left[\tfrac{1}{24}\pi^2 - 4n(n+1) (\ln2)^2 \right] \pi^2\delta ^4 . \end{align} \end{subequations} For a sequence of pulses of alternating phases, it reads \begin{subequations} \label{P-m-a} \begin{align}\label{Peven-m-a} P_{2n} & \approx n^2 \pi^2 \delta^2 , \\ \label{Podd-m-a} P_{2n+1} & \approx 1 - (n+\tfrac12)^2 \pi^2 \delta^2. \end{align} \end{subequations} \begin{figure}[t] \includegraphics[width=0.75\columnwidth]{fig4.eps} \caption{ Transition probability vs the static detuning shift $\Delta_0$ for sequences of $N=8$ and 9 identical pulses for sech pulses with a pulse area of $\pi$ and width $T$ (Rosen-Zener model). The two frames are zoomed-in and modified versions of the corresponding frames in Figs.~\ref{fig:rz-even} and \ref{fig:rz-odd}. In each frame, the transition probability is plotted for a sequence of pulses with the same phase (thin blue solid curve), and a sequence of pulses with alternating phases (thick red solid curve). The dashed curves show the approximations for $|\delta|\ll 1$: Eqs.~\eqref{Peven-p-a} and \eqref{Peven-m-a} for $N=8$ pulses, and Eqs.~\eqref{Podd-p-a} and \eqref{Podd-m-a} for $N=9$ pulses. } \label{fig:rz-approx} \end{figure} The four approximate formulas \eqref{Peven-p-a}-\eqref{Podd-m-a} are plotted in Fig.~\ref{fig:rz-approx} and compared to the exact values. A very good agreement between exact and approximate values is observed for $|\delta| \ll 1$, as it should be the case. For a sequence of pulses with alternating phases, the dependence on $N$ and $\delta$ near resonance is the same for an even and odd number of pulses, albeit inverted, see Eqs.~\eqref{Peven-m-a} and \eqref{Podd-m-a}: the transition probability departs as $\propto N^2 \delta^2$ from its resonant value. This means that the excitation profile can be squeezed as much as desired by merely increasing the number of pulses $N$. The behavior of the transition probability for a sequence of pulses with the same phases is rather different: the transition probability departs as $\propto N^2 \delta^4$ from zero for an even number of pulses [Eq.~\eqref{Peven-p-a}], while for an odd number of pulses, its departure from unity is described by a sum of a $N$-independent term $\propto \delta^2$ and a term $\propto N^2 \delta^4$ [Eq.~\eqref{Podd-p-a}]. Consequently, the transition probability is much flatter in the range near resonance than for sequences of pulses with alternating phases. Moreover, the presence of the $N$-independent term $\propto \delta^2$ for odd $N$ impedes the squeezing of the excitation profile by increasing $N$. In either cases, due to the different departure law from resonance, the profile is much broader than for alternating-phase sequences, as indeed seen in Figs.~\ref{fig:rz-even}-\ref{fig:rz-approx}. The conclusion is that, as far as squeezing of the excitation profile near resonance is concerned, the sequences of pulses with alternating phases outperform the sequences of pulses with equal phases and therefore, are much more efficient for sensing of small detuning shifts. Moreover, the simple and accurate approximate formulas, Eqs.~\eqref{Peven-p-a}-\eqref{Podd-m-a}, allow ones to not only sense a detuning shift but also measure it by measuring populations. Any ambiguity, which may arise for a larger detuning shift due to the multiple oscillations for large $N$, can be resolved be making measurements for different $N$. In order to estimate the sensitivity of this technique, consider the sequences with alternating phases, for which the deviation from the resonant value is, in the lowest order, $(N\pi\delta/2)^2$, see Eqs.~\eqref{P-m-a}. By setting this deviation to $\frac12$ and recalling that $\delta = \Delta_0 T$, we find that the half-width-at-half-maximum of the spike or the dip is \begin{equation} \label{delta-rz} \Delta_{\frac12} = \frac{\sqrt{2}}{N\pi T} \approx \frac{0.45}{N T} . \end{equation} Therefore, the sensitivity can be increased by increasing $T$ (which is the well-known Fourier bandwidth argument) or by increasing $N$. For example, a sequence of ten $\pi$ pulses with alternating phases, each of duration 10 $\mu$s, allows one to sense and measure a detuning shift of 4.5 kHz. If, instead of a deviation of $\frac12$, we can measure a population deviation of $\frac1{10}$, then the sensitivity of the same arrangement improves to 2 kHz. As it is clear from the approximations \eqref{Peven-p-a} and \eqref{Podd-p-a}, the excitation profile can be squeezed by sequences of pulses with the same phases too. However, the scaling is much less efficient, $\Delta_{\frac12} \propto 1/N^{\frac12}$, rather than $1/N$. By setting the deviation to $\frac12$, we find from Eq.~\eqref{Peven-p-a} that \begin{equation} \Delta_{\frac12} = \frac{1}{\sqrt{N\pi \sqrt{2}\, \ln2}\, T} \approx \frac{0.57}{\sqrt{N}\, T}. \end{equation} Hence a sequence of ten $\pi$ pulses with the same phases, each of duration 10 $\mu$s, allows one to sense and measure a detuning shift of 18 kHz, a factor of 4 larger than for sequences of alternating phases. Therefore, by merely flipping the phase of every other pulse in the sequence, one can achieve much stronger (quadratically enhanced) squeezing and hence much better sensitivity. \section{Rectangular pulses: Rabi model}\label{Sec:Rabi} One can apply the same approach using pulses of rectangular shape, to the so-called Rabi model, \begin{equation} \label{Rabi model} \Omega(t) = \left\{ \begin{array}{cc} \Omega_0, & |t|\leqq T/2 \\ 0, & |t|> T/2 \end{array} \right. , \quad \Delta(t) = \Delta_0. \end{equation} The Cayley-Klein parameters of the Rabi propagator are far simpler than for the Rosen-Zener model \cite{Shore1990}, \begin{subequations} \begin{align} a &= \cos \left(\tfrac{1}{2} \sqrt{\alpha ^2+\delta ^2}\right)+\frac{i \delta \sin \left(\frac{1}{2} \sqrt{\alpha ^2+\delta ^2}\right)}{\sqrt{\alpha ^2+\delta ^2}} , \\ b &= -\frac{i \alpha \sin \left(\frac{1}{2} \sqrt{\alpha ^2+\delta ^2}\right)}{\sqrt{\alpha ^2+\delta ^2}}, \end{align} \end{subequations} with $\alpha = \Omega_0 T$ and $\delta = \Delta_0 T$. Note that the pulse area is equal to $\alpha$; therefore, a resonant $\pi$ pulse is realized with $\alpha = \pi$. For such a $\pi$ pulse, the Taylor expansion of the Cayley-Klein parameter $a$ reads \begin{equation} a \approx i\frac{\delta}{\pi} - \frac{\delta^2}{4\pi} - i \frac{\delta^3}{2\pi^3} + \frac{\delta^4}{16\pi^3} + O(\delta^5). \end{equation} Hence to the lowest order in $\delta$ we have $a_r = -\delta^2/(4\pi)$ and $a_i = \delta/\pi$. The transition probability for the pulse sequences with alternating phases can be calculated from Eqs.~\eqref{P-N-pm}. For small $\delta$ we find from Eqs.~\eqref{PN+-} \begin{subequations} \label{PN+-Rabi} \begin{align} \mathcal{P}_{2n}^\pm &\approx (2n)^2 \frac{\delta^2}{\pi^2}, \label{PN+-even-Rabi} \\ \mathcal{P}_{2n+1}^\pm &\approx 1 - (2n+1)^2 \frac{\delta^2}{\pi^2}, \label{PN+-odd-Rabi} \end{align} \end{subequations} By setting the transition probability to $\frac12$ and recalling that $\delta = \Delta_0 T$, we find for both odd and even $N$ \begin{equation} \label{delta-rabi} \Delta_{\frac12} \approx \frac{\pi}{\sqrt{2}\, N T} \approx \frac{2.22}{N T} . \end{equation} This value is a factor of about 5 larger than that for the Rosen-Zener model, Eq.~\eqref{delta-rz}. \begin{figure}[t] \includegraphics[width=0.70\columnwidth]{fig5.eps} \caption{ Comparison of the Rosen-Zener and Rabi models for sequences of 10 pulses with alternating phases. The sech and rectangular pulse have the same peak amplitude and the same area $\pi$. } \label{fig:rabi-rz} \end{figure} Figure \ref{fig:rabi-rz} compares the excitation profiles for the Rabi and Rosen-Zener models generated by sequences of 10 $\pi$ pulses of alternating phases. The pulse area of each pulse is $\pi$ and the peak Rabi frequency is the same for each model. Obviously, the feature near resonance is much more narrow for the Rosen-Zener model (a factor of about 5, as noted above). The physical reason is that the rectangular pulses in the Rabi model exhibit typical power broadening, due to its sharp edges, while the smooth pulse in the Rosen-Zener model has no power broadening at all \cite{Vitanov2001,Halfmann2003,Boradjiev2013}. Note that some pulse shapes, with wings vanishing as $t^{-n}$ exhibit even \emph{power narrowing} \cite{Boradjiev2013} and they might provide even better sensitivity than sech pulses. \section{Adiabatic chirped pulses: Demkov-Kunike model}\label{Sec:DK} The level-crossing {Demkov-Kunike model} is defined as \cite{Demkov1969} \begin{equation} \label{DK model} \Omega(t) = \Omega_0\, \text{sech}\, (t/T),\quad \Delta(t) = \Delta_0 + B \tanh(t/T). \end{equation} For $\Delta_0=0$, the Demkov-Kunike model reduces to the Allen-Eberly-Hioe model \cite{Allen1975,Hioe1984,Silver1985}, also known as the \emph{complex-sech pulse} in NMR. It is the most beautiful example of chirped adiabatic passage involving a level crossing. The addition of the static detuning $\Delta_0$ allows to simulate a detuning shift in this model. For $B=0$, the Demkov-Kunike model reduces to the Rosen-Zener model. For $B=\Delta_0$, the Demkov-Kunike model turns into the Bambini-Berman model \cite{Bambini1981}. The Cayley-Klein parameters for this model are \cite{Demkov1969,Vitanov2012} \begin{subequations} \begin{align} a &= \frac{\Gamma(\nu) \Gamma(\nu-\lambda-\mu)} { \Gamma(\nu-\lambda) \Gamma(\nu-\mu)} , \\ b &= -\frac{i\alpha \Gamma(1-\nu) \Gamma(\nu-\lambda-\mu)} {2 \Gamma(1-\lambda) \Gamma(1-\mu)}, \end{align} \end{subequations} where \begin{subequations} \begin{align} \lambda &= \left(\sqrt{\alpha^2-\beta^2} - i\beta\right) / 2, \\ \mu &= -\left(\sqrt{\alpha^2+\beta^2} + i\beta\right) / 2, \\ \nu &= (1+i\delta-i\beta) / 2, \end{align} \end{subequations} with $\alpha = \Omega_0 T$, $\beta = B T$, and $\delta = \Delta_0 T$. In the Demkov-Kunike model, we have for $\alpha=\beta=2$ (corresponding to pulse area of $2\pi$ and chirp rate of $B=2/T$) \begin{equation} \label{a-dk} a \approx 0.086 + 0.165i\delta - 0.052\delta^2 + 0.036i\delta^3 + O(\delta^4), \end{equation} from where we find \begin{subequations} \begin{align} \theta &\approx 1.484 + 0.052\delta ^2 + O(\delta ^4), \\ \Theta &\approx 0.33 \delta + O(\delta ^3). \end{align} \end{subequations} \begin{figure}[t] \includegraphics[width=0.96\columnwidth]{fig6.eps} \caption{ Transition probability in the Demkov-Kunike model vs the static detuning shift $\Delta_0$ for sequences of $N$ identical pulses indicated by the boxed numbers. The pulse shape is hyperbolic secant, with a pulse area of $2\pi$ and width $T$, and the detuning given by a hyperbolic-tangent chirp (with $\beta = 2/T$) and a constant term $\Delta_0$. In each frame, the transition probability is plotted for a single pulse (dashed curve), a sequence of pulses with the same phase (thin blue solid curve), and a sequence of pulses with alternating phases (thick red solid curve). The dotted curves illustrate the approximations \eqref{Pdk-even-m-a} for even $N$ and \eqref{Pdk-odd-m-a} for odd $N$. } \label{fig:dk} \end{figure} The transition probability for the Demkov-Kunike model can be calculated from Eqs.~\eqref{P-N} and \eqref{P-N-pm} and it is plotted in Fig.~\ref{fig:dk}. We are interested again in the feature near zero static detuning which emerges as a spike (for odd $N$) or a dip (for even $N$). Once again, the features produced by the sequences with alternating phases are much more narrow than the features produced by both the single pulse and the sequences of pulses with the same phases. The single-pulse profile is much broader near zero detuning compared to the single-pulse profile of the Rosen-Zener model in Figs.~\ref{fig:rz-even} and \ref{fig:rz-odd}, which manifests the robustness characteristic of adiabatic passage methods. Consequently, it is more difficult to squeeze the excitation profile with a small number of pulses (the frames with 2 and 3 pulses), but for longer pulse sequences the desired squeezing still occurs, especially for pulse sequences with alternating phases. As for resonant pulses, it is possible to sense a detuning shift by using the same chirped pulse used for population inversion, without changing anything except for the sign of the Rabi frequency. One can derive the behavior of the transition probability for small $\delta$ using Eqs.~\eqref{PN++}, \eqref{PN+-} and \eqref{a-dk}. For a sequence of pulses with the same phases, the picture is rather messy (see Fig.~\ref{fig:dk}) and the approximation is not very meaningful. For a sequence of pulses of alternating phases, Eqs.~\eqref{PN+-} and \eqref{a-dk} give the asymptotics $(|\delta|\ll 1)$ \begin{subequations} \begin{align}\label{Pdk-even-m-a} P_{2n} & \approx (0.33n)^2 \delta^2 (1+0.438\delta^2) , \\ \label{Pdk-odd-m-a} P_{2n+1} & \approx 1 - [0.33 (n+\tfrac12)]^2 \delta^2 (1+0.438\delta^2), \end{align} \end{subequations} where higher terms in $\delta$ are retained for better accuracy. These approximations are plotted in Fig.~\ref{fig:dk} by dashed curves. They allow one, as in the Rosen-Zener model, to estimate the sensitivity of this technique. By setting the transition probability to $\frac12$ and recalling that $\delta = \Delta_0 T$, we find for both odd and even $N$ \begin{equation} \label{delta-dk} \Delta_{\frac12} \approx \frac{4.3}{N T}. \end{equation} For example, a sequence of ten $\pi$ pulses with alternating phases, each of duration 10 $\mu$s, allows one to sense and measure a detuning shift of 43 kHz. This is almost a factor of 10 larger than for resonant $\pi$ pulses in the Rosen-Zener model, cf.~Eq.~\eqref{delta-rz} versus Eq.~\eqref{delta-dk}. This is not surprising because the adiabatic passage techniques, as described here by the Demkov-Kunike model, are resilient to parameter variations, including the detuning. For larger values of the pulse area and the chirp, the transition probability becomes even more robust to parameter errors and the sensitivity decreases even more. \section{Conclusions\label{Sec:conclusions}} This paper presented a method for detection and measurement of small detuning shifts generated, e.g., by weak external electric or magnetic fields. The method uses coherent amplification of transition probability errors by a train of identical pulses in two setups: with the same phase of each subsequent pulse, and with an alternating phase shift of $\pi$ from pulse to pulse. Two kinds of pulses were considered: a resonant $\pi$ pulse and an adiabatic chirped pulse, both of which are standard quantum control tools for complete population inversion. In either cases, small detuning shifts do not change the transition probability very much; however, they modify the dynamical phases in the propagator much more significantly, which are amplified and mapped onto the populations by the repeated application of the same pulse. Explicit analytic estimates were derived using the well-known non-crossing Rosen-Zener and Rabi models and the level-crossing Demkov-Kunike model. Based on the analytical results and numerical simulations, it was concluded that sequences of pulses with alternating phases outperform those with the same phases, as far as sensing is concerned: they generate much steeper, and hence much narrower, excitation profiles around resonance, thereby providing much higher sensitivity to detuning shifts. Smooth resonant $\pi$ pulses, exemplified by the Rosen-Zener model, are by far the best performer, with the greatest sensitivity. Alternatively, Gaussian pulses, for which analytic results (albeit not so simple) are also available \cite{Vasilev2004}, deliver similar performance. It is worth considering also pulses of Lorentzian-type shapes (with wings vanishing as $\propto t^{-n}$), which exhibit power narrowing \cite{Boradjiev2013} and may deliver even better sensitivity than sech pulses. Rectangular pulses, represented by the Rabi model, exhibit broader profiles (by a factor of 5) than the Rosen-Zener model due to power broadening generated by the sharp edges of the pulse. Chirped adiabatic pulses, exemplified by the Demkov-Kunike model considered here, are far less suitable for sensing than resonant pulses because of the inherent robustness of adiabatic techniques to parameter variations. Similar results can be obtained for linearly chirped Gaussian pulses for which analytical results are available \cite{Vasilev2005}. Furthermore, using rectangular pulses with a linear chirp, as in the popular Landau-Zener-St\"uckelberg-Majorana model \cite{Landau1932,Zener1932,Stueckelberg1932,Majorana1932} model in its finite version \cite{Vitanov1996}, is an inappropriate option for sensing either because it features both adiabaticity and sharp pulse edges. Therefore, \emph{sequences of smooth resonant $\pi$ pulses with alternating phases} are identified as the most suitable for sensing of small detuning shifts. It is worth emphasizing that the proposed sensing method uses identical pulses (except for the possible $\pi$ phase shift from pulse to pulse). In this manner, no additional uncertainties and errors are introduced which might mask the effects of small level shifts. The proposed technique is very convenient for practical use because the same pulse used as a NOT gate in a quantum circuit can be used to detect detuning shifts. Hence this simple recipe provides an efficient tool for rapid sensing of weak electric and magnetic fields, without sophisticated tomography setups or entangling operations. It is applicable to all kinds of experimental platforms wherein the environment causes energy level shifts. Particularly promising are Rydberg atoms and ions \cite{Saffman2010,Fan2015,Mokhberi2020,Adams2020} due to their increased sensitivity to electric field variations. The present idea is basically the opposite of the idea that the NMR community is pursuing since many years, that is, alternating pulse phases such that the detuning effects are suppressed, termed dynamical rephasing or dynamical decoupling (DD). The simplest such sequence is the Carr-Purcell-Meiboom-Gill (CPMG) two-pulse sequence \cite{Carr1954,Meiboom1958}. An important development is the XY-4 sequence \cite{Maudsley1986,Viola1999}, which uses 4 phase-shifted pulses. The XY-4 sequence is used as a building block for periodic DD, in which it is applied sequentially \cite{Gullion1990}, and concatenated DD (CDD) sequences, which concatenate lower-order CDDs recursively, starting from XY-4 \cite{Khodjasteh2005, Piltz2013,Casanova2015}. Another important development is the concept of robust DD sequences \cite{Souza2011,Genov2017}, which are resilient to various pulse errors. Yet another development is the extension to quantum systems with more than two states \cite{Vitanov2015}. Further details can be found in a comprehensive review \cite{Souza2012}. It is likely that, taking inspiration from the vast DD literature, the straightforward flips in the pulse phase in the sensing method proposed here could be optimized further, by letting the relative phases from pulse to pulse be free control parameters. However, this task is well beyond the scope of the present work. It should be clear that the sensing method presented here is based upon quantum interference. Therefore, the measurement should be fast enough in order to avoid noise and hence dephasing. Finally, in this work only detuning errors have been considered. In a real experiment, Rabi frequency errors might occur too, which would turn the parameter estimation problem into a multi-parameter one. Therefore, the method is strictly applicable in the absence of such errors. However, the underlying assumption is that the $\pi$ pulses used in the sensing sequence is of very high fidelity because it is used in some quantum circuit, which would be the main experiment, i.e. the loss of probability does not come from it but from the external ambient fields. The present method allows one to quickly measure, once in a while, whether the external field has changed. More importantly, a closer inspection shows that small Rabi frequency errors do not affect the method as far as detuning sensing is concerned. It is only important that the Rabi frequency does not change during the sensing sequence, even if it is slightly different from its nominal $\pi$ pulse value. \acknowledgments This work is supported by the Bulgarian Science Fund Grant DO02/3 (Quant-ERA Project ERyQSenS).

1. Introduction and Background {#sec1} ============================== Increasing usage and production of handheld devices and smart-phone have pushed Muslim community to create the Holy Quran in the digital form. Thus, authentication of digital Quran becomes an emerging issue because the current digital Quran that is mostly in mobile applications is unverified. Even though the digital Quran is verified by the Islamic authority, still there is a problem such as falsifying some parts of Quran\'s verses. In such case, the readers are unable to validate the verse, whether it is correct or fake, due to an accidental typo or an intentional act. History tells that according to investigation in July, 2005, several secret scanned documents of World War II at the National Archives have been altered during or after the 2000s \[[@B1]\]. Hence, protecting a digital document such as the Holy Quran is important nowadays. In this regard, this study proposes a solution to protect the digital Quran and to localize the tamper region on the digital Quran images if present. The proposed solution is inspired by the digital watermarking concept. Digital watermarking means that the digital data embedded with a predefined authentication code remains undetectable to human eye but effortlessly identified by the specified algorithm. The major focus is to shield the integrity, security, and fidelity of the digital content such as that of the digital watermarking promisingly applicable for the present electronically driven world \[[@B2]\]. Particularly, a schema known as fragile watermarking has attracted great attention for authentication and integrity of the digital content \[[@B3]\]. There are several known requirements for a fragile watermarking method that must be considered including detecting common image forgery, geometric transformations, signaling elimination of original image, putting new objects, and notifying once image processing operations present. Moreover, it is better to authenticate the media without referring to the original image; this is known as blind detection \[[@B4]\]. Numerous methods recognized that attacks such as the copy-and-paste \[[@B5]\], vector quantization \[[@B6]\], the Holliman-Memon counterfeiting attack \[[@B7]\], or collage attack \[[@B8], [@B9]\] should be considered by the fragile watermarking method. It is known that watermarking methods alter the host image after inserting the authentication code. This situation is unsuitable for sensitive applications such as medical imaging and military purpose \[[@B10]\]. However, most of the image applications can accept the degradation of the image quality as long as the original and watermarked images are perceptually comparable or the content is not influenced. Thus, watermarking the Holy Quran should be done carefully to avoid any influence on the verses. 2. Related Works {#sec2} ================ Digital watermarking is generally categorized into three kinds named fragile, semifragile, and robust watermarking. Fragile watermarking has the characteristic of being easily broken even if small forgery is present. Quite similar to fragile watermarking, semifragile watermarking has small difference, which is being robust to nonhuman forgery such as JPEG compression. Attractiveness of JPEG compression becomes a key of the exploration aspect on semifragile watermarking. The last category named robust watermarking is intended to make it hard to break against any tamper activities. Thus, robust watermarking is suitable for copyright protection of the digital images \[[@B25]\]. Meanwhile, the fragile and semifragile watermarking are commonly applied for authentication of the multimedia content, including video, audio, and still image \[[@B26], [@B27]\]. Some watermarking methods in the literature show that it can localize the tamper regions, and other methods are only able to inform if the still image is authentic or tampered. Two concepts, namely, pixel-based and block-based approach, are adopted in watermarking techniques when handling the host image. Pixel-based concept treats the host image pixel by pixel to embed the watermark. The block-based concept treats the host image block by block. These approaches, pixel-based and block-based, can be implemented in two domains known as spatial and frequency domain \[[@B10]--[@B33]\]. Watermarking technique in the spatial domain has several security issues other than the frequency domain, including the following.Such techniques work on very limited space of an image at the pixel level, depending on the color depth.Authentication can be easily passed using some image distortion such as compression or noise.Natural statistical property in the color image that has linked components in the color information (e.g., RGB and CMYK) would be distorted after watermarking process makes it open to image attacks. Watermarking method in pixel-based concept \[[@B26], [@B29]--[@B31]\] is exposed to the brute-force attack because the watermark is commonly hidden into the least significant bits (LSBs) \[[@B23]\]. In addition, such method is suitable only to localize the image forgery on some of the most significant bits. On the other hand, the block-based concept has issues for parameter of the block sizes and watermark payloads. Some experiments are required to determine the proper parameter that facilitates acceptable tamper detection while maintaining the image quality. However, the main drawback in block-based concept is being unable to locate the tampered pixels accurately; this might be important for specific applications such as in the military communication \[[@B24]\]. In addition, Preda \[[@B27]\] found that the latest watermarking methods are susceptible to forging attacks and inaccurate when dealing with unintentional image tampering. Recent fragile watermarking methods are reviewed to comprehend the state of the art. Many researchers focused on watermarking method for binary image \[[@B34]\] and grayscale image \[[@B35]--[@B38]\] and according to \[[@B39], [@B40]\] only limited are concerned about color images. Even though their method is robust with minimum complexity, the original image must be present during authentication process or extracting the watermark information. Wong \[[@B41]\] introduced a block-based approach along with the public-key scheme. The host image is presented as LSB-zeroed prior to embedding the binary watermark image. Such method has high security and can handle crop and scale attacks. Their block-based approach does not have any correlation between the neighboring blocks or any blocks within the watermarked image. Thus, this method is exposed to several attacks, including vector quantization, cover-up, and transplantation. Wu and Liu proposed DCT-based watermarking method \[[@B42]\] that embeds a binary watermark into the DCT coefficients. A look-up table is defined and utilized to map the DCT coefficients into zero or one. There are two issues in their method. First, they must transmit the look-up table in a secure channel to perform authentication everywhere. Second, their block-wise approach is independent and hence vulnerable to the same attacks like Wong\'s method \[[@B41]\]. Li et al. \[[@B43]\] proposed a block-wise method that has dependence between neighboring blocks. Neighboring blocks are combined together to establish dependence among them. The authentication information is extracted from the host image as a binary feature map. Authentication can be done without the original image. They claimed that the method was resistant to cover-up attacks and vector quantization. Unfortunately, the watermark is embedded into LSB that is known weak to brute-force attack. Moreover, contextual dependence that was generated based on deterministic information between blocks is vulnerable to transplantation attack because the contextual dependence is established based on deterministic information. Li et al. \[[@B44], [@B45]\] tried to avoid transplantation attack and reported that their method provided tamper detection at the pixel level. However, their method worked on spatial domain, which is also less secure to brute-force attack. Barreto et al. \[[@B46]\] focused on transplantation attack; they generated the watermark from a block along with their neighbor using hash function plus some random data. It makes the watermark information nondeterministic and distinctive. Nevertheless, accuracy of the tampering detection is influenced by the block size. In addition, their method is applied on spatial domain and hence vulnerable to brute-force attack. Other researchers attempt to avoid spatial domain by introducing transform-domain schemes \[[@B47]--[@B49]\]. Winne at el. \[[@B47]\] watermarked the coefficient of high-frequency subbands of luminance component. Their method achieved better localization and less embedding distortion. Quite similar \[[@B49]\], generating the watermark from LL component and embedding it into LL component too, The watermarked image was reported imperceptible by Xie and Arce \[[@B49]\]. Fridrich et al. \[[@B50]\] considered quantized DCT coefficients to generate the watermark. Then, the watermark is embedded into DCT coefficients. However, their method is not intended to locate the tampering pixels but rather it can tell whether the image is authentic or has been tampered. As discussed above, the performance of watermarking method can be analyzed through the image quality metrics and performing image attacks. For example, fragility of the watermark can be evaluated using collage attack \[[@B4]\]. In such attack, authenticated blocks of known watermarked image are copied into another image to create forgery content that can pass the authentication process. In the literature, a watermark hidden in wavelet domain shows resistance to brute-force attack \[[@B26], [@B32], [@B33]\]. Moreover, imperceptibility of the watermark has reported promisingly with satisfactory PSNR (peak signal-to-noise ratio) in the wavelet-based strategy \[[@B25], [@B11], [@B12], [@B28]--[@B33]\]. Hence, wavelet-based strategy has been recognized to shield the digital content against forgery. In this regard, the proposed method prefers wavelet-based strategy to achieve better authentication schema. In addition, chaotic maps \[[@B59]\] have appealed further attention to improve the digital watermarking \[[@B51]--[@B53]\]. It is because chaotic maps have properties that are sensitive to initial and parameter values and show chaotic behavior. In this regard, chaotic maps are considered in this study to increase the security of proposed fragile watermarking. The contributions of this paper are a novel fragile watermarking method that utilized the discrete wavelet transforms (DWT) prior to embedding the watermark and chaotic maps to encrypt the watermark information. This method is applied to protect and authenticate the digital Holy Quran and to be able to locate the tamper region if present. The rest of the paper is structured as follows: [Section 2](#sec2){ref-type="sec"} explains the proposed watermarking method in the wavelet domain and illustrates the diagrams of embedding and authenticating the method, [Section 3](#sec3){ref-type="sec"} covers the experimental results, and the last section concludes the paper. 3. The Proposed Watermarking Method in the Wavelet Domain {#sec3} ========================================================= This section describes the proposed fragile watermarking method. [Figure 1](#fig1){ref-type="fig"} presents the diagram of the proposed embedding process and [Figure 2](#fig2){ref-type="fig"} shows the diagram of the proposed authentication process. The proposed watermarking method is secured by using the chaotic map to blur the authentication code. Hence, intruder is impossible to generate correct authentication code even if they know the initial parameter of the chaotic map because it has random-like behavior. Chaotic maps are attractive because a small difference in the initial condition would produce a huge difference and they have large variation range \[[@B54]\]. Such characteristics fulfill the classic Shannon\'s theory for information hiding \[[@B55]\]. Recently, chaotic maps have become popular because they have been proved to enhance the security for information hiding \[[@B56]\]. In this study, two chaotic maps are combined to encrypt the watermark information prior to embedding into the wavelet domain. First chaotic map is used to produce a sequence key and the second is used for data encryption. Such combination offers the subsequent advantages, including being resilient to the fixed length of word that influences by the chaotic sequence, it greatly volatile and it resistant to attacks \[[@B54]\]. The major idea behind this study is to embed the authentication code in the first level of 2D Daubechies discrete wavelet transform. The DWT is a distinguished transformation method that has drawn attention particularly because of the image compression (JPEG2000). 2D DWT used high-pass and low-pass filters to decompose the image into wavelet coefficient, horizontal, vertical, and diagonal details. The DWT-based is considered in this paper rather than DCT- based as the wavelet transform mimics the human vision system (HVS) more similar than the DCT \[[@B21]\]. DWT does not decompose and process the image block by block thus minimizing the image artefacts unlike DCT. DWT clearly splits high- and low-frequency information with respect to pixel by pixel basis \[[@B27]\]. Embedding the watermark into the transform values will only alter the image locally since DWT has a spatial frequency locality characteristic. It is known that changing the coefficients in the wavelet domain is more likely to be undetectable contrasting DCT and FFT. DWT also offers spatial and frequency description of an image. Hence, DWT provides a good basis for hiding the watermark while preserving the image quality \[[@B27]\]. The proposed method is designed to work block by block on the wavelet domain. Each block that consists of the wavelet coefficients is processed consecutively with encrypted watermarks entirely over the image. Hence, particular blocks are able to share duplicate authentication code and create relation between those blocks. This relation makes it hard for the intruder to tamper the watermarked image without breaking the watermarks. 3.1. Encryption of the Authentication Code Based on Chaotic Map {#sec3.1} --------------------------------------------------------------- Each image of the digital Holy Quran is different for each page. This is because a single page of the Quran image is formed by numerous unique verses. In addition, Quran images include a border to prettify the pages; the border is commonly exclusive among the digital Quran. Hence, such characteristics make the Quran image have a great chance to produce the secret key that is required by the chaotic map. The chaotic trajectory is sensitive to its parameter value and the initial condition. In this regard, random pixel values are selected from the Quran image. The encryption of authentication code begins by secret key\'s generation, then followed by the encryption process. The steps are detailed as follows.(i)Take randomly one pixel value from *x*-axis and one pixel value from *y*-axis of the Quran image, defined as *p* ~*x*~ and *p* ~*y*~. Nonzero pixels are considered for those values.(ii)Apply equation below to obtain *Q* value that is later used for generating the parameter value and initial condition, the equation being as follows: $$\begin{matrix} {Q = \frac{\left| {p_{x} - p_{y}} \right|}{N},\quad 0 \leq Q \leq 1,} \\ \end{matrix}$$ where *N* is the pixels amount of the Quran image.(iii)Equations below proposed by Phan \[[@B17]\] are utilized to generate the parameter value: $$\begin{matrix} {S_{k} = \begin{cases} {3 + Q,} & {Q > 0.57,} \\ {3.75 + \left( {0.43 - Q} \right),} & {Q \leq 0.57.} \\ \end{cases}} \\ \end{matrix}$$(iv)Initial condition for the chaotic map is defined using equation as follows: $$\begin{matrix} {R_{n} = \left| {\left\lceil Q \right\rceil - Q} \right|,\quad 0 < R_{n} \leq 1.} \\ \end{matrix}$$(v)Generate a sequence of real numbers using logistic map. That sequence is later used for parameter value of Henon map, $$\begin{matrix} {R_{n + 1} = \mu R_{n}\left( {1 - R_{n}} \right),\quad n = 1,2,3,\ldots,} \\ \end{matrix}$$ where *n* is the map iteration index and the previously calculated secret keys *S* ~*k*~ and *R* ~*n*~ are used as the parameter value and initial condition of the logistic map, respectively. The logistic map has proven to be very sensitive to initial value, nonconvergent, and nonperiodic when 3.57 \< *μ* ≤ 4.0 \[[@B58]\].(vi)Henon map \[[@B14], [@B18]\] is employed to encrypt the authentication code; the generalized equation is presented below: $$\begin{matrix} {t_{n + 1} = \left\lbrack {1 + b\left( {t_{n - 1} - c} \right) + 379R_{n}^{2}} \right\rbrack\left( {{mod}{\,\,}1} \right),} \\ \end{matrix}$$ where *b* and *c* are Henon map parameters that are specified by *b* = 0.3 and 1.07 ≤ *c* ≤ 1.09. Chaotic maps are constrained within the limit by performing modulo operation (mod⁡  1). In addition, such operation retains the sequence convergence. The generated sequence is real numbers and hence it is quantized into binary system using simple threshold as below: $$\begin{matrix} {c\left( n \right) = \begin{cases} {1,} & {t_{n} \geq 0.5,} \\ {0,} & {t_{n} < 0.5.} \\ \end{cases}} \\ \end{matrix}$$(vii)Finally, the binary sequence *c*(*n*) generated above is used to encrypt the generated authentication code. The authentication code is produced by feeding the Quran image into hash function; the authentication code is defined in binary systems as *h*(*n*) ∈ {0, 1}. Then, using equation below the encrypted authentication code is attained: $$\begin{matrix} {l\left( n \right) = {\sum\limits_{n = 1}^{N}{h\left( n \right) \oplus c\left( n \right)}},} \\ \end{matrix}$$ where *N* is length of the authentication code. The decryption of authentication code is simply an opposite procedure of encryption process as explained above. Using same secret keys, the decryption is using the following formula: $$\begin{matrix} {h^{\prime}\left( n \right) = {\sum\limits_{n = 1}^{N}{l^{\prime}\left( n \right) \oplus c\left( n \right).}}} \\ \end{matrix}$$ 3.2. The Proposed Watermark Embedding Process {#sec3.2} --------------------------------------------- The encrypted authentication code is embedded into the host image according to steps below.(i)Firstly, the host image is brought into wavelet domain by performing 2D discrete wavelet transforms using Daubechies function. The decomposition generates *L* resolution levels.(ii)Decomposition process using DWT will produce four matrices defined as wavelet coefficients, *LH* ~*p*~, *HL* ~*p*~, and *HH* ~*p*~. The resolution level is determined with *p* × *L*. In this paper, the decomposition level of the *p* = 1 is considered. According to Run et al. \[[@B20]\], selecting higher *p* value to produce higher subbands level tolerates a greater tampering detection, but it decreases the localization accuracy against image tampering.(iii)After decomposition, one decompose matrix called wavelet coefficients *CW*  are rounded into closest integer. Then, the matrix is fragmented into small size of nonoverlapping block. Such fragmented blocks ensure that matrix of wavelet coefficients parallel to the same spatial spot will be inserted with watermark code. The payload of watermark code in a block is controlled using *n* parameter. The block size should be determined properly to allow sufficient watermark payload and to maintain fine image quality. The suitable block size also ensures optimum localization capability on tampered region. Equation below is used to calculate the block size *b*: $$\begin{matrix} {b\left( n \right) = \left\lceil \sqrt{2^{n}} \right\rceil.} \\ \end{matrix}$$(iv)The watermark information that is already prepared as described in [Section 3.1](#sec3.1){ref-type="sec"} is utilized. Such watermark resists local attacks problem because the information has been blurred using chaotic map \[[@B20]\]. The bits of watermark are defined as *l* ~*m*~. It is known that the watermark length of *l* ~*m*~ is influenced by the selected hash function. In this regard, simple function is required to select the bits sequentially with respect to *n* parameter (*n* value is same as in step (iii)) as defined below: $$\begin{matrix} {W = \begin{cases} {\text{LSBit}_{1\ldots n},} & {n \leq 8,} \\ {\text{LSBit}_{1\ldots 8},} & {n > 8.} \\ \end{cases}} \\ \end{matrix}$$(v)The selected authentication bits *W* are hidden into the wavelet coefficients using weighted-sum function, *f*(*x*) \[[@B13]\]. The block of wavelet coefficients is fed into weighted-function and then the resulting value is compared with selected authentication bits; both values are treated in decimal form. A loop is performed until the weighted-sum value is equal to the selected authentication bits. The value in a block *B* of wavelet coefficient is modified ±1 with respect to function as follows: $$\begin{matrix} {S = W - f\left( x \right),} \\ {B\left( d \right) = \begin{cases} {B_{|s|} + 1,} & {0 < \left| S \right| \leq 2^{n},S > 0} \\ {B_{|s|} - 1,} & {0 < \left| S \right| \leq 2^{n},S < 0} \\ {B_{|{s - 2^{m + 1}}|} + 1,} & {2^{n} < \left| S \right| \leq 0,S > 0} \\ {B_{|{s - 2^{m + 1}}|} - 1,} & {2^{n} < \left| S \right| \leq 0,S < 0.} \\ \end{cases}} \\ \end{matrix}$$(vi)After processing all blocks of the wavelet coefficients, the new wavelet coefficients are stored as *CW* ~1~ *\'*. Afterward, *CW* ~1~ *\'* along with three decomposition matrixes (*LH*, *HL*, and *HH*) are inverted back into spatial domain using 2D-Inverse DWT. The new image *I* ~1~ *\'* is decomposed using 2D-DWT to obtain the wavelet coefficient, stored as *CW* ~2~. Using equation below matrix *D* is calculated: $$\begin{matrix} {D = CW_{1}^{\prime} - CW_{1}^{\prime\prime}.} \\ \end{matrix}$$(vii)The matrix *D* as generated above consists only of three kinds of values {−1, 0, + 1}. Since matrix *D* is halftone of the new image size then it requires stretching the matrix such that it has the same size of the image size. Such aim can be done by multiplying the matrix size by two and filling the gap with zero value. Further, the stretched matrix *D* is embedded into spatial domain of new image *I* ~2~ by altering the least significant bits of its pixels with value in matrix *D*. 3.3. The Watermark Authentication Process {#sec3.3} ----------------------------------------- The authentication process is allowed by tamper localization on the protected image if present. This process can be seen as reverse of the embedding process. The following steps describe the authentication process.As explained above, two watermarks are embedded separately on wavelet domain and then on spatial domain. It requires extracting them in reverse way. In spatial domain, the watermarks are hidden in least significant bits of pixels with respect to the matrix *D*. In this regard, the LSB of each pixel is extracted by taking the two insignificant bits; the value is then saved in matrix *E*.Matrix *E* must be suppressed into half size because it contains gap values of zeros. The new image after removing the insignificant bits is stored as *I* ~*e*~.The new image *I* ~*e*~ is decomposed into wavelet domain using 2D-DWT. It decomposes into first level subbands same as on watermarking process. Hence, it generates first wavelet coefficient, *LH* ~1~, *HL* ~1~, and *HH* ~1~. The generated wavelet coefficient *CW* ~1~ is floored into nearest integer. Afterward, *CW* ~1~ is added with matrix *E* to generate new wavelet coefficient *CW* ~2~.The *CW* ~2~ is processed block-by-block with block size same as defined during watermarking process. The weighted-sum function is utilized to obtain weight value of the block. The weight value is decrypted prior to comparing with authentication bits that are stored in database. The valid block should have same value with the selected authentication bits; otherwise it is categorized as tampered block. The wavelet coefficient value of tampered block is modified into zero values to localize the tamper region. Finally, authenticated image is generated by transforming the wavelet coefficient using 2D-DWT. A black box will exist if any forgery is present. 4. Implementation and Experimental Results {#sec4} ========================================== The performance of the proposed method is assessed with two datasets taken from popular android applications. Each one of the datasets consists of 604 images of the Quran pages. The file formats for the datasets A and B are JPEG and PNG, respectively. Dataset A has a border for each page and the color defined in RGB spaces. Dataset B does not have a border and is presented in the indexed grayscale image. [Figure 3](#fig3){ref-type="fig"} presents some Quran images and [Table 1](#tab1){ref-type="table"} detailed the characteristics of each dataset. The experiment is conducted based on the common methodology as reported throughout the literature \[[@B11]--[@B16], [@B28]\]. The assessment includes localization accuracy against image forgery and measures image quality after watermarking process. The result is presented for four pages only, selected from pages 1, 3, 50, and 601, for the sake of simplicity of the presentation. These four pages are considered because they represent the whole Quran pages. Page 1 consists only of one chapter and it is like page 50. Page 3 is almost similar with all pages of the Holy Quran. Page 50 shows the beginning of a new chapter and page 601 consists of several chapters in the page. The main concern of fragile watermarking method is the ability to detect the manipulation while preserving the host image after watermarking. Hence, several methods of image quality measurement are utilized to determine the quality of the watermarked image. The image quality methods are as in [Table 2](#tab2){ref-type="table"}. The image quality after the watermarking process is measured using the above-mentioned equations. In the literature, some methods have reported the reference value that indicates adequate image quality. The known acceptable values are shown in [Table 3](#tab3){ref-type="table"}. Proposed fragile watermarking method should be capable of detecting any manipulation and locate the tampered location. Certain form of manipulation can be unintended; for example, the usage of the compression schema and other manipulations can be intended including resizing, cropping, rotation, and other image manipulations. Those image manipulations can be summarized as follows.Lossy compression such as JPEG and MPEG practically damages the image quality through irrecoverable loss of information.Geometric distortions alter the image symmetry and include operations such as cropping, rotations, scaling, and translation.General signal processing procedures are such as image filtering, resampling, color reduction, local conversation of pixels, and adding of constant offset to the pixel values.Other intentional attacks are done by human hand, for example, performing watermarking on watermarked image or scanning watermarked image to produce same image that bypassed the watermark information. As summarized above, six common image manipulations that covered the above-mentioned image manipulations are considered to demonstrate the performance of the proposed approach. The considered manipulations are only applicable for Quran pages, which should not alter the content extreme (e.g., rotation or flip the image). The considered image manipulations applied on the dataset are presented as in [Table 4](#tab4){ref-type="table"}. Please note that the pixel manipulation is too small. Therefore, it is hard to see by using naked eye. Then, replacement manipulation is executed by exchanging one verse of watermarked Quran image with another verse. It is clear that such manipulation will distort the content. Such replacement is a challenge to the nonprofessional to be discovered because of the lack of knowledge regarding the authentic Quran. The five image manipulations (except JPEG compression) that are applied on Quran image page number 3 are presented in [Table 5](#tab5){ref-type="table"}. As mentioned in [Section 3](#sec3){ref-type="sec"}, the proposed method requires a parameter for the block size, defined as *n*. The experiment will be conducted for a range of *n* values, where *n* = {1,2,..., 15}. Each of *n* values is subjected to the above-mentioned image manipulations. Initially, the experiment is carried out to evaluate the PSNR value of watermarked image, because this metrics is commonly used and suitable to measure the performance of digital watermarking method. The proposed method can achieve PSNR value above 42 dB \[[@B3], [@B13]\] after *n* \> 3 as shown in [Table 6](#tab6){ref-type="table"} and [Figure 4](#fig4){ref-type="fig"}, for both datasets A and B. It confirms that the level of image distortion following watermarking process is low enough even when considering a small *n* value. The first important feature of the fragile watermark is imperceptibility. In order to discover the difference between the watermarked image and original image, hence the SSIM metric can be utilized to inspect the visual similarity \[[@B21]\]. SSIM has proved to be in line with the human visual system and is able to evaluate the relationship of the two different images, including image contrast, image brightness, and three aspects in the image structure. The SSIM value during experiment shows near to one when *n* \> 8, as depicted in [Table 7](#tab7){ref-type="table"} and [Figure 4](#fig4){ref-type="fig"}. Hence, the proposed method is able to preserve the image quality positively. Afterward, the experiment is conducted by performing the dewatermarking process minus any image attacks to verify whether the authentication process can be accomplished properly or not. The result indicates that the watermarked image can pass the authentication perfectly for every *n* value to the parameter of proposed method. The *n* parameter is ranging from 1 to 15. The percentage of authentic blocks calculated by dividing the authenticated blocks with overall quantity of all blocks is 100% for both datasets A and B. It can be concluded that the proposed method can perform embedding and extracting of the watermark on host image perfectly. The next evaluation is carried out to study the capability of proposed method with respect to the pixel manipulation as reported in [Table 5](#tab5){ref-type="table"}. [Table 5](#tab5){ref-type="table"} of the pixel manipulation shows a particular region of the watermarked Quran image that has been altered, which is a single dot of the verse deleted. Such manipulation undoubtedly annoys the integrity of the Quran\'s content as the dot(s) has significance in Arabic alphabets. The parameter *n* value within range {1, 15} is tested to understand the influence against tamper detection. Tables [8](#tab8){ref-type="table"} and [9](#tab9){ref-type="table"} show PSNR and BER results. Figures [5(i)](#fig5){ref-type="fig"}--[5(vi)](#fig5){ref-type="fig"} show the location of the tampered region in the image as black box. Image quality metrics confirm that after attack the quality is decreased and tends to have a constant value as depicted in [Figure 6](#fig6){ref-type="fig"}. The fragility reported in [Figure 7](#fig7){ref-type="fig"} that shows on JPEG dataset pixel manipulation can be detected under any *n* value. Meanwhile, fragility on PNG dataset can achieve 100% only after *n* ≥ 5. The Gaussian filter is applied on the watermarked image, and the fragility is measured. The filter is blurring the Quran image that makes some verses unreadable or vague. [Table 10](#tab10){ref-type="table"} shows the image quality measurement to indicate the image quality after attack and [Table 11](#tab11){ref-type="table"} reports that the BER results indicated how fragile the proposed method is. As expected, changes in the magnitude of gradient vectors due to Gaussian filter can be detected properly. Hence, the proposed method is fragile to blurring attack. Blurring attack has decreased the image quality as reported in [Figure 8](#fig8){ref-type="fig"}, the PSNR result and other metrics shows obtain lower values compare with metrics\' values before apply blurring attack. BER for both datasets are shown in [Figure 9](#fig9){ref-type="fig"} and finest fragility can be achieved only after *n* ≥ 8. The Quran image should not contain any noise because a dot is a matter in Arabic letters. Hence, a Gaussian noise that is commonly used to add noises in image processing is considered. The Gaussian noise does not have structure and it is difficult to eliminate such noises without suffering modifications to the image itself. Tables [12](#tab12){ref-type="table"} and [13](#tab13){ref-type="table"} show the PSNR and BER results of the proposed method, respectively, after Gaussian noise is inserted to the watermarked images. The result confirms that the proposed fragile watermarking method is capable of localizing the noise attack. Image quality is decreased after this attack as presented in [Figure 10](#fig10){ref-type="fig"}. [Figure 11](#fig11){ref-type="fig"} reported fragility on PNG dataset better than JPEG dataset. Fragility on JPEG dataset is superior after *n* \> 6. The succeeding considered attack is median filtering attack. Such attack is known as nonlinear attack and alters the gradient vector of image much more than Gaussian filter. The median filter does not change the edge points unlike Gaussian filter. The PSNR and the BER results of proposed method under median filtering attacks are presented in Tables [14](#tab14){ref-type="table"} and [15](#tab15){ref-type="table"}, respectively. As mentioned before, the median filters are executed with 3 × 3 window. The result demonstrates that tamper localization after a median filtering attack is very efficient in both dataset A and dataset B. This attack degrades the image quality more than previous attacks as presented in [Figure 12](#fig12){ref-type="fig"}. The proposed method can attain acceptable fragility only after *n* ≥ 8 for both datasets as depicted in [Figure 13](#fig13){ref-type="fig"}. The collage attack is one of the security issues on the fragile watermarking method that needs to be concerned. In this regard, the watermarked image is altered by replacing the verse line number 9 with another verse. The pasted versed is taken from watermarked image and hence it has authentication code. Figures [14(i)](#fig14){ref-type="fig"}--[14(iv)](#fig14){ref-type="fig"} and [15(i)](#fig15){ref-type="fig"}--[15(iv)](#fig15){ref-type="fig"} report the position of the altered region in both datasets as black box. The result demonstrates even with small *n* value the altered region can be detected but not accurately. The proposed method can localize the collage regions correctly when the parameter *n* \> 4. The NHD metric is utilized to estimate the localization performance by considering the modification between watermarked image and attacked image. The NHD is shown in [Figure 16](#fig16){ref-type="fig"}; the values expressing the proposed method can detect the collage attack as long as the *n* parameter is greater than or equal to 8. Finally, JPEG compression is tested on proposed method to evaluate the fragility against nonmalicious attack. The evaluation is considered from soft to modest JPEG compression with quality factor (QF) of JPEG compression within range {50, 100}. PSNR results are presented in [Figure 17](#fig17){ref-type="fig"}; the compression does not affect the image quality. The BER of the watermarked image following authentication process is measured. [Figure 18](#fig18){ref-type="fig"} reported the BER results for *n* = 1 to *n* = 15. As expected, the BER value is high which means proposed method is fragile to JPEG compression with respect to any *n* value. Such fragility is contributed from embedding process that considers the coefficients matrix of the first wavelet subbands. The first wavelet subbands are altered during the JPEG compression. A proper value for *n* parameters can vary from one application to another. Nevertheless, it is suggested that *n* value equal to eight (8) is applicable for general digital watermarking purposes. 5. Conclusion and Future Work {#sec5} ============================= Based on the literature review, authors found no study attempt to explore the digital watermarking for the Holy Quran images. Most of the recent digital watermarking methods are applied on natural images, for example, Lena, baboon, and other popular images. Hence, there are a gap and an open issue related to preserving the integrity of the Holy Quran. This paper brings novel issue to the society and presents one of the solutions to protect the Holy Quran from common image tampering. The proposed method is intended as fragile watermarking that considers both wavelet domain and spatial domain. The first layer of watermarking method is applied on wavelet domain and second layer on spatial domain. The host image is transformed into wavelet domain using discrete wavelet transformation (DWT) prior to hiding the watermark. The recent methods demonstrated that such watermarking schema in the wavelet domain is robust to brute-force attack rather than watermarking method in the spatial domain. Furthermore, proposed method can be categorized as a block-based approach and it introduces correlation between the blocks that make resistance to collage attack. The two-layer watermarks also guarantee challenging the intruder to bypass the authentication without breaking the watermarked image. Thanks are due to chaotic maps that are very sensitive to initial value, demonstrate greatly blurred authentication code, and protect the watermark against local attacks. According to the result of image quality metrics, the proposed method obtained satisfactory image quality after embedding the procedure. Finally, several image attacks that applied into the watermarked image demonstrated that the proposed method yields promising localization of the image tampering at a minimum watermark payload. In the future, analysis on other attacks can be considered such as blurring attack, noise attack, and other known attacks. Further, the study should work on various image formats, for example, JPEG2000, TIFF or BMP formats. Investigation on the image quality after watermarking process, using metrics other than PSNR and SSIM, is important. Implementing watermarks on higher level of decomposition can also be explored. Finally, the security can also be improved by improving the encryption part. The authors would like to thank and acknowledge the IT Research Center for the Holy Quran (NOOR) at Taibah University for their financial support during the academic year 2012/2013 under research Grant Reference no. NRC1-75. Conflict of Interests ===================== The authors declare that there is no conflict of interests regarding the publication of this paper. {#fig1} {#fig2} {#fig3} {#fig4} {#fig5} {#fig6} {#fig7} {#fig8} {#fig9} {#fig10} {#fig11} {#fig12} {#fig13} {#fig14} {#fig15} ![NHD result on collage attack, *n* = \[1,15\].](TSWJ2014-803983.016){#fig16} ![PSNR before and after JPEG compression, QF = \[50,100\].](TSWJ2014-803983.017){#fig17} ![Bit error rate result against JPEG compression, QF = \[50,100\].](TSWJ2014-803983.018){#fig18} ###### Characteristics of the datasets.   Dataset A Dataset B ----------------------------- ------------------------- -------------------------- Source Quran Kareem \[[@B22]\] Quran Android \[[@B22]\] Format JPEG PNG Image compression Lossy Lossless Dimensions (width × height) 547 × 793 800 × 1294 Number of pages 604 604 Min. file size 67 KB 30 KB Max. file size 78 KB 100 KB Total dataset file size 42.6 MB 51.1 MB Border on each page Present No border Color image Yes (full color) Grayscale (limited) ###### Common image quality metrics. -- -- ###### Known acceptable value in image quality measurement. Number Image quality measurement Known acceptable value -------- ------------------------------------ -------------------------------------------------- 1 Peak signal-to-noise ratio (PSNR) ≥42 dB, greater than or equal to 42 db is better 2 Universal image quality index \~1.0, near to 1.0 is perfect 3 Signal-to-signal noise ratio (SNR) ≥20 dB, greater than or equal to 20 db is better 4 Structural similarity (SSIM) \~1.0000, near to 1.0000 is perfect 5 Normalized Hamming distance (NHD) =1, equal to 1 is perfect ###### Six image manipulations applied on datasets. Number Image manipulation Parameter JPEG PNG -------- --------------------- --------------------------------------------------------------------------- ------ ----- 1 Pixel manipulation Window size 10 × 10 *✓* *✓* 2 Blurring Gaussian filter, sigma = 0.5 *✓* *✓* 3 Noise Gaussian noise, 150 blocks, block size 3 × 3, mean = 0.4, variance = 0.01 *✓* *✓* 4 Median filtering Filter size 3 × 3 *✓* *✓* 5 Replacement/collage --- *✓* *✓* 6 JPEG compression Quality factor {50, 100} *✓* *✗* ###### Examples of image manipulations on page-050.png. -- -- ###### PSNR results of proposed watermarking method. \# 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 ------ ------- ------- ----------- ----------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- JPEG 001 40.74 40.12 **43.06** 45.01 47.61 48.99 50.65 51.20 51.80 52.05 52.19 52.25 52.27 52.28 52.30 003 42.50 41.93 **44.86** 46.75 49.27 50.69 51.84 52.69 53.00 52.99 53.24 53.27 53.27 53.29 53.28 050 42.20 41.61 **44.55** 46.37 48.90 50.26 51.62 52.35 52.71 52.88 52.98 53.01 53.04 53.06 53.06 601 41.97 41.22 **44.15** 46.14 48.68 49.97 51.52 52.12 52.48 52.75 52.89 52.90 52.93 52.94 52.95 PNG 001 38.78 38.05 **42.94** 43.94 48.10 50.04 53.84 56.05 59.64 62.35 66.37 68.80 72.36 74.01 80.24 003 38.02 37.39 41.30 **43.22** 47.20 49.35 52.95 55.29 58.69 61.17 65.52 68.35 70.81 73.75 78.57 050 37.34 36.81 41.30 **42.74** 46.43 48.89 52.32 54.77 58.10 60.86 64.01 67.77 70.21 73.00 79.44 601 37.15 36.55 40.31 **42.64** 46.30 48.58 52.12 54.49 57.70 60.66 64.11 67.71 70.96 74.78 76.58 ###### SSIM results of proposed watermarking method. \# 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 ------ ------- ------- ------- ------- ------- ------- ----------- ----------- ----------- ----------- ------- ------- ------- ------- ------- JPEG 001 0.905 0.900 0.946 0.966 0.985 0.992 0.997 0.998 0.999 **1.000** 1.000 1.000 1.000 1.000 1.000 003 0.925 0.924 0.959 0.974 0.990 0.994 0.998 0.999 **1.000** 1.000 1.000 1.000 1.000 1.000 1.000 050 0.932 0.930 0.962 0.976 0.990 0.995 0.998 0.999 0.999 **1.000** 1.000 1.000 1.000 1.000 1.000 601 0.928 0.925 0.959 0.974 0.989 0.994 0.998 0.999 0.999 **1.000** 1.000 1.000 1.000 1.000 1.000 PNG 001 0.992 0.988 0.995 0.996 0.998 0.999 **1.000** 1.000 1.000 1.000 1.000 1.000 1.000 1.000 1.000 003 0.990 0.987 0.993 0.995 0.998 0.999 0.999 **1.000** 1.000 1.000 1.000 1.000 1.000 1.000 1.000 050 0.989 0.986 0.993 0.995 0.998 0.999 0.999 **1.000** 1.000 1.000 1.000 1.000 1.000 1.000 1.000 601 0.989 0.983 0.992 0.995 0.998 0.999 0.999 **1.000** 1.000 1.000 1.000 1.000 1.000 1.000 1.000 ###### PSNR result before and after pixel manipulation. \#   1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 ------ ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- JPEG 001 Bef. 40.74 40.12 43.06 45.01 47.61 48.99 50.65 51.20 51.80 52.05 52.19 52.25 52.27 52.28 52.30 Aft. 40.98 40.40 43.16 44.98 47.33 48.53 49.89 50.33 50.79 50.97 51.07 51.11 51.13 51.13 51.15 003 Bef. 42.50 41.93 44.86 46.75 49.27 50.69 51.84 52.69 53.00 52.99 53.24 53.27 53.27 53.29 53.28 Aft. 42.67 42.13 44.91 46.64 48.87 50.06 50.95 51.59 51.81 51.81 51.98 52.00 52.00 52.01 52.01 050 Bef. 42.20 41.61 44.55 46.37 48.90 50.26 51.62 52.35 52.71 52.88 52.98 53.01 53.04 53.06 53.06 Aft. 42.42 41.87 44.66 46.38 48.68 49.86 50.97 51.55 51.82 51.95 52.04 52.06 52.08 52.09 52.09 601 Bef. 41.97 41.22 44.15 46.14 48.68 49.97 51.52 52.12 52.48 52.75 52.89 52.90 52.93 52.94 52.95 Aft. 42.19 41.47 44.25 46.09 48.39 49.49 50.75 51.21 51.48 51.68 51.78 51.79 51.81 51.82 51.82 PNG 001 Bef. 38.78 38.05 42.94 43.94 48.10 50.04 53.84 56.05 59.64 62.35 66.37 68.80 72.36 74.01 80.24 Aft. 76.38 75.57 76.50 75.44 78.33 79.67 79.67 79.67 79.67 79.67 79.67 79.67 79.67 79.67 79.67 003 Bef. 38.02 37.39 41.30 43.22 47.20 49.35 52.95 55.29 58.69 61.17 65.52 68.35 70.81 73.75 78.57 Aft. 73.43 76.73 71.22 71.60 72.42 72.42 72.42 72.42 72.42 70.92 72.42 72.42 72.42 72.42 72.42 050 Bef. 37.34 36.81 41.30 42.74 46.43 48.89 52.32 54.77 58.10 60.86 64.01 67.77 70.21 73.00 79.44 Aft. 73.31 74.07 77.13 77.80 78.51 73.47 76.90 76.90 76.90 76.90 76.90 76.90 76.90 76.90 76.90 601 Bef. 37.15 36.55 40.31 42.64 46.30 48.58 52.12 54.49 57.70 60.66 64.11 67.71 70.96 74.78 76.58 Aft. 71.00 70.84 72.44 72.44 74.74 74.69 75.02 75.02 75.02 75.02 75.02 75.02 75.02 75.02 75.02 ###### Bit error rates of proposed method. \# 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 ----- ----- ------ ------ ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- 001 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 003 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 050 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 601 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 001 44 100 53.8 100 100 100 100 100 100 100 100 100 100 100 100 003 100 88.9 100 100 100 100 100 100 100 100 100 100 100 100 100 050 63 100 38.1 100 100 100 100 100 100 100 100 100 100 100 100 601 100 79.5 100 100 100 100 100 100 100 100 100 100 100 100 100 ###### PSNR result on Gaussian blurring. \#   1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 ------ ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- JPEG 001 Bef. 40.74 40.12 43.06 45.01 47.61 48.99 50.65 51.20 51.80 52.05 52.19 52.25 52.27 52.28 52.30 Aft. 43.91 43.49 45.62 46.85 48.24 48.82 49.39 49.54 49.70 49.76 49.80 49.81 49.81 49.82 49.82 003 Bef. 42.50 41.93 44.86 46.75 49.27 50.69 51.84 52.69 53.00 52.99 53.24 53.27 53.27 53.29 53.28 Aft. 44.85 44.49 46.49 47.54 48.64 49.08 49.39 49.58 49.64 49.64 49.68 49.69 49.69 49.69 49.69 050 Bef. 42.20 41.61 44.55 46.37 48.90 50.26 51.62 52.35 52.71 52.88 52.98 53.01 53.04 53.06 53.06 Aft. 44.59 44.21 46.17 47.17 48.23 48.69 49.04 49.20 49.27 49.30 49.32 49.32 49.33 49.33 49.33 601 Bef. 41.97 41.22 44.15 46.14 48.68 49.97 51.52 52.12 52.48 52.75 52.89 52.90 52.93 52.94 52.95 Aft. 44.42 43.92 45.85 46.94 48.02 48.42 48.82 48.94 49.00 49.06 49.08 49.08 49.09 49.09 49.09 PNG 001 Bef. 38.78 38.05 42.94 43.94 48.10 50.04 53.84 56.05 59.64 62.35 66.37 68.80 72.36 74.01 80.24 Aft. 46.33 44.92 49.87 50.55 54.47 55.40 57.63 58.44 59.43 59.74 60.03 60.13 60.18 60.20 60.22 003 Bef. 38.02 37.39 41.30 43.22 47.20 49.35 52.95 55.29 58.69 61.17 65.52 68.35 70.81 73.75 78.57 Aft. 45.39 44.32 48.04 49.64 52.80 53.91 55.47 56.12 56.68 56.90 57.10 57.15 57.17 57.19 57.20 050 Bef. 37.34 36.81 41.30 42.74 46.43 48.89 52.32 54.77 58.10 60.86 64.01 67.77 70.21 73.00 79.44 Aft. 44.18 44.01 47.91 49.11 51.49 52.88 54.08 54.63 55.04 55.22 55.34 55.40 55.42 55.44 55.45 601 Bef. 37.15 36.55 40.31 42.64 46.30 48.58 52.12 54.49 57.70 60.66 64.11 67.71 70.96 74.78 76.58 Aft. 43.94 43.63 46.92 48.77 50.96 52.14 53.25 53.70 54.04 54.21 54.31 54.35 54.37 54.38 54.39 ###### Bit error rates on Gaussian blurring. \# 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 ----- ---- ------ ------ ------ ------ ------ ------ ------ ------ ----- ----- ----- ----- ----- ----- 001 77 88.9 94.3 97 98.6 98.9 99.7 100 99.5 100 100 100 100 100 100 003 76 86.9 93.9 97.2 98.3 99.3 99.7 100 99.5 100 100 100 100 100 100 050 75 89.2 94.7 96.6 98.4 99.1 99.9 100 99.5 100 100 100 100 100 100 601 76 86.4 94.6 96.7 98.5 99.2 99.4 100 99.5 100 100 100 100 100 100 001 97 98.7 56.1 99.4 99.7 89.8 99.8 99.9 100 100 100 100 100 100 100 003 94 96.4 63 98.4 99.3 99 100 100 100 100 100 100 100 100 100 050 93 96.1 69.1 98.3 99 97.3 98.2 99.8 100 100 100 100 100 100 100 601 93 95.3 70.5 98.1 98.6 99.6 99.5 100 99.8 100 100 100 100 100 100 ###### PSNR results on Gaussian noise attack. \#   1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 ------ ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- JPEG 001 Bef. 40.74 40.12 43.06 45.01 47.61 48.99 50.65 51.20 51.80 52.05 52.19 52.25 52.27 52.28 52.30 Aft. 40.80 40.25 42.87 44.53 46.59 47.59 48.64 48.97 49.29 49.42 49.49 49.52 49.53 49.54 49.55 003 Bef. 42.50 41.93 44.86 46.75 49.27 50.69 51.84 52.69 53.00 52.99 53.24 53.27 53.27 53.29 53.28 Aft. 42.07 41.60 43.92 45.24 46.73 47.42 47.88 48.19 48.29 48.28 48.36 48.37 48.37 48.37 48.37 050 Bef. 42.20 41.61 44.55 46.37 48.90 50.26 51.62 52.35 52.71 52.88 52.98 53.01 53.04 53.06 53.06 Aft. 41.83 41.34 43.69 45.01 46.55 47.24 47.81 48.08 48.20 48.25 48.29 48.30 48.31 48.31 48.31 601 Bef. 41.97 41.22 44.15 46.14 48.68 49.97 51.52 52.12 52.48 52.75 52.89 52.90 52.93 52.94 52.95 Aft. 41.70 41.06 43.47 44.96 46.61 47.30 48.01 48.25 48.39 48.48 48.53 48.54 48.55 48.55 48.55 PNG 001 Bef. 38.78 38.05 42.94 43.94 48.10 50.04 53.84 56.05 59.64 62.35 66.37 68.80 72.36 74.01 80.24 Aft. 28.95 28.96 28.92 28.92 28.91 28.91 28.91 28.90 28.90 28.90 28.90 28.90 28.90 28.90 28.90 003 Bef. 38.02 37.39 41.30 43.22 47.20 49.35 52.95 55.29 58.69 61.17 65.52 68.35 70.81 73.75 78.57 Aft. 28.99 29.00 28.97 28.96 28.95 28.95 28.95 28.95 28.95 28.95 28.95 28.95 28.95 28.95 28.95 050 Bef. 37.34 36.81 41.30 42.74 46.43 48.89 52.32 54.77 58.10 60.86 64.01 67.77 70.21 73.00 79.44 Aft. 29.00 29.00 28.97 28.97 28.96 28.96 28.95 28.95 28.95 28.95 28.95 28.95 28.95 28.95 28.95 601 Bef. 37.15 36.55 40.31 42.64 46.30 48.58 52.12 54.49 57.70 60.66 64.11 67.71 70.96 74.78 76.58 Aft. 29.02 29.03 28.99 28.99 28.98 28.98 28.98 28.97 28.97 28.97 28.97 28.97 28.97 28.97 28.97 ###### Bit error rates result on Gaussian noise attack. \# 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 ----- ----- ------ ------ ------ ------ ----- ------ ----- ----- ----- ----- ----- ----- ----- ----- 001 96 96.5 100 100 100 100 100 100 100 100 100 100 100 100 100 003 83 95.6 98.3 99.8 99.8 100 99.8 100 100 100 100 100 100 100 100 050 83 94.3 95.6 97 99.5 100 100 100 100 100 100 100 100 100 100 601 91 94.3 97.5 94.3 100 100 100 100 100 100 100 100 100 100 100 001 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 003 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 050 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 601 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 ###### PSNR results on median filtering attack. \#   1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 ------ ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- JPEG 001 Bef. 40.74 40.12 43.06 45.01 47.61 48.99 50.65 51.20 51.80 52.05 52.19 52.25 52.27 52.28 52.30 Aft. 41.81 41.75 42.11 42.20 42.28 42.29 42.31 42.32 42.32 42.32 42.32 42.32 42.33 42.33 42.33 003 Bef. 42.50 41.93 44.86 46.75 49.27 50.69 51.84 52.69 53.00 52.99 53.24 53.27 53.27 53.29 53.28 Aft. 41.51 41.44 41.87 42.00 42.10 42.13 42.15 42.16 42.16 42.16 42.16 42.16 42.16 42.16 42.16 050 Bef. 42.20 41.61 44.55 46.37 48.90 50.26 51.62 52.35 52.71 52.88 52.98 53.01 53.04 53.06 53.06 Aft. 41.23 41.17 41.57 41.69 41.78 41.80 41.83 41.84 41.84 41.84 41.84 41.84 41.84 41.84 41.84 601 Bef. 41.97 41.22 44.15 46.14 48.68 49.97 51.52 52.12 52.48 52.75 52.89 52.90 52.93 52.94 52.95 Aft. 40.93 40.83 41.22 41.32 41.39 41.41 41.44 41.44 41.44 41.45 41.45 41.45 41.45 41.45 41.45 PNG 001 Bef. 38.78 38.05 42.94 43.94 48.10 50.04 53.84 56.05 59.64 62.35 66.37 68.80 72.36 74.01 80.24 Aft. 38.66 37.95 42.64 43.57 47.19 48.69 51.13 52.15 53.27 53.78 54.15 54.27 54.35 54.37 54.41 003 Bef. 38.02 37.39 41.30 43.22 47.20 49.35 52.95 55.29 58.69 61.17 65.52 68.35 70.81 73.75 78.57 Aft. 37.83 37.23 40.90 42.60 45.79 47.22 49.07 49.88 50.62 50.91 51.18 51.26 51.29 51.32 51.34 050 Bef. 37.34 36.81 41.30 42.74 46.43 48.89 52.32 54.77 58.10 60.86 64.01 67.77 70.21 73.00 79.44 Aft. 37.09 36.58 40.71 41.92 44.72 46.21 47.73 48.45 49.01 49.28 49.44 49.53 49.56 49.57 49.59 601 Bef. 37.15 36.55 40.31 42.64 46.30 48.58 52.12 54.49 57.70 60.66 64.11 67.71 70.96 74.78 76.58 Aft. 36.84 36.25 39.69 41.65 44.28 45.56 46.99 47.61 48.09 48.33 48.47 48.54 48.57 48.58 48.59 ###### Bit error rates result on median filtering attack. \# 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 ----- ---- ------ ------ ------ ------ ------ ------ ------ ------ ----- ----- ----- ----- ----- ----- 001 79 87.4 92.2 95.3 98 99.2 99.7 99.7 98.9 100 100 100 100 100 100 003 80 86.4 86.5 92.7 97.7 98.7 99.7 99.7 100 100 100 100 100 100 100 050 79 82.1 93.8 95.1 98.5 98.9 99.7 99.7 100 100 100 100 100 100 100 601 79 87.4 92.5 96.1 98.3 99.1 99.2 100 99.5 100 100 100 100 100 100 001 97 98.6 71.5 99.6 99.7 99.8 99.9 100 100 100 100 100 100 100 100 003 95 96.4 85.1 98.6 99.3 99.4 99.6 100 100 100 100 100 100 100 100 050 94 96.3 73.5 98.5 99.2 99.7 98.1 99.8 99.8 100 100 100 100 100 100 601 93 95.5 93.9 98.6 98.9 99.6 99.7 99.8 99.8 100 100 100 100 100 100 [^1]: Academic Editors: J.-H. Park and Y. Pan

1. Introduction =============== The Third International Conference on Information and Knowledge Management (CIKM-94) was held at NIST on November 29--December 1, 1994. This conference is sponsored by Association for Computing Machinery (ACM) Special Interest Group on Artificial Intelligence (SIGART) and Special Interest Group on Information Retrieval (SIGIR), with cooperation from American Association of Artificial Intelligence (AAAI), Center for Applied Information Technology at NIST, National Science Foundation, and the University of Maryland Baltimore Campus. The purpose of the conference was to identify challenging problems facing the development of future knowledge and information systems, and to shape future directions of research by soliciting and reviewing high quality, applied and theoretical research findings. An important part of the conference was the workshops program which focused on timely research challenges and initiatives. In conjunction with the conference, four separate workshops were held at NIST on December 2, 1994. A separate exhibit program was also held at NIST during the conference hours. The conference co-chairs were Bharat Bhargava, Purdue University and Yelena Yesha, NIST/UMBC. The program chair was Nabil Adam of Rutgers University. David Jefferson and Shirley Hurwitz of NIST served on the steering committee and program vice chair respectively, and Elizabeth Fong of NIST served as the local arrangement chair. The conference consisted of six keynote addresses of which one was a banquet speaker, seven invited talks, four panels, and 55 papers. These are summarized as follows: 2. Keynote Addresses ==================== Arati Prabhakar, Director of NIST opened the conference with a warm welcome and presented the National Information Infrastructure program. Jacob Slonim, Head of Research, Centre for Advanced Studies, IBM Canada, presented a talk on digital libraries---is it reality or myth? Milton Halem, Chief, Space Data and Computing Division, NASA, spoke on the paradigm shift from numerically intensive to data intensive computing. Al Aho, Vice President, Bellcore, presented a keynote address on issues in large knowledge repositories. Robert Ewald, Chief Operating Officer of Cray Research, gave a keynote on computing into the 21st century, and Jeffrey Ullman, Professor and Chair, Stanford University, spoke on approaches to high-level interoperability. The banquet speaker was Thomas Kalil, Director of the National Economic Council, The White House. Mr. Kalil addressed the dinner audience on the Clinton-Gore Administration and the National Information Infrastructure. 3. Panels ========= This conference featured four panel sessions. The panel, "Making money on the Internet," was coordinated by Charles Nicholas, UMBC. The panel speakers were Andrew Anker, Wired Magazine, and Harold Stone, NEC Research Laboratory. Steve Ray of NIST coordinated a panel on "Manufacturing." The invited panelists were Jeff Sutherland, Easel Corporation; Pradeep Khosla, ARPA; Dana Nau, University of Maryland; Leo Obrst, Boeing Helicopters; and Ram Sriram, NIST. The focus of the panel was to identify issues associated with computer integrated manufacturing and the information management requirements that are necessary to support manufacturing problems. Milton Halem of NASA coordinated a panel on "Federal Program Initiatives in Computer and Information Sciences." The panelists represented ARPA, NASA, NSF, DOE, and NIST. Each panelist described their respective technology solicitation program in computer and information sciences. The panel provided useful information to academia in relevant research areas. Elizabeth Fong of NIST and Abdelsalam Helal of the University of Texas at Arlington coordinated a panel on "Standards---Interoperability." Panelists were invited to speak on the following selected emerging standards: Knowledge Query Manipulation Language (KQML) by Dan McKay; Object Database Management Standard by Doug Barry; SQL-3 by Andrew Eisenberg; and STEP/EXPRESS language by David Price. The final panelist, Son Dao of Hughes Corporation, presented a reference model architecture for the emerging standards where interoperation and integration might occur. 4. Technical Papers =================== Parallel technical paper sessions were held in the following topics: object-oriented conceptual modeling, query processing, knowledge bases, text databases information retrieval, transaction processing, intelligent agents, logical/deductive databases, schema integration, temporal databases storage systems, schema evolution, spatial databases, mobile computing, knowledge discovery, data mining, medical applications, and indexing/caching. 5. Exhibits =========== Mary Brady of NIST coordinated the conference exhibit programs. The following information technology demonstration systems were presented: **ALIBI** (Adaptive Location of Internetworked Bases of Information) is a resource discovery and information retrieval tool. It allows users to quickly and easily retrieve data using a query-based interface. **FAST**---An Internet based electronic commerce system is developed by Information Sciences Institute of the University of Southern California (ISI) under the sponsorship of ARPA. FAST parses the buyers' request for quotes, routes them to appropriate online sellers, and return the sellers' quotes to the buyers. **Fazzt** is a digital delivery service developed by KenCast, Inc. It demonstrates a satellite broadcast service for the delivery of large digital objects (e.g., electronic catalogs and documents). **Remote Database Access (RDA)** is a communication protocol for remote database access that has been adopted as an ISO, ANSI, and FIPS standard. The RDA demonstration shows a database client receiving a complex query statement in SQL from a user and accessing multiple, heterogeneous database servers that are geographically dispersed. This demonstration is conducted by the Information Systems Engineering Division of NIST. **Secure Electronic Mail** is a prototype secure e-mail which incorporates confidentiality and digital signature services via public key and secret key cryptography using a standards based infrastructure. This demonstration is conducted by the NIST's Computer Security Division and the Systems and Network Architecture Division. **The Smart Procurement System** is an innovative application of two evolving computer technologies: the World Wide Web (WWW) and Intelligent Agent (IA). The system enables a purchaser to execute the government's procurement process electronically, obtaining competitive bids from participating vendors in hours instead of weeks. The initial prototype being demonstrated is conducted by the Center for Applied Information Technology (CAIT) of NIST, under a Cooperative Research and Development Agreement (CRADA) with Enterprise Integration Technologies (EIT). 6. Workshops ============ Four separate workshops were conducted in association with CIKM on December 2, 1994, at NIST. The *Intelligent Agent* workshop, chaired by Tim Finin of the University of Maryland Baltimore Campus, drew a crowd of over 60 researchers. Twelve papers and a panel session were presented. The workshop on *Electronic Commerce*, co-chaired by Nabil Adam, Rutgers University and Yelena Yesha, NIST, featured a keynote talk by Robert Neches, ARPA, followed by the presentation of ten papers and a lively panel discussion on examples of electronic commerce services on the Internet. The workshop on *Advances in Geographic Information Systems (GIS)* was organized by Niki Pissinou, SW Louisiana University. Many papers were presented along with panel sessions and group discussions on topics involving GIS research and the applicability of computer science. The *Intelligent Hypertext* workshop was the continuation of last years' workshop on the same topic, organized by Charles Nicholas of UMBC and NIST. This workshop focused on presentation of a few selected position papers. The workshop broke into small working groups and concluded with final reports and a wrap-up discussion. 7. Conclusions ============== This conference succeeded in fostering direct interactions between industry, government, and academic researchers in identifying challenging problems facing the development of future knowledge and information systems. All of the refereed papers for the CIKM conference appeared in a proceedings which is available from the ACM Headquarters. Information on CIKM-94 is available on-line via ftp, Gopher, and World-Wide-Web. Access information is: ftp: from [ftp.cs.umbc.edu.in.pub/cikm](ftp.cs.umbc.edu.in.pub/cikm)gopher: from [gopher.cs.umbc.edu.in.conferences/cikm](http://gopher.cs.umbc.edu.in.conferences/cikm)www: from <http://www.cs.umbc.edu/cikm>

\section{From continuous dynamics to multivalued mappings} \label{sec:cont-graph} In this section we recall how the dynamics of multivalued mappings can be linked to dynamical systems as described in \cite{KMV0}. Since multivalued mappings are discrete in both time and space, we need to address the issues of both time and space discretization. \subsection{Grids and outer approximations} \label{sec:cgt} To represent continuous dynamics in terms of the combinatorial structures described in Section~\ref{sec:comb-syst} requires discretizing phase space. We wish this discretization to be as generally applicable and as topologically nice as possible. With this in mind we choose the basic elements of our discretization to be {\em regular closed sets}, i.e. sets $A\subset X$ such that $A ={\rm cl}\,}%{\text{cl}(\mbox{\rm int\,}(A))$. \begin{proposition} \label{prop:walker} \cite[Proposition 2.3]{Walker} Let $X$ be a topological space. The family ${\mathscr R}(X)$ of regular closed subsets of $X$ is a Boolean algebra with the following operations: \begin{enumerate} \item[(i)] $A\leq B$ if and only if $A\subset B$; \item[(ii)] $A\vee B := A\cup B$; \item[(iii)] $A\wedge B := {\rm cl}\,}%{\text{cl} ( \mbox{\rm int\,}(A\cap B))$; \item[(iv)] $A^{\bigstar} := {\rm cl}\,}%{\text{cl}(A^c)$; \end{enumerate} where $0=\varnothing$ and $1=X$. \end{proposition} \begin{remark} Proposition 2.3 in \cite{Walker} proves that ${\mathscr R}(X)$ is a complete Boolean algebra, i.e. $\bigvee_\alpha A_\alpha := {\rm cl}\,}%{\text{cl}\left( \bigcup_\alpha \mbox{\rm int\,}(A_\alpha) \right)$ and $\bigwedge_\alpha A_\alpha := {\rm cl}\,}%{\text{cl}\left( \mbox{\rm int\,}(\bigcap_\alpha A_\alpha) \right)$ are well-defined. \end{remark} \begin{lemma} \label{lem:equivregcl} Let $A,A' \in {\mathscr R}(X)$, then $A\wedge A' = \varnothing$ if and only if $A\cap \mbox{\rm int\,} (A') = \varnothing$. \end{lemma} \begin{proof} By definition \[ A \wedge A' = {\rm cl}\,}%{\text{cl}(\mbox{\rm int\,}(A\cap A')). \] Using the property $\mbox{\rm int\,}(A\cap A') = \mbox{\rm int\,}(A)\cap \mbox{\rm int\,}(A')$, \[ A \wedge A' = {\rm cl}\,}%{\text{cl}(\mbox{\rm int\,}(A)\cap \mbox{\rm int\,}(A')). \] Also, if $U\subset X$ is open and $B,B'\subset X$ with ${\rm cl}\,}%{\text{cl}(B) = {\rm cl}\,}%{\text{cl}(B')$, then ${\rm cl}\,}%{\text{cl}(B\cap U) = {\rm cl}\,}%{\text{cl}(B'\cap U)$. Taking $U= \mbox{\rm int\,}(A')$, $B = \mbox{\rm int\,}(A),$ and $B'= A$ implies \[ A \wedge A' = {\rm cl}\,}%{\text{cl}(A\cap \mbox{\rm int\,}(A')). \] Therefore \[ A \wedge A'=\varnothing \quad\hbox{iff}\quad {\rm cl}\,}%{\text{cl}(A\cap \mbox{\rm int\,}(A'))=\varnothing \quad\hbox{iff}\quad A\cap \mbox{\rm int\,}(A')=\varnothing, \] which proves the equivalence. \end{proof} Sets $A,A'\subset X$ for which $A\wedge A' = \varnothing$ will be referred to as \emph{regularly disjoint sets}. \begin{lemma} \label{lem:intregdisj} Let $A,B,C \in {\mathscr R}(X)$ be mutually regularly disjoint sets. Then \[ A = {\rm cl}\,}%{\text{cl}\bigl( (A\cup B)\setminus (B\cup C)\bigr). \] \end{lemma} \begin{proof} We start with the observation that if $A,A'\in {\mathscr R}(X)$ are mutually regularly disjoint, then ${\rm cl}\,}%{\text{cl}(A\setminus A') = A$. Indeed, by Lemma \ref{lem:equivregcl}, $A\wedge A'=\varnothing$, is equivalent to $A\cap \mbox{\rm int\,}(A') =\mbox{\rm int\,}(A) \cap A' = \varnothing$. This implies $\mbox{\rm int\,}(A) \subset A\setminus A' \subset A$ and therefore $A= {\rm cl}\,}%{\text{cl}(\mbox{\rm int\,}(A)) \subset {\rm cl}\,}%{\text{cl}(A\setminus A') \subset {\rm cl}\,}%{\text{cl}(A) = A$, which proves the statement. Note that $(A\cup B)\setminus (B\cup C) = (A\cup B) \setminus B\setminus C = A\setminus B\setminus C = A\setminus (B\cup C)$. By assumption $A\wedge (B\cup C) = \varnothing$. By the previous statement we then have \[ A = {\rm cl}\,}%{\text{cl} (A\wedge (B\cup C)) = {\rm cl}\,}%{\text{cl}\bigl( (A\cup B)\setminus (B\cup C)\bigr), \] which proves the lemma. \end{proof} For the purpose of computation we are only interested in finite collections of regular closed sets. \begin{proposition} \label{prop:grid-sub} Let ${\mathscr R}_0\subset {\mathscr R}(X)$ be a finite subalgebra of the Boolean algebra ${\mathscr R}(X)$ of regular closed subsets of $X$ and let ${\mathsf J}({\mathscr R}_0)$ denote the set of atoms of ${\mathscr R}_0$. Then \begin{enumerate} \item[(i)] $X=\bigcup\setof{A\mid A \in {\mathsf J}({\mathscr R}_0)}$. \item[(ii)] If $A,A'\subset X$ are atoms of ${\mathscr R}_0$, then $ A\cap \mbox{\rm int\,}(A') = \varnothing. $ \end{enumerate} Conversely, every finite set ${\mathsf J}= \{A\mid A\subset X\}$ of mutually regularly disjoint subsets which satisfies (i) generates a subalgebra of ${\mathscr R}(X)$ for which ${\mathsf J}$ is the set of atoms. \end{proposition} \begin{proof} The proof of (i) follows from the fact that $1=X$ and Proposition~\ref{prop:walker}(ii). Property (ii) follows from Lemma \ref{lem:equivregcl}. The converse statement follows from Stone's Representation Theorem, cf.\ \cite{Davey:2002p5923}. \end{proof} Proposition~\ref{prop:grid-sub} implies that if $X$ is a compact metric space, then any finite subalgebra of ${\mathscr R}(X)$ defines a {\em grid} on $X$ (see \cite{KMV0}, \cite{Mrozek-grid}), and conversely a grid defines a finite subalgebra of ${\mathscr R}(X)$. We denote the space of a grids on $X$ by ${\mathsf{Grid}}(X)$, which is a lattice dual to the lattice of finite subalgebras ${{\rm sub}_F\mathscr{R}}(X)$. Since we make use of grids to pass from the computations to dynamics, we recall and establish several fundamental properties. First, by \cite[Theorem 2.2]{KMV0} given a compact metric space there exists a grid with elements of arbitrarily small diameter. Second, as is discussed in detail in this section, grids provide a natural correspondence between the combinatorial systems of Section~\ref{sec:comb-syst} and the continuous systems of interest. To begin to set up the relationship between combinatorial and continuous systems, consider a grid on $X$ indexed by a finite set ${\mathcal X}$. In particular, given $\xi\in{\mathcal X}$ the corresponding grid element is denoted by $\supp{\xi}\in {\mathscr R}(X)$. The \emph{evaluation mapping} $\supp{\cdot}\colon {\mathsf{Set}}({\mathcal X}) \to {\mathscr R}(X)$ is defined by \[ |{\mathcal U}| := \bigcup_{\xi\in {\mathcal U}} \supp{\xi}. \] The range of ${\mathsf{Set}}({\mathcal X})$ under $|\cdot|$ is the subalgebra whose atoms are the grid elements, and this subalgebra will be denoted by ${\mathscr R}_{{{\mathcal X}}}(X)$. Proposition~\ref{prop:grid-sub} immediately implies the following. \begin{corollary} \label{cor:grid->lattice}\label{prop:grid->lattice} Given a grid on a compact metric space $X$ indexed by ${\mathcal X}$, then the evaluation mapping $\supp{\cdot}\colon {\mathsf{Set}}({\mathcal X}) \to {\mathscr R}(X)$ is a Boolean isomorphism onto ${\mathscr R}_{{{\mathcal X}}}(X)$. \end{corollary} \begin{proof} By construction the evaluation map is a lattice homomorphism. Since $|\varnothing| = \varnothing$ and $|{\mathcal X}| = X$, Lemma 4.17 in \cite{Davey:2002p5923} shows that the evaluation map is Boolean. \end{proof} Before explicitly describing the discretization of a general dynamical system $\varphi\colon {\mathbb{T}}^+ \times X\to X$, we consider the simple setting of approximating the dynamics generated by a continuous map $f\colon X\to X$. \begin{definition} \label{defn:outerapprox} Let $f\colon X \to X$ be a continuous map. Let ${\mathcal X}$ be the indexing set for a grid on $X$. A multivalued mapping ${\mathcal F}\colon{\mathcal X} \raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$} {\mathcal X}$ is an \emph{outer approximation} of $f$ if \begin{equation} \label{eq:outerapprox} f(\supp{\xi}) \subset \mbox{\rm int\,} |{\mathcal F}(\xi)| \text{ for all }\xi\in {\mathcal X}. \end{equation} A multivalued mapping ${\mathcal F}\colon{\mathcal X} \raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$} {\mathcal X}$ is a \emph{weak outer approximation} for $f$ if \begin{equation} \label{eq:weakouterapprox} f(\supp{\xi}) \subset \mbox{\rm int\,} \supp{\bigcup_{n\geq 0}{\mathcal F}^n(\xi)}\text{ for all }\xi\in {\mathcal X}. \end{equation} \end{definition} \begin{remark} By definition outer approximations ${\mathcal F}$ are necessarily left-total, and therefore combinatorial omega limit sets and attractors are invariant sets for outer approximations ${\mathcal F}$. \end{remark} \subsection{Attractors, repellers, and their neighborhoods} \label{subsec:at-rep-neigh} Recall that a set $U\subset X$ is an \emph{attracting neighborhood} for a continuous function $f\colon X \to X$ if $\omega(U,f)\subset \mbox{\rm int\,}(U)$. A \emph{trapping region} $U$ is an attracting neighborhood with the additional property that $f({\rm cl}\,}%{\text{cl}(U)) \subset \mbox{\rm int\,}(U)$. A set $A\subset X$ is an \emph{attractor} if there exists a trapping region $U$ such that $A=\Inv (U,f)$ in which case $A = \omega(U,f)\subset \mbox{\rm int\,}(U)$. A set $U\subset X$ is a \emph{repelling neighborhood} for a continuous function $f\colon X \to X$ if $\alpha(U,f)\subset \mbox{\rm int\,}(U)$. A \emph{repelling region} $U$ is an repelling neighborhood with the additional property that $f^{-1}({\rm cl}\,}%{\text{cl}(U)) \subset \mbox{\rm int\,}(U)$. A set $R\subset X$ is an \emph{repeller} if there exists a repelling region $U$ such that $R=\Inv^+ (U,f)$ in which case $R = \alpha(U,f)\subset \mbox{\rm int\,}(U)$ cf.\ \cite{KMV-1a}. The sets of all attracting neighborhoods and repelling neighborhoods are denoted by ${\mathsf{ ANbhd}}(X,f)$ and ${\mathsf{ RNbhd}}(X,f)$ respectively. As is shown in \cite{KMV-1a} these sets are lattices under the operations union and intersection. The following propositions indicate that attracting and repelling neighborhoods can be identified using weak outer approximations. \begin{proposition} \label{prop:discr-att-to-varphi1} Let ${\mathcal F}\colon {\mathcal X} \raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$} {\mathcal X}$ be an weak outer approximation for $f$. If ${\mathcal U}\subset {\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})$, then $|{\mathcal U}|$ is a trapping region for $f$, and therefore $|{\mathcal U}|\in{\mathsf{ ANbhd}}(X,f)$. \end{proposition} \begin{proof} Since ${\mathcal F}$ is a weak outer approximation, for $\xi\in{\mathcal U}$ we have \[ f(|\xi|) \subset \mbox{\rm int\,}\left|\bigcup_{n\ge0}{\mathcal F}^n(\xi)\right| \subset \mbox{\rm int\,}|{\mathcal U}| \] because ${\mathcal F}^n({\mathcal U})\subset {\mathcal U}$ for all $n\ge0$. Therefore $ f(|{\mathcal U}|) \subset \mbox{\rm int\,}|{\mathcal U}|, $ which implies that $|{\mathcal U}|$ is a trapping region for $f$. \end{proof} \begin{proposition} \label{prop:prop:discr-att-to-varphi2} Let ${\mathcal F}\colon{\mathcal X} \raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$} {\mathcal X}$ be an weak outer approximation for $f$. If ${\mathcal U}\in{\mathsf{ Invset}}^-({\mathcal X},{\mathcal F})$, then $|{\mathcal U}|\in {\mathsf{ RNbhd}}(X,f)$. \end{proposition} \begin{proof} Let ${\mathcal U}\in{\mathsf{ Invset}}^-({\mathcal X},{\mathcal F})$. By Proposition~\ref{prop:cms-forback}, ${\mathcal U}^c\in{\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})$, and thus by Proposition~\ref{prop:discr-att-to-varphi1}, $\supp{{\mathcal U}^c}\in {\mathsf{ ANbhd}}(X,f)$. By \cite[Corollary 3.24]{KMV-1a} $\supp{{\mathcal U}^c}^c \in {\mathsf{ RNbhd}}(X,f)$ and thus by \cite[Corollary 3.26]{KMV-1a} ${\rm cl}\,}%{\text{cl}\left(\supp{{\mathcal U}^c}^c\right) \in {\mathsf{ RNbhd}}(X,f)$. Finally, by Corollary~\ref{prop:grid->lattice}, $|{\mathcal U}|\in {\mathsf{ RNbhd}}(X,f)$. \end{proof} \begin{remark} \label{rmk:other1} Another approach is to achieve the latter directly. In that case a negative time variation on (\ref{eq:weakouterapprox}) is needed. The duality approach used here does not require additional assumptions and is therefore preferable. However, the duality approach implies that $|{\mathcal U}|$ is a repelling neighborhood, and it is \emph{not} clear whether $|{\mathcal U}|$ is a repelling \emph{region}. \end{remark} As the following proposition indicates, outer approximations, as opposed to weak outer approximations, allow one to obtain the same results using a larger variety of sets. The proof makes use of the following observation. If ${\mathcal F}$ is an outer approximation for $f$, then \begin{equation} \label{eqn:iterate} f^n(|\xi|) \subset \mbox{\rm int\,} |{\mathcal F}^n(\xi)|\quad \forall \xi\in {\mathcal X},\; \forall n\ge 0. \end{equation} This property can be derived as follows. Observe that \[ f^2(|\xi|) = f(f(|\xi|))\subset f(\mbox{\rm int\,} |{\mathcal F}(\xi)|) \subset f(|{\mathcal F}(\xi)|), \] and by definition \[ f(|{\mathcal F}(\xi)|) = \bigcup_{\xi' \in{\mathcal F}(\xi)} f(|\xi'|) \subset \bigcup_{\xi' \in{\mathcal F}(\xi)} \mbox{\rm int\,} \supp{{\mathcal F}(\xi')} \subset \mbox{\rm int\,} \supp{{\mathcal F}^2(\xi)}. \] Equation \eqref{eqn:iterate} follows by proceeding inductively. \begin{proposition} \label{prop:discr-att-to-varphi2} Let ${\mathcal F}\colon{\mathcal G} \raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$} {\mathcal G}$ be an outer approximation for $f$. If ${\mathcal U}\in {\mathsf{ASet}}({\mathcal X},{\mathcal F})$, then $|{\mathcal U}| \in {\mathsf{ ANbhd}}(X,f)$. If ${\mathcal U}\in {\mathsf{RSet}}({\mathcal X},{\mathcal F})$, then $|{\mathcal U}| \in {\mathsf{ RNbhd}}(X,f)$. \end{proposition} \begin{proof} By Proposition~\ref{prop:char-attset} if ${\mathcal U}\in {\mathsf{ASet}}({\mathcal X},{\mathcal F})$, then there exists an $n\ge 1$ such that ${\mathcal F}^k({\mathcal U}) \subset {\mathcal U}$ for all $k\ge n$. By \eqref{eqn:iterate} $f^k(\xi) \subset \mbox{\rm int\,} |{\mathcal F}^k(\xi)|$ for all $k\ge n$ and all $\xi\in {\mathcal U}$. Therefore, \begin{eqnarray*} f^k(\supp{{\mathcal U}}) &=& \bigcup_{\xi\in {\mathcal U}} f^k(\supp{\xi}) \subset \bigcup_{\xi\in {\mathcal U}} \mbox{\rm int\,} \bigl(\supp{{\mathcal F}^k(\xi)}\bigr)\\ &\subset& \mbox{\rm int\,} \left[ \bigcup_{\xi\in {\mathcal U}}\supp{{\mathcal F}^k(\xi)} \right] = \mbox{\rm int\,} \supp{ {\mathcal F}^k({\mathcal U})} \subset \mbox{\rm int\,} |{\mathcal U}|\;\;\forall k\ge n, \end{eqnarray*} and hence $\supp{{\mathcal U}}\in {\mathsf{ ANbhd}}(X,f)$. To prove the second part, let ${\mathcal U}\in {\mathsf{RSet}}({\mathcal X},{\mathcal F})$. By Proposition~\ref{prop:invo} ${\mathcal U}^c\in {\mathsf{ASet}}({\mathcal X},{\mathcal F})$ and thus $|{\mathcal U}^c| \in {\mathsf{ ANbhd}}(X,f)$. By Corollary~\ref{prop:grid->lattice} \[ |{\mathcal U}|^{\bigstar} = \supp{{\mathcal U}^c}, \] and thus by \cite[Corollary 3.26]{KMV-1a} $|{\mathcal U}| \in {\mathsf{ RNbhd}}(X,f)$. \end{proof} Let ${\mathsf{ ANbhd}}_{\scrR}(X,f)$ and ${\mathsf{ RNbhd}}_{\scrR}(X,f)$ denote the sets of regular closed attracting and repelling neighborhoods, respectively. \begin{proposition} \label{prop:ANbhdR} Given a continuous function $f\colon X\to X$ on a compact metric space, ${\mathsf{ ANbhd}}_{\scrR}(X,f)$ and ${\mathsf{ RNbhd}}_{\scrR}(X,f)$ are sublattices of ${\mathscr R}(X)$. \end{proposition} \begin{proof} Since the elements of ${\mathsf{ ANbhd}}_{\scrR}(X,f)$ are regular closed sets, then ${\mathsf{ ANbhd}}_{\scrR}(X,f)\subset {\mathscr R}(X)$. Thus it only needs to be shown that ${\mathsf{ ANbhd}}_{\scrR}(X,f)$ is a bounded lattice. Let $U,U' \in {\mathsf{ ANbhd}}_{\scrR}(X,f)$. Observe that $\varnothing, X\in {\mathsf{ ANbhd}}_{\scrR}(X,f)$. By \cite[Lemma 3.2]{KMV-1a}, $\mbox{\rm int\,}(U) \cap \mbox{\rm int\,}(U') \in {\mathsf{ ANbhd}}_{\scrR}(X,f)$, and therefore \[ U\wedge U' = {\rm cl}\,}%{\text{cl} \bigl( \mbox{\rm int\,}(U) \cap \mbox{\rm int\,}(U') \bigr) \in {\mathsf{ ANbhd}}(X,f), \] which proves that ${\mathsf{ ANbhd}}_{\scrR}(X,f)$ is closed under the operations $\vee$ and $\wedge$ of ${\mathscr R}(X)$. The same argument applies to ${\mathsf{ RNbhd}}_{\scrR}(X,f)$. \end{proof} \begin{remark}\label{rmk:eval} As indicated in \cite{KMV-1a} ${\mathsf{ ANbhd}}(X,f)$ is a sublattice of ${\mathsf{Set}}(X)$ and by Proposition~\ref{prop:ANbhdR} ${\mathsf{ ANbhd}}_{\scrR}(X,f)$ is a sublattice of ${\mathscr R}(X)$. Since the operations in these two lattices are different ${\mathsf{ ANbhd}}(X,f)$ and ${\mathsf{ ANbhd}}_{\scrR}(X,f)$ are not interchangeable. The same comment applies to ${\mathsf{ RNbhd}}(X,f)$ and ${\mathsf{ RNbhd}}_{\scrR}(X,f)$. \end{remark} \begin{theorem} \label{thm:first-evalf} Let ${\mathcal X}$ be an indexing set for a grid on $X$ and let ${\mathcal F}\colon{\mathcal X}\raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$} {\mathcal X}$ be a weak outer approximation of $f$. Then, \begin{equation}\label{diag:AR2} \begin{diagram} \node{{\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})} \arrow{e,l,<>}{^c} \arrow{s,l,V}{ |\cdot|}\node{{\mathsf{ Invset}}^-({\mathcal X},{\mathcal F})} \arrow{s,r,V}{ |\cdot|}\\ \node{{\mathsf{ ANbhd}}_{\scrR}(X,f)} \arrow{e,l,<>}{^{\bigstar}} \arrow{s,l,A}{\displaystyle \omega}\node{{\mathsf{ RNbhd}}_{\scrR}(X,f)} \arrow{s,r,A}{\displaystyle \alpha}\\ \node{{\mathsf{ Att}}(X,f)} \arrow{e,l,<>}{^*} \node{{\mathsf{ Rep}}(X,f)} \end{diagram} \end{equation} is a commuting diagram of distributive lattices. The same statement holds if we replace ${\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})$ and ${\mathsf{ Invset}}^-({\mathcal X},{\mathcal F})$ by ${\mathsf{ASet}}({\mathcal X},{\mathcal F})$ and ${\mathsf{RSet}}({\mathcal X},{\mathcal F})$ respectively. \end{theorem} The proof of this result makes use of the following lemma. \begin{lemma} \label{lem:omega-int} Let $U\in {\mathsf{ ANbhd}}(X,f)$, then $\omega(U) = \omega({\rm cl}\,}%{\text{cl}(U)) = \omega(\mbox{\rm int\,}(U)) \subset \mbox{\rm int\,}(U)$. \end{lemma} \begin{proof} If $U$ is attracting, then $\mbox{\rm int\,}(U)$ is attracting. This implies $\omega(\mbox{\rm int\,}(U)) = \Inv(\mbox{\rm int\,}(U))$, see \cite[Corollary 3.6]{KMV-1a}. Moreover, $\omega(\mbox{\rm int\,}(U)) \subset \omega(U) = \Inv(U,f) \subset \mbox{\rm int\,}(U)$, which implies that $\omega(\mbox{\rm int\,}(U)) = \omega(U)$. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:first-evalf}] The proof of the upper square follows from Propositions~\ref{prop:discr-att-to-varphi2} and \ref{prop:ANbhdR} and the relation $|{\mathcal U}|^{\bigstar} = |{\mathcal U}^c|$. The first step in the proof of the lower square is to show that $\omega\colon {\mathsf{ ANbhd}}_{\scrR}(X,f) \to {\mathsf{ Att}}(X,f)$, and $\alpha\colon {\mathsf{ RNbhd}}_{\scrR}(X,f) \to {\mathsf{ Rep}}(X,f)$ are lattice homomorphisms. For $\vee = \cup$ the homomorphism property is obvious. As for $\wedge$, applying Lemma~\ref{lem:omega-int} and \cite[Proposition 4.1]{KMV-1a} results in \[ \omega(U\wedge U') = \omega\bigl({\rm cl}\,}%{\text{cl}(\mbox{\rm int\,}(U\cap U'))\bigr) = \omega(\mbox{\rm int\,}(U\cap U')) =\omega(U\cap U') = A\wedge A'. \] The same argument applies to repelling neighborhoods. The surjectivity of $\alpha$ and $\omega$ follows from \cite[Proposition 5.5]{KMV0} \end{proof} Strengthening the assumptions on the outer approximation ${\mathcal F}$ allows one to extend Theorem~\ref{thm:first-evalf}. \begin{theorem} \label{thm:commutative-diagram} Let ${\mathcal X}$ be an indexing set for a grid on $X$ and let ${\mathcal F}\colon{\mathcal X}\raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$} {\mathcal X}$ be an outer approximation of $f$. Then \begin{equation} \label{diag:comm-liftfull} \begin{diagram} \node{}\node{{\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})} \arrow{e,l,<>}{^c} \arrow{s,l,V}{\imath} \arrow{sw,l,A}{{\bm{\omega}}}\node{{\mathsf{ Invset}}^-({\mathcal X},{\mathcal F})} \arrow{s,r,V}{\imath}\arrow{se,l,A}{{\bm{\alpha}}} \\ \node{{\mathsf{ Att}}({\mathcal X},{\mathcal F})}\arrow{sse,r}{\omega(|\cdot|)}\node{{\mathsf{ASet}}({\mathcal X},{\mathcal F})}\arrow{s,l,V}{|\cdot|}\arrow{w,l,A}{{\bm{\omega}}}\arrow{e,l,<>}{^c} \node{{\mathsf{RSet}}({\mathcal X},{\mathcal F})}\arrow{s,l,V}{|\cdot|}\arrow{e,l,A}{{\bm{\alpha}}}\node{{\mathsf{ Rep}}({\mathcal X},{\mathcal F})} \arrow{ssw,r}{\alpha(|\cdot|)} \\ \node{}\node{{\mathsf{ ANbhd}}_{\scrR}(X,f)}\arrow{s,l,A}{\omega}\arrow{e,l,<>}{^{\bigstar}} \node{{\mathsf{ RNbhd}}_{\scrR}(X,f)}\arrow{s,l,A}{\alpha} \\ \node{}\node{{\mathsf{ Att}}(X,f)} \arrow{e,l,<>}{^*} \node{{\mathsf{ Rep}}(X,f)} \end{diagram} \end{equation} is a commutative diagram of distributive lattices where $\imath$ denotes inclusion. \end{theorem} The proof of Theorem~\ref{thm:commutative-diagram} follows directly from Theorem~\ref{thm:first-evalf}, the following two lemmas, and the use of duality between attractors and repellers to obtain the same lemmas for repellers. \begin{lemma} \label{lem:commute1}\label{prop:commute1} Let ${\mathcal F}\colon{\mathcal X}\raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$} {\mathcal X}$ be an outer approximation for a continuous mapping $f\colon X \to X$, and let ${\mathcal U}\in {\mathsf{ASet}}({\mathcal X},{\mathcal F})$. Then $\omega(|{\mathcal U}|) = \omega(|{{\bm{\omega}}}({\mathcal U})|)$. \end{lemma} \begin{proof} From Proposition \ref{prop:combomega}(iv) we have that ${\bm{\omega}}({\mathcal U}) = {\mathcal F}^k({\mathcal U})$, for some $k$ large enough. Set $S = \omega(|{\mathcal U}|,f)$. Then $f(S) = S$, and \[ S = f^k(S) \subset f^k(|{\mathcal U}|) \subset |{\mathcal F}^k({\mathcal U})| = |{\bm{\omega}}({\mathcal U})|, \] which proves the lemma. \end{proof} \begin{lemma} \label{lem:latt-latt2} The mapping $\omega(|\cdot|)\colon{\mathsf{ Att}}({\mathcal X},{\mathcal F}) \to {\mathsf{ Att}}(X,f)$ is a lattice homomorphism. \end{lemma} \begin{proof} The property for $\vee = \cup$ is obvious. Therefore, we restrict the proof to $\wedge$. Let ${\mathcal A},{\mathcal A}'\in {\mathsf{ Att}}({\mathcal X},{\mathcal F})$. We have that \[ \omega(|{\mathcal A}\wedge {\mathcal A}'|) \subset \omega(|{\mathcal A}\cap {\mathcal A}'|) = \omega(|{\mathcal A}|\wedge|{\mathcal A}'|) = \omega(|{\mathcal A}|) \wedge \omega(|{\mathcal A}'|) =A\wedge A', \] where $A = \omega(|{\mathcal A}|)$ and $A' = \omega(|{\mathcal A}|')$. Conversely, since $A\subset \mbox{\rm int\,}(|{\mathcal A}|)$ and $A' \subset \mbox{\rm int\,}(|{\mathcal A}'|)$, we have \[ A\cap A' \subset \mbox{\rm int\,}(|{\mathcal A}|) \cap \mbox{\rm int\,}(|{\mathcal A}'|) = \mbox{\rm int\,}(|{\mathcal A}|\cap|{\mathcal A}'|) \subset {\rm cl}\,}%{\text{cl}\bigl(\mbox{\rm int\,}(|{\mathcal A}|\cap|{\mathcal A}'|) \bigr) = |{\mathcal A}\cap{\mathcal A}'|, \] and therefore $A\wedge A' \subset \omega(|{\mathcal A}\cap{\mathcal A}'|) = \omega(|{\bm{\omega}}({\mathcal A}\cap{\mathcal A}')|) = \omega(|{\mathcal A}\wedge{\mathcal A}'|)$ by Lemma \ref{lem:commute1}. Combining the inclusions proves the lemma. \end{proof} \begin{remark} \label{rmk:other2} Note that the evaluation map $|\cdot|\colon{\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})\to{\mathsf{ ANbhd}}_{\scrR}(X,f)$ can be restricted to ${\mathsf{ Att}}({\mathcal X},{\mathcal F})$, since every attractor is forward invariant. However, ${\mathsf{ Att}}({\mathcal X},{\mathcal F})$ is not a sublattice of ${\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})$, since the lattice operations are different; ${\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})$ is a lattice under union and intersection, but the $\wedge$ operation for ${\mathsf{ Att}}({\mathcal X},{\mathcal F})$ is ${\mathcal A}\wedge{\mathcal A}'={\bm{\omega}}({\mathcal A}\cap{\mathcal A}')$. In particular $$ |{\mathcal A}\wedge{\mathcal A}'|=|{\bm{\omega}}({\mathcal A}\cap{\mathcal A}')|\subset|{\mathcal A}\cap{\mathcal A}'|=|{\mathcal A}|\wedge|{\mathcal A}'| $$ but $|{\mathcal A}\wedge{\mathcal A}'|$ need not be equal to $|{\mathcal A}|\wedge|{\mathcal A}'|$ in general. Therefore we cannot replace ${\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})$ by ${\mathsf{ Att}}({\mathcal X},{\mathcal F})$ in Diagram~$(\ref{diag:AR2})$. The relationship between ${\mathsf{ Att}}({\mathcal X},{\mathcal F})$ and ${\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})$ is shown in Diagram (\ref{diag:comm-liftfull}). \end{remark} \subsection{Approximating dynamical systems} \label{subsec:approximationDS} In this section we address the question of how to approximate a general dynamical system $\varphi\colon {\mathbb{T}}^+\times X \to X$. If the time parameter ${\mathbb{T}} ={\mathbb{Z}}$, then the dynamical system is generated by the continuous map \[ f(\cdot) := \varphi(1,\cdot)\colon X\to X. \] In this case the dynamical system is represented by a (weak) outer approximation (Definition~\ref{defn:outerapprox}) and the results of Section~\ref{subsec:at-rep-neigh} apply. Thus, for the remainder of this section we assume that ${\mathbb{T}} ={\mathbb{R}}$, for which the question of choosing an appropriate representation is more subtle. For the definition of alpha and omega limit sets, and attractors and repellers we refer to \cite{KMV-1a}. Recall that a set $U\subset X$ is an {\em attracting neighborhood} for $\varphi\colon {\mathbb{R}}^+\times X\to X$ if $\omega(U,\varphi) \subset \mbox{\rm int\,}(U)$. A \emph{trapping region} is a forward invariant attracting neighborhood. A set $A\subset X$ is an \emph{attractor} if there exists an attracting neighborhood $U$ such that $A=\Inv (U,\varphi)$ in which case $A = \omega(U,\varphi)$. Repelling regions/neighborhoods and repellers can be define analogously, cf.\ \cite{KMV-1a}. The notion for attracting and repelling neighborhoods is ${\mathsf{ ANbhd}}(X,\varphi)$ and ${\mathsf{ RNbhd}}(X,\varphi)$, cf.\ \cite{KMV-1a}. \begin{remark} \label{rmk:isol} Attractors and repellers are examples of isolated invariant sets. In general, an invariant set $S\subset X$ is an \emph{isolated invariant set} if there exist a neighborhood $U\subset X$ such that $ S = \Inv(U,\varphi) \subset \mbox{\rm int\,}(U)$. The latter is called an i\emph{solating neighborhood}. The notion of isolated invariant set is also of importance beyond attractors and repellers. \end{remark} The following lemma provides a relationship between trapping regions for time-$\tau$ maps and attracting neighborhoods for $\varphi$. \begin{lemma} \label{lem:att-aux1} If $U$ is a trapping region for the time-$\tau$ map $f(\cdot)=\varphi(\tau,\cdot)$, then $U$ is an attracting neighborhood for $\varphi$. \end{lemma} \begin{proof} Let $U$ be a trapping region for $f$ and let $A = \omega(U,f)$ denote the associated attractor. Set $U^\tau = \varphi([0,\tau],U)$. The first step of the proof is to show that $U^\tau$ is forward invariant under $\varphi$. Observe that \begin{equation*} \begin{aligned} \varphi([n\tau,(n+1)\tau],U) &= \varphi(n\tau,\varphi([0,\tau],U))\\ &= \varphi([0,\tau],\varphi(n\tau,U)) \\ &= \varphi([0,\tau],f^n(U)) \subset \varphi([0,\tau],U) = U^\tau, \end{aligned} \end{equation*} where the inclusion follows from the forward invariance of $U$ under $f$. Thus $\varphi([0,\infty),U^\tau) = \varphi([1,\infty),\varphi([0,\tau],U))\subset U^\tau$. Since $U^\tau$ is forward invariant and $U^\tau = \varphi([0,\tau],U)$, \[ \omega(U,\varphi) = \omega(U^\tau,\varphi) \subset U^\tau. \] Observe that \[ {\rm cl}\,}%{\text{cl}\left(\bigcup_{k\geq n} f^k(U)\right)\subset {\rm cl}\,}%{\text{cl}(\varphi([n,\infty),U)). \] Thus \[ A = \bigcap_{n\ge 0} {\rm cl}\,}%{\text{cl}\bigl( \bigcup_{k\ge n} f^k(U)\bigr) \subset \bigcap_{n\ge 0}{\rm cl}\,}%{\text{cl}(\varphi([n,\infty),U)) = \omega(U,\varphi). \] Since $A$ is the maximal isolated invariant set for the time $\tau$-map $f$ in $U$, it follows from \cite[Theorem 1]{Mrozek1} that $A$ is the maximal isolated invariant set for $\varphi$ in $U$. In particular, \[ \omega(U,\varphi)= A \subset \mbox{\rm int\,}(U) \] and hence $U$ is an attracting neighborhood. \end{proof} \begin{proposition} \label{prop:discr-att-to-varphi3} Let ${\mathcal F}\colon {\mathcal X} \raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$} {\mathcal X}$ be an weak outer approximation for a time-$\tau$ mapping $f=\varphi(\tau,\cdot)$. Then, \begin{enumerate} \item[(i)] if ${\mathcal U}\subset {\mathcal X}$ is a forward invariant set for ${\mathcal F}$, then $U=|{\mathcal U}|$ is an attracting neighborhood for $\varphi$, and \item[(ii)] if ${\mathcal U}\subset {\mathcal X}$ is a backward invariant set for ${\mathcal F}$, then $U=|{\mathcal U}|$ is a repelling neighborhood for $\varphi$. \end{enumerate} \end{proposition} \begin{proof} Since $|{\mathcal U}|$ is a trapping region for $f$ by Proposition~\ref{prop:discr-att-to-varphi1}, Lemma~\ref{lem:att-aux1} imply that $\omega(|{\mathcal U}|,\varphi) \subset \mbox{\rm int\,}|{\mathcal U}|$, which proves that $U= |{\mathcal U}|$ is an attracting neighborhood for $\varphi$. If ${\mathcal U}\in {\mathsf{ Invset}}^-({\mathcal X},{\mathcal F})$, then ${\mathcal U}^c\in {\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})$, and thus $|{\mathcal U}^c| \in {\mathsf{ ANbhd}}(X,\varphi)$. By \cite[Corollary 3.26]{KMV-1a} we have that ${\rm cl}\,}%{\text{cl} \left(\supp{{\mathcal U}^c}^c\right)\in {\mathsf{ RNbhd}}(X,\varphi)$. Therefore, \[ |{\mathcal U}^c|^{\bigstar} ={\rm cl}\,}%{\text{cl} |{\mathcal U}^c|^c = |{\mathcal U}^{cc}| = |{\mathcal U}| \in {\mathsf{ RNbhd}}(X,\varphi), \] which proves the second statement. \end{proof} For attracting and repelling sets we can prove a similar statement, if we consider strong outer approximations instead of weak outer approximations. \begin{proposition} \label{prop:discr-att-to-varphi4} Let ${\mathcal F}\colon {\mathcal X} \raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$} {\mathcal X}$ be a outer approximation for a time-$\tau$ mapping $f=\varphi(\tau,\cdot)$. Then, \begin{enumerate} \item[(i)] if ${\mathcal U}\subset {\mathcal U}$ is an attracting set for ${\mathcal F}$, then $U=|{\mathcal U}|$ is an attracting neighborhood for $\varphi$, and \item[(ii)] if ${\mathcal U}\subset {\mathcal U}$ is a repelling set for ${\mathcal F}$, then $U=|{\mathcal U}|$ is a repelling neighborhood for $\varphi$. \end{enumerate} \end{proposition} \begin{proof} If ${\mathcal U}\in {\mathsf{ASet}}({\mathcal X},{\mathcal F})$, then from the proof of Proposition \ref{prop:discr-att-to-varphi2} it follows that $\varphi(k\tau,{\rm cl}\,}%{\text{cl} |{\mathcal U}|) \subset \mbox{\rm int\,} |{\mathcal U}|$. From the proof of Lemma \ref{lem:att-aux1} we then derive that $|{\mathcal U}|$ is an attracting neighborhood for $\varphi$. Proving that the same holds for ${\mathcal U}\in {\mathsf{RSet}}({\mathcal X},{\mathcal F})$ follows in the same way as in the proof of Proposition \ref{prop:discr-att-to-varphi3}. \end{proof} From the above results we conclude that \[ {\mathsf{ Att}}(X,\varphi) = {\mathsf{ Att}}(X,f)\quad\hbox{and}\quad {\mathsf{ Rep}}(X,\varphi) = {\mathsf{ Rep}}(X,f). \] The fact that ${\mathsf{ Rep}}(X,\varphi) = {\mathsf{ Rep}}(X,f)$ as sets also implies that they are the same as lattices, since the binary operations are $\cap$ and $\cup$. This implies that $\mathop{\rm id }\nolimits\colon {\mathsf{ Rep}}(X,\varphi) \to {\mathsf{ Rep}}(X,f)$ is a lattice isomorphism. For attractors, we have the same result. \begin{corollary} \label{for:sameA} The identity $\mathop{\rm id }\nolimits\colon {\mathsf{ Att}}(X,\varphi) \to {\mathsf{ Att}}(X,f)$ is a lattice isomorphism. \end{corollary} \begin{proof} Since as posets $\bigl({\mathsf{ Att}}(X,\varphi),\subset\bigr) = \bigl({\mathsf{ Att}}(X,f),\subset\bigr)$, it follows that $A \wedge A'$ in ${\mathsf{ Att}}(X,\varphi)$ is the same as $A\wedge A'$ in ${\mathsf{ Att}}(X,f)$. \end{proof} Since the evaluation mapping ${\mathcal U} \mapsto |{\mathcal U}|$ yields regular closed sets and is a Boolean homomorphism, we can summarize the above propositions in the following commuting diagram. \begin{equation} \label{diag:AR2a} \begin{diagram} \node{}\node{{\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})} \arrow{e,l,<>}{^c} \arrow{s,l,V}{\imath} \arrow{sw,l,A}{{\bm{\omega}}}\node{{\mathsf{ Invset}}^-({\mathcal X},{\mathcal F})} \arrow{s,r,V}{\imath}\arrow{se,l,A}{{\bm{\alpha}}} \\ \node{{\mathsf{ Att}}({\mathcal X},{\mathcal F})}\arrow{sse,r}{\omega(|\cdot|)}\node{{\mathsf{ASet}}({\mathcal X},{\mathcal F})}\arrow{s,l,V}{|\cdot|}\arrow{w,l,A}{{\bm{\omega}}}\arrow{e,l,<>}{^c} \node{{\mathsf{RSet}}({\mathcal X},{\mathcal F})}\arrow{s,l,V}{|\cdot|}\arrow{e,l,A}{{\bm{\alpha}}}\node{{\mathsf{ Rep}}({\mathcal X},{\mathcal F})} \arrow{ssw,r}{\alpha(|\cdot|)} \\ \node{}\node{{\mathsf{ ANbhd}}_{\scrR}(X,\varphi)}\arrow{s,l,A}{\omega}\arrow{e,l,<>}{^{\bigstar}} \node{{\mathsf{ RNbhd}}_{\scrR}(X,\varphi)}\arrow{s,l,A}{\alpha} \\ \node{}\node{{\mathsf{ Att}}(X,\varphi)} \arrow{e,l,<>}{^*} \node{{\mathsf{ Rep}}(X,\varphi)} \end{diagram} \end{equation} \section{Convergence and realization of algebraic structures via multivalued maps} \label{sec:str-conv} Convergent sequences of outer approximations can be constructed as indicated in \cite{KMV0}. For our purposes, we will use a modified notion of a convergent sequence of outer approximations and establish that arbitrary finite attractor lattices can be realized along such a sequence. We start with recalling some facts about outer approximations from \cite{KMV0}. \subsection{Convergent sequences of outer approximations} \label{subsec:conv-fam} Given a continuous map $f\colon X\to X$ and a grid indexed by ${\mathcal X}$, \cite[Proposition 2.5]{KMV0} implies that there is a natural choice of outer approximation which is minimal, namely \[ {\mathcal F}_{{\rm o}}(\xi) := \{ \eta\in {\mathcal X}~|~ f(|\xi|) \cap |\eta| \not =\varnothing\}. \] We refer to this as the \emph{minimal multivalued mapping} for $f$ with respect to ${\mathcal X}$. A multivalued mapping ${\mathcal F}$ \emph{encloses} a multivalued mapping ${\mathcal F}'$ if ${\mathcal F}'(\xi) \subset {\mathcal F}(\xi)$ for all $\xi\in {\mathcal X}$. Observe that this defines a partial order on multivalued mappings, which we denote by ${\mathcal F}'\le {\mathcal F}$. \begin{lemma} \label{lem:enclose10} (\cite[Corollary 2.6]{KMV0}) A multivalued mapping ${\mathcal F}\colon{\mathcal X}\raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$} {\mathcal X}$ is a outer approximation for $f\colon X\to X$ if and only if ${\mathcal F}$ encloses the minimal multivalued mapping ${\mathcal F}_{{\rm o}}$. \end{lemma} Outer approximations are naturally generated by numerical approximations of $f$. For $U\subset X$ let $\mathop{\rm cov}_{{\mathcal X}}(U):=\{\eta\in{\mathcal X}~|~U\cap |\eta|\neq\varnothing\}$. In this notation ${\mathcal F}_{{\rm o}}(\xi)=\mathop{\rm cov}_{{\mathcal X}}(f(|\xi|))$. More generally, let ${\bm{\varrho}}\colon {\mathcal X} \to [0,\infty)$ then by Lemma~\ref{lem:enclose10} \[ {\mathcal F}_{\bm\varrho}(\xi) := \bigl\{ \eta\in {\mathcal X}~|~ B_{\rho(\xi)}\bigl(f(|\xi|)\bigr) \cap |\eta| \not =\varnothing\bigr\}=\textstyle{\mathop{\rm cov}_{{\mathcal X}}}(B_{\rho(\xi)}(f(|\xi|))) \] is an outer enclosure. In the case ${\bm{\varrho}}(|\xi|)=\rho$ is constant for all $\xi\in{\mathcal X}$, we call ${\mathcal F}_\rho$ the \emph{$\rho$-minimal multivalued mapping} for $f$. Observe that with this multivalued map the 'errors' in the image of $f$ are always smaller than $\rho$ plus the grid size expressed by \begin{equation} \label{eqn:diam} {\rm diam}\,({\mathcal X}) = \max \{{\rm diam}\,(|\xi|)~|~\forall~\xi\in {\mathcal X}\}. \end{equation} To define convergent sequences of outer approximations we make use of multivalued mappings ${\mathcal F}$ that satisfy the {\em squeezing} condition \begin{equation} \label{eqn:squeeze1} {\mathcal F}_{{\rm o}} \le {\mathcal F}\le {\mathcal F}_{{\rho}}. \end{equation} Outer approximations can always be enclosed by some $\rho$-multivalued mapping. Indeed, by choosing $\rho = {\rm diam}\,(X)$, we have that ${\mathcal F}_{{\rho}}$ encloses every outer approximation ${\mathcal F}$. \begin{definition} \label{defn:conv10} Let ${\mathcal F}_n\colon {\mathcal X}_n\raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$}{\mathcal X}_n$ be a sequence of outer approximations for $f\colon X\to X$. Then ${\mathcal F}_n$ \emph{converges} if ${\rm diam}\,({\mathcal X}_n)\to 0$ and if there exist $\rho_n$-minimal maps ${\mathcal F}_{{\rho}_n}$ with $\rho_n\to 0$ such that \[ {\mathcal F}_{{\rm o},n} \le {\mathcal F}_n\le {\mathcal F}_{{\rho}_n}\;\hbox{on ${\mathcal X}_n$.} \] \end{definition} The following proposition extends the convergence result for minimal multivalued mappings of \cite[Proposition~5.4]{KMV0} to $\rho$-minimal multivalued mappings. \begin{proposition} \label{prop:conv2} Let $\epsilon>0$ and $k>0$. There exists $\delta >0$, such that for every grid indexed by ${\mathcal X}$, with ${\rm diam}\,({\mathcal X})<\delta$, and every $\rho$-minimal mapping with $\rho<\delta$ we have \begin{equation} \label{eqn:enclose3} |{\mathcal F}_{{\rho}}^k(\xi)| \subset B_\epsilon (f^k(|\xi|)) \quad\hbox{and}\quad |{\mathcal F}_{{\rho}}^{-k}(\xi)| \subset B_\epsilon (f^{-k}(|\xi|)) \end{equation} for all $\xi\in {\mathcal X}$. \end{proposition} \begin{proof} First consider forward dynamics. Since $X$ is a compact metric space, $f\colon X\to X$ is uniformly continuous so that for every $e>0$ there exists a $d>0$ such that \begin{equation} \label{eqn:uniform1} f(B_{d}(U)) \subset B_e(f(U)) \end{equation} for every $U\subset X$. Let $\delta_0 = \epsilon$. From Equation (\ref{eqn:uniform1}) choose $\delta_i>0$ for $i=1,\cdots, k-1$ inductively so that given $\delta_{i-1}$ \begin{equation} \label{eqn:uniform2} f\bigl(B_{\delta_i}(f^{k-i}(|\xi|)) \bigr)\subset B_{\delta_{i-1}/3}(f^{k-i+1}(|\xi|)) \end{equation} for all $\xi\in {\mathcal X}$. Define $\delta = \min_{0\le i\le k-1}\{\delta_i/3\}$. Let ${\mathcal X}$ be the indexing set for a grid on $X$ with ${\rm diam}\,({\mathcal X}) <\delta$, and consider ${\mathcal F}_\rho$ on ${\mathcal X}$ with $\rho<\delta$. For the evaluation $|{\mathcal F}_{{\rho}}(\xi)|$ we have \begin{equation} \label{eqn:uniform3} |{\mathcal F}_{{\rho}}(\xi)| \subset B_{\delta+\rho}(f(|\xi|)) \subset B_{2\delta}(f(|\xi|))\subset B_{\delta_{k-1}}(f(|\xi|)). \end{equation} We now proceed inductively. Recall that ${\mathcal F}_{{\rho}}^2(\xi) = \bigcup_{\eta\in {\mathcal F}_{{\rho}}(\xi)} {\mathcal F}_{{\rho}}(\eta)$, and for $\eta\in {\mathcal F}_{{\rho}}(\xi)$ Equation (\ref{eqn:uniform3}) implies $\eta\subset B_{\delta_{k-1}}(f(|\xi|))$. Combining this with (\ref{eqn:uniform2}) we obtain \[ f(|\eta|) \subset f\bigl( B_{\delta_{k-1}}(f(|\xi|)) \bigr) \subset B_{\delta_{k-2}/3}(f^2(|\xi|)), \] and by (\ref{eqn:uniform3}) applied to $\eta$ we have \[ |{\mathcal F}_{{\rho}}(\eta)| \subset B_{2\delta}(f(|\eta|)). \] Therefore, \[ |{\mathcal F}_{{\rho}}^2(\xi)| = \bigcup_{\eta\in {\mathcal F}_{{\rho}}(\xi)} |{\mathcal F}_{{\rho}}(\eta)| \subset B_{\delta_{k-2}/3+2\delta}(f^2(|\xi|))\subset B_{\delta_{k-2}}(f^2(|\xi|)). \] In the general case $i\ge 2$, for $\eta\in {\mathcal F}_{{\rho}}^{i-1}(\xi)$ we have $\eta\subset B_{\delta_{k-i+1}}(f^{i-1}(|\xi|))$ and $f(\eta) \subset B_{\delta_{k-i}/3}(f^i(|\xi|))$. This yields \[ |{\mathcal F}_{{\rho}}^i(\xi)| = \bigcup_{\eta\in {\mathcal F}_{{\rho}}^{i-1}(\xi)} |{\mathcal F}_{{\rho}}(\eta)| \subset B_{\delta_{k-i}/3+2\delta}(f^i(|\xi|))\subset B_{\delta_{k-i}}(f^i(|\xi|)). \] After $k$ steps, we obtain $|{\mathcal F}_{{\rho}}^k(\xi)|\subset B_{\delta_{0}}(f^k(|\xi|))= B_{\epsilon}(f^k(|\xi|))$, which completes the proof for forward dynamics. In the case of backward dynamics, we use the fact that for every $e>0$ there exists a $d>0$ such that \begin{equation} \label{eqn:uniform1back} f^{-1}(B_{d}(U)) \subset B_e(f^{-1}(U)) \end{equation} for every $U\subset X$, by continuity of $f$ and compactness of $X$. Moreover, we will use the following characterization of ${\mathcal F}_\rho^{-1}$ \begin{eqnarray*} {\mathcal F}_\rho^{-1}(\xi)&=&\{\eta\in{\mathcal X}~|~B_\rho(f(\eta))\cap|\xi|\neq\varnothing\}=\{\eta\in{\mathcal X}~|~B_\rho(|\xi|)\cap |\eta|\neq\varnothing\}\\ &=&\{\eta\in{\mathcal X}~|~|\eta|\cap f^{-1}(B_\rho(|\xi|))\neq\varnothing\} =\textstyle{\mathop{\rm cov}_{{\mathcal X}}}(f^{-1}(B_\rho(|\xi|))). \end{eqnarray*} Now we proceed similarly to the previous case. Let $\delta_0 = \epsilon$. From Equation (\ref{eqn:uniform1back}) choose $\delta_i>0$ for $i=1,\cdots, k$ inductively so that given $\delta_{i-1}$ \begin{equation}\label{eqn:uniform2back} f^{-1}\bigl(B_{\delta_i}(f^{-k+i}(|\xi|)) \bigr)\subset B_{\delta_{i-1}/3}(f^{-k+i-1}(|\xi|)) \end{equation} for all $\xi\in {\mathcal X}$. Define $\delta = \min_{0\le i\le k}\{\delta_i/3\}$. Then, since $\rho<\delta<\delta_i$, \begin{eqnarray} |{\mathcal F}_\rho^{-1}(\xi)| &=& \bigl|\{\eta\in{\mathcal X}~|~|\eta|\cap f^{-1}(B_\rho(|\xi|))\neq\varnothing\}\bigr|\nonumber \\ &\subset& \bigl|\{\eta\in{\mathcal X}~|~|\eta|\cap f^{-1}(B_{\delta_{k-i+2}}(|\xi|))\neq\varnothing\}\bigr|\nonumber \\ &\subset& B_{\delta_{k-i+1}/3+\delta}(f^{-1}(|\xi|) \label{eqn:uniform3back} \end{eqnarray} for all $i=2,\ldots,k$ and $\xi\in{\mathcal X}$. We now proceed inductively. Recall that ${\mathcal F}_{{\rho}}^{-2}(\xi) = \bigcup_{\eta\in {\mathcal F}_{{\rho}}^{-1}(\xi)} {\mathcal F}_{{\rho}}^{-1}(\eta)$, and for $\eta\in {\mathcal F}_{{\rho}}^{-1}(\xi)$ Equation (\ref{eqn:uniform3back}) implies $\eta\subset B_{\delta_{k-1}/3+\delta}(f^{-1}(|\xi|))$. Combining this with (\ref{eqn:uniform2back}) we obtain \begin{eqnarray*} f^{-1}(B_\rho(|\eta|)) &\subset& f^{-1}\bigl( B_{\delta_{k-1}/3+2\delta}(f^{-1}(|\xi|)) \bigr) \\ &\subset& f^{-1}\bigl( B_{\delta_{k-1}}(f^{-1}(|\xi|)) \bigr) \subset B_{\delta_{k-2}/3}(f^{-2}(|\xi|)). \end{eqnarray*} Therefore, \[ |{\mathcal F}_{{\rho}}^{-2}(\xi)| = \bigcup_{\eta\in {\mathcal F}_{{\rho}}^{-1}(\xi)} |{\mathcal F}_{{\rho}}^{-1}(\eta)| \subset B_{\delta_{k-2}/3+\delta}(f^{-2}(|\xi|)). \] In the general case $i\ge 2$, for $\eta\in {\mathcal F}_{{\rho}}^{-i+1}(\xi)$ we have $\eta\subset B_{\delta_{k-i+1}/3+\delta}(f^{-i+1}(|\xi|))$ and $f^{-1}(B_\rho(|\eta|)) \subset B_{\delta_{k-i}/3}(f^{-i}(|\xi|))$. This yields \[ |{\mathcal F}_{{\rho}}^{-i}(\xi)| = \bigcup_{\eta\in {\mathcal F}_{{\rho}}^{-i+1}(\xi)} |{\mathcal F}_{{\rho}}^{-1}(\eta)| \subset B_{\delta_{k-i}/3+\delta}(f^{-i}(|\xi|)). \] After $k$ steps, we obtain $|{\mathcal F}_{{\rho}}^{-k}(\xi)|\subset B_{\delta_{0}/3+\delta}(f^{-k}(|\xi|))\subset B_{\delta_0}(f^{-k}(|\xi|))=B_{\epsilon}(f^{-k}(|\xi|))$, which completes the proof for backward dynamics. \end{proof} \begin{proposition} \label{prop:conv11} Let ${\mathcal F}_n\colon{\mathcal X}_n\raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$}{\mathcal X}_n$ be a convergent sequence of outer approximations for $f\colon X\to X$. For every $\epsilon>0$ and every $k>0$, there exists $N>0$ such that \begin{equation*} |{\mathcal F}_n^k(\xi)| \subset B_\epsilon (f^k(|\xi|)) \quad\hbox{and}\quad |{\mathcal F}_n^{-k}(\xi)| \subset B_\epsilon (f^{-k}(|\xi|)) \end{equation*} for all $n\ge N$ and for all $\xi\in {\mathcal X}_n$. \end{proposition} \begin{proof} We start with the observation that ${\mathcal F}_n^k \le{\mathcal F}_{{\rho}_n}^k$. Indeed, suppose true for $k-1$, then \[ {\mathcal F}_n^k(\xi) = \bigcup_{\eta\in {\mathcal F}_n^{k-1}(\xi)} {\mathcal F}_n(\eta) \subset \bigcup_{\eta\in {\mathcal F}_n^{k-1}(\xi)}{\mathcal F}_{{\rho}_n}(\eta) \subset \bigcup_{\eta\in {\mathcal F}_{{\rho}_n}^{k-1}(\xi)} {\mathcal F}_{{\rho}_n}(\eta) = {\mathcal F}_{{\rho}_n}^k(\xi). \] To complete the proof we choose $\delta>0$ such that the conclusion of Proposition \ref{prop:conv2} holds. Choose $N>0$ such that $\rho_n<\delta$ for all $n\ge N_1$, and choose $N_2>0$ such that ${\rm diam}\,({\mathcal X}_n) \le \delta$ for all $n\ge N_2$. Choosing $N= \max\{N_1,N_2\}$ completes the proof. \end{proof} \subsection{Realization of attractors and repellers} \label{subsec:resol} Theorem~\ref{thm:commutative-diagram} guarantees that forward invariant sets and attractors for an outer approximation ${\mathcal F}$ of $f$ yield attracting neighborhoods for $f$. The converse statement is that every attractor of a dynamical system can be realized by an outer approximation provided the diameter of the grid is sufficiently small. Our goal is the stronger result that the lattice structure of attractors can be realized. We start by generalizing \cite[Proposition~5.5]{KMV0} from the context of minimal multivalued maps to the setting of convergent sequences of outer approximations. \begin{proposition} \label{prop:conv-latt-3} Let ${\mathcal F}_n\colon {\mathcal X}_n \raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$} {\mathcal X}_n$ be a convergent sequence of outer approximations for $f$, and let $U\in {\mathsf{ ANbhd}}(X,f)$ or $U\in {\mathsf{ RNbhd}}(X,f)$. Then there exists $N>0$ such that for all $n\ge N$ the set ${\mathcal U} = \mathop{\rm cov}_{{\mathcal X}_n}(U)$ is an attracting or repelling set for ${\mathcal F}_n$, respectively. \end{proposition} \begin{proof} We consider the case $U\in {\mathsf{ ANbhd}}(X,f)$, the other case is analogous. Let $A = \omega(U)$ and let $0<d < \frac{1}{2} {\rm dist}\,(U,A^*)$. Since $U$ is an attracting neighborhood by \cite[Proposition 3.21]{KMV-1a}, there exists $K>0$ such that $f^k(B_d(U)) \subset \mbox{\rm int\,}(U)$ for all $k\ge K$. This implies that for $K \le k\le 2K$ there exists an $\epsilon>0$ such that $B_\epsilon\bigl(f^k(B_d(U))\bigr) \subset \mbox{\rm int\,}(U)$. By Proposition \ref{prop:conv11} we can choose $N$ such that $|{\mathcal F}_n^k(|\xi|)| \subset B_\epsilon(f^k(|\xi|))$ for all $K \le k \le 2K, \xi \in {\mathcal X}_n,$ and $n\ge N$. We also choose $N$ such that ${\mathcal U}=\mathop{\rm cov}_{{\mathcal X}_n}(U) \subset B_d(U)$. This yields \[ |{\mathcal F}_n^k({\mathcal U})|\subset \bigcup_{\xi\in {\mathcal U}} B_\epsilon(f^k(|\xi|))\subset \mbox{\rm int\,}(U) \subset |{\mathcal U}|, \] which implies that ${\mathcal F}_n^k({\mathcal U})\subset {\mathcal U}$ for all $K\le k\le 2K$. Thus ${\mathcal F}_n^k({\mathcal U})\subset {\mathcal U}$ for all $k \ge K$, since, for example, ${\mathcal F}_n^{2K+k}=({\mathcal F}_n^K)^{K+k}$ for all $0<k\le K$. Using Proposition \ref{prop:char-attset}, this proves that ${\mathcal U}$ is an attracting set when $n$ is sufficiently large, i.e. ${\rm diam}\,({\mathcal X}_n)$ is sufficiently small. \end{proof} This leads to the following corollary, which in the case of the minimal multivalued map is also a consequence of \cite[Proposition~5.5]{KMV0}. \begin{corollary} \label{cor:resolution1} Let ${\mathcal F}_n\colon {\mathcal X}_n\raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$} {\mathcal X}_n$ be convergent sequence of outer approximations for $f$, and let $A\in {\mathsf{ Att}}(X,f)$ be an attractor for $f$. For every $0< d < \frac{1}{2} {\rm dist}\,(A,A^*)$ there exists an $N>0$ such that for every $n\ge N$ there is an attractor ${\mathcal A}_n \in {\mathsf{ Att}}({\mathcal X}_n,{\mathcal F}_n)$ and a repeller ${\mathcal R}_n \in {\mathsf{ Rep}}({\mathcal X}_n,{\mathcal F}_n)$ with \[ A = \omega(|{\mathcal A}_n|)\subset |{\mathcal A}_n| \subset B_d(A)\quad\hbox{and}\quad A^* = \alpha(|{\mathcal R}_n|)\subset |{\mathcal R}_n| \subset B_d(A^*). \] \end{corollary} \begin{proof} Fix $0< d < \frac{1}{2} {\rm dist}\,(A,A^*)$. By Proposition~\ref{prop:conv-latt-3}, there exists $N>0$ such that ${\mathcal U}_n = \mathop{\rm cov}_{{\mathcal X}_n}(B_{d/2}(A))$ is an attracting set for ${\mathcal F}_n$ and ${\mathcal V}_n = \mathop{\rm cov}_{{\mathcal X}_n}(B_{d/2}(A^*))$ is a repelling set for ${\mathcal F}_n$ for all $n\ge N$, since $B_{d/2}(A)$ and $B_{d}(A)$ are attracting neighborhoods, and $B_{d/2}(A^*)$ and $B_{d}(A^*)$ are repelling neighborhoods. Choosing $N$ large enough so that ${\rm diam}\,({\mathcal X}_n)<d/2$ as well implies that $|{\mathcal U}_n|\subset B_d(A)$ and $|{\mathcal V}_n|\subset B_d(A^*)$. Moreover, if ${\mathcal A}_n={\bm{\omega}}({\mathcal U}_n)$ and ${\mathcal R}_n={\bm{\alpha}}({\mathcal V}_n)$, then ${\mathcal A}_n\subset {\mathcal U}_n$ and ${\mathcal R}_n\subset {\mathcal V}_n$ implies $|{\mathcal A}_n|\subset B_d(A)$ and $|{\mathcal R}_n|\subset B_d(A^*)$. By Proposition~\ref{lem:commute1}, $\omega(|{\mathcal A}_n|)=\omega(|{\bm{\omega}}({\mathcal U}_n)|)=\omega(|{\mathcal U}_n|)$. Moreover $A=\omega(B_{d/2}(A))\subset\omega(|{\mathcal U}_n|)\subset\omega(B_d(|{\mathcal U}_n|)=A$ so that $A=\omega(|{\mathcal A}_n|)$. By Proposition~\ref{prop:discr-att-to-varphi2}, $|{\mathcal A}_n|$ is an attracting neighborhood so that $\omega(|{\mathcal A}_n|)\subset|{\mathcal A}_n|$, which completes the proof for the attractor. The same argument holds for the repeller. \end{proof} \subsection{Posets, Lattices and Grids} \label{subsec:PLG} In the previous subsection we established that any attractor or repeller in a system can be realized via a multivalued map if the diameter of the grid is sufficiently small. Furthermore, these discrete attractors and repellers correspond to arbitrarily narrow attracting and repelling neighborhoods, respectively. To prove that the lattice structures can be realized via multivalued maps requires more subtle constructions based on the lattice and poset structures of grids and multivalued maps. We begin by providing a systematic means of generating convergent sequences of outer approximations. For the sake of simplicity we will abuse notation throughout this subsection and often refer to the grid $\setof{\supp{\xi}\mid \xi\in{\mathcal X}}$ by its indexing set ${\mathcal X}$. \begin{definition} \label{defn:refinement} A grid ${\mathcal X}'$ on $X$ is a {\em refinement of ${\mathcal X}$}, denoted by ${\mathcal X}'\le {\mathcal X}$, if for every $\xi'\in {\mathcal X}'$ there exists exactly one $\xi\in {\mathcal X}$ such that $|\xi'|\subset |\xi|$. \end{definition} Refinement defines a partial order on the space of grids ${\mathsf{Grid}}(X)$, which can be used to compare multivalued maps. \begin{definition} \label{defn:cofil1} Let ${\mathcal X}' \le {\mathcal X}$ be grids on $X$ and let ${\mathcal F}\colon {\mathcal X} \raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$} {\mathcal X}$ and ${\mathcal F}\colon{\mathcal X}' \raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$} {\mathcal X}'$ be multivalued mappings. A partial order on multivalued mappings and grids is given by \[ {\mathcal F}'\le {\mathcal F} \quad\text{if}\quad \supp{{\mathcal F}'({\mathcal U}')} \subseteq \supp{{\mathcal F}({\mathcal U})}\ \text{for all}\ \supp{{\mathcal U}'} = \supp{{\mathcal U}} \] where ${\mathcal U}'\in {\mathsf{Set}}({\mathcal X}')$ and ${\mathcal U}\in {\mathsf{Set}}({\mathcal X})$. \end{definition} \begin{definition} \label{defn:commref} The \emph{common refinement} of ${\mathcal X}$ and ${\mathcal X}'$ is the grid \[ \{|\xi|\wedge|\xi'|~|~\hbox{$\xi\in{\mathcal X}$ and $\xi'\in{\mathcal X}'$ with $|\xi|\wedge|\xi'|\neq\varnothing$}\}. \] The set of all pairs $\xi$ and $\xi'$ for which $|\xi|\wedge |\xi'|\neq\varnothing$ is an indexing set for this grid. We denote this indexing set by ${\mathcal X}\wedge{\mathcal X}'$ and an individual index by $\xi\wedge\xi'$. \end{definition} Note that whenever the index $\xi\wedge\xi'$ is used, it is implied that $|\xi|\wedge|\xi'|\neq\varnothing.$ The \emph{common refinement of multivalued mappings} ${\mathcal F}\colon {\mathcal X} \raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$} {\mathcal X}$ and ${\mathcal F}\colon{\mathcal X}' \raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$} {\mathcal X}'$ is given by \begin{equation} \label{eqn:common-milt} ({\mathcal F}\wedge {\mathcal F}')(\xi\wedge\xi') := \{ \eta\wedge\eta' ~|~\eta\in {\mathcal F}(\xi),~\eta'\in {\mathcal F}'(\xi)\}. \end{equation} Observe that ${\mathcal F}\wedge {\mathcal F}'\colon {\mathcal X}\wedge {\mathcal X}'\raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$} {\mathcal X}\wedge {\mathcal X}'$. \begin{definition} \label{defn:cofilt} A \emph{cofiltration of grids} is a sequence $\{{\mathcal X}_n\}_{n\in {\mathbb{N}}_0} \subset {\mathsf{Grid}}(X)$ of refinements so that \[ {\mathcal X}_0 \ge {\mathcal X}_1 \ge \cdots \ge {\mathcal X}_n\ge \cdots \] Furthermore, given a cofiltration of grids $\{{\mathcal X}_n\}_{n\in {\mathbb{N}}_0}$, a sequence of multivalued mappings ${\mathcal F}_n\colon {\mathcal X}_n \raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$} {\mathcal X}_n$, which satisfies \[ {\mathcal F}_0 \ge {\mathcal F}_1 \ge \cdots \ge {\mathcal F}_n \ge \cdots \] is called a \emph{cofiltration of multivalued mappings}. The function ${\rm diam}\,\colon {\mathsf{Grid}}(X) \to {\mathbb{R}}^+$ is order-preserving so that ${\rm diam}\,({\mathcal X}') \le {\rm diam}\,({\mathcal X})$ for any pair ${\mathcal X}'\le {\mathcal X}$. If ${\rm diam}\,({\mathcal X}_n) \to 0$ as $n\to \infty$, then a cofiltration $\{{\mathcal X}_n\}_{n\in {\mathbb{N}}_0}$ of grids is said to be \emph{contracting}. \end{definition} Given any sequence of grids $\{{\mathcal X}_n\}$ for which ${\rm diam}\,({\mathcal X}_n) \to 0$, we can construct a contracting cofiltration as follows \[ {\mathcal X}_0 \ge {\mathcal X}_0\wedge {\mathcal X}_1 \ge {\mathcal X}_0\wedge {\mathcal X}_1 \wedge {\mathcal X}_2 \ge \cdots \ge \bigwedge_{i=0}^n {\mathcal X}_i \ge \cdots . \] If ${\mathcal F}_n\colon {\mathcal X}_n \raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$} {\mathcal X}_n$ is a sequence of multivalued mappings with ${\rm diam}\,({\mathcal X}_n) \to 0$, then \begin{equation} \label{eqn:cofil2} \bigwedge_{i=1}^n {\mathcal F}_i \colon \bigwedge_{i=1}^n {\mathcal X}_i \raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$} \bigwedge_{i=1}^n {\mathcal X}_i \end{equation} is a cofiltration of multivalued mappings. From an algorithmic point of view, given a grid ${\mathcal X}_n$, one designs an algorithm to construct ${\mathcal F}_n:{\mathcal X}_n\raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$}{\mathcal X}_n$. The monotonicity of images of ${\mathcal F}_n$ required for a cofiltration does not automatically follow from the fact that ${\mathcal X}_n$ is a cofiltration of grids, and the construction in equation~$(\ref{eqn:cofil2})$ is often inefficient in practical applications. From a theoretical point of view, an important example of a convergent cofiltration of multivalued mappings is given by the $\rho$-minimal multivalued mappings on a contracting cofiltration of grids. Theorems~\ref{thm:conv-latt-2} and~\ref{thm:conv-latt-3} contrast what is attainable through a convergent cofiltration versus simply a convergent sequence of multivalued mappings, see Remark~\ref{rmk:mono}. Some properties of dynamics are preserved through cofiltrations. We only present the following which we make use of in the proof of Theorem~\ref{thm:conv-latt-2}. \begin{proposition} \label{prop:cofiltmap} Let $\{{\mathcal X}_n\}_{n\in {\mathbb{N}}_0}$ be a cofiltration of grids and let ${\mathcal F}_n\colon {\mathcal X}_n\raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$} {\mathcal X}_n$ be a cofiltration of multivalued mappings. Consider a collection of subsets ${\mathcal W}_n\subset {\mathcal X}_n$ such that $\supp{{\mathcal W}_n} = \supp{{\mathcal W}_m}\subset X$. If $m > n$ and ${\mathcal W}_n \in {\mathsf{ Invset}}^-({\mathcal X}_n,{\mathcal F}_n)$, then ${\mathcal W}_m \in {\mathsf{ Invset}}^-({\mathcal X}_m,{\mathcal F}_m)$. \end{proposition} \begin{proof} We need to show that ${\mathcal F}_m^{-1}({\mathcal W}_m)\subset {\mathcal W}_m$. Since ${\mathcal W}_n \in {\mathsf{ Invset}}^-({\mathcal X}_n,{\mathcal F}_n)$, it is sufficient to show that $\supp{{\mathcal F}_m^{-1}({\mathcal W}_m)} \subset \supp{{\mathcal F}_n^{-1}({\mathcal W}_n)}$ so that \[ \supp{{\mathcal F}_m^{-1}({\mathcal W}_m)} \subset \supp{{\mathcal F}_n^{-1}({\mathcal W}_n)} \subset \supp{{\mathcal W}_n} = \supp{{\mathcal W}_m}. \] Let $\beta_n\in{\mathcal W}_n$ and $\beta_m\in{\mathcal W}_m$ satisfy $\supp{\beta_m} \subset \supp{\beta_n}$. Consider $\eta\in{\mathcal F}_m^{-1}(\beta_m)$. By definition of cofiltration, there exists $\xi\in{\mathcal X}_n$ such that $\supp{\eta}\subset\supp{\xi}$. Now $\beta_m\in{\mathcal F}_m(\eta)$ which implies that \[ \supp{\beta_m}\subset \supp{{\mathcal F}_m(\eta)}\subset\bigcup_{\supp{\zeta}\subset\supp{\xi}}|{\mathcal F}_m(\zeta)|= \supp{{\mathcal F}_m\left(\bigcup_{\supp{\zeta}\subset\supp{\xi}}\zeta\right)} \subset |{\mathcal F}_n(\xi)| \] where the last inclusion follows from the definition of cofiltration. Since ${\mathcal F}_n(\xi)$ is a union of elements of ${\mathcal X}_n$, and $\supp{\beta_m}\subset \supp{\beta_n}$, we must have $\supp{\beta_n}\subset|{\mathcal F}_n(\xi)|$, which implies $\beta_n\in{\mathcal F}_n(\xi)$ and equivalently $\xi\in{\mathcal F}_n^{-1}(\beta_n)$. Hence $\xi\in{\mathcal F}_n^{-1}({\mathcal W}_n)$ and $\supp{\eta}\subset\supp{\xi}\subset \supp{{\mathcal F}_n^{-1}({\mathcal W}_n)}$. Thus, if $\eta\in{\mathcal F}_m^{-1}({\mathcal W}_m)$, then $\supp{\eta}\subset \supp{{\mathcal F}_n^{-1}({\mathcal W}_n)}$, and therefore $\supp{{\mathcal F}_m^{-1}({\mathcal W}_m)} \subset \supp{{\mathcal F}_n^{-1}({\mathcal W}_n)}$. \end{proof} The realization of the lattice structures of attractors and repellers is presented in the language of lifts which are defined as follows. \begin{definition} \label{defn:lift} Let ${\mathsf L}$, ${\mathsf K}$, and ${\mathsf H}$ be bounded distributive lattices. Let $g\colon {\mathsf L} \rightarrowtail {\mathsf K}$ be a lattice monomorphism and $h\colon {\mathsf H}\twoheadrightarrow {\mathsf K}$ be a lattice epimorphism. A lattice homomorphism $\ell\colon {\mathsf L}\to {\mathsf H}$ is a {\em lift of $g$ through $h$} if $g = h\circ \ell$. \end{definition} Observe that a lift is necessarily a lattice monomorphism. As is made clear in the next section our goal is to construct a lift. To do this we make use of concepts from the theory of distributive lattices. Recall that an element $c\in{\mathsf L}$ is {\em join-irreducible} if \begin{enumerate} \item[(a)] $c\neq 0$ and \item[(b)] $c=a\vee b$ implies $c=a$ or $c=b$ for all $a,b\in{\mathsf L}$. \end{enumerate} The set of join-irreducible elements in ${\mathsf L}$ is denoted by ${\mathsf J}({\mathsf L})$. Note that ${\mathsf J}({\mathsf L})$ is a poset as a subset of ${\mathsf L}$. Observe that $c$ is join-irreducible if and only if there exists a unique element $a\in{\mathsf L}$ satisfying $a<c$ and there does not exist $b\in{\mathsf L}$ such that $a<b<c$. The element $a\in {\mathsf L}$ is called the \emph{immediate predecessor} of $c$ and denoted by \begin{equation} \label{eqn:immpred} a = ^\leftarrow c. \end{equation} Given a finite poset ${\mathsf P}$ with partial order $\le$, then the down-set of $p\in {\mathsf P}$ is given by $\downarrow\! p=\{q\in{\mathsf P}~|~q\le p\}$. These sets generate a finite distributive lattice ${\mathsf O}({\mathsf P})$ in ${\mathsf{Set}}({\mathsf P})$ called the lattice of {\em down-sets}. The elements $\downarrow\! p$ are the join-irreducible elements in ${\mathsf O}({\mathsf P})$. Birkhoff's representation theorem for finite distributive lattices ${\mathsf L}$ states that ${\mathsf L} \cong {\mathsf O}({\mathsf J}({\mathsf L}))$ cf.\ \cite{Davey:2002p5923}. \begin{remark} \label{rem:birkhoff} Birkhoff's representation theorem allows us to recast the definition of $\ell$ being a lift of $g$ through $h$ via the following commutative diagram \begin{equation} \label{diag:lift12} \begin{diagram} \node{~} \node{{\mathsf H}} \arrow{s,r,A}{h}\\ \node{{\mathsf O}({\mathsf P})} \arrow{ne,l,..,V}{\ell} \arrow{e,l,V}{g} \node{{\mathsf K}} \end{diagram} \end{equation} for any poset ${\mathsf P}$ isomorphic to ${\mathsf J}({\mathsf L})$. For the sake of simplicity we will abuse notation and use $\ell\colon {\mathsf L}\to {\mathsf H}$ and $\ell\colon {\mathsf O}({\mathsf P})\to {\mathsf H}$ to denote two distinct, but equivalent homomorphisms. \end{remark} We are interested in the case in which ${\mathsf H}$ is a Boolean algebra, or ${\mathsf H}$ is embedded in a Boolean algebra, and thus we want to extend the lift $\ell\colon {\mathsf L} \to {\mathsf H}$ to a Boolean homomorphism. To do this we make use of the Booleanization functor. The natural extension ${\mathsf L} \hookrightarrow {\mathsf{Set}}({\mathsf J}({\mathsf L}))$ is called the Booleanization of ${\mathsf L}$ and is denoted by ${\mathsf B}({\mathsf L}) = {\mathsf{Set}}({\mathsf J}({\mathsf L}))$. Booleanization is a covariant functor and the induced homomorphism ${\mathsf B}(\ell) \colon {\mathsf B}({\mathsf L}) \to {\mathsf H}$ is Boolean and ${\mathsf B}(\ell)|_{\mathsf L} = \ell$, cf.\ \cite[Definition 9.5.5]{vickers} and \cite[Corollary 20.11]{miraglia}. The combination of Birkhoff's representation theorem and the Booleanization functor allows one to give the following representation of $\ell$: \begin{equation} \label{eqn:inatom} \ell(\alpha) = \bigvee_{p\in \alpha} c_p \end{equation} where $c_p := \ell(\gamma) \setminus \ell(\beta)\in{\mathsf H}$ and is independent of the choice of $\beta,\gamma\in {\mathsf O}({\mathsf J}({\mathsf L}))$ for which $\gamma \setminus \beta = p$, cf.\ Theorem 2.1 and Proposition 2.3 in \cite{KMV-1a}. Observe that the $c_p$ are atoms of ${\mathsf H}$, i.e.\ if $p\not = p'$, then $c_p \wedge c_{p'} = 0$. With these abstract constructions in mind, we now turn to the objects of interest. As is detailed in the next section, we are interested in lifts of the form $\ell\colon {\mathsf R} \to {\mathsf{ Invset}}^-({\mathcal X},{\mathcal F})$ where ${\mathsf R}\subset {\mathsf{ Rep}}(X,f)$ is a finite sublattice of repellers and such that $\alpha(|\ell(R)|) = R$ for all $R \in {\mathsf R}$. Since ${\mathsf{ Invset}}^-({\mathcal X},{\mathcal F})$ embeds (as a lattice) into the Boolean algebra ${\mathsf{Set}}({\mathcal X})$, we can adopt the perspective that $\ell\colon {\mathsf R} \to {\mathsf{Set}}({\mathcal X})$ is a lattice monomorphism. If we represent ${\mathsf R}$ by a lattice isomorphism ${\mathsf O}({\mathsf P}) \cong {\mathsf R}$, where ${\mathsf P}\cong{\mathsf J}({\mathsf R})$, then application of the Booleanization functor to $\ell\colon{\mathsf O}({\mathsf P}) \to {\mathsf{ Invset}}^-({\mathcal X},{\mathcal F})$ yields the Boolean monomorphism ${\mathsf B}(\ell) \colon {\mathsf{Set}}({\mathsf P})\to {\mathsf{Set}}({\mathcal X})$. This allows us to represent $\ell$ by \begin{equation} \label{eqn:props-10} \ell(\alpha) = \bigcup_{p\in\alpha} {\mathcal V}_p \end{equation} where \begin{equation} \label{eq:Vp} {\mathcal V}_p := \ell(\gamma)\setminus \ell(\beta)\subset {\mathcal X} \end{equation} for any choice of $\beta,\gamma\in{\mathsf O}({\mathsf P})$ such that $\setof{p}=\gamma\setminus\beta$. Since $\setof{{\mathcal V}_p\mid p\in{\mathsf P}}$ are atoms, ${\mathcal V}_p\cap {\mathcal V}_{p'} = \varnothing$ if $p\neq p'$. \begin{proposition} \label{prop:Boolappl} Let ${\mathsf O}({\mathsf P})$ be a finite distributive lattice, and let $\ell\colon {\mathsf O}({\mathsf P}) \to {\mathsf{Set}}({\mathcal X})$ be a lattice monomorphism. Then ${\mathsf P}$ is an indexing set for a grid on $X$ whose elements are $ \supp{{\mathcal V}_p} $ under the evaluation map $\supp{\cdot}\colon {\mathsf{Set}}({\mathcal X})\to {\mathscr R}(X)$ and ${\mathcal V}_p := \ell(\gamma)\setminus \ell(\beta)\subset {\mathcal X}$ for any choice of $\beta,\gamma\in{\mathsf O}({\mathsf P})$ such that $\setof{p}=\gamma\setminus\beta$. \end{proposition} \begin{proof} By Corollary~\ref{cor:grid->lattice} the evaluation map $|\cdot|\colon {\mathsf{Set}}({\mathcal X})\to {\mathscr R}(X)$ is Boolean, and thus the composition $\supp{{\mathsf B}(\ell)}\colon {\mathsf{Set}}({\mathsf P})\to {\mathscr R}(X)$ is Boolean. In particular $\supp{{\mathsf B}(\ell)({\mathsf P})}$ is a finite subalgebra of ${\mathscr R}(X)$. Hence the atoms of $\supp{{\mathsf B}(\ell)({\mathsf P})}$, which are $\setof{\supp{{\mathcal V}_p} \mid p\in {\mathsf P}}$, form a grid of $X$. \end{proof} \begin{proposition} Let ${\mathsf O}({\mathsf P})$ be a finite distributive lattice and let $\ell\colon {\mathsf O}({\mathsf P}) \to {\mathsf{Set}}({\mathcal X})$ be a lattice monomorphism. Then \[ \supp{{\mathcal V}_p} \cap \mbox{\rm int\,} \supp{\ell(\alpha)} = \varnothing \quad\text{for all}\quad p\not\in\alpha \in {\mathsf O}({\mathsf P}). \] \end{proposition} \begin{proof} By \eqref{eqn:props-10} and \eqref{eq:Vp} we have that ${\mathcal V}_p \cap l(\alpha) = \varnothing$. Because $\setof{{\mathcal V}_p\mid p\in{\mathsf P}}$ is a grid for $X$ we obtain \[ \supp{{\mathcal V}_p} \wedge \supp{\ell(\alpha)} = \varnothing,\quad \forall p\not\in \alpha \in {\mathsf O}({\mathsf P}) \] which is equivalent to $\supp{{\mathcal V}_p} \cap \mbox{\rm int\,} \supp{\ell(\alpha)} = \varnothing$ by Lemma~\ref{lem:equivregcl}. \end{proof} Because $\setof{{\mathcal V}_p\mid p\in {\mathsf P}}$ are atoms, $\supp{{\mathcal V}_p}\wedge \supp{{\mathcal V}_{p'}} = 0$ under the lattice operation of ${\mathscr R}(X)$. Since in this lattice $\wedge \neq \cap$, we cannot conclude that $\supp{{\mathcal V}_p}\cap \supp{{\mathcal V}_{p'}} = \varnothing$. More generally, since $\supp{{\mathsf B}(\ell)}\colon {\mathsf O}({\mathsf P})\to {\mathscr R}(X)$ is a lattice homomorphism $\supp{\ell(\gamma)} \wedge \supp{\ell(\alpha)} = \supp{\ell(\gamma\cap \alpha)}$, but this does not imply that $\supp{\ell(\gamma)} \cap \supp{\ell(\alpha)} = \supp{\ell(\gamma\cap \alpha)}$ since $\supp{\cdot}\colon {\mathsf{ Invset}}^-({\mathcal X},{\mathcal F})\to {\mathsf{ RNbhd}}(X,f)$ is not a lattice homomorphism. It is a homomorphism if we replace ${\mathsf{ RNbhd}}(X,f)$ by ${\mathsf{ RNbhd}}_{\scrR}(X,f)$. In order to obtain results that hold in ${\mathsf{ RNbhd}}(X,f)$, we introduce the following concept. \begin{definition} Let ${\mathsf O}({\mathsf P})$ be a finite distributive lattice. A lattice monomorphism $\ell\colon {\mathsf O}({\mathsf P}) \to {\mathsf{Set}}({\mathcal X})$ is {\em well-separated} if \begin{equation} \label{eqn:props-11} |{\mathcal V}_p| \cap |{\mathcal V}_{p'}| = \varnothing\quad\hbox{for all $p~\Vert~{p'}$},\quad p,p'\in{\mathsf P} \end{equation} where the $p~\Vert~{p'}$ indicates that $p$ and ${p'}$ are incomparable, i.e. $p\not\le {p'}$ and $p'\not \le {p}$. \end{definition} \begin{proposition} \label{prop:well-sep} If ${\mathsf O}({\mathsf P})$ is a finite distributive lattice and $\ell\colon {\mathsf O}({\mathsf P}) \to {\mathsf{Set}}({\mathcal X})$ is well-separated, then \begin{equation} \label{eqn:well-sep} \supp{\ell(\gamma)} \cap \supp{\ell(\alpha)} = \supp{\ell(\gamma)} \wedge \supp{\ell(\alpha)}. \end{equation} \end{proposition} \begin{proof} Observe that \[ \supp{\ell(\gamma)} \cap \supp{\ell(\alpha)}= \supp{ \bigcup_{p\in\gamma} {\mathcal V}_p } \cap \supp{ \bigcup_{q\in\alpha} {\mathcal V}_q } = \bigcup_{p\in\gamma\atop q\in\alpha} \left( \supp{{\mathcal V}_p} \cap \supp{{\mathcal V}_q} \right). \] By \eqref{eqn:props-11}, if $\supp{{\mathcal V}_p} \cap \supp{{\mathcal V}_q} \neq\varnothing$, then either $p\le q$ or $q\le p$. Since $\gamma$ and $\alpha$ are down sets, this implies that $p\in\gamma\cap\alpha$ or $q\in\gamma\cap\alpha$ respectively, and hence $\supp{{\mathcal V}_p} \cap \supp{{\mathcal V}_q} \subset \supp{{\mathcal V}_r}$ for some $r\in\gamma\cap\alpha$. Therefore \begin{eqnarray*} \supp{\ell(\gamma)} \cap \supp{\ell(\alpha)} &=& \bigcup_{p\in\gamma\atop q\in\alpha} \supp{{\mathcal V}_p} \cap \supp{{\mathcal V}_q} =\bigcup_{r\in\gamma\cap\alpha} \supp{{\mathcal V}_r} = \supp{ \bigcup_{r\in\gamma\cap\alpha}{\mathcal V}_r }\\ &=& \supp{\ell({\gamma\cap\alpha})} = \supp{\ell(\gamma)\cap \ell(\alpha)} = \supp{\ell(\gamma)} \wedge \supp{\ell(\alpha)} \end{eqnarray*} where the last two equalities follow from the fact that $\ell$ is a lattice homomorphism and Corollary~\ref{prop:grid->lattice}, respectively. \end{proof} Let $\lambda\in{\mathsf O}({\mathsf P})$ which is a subposet of ${\mathsf P}$. Note that $0\in{\mathsf O}(\lambda)$. However, if $\lambda\neq{\mathsf P}$, then ${\mathsf P}\notin{\mathsf O}(\lambda)$, and hence ${\mathsf O}(\lambda)$ is not a sublattice of ${\mathsf O}({\mathsf P})$. Therefore we define $\lambda^\top$ to be the poset $\lambda\cup\{\top\}$ where the additional top element $\top$ has relations $p\le\top$ for all $p\in\lambda$. Observe that as a set $\top ={\mathsf P}$. Then $$ {\mathsf O}(\lambda^\top)\approx\{\alpha\in{\mathsf O}({\mathsf P})~|~\alpha\subset\lambda\;\hbox{or}\;\alpha={\mathsf P}\} $$ making ${\mathsf O}(\lambda^\top)$ a sublattice of ${\mathsf O}({\mathsf P})$. Booleanization implies ${\mathsf B}({\mathsf O}(\lambda^\top))\subset{\mathsf B}({\mathsf O}({\mathsf P}))={\mathsf{Set}}({\mathsf P})$. \begin{definition} \label{defn:part-lift} Let $\lambda\in {\mathsf O}({\mathsf P})$. A lattice homomorphism $\ell:{\mathsf O}(\lambda^\top) \to {\mathsf H}$ is a {\em partial lift of $g$ on ${\mathsf O}(\lambda^\top)$} in Diagram \eqref{diag:lift12} if \[ h(\ell(\beta)) = g(\beta)\;\;\hbox{for all $\beta \leq \lambda$.} \] \end{definition} Note that by the above definition $\ell(1) = 1$, since $\ell$ is a lattice homomorphism. \subsection{Realization of attractor and repeller lattices} \label{subset:lift-grid} In this section, using ideas from \cite{KMV-1a}, we prove that the lattice structures can be realized via multivalued maps. \begin{theorem} \label{thm:conv-latt-1} Let $f\colon X \to X$ be a continuous mapping on a compact metric space $X$. Let ${\mathcal F}_n\colon {\mathcal X}_n \raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$} {\mathcal X}_n$ be a convergent cofiltration of outer approximations for $f$ defined on a contracting cofiltration of grids on $X$. If ${\mathsf A} \subset {\mathsf{ Att}}(X,f)$ is a finite sublattice, then there exists an $n_{\mathsf A}$ such that for all $n\ge n_{\mathsf A}$ there exists a lift $\ell_n\colon {\mathsf A} \to {\mathsf{ Invset}}^+({\mathcal X}_n,{\mathcal F}_n)$ of the inclusion map $\iota\colon {\mathsf A} \rightarrowtail {\mathsf{ Att}}(X,f)$ through $\omega(\supp{\cdot})\colon {\mathsf{ Invset}}^+({\mathcal X}_n,{\mathcal F}_n) \to {\mathsf{ Att}}(X,f)$, i.e.\ the following diagram commutes \[ \begin{diagram} \node{~} \node{{\mathsf{ Invset}}^+({\mathcal X}_n,{\mathcal F}_n)} \arrow{s,r}{\omega(|\cdot|)}\\ \node{{\mathsf A}} \arrow{e,l,V}{\imath} \arrow{ne,l,..,V}{\ell_n} \node{{\mathsf{ Att}}(X,f).} \end{diagram} \] Furthermore, $\ell_n$ can be chosen such that $\supp{\ell_n({\mathsf A})}$ is a sublattice of ${\mathsf{ ANbhd}}(X,f)$. \end{theorem} We do not know of a direct proof of Theorem~\ref{thm:conv-latt-1}. The difficulty arises from the fact that $\wedge = \cap$ for the lattice ${\mathsf{ Invset}}^+({\mathcal X}_n,{\mathcal F}_n)$, but $\wedge\neq \cap$ for the lattice ${\mathsf{ Att}}(X,f)$. Recall, however, that $\wedge = \cap$ for the lattice ${\mathsf{ Rep}}(X,f)$. With this in mind we prove the following analogous theorem for repellers. By the proof of \cite[Theorem 1.2]{KMV-1a} and in particular \cite[commutative diagram (24)]{KMV-1a}, the Theorem~\ref{thm:conv-latt-2} for repellers implies Theorem~\ref{thm:conv-latt-1} for attractors by a duality argument. \begin{theorem} \label{thm:conv-latt-2} Let $f\colon X \to X$ be a continuous mapping on a compact metric space $X$. Let ${\mathcal F}_n\colon {\mathcal X}_n \raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$} {\mathcal X}_n$ be a convergent cofiltration of outer approximations for $f$ defined on a contracting cofiltration of grids on $X$. If ${\mathsf R} \subset {\mathsf{ Rep}}(X,f)$ is a finite sublattice, then there exists an $n_{\mathsf R}$ such that for all $n\ge n_{\mathsf R}$ there exists a lift $\ell_n\colon {\mathsf R} \to {\mathsf{ Invset}}^-({\mathcal X}_n,{\mathcal F}_n)$ of the inclusion map $\iota\colon {\mathsf R} \rightarrowtail {\mathsf{ Rep}}(X,f)$ through $\alpha(\supp{\cdot})\colon {\mathsf{ Invset}}^-({\mathcal X}_n,{\mathcal F}_n) \twoheadrightarrow {\mathsf{ Rep}}(X,f)$, i.e.\ the following diagram commutes \begin{equation} \label{eq:repellerliftdiagram} \begin{diagram} \node{~} \node{{\mathsf{ Invset}}^-({\mathcal X}_n,{\mathcal F}_n)} \arrow{s,r}{\alpha(|\cdot|)}\\ \node{{\mathsf R}} \arrow{e,l,V}{\imath} \arrow{ne,l,..,V}{\ell_n} \node{{\mathsf{ Rep}}(X,f).} \end{diagram} \end{equation} Furthermore, $\ell_n$ can be chosen such that $\supp{\ell_n({\mathsf R})}$ is a sublattice of ${\mathsf{ RNbhd}}(X,f)$. \end{theorem} \begin{proof} Observe that by Remark~\ref{rem:birkhoff} to prove \eqref{eq:repellerliftdiagram} it is sufficient to prove the existence of ${\mathcal U}_n$ such that the following diagram commutes \begin{equation} \label{eq:diagramproof} \begin{diagram} \node{~} \node{{\mathsf{ Invset}}^-({\mathcal X}_n,{\mathcal F}_n)} \arrow{s,r,A}{\alpha(|\cdot|)} \arrow{e,l,J}{}\node{{\mathsf{Set}}({\mathcal X}_n)}\\ \node{{\mathsf O}({\mathsf P})} \arrow{ne,l,..,V}{{\mathcal U}_n} \arrow{e,l,V}{R_n} \node{{\mathsf R}_n} \arrow{e,J}\node{{\mathsf{ Rep}}(X,f)} \end{diagram} \end{equation} where ${\mathsf R}_n := \alpha\left(\supp{{\mathsf{ Invset}}^-({\mathcal X}_n,{\mathcal F}_n)} \right)$. Viewing ${\mathcal U}_n$ as a map into ${\mathsf{Set}}({\mathcal X}_n)$, by Proposition~\ref{prop:Boolappl} we obtain a grid $\setof{\supp{{\mathcal V}_{n,p}}\mid p\in {\mathsf P}}$ for $X$. Observe that if $\alpha, \alpha'\in {\mathsf O}({\mathsf P})$ satisfy $\alpha\cap\alpha' = \varnothing$, then $R_n(\alpha)\cap R_n(\alpha')=\varnothing$, and hence there exists $d_\varnothing >0$ such that \begin{equation} \label{eq:sepRep} B_{d_\varnothing}(R_n(\alpha)) \cap B_{d_\varnothing}(R_n(\alpha')) = \varnothing. \end{equation} Since ${\mathsf O}({\mathsf P})$ is finite, we can choose $d_\varnothing >0$ such that \eqref{eq:sepRep} is satisfied for all $\alpha, \alpha'\in {\mathsf O}({\mathsf P})$ satisfying $\alpha\cap\alpha' = \varnothing$. By Corollary~\ref{cor:resolution1} there exists $n_{d_\varnothing}>0$ such that for each $\alpha\in {\mathsf O}({\mathsf P})$ there is an associated discrete repeller ${\mathcal R}_n(\alpha)\in {\mathsf{ Invset}}^-({\mathcal X}_n,{\mathcal F}_n)$ satisfying \[ R_n(\alpha) \subset |{\mathcal R}_n(\alpha)|\le B_{d_\varnothing}(R_n(\alpha)) \] for all $n\ge n_{d_\varnothing}$. Having established the necessary notation we provide a proof by induction making use of partial lifts. To establish the initial induction step choose $n\ge n_{d_\varnothing}$ and let $q\in {\mathsf P}$ be minimal. Observe that $\setof{q}\in {\mathsf O}({\mathsf P})$. Define \[ {\mathcal U}_n(\setof{q}) := {\mathcal R}_n(\setof{q}). \] Observe that ${\mathcal U}_n\colon {\mathsf O}(\setof{q}^\top)\to {\mathsf{ Invset}}^-({\mathcal X}_n,{\mathcal F}_n)$ defines a partial lift of $R_n$ through $\alpha(\supp{\cdot})$ that satisfies the following three conditions: \begin{description} \item[C1] ${\mathcal V}_{n,p}\cap{\mathcal V}_{n,p'} = \varnothing$ for all $p\not = {p'}$; \item[C2] $\supp{{\mathcal V}_{n,p}} \cap R_n(\alpha) = \varnothing$ if $p \not\in\alpha$; \item[C3] $\supp{{\mathcal V}_{n,p}} \cap \supp{{\mathcal V}_{n,p'}} = \varnothing$ for all $p~\Vert~{p'}$, \end{description} where $p,p'\in \setof{q}, \alpha\in{\mathsf O}({\mathsf P}),$ and ${\mathcal V}_{n,p}$ is defined by \eqref{eq:Vp}. Assume that for some $\lambda \in {\mathsf O}({\mathsf P})$ and some $n_\lambda\ge n_{d_\varnothing}$ there exists a partial lift ${\mathcal U}_{n_\lambda}\colon {\mathsf O}(\lambda^\top)\to {\mathsf{ Invset}}^-({\mathcal X}_{n_\lambda},{\mathcal F}_n)$ of $R_n$ through $\alpha(\supp{\cdot})$ which satisfies Conditions {\bf C1} - {\bf C3} for $p,p'\in \lambda$ and $\alpha\in{\mathsf O}({\mathsf P}).$ Furthermore, given ${\mathcal U}_{n_\lambda}$ \eqref{eq:Vp} defines $\setof{{\mathcal V}_{n_\lambda,p}\mid p\in {\mathsf P}}$. We now show that for $n$ sufficiently large a new partial lift can be constructed on a down set in ${\mathsf O}({\mathsf P})$ with one additional element. Let $q\in {\mathsf P}\setminus \lambda$ be minimal. Define $\mu = \downarrow\! q$. By condition {\bf C2}, \begin{enumerate} \item [(i)] if $p\in \lambda\setminus \mu$ then $\supp{{\mathcal V}_{n_\lambda,p}} \cap R_{n_\lambda}(\mu) = \varnothing$. \end{enumerate} Since $R_{n_\lambda}$ is a lattice homomorphism, and ${\mathcal U}_{n_\lambda}$ is a partial lift, \begin{enumerate} \item [(ii)] if $\mu\cap \alpha\subset \lambda$, i.e.\ if $q\not\in\alpha\in{\mathsf O}({\mathsf P})$, then \[ R_{n_\lambda}(\mu) \cap R_{n_\lambda}(\alpha) = R_{n_\lambda}({\mu\cap \alpha})\subset R_{n_\lambda}(\lambda) \subset \mbox{\rm int\,} |{\mathcal U}_{n_\lambda}(\lambda)|. \] \end{enumerate} Property (i) implies that there exists a $d_\lambda>0$ such that $\supp{{\mathcal V}_{n_\lambda,p}} \cap B_{d_\lambda}(R_{n_\lambda}(\mu)) =\varnothing$ for all $p\in \lambda\setminus \mu$. Property (ii) is equivalent to $\bigl( R_{n_\lambda}(\mu) \setminus \mbox{\rm int\,} |{\mathcal U}_{n_\lambda}(\lambda)|\bigr) \cap R_{n_\lambda}(\alpha) = \varnothing$, which implies that if $q\not\in\alpha$, then ${\rm cl}\,}%{\text{cl}\bigl( R_{n_\lambda}(\mu)\setminus |{\mathcal U}_{n_\lambda}(\lambda)|\bigr) \cap R_{n_\lambda}(\alpha) = \varnothing$. We can therefore choose $d_\lambda$ small enough such that \begin{enumerate} \item [(i)'] if $p\in \lambda\setminus \mu$, then $\supp{{\mathcal V}_{n_\lambda,p}} \cap B_{d_\lambda}(R_{n_\lambda}(\mu)) = \varnothing$, and \item [(ii)'] if $q\not\in\alpha\in{\mathsf O}({\mathsf P})$, then ${\rm cl}\,}%{\text{cl} \bigl(B_{d_\lambda}(R_{n_\lambda}(\mu)) \setminus \supp{{\mathcal U}_{n_\lambda}(\lambda)}\bigr) \cap R_{n_\lambda}(\alpha) = \varnothing$. \end{enumerate} Observe that throughout this discussion we have been working with $n = n_\lambda$ and thus the fixed evaluation map $\supp{\cdot} = \supp{\cdot}_{n_\lambda} \colon {\mathcal X}_{n_\lambda} \to {\mathscr R}(X)$. Now we must change $n$, and thus the evaluation map also changes. Whenever the choice of evaluation map is clear, we continue to denote it by $\supp{\cdot}$. By Corollary~\ref{cor:resolution1} we can choose $n_{d_\lambda} \ge n_\lambda$ such that for any $n\ge n_{d_\lambda}$ the repeller ${\mathcal R}_{n}(\mu)$ guaranteed by this corollary satisfies $R_n(\mu) \subset \supp{{\mathcal R}_{n}(\mu)}_n \subset B_{d_\lambda}(R_n(\mu))$ . Observe that this in turn implies that for all $n\ge n_{d_\lambda}$ \begin{enumerate} \item [(i)"] if $p\in \lambda\setminus \mu$, then $\supp{{\mathcal V}_{n_\lambda,p}}_{n_\lambda} \cap \supp{{\mathcal R}_n(\mu)}_n = \varnothing$, and \item [(ii)"] if $q\not\in\alpha\in{\mathsf O}({\mathsf P})$, then ${\rm cl}\,}%{\text{cl} \bigl(\supp{{\mathcal R}_n(\mu)}_n \setminus \supp{{\mathcal U}_{n_\lambda}(\lambda)}_{n_\lambda} \bigr) \cap R_n(\alpha) = \varnothing$. \end{enumerate} Recall that for $n\ge n_{\lambda}$, ${\mathcal X}_n$ is a refinement of ${\mathcal X}_{n_\lambda}$. Thus for each $n$ there exists unique sets ${\mathcal V}^n_{n_\lambda,p}, {\mathcal U}^n_{n_\lambda}(\alpha) \subset {\mathcal X}_{n}$ such that \begin{equation} \label{eq:defnVU} \supp{{\mathcal V}^n_{n_\lambda,p}}_n = \supp{{\mathcal V}_{n_\lambda,p}}_{n_\lambda} \quad\text{and}\quad \supp{{\mathcal U}^n_{n_\lambda}(\alpha)}_n = \supp{{\mathcal U}_{n_\lambda}(\alpha)}_{n_\lambda}. \end{equation} By assumption ${\mathcal U}_{n_\lambda}(\lambda)\in {\mathsf{ Invset}}^-({\mathcal X}_{n_\lambda},{\mathcal F}_{n_\lambda})$. By Proposition~\ref{prop:cofiltmap} ${\mathcal U}^n_{n_\lambda}(\lambda)\in {\mathsf{ Invset}}^-({\mathcal X}_{n},{\mathcal F}_n)$. Let $n_{\mu\cup\lambda}\ge n_{d_\lambda}$. Then ${\mathcal U}^{n_{\mu\cup\lambda}}_{n_\lambda}\colon {\mathsf O}(\lambda^\top)\to {\mathsf{ Invset}}^{-}({\mathcal X}_{n_{\mu\cup\lambda}},{\mathcal F}_{n_{\mu\cup\lambda}})$ is a partial lift of $R_{n_{\mu\cup\lambda}}$ through $\alpha(\supp{\cdot})\colon {\mathsf{ Invset}}^{-}({\mathcal X}_{n_{\mu\cup\lambda}},{\mathcal F}_{n_{\mu\cup\lambda}})\to {\mathsf R}_{n_{\mu\cup\lambda}}$. To complete the induction step we must show that this partial lift can be extended to ${\mathsf O}((\lambda\cup\mu)^\top)$. \vskip 6pt \noindent {\em Claim:} This partial lift can be extended to ${\mathsf O}((\lambda\cup\mu)^\top)$ via the following definition. \vskip 6pt\noindent Given $\alpha\in\lambda\cup\mu$ define ${\mathcal U}_{n_{\mu\cup\lambda}}\colon {\mathsf O}((\lambda\cup\mu)^\top) \to {\mathsf{ Invset}}^-({\mathcal X}_{n_{\mu\cup\lambda}},{\mathcal F}_{n_{\mu\cup\lambda}})$ by \begin{equation}\label{eqn:liftk} {\mathcal U}_{n_{\mu\cup\lambda}}(\alpha)=\bigcup_{p\in\alpha}{\mathcal V}_{n_{\mu\cup\lambda},p} \end{equation} where \begin{equation} \label{eqn:defnext} {\mathcal V}_{n_{\mu\cup\lambda},q} := {\mathcal R}_{n_{\mu\cup\lambda}}(\mu)\setminus {\mathcal U}^{n_{\mu\cup\lambda}}_{n_\lambda}(\lambda) \quad\text{and}\quad {\mathcal V}_{n_{\mu\cup\lambda},p} := {\mathcal V}^{n_{\mu\cup\lambda}}_{n_\lambda,p}\ \text{for}\ p\in\lambda. \end{equation} Our induction hypothesis and the proof of the claim makes use of conditions {\bf C1} - {\bf C3}, thus we begin by verifying that they are satisfied. By the induction hypothesis to prove {\bf C1} for all $p,p' \subset \lambda \cup \mu$ it is sufficient to show that ${\mathcal V}_{n_{\mu\cup\lambda},q} \cap {\mathcal V}_{n_{\mu\cup\lambda},p} = \varnothing$ for $p\in\lambda$. This follows from the fact that \[ {\mathcal V}_{n_{\mu\cup\lambda},p} = {\mathcal V}^{n_{\mu\cup\lambda}}_{n_\lambda,p} \subset {\mathcal U}^{n_{\mu\cup\lambda}}_{n_\lambda}(\lambda) \] for all $p\in \lambda$. To prove Condition {\bf C2} observe that by definition \[ \supp{{\mathcal V}_{n_{\mu\cup\lambda},q}} = \supp{{\mathcal R}_{n_{\mu\cup\lambda}}(\mu)\setminus {\mathcal U}^{n_{\mu\cup\lambda}}_{n_\lambda}(\lambda)} = {\rm cl}\,}%{\text{cl} \bigl( \supp{{\mathcal R}_{n_{\mu\cup\lambda}}(\mu)} \setminus |{\mathcal U}_{n_{\mu\cup\lambda}}(\lambda)|\bigr) \] where the latter equality follows from Lemma~\ref{lem:intregdisj}, and then apply (ii)''. Turning to Condition {\bf C3}, by definition ${\mathcal V}_{n_{\mu\cup\lambda},q} \subset {\mathcal R}_{n_{\mu\cup\lambda}}(\mu)$ and thus by (i)", \eqref{eq:defnVU}, and \eqref{eqn:defnext} we have that $\supp{{\mathcal V}_{n_{\mu\cup\lambda},p}}\cap \supp{{\mathcal V}_{n_\lambda,q}} = \varnothing$ for all $p\in \lambda \setminus \mu$. Let $p\in \lambda\cup \mu$. Note that $p\le q$ if and only if $p\in \mu$, and thus $p\not \le q$ if and only if $p\in (\lambda\cup \mu)\setminus \mu$. Moreover, $q\le p$ if and only if $p=q$, and thus $q\not \le p$ if and only if $p\not = q$. We conclude that $p~\Vert~q$ if and only if $p\in \lambda \setminus \mu$, which proves {\bf C3} for all $p,p'\in \lambda \cup \mu$ satisfying $p~\Vert~p'$. To prove the claim we need to verify four statements: \begin{enumerate} \item ${\mathcal U}_{n_{\mu\cup\lambda}}$ is an extension of ${\mathcal U}_{n_\lambda}$, \item ${\mathcal U}_{n_{\mu\cup\lambda}}$ maps into ${\mathsf{ Invset}}^{-}({\mathcal X}_{n_{\mu\cup\lambda}},{\mathcal F}_{n_{\mu\cup\lambda}})$, \item ${\mathcal U}_{n_{\mu\cup\lambda}}$ is a lattice homomorphism, and \item ${\mathcal U}_{n_{\mu\cup\lambda}}$ is a partial lift of $R_{n_{\mu\cup\lambda}}$. \end{enumerate} To prove the first statement observe that each $\alpha\subset\lambda\cup\mu$ can be expressed as $\alpha=\beta\cup\nu$ for $\beta\subset\lambda$ and $\nu=\varnothing$ or $\nu=\mu$. If $\nu=\varnothing$, then ${\mathcal U}_{n_{\mu\cup\lambda}}(\alpha)={\mathcal U}_{n_\lambda}(\beta)$, and if $\nu=\mu$, then ${\mathcal U}_{n_{\mu\cup\lambda}}(\alpha)={\mathcal U}_{n_\lambda}(\beta) \cup {\mathcal U}_{n_{\mu\cup\lambda}}(\mu)$. The second statement follows from the fact that $\vee = \cup $ as the lattice operation in ${\mathsf{ Invset}}^{-}({\mathcal X}_{n_{\mu\cup\lambda}},{\mathcal F}_{n_{\mu\cup\lambda}})$. The third statement follows from \eqref{eqn:liftk} and {\bf C1}. See \cite[Theorem 4.8 {\em Proof of (a)}]{KMV-1a} for details. To demonstrate the fourth statement, note that \begin{eqnarray*} {\mathcal U}_{n_{\mu\cup\lambda}}(\mu) &=& {\mathcal V}_{n_{\mu\cup\lambda},q}\cup \left( \bigcup_{p\in\lambda\cap\mu}{\mathcal V}_{n_{\mu\cup\lambda},p}\right) \\ &=& \left( {\mathcal R}_{n_{\mu\cup\lambda}}(\mu)\setminus \left( \bigcup_{p\in\lambda}{\mathcal V}_{n_{\mu\cup\lambda},p} \right) \right) \cup \left( \bigcup_{p\in\lambda\cap\mu}{\mathcal V}_{n_{\mu\cup\lambda},p} \right)\\ &=&\left( {\mathcal R}_{n_{\mu\cup\lambda}}(\mu)\setminus \left( \bigcup_{p\in\lambda\cap\mu}{\mathcal V}_{n_{\mu\cup\lambda},p} \right) \right)\cup\left( \bigcup_{p\in\lambda\cap\mu}{\mathcal V}_{n_{\mu\cup\lambda},p} \right)\\ &=&{\mathcal R}_{n_{\mu\cup\lambda}}(\mu)\cup \left( \bigcup_{p\in\lambda\cap\mu}{\mathcal V}_{n_{\mu\cup\lambda},p} \right) \\ &=& {\mathcal R}_{n_{\mu\cup\lambda}}(\mu)\cup {\mathcal U}_{n_{\mu\cup\lambda}}({\lambda\cap\mu}). \end{eqnarray*} Therefore, \begin{eqnarray*} \alpha(\supp{{\mathcal U}_{n_{\mu\cup\lambda}}(\mu)}) &=& \alpha( \supp{{\mathcal R}_{n_{\mu\cup\lambda}}(\mu)}) \cup \alpha(\supp{{\mathcal U}_{n_{\mu\cup\lambda}}({\lambda\cap\mu})})\\ &=&R_{n_{\mu\cup\lambda}}(\mu)\cup R_{n_{\mu\cup\lambda}}({\lambda\cap\mu})=R_{n_{\mu\cup\lambda}}(\mu), \end{eqnarray*} and thus ${\mathcal U}_{n_{\mu\cup\lambda}}$ is a partial lift of $R_{n_{\mu\cup\lambda}}$. We have now proved the claim and completed the induction step. Since the lattice of repellers ${\mathsf R}$ is finite, a finite application of the induction argument gives rise to the commutative diagram \eqref{eq:diagramproof} and hence diagram \eqref{eq:repellerliftdiagram}. The commutative diagram \eqref{diag:comm-liftfull} guarantees that $\supp{\ell_n(\cdot)}\colon {\mathsf R} \to {\mathsf{ RNbhd}}_{\scrR}(X,f)$ or equivalently that $\supp{\ell_n}$ can be viewed as a lift of the embedding ${\mathsf R} \to {\mathsf{ Rep}}(X,f)$ through $\alpha\colon {\mathsf{ RNbhd}}_{\scrR}(X,f) \to {\mathsf{ Rep}}(X,f)$. The careful reader will note that Condition {\bf C3} has not yet been used. {\bf C3} implies that ${\mathcal U}_n$ is a well-separated lift. The definition of well-separated guarantees that $\supp{\ell_n}({\mathsf R})$ is a sublattice in ${\mathsf{ RNbhd}}(X,f)$, which is essential for the final claim of the theorem. Observe that this implies that $\supp{\ell_n}$ can be viewed as a lift of the embedding ${\mathsf R} \to {\mathsf{ Rep}}(X,f)$ through $\alpha\colon {\mathsf{ RNbhd}}(X,f) \to {\mathsf{ Rep}}(X,f)$. \end{proof} Theorem~\ref{thm:conv-latt-1} indicates that given a convergent cofiltration of multivalued maps obtained by refinement that the lattice structure of attractors can be realized as a lift to ${\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})$. The following theorem shows that a similar result holds if one works with an arbitrary convergent sequence of multivalued maps. Note that since ${\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})\subset {\mathsf{ASet}}({\mathcal X},{\mathcal F})$, the conclusion of this theorem is weaker than that of Theorem~\ref{thm:conv-latt-1}. \begin{theorem} \label{thm:conv-latt-3} Let $f\colon X \to X$ be a continuous mapping on a compact metric space $X$. Let ${\mathcal F}_n\colon {\mathcal X}_n \raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$} {\mathcal X}_n$ be a convergent sequence of outer approximations. Then for every finite sublattice ${\mathsf A} \subset {\mathsf{ Att}}(X,f)$ there exists an $n_{\mathsf A}$ such that for all $n\ge n_{\mathsf A}$ there exists a lift $\ell_n\colon {\mathsf A} \to {\mathsf{ASet}}({\mathcal X}_n,{\mathcal F}_n)$ of the inclusion map $\iota\colon {\mathsf A} \rightarrowtail {\mathsf{ Att}}(X,f)$ through $\omega(\supp{\cdot})\colon {\mathsf{ASet}}({\mathcal X}_n,{\mathcal F}_n) \to {\mathsf{ Att}}(X,f)$, i.e.\ the following diagram commutes \[ \begin{diagram} \node{~} \node{{\mathsf{ASet}}({\mathcal X}_n,{\mathcal F}_n)} \arrow{s,r}{\omega(|\cdot|)}\\ \node{{\mathsf A}} \arrow{e,l,V}{\imath} \arrow{ne,l,..,V}{\ell_n} \node{{\mathsf{ Att}}(X,f).} \end{diagram} \] Furthermore, $\ell_n$ can be chosen such that $\supp{\ell_n({\mathsf A})}$ is a sublattice of ${\mathsf{ ANbhd}}(X,f)$. Similar statements hold for finite sublattices ${\mathsf R} \subset {\mathsf{RSet}}(X,f)$. \end{theorem} The proof for Theorem~\ref{thm:conv-latt-3} is similar in spirit to that of Theorem~\ref{thm:conv-latt-2}. However, because we are not assuming a cofiltration of grids we cannot make use of Proposition~\ref{prop:cofiltmap}. We make use of the following abstract result to circumvent this difficulty. \begin{theorem}\label{thm:lift} Let $f\colon X \to X$ be a continuous mapping on a compact metric space $X$. Let ${\mathsf P}$ be a poset with $I$ elements. Let $R\colon{\mathsf O}({\mathsf P})\to{\mathsf R}\subset{\mathsf{ Rep}}(X,f)$ be a lattice isomorphism, and let $\pi:{\mathsf P}\to\{1,2,\ldots,I\}$ be a bijective, order-preserving map. Let $p_i :=\pi^{-1}(i)$ and $\mu_i :=\downarrow\! p_i \in {\mathsf O}({\mathsf P})$ for $i=\setof{1,\ldots,I }$. Then there exist $\setof{\epsilon_i}_{i=1}^I$ with $\epsilon_i>0$, such that if $\setof{N_i}_{i=1}^I$ is a collection of compact sets satisfying \begin{equation}\label{eqn:stable} B_{\epsilon_i/2}(R(\mu_i)) \subset N_{i} \subset B_{\epsilon_i}(R(\mu_i)), \end{equation} then each $N_i$ is a repelling neighborhood. Furthermore, $U\colon{\mathsf O}({\mathsf P})\to{\mathsf{ RNbhd}}(X,f)$, determined by $U(\mu_1)=N_1$ and $U(\mu_{i+1})=N_{i+1}\cup U(^\leftarrow{\mu_{i+1}})$, is a lift of $R$ through $\alpha \colon {\mathsf{ RNbhd}}(X,f) \to {\mathsf{ Rep}}(X,f)$. \end{theorem} \begin{proof} We use an inductive argument to prove the existence of $\{\epsilon_i\}_{i=1}^I$. Simultaneously we prove that at each stage of the induction argument the restriction of $U$ to ${\mathsf O}(\mu_i^\top)$, which we denote by $U_i$, is a partial lift of $R$ through $\alpha \colon {\mathsf{ RNbhd}}(X,f) \to {\mathsf{ Rep}}(X,f)$. More precisely, once $\epsilon_i$ is determined we choose a compact set $N_i\subset X$ satisfying \eqref{eqn:stable} at which point $U_i$ is well defined. Given $U_i$ define \[ V_{i,p} := U_i(\beta)\setminus U_i(\gamma)\quad \text{for }\beta\setminus\gamma = \setof{p} \] This definition is independent of the particular choice of $\beta$ and $\gamma$. See \eqref{eqn:inatom} and the associated discussion. Observe that \[ U_i(\alpha) = \bigcup_{p\in \alpha} V_{i,p}. \] Choose $\epsilon_1 >0$ such that $B_{\epsilon_1}(R(\mu_1)) \cap R(\mu_1)^* = \varnothing$. Choose $N_1$ satisfying \eqref{eqn:stable}. By \cite[Proposition 3.25]{KMV-1a} $N_1$ is a repelling neighborhood of $R(\mu_1)$. This defines $U_1$ on ${\mathsf O}(\mu_1^\top)={\mathsf O}(\setof{p_1}^\top)$. We leave it to the reader to check that the following three conditions (cf.\ the proof of Theorem~\ref{thm:conv-latt-2}) are trivially satisfied: \begin{description} \item[C1] $V_{k,p}\cap V_{k,p'} = \varnothing$ for all $p\not = {p'}$; \item[C2] ${\rm cl}\,}%{\text{cl}(V_{k,p}) \cap R(\alpha) = \varnothing$ if $p \not\in\alpha$; \item[C3] ${\rm cl}\,}%{\text{cl}(V_{k,p}) \cap {\rm cl}\,}%{\text{cl}(V_{k,p'}) = \varnothing$ for all $p~\Vert~{p'}$, \end{description} where $k=1$, $p,p'\in \setof{p_1},$ and $\alpha\in{\mathsf O}({\mathsf P})$. Since \[ \alpha(N_1,f) = \alpha(U_1(\setof{p_1}),f) = R(\setof{p_1}), \] $U_1$ is a partial lift of $R$ through $\alpha(\cdot,f)$. To carry out the induction argument, assume that $\{\epsilon_i\}_{i=1}^k$ and $\setof{N_i}_{i=1}^k$ have been chosen such that \eqref{eqn:stable} is satisfied and that the resulting lattice homomorphism $U_k$ defined on ${\mathsf O}(\lambda_k^\top)$, where $\lambda_k := \setof{p_1, \cdots p_k}$, is a partial lift of $R$ through $\alpha(\cdot,f)$ satisfying conditions {\bf C1} - {\bf C3} for $p,p'\in\lambda_k$ and $\alpha\in{\mathsf O}({\mathsf P}).$ Choose $\epsilon^0_{k+1}$ such that $B_{\epsilon^0_{k+1}}(R(\mu_{k+1}))\cap R(\mu_{k+1})^* =\varnothing$. By \cite[Proposition 3.25]{KMV-1a} if $R({\mu_{k+1}})\subset \mbox{\rm int\,}(D) \subset B_{\epsilon^0_{k+1}}(R(\mu_{k+1}))$, then $D$ is a repelling neighborhood for $R({\mu_{k+1}})$. We claim that there exists $\epsilon_{k+1}\in (0, \epsilon^0_{k+1})$ such that if $R({\mu_{k+1}})\subset \mbox{\rm int\,}(D) \subset B_{\epsilon_{k+1}}(R(\mu_{k+1}))$, then $D$ satisfies the following two conditions: \begin{enumerate} \item [(i)] if $p\in \lambda_k\setminus {\mu_{k+1}}$, then ${\rm cl}\,}%{\text{cl}(V_{k,p}) \cap D = \varnothing$, and \item [(ii)] if $p_{k+1}\not\in\alpha\in{\mathsf O}({\mathsf P})$, then ${\rm cl}\,}%{\text{cl} \left(D\setminus U_k(\lambda_k)\right) \cap R(\alpha) = \varnothing$. \end{enumerate} To establish (i) we use the induction hypothesis {\bf C2}, which implies that ${\rm cl}\,}%{\text{cl} (V_{k,p}) \cap R(\mu_{k+1}) = \varnothing$ for all $p\in {\lambda_k}\setminus {\mu_{k+1}}$. Since ${\rm cl}\,}%{\text{cl} (V_{k,p})$ and $R(\mu_{k+1})$ are compact, we can choose $\epsilon^1_{k+1}\in (0,\epsilon^0_{k+1})$ such that ${\rm cl}\,}%{\text{cl}(V_{k,p}) \cap D = \varnothing$ for all neighborhoods $D$ such that $R({\mu_{k+1}})\subset \mbox{\rm int\,}(D) \subset B_{\epsilon^1_{k+1}}(R(\mu_{k+1}))$. To establish (ii) note that the inclusion \begin{equation} \label{eqn:alt2} D \cap R(\alpha) \subset \mbox{\rm int\,} U_k(\lambda_k) \end{equation} is equivalent to \begin{equation} \label{eqn:alternative} (D\setminus \mbox{\rm int\,} U_k(\lambda_k)) \cap R(\alpha) = \varnothing. \end{equation} Observe that (\ref{eqn:alternative}) implies (ii) and thus it is sufficient to verify (\ref{eqn:alt2}) under the assumption that $p_{k+1}\not\in\alpha$. Since, $R(\mu_{k+1})\cap R(\alpha) = R({\mu_{k+1}\cap \alpha}) \subset R(\lambda_k)\subset \mbox{\rm int\,} U_k(\lambda_k)$, we can choose $\epsilon_{k+1}\in (0, \epsilon^1_{k+1})$ such that $R({\mu_{k+1}})\subset \mbox{\rm int\,}(D) \subset B_{\epsilon_{k+1}}(R(\mu_{k+1}))$ satisfies (\ref{eqn:alt2}). For the above choice of $\epsilon_{k+1}$ choose $N_{k+1}$ satisfying \eqref{eqn:stable}. This defines $U_{k+1}$. The proof that $U_{k+1}$ is a partial lift is identical in form to that of the proof of Theorem~\ref{thm:conv-latt-2} and thus left to the reader. The induction argument terminates after $I$ steps. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:conv-latt-3}] We first prove the result in the context of repellers, i.e.\ we prove the existence of a lift $\ell_n\colon {\mathsf R} \to {\mathsf{RSet}}({\mathcal X}_n,{\mathcal F}_n)$ of the inclusion map $\iota\colon {\mathsf R} \rightarrowtail {\mathsf{ Rep}}(X,f)$ through $\alpha(\supp{\cdot})\colon {\mathsf{RSet}}({\mathcal X}_n,{\mathcal F}_n) \to {\mathsf{ Rep}}(X,f)$. Let $\setof{\epsilon_i}_{i=1}^I$, be the set of radii produced by applying Theorem~\ref{thm:lift} with ${\mathsf P} = {\mathsf J}({\mathsf R})$. Choose $n$ sufficiently large so that ${\rm diam}\,({\mathcal X}_{n})<\min_{i}\{\epsilon_i/2\}$. For $i = 1,\ldots, I$ define ${\mathcal N}_i=\mathop{\rm cov}_{{\mathcal X}_n}(B_{\epsilon_i/2}(R(\mu_i)))$. By Proposition~\ref{prop:conv-latt-3}, if $n$ is chosen sufficiently large, then each ${\mathcal N}_i\in{\mathsf{RSet}}({\mathcal X}_n,{\mathcal F}_n)$. Similarly to Theorem~\ref{thm:lift}, the map $\ell_n\colon{\mathsf O}({\mathsf P})\to{\mathsf{RSet}}({\mathcal X}_n,{\mathcal F}_n)$ given by $$ \ell_n(\mu_1)={\mathcal N}_1\;\;\hbox{and}\;\; \ell_n(\mu_{i+1})={\mathcal N}_{i+1}\cup \ell_n(^\leftarrow{\mu_{i+1}}) $$ is a lift. By Theorem~\ref{thm:lift} $|\ell_n|\colon {\mathsf O}({\mathsf P})\to {\mathsf{ RNbhd}}(X,f)$ is a lift. The statement for attractors follows from duality, i.e.\ by the proof of \cite[Theorem 1.2]{KMV-1a} and in particular \cite[commutative diagram (24)]{KMV-1a}. \end{proof} \begin{remark}\label{rmk:mono} To put Theorems~\ref{thm:conv-latt-2} and~\ref{thm:conv-latt-3} into perspective, we recall that the monotonicity of images of ${\mathcal F}_n$ required for a cofiltration of mappings may in some applications be computationally expensive to attain, even though most practical algorithms construct ${\mathcal F}_n$ on a cofiltration of grids ${\mathcal X}_n$ through successive refinement. Theorem~\ref{thm:conv-latt-2} implies that if one does indeed compute a cofiltration of mappings, then the structure of attractors of $f$ can be realized in forward invariant sets of ${\mathcal F}$. Without a cofiltration, Theorem~\ref{thm:conv-latt-3} still implies the weaker result that the structure of attractors of $f$ can be realized in attracting sets of ${\mathcal F}$. \end{remark} \begin{remark} \label{rem:comm-lift11} By Proposition \ref{prop:discr-att-to-varphi3} and \ref{prop:discr-att-to-varphi4} we can restate the Diagram (\ref{diag:comm-liftfull}) by Diagram \eqref{diag:AR2a} for $\varphi$. For attractors this reads: \begin{equation*} \begin{diagram} \node{{\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})}\arrow{se,l,A}{{\bm{\omega}}}\arrow{e,l,V}{\imath}\node{{\mathsf{ASet}}({\mathcal X},{\mathcal F})}\arrow{e,l,V}{|\cdot|}\arrow{s,l,A}{{\bm{\omega}}}\node{{\mathsf{ ANbhd}}_{\scrR}(X,\varphi)}\arrow{e,l,A}{\omega}\node{{\mathsf{ Att}}(X,\varphi)}\\ \node{}\node{{\mathsf{ Att}}({\mathcal X},{\mathcal F})}\arrow{ene,r}{\omega(|\cdot|)} \end{diagram} \end{equation*} By Corollary \ref{for:sameA} we have that ${\mathsf{ Att}}(X,f) = {\mathsf{ Att}}(X,\varphi)$ as lattices, and therefore $\ell$ also provides a lift for $\varphi$ in Theorem~\ref{thm:conv-latt-1} and Theorem~\ref{thm:conv-latt-2}. In the case of Theorem~\ref{thm:conv-latt-1}, when we construct lifts under refinements, then the lift $\ell$ yields opposite arrows for all arrows in Diagram (\ref{diag:comm-lift11}). In the case of Theorem~\ref{thm:conv-latt-2} we only provide opposite arrows to ${\mathsf{ASet}}({\mathcal X},{\mathcal F})$, which implies that lifts to ${\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})$ may not exist. The same reasoning holds for repellers and their lifts. \end{remark} \section{Combinatorial systems} \label{sec:comb-syst} In this section we discuss the dynamics of combinatorial multivalued maps. We begin with basic properties, especially those related to the asymptotic dynamics. We then discuss attractors, repellers and the combinatorial equivalences of their neighborhoods. Finally we discuss the concept of attractor-repeller pairs in this combinatorial setting. \subsection{Combinatorial multivalued maps} Let ${\mathcal X}$ be a finite set of vertices. To emphasize the fact that we are interested in dynamics we denote mappings ${\mathcal F} \colon {\mathcal X} \to {\mathsf{Set}}({\mathcal X})$ by ${\mathcal F}\colon {\mathcal X} \raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$} {\mathcal X}$ and refer to them as \emph{combinatorial multivalued mappings} on ${\mathcal X}$. The inverse image of a element $\xi \in {\mathcal X}$ is defined by \begin{equation} \label{eqn:inverse-cms} {\mathcal F}^{-1}(\xi) := \{ \eta\in {\mathcal X}~|~ \xi\in {\mathcal F}(\eta)\}, \end{equation} which generates a combinatorial multivalued mapping denoted by ${\mathcal F}^{-1}\colon{\mathcal X}\raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$} {\mathcal X}$. \begin{definition} \label{defn:total} A multivalued mapping is \emph{left-total} if ${\mathcal F}(\xi)\neq\varnothing$ for all $\xi \in {\mathcal X}$ and \emph{right-total} if ${\mathcal F}^{-1}(\xi)\neq\varnothing$ for all $\xi \in {\mathcal X}$. A multivalued mapping is \emph{total} if it is both left- and right-total. \end{definition} The {\em ${\bm{\omega}}$-limit set} and {\em ${\bm{\alpha}}$-limit set} capture the asymptotic dynamics of a set ${\mathcal U}\subset {\mathcal X}$ under ${\mathcal F}\colon {\mathcal X}\raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$} {\mathcal X}$ and are defined by \[ {\bm{\omega}}({\mathcal U}) = \bigcap_{k\ge 0} \bigcup_{n\geq k} {\mathcal F}^n({\mathcal U}) \quad\hbox{and}\quad {\bm{\alpha}}({\mathcal U}) = \bigcap_{k\le 0} \bigcup_{n\leq k} {\mathcal F}^n({\mathcal U}), \] respectively. Observe that omega and alpha limit sets of nonempty sets may be empty, but they satisfy the following properties. \begin{proposition} \label{prop:combomega} Let ${\mathcal F}:{\mathcal X}\raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$}{\mathcal X}$ be a multivalued mapping and let ${\mathcal U}\subset {\mathcal X}$. Then, \begin{enumerate} \item[(i)] there exists a $k_*\ge 0$ such that ${\bm{\omega}}({\mathcal U}) = \bigcup_{n\geq k} {\mathcal F}^n({\mathcal U})$ for all $k\ge k_*$; \item[(ii)] ${\mathcal F}({\bm{\omega}}({\mathcal U})) = {\bm{\omega}}({\mathcal U})$ and ${\bm{\omega}}({\mathcal F}({\mathcal U})) = {\bm{\omega}}({\mathcal U})$, and thus ${\bm{\omega}}({\mathcal U}) \in {\mathsf{ Invset}}^+({\mathcal X})$; \item[(iii)] ${\mathcal F}$ left-total and ${\mathcal U} \not = \varnothing$ implies that ${\bm{\omega}}({\mathcal U})$ is invariant and ${\bm{\omega}}({\mathcal U})\not = \varnothing$; \item[(iv)] if there exists $k_*>0$ such that ${\mathcal F}^n({\mathcal U}) \subset {\mathcal U}$ for $k\ge k_*$, then ${\bm{\omega}}({\mathcal U}) \subset {\mathcal U}$; \item[(v)] ${\mathcal V}\subset {\mathcal U}$ implies ${\bm{\omega}}({\mathcal V}) \subset {\bm{\omega}}({\mathcal U})$, and in particular ${\bm{\omega}}({\mathcal V} \cap {\mathcal U}) \subset {\bm{\omega}}({\mathcal V}) \cap {\bm{\omega}}({\mathcal U})$; \item[(vi)] ${\bm{\omega}}({\mathcal V}\cup {\mathcal U}) = {\bm{\omega}}({\mathcal V}) \cup {\bm{\omega}}({\mathcal U})$, and in particular ${\bm{\omega}}({\mathcal U}) = \bigcup_{\xi\in {\mathcal U}} {\bm{\omega}}(\xi)$; \item[(vii)] ${\bm{\omega}}({\bm{\omega}}({\mathcal U})) = {\bm{\omega}}({\mathcal U})$. \end{enumerate} The same properties hold for ${\bm{\alpha}}$-limit sets via time-reversal, i.e. replace ${\mathcal F}$ by ${\mathcal F}^{-1}$. \end{proposition} \begin{proof} All properties can essentially be derived from Property (i), which we prove now. Forward images are nested sets. Since ${\mathcal X}$ is finite, it follows that there exists $k_*$ such that \[ \bigcup_{n\geq k} {\mathcal F}^n({\mathcal U}) = \bigcup_{n\geq k_*} {\mathcal F}^n({\mathcal U}) \] for all $k\geq k_*$. \end{proof} \subsection{Attractors and repellers} \label{sec:cds-AR} Alpha and omega limit sets capture the asymptotic dynamics of individual sets. Our goal for the remainder of this section is to understand the structure of the asymptotic dynamics of all sets. We begin with the concept of forward and backward invariance. A set ${\mathcal S}\subset {\mathcal X}$ is \emph{forward invariant} if ${\mathcal F}({\mathcal S}) \subset {\mathcal S}$ and it is \emph{backward invariant} if ${\mathcal F}^{-1}({\mathcal S}) \subset {\mathcal S}$. The sets of forward and backward invariant sets in ${\mathcal X}$ are denoted by ${\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})$ and ${\mathsf{ Invset}}^-({\mathcal X},{\mathcal F})$ respectively. \begin{proposition} \label{prop:cms-forback} The sets ${\mathsf{ Invset}}^-({\mathcal X},{\mathcal F})$ and ${\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})$ are finite distributive lattices with respect to intersection and union. The mapping ${\mathcal U}\mapsto {\mathcal U}^c$ is an involute lattice anti-isomorphism between ${\mathsf{ Invset}}^-({\mathcal X},{\mathcal F})$ and ${\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})$. \end{proposition} \begin{proof} We leave the proof that ${\mathsf{ Invset}}^-({\mathcal X},{\mathcal F})$ and ${\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})$ are finite distributive lattices with respect to intersection and union to the reader. To show that set complement maps ${\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})$ to ${\mathsf{ Invset}}^-({\mathcal X},{\mathcal F})$ consider ${\mathcal U}\in {\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})$ and $\xi\in {\mathcal U}^c$. Suppose $\eta \in {\mathcal F}^{-1}(\xi)\cap {\mathcal U}$. Then \[ \xi \in {\mathcal F}\bigl( \eta\bigr) \subset {\mathcal F}({\mathcal U}) \subset {\mathcal U}, \] which contradicts the fact that $\xi\in {\mathcal U}^c$. We conclude that ${\mathcal F}^{-1}({\mathcal U}^c) \subset {\mathcal U}^c$, and therefore ${\mathcal U}^c\in {\mathsf{ Invset}}^-({\mathcal X},{\mathcal F})$. The same arguments hold when ${\mathcal U}\in {\mathsf{ Invset}}^-({\mathcal X},{\mathcal F})$ is backward invariant. The fact that the map ${\mathcal U} \mapsto {\mathcal U}^c$ is a lattice anti-isomorphism follows from De Morgan's laws. \end{proof} To characterize the asymptotic dynamics of forward and backward invariant sets we make use of the following structures. \begin{definition} \label{defn:combattractor} Let ${\mathcal F}\colon{\mathcal X}\raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$}{\mathcal X}$ be a combinatorial multivalued mapping. A set ${\mathcal A} \subset {\mathcal X}$ is an \emph{attractor} for ${\mathcal F}$ if ${\mathcal F}({\mathcal A})={\mathcal A}$. A set ${\mathcal R}\subset {\mathcal X}$ is a \emph{repeller} for ${\mathcal F}$ if ${\mathcal F}^{-1}({\mathcal R}) = {\mathcal R}$. The sets of all attractors and repellers in ${\mathcal X}$ are denoted by ${\mathsf{ Att}}({\mathcal X},{\mathcal F})$ and ${\mathsf{ Rep}}({\mathcal X},{\mathcal F})$ respectively. \end{definition} For ${\mathcal A},{\mathcal A}' \in {\mathsf{ Att}}({\mathcal X},{\mathcal F})$ define \[ {\mathcal A}\vee{\mathcal A}' = {\mathcal A} \cup {\mathcal A}'\quad {\rm and}\quad{\mathcal A} \wedge {\mathcal A}' = {\bm{\omega}}({\mathcal A}\cap{\mathcal A}'). \] Similarly, for ${\mathcal R},{\mathcal R}' \in {\mathsf{ Rep}}({\mathcal X},{\mathcal F})$ define $$ {\mathcal R}\vee{\mathcal R}' = {\mathcal R} \cup {\mathcal R}'\quad {\rm and}\quad{\mathcal R} \wedge {\mathcal R}' = {\bm{\alpha}}({\mathcal R}\cap{\mathcal R}'). $$ \begin{proposition} \label{prop:cms-attrep} The sets $\bigl({\mathsf{ Att}}({\mathcal X},{\mathcal F}),\wedge,\vee\bigr)$ and $\bigl({\mathsf{ Rep}}({\mathcal X},{\mathcal F}),\wedge,\vee\bigr)$ are finite, distributive lattices. \end{proposition} \begin{proof} Let ${\mathcal A},{\mathcal A}'\in {\mathsf{ Att}}({\mathcal X},{\mathcal F})$ be attractors. Then ${\mathcal F}({\mathcal A} \cup {\mathcal A}') = {\mathcal F}({\mathcal A}) \cup {\mathcal F}({\mathcal A}') = {\mathcal A} \cup {\mathcal A}'$, and thus ${\mathcal A}\cup {\mathcal A}' \in {\mathsf{ Att}}({\mathcal X},{\mathcal F})$. Similarly, ${\mathcal A}\wedge {\mathcal A}' = {\bm{\omega}}({\mathcal A}\cap {\mathcal A}')$, and therefore ${\mathcal F}({\mathcal A}\wedge{\mathcal A}') = {\mathcal A}\wedge{\mathcal A}'$ by Proposition \ref{prop:combomega}(ii), which proves ${\mathcal A}\wedge {\mathcal A}' \in {\mathsf{ Att}}({\mathcal X},{\mathcal F})$. This proves that ${\mathsf{ Att}}({\mathcal X},{\mathcal F})$ is a lattice. The same holds for ${\mathsf{ Rep}}({\mathcal X},{\mathcal F})$. It remains to show that both sublattices are distributive. Let ${\mathcal A},{\mathcal A}'{\mathcal A}'' \in {\mathsf{ Att}}({\mathcal X},{\mathcal F})$. Then \begin{eqnarray*} ({\mathcal A}\wedge {\mathcal A}') \vee ({\mathcal A}\wedge {\mathcal A}'') & = & {\bm{\omega}}({\mathcal A}\cap {\mathcal A}') \cup {\bm{\omega}}({\mathcal A}\cap {\mathcal A}'')\\ &= & {\bm{\omega}}\bigl(({\mathcal A}\cap {\mathcal A}') \cup ({\mathcal A}\cap {\mathcal A}'')\bigr)\\ &=& {\bm{\omega}}\bigl({\mathcal A}\cap ({\mathcal A}'\cup {\mathcal A}'')\bigr) = {\bm{\omega}}({\mathcal A}) \wedge {\bm{\omega}}({\mathcal A}'\cup {\mathcal A}'')\\ &=& {\mathcal A}\wedge ({\mathcal A}'\vee {\mathcal A}''), \end{eqnarray*} which proves distributivity. The arguments for ${\mathsf{ Rep}}({\mathcal X},{\mathcal F})$ are symmetric. \end{proof} \begin{remark} Because the lattice operations are distinct, ${\mathsf{ Att}}({\mathcal X},{\mathcal F})$ and ${\mathsf{ Rep}}({\mathcal X},{\mathcal F})$ cannot be viewed as sublattices of ${\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})$ and ${\mathsf{ Invset}}^-({\mathcal X},{\mathcal F})$, respectively. \end{remark} By Proposition \ref{prop:combomega}(ii) every omega limit set is an attractor, and similarly, every alpha limit set is a repeller. The following result adds structure to this observation. \begin{proposition} \label{att-rep-fbi} The functions \[{\bm{\omega}}\colon {\mathsf{ Invset}}^+({\mathcal X},{\mathcal F}) \to {\mathsf{ Att}}({\mathcal X},{\mathcal F})\quad \hbox{ and}\quad {\bm{\alpha}}\colon {\mathsf{ Invset}}^-({\mathcal X},{\mathcal F}) \to {\mathsf{ Rep}}({\mathcal X},{\mathcal F}), \] are lattice epimomorphisms. \end{proposition} \begin{proof} We give the proof for ${\bm{\omega}}\colon {\mathsf{ Invset}}^+({\mathcal X},{\mathcal F}) \to {\mathsf{ Att}}({\mathcal X},{\mathcal F})$. From Proposition \ref{prop:combomega}(vi) it follows that ${\bm{\omega}}({\mathcal U}\cup {\mathcal U}') = {\bm{\omega}}({\mathcal U}) \cup {\bm{\omega}}({\mathcal U}')$. For the meet operation we argue as follows. Since ${\mathcal U}$ and ${\mathcal U}'$ are forward invariant, we have ${\mathcal A} = {\bm{\omega}}({\mathcal U})\subset {\mathcal U}$ and ${\mathcal A}' = {\bm{\omega}}({\mathcal U}')\subset {\mathcal U}'$. Then \begin{eqnarray*} {\mathcal A} \wedge {\mathcal A}' &=& {\bm{\omega}}({\mathcal A} \cap {\mathcal A}') \subset {\bm{\omega}}({\mathcal U}\cap {\mathcal U}') = {\bm{\omega}}\bigl( {\bm{\omega}}({\mathcal U}\cap {\mathcal U}')\bigr)\\ &\subset& {\bm{\omega}}\bigl( {\bm{\omega}}({\mathcal U})\cap {\bm{\omega}}({\mathcal U}')\bigr) = {\mathcal A}\wedge {\mathcal A}', \end{eqnarray*} which proves that ${\bm{\omega}}({\mathcal U}\cap {\mathcal U}') = {\mathcal A} \wedge {\mathcal A}' = {\bm{\omega}}({\mathcal U}) \wedge {\bm{\omega}}({\mathcal U}')$, and therefore ${\bm{\omega}}$ is a lattice homomorphism. The map is surjective, since ${\mathsf{ Att}}({\mathcal X},{\mathcal F}) \subset {\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})$ and ${\bm{\omega}}|_{{\mathsf{ Att}}({\mathcal X},{\mathcal F})} = \mathop{\rm id }\nolimits$. Moreover, ${\bm{\omega}}(\varnothing) = \varnothing = 0$ in ${\mathsf{ Att}}({\mathcal X},{\mathcal F})$, and ${\bm{\omega}}({\mathcal X},{\mathcal F}) = 1$ in ${\mathsf{ Att}}({\mathcal X},{\mathcal F})$. \end{proof} While the previous proposition demonstrates that all attractors and repellers can be obtained as the omega and alpha limit sets of a forward and backward invariant sets, there are larger collections of sets that lead to attractors and repellers. An \emph{attracting set} ${\mathcal U}$ has the property that ${\bm{\omega}}({\mathcal U}) \subset {\mathcal U}$, and the attracting sets are denoted by ${\mathsf{ASet}}({\mathcal X},{\mathcal F})$. Similarly, a \emph{repelling set} is defined by ${\bm{\alpha}}({\mathcal U}) \subset {\mathcal U}$, and the repelling sets are denoted by ${\mathsf{RSet}}({\mathcal X},{\mathcal F})$. Observe that forward and backward invariant sets are attracting and repelling sets, respectively, but not vice-versa. \begin{proposition} The sets ${\mathsf{ASet}}({\mathcal X},{\mathcal F})$ and ${\mathsf{RSet}}({\mathcal X},{\mathcal F})$ are finite sublattices of ${\mathsf{Set}}({\mathcal X},{\mathcal F})$ and therefore finite distributive lattices. \end{proposition} \begin{proof} The proof for ${\mathsf{ASet}}({\mathcal X},{\mathcal F})$ follows from the following containments \[ \begin{aligned} {\bm{\omega}}({\mathcal U}\cup{\mathcal U}') &= {\bm{\omega}}({\mathcal U}) \cup {\bm{\omega}}({\mathcal U}') \subset {\mathcal U}\cup {\mathcal U}',\\ {\bm{\omega}}({\mathcal U}\cap{\mathcal U}') &\subset {\bm{\omega}}({\mathcal U}) \cap {\bm{\omega}}({\mathcal U}') \subset {\mathcal U}\cap {\mathcal U}'. \end{aligned} \] The proof for ${\mathsf{RSet}}({\mathcal X},{\mathcal F})$ is similar. \end{proof} The same proof as that of Proposition~\ref{att-rep-fbi} leads to the following result. \begin{proposition} \label{att-rep-aset} The mappings ${\bm{\omega}}: {\mathsf{ASet}}({\mathcal X},{\mathcal F}) \to {\mathsf{ Att}}({\mathcal X},{\mathcal F})$ and ${\bm{\alpha}}: {\mathsf{RSet}}({\mathcal X},{\mathcal F}) \to {\mathsf{ Rep}}({\mathcal X},{\mathcal F})$ are lattice epimorphisms. \end{proposition} Proposition \ref{prop:cms-forback} establishes ${\mathcal U} \mapsto {\mathcal U}^c$ as a lattice anti-isomorphism between ${\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})$ and ${\mathsf{ Invset}}^-({\mathcal X},{\mathcal F})$. The result is true for attracting and repelling sets. To prove this we make use of the following result. \begin{proposition} \label{prop:char-attset} A set ${\mathcal U}$ is an attracting set if and only if there exists $k>0$ such that ${\mathcal F}^n({\mathcal U}) \subset {\mathcal U}$ for all $n\ge k$. Similarly, a set ${\mathcal U}$ is a repelling set if and only if there exists $k\ge 0$ such that ${\mathcal F}^{-n}({\mathcal U}) \subset {\mathcal U}$ for all $n\ge k$. \end{proposition} \begin{proof} If ${\mathcal U} \in {\mathsf{ASet}}({\mathcal X},{\mathcal F})$, then ${\bm{\omega}}({\mathcal U}) \subset {\mathcal U}$. By Proposition \ref{prop:combomega}(i), there exists $k>0$ such that ${\bm{\omega}}({\mathcal U}) = \Gamma^+_k({\mathcal U}) = \bigcup_{n\ge k} {\mathcal F}^n({\mathcal U}) \subset {\mathcal U}$, which implies that ${\mathcal F}^n({\mathcal U}) \subset {\mathcal U}$ for all $n\ge k$. Conversely, if there exists $k>0$ such that ${\mathcal F}^n({\mathcal U}) \subset {\mathcal U}$ for all $n\ge k$, then Proposition \ref{prop:combomega}(iv) implies that ${\bm{\omega}}({\mathcal U}) \subset {\mathcal U}$, which proves that ${\mathcal U}\in {\mathsf{ASet}}({\mathcal X},{\mathcal F})$. \end{proof} \begin{proposition} \label{prop:invo} The mapping ${\mathcal U}\mapsto {\mathcal U}^c$ is a lattice anti-isomorphism between ${\mathsf{ASet}}({\mathcal X},{\mathcal F})$ and ${\mathsf{RSet}}({\mathcal X},{\mathcal F})$. \end{proposition} \begin{proof} Let ${\mathcal U}\in {\mathsf{ASet}}({\mathcal X},{\mathcal F})$. Then by Proposition \ref{prop:char-attset}, there exists $k>0$ such that ${\mathcal F}^n({\mathcal U}) \subset {\mathcal U}$ for all $n\ge k$. As in the proof of Proposition \ref{prop:cms-forback}, assume $\xi\in {\mathcal U}^c$. Suppose that there exists $k>0$ such that $ {\mathcal F}^{-n}(\xi)\cap {\mathcal U}\neq \varnothing$ for all $n\ge k$. Let $\eta\in{\mathcal F}^{-n}(\xi)\cap {\mathcal U}$ for $n\ge k$. Then we have $$ \xi \in {\mathcal F}^n\bigl( \eta\bigr) \subset {\mathcal F}^n({\mathcal U}) \subset {\mathcal U}, $$ which contradicts the fact that $\xi\in {\mathcal U}^c$. We conclude that there exists $k>0$ such that ${\mathcal F}^{-n}({\mathcal U}^c) \subset {\mathcal U}^c$ for all $n\ge k$, and therefore ${\mathcal U}^c \in {\mathsf{RSet}}({\mathcal X},{\mathcal F})$. \end{proof} \subsection{Attractor-repeller pairs} \label{subsec:attrep-pairs} For a multivalued mapping ${\mathcal F}\colon {\mathcal X}\raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$}{\mathcal X}$ one can introduce the notions of dual repeller and dual attractor. \begin{definition} \label{defn:dual-cms} Let ${\mathcal F}\colon {\mathcal X}\raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$}{\mathcal X}$ be a multivalued mapping. The \emph{dual repeller} ${\mathcal A}^*$ to an attractor ${\mathcal A}$ is defined by ${\mathcal A}^* = {\bm{\alpha}}({\mathcal A}^c)$. Similarly the {\em dual attractor} to a repeller ${\mathcal R}$ is ${\mathcal R}^* = {\bm{\omega}}({\mathcal R}^c)$. The pairs $({\mathcal A},{\mathcal A}^*)$ and $({\mathcal R}^*,{\mathcal R})$ are called \emph{attractor-repellers pairs} in ${\mathcal X}$. \end{definition} It follows from Proposition \ref{prop:cms-forback} and Proposition \ref{prop:combomega}(vii) that if ${\mathcal A}\in{\mathsf{ Att}}({\mathcal X},{\mathcal F})$, then ${\mathcal A}^c \in {\mathsf{ Invset}}^-({\mathcal X},{\mathcal F})$ and thus ${\mathcal A}^* = {\bm{\alpha}}({\mathcal A}^c)\in{\mathsf{ Rep}}({\mathcal X},{\mathcal F})$. Similarly, if ${\mathcal R}\in{\mathsf{ Rep}}({\mathcal X},{\mathcal F})$, then ${\mathcal R}^* = {\bm{\omega}}({\mathcal R}^c)\in {\mathsf{ Att}}({\mathcal X},{\mathcal F})$. \begin{proposition} \label{prop:cms-star} Let $({\mathcal A},{\mathcal A}^*)$ be a attractor-repeller pair. Then, $$ {\mathcal A} = {\bm{\omega}}({{\mathcal A}^*}^c)\hbox{~and~} {\mathcal A}^* = {\bm{\alpha}}({\mathcal A}^c). $$ The operator ${\mathcal A} \mapsto {\mathcal A}^*$ is a lattice anti-isomorphism from ${\mathsf{ Att}}({\mathcal X},{\mathcal F})$ to ${\mathsf{ Rep}}({\mathcal X},{\mathcal F})$. \end{proposition} \begin{proof} The set ${\mathcal A}^{**} = {\bm{\omega}}\bigl(({\mathcal A}^*)^c\bigr)$ is forward-backward invariant, and hence for every $\xi \in {\mathcal A}^{**}$ we have ${\bm{\alpha}}(\xi)\subset {\mathcal A}^{**}$ and ${\bm{\omega}}(\xi)\subset {\mathcal A}^{**}$. Moreover, ${\mathcal A} = ({\mathcal A}^c)^c \subset ({\mathcal A}^*)^c$, and thus ${\mathcal A} = {\bm{\omega}}({\mathcal A}) \subset {\bm{\omega}}\bigl(({\mathcal A}^*)^c\bigr) = {\mathcal A}^{**}$. Let $\xi \in {\mathcal A}^{**}\setminus {\mathcal A} = {\mathcal A}^{**} \cap {\mathcal A}^c$. Since ${\mathcal A}^c$ is backward invariant, ${\bm{\alpha}}(\xi) \subset {\mathcal A}^c$, and thus ${\bm{\alpha}}(\xi) \in {\mathcal A}^{**} \setminus {\mathcal A}$. Also ${\bm{\alpha}}(\xi) \subset {\bm{\alpha}}({\mathcal A}^c) = {\mathcal A}^*$, which implies that ${\bm{\alpha}}(\xi) \in {\mathcal A}^{**} \cap {\mathcal A}^*$. The forward invariance of $({\mathcal A}^*)^c$ implies that ${\mathcal A}^{**}\cap {\mathcal A}^* = \varnothing$. We conclude that that ${\bm{\alpha}}(\xi) = \varnothing$ for all $\xi\in {\mathcal A}^{**} \setminus {\mathcal A}$. By definition ${\mathcal A}^{**}$ is an attractor, and therefore ${\mathcal F}^{-1}(\xi) \cap {\mathcal A}^{**} \not = \varnothing$ for all $\xi\in {\mathcal A}^{**}$, and consequently ${\bm{\alpha}}(\xi) \not = \varnothing$ for all $\xi \in {\mathcal A}^{**}$, a contradiction. This shows that ${\mathcal A}^{**} = {\mathcal A}$. Similar arguments also apply to repellers. The mapping ${\mathcal A}\mapsto {\mathcal A}^*$ is an involution, and the lattices ${\mathsf{ Att}}({\mathcal X},{\mathcal F})$ to ${\mathsf{ Rep}}({\mathcal X},{\mathcal F})$ are isomorphic. To show that ${\mathcal A} \mapsto {\mathcal A}^*$ is a lattice anti-isomorphism we argue as follows. Let ${\mathcal A}^* = {\bm{\alpha}}({\mathcal A}^c)$ and ${\mathcal A}'^* = {\bm{\alpha}}({{\mathcal A}'}^c)$, then by Proposition \ref{att-rep-fbi} and De Morgans' laws \begin{eqnarray*} ({\mathcal A} \vee {\mathcal A}')^* &=& {\bm{\alpha}}\bigl(({\mathcal A}\vee{\mathcal A}')^c\bigr) = {\bm{\alpha}}\bigl(({\mathcal A}\cup{\mathcal A}')^c\bigr)\\ &=& {\bm{\alpha}}\bigl({\mathcal A}^c\cap {{\mathcal A}'}^c\bigr) = {\bm{\alpha}}({\mathcal A}^c)\wedge {\bm{\alpha}}({{\mathcal A}'}^c)\\ &=& {\mathcal A}^* \wedge {\mathcal A}'^*. \end{eqnarray*} The same holds for ${\mathcal A}^*$ and ${\mathcal A}'^*$, i.e. $({\mathcal A}^* \vee {\mathcal A}'^*)^* = {\mathcal A}^{**} \wedge {\mathcal A}'^{**} = {\mathcal A} \wedge {\mathcal A}'$. Observe that $$ ({\mathcal A} \wedge {\mathcal A}')^* = ({\mathcal A}^* \vee {\mathcal A}'^*)^{**} = {\mathcal A}^* \vee {\mathcal A}'^*, $$ which proves the proposition. \end{proof} Much of the discussion of this section up to this point can be summarized in the following commutative diagram of lattice homomorphisms. \begin{equation} \label{eq:diag+-} \begin{diagram} \node{{\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})}\arrow{se,l,V}{\imath}\arrow{sse,l,A}{{\bm{\omega}}} \arrow[3]{e,l,<>}{^c} \node{}\node{} \node{{\mathsf{ Invset}}^-({\mathcal X},{\mathcal F})}\arrow{ssw,r,A}{{\bm{\alpha}}}\arrow{sw,l,V}{\imath}\\ \node{}\node{{\mathsf{ASet}}({\mathcal X},{\mathcal F})} \arrow{e,l,<>}{^c} \arrow{s,l,A}{{\bm{\omega}}}\node{{\mathsf{RSet}}({\mathcal X},{\mathcal F})} \arrow{s,r,A}{{\bm{\alpha}}}\\ \node{}\node{{\mathsf{ Att}}({\mathcal X},{\mathcal F})} \arrow{e,l,<>}{^*} \node{{\mathsf{ Rep}}({\mathcal X},{\mathcal F}).} \end{diagram} \end{equation} \section{Introduction} \label{sec:intro} The issue of computability in the context of nonlinear dynamics has recently received considerable attention; see for example \cite{braverman,bournez} and references therein. An important implication of this work is that the topological structure of invariant sets need not be computable. Perhaps this is not surprising, given that the work over the last century has clearly demonstrated the incredible diversity and complexity of invariant sets. One interpretation of these results for practical applications is that analyzing dynamics by computing invariant sets may lead to a level of computations that is too fine to be useful and perhaps ultimately unattainable. With this in mind we consider the question of the computation of coarse dynamical structures for the following rather general setting. A \emph{dynamical system} on a topological space $X$ is a continuous map $\varphi: {\mathbb{T}}^+ \times X \to X$ that satisfies \begin{enumerate} \item [(i)] $\varphi(0,x) = x$ for all $x\in X$, and \item [(ii)] $\varphi(t,\varphi(s,x)) = \varphi(t+s,x)$ for all $s,t\in {\mathbb{T}}^+$ and for all $x\in X$, \end{enumerate} where ${\mathbb{T}}$ denotes the time domain, which is either ${\mathbb{Z}}$ or ${\mathbb{R}}$ and ${\mathbb{T}}^+:=\setof{t\in{\mathbb{T}}\mid t\geq 0}$. As is discussed in detail in Section~\ref{subsec:approximationDS}, for the results presented in this paper there is no loss of generality in assuming the dynamics is generated by the continuous function $f\colon X\to X$ where $f(\cdot) := \varphi(1,\cdot)$. The most significant assumption we make is that $X$ is a compact metric space. We emphasize that we do {\em not} assume that $f$ is injective nor surjective. Recall that a set $U\subset X$ is an \emph{attracting neighborhood} for $f$ if \[ \omega(U,f):= \bigcap_{n\in{\mathbb{Z}}^+} {\rm cl}\,}%{\text{cl}\left(\bigcup_{k=n}^\infty f^k(U) \right) \subset \mbox{\rm int\,}(U). \] A set $A\subset X$ is an \emph{attractor} if there exists an attracting neighborhood $U$ such that $A = \omega(U,f)$. The sets of all attracting neighborhoods and all attractors are denoted by ${\mathsf{ ANbhd}}(X,f)$ and ${\mathsf{ Att}}(X,f)$, respectively. We remark that in general a given system can have at most a countably infinite number of attractors. Attractors are central to the study of nonlinear dynamics for at least two reasons. First, they are the invariant sets that arise from the asymptotic dynamics of regions of phase space, thus they capture the ``observable" dynamics. Second, they are intimately related to the structure of the global dynamics. More precisely, recall that Conley's fundamental decomposition theorem \cite{robinson} states that the dynamics is gradient like outside of the chain recurrent set. Furthermore, the chain recurrent set can be characterized using the set of attractors and their dual repellers. With this in mind in \cite{KMV0} we discuss a combinatorial approach for identifying attracting neighborhoods and demonstrate that, even though there may be infinitely many attractors, it is possible to obtain arbitrarily good approximations in phase space of these attractors. This in turn provides a constructive method for obtaining arbitrarily good approximations of the chain recurrent set. From the perspective of understanding the dynamics of nonlinear models one encounters the issue of minimal scales. Every model has a scale below which the model is no longer valid. Any given numerical simulation has a minimal scale, and there is a maximal resolution for experimental measurements. Especially in the latter case, the maximal relevant resolution is often dependent on the location in phase space. This issue of scale motivates recent work \cite{database,bush2012combinatorial,Ban,lyapunov} that focuses on rigorously computing global dynamical structures with an a priori choice of maximal resolution of measurement. Given a fixed scale there can be at most a finite subset of attractors that are observable. This suggests that understanding finite resolution dynamics requires a deeper understanding of the structure of the set of all attractors. In \cite{KMV-1a} we prove that ${\mathsf{ Att}}(X,f)$ and ${\mathsf{ ANbhd}}(X,f)$ are bounded, distributive lattices. The lattice operations for ${\mathsf{ ANbhd}}(X,f)$ are straightforward, $\vee = \cup$ and $\wedge = \cap$. The operations for ${\mathsf{ Att}}(X,f)$ are more subtle; $\vee = \cup$, but the $\wedge$ operation is given by $A_1\wedge A_2=\omega(A_1\cap A_2,f)$. A consequence of the lattice structure is that a finite set of attractors generates a finite sublattice ${\mathsf A}\subset {\mathsf{ Att}}(X,f)$ of attractors. By definition $\omega(\cdot,f)\colon {\mathsf{ ANbhd}}(X,\varphi)\to {\mathsf{ Att}}(X,f)$ is surjection. In \cite{KMV-1a} we show that this is, in fact, a lattice epimorphism. However, this fact does not provide any information, in and of itself, as to the structural relationship between a given finite lattice of attractors ${\mathsf A}$, the dynamic information of interest, and the set of attracting neighborhoods $\omega(\cdot,f)^{-1}({\mathsf A}) \subset {\mathsf{ ANbhd}}(X,f)$, which consist of the potentially observable or computational objects. This is resolved by the following theorem which proves that the lattice structure of invariant dynamics of interest, namely attractors, is contained within the lattice structure of the observable or computable dynamics, namely attracting neighborhoods. \begin{theorem} \label{thm:kmv1} \cite[Theorem 1.2]{KMV-1a} Let $\imath$ denote the inclusion map. For every finite sublattice ${\mathsf A} \subset {\mathsf{ Att}}(X,f)$, there exists a lattice monomorphism $k$ such that the following diagram \[ \begin{diagram} \node{~} \node{{\mathsf{ ANbhd}}(X,f)} \arrow{s,r,A}{\omega(\cdot,f)}\\ \node{{\mathsf A}} \arrow{e,l,V}{\imath} \arrow{ne,l,..,V}{k} \node{{\mathsf{ Att}}(X,f)} \end{diagram} \] commutes. \end{theorem} The homomorphism $k$ is called a lift of $\imath$ through $\omega(\cdot,f)$. The proof of Theorem~\ref{thm:kmv1} is nontrivial, especially since we are not assuming that $f$ is injective nor surjective. Thus \cite{KMV-1a} contains a detailed discussion and development of the definitions and properties of many of the standard dynamical concepts such as attractors, repellers, invariant sets, etc.~as well as corresponding neighborhoods of these objects. We do not repeat them in this paper, but recall them as necessary. As suggested at the beginning of this introduction, the focus of this paper is on computation. The computational methods we are interested in analyzing are based on a finite discretization, indexed by ${\mathcal X}$, of the phase space $X$ and the computation of an outer approximation of a map $f\colon X\to X$ by a combinatorial multivalued map ${\mathcal F}\colon{\mathcal X}\raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$}{\mathcal X}$ (see Section~\ref{sec:cgt}). Observe that a combinatorial multivalued map is equivalent to a finite directed graph. The latter interpretation is useful from the perspective of algorithms, but treating ${\mathcal F}$ as a map provides intuition as to how to define important dynamical analogues in the discrete setting. This is discussed in detail in Section~\ref{sec:comb-syst}, but we also point out the work in \cite{McG,McGW} on closed relations. Given our focus on attractors there are three structures arising from combinatorial dynamics that are of particular interest:~forward invariant sets, ${\mathsf{ Invset}}^+({\mathcal X},{\mathcal F}):=\setof{{\mathcal S}\subset {\mathcal X}\mid {\mathcal F}({\mathcal S})\subset ({\mathcal S})}$; attracting sets, ${\mathsf{ASet}}({\mathcal X},{\mathcal F}):=\setof{{\mathcal U}\subset{\mathcal X} \mid {\bm{\omega}}({\mathcal U},{\mathcal F}) \subset {\mathcal U}}$; and attractors, ${\mathsf{ Att}}({\mathcal X},{\mathcal F}) := \setof{{\mathcal A}\subset{\mathcal X} \mid {\mathcal F}({\mathcal A})={\mathcal A}}$. In Section~\ref{sec:comb-syst} we assign appropriate lattice structures to these sets. Note that these lattices are explicitly computable, since they are defined in terms of elementary operations on a finite directed graph. To easily pass between the combinatorial and continuous dynamics, we insist that the discretization of phase space be done with regular closed sets. Given a compact metric space $X$, the family of all regular closed sets ${\mathscr R}(X)$ forms a Boolean algebra. As this gives rise to technical issues, it is important to note that the lattice operations for ${\mathscr R}(X)$ differ from those of ${\mathsf{Set}}(X)$, in particular $\vee=\cup$ and $\wedge={\rm cl}\,}%{\text{cl}(\mbox{\rm int\,}(\cdot)\cap\mbox{\rm int\,}(\cdot))$. The atoms of any finite sublattice of ${\mathscr R}(X)$ form a {\em grid}, which provides an appropriate discretization of the phase space $X$. As indicated above the grid is indexed by ${\mathcal X}$. We pass from subsets of ${\mathcal X}$ to subsets of $X$, by means of an evaluation map $\supp{\cdot}\colon {\mathcal X} \to {\mathscr R}(X)$. Observe that, as a consequence of the discretization procedure, our computations can only represent elements of ${\mathscr R}(X)$, which is a strict subset of ${\mathsf{Set}}(X)$. We denote the family of attracting neighborhoods of $f$ that are regular closed sets of $X$ by ${\mathsf{ ANbhd}}_{\scrR}(X,f)$. Even though ${\mathsf{ ANbhd}}_{\scrR}(X,f)\subset {\mathsf{ ANbhd}}(X,f)$, this inclusion is not a lattice homomorphism since, the lattice operations are different. Furthermore, for a fixed multivalued map ${\mathcal F}$, an outer approximation for $f$, the evaluation map $\supp{\cdot}$ maps ${\mathsf{ASet}}({\mathcal X},{\mathcal F})$ to a strict subset of ${\mathsf{ ANbhd}}_{\scrR}(X,f)$. The lattice homomorphisms that relate the above mentioned lattices produce the following commutative diagram (see Remark~\ref{rem:comm-lift11} for the analogue for $\varphi$) \begin{equation} \label{diag:comm-lift11} \begin{diagram} \node{{\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})}\arrow{se,r,A}{{\bm{\omega}}(\cdot,{\mathcal F})}\arrow{e,l,V}{\imath}\node{{\mathsf{ASet}}({\mathcal X},{\mathcal F})}\arrow{e,l,V}{|\cdot|}\arrow{s,r,A}{{\bm{\omega}}(\cdot,{\mathcal F})}\node{{\mathsf{ ANbhd}}_{\scrR}(X,f)}\arrow{e,l,A}{\omega(\cdot,f)}\node{{\mathsf{ Att}}(X,f)}\\ \node{}\node{{\mathsf{ Att}}({\mathcal X},{\mathcal F})}\arrow{ene,r}{\omega(|\cdot|,f)} \end{diagram} \end{equation} where $\imath$ is the inclusion map. Returning to the question of computability, the analoguous result to Theorem~\ref{thm:kmv1} in the context of a multivalued outer approximation ${\mathcal F}$ of $f$ is the existence of a lifting for either of the following commutative diagrams \begin{equation} \label{eq:twodiagrams} \begin{diagram} \node{~} \node{{\mathsf{ASet}}({\mathcal X},{\mathcal F})} \arrow{s,r,A}{\omega(\cdot,f)}\\ \node{{\mathsf A}} \arrow{e,l,V}{\imath} \arrow{ne,l,..,V}{} \node{{\mathsf{ Att}}(X,f)} \end{diagram} \qquad \text{or} \quad \begin{diagram} \node{~} \node{{\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})} \arrow{s,r,A}{\omega(\cdot,f)}\\ \node{{\mathsf A}} \arrow{e,l,V}{\imath} \arrow{ne,l,..,V}{} \node{{\mathsf{ Att}}(X,f)} \end{diagram} \end{equation} for a given finite sublattice of attractors ${\mathsf A}$. Observe that by \eqref{diag:comm-lift11} a lift for the second diagram implies a lift for the first. If ${\mathsf{ Att}}(X,f)$ is an infinite lattice, then any given approximation ${\mathcal F}\colon {\mathcal X}\raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$}{\mathcal X}$ captures only a finite sublattice ${\mathsf A}\subset {\mathsf{ Att}}(X,f)$. With this in mind we consider sequences of multivalued maps ${\mathcal F}_n\colon {\mathcal X}_n\raisebox{-0.2ex}{$\,\overrightarrow{\to}\,$}{\mathcal X}_n$ that provide arbitrarily close approximations of $f$. There are essentially two types of sequences that we consider, one that involves a coherent refinement of the grids associated with ${\mathcal X}_n$ and the other where it is only assumed that the diameter of the grid can be made arbitrarily small. The first is most relevant if one considers a numerical scheme based on a systematic refinement of phase space. The second is relevant if one wants to compare approximations performed using different types of discretizations of phase space. With a coherent refinement scheme we are able to prove the existence of a lifting, Theorem~\ref{thm:conv-latt-1}, based on the second diagram of \eqref{eq:twodiagrams}. The more general result, Theorem~\ref{thm:conv-latt-3}, is based on the first diagram. We conclude this introduction with a brief outline of the paper. Section~\ref{sec:comb-syst} introduces the dynamics of combinatorial multivalued maps and the appropriate lattice structures. This includes the lattices of backward invariant sets ${\mathsf{ Invset}}^-({\mathcal X},{\mathcal F})$, repelling sets ${\mathsf{RSet}}({\mathcal X},{\mathcal F})$, and repellers ${\mathsf{ Rep}}({\mathcal X},{\mathcal F})$. The duality between these lattices, associated with backward dynamics, and those associated with forward dynamics is presented in the commutative diagram \eqref{eq:diag+-}. In Section~\ref{sec:cont-graph} we focus on combinatorial multivalued maps as an approximation scheme for continuous nonlinear dynamics. We begin in Section~\ref{sec:cgt} by recasting the concept of grid \cite{Mrozek-grid} into the more general setting of regular closed subsets, cf.\ \cite{Walker}. Section~\ref{subsec:at-rep-neigh} contains results, summarized for the most part by Theorems~\ref{thm:first-evalf} and \ref{thm:commutative-diagram}, that relate the lattice structures of combinatorial systems with those of continuous systems. Section~\ref{subsec:approximationDS} deals with the issue of the approximation of dynamical systems where the time variable ${\mathbb{T}} = {\mathbb{R}}$. As mentioned earlier in the introduction, one approach is to set $f(\cdot) = \varphi(\tau, \cdot)$, where if ${\mathbb{T}} = {\mathbb{R}}$ then it is permissible to choose any fixed $\tau >0$. For this approach the concept of outer approximation is sufficient. The weakness of this approach is that from the perspective of obtaining optimal approximations it may be desirable to choose different values of $\tau$ on different regions of phase space. An alternative approach developed in \cite{BKM} involves combinatorializing the flow via a triangulation of space and the multivalued mapping is defined by considering the behavior of the associated vector field on the vertices of the triangulation. This method fits into our framework but requires the notion of a weak outer approximation as is demonstrated via the commutative diagram \eqref{diag:AR2a}. Section~\ref{sec:str-conv} brings together the ideas of Sections~\ref{sec:comb-syst} and \ref{sec:cont-graph} to demonstrate the general computability of the lattices of interest. We begin in Section~\ref{subsec:conv-fam} with a discussion concerning the convergence of outer approximations from a more classical numerical analysis perspective, i.e.\ tracking of individual orbits. Section~\ref{subsec:resol} discusses the identification individual attractors or repellers using a outer approximations. Section~\ref{subsec:PLG} returns to the issue of convergent sequences of outer approximations but from a lattice theoretic perspective. We also discuss Birkhoff's representation theorem for finite distributive lattices and discuss its application in the context of lifts. Finally in Section~\ref{subset:lift-grid} we prove the desired lifting theorems, Theorem~\ref{thm:conv-latt-1} and \ref{thm:conv-latt-3}. The reader will immediately note that the details of the proofs are carried out in the context of repeller structures, Theorems~\ref{thm:conv-latt-2} and \ref{thm:lift}, and then duality is used to obtain Theorem~\ref{thm:conv-latt-1} and \ref{thm:conv-latt-3}, respectively. The reason for this is related to the lattice structures of ${\mathsf{ Att}}(X,f)$ and ${\mathsf{ Rep}}(X,f)$. In particular, $\wedge = \cap$ for ${\mathsf{ Rep}}(X,f)$, but not for ${\mathsf{ Att}}(X,f)$. This lack of symmetry arises from the fact that we do not assume that $f$ is either injective or surjective, and thus attractors and repellers have fundamentally different properties. The duality between between attractor/attracting neighborhoods and repellers/repelling neighborhoods is expressed in following diagram, cf.\ \eqref{diag:AR2} \begin{equation}\label{diag:AR2b} \begin{diagram} \node{{\mathsf{ ANbhd}}(X,f)} \arrow{e,l,<>}{^c} \arrow{s,l,A}{\displaystyle \omega}\node{{\mathsf{ RNbhd}}(X,f)} \arrow{s,r,A}{\displaystyle \alpha}\\ \node{{\mathsf{ Att}}(X,f)} \arrow{e,l,<>}{^*} \node{{\mathsf{ Rep}}(X,f)} \end{diagram} \end{equation} where ${\mathsf{ Rep}}(X,f)$ and ${\mathsf{ RNbhd}}(X,f)$ are the lattices of repellers and repelling neighborhoods, respectively. \begin{remark} We include a variety of different lattices in this paper. In each case the $\vee$ operation is simply the union of sets, but there are five different $\wedge$ operations. For the benefit of the reader we include the following tables, the first for topological structures and the second for combinatorial structures, as a simple summary. \vskip 12pt \begin{center} \begin{tabular}{|l|c|} \hline Lattice & $U \wedge V$\\ \hline ${\mathsf{ Att}}(X,f)$ & $\omega(U\cap V)$\\ ${\mathsf{ Rep}}(X,f)$ & $U\cap V$\\ ${\mathsf{ ANbhd}}(X,f)$ & $U\cap V$\\ ${\mathsf{ RNbhd}}(X,f)$ & $U\cap V$\\ ${\mathsf{ ANbhd}}_{\scrR}(X,f)$ & ${\rm cl}\,}%{\text{cl}(\mbox{\rm int\,}(U)\cap\mbox{\rm int\,}(V))$\\ ${\mathsf{ RNbhd}}_{\scrR}(X,f)$ & ${\rm cl}\,}%{\text{cl}(\mbox{\rm int\,}(U)\cap\mbox{\rm int\,}(V))$\\ ${\mathsf{ Invset}}^\pm(X,f)$ & $U\cap V$\\ \hline \end{tabular} \qquad \begin{tabular}{|l|c|} \hline Lattice & ${\mathcal U} \wedge {\mathcal V}$\\ \hline ${\mathsf{ Att}}({\mathcal X},{\mathcal F})$ & ${\bm{\omega}}({\mathcal U}\cap {\mathcal V})$\\ ${\mathsf{ Rep}}({\mathcal X},{\mathcal F})$ & ${\bm{\alpha}}({\mathcal U}\cap {\mathcal V})$\\ ${\mathsf{ASet}}({\mathcal X},{\mathcal F})$ & ${\mathcal U}\cap {\mathcal V}$\\ ${\mathsf{RSet}}({\mathcal X},{\mathcal F})$ & ${\mathcal U}\cap {\mathcal V}$\\ ${\mathsf{ Invset}}^+({\mathcal X},{\mathcal F})$ & ${\mathcal U}\cap {\mathcal V}$\\ ${\mathsf{ Invset}}^-({\mathcal X},{\mathcal F})$ & ${\mathcal U}\cap {\mathcal V}$\\ \hline \end{tabular} \end{center} \vskip 12pt \end{remark}

\section{Introduction} Phase I dose-finding trials aim to identify the maximum tolerated dose (MTD), defined as the highest dose with probability of toxicity close to or no higher than a targeted rate $p_T$ (say, $p_T = 1/6$ or $1/3$). Classical dose-finding trials usually prespecify a set of $D$ ($D>1$) candidate dose levels, and each dose has an unknown probability $p_i$ of inducing dose-limiting toxicity (DLT) among patients. The trial starts from a low dose, treats a cohort of patients at the dose, records their binary DLT outcomes up to a period of time, and decides the next dose level for treating future cohorts of patients. Iteratively, the trial proceeds by sequentially treating future \color{black} \rm patients at a grid of predetermined dose levels according to an appropriate design. Usually additional safety rules are implemented to safe-guard the enrolled patients (Goodman et al., 1995), e.g., no skipping in dose escalation. Importantly, designs for the phase I trials are confined by the following basic principle: \vskip 1.5em \noindent {\bf Ethical Principle:} {\it One should not treat patients at doses perceived overly toxic, such as those believed to be above the MTD. } \vskip 1.5em Since 1989, the 3+3 design (Storer, 1989) and the CRM design (O"Quigley et al., 1990) have been the two main-stream statistical methods for phase I dose-finding trials. The 3+3 design has been the most frequently applied approach in practice and CRM the most polished model-based methods in the literature. CRM is based on a dose-response model that borrows information across all the doses for decision making. However, the main issues of CRM were lack of standard implementation in practice and challenges in interpreting the method to clinicians. These issues have been improved recently as many useful tools (e.g., the dfCRM and bcrm R packages) for CRM have been developed. We refer to Cheung (2011) for an extensive discussion of CRM. Since 10 years ago, a new development in the design for dose-finding clinical trials has been gradually transforming the research and practice. In particular a new class of interval-based dose-finding designs have been recognized, extended, and implemented by researchers and practitioners. In Cheung and Chappell (2002), the authors introduced the concept of indifference interval in defining the MTD, and Ji et al. (2007ab) first introduced a design that solely depends on properties of toxicity probability intervals (TPI). That is, the statistical inference for dose-finding process is dependent on the posterior probability of the TPIs, instead of a point estimate, that describes the toxicity of a dose. Ji et al. (2010), Ji and Wang (2013) and Yang et al. (2015) further developed the modified TPI (mTPI) designs to improve the performance of TPI. These methods share the same idea of assessing the posterior probabilities that the toxicity rate $p_i$ of a dose $i$ falls into three intervals $(0, p_T - \epsilon_1)$, $(p_T - \epsilon_1, p_T + \epsilon_2)$, and $(p_T + \epsilon_2, 1)$, represented as $Pr\{p_i \in (p_T - \epsilon_1, p_T + \epsilon_2) \mid data\}$ for example, and conducting up-and-down type of dose-finding decisions using the posterior probabilities of the three intervals. Here $\epsilon_{1} $ and $\epsilon_2$ \color{black} \rm are small fractions that reflect investigators' desire about how accurate they want the MTD to be around $p_T$, the target value. Usually $\epsilon_{1,2} \le 0.05.$ For example, when $p_T = 0.3$ and $\epsilon_{1,2} = 0.05$, the three intervals become $(0, 0.25)$, $(0.25, 0.35)$, and $(0.35, 1)$. This means that doses with toxicity probabilities between 0.25 and 0.35 can be considered as the MTDs, and those lower than 0.25 or higher than 0.35 are too low or too high relative to the target value $p_T=0.3$. Based on this simple logic, mTPI calculates the posterior probabilities of these three intervals for a given dose and use a decision theoretic framework to guide the optimal dose finding in mTPI. The authors show that the decisions of mTPI are optimal, minimizing a posterior expected risk. Recently, Guo et al. (2017) proposed an extended version called mTPI-2, which is based on the same interval decisions but with finer intervals to mitigate some practical concerns in mTPI. During the same time period since 2007, an alternative class of interval designs was developed, represented by the cumulative cohort design (CCD, Ivanova et al., 2007) and more recently an extended version as the Bayesian optimal interval design (BOIN, Liu and Yuan, 2015). In CCD, the authors propose to calculate a point estimate $\hat{p}_i = x_i /n_i$ where $x_i$ is the number of patients experiencing the dose limiting toxicity (DLT) events and $n_i$ is the number of patients that have been treated at current dose $i$. Then CCD compares $\hat{p}_i$ to three intervals of the same format as in mTPI. If $\hat{p}_i$ falls in the lower interval $(0, \phi_1),$ middle interval $(\phi_1, \phi_2)$, or upper interval $(\phi_2, 1)$, the decision is to escalate to dose $(i+1)$, stay at dose $i$, or de-escalate to dose $(i-1)$, respectively. The BOIN design uses the same type of decision rules except that it re-calibrates the middle interval of $(\phi_1, \phi_2)$ to a new interval $(\lambda_{e}, \lambda_{d})$ which typically satisfies $\phi_1 < \lambda_e < \lambda_d < \phi_2$. The authors argue that the new boundaries $\lambda_{e}$ and $\lambda_d$ are optimal as they minimize an error function. We will review these two developments of interval-based designs and demonstrate the key similarities and differences among them. In Section 2, we review TPI, mTPI, and mTPI-2 as the ``iDesigns", with ``i" standing for ``intervals", which use probabilities of the toxicity probability intervals in a Bayesian framework. We also review CCD and BOIN as the ``IB-Designs", with "IB" standing for "interval boundaries". , which use the point estimate $\hat{p}_i$ and fixed interval boundaries for inference. In addition, we delineate an important difference between the interval-based designs and the CRM design. In Section 3 we report operating characteristics of different designs based on objective, crowd-sourcing, and large-scale simulation studies. In Section 4 we propose a new criterion for the evaluation of different designs other than the operating characteristics from simulated trials. We end with a discussion in Section 5. \section{ Review of Designs } \label{sec:review} \subsection{The iDesigns} \label{sec:iDesign} We first review a class of designs based on probability of toxicity probability intervals (TPIs). We call these designs the iDesigns. \paragraph*{The TPI Design} Consider three specific TPIs, the under-dosing interval (UI) $(0, p_T - \epsilon_1)$, the equivalence interval (EI) $(p_T - \epsilon_1, p_T + \epsilon_2)$, and the over-dosing interval (OI) $(p_T + \epsilon_2, 1)$. Suppose the toxicity rate of a dose $i$ is $p_i$. When $p_i$ falls into UI, EI, or OI, dose $i$ is considered below, equivalent to, or above the MTD, respectively. If dose $i$ is currently used for treating patients, then the decision is to ``escalate'' (E) to dose $(i+1)$, ``stay'' (S) at dose $i$, or ``de-escalate'' (D) to dose $(i-1)$, respectively. In other words, {\bf decision E is associated with UI, S associated with EI, and D associated with OI.} This mimics a typical up-and-down type of decision in dose finding (see Stylianou and Flournoy, 2002 for a discussion). Therefore, as long as one can accurately infer to which interval $p_i$ belongs, one can easily decide the dose for future patients in the trial. In the TPI design, Ji et al., (2007a) use $(0, p_T - K_1 \sigma_i)$, $(p_T - K_1 \sigma_i, p_T + K_2 \sigma_i)$, and $(p_T+ K_2 \sigma_i, 1)$ to describe the EI, UI, and OI, respectively, where $K_{1,2}$ are two prespecified constants and $\sigma_i$ is the posterior standard deviation of $p_i$. They define a penalty (loss) function that maps decisions E, S, or D and the value of $p_i$ to a value of loss. For example, if $p_i$ is in UI and the decision is $E$, there is no loss. However, if $p_i$ is in UI and the decision is $D$ or $S$, there is a positive loss. The positive loss values quantify the violation of the underlying ethical constraints that patients must not be treated at doses above the MTD. Under this decision theoretic framework, the authors derived the Bayes rule which chooses one of the three decisions (E, S, D) that minimizes the posterior expected loss. For statistical inference, the TPI design models the observed binary toxicity outcomes use a binomial likelihood, and models $p_i$ with independent beta prior distributions. This anti-intuitive choice of prior models have been show to work well with up-and-down rules, and are in sharp contrast to other model-based designs, such as the CRM design. However, it appears the use of independent beta priors is being increasingly adopted by various new methods, including the most recent development of a semi-parametric CRM design (Clertant and O'Quigley, 2017). \paragraph*{The mTPI Design} In the mTPI design (Ji et al, 2010), the authors made $\epsilon_{1}$ and $\epsilon_2$ two user-specified constants, with the notion that the equivalence interval $(p_T - \epsilon_1, p_T + \epsilon_2)$ should be provided by the trial clinicians. This is similar to the concept of user-provided effect sizes in designing randomized clinical trials. In particular, $(p_T - \epsilon_1)$ should be the lowest toxicity probability that the clinician would feel comfortable not to escalate, and $(p_T + \epsilon_2)$ should be the highest toxicity probability that the clinician would feel comfortable not to de-escalate. Therefore, any dose with probability in between $(p_T - \epsilon_1, p_T + \epsilon_2)$ can be considered as the MTD. Under the same decision theoretic framework, mTPI derives the optimal rule (Bayes' rule) that minimizes the posterior expected loss. The optimal rule is simple and is \color{black} \rm equivalent to 1) identifying the interval among UI, EI and OI that has the largest unit probability mass (UPM, defined below) and 2) choosing the decision E, S, or D that is associated with the interval. Here, UPM of an interval is defined as the ratio of the posterior probability $p_i$ belonging to the interval and the length of the interval. For example, \begin{equation} \mbox{UPM(EI)} = \frac{Pr\{p_i \in (p_T - \epsilon_1, p_T + \epsilon_2) \mid data\}}{\epsilon_1+\epsilon_2}. \label{eq:UPM} \end{equation} \paragraph*{The mTPI-2 Design} Although mTPI uses the Bayes' rule -- an optimal decision rule that minimizes the posterior expected loss to guide decision making, the statistically optimal decisions may be at odds with the {\bf Ethical Principle} stated in Section 1 for clinical trials. For example, suppose at the current dose 3 out of 6 patients experienced DLT and the trial targets probability $p_T=0.3$ for the true MTD; under the Bayes' rule, the mTPI design would instruct to ``S'', stay at the current dose. Since the empirical rate is 3/6=0.5 and the target rate is 0.3, oftentimes the dose would be considered above the MTD and therefore the desirable decision is ``D'', de-escalate. However, there is not clear guidelines in the community on what decisions are acceptable. And even though there is a large uncertainty associated with the data of only 6 patients, many safety review boards would express concerns if a design chooses ``S'' when the empirical data is 3/6. Interestingly, when the trial data shows 4 out of 8 or 5 out of 10 patients experienced DLT, even though the empirical rate is still 0.5, mTPI would instruct to ``D'', de-escalate to the lower dose. Therefore, the mTPI design automatically accounts for the variability of the data in the decision making, which is considered as a key principle in statistics inference. However, in trial implementation, patient safety should be strictly protected. To this end, Guo et al. (2017) conduct a detailed investigation. They note that just as model selection criteria like AIC (Akaike, 1974) or BIC (Schwarz, 1978) include a penalty for model size and a score (log-likelihood) for model fitting, the Bayesian decision theoretic framework in mTPI intrinsically treats the three intervals as three models, and penalize models based on the model size which is the length of each interval. As a result, the Bayesian Occam's razor principle favoring parsimonious models (Jefferys and Berger, 1992) is seen in mTPI. Guo et al. (2017) propose a simple remedy known as the mTPI-2 design, in which the UI and OI are broken into smaller subintervals having the same length as the EI. Then the interval having the largest UPM value will be associated with the corresponding dose-finding decision. For example, suppose $p_T=0.3$ and the EI is $(0.25, 0.35)$; mTPI-2 then breaks the OI $(0.35, 1)$ into subintervals $(0.35, 0.45), (0.45, 0.55), $ and so on. If any subinterval in the OI has the largest UPM, the decision is ``D'', to de-escalate; if any subinterval in the UI has the largest UPM, the decision is ``E'', to escalate; and if the EI has the largest UPM, the decision is ``S'', to stay. Guo et al. (2017) show that this again is an optimal decision that minimizes the posterior expected loss for a 0-1 loss function. In addition, they demonstrate that the new Bayesian inference is now not affected by the Occam's razor, as all the subintervals (except for the two boundary ones) have the same length. Consequently, when $p_T=0.3$ and 3 out of 6 patients experienced DLT, mTPI-2's decision is now ``S'', to stay. \subsection{The IB-Designs}\label{sec:ibdesign} Almost parallel to the development of the TPI, mTPI, and mTPI-2 methods, a different class of interval-based designs has been developed and extended. These methods are based on predetermined decision rules that can be summarized in three steps: \begin{enumerate} \item Obtain a point estimate of $p_i$, say $\hat{p}_i = x_i/n_i$. \item Decide three toxicity probability intervals $(0, \phi_1), (\phi_1, \phi_2), (\phi_2, 1)$ so that $\phi_1 < p_T < \phi_2$ are two boundaries that define the equivalence interval $(\phi_1, \phi_2)$ (they are similar to $\epsilon_{1,2}$ in the mTPI and mTPI-2 designs), \item Escalate, stay, or de-escalate if $\hat{p}_i$ falls into the three intervals, respectively. \end{enumerate} As can be seen, these rules use intervals differently from the iDesigns. The iDesigns use $Pr(p_i \in \mbox{an interval} \mid data)$ for decision making, while here the decisions are based on comparison of $\hat{p}_i$ and intervals. In other words, here the intervals serve as boundaries for assessing the point estimate $\hat{p}_i$. Therefore, we call these interval-based designs the IB-Designs where ``IB" stands for interval boundaries. \paragraph*{The CCD Design} Under the CCD design (Ivanova et al., 2007), if $\hat{p}_i$ falls into the lower interval $(0, \phi_1)$, escalate to dose $(i+1)$; the middle interval $(\phi_1, \phi_2)$, stay at dose $i$; the upper interval $(\phi_2, 1)$, de-escalate to dose $(i-1)$. Because only change of one dose level is allowed in escalation or de-escalation, the CCD design generates a Markov chain of toxicity outcomes, and the authors prove asymptotic convergence of the design. In the CCD design, the authors recommend a simpler form by expressing the lower cutoff toxicity probability $\phi_1 = (p_T - \Delta)$ and the upper cutoff $\phi_2 = (p_T + \Delta),$ which have equal distance to $p_T$. Then the above 1-3 steps are applied to guide the dose finding trial. Choices of $\Delta$ are given by Ivanova et al. (2007) for moderate sample size ($\le 20$) and six dose levels. We list the recommended values below: 1) $\Delta = 0.09$ for $p_T \in \{0.10, 0.15, 0.20, 0.25\}$; 2) $\Delta = 0.10$ for $p_T \in \{0.30, 0.35\}$; 3) $\Delta = 0.12$ for $p_T = 0.40$; and 4) $\Delta=0.13$ for $p_T \in \{0.45, 0.50\}$. \paragraph*{The BOIN Design} Liu and Ying (2015) extend CCD and develop the BOIN designs, with local and global BOIN as two versions. They state that BOIN is an improvement of CCD since it uses interval boundaries that are optimal and vary with enrollment counts and dose levels. In BOIN, the dose-finding problem is cast slightly differently. Now consider $0<\phi_1 <p_T < \phi_2 <1$ two user-provided values (instead of theoretically derived in CCD) that represent the lower and upper bound of the equivalence interval for $p_T$. In this sense, $\phi_1$ and $\phi_2$ are given and decided by the trial investigators, which is similar to the mTPI and mTPI-2 designs. Therefore, a dose with a toxicity probability falling into $(\phi_1, \phi_2)$ can be considered as an MTD candidate. Suppose dose $i$ is currently used to treat patients in the trial. In the local BION design, an ad-hoc optimization procedure changes the two values of $\phi_1$ and $\phi_2$ to two new values called $\lambda_e(i)$ and $\lambda_d(i)$ so that the interval $(\lambda_e(i), \lambda_d(i))$ is nested in $(\phi_1, \phi_2)$, i.e., $\phi_1 < \lambda_e(i) < \lambda_d(i) < \phi_2$. Interestingly, the optimal values of $\lambda_e(i)$ and $\lambda_d(i)$ do not depend on dose level $i$ based on the local BOIN design. Therefore, we drop index $i$ in $\lambda_e(i)$ and $\lambda_d(i)$ hereafter for local BOIN. The local BOIN design first examines if $\hat{p}_i$ falls into one of the three intervals $(0, \lambda_e]$, $(\lambda_e, \lambda_d)$, $[\lambda_d, 1)$, and escalates to dose $(i+1)$, stays at dose $i$, or de-escalates to dose $(i-1)$, accordingly. In other words, the local BOIN design uses the same concept for dose escalation as the CCD design, except BOIN changes the original user-provided boundary $\phi_{1,2}$ to $\lambda_{e,d}$ based on an optimization criterion. Liu and Ying (2015) also develop a global BOIN design, in which the two boundary values depend on the number of treated patients at each dose level $i$. That is, $\lambda_{e,d}$ in global BOIN do depend on dose level $i$. Although conceptually the global BOIN design is more advanced, the authors conclude that the global BOIN design does not perform as well as the local BOIN design in their simulations. Therefore, the authors recommend the local BOIN design for practical implementation, which is quite similar to the CCD design since the boundaries of both designs are not dependent on dose levels. We refer to BOIN as the local BOIN design hereafter. \subsection{The iDesigns and IB-Designs}\label{two iDesigns} Both iDesigns and IB-Designs use an equivalence interval $(a, b),$ $a < p_T < b,$ to convey the fundamental concept of the interval designs -- {\bf to represent the MTD with an interval instead of a single value.} BOIN and CCD both use an equivalence interval $(\phi_1, \phi_2)$ as the boundaries for $\hat{p}_i$. While CCD uses theoretical derivation to determine $\phi_{1,2}$, BOIN requires users provide $\phi_{1,2}$ first and then derives a new set of boundaries $\lambda_{e,d}$ based on $\phi_{1,2}$ as the equivalence interval. Table \ref{tab:designs} lists the main common and different features of the designs. The main difference between iDesigns and IB-Designs is that iDesigns use the posterior probability that $p_i$ falls into the intervals as the foundation for statistical inference and decision making, while the IB-Designs relies on the point estimate $\hat{p}_i$ and directly compares it with the interval boundaries. Also, the IB-designs do not use a classical decision-theoretic framework (Berger, 2013) that involves loss, risk, and decision rules as the essential elements. The optimality suggested by BOIN is based on an ``error'' function $\alpha(\lambda_e, \lambda_d)$ (Liu and Ying, 2015) to mimic the type-I error in hypothesis testing. In contrast, the iDesigns use a formal decision theoretic setup, with a loss/penalty function, and aims to find the optimal decision rule that takes D, S, or E as {\it actions} and minimizes the posterior expected loss. Ying et al. (2016) argue that the use of UPM in mTPI is not desirable since it produces questionable decisions such as ``S'' when 3 out of 6 patients experienced toxicity at a given dose. However, Guo et al. (2017) show that these decisions in mTPI are theoretically justified under the principle of Occam's razor. But because of the {\bf Ethical Principle} in phase I clinical trials, the mTPI decisions might not be perceived as ethically sound in practice. Therefore, Guo et al. (2017) propose the mTPI-2 design which blunts the Occam's razor in order to achieve theoretical optimality and ethical satisfaction at the same time. Later in this manuscript, we show that mTPI-2 and BOIN perform quite similarly. \begin{table} \resizebox{1\textwidth}{!}{ \begin{tabular}{|l|c|c|c|c|c|}\hline Criterion & \multicolumn{3}{|c|}{The iDesigns} & \multicolumn{2}{|c|}{The IB-Design} \\ \cline{2-6} & TPI (2007) & mTPI (2010) & mTPI-2 (2016) & CCD (2007) &BOIN (2015) \\ \hline 1. $p_T \pm constant$ as equivalence interval & Y & Y & Y & Y & Y \\ 2. Loss/penalty function & Y & Y & Y & \fbox{N} & \fbox{N} \\ 3. Safety rule $Pr(p_i > p_T | data) < 0.95$ & Y & Y & Y & \fbox{N} & Y \\ 4. Isotonic$^*$ for MTD selection & Y & Y & Y & Y & Y \\ 5. A decision table of (D, S, E) & Y & Y & Y & Y & Y \\ 6. Inference using probabilities of the intervals & Y & Y & Y & \fbox{N} & \fbox{N}\\ 7. Additional safety rule and dose exclusion rule & Y & Y & Y & \fbox{N} & Y\\ \hline \end{tabular} } \caption{Summary of different features between the three iDesigns, TPI, mTPI, mTPI-2, and two IB-Designs, CCD and BOIN. "Y" or "N" represents the feature is or is not present in the design. $^*$: All the methods use isotonic transformation to select the MTD at the end of the trial. }\label{tab:designs} \end{table} There are a few other iDesigns and IB-Designs in the literature with similar features as the ones we reviewed so far. For example, there is a frequentist-version of the mTPI design, called TEQR (Blanchard and Longmate, 2011) as an iDesign. The Group design (Gezmu and Flournoy, 2006) that precedes the CCD design belongs to the category of IB-Designs. TEQR uses the same concept as the mTPI but resorts to a frequentist inference. The Group design decides the next dose assignment only using the outcomes of the current cohort of subjects at a dose level. We omit reviewing these methods in details as they largely overlap with the iDesigns and IB-Designs that have been discussed. We also omit the detailed review of Cheung and Chappell (2002) who pioneered the use of indifference interval boundaries in their work, as their work is largely aimed at the sensitivity of the CRM design, a non-interval dose-finding design. \subsection{A Key Difference to CRM}\label{sec:compCRM} The iDesigns and IB-Designs differ from the CRM design in one important aspect. The interval designs use up-and-down and fixed decision rules when patient outcomes are observed at a given dose. For example, suppose dose $i$ is the current dose and for a given set of observations $(x_i, n_i)$ at the current dose $i$, the next dose level for treating future patients under the interval designs is fixed, regardless of data on other doses and the dose level $i$. CRM is different and uses random rules. That is, given $(x_i, n_i)$ at dose $i$, the decision of escalation, de-escalation, or staying can vary and depends on $(x_j, n_j)$ on other doses $j$'s. That is, CRM ``borrows information'' from all the doses, which is a sound statistical principle. However, as seen in mTPI, for dose-finding trials with the {\bf Ethical Principle} and small sample size, sound statistical principles do not necessarily lead to sound ethical decisions. For example, the original CRM method would give ethically unacceptable decisions, which led to many additional ethical rules later on (Goodman et al., 1995; Babb et al., 1998). In contrast, the interval-based designs do not formally borrow information across doses in the statistical estimation. However, because they use up-and-down rules (D, S, and E), there is intrinsic borrowing information in the Markov-dependent manner (Ivanova et al., 2007) across different cohorts of patients. To see this, let the current dose be dose level $i$. Then under iDesigns and IB-Designs, dose $(i-1)$ must have been perceived as below the MTD and dose $(i+1)$ is either un-used or must have been perceived as above the MTD. \section{Numerical Examples}\label{sec:example} \subsection{Designs Under Comparison} We conducted three extensive simulation studies to assess the performance of four different types of designs, 1) 3+3 as an algorithmic design, 2) CRM as the classical non-interval-based design, 3) mTPI and mTPI-2 as iDesigns, and 4) BOIN as an IB-Design. For BOIN, we implemented three versions of the local BOIN design to examine the sensitivity of the design to different specifications of the equivalence boundaries. Recall that BOIN uses the simple rule that compares $\hat{p}_i = x_i / n_i$ with two boundaries, $\lambda_e$ and $\lambda_d$. If $\hat{p}_i$ is in $(0, \lambda_e]$, $(\lambda_e, \lambda_d)$, or $[\lambda_d, 1)$, the decision is E, S, or D, respectively. The two $\lambda$'s are calculated as a linear and one-to-one functions of two user-provided values $\phi_1$ and $\phi_2$ where $0<\phi_1 < \lambda_e < \lambda_d < \phi_2 < 1$. The values of $\phi_1$ and $\phi_2$ are determined so that doses with toxicity probabilities smaller than $\phi_1$ or larger than $\phi_2$ will not be considered as MTD. In order to compare BOIN with mTPI and mTPI-2 under different BOIN assumptions, we implemented three versions of BOIN. In the first version we used the default choice in which Liu and Ying (2015) suggested setting $\phi_1 = 0.6 p_T$ and $\phi_2 = 1.4 p_T$ regardless of trials, diseases and dose levels. We call this version $\mbox{BOIN}_{\mbox{default}}.$ The second version is to set $(p_T - \phi_1) = \epsilon_1$ and $(\phi_2 - p_T) = \epsilon_2$, so that the user-provided equivalence interval is the same for BOIN and the iDesigns. We call the second version $\mbox{BOIN}_{\mbox{epsilon}}.$ However, as can be seen above BOIN actually uses $\lambda_{e}$ and $\lambda_{d}$, not the user-provided $\phi_{1,2}$ for decision making. For this reason, we tried the third version of BOIN, called $\mbox{BOIN}_{\mbox{lambda}},$ which sets $(p_T - \lambda_e) = \epsilon_1$ and $(\lambda_d - p_T) = \epsilon_2$. This way, the actual intervals for decision making between BOIN and iDesigns are identical. For example, if $p_T = 0.3$, the mTPI and mTPI-2 designs would by default assume $(0.25, 0.35)$ as the equivalence interval, corresponding to $\epsilon_1 = \epsilon_2 =0.05.$ In contrast, BOIN would have three versions: $\mbox{BOIN}_{\mbox{default}}$ assumes $\phi_1 = 0.18$ and $\phi_2 = 0.42$, which gives $\lambda_e = 0.236$ and $\lambda_d = 0.358.$ Therefore, $\mbox{BOIN}_{\mbox{default}}$ escalates if the current dose $i$ has $\hat{p}_i < 0.236$, de-escalates if $\hat{p}_i > 0.358$, or stay if $0.236 < \hat{p}_i < 0.358.$ For $\mbox{BOIN}_{\mbox{epsilon}},$ it has $\phi_1=0.25$ and $\phi_2=0.35$ which gives $\lambda_e=0.275$ and $\lambda_d = 0.325$. And for $\mbox{BOIN}_{\mbox{lambda}}$, one would use $\lambda_e=0.25$ and $\lambda_d = 0.35$ directly. Note this implies that $\phi_1 = 0.205$ and $\phi_2 = 0.402.$ \subsection{Simulation Scenarios}\label{sec:scenario} We considered three sets of scenarios to comprehensively compare the designs. \paragraph{Set one.} The first set of scenarios was generated via ``crowd-sourcing'' (Yang et al., 2015). Specifically, we recorded the scenarios generated by the real users of the online tool Next-Gen DF (NGDF, Yang et al., 2015), which provided web-based software for simulating phase I dose-finding trials using mTPI-2, mTPI, CRM, and 3+3. We collected 2,447 scenarios generated by NGDF users from April 16, 2014 to September 8, 2016, which are unique in sample size, target toxicity rate $p_T$, cohort size, and true toxicity probability. Figure \ref{fig:ngdf-sc} summarizes the scenarios using six different characteristics. It can be seen the scenarios are quite different in terms of these characteristics. \begin{figure}[htbp] \begin{center} \includegraphics[scale=0.35]{ngdf_scenario_summary.png} \caption{Summary of 2,447 scenarios from the NGDF users. Panels (a)-(e) are the sample size, target toxicity probability $p_T$, cohort size of the trial, the true toxicity probability of the doses in all the scenarios, and the number of doses in all the scenarios, respectively. Panel (f) gives the relative position of each dose in each scenario. Suppose a scenario has $d$ doses, and each dose is indexed by $i$, $i=1, \ldots, d$. Then the relative position of dose $i$ is $i/d$. }\label{fig:ngdf-sc} \end{center} \end{figure} \paragraph{Set two.} We adopted the 42 scenarios in Ji and Wang (2013), which provides 14 scenarios for each of the three $p_T$ values, 0.1, 0.2, and 0.3. Although the number of scenarios is small, they are representative of different dose-response shapes and do not over- or under-represent any particular shapes. See Appendix B for details. \paragraph{Set three.} Following Paoletti et al. (2004), we generated 1,000 scenarios based on a model-based approach. That is, we use an ad-hoc probability model (see below) to generate the scenarios. \begin{description} \item[\textbullet] Assume that there are $d$ doses in the scenario, $d>1.$ Fix the target toxicity probability value $p_T$, say $p_T=0.2.$ \item[\textbullet] With equal probabilities select one of the $d$ doses as the MTD. Without loss of generality, denote this selected dose level by $i$. \item[\textbullet] Generate a random value $p_i = \Phi(\xi_i)$ where $\Phi(\cdot)$ is the cumulative density function (CDF) of the standard normal distribution, and assign $p_i$ to be the toxicity rate of the MTD for this scenario. Here $\xi_i\sim N(\Phi^{-1}(p_T),0.01^{2})$, where $\Phi^{-1}(\cdot)$ denotes the inverse CDF of the standard normal distribution. The term $\xi_i$ perturbs the toxicity probability of the MTD so that it is not identical to $p_T$. \item[\textbullet] Generate the toxicity rates of the two adjacent doses, $p_{i-1}$ and $p_{i+1}$ by $$ p_{i-1} = \Phi \left[\Phi^{-1}(p_i)-\left\{\Phi^{-1}(p_i)-\Phi^{-1}(2p_T-p_i)\right\}\times I\{\Phi^{-1}(p_i)>\Phi^{-1}(p_T)\}-\xi_{i-1}^2 \right] $$ and $$p_{i+1} = \Phi\left[\Phi^{-1}(p_i)+\left\{\Phi^{-1}(2 p_T-p_i)-\Phi^{-1}(p_i)\right\}\times I\{\Phi^{-1}(p_i)<\Phi^{-1}(p_T)\}+\xi_{i + 1}^2\right]$$ where $\xi_{i-1}\sim N(\Phi^{-1}(\mu_1),0.1^{2})$, $\xi_{i+1}\sim N(\Phi^{-1}(\mu_2),0.1^{2})$, and $I(\cdot)$ is an indicator function. \item[\textbullet] Generate the toxicity probabilities iteratively for the remaining dose levels according to $p_{i-2}=\Phi[\Phi^{-1}(p_{i-1})-\xi_{i-2}^2]$ and $p_{i+2}=\Phi[\Phi^{-1}(p_{i+1})+\xi_{i+2}^2]$, and $p_{i-3}=\Phi[\Phi^{-1}(p_{i-2})-\xi_{i-3}^2]$ and $p_{i+3}=\Phi[\Phi^{-1}(p_{i+2})+\xi_{i+3}^2]$, and so on, where $\xi_{i-2},\xi_{i-3},\ldots,\sim N(\Phi^{-1}(\mu_3),0.25^{2})$ and $\xi_{i+2},\xi_{i+3},\ldots,\sim N(\Phi^{-1}(\mu_4),0.25^{2})$. \end{description} We set the values of $\mu$'s such that the average distance at each side of MTD is less than 0.1. This generates a set of ``smooth" scenarios where toxicity gradually increase along the dose levels. See Figure \ref{fig:1000sc} for details. \begin{figure}[htbp] \centering \includegraphics[scale=0.36]{1000_scenarios.png}\\ \caption{Summary of the 1,000 scenarios with $p_T=0.2$ generated from the random scheme in Paoletti et al. (2004). Left: a randomly selected 50 scenarios. Right: the barplots of the true toxicity probability for each of the six dose levels across 1,000 generated scenarios.}\label{fig:1000sc}s \end{figure} \subsection{Results} \paragraph{Definition of true MTD} In the interval designs, because the doses with probability of toxicity falls into the equivalence interval $(p_T-\varepsilon_1, p_T+\varepsilon_2)$ are acceptable MTDs, we provide a clear definition of true MTDs for a given scenario. \begin{enumerate} \item If $(p_T-\varepsilon_1) \leq p_i \leq (p_T+\varepsilon_2)$, then dose $i$ is the true MTD. If more than one dose satisfies the condition, then all of them are considered the true MTDs. \item If there is no $p_i$ meeting condition 1 above, the true MTD is the maximum dose level $i$ of which the true toxicity probability $p_i < p_T$. \item If the MTD could not be identified (e.g., if all the doses have toxicity probabilities $> p_T$), the correct decision is not to select any dose and the true MTD is set as `none'. In other words, selecting any dose as the MTD would be considered as a mistake. \end{enumerate} \paragraph{Results for set one.} We implemented mTPI-2, mTPI, all three BOIN versions, CRM and 3+3 based on the 2,447 crowd-sourcing scenarios. For each scenario, 2,000 simulated trials were conducted on computer. We compared designs' operating characters in terms of two simple metrics, {\it safety} and {\it reliability} (Ji and Wang, 2013). These two metrics capture the most important and fundamental properties of a dose-finding design: patient safety and ability to identify the true MTD. {\it Safety} is the average percentage of the patients treated at or below the true MTD across the simulated trials for a given scenario, and {\it reliability} is the percentage of simulated trials selecting the true MTD for a given scenario. In some literature, {\it reliability} is also called PCS, standing for the percentage of correct selection of the true MTD. Usually there is a tradeoff between the two metrics. That is, a design with better safety is usually associated with less reliability and vice versa. This is because in order to correctly identify the true MTD, a design must accurately infer doses both below and above the MTD, which means assigning some patients to doses above the MTD to learn their high toxicity probabilities. Doing that would lower the {\it safety} metric of the design. \begin{figure}[htbp] \begin{center} \resizebox{0.9\textwidth}{3.5in}{ \begin{tabular}{c} \includegraphics{ngdf_safety.png} \\ \includegraphics{ngdf_reliability.png} \end{tabular} } \caption{Comparison of {\it safety} and {\it reliability} using the 2,447 crowd-sourcing scenario set one. Five designs are compared which are mTPI2, mTPI, BOIN, CRM, and 3+3. Upper panel [{\it Safety}]: each boxplot describes the differences (Design1 $-$ Design2) in the {\it safety} across all 2,447 scenarios. A value greater than zero means Design 1 puts more percentages of patients on doses at or below the true MTD than Design 2. Lower panel [{\it Reliability}]: each boxplot describes the differences (Design1 $-$ Design2) in the {\it reliability} of two designs across all 2,447 scenarios. A value greater than zero means Design 1 is more likely to identify the true MTD than Design 2. }\label{fig:res-rs} \end{center} \end{figure} Figure \ref{fig:res-rs} presents the comparison results. For ease of exposition, we only show results from $\mbox{BOIN}_{\mbox{lambda}}$ so that the interval boundaries among BOIN, mTPI, and mTPI-2 are the same. Results including $\mbox{BOIN}_{\mbox{default}}$ and $\mbox{BOIN}_{\mbox{epsilon}}$ are shown in Appendix A. Overall, mTPI-2 is the safest design, putting more patients on average at doses at or below the MTD than the other four methods. BOIN is slightly worse than the mTPI-2 design on the safety measure, followed by mTPI, CRM, and 3+3. Except for 3+3, all the four model-based designs perform similarly in terms of {\it reliability}, i.e., the probability of identifying the true MTD, with mTPI-2 having slight advantage. The 3+3 design is the worst in the identification of the true MTD. \paragraph{Results for set two.} Figure \ref{fig:res-jco} shows the same comparison of the five designs based on the 42 scenarios (Appendix B) in Ji and Wang (2013), with 14 scenarios for each of the three $p_T$ values, 0.1, 0.2, and 0.3. As can be seen, mTPI-2 is the safest design while CRM edges mTPI in {\it reliability}, followed closely by mTPI-2 and BOIN. Again, the 3+3 design is the worst in both safety and reliability. \begin{figure}[htbp] \begin{center} \resizebox{0.9\textwidth}{3.8in}{ \begin{tabular}{c} \includegraphics{boxplot_jco_safety.png} \\ \includegraphics{boxplot_jco_reliability.png} \end{tabular} } \caption{Comparison of {\it safety} and {\it reliability} using the 42 scenarios in scenario set two. Five designs are compared which are mTPI2, mTPI, BOIN, CRM, and 3+3. Upper panel[Safety]: each boxplot describes the differences (Design1 $-$ Design2) in the {\it safety} across all 42 scenarios. A value greater than zero means Design 1 puts more percentages of patients on doses at or below the true MTD than Design 2. Lower panel[Reliability]: each boxplot describes the differences (Design1 $-$ Design2) in the {\it reliability} of two designs across all 42 scenarios. A value greater than zero means Design 1 is more likely to identify the true MTD than Design 2. }\label{fig:res-jco} \end{center} \end{figure} \paragraph{Results for set three.} The purpose of having set three is to examine the performance of all the designs when scenarios are systematically generated from the stochastic model in Paoletti et al. (2004). For illustration purpose, we assumed $p_T=0.2$, cohort size 1, and six doses per scenario. We simulated 2,000 trials per scenario and presented the pair-wise comparison of the mTPI-2, mTPI, CRM, and BOIN designs using two criteria : 1) PCS: the percentage of trials of correct selection of the true MTD (this is the same as {\it reliability}); 2) Accuracy index for patient allocation (Cheung, 2011): a value $<1$ that measures how accurate patients are allocated to the MTD and equals 1 when all the patients are allocated at the true MTD. Results are presented in Figure \ref{fig:third}. Overall, mTPI-2, CRM and BOIN perform very similarly. As shown in previous work (Horton et al., 201 ) the CRM and BOIN designs perform well in these scenarios. The mTPI design is slightly worse than the other three designs due to its stickiness (Guo et al., 2017) as a result of Occam's razor. The mTPI-2 design presents comparable performance. \begin{figure}[htbp] \centering \begin{tabular}{|c|}\hline\hline \includegraphics[scale=0.31]{PCS.png} \\ Scatter plots of PCS \\ \hline\hline \includegraphics[scale=0.31]{AIsub.png} \\ Scatter plots of Accuracy Index for subjects allocation \\ \hline \end{tabular} \caption{Smooth scatter plots comparing the mTPI-2, mTPI, CRM and BOIN designs using 1,000 randomly generated scenarios according to Paoletti et al. (2004). The red dot correspond to the two mean values of the two methods being compared within each plot. Top panel [PCS]: The percentage of trials of correct selection of the true MTD, which is the same as the reliability criterion. Bottom panel [Accuracy index for subjects allocation]: a criterion $<1$ that measures how accurate patients are allocated surrounding doses near the true MTD. A value closer to 1 means more accurate allocation.}\label{fig:third} \end{figure} \section{Evaluation of Decision Tables}\label{sec:table} As in Section \ref{sec:example} and the vast literature, the gold standard of comparing different dose-finding methods is the operating characteristics (OC) from running a large number of simulated clinical trials on computer, with randomly generated data based on prespecified scenarios that determine the true toxicity probabilities of the candidate doses. While this provides a large-sample (of trials) and population-level assessment on the performance of the methods, it does not evaluate designs in terms of per-trial and per-patient behavior. For example, results in Figures \ref{fig:res-rs} compare the distributions of the safety and reliability values for each design when tested on millions of simulated trials. But for an individual investigator, the trial at hand is of the most consideration. Practically, these OC results are not well understood by the clinical society and are often perceived as a ``black-box'' operation. Also, due to lack of user-friendly tools, it is challenging to generate the OC results for multiple dose-finding designs in practice. One of the key features for the 3+3 design is its transparent nature as all the decision rules can be prespecified and assessed without the need of running simulation. This same feature is also enjoyed by interval-based designs. For this reason, we propose to simply evaluate different designs based on the decision tables that consist of the dose-assignment decisions of each design. This evaluation can be added to the classical OC tables. For example, Figure \ref{fig:mtpi2-table} provides a decision table of mTPI-2 based on $p_T = 0.3$ and $\epsilon_1 = \epsilon_2 = 0.05$ for up to 15 patients. Such a table allows investigators to examine all the possible decisions before the trial starts. We denote such a table as $\{R_{x,n}(\mbox{mTPI-2}, p_T, \epsilon_1, \epsilon_2)\}$ which consists of decisions of mTPI-2 for values of $p_T$, $\epsilon_{1,2}$, and up to $n$ patients and $x \le n$ possible DLT outcomes. Here $R_{x,n}$ takes three values \{D, S, E\} to denote the three up-and-down dose-assignment decisions. All the interval-based designs, mTPI-2, mTPI, and BOIN can provide such a table prior to the trial starts, and one common feature across the three designs is that the table entry $R_{x,n}$ is fixed given $x$ and $n$ for fixed $p_T$ and $\epsilon_{1,2}$. For example, when 1 out of 3 patients experiences DLT, the table in Figure \ref{fig:mtpi2-table} shows the decision S, which is to stay at the current dose. This decision is applied to any dose level and any trial with $p_T=0.3, \epsilon_1=\epsilon_2=0.5$. For CRM, the table $\{R_{x,n}(\mbox{CRM}, p_T)\}$ does not depend on $\epsilon_{1,2}$ since CRM does not require input of these two values. More importantly, each entry $R_{x,n}$ is not fixed but a random variable that follows a probability distribution depending on the probability model in the CRM design and the true toxicity probabilities of the doses. For example, consider a trial targeting $p_T=0.3$. If 1 out of 3 patients experiences DLT, CRM will have a large probability (not equal to 1) to S, stay, but also small probabilities to E or D, depending on the patients data on other doses. Unfortunately, the probability distribution of $R_{x,n}$ cannot be derived analytically as it requires integration of all the possible outcomes of the trials that would have 1 DLT out of 3 patients at any dose. Instead, an investigator can obtain a numerical approximation of the random decision table under CRM based on the same computer simulated trials for generating the operating characteristics, and examine the approximated decision table carefully before proceeding. To see this, we provide an example next. Suppose we want to design a dose-finding trial with six doses, cohort size 3, and MTD target $p_T=0.3$. We apply CRM to the 14 scenarios in Ji and Wang (2014) with $p_T=0.3$ (Appendix B) and sample size 51. We simulate 10,000 trials per scenario on computer using the R package {\tt dfCRM} (\url{https://cran.r-project.org/web/packages/dfcrm/index.html}). We then tabulate the frequencies of the decisions D, S, or E that CRM takes whenever $x$ out of $n$ patients experience DLT in the simulated trials. These frequencies are the empirical distribution of the random decision $R_{x,n}$ under CRM since they are integrated over the data from a large number of simulated trials under various scenarios. In other words, this can be considered as a Monte Carlo approximation of the true distribution of $R_{x,n}(\mbox{CRM}, 0.3).$ Note that the actual CRM decisions specify the exact dose level for the next cohort of patients, but these decisions can be easily converted to D, S, or E based on the decided dose level and the current dose level. Therefore $R_{x,n}$ only takes D, S, or E, and we obtain three proportions $\{q_{x,n}(D), q_{x,n}(S), q_{x,n}(E)\}$ for each $x$ and $n$, where $q_{x,n}(D) + q_{x,n}(S) + q_{x,n}(E)=1.$ The results are summarized as a decision table in Figure \ref{fig:crm-table} for up to 15 patients. We removed results for more than 15 patients for ease of exposition. This table allows investigators to examine the specific CRM method (in our case, the CRM is the one from the dfCRM R package) and its performance under the specific trial setting ($p_T=0.3$, cohort size 3, and sample size 51) using a specific set of scenarios. For example, it can be seen from the table that more than half of the times the CRM design would stay when 1 out of 3 patients experienced DLT at a given dose (the yellow segment being longer than the blue segment in the bar for 1 DLT and 3 patients). When 2 out of 3 patients experienced DLT, about half of the times CRM would de-escalate but the other half of the times it would stay. \color{black} \rm These percentages various decisions are important for investigators and review boards to assess the performance of CRM and any designs being considered. And importantly, they are more intuitive and transparent than the OC results. As a comparison, we also obtain the fixed decision tables (not shown) for a sample size of $n=51$ patients for the mTPI-2 and BOIN$_{\mbox{lambda}}$ designs given $p_T=0.3$ value and a pair of $\epsilon_{1,2}$ values, denoted as $\{R_{x,n}(\mbox{mTPI2}, p_T, \epsilon_1, \epsilon_2)\}$ and $\{R_{x,n}(\mbox{BOIN}, p_T, \epsilon_1, \epsilon_2)\}$. Since these tables not only depend on $p_T$ values, but also $\epsilon_{1,2}$, we vary both $\epsilon$'s from 0.005 to 0.05. For CRM, for each $(x, n)$ value we obtain three proportions $\{q_{x,n}(D), q_{x,n}(S), q_{x,n}(E)\}$ which represent the proportions of the three decisions. As a summary, we compute the mean decision score defined as $E(R_{x,n}, p_T)=1*q_{x,n}(E) + 2*q_{x,n}(S) + 3*q_{x,n}(D)$ for each $x$ and $n$ value, where values $1$, $2$, and $3$ are assigned to decision $E$, $S$, and $D$, respectively. Therefore, a higher value is associated with a more conservative decision. Figure \ref{fig:decision-diff} shows the sum of the differences (Design1 - Design2) of $R$ for mTPI-2 and BOIN and $E(R)$ for CRM between any two designs for each of three $p_T$ values, $0.1, 0.2,$ and $0.3$. For example, the top left panel presents the differences $\sum_{n=1}^{51}\sum_{x=1}^n \{R_{x,n}(\mbox{mTPI-2}, 0.1,, \epsilon_1, \epsilon_2) - R_{x,n}(\mbox{BOIN}, 0.1, \epsilon_1, \epsilon_2)\}$ for all the $\epsilon_{1,2}$ values. A positive or negative value in the plot implies that Design1 is more likely to de-escalate or escalate, respectively. Overall mTPI-2 appear to have more de-escalation decisions than the other two designs in most cases, and the two interval-based designs do not differ much in general. The differences of the two designs get larger for larger $\epsilon$ values. CRM is in general more aggressive in its decisions than mTPI-2 and BOIN. However, when $\epsilon_2$ is large and $\epsilon_1$ is small, the equivalence interval $(p_T - \epsilon_1, p_T + \epsilon_2)$ becomes more skewed towards right, which means mTPI-2 and BOIN will be less likely to de-escalate and more likely to escalate. Therefore, we see the decisions of mTPI-2 and BOIN are more aggressive than CRM in these $\epsilon$ values (upper left corner of the heatmaps in Figure \ref{fig:decision-diff}). \begin{figure}[htbp] \centering \includegraphics[scale=0.35]{mTPI2_decision.png} \caption{A decision table under the mTPI-2 design for $p_T=0.3$ and $\epsilon_1=\epsilon_2=0.05.$ Each column represents $n$ number of patients treated at the current dose and each row represents $x$ number of patients with DLTs. }\label{fig:mtpi2-table} \end{figure} \begin{figure}[htbp] \centering \includegraphics[scale=0.45]{fig7a.png} \caption{A decision table under the CRM design for $p_T=0.3$. Each column represents $n$ number of patients treated at the current dose and each row represents $x$ number of patients with DLTs. For each entry, the decisions E, S, and D are colored blue, stay, and purple, respectively. The length of the colored segments within each bar are proportional to the three proportions $(q_{x,n}(E), q_{x,n}(S),q_{x,n}(D))$ of the three decisions taken by CRM for a given $(x, n)$ data point from a simulation study using the 14 scenarios (Ji and Wang, 2013) and 10,000 simulated trials per scenarios. }\label{fig:crm-table} \end{figure} \begin{figure}[htbp] \hspace{-0.2in} \begin{tabular}{ccc} \includegraphics[scale=0.13]{mtpi2_boin_pt01.png} & \includegraphics[scale=0.13]{mtpi2_crm_pt01.png} & \includegraphics[scale=0.13]{boin_crm_pt01.png} \\ \includegraphics[scale=0.13]{mtpi2_boin_pt02.png} & \includegraphics[scale=0.13]{mtpi2_crm_pt02.png} & \includegraphics[scale=0.13]{boin_crm_pt02.png} \\ \includegraphics[scale=0.13]{mtpi2_boin_pt03.png} & \includegraphics[scale=0.13]{mtpi2_crm_pt03.png} & \includegraphics[scale=0.13]{boin_crm_pt03.png} \\ \end{tabular} \caption{Pairwise difference decision tables of the mTPI-2, CRM and BOIN$_{\mbox{lambda}}$ designs for different $p_T$ values of 0.1, 0.2, and 0.3, and different $\epsilon_{1,2}$ values. The decisions E, S, and D are coded as 1, 2, and 3, respectively. Each table entry is the sum of the differences (Design1 - Design2) of the decisions $R_{x,n}(Design1,p_T,\epsilon_1, \epsilon_2) - R_{x,n}(Design2,p_T,\epsilon_1, \epsilon_2)$ between two designs for $n=51$ patients. For CRM, we use $E(R_{x,n}(CRM, p_T))$ in the calculation. A positive/negative value or green/red color means Design1 is more likely to de-escalate/escalate. }\label{fig:decision-diff} \end{figure} \section{Discussion} We show using crowd-sourcing results that mTPI, mTPI-2, and BOIN designs are superior than 3+3 and compare well with other non-interval designs, such as CRM. Due to the simplicity in the implementation of the interval designs, we recommend the use of these designs for practical trials. In particular, mTPI-2 seems to stand out with its rigorous theoretical development, superior numeric performance, simplicity in the implementation, and adherence to the ethical principle in its dose escalation decisions. The interval designs can be categorized as the iDesigns, including TPI, mTPI, and mTPI-2, and IB-Designs including CCD and BOIN. They differ in the statistical inference and underlying decision strategies. Interestingly, they are very similar in implementation which produces a fixed decision table and is attractive to practitioners due to its simplicity and transparency. The use of deterministic decisions allow their decision tables to be calculated and examined prior to the trial onset, allowing investigators to see the behavior of the designs given various patient outcomes in terms of $x$ number of DLTs out of $n$ patients treated at a dose. The mTPI-2 design appears to have the best overall performance in our comparison using the the 42 scenarios in Ji and Wang (2013). The performance of mTPI-2 is similar to BOIN on overall reliability or PCS using the randomly generated scenarios. However, the decision tables give clear contrast that mTPI-2 is a safer design. We also show that running a large number of computer simulated clinical trials and comparing operating characteristics of the designs is not sufficient to fully assess dose-finding designs in practice. We recommend to examine the decision tables, even for methods like CRM when its decision tables are not deterministic. Since a large number of simulated trials are usually conducted on computer as a standard practice, one can quickly generate the empirical distributions of the CRM decisions $R_{x,n}$ as in Figure \ref{fig:crm-table} to assess the performance of CRM for a given pair of $(x, n)$ values. Decision tables and operating characteristics tables will jointly allow investigators to evaluate the population-level (across many simulated trials) and individual-level (across patients) behavior of a design. Therefore, both of them should be reviewed when these designs are considered for a practical trial. \section*{Acknowledgement} This is a collaborative research with Dr. Sue-Jane Wang from US FDA. She provided numerous insightful critiques, did some technical investigations and gave constructive inputs throughout the research. We sincerely thank her for her professional expertise and scientific collaboration. \bibliographystyle{apa}

Extragonadal teratomas are thought to arise from misplaced, pluripotential, primordial germ cells. Childhood teratomas occur primarily in extragonadal sites, mainly the spine, head, and neck while most adult teratomas are gonadal. The head and neck location is the second most common site (after the sacrococcygeal region), accounting for 5% to 14% of cases in infancy.[@b1-asm-6-623],[@b2-asm-6-623] The orbit is a rare site either primarily or secondarily as an intracranial extension. Females are approximately doubly affected compared to the males, and there is a propensity of the tumor to occur more frequently in the left orbit.[@b3-asm-6-623] CASE ==== A 5-month-old female healthy baby referred as a case of left orbital cellulitis for the drainage of an orbital abscess following treatment with intravenous antibiotics with no improvement. She had a history of sudden onset proptosis of 1-week duration. Examination showed that she was able to fix and follow objects with her right eye while fixing with limitation of movements on the left. There was left eye proptosis, severe eyelid swelling, conjunctival chemosis, and exposure keratopathy ([Figure 1a](#f1-asm-6-623){ref-type="fig"}). The fundus view was hazy. She was otherwise afebrile and healthy apart from anemia and mild leukocytosis. The B-scan showed a normal left globe and large diffuse orbital lesion of low reflectivity pushing the globe forward. Magnetic resonance imaging (MRI) showed multicompartmental cystic collection with fluid signal intensity, slightly higher signal intensity on T1-weighed images, and marginal rim of enhancement occupying the left orbital cavity ([Figure 1b](#f1-asm-6-623){ref-type="fig"}) in addition to sinusitis. The initial clinical and radiological diagnosis was orbital cellulitis complicated by abscess formation. She was admitted for intravenous gentamicin and cefazolin with drainage of intraconal and extraconal abscess under general anesthesia. The patient was discharged home following initial improvement. Underlying orbital pathologic lesion was suspected on her follow-up and was readmitted two months later. Her systemic evaluation revealed persistent anemia. Her eye examination showed less left eye proptosis with conjunctival prolapse and faint corneal scar ([Figure 1c](#f1-asm-6-623){ref-type="fig"}). Her repeated MRI revealed left orbital extraconal and inferonasal multicompartmental cystic lesion with a marginal contrast enhancement. The radiologic diagnosis confirmed a possible hematic cyst. Orbital exploration showed a huge mass with cystic component, which was excised. The histopathologic findings showed incomplete irregular cystic cavities lined by focally ciliated pseudostratified columnar epithelium with goblet cells representing the respiratory type of epithelium ([Figure 2a](#f2-asm-6-623){ref-type="fig"}), adjacent epidermal component with skin including skin appendages ([Figure 2b](#f2-asm-6-623){ref-type="fig"}), cartilage, adipose tissue ([Figure 2c](#f2-asm-6-623){ref-type="fig"}) and smooth muscle fibers. Several mucoserous, apocrine, and sweat glands were seen ([Figure 2d](#f2-asm-6-623){ref-type="fig"}). The final diagnosis was a mature teratoma. The patient was followed up with no recurrence for a period of 16 months; however, her follow-up examination revealed left esotropia due to documented amblyopia in the left eye, which was managed by patching. DISCUSSION ========== Extragonadal teratomas arise from pluripotential primodial germ cells that are thought to become misplaced, or have aberrant migration to several structures near the midline (such as in the head, mediastinum, or sacrococcygeal region). Another hypothesis suggests an origin from ordinary somatic cells with full genetic code. Theoretically, these cells can produce any other type of cell without being a "germ cell."[@b1-asm-6-623] Holme reported the first case of orbital teratoma in 1862.[@b4-asm-6-623] Since then several cases were reported.[@b3-asm-6-623],[@b5-asm-6-623]--[@b7-asm-6-623] The review of published reports indicated the preponderance of left-sided orbital teratomas and a female-to-male ratio of 2:1.[@b8-asm-6-623] Our patient was a male with left orbital involvement. Childhood teratomas occur primarily in extragonadal locations, and the majority of head and neck teratomas were identified during the first year of life or even at birth, particularly the ones affecting the orbit.[@b1-asm-6-623],[@b2-asm-6-623] Orbital teratomas in contrast to the frequent orbital dermoid cysts are extremely rare.[@b8-asm-6-623] Orbital teratomas are usually benign and well differentiated; however, they may cause facial deformity due to rapid growth as well as destructive proptosis and exposure keratopathy.[@b8-asm-6-623] Proptosis at birth due to a congenital teratoma is unusual, but has been reported.[@b7-asm-6-623],[@b9-asm-6-623] Our patient presented with rapidly increasing proptosis over 1 week at the age of 5 months. An extension into the orbit from an adjacent space such as paranasal sinuses, cavernous sinus, or pterygopalatine fossa can also occur. In this case the presentation can be delayed and an associated airway obstruction may be encountered.[@b10-asm-6-623] Teratomas can extend into the periorbital sinus or through the superior orbital fissure, which can be further complicated by intracranial extension of the tumor.[@b3-asm-6-623],[@b11-asm-6-623] In addition, orbital teratomas can present with recurrent orbital cellulitis similar to our case, which might delay the diagnosis.[@b12-asm-6-623] Histopathologically, orbital teratomas are mostly benign and well differentiated. A total of 90% of teratomas contain all 3 embryonic layers: ectoderm, mesoderm, and neural tissue. Between 20% to 40% of childhood teratomas contain some immature tissues. [@b1-asm-6-623],[@b2-asm-6-623] Overall, neuroectodermal elements "both mature and immature" frequently dominate the childhood teratoma, particularly in the head and neck region compared with the sacrococcygeal teratomas. It was reported that up to 95% of childhood teratomas have some neuroectodermal elements.[@b2-asm-6-623] Such teratomas may be confused with primitive neuroectodermal tumors (PNETs), since immature teratomas might include regions that resemble the developing neural tube or cells forming rosettes, similar to those seen in PNETs, neuroblastoma, and medulloblastoma.[@b2-asm-6-623] Benign teratomas may contain mature tissues or smaller amounts (less than 50%) of immature elements.[@b13-asm-6-623] Immature neural tissue in a young-child teratoma does not reduce the prognosis; however, when they are found in an adult, they are usually associated with a malignant behavior and poor outcome. [@b2-asm-6-623] Malignant extragonadal teratomas usually have carcinomatous features. They constitute a small number and tend to occur in older children. Few teratomas also show malignant degeneration over time.[@b2-asm-6-623],[@b12-asm-6-623] Orbital teratomas are heterogenous complex lesions containing adipose tissue and occasional bone formation. Calcification can be dystrophic, but may also occur as mineralization of osteoid matrix (true bone), cartilage or well-formed teeth.[@b12-asm-6-623],[@b13-asm-6-623] Therefore, preoperative cross-sectional imaging is crucial, and computed tomography (CT) is required to define bony anatomy. Teratomas appear as multiloculated, cystic masses and solid areas within a well-delineated mass showing the areas of low attenuation and high signal intensity (representing lipid) on CT and MRI. This was observed in our case with slightly higher signal intensity on T1-weighted images.[@b7-asm-6-623],[@b12-asm-6-623],[@b14-asm-6-623] The differential diagnosis for an orbital mass with low attenuation includes vascular lesions, microphthalmia with cyst, coloboma and congenital cystic eye, epidermoid cyst, dermoid cyst, and meningoencephalocele.[@b15-asm-6-623] The clinical management of these lesions is unclear, due in part to their low incidence. Early surgery is now recommended aiming at complete surgical excision, thus saving the eye if possible. This is essential to avoid the risks of rapid growth resulting in necrosis, infection, hemorrhage, or spontaneous rupture.[@b7-asm-6-623],[@b10-asm-6-623],[@b13-asm-6-623],[@b16-asm-6-623] Orbital teratomas may rarely contain nongerm cell malignancies, such as neuroblastoma that has been reported, and in such cases, decision regarding adjunctive chemotherapy or radiotherapy should be based on the pathologic findings for the individual case.[@b17-asm-6-623] Teratomas with a malignant change are resistant to chemotherapy, and this explains the lack of initial response to chemotherapy and radiotherapy.[@b13-asm-6-623] The prognosis of teratoma is related to several factors including the age of the patient site, immaturity of the tumor, and extent of the teratoma.[@b10-asm-6-623] The prognosis is usually good; however, a complete surgical removal might be difficult and postoperative complications such as orbital asymmetry, ocular injury, and loss of vision can occur.[@b7-asm-6-623],[@b13-asm-6-623] Some reports have related the early surgical intervention to preservation of visual acuity.[@b5-asm-6-623],[@b18-asm-6-623] The eye-conserving approach might necessitate decompressing the tumor before dissecting it from the periorbital tissues if the tumor has a major cystic component.[@b19-asm-6-623] In conclusion, ophthalmologists should be aware of the occurrence of orbital teratomas that have variable heterogenous histologic characteristics by their nature. Both the pathologic and radiologic diagnosis of a teratoma may be suspected when a complex lesion contains fatty regions and calcifications. Early diagnosis is helpful to aim at a complete surgical excision of the lesion and to avoid the disadvantages of partial resection such as regrowth of the tumor and the potential for a malignant change at a later date.[@b10-asm-6-623],[@b20-asm-6-623] {#f1-asm-6-623} {#f2-asm-6-623}

\section{Introduction} The Klein-Gordon equation is a fundamental physical equation describing the motion of a spin-less particle, see for instance \cite{BjoD65}, \cite{Sak67} and \cite{Bog59} for a physical introduction on this topic and \cite{GinVe85}, \cite{BreWa81} and the references therein for the mathematical analysis, and in particular \cite{Bou96,Bam03} for some long time results when the equation is set on a compact domain, the case mainly considered in the present work. In the relativistic regime, i.e. the speed of light $c=1$, the Klein-Gordon equation (as a non-linear wave equation) is numerically well studied, see for instance \cite{StVaz78}, \cite{JimVaz78}, \cite{PasJimVaz95}, and, for long time behavior results in the weakly nonlinear regime, \cite{CoHaLu08}, \cite{FG11} and \cite{FGP1,FGP2}. Here we are interested in the Klein-Gordon equation in the non-relativistic limit regime, i.e. the speed of light $c$ tending to infinity. Thus, our model problem reads \begin{equation} \label{eq:kgr} \frac{1}{c^2} \partial_{tt} z - \Delta z + c^2 z = f(z), \quad z(0) = \phi, \quad \partial_t z(0) = c^2 \gamma,\quad c\gg 1, \end{equation} where we will consider essentially the case $f(z) =\lambda |z|^{2p} z$, $\lambda \in \mathbb{R} $, assume that the initial condition $\varphi$ and $\gamma$ do not depend on $c$, and consider periodic boundary conditions in the space variable. In this regime the Klein-Gordon equation has recently gained a lot of attention, see for instance \cite{MaNakOz02}, \cite{MaNak02} and \cite{Tsu84} (when Equation \eqref{eq:kgr} is set on the whole space $ \mathbb{R} ^d$). Numerically, the highly oscillatory nature of the exact solution is very challenging as standard numerical schemes require severe time step restrictions depending on the large parameter $c^2$, see for instance \cite{HLW}, \cite{EFHI09} and \cite{PetzJ97} for a numerical overview on highly oscillatory problems. In the recent work \cite{BaoD12}, Gautschi-type exponential integrators (proposed in \cite{HoLu99} for equations with high frequencies generated by the linear part) were applied to the non-relativistic Klein-Gordon equation, which allow time steps of order $\mathcal{O}(c^{-2})$. Following the one-term asymptotics results given in \cite{MaNak02} we study in this paper the existence of a complete asymptotic expansion of the exact solution of \eqref{eq:kgr} in terms of $c^{-2}$ over a $c$ -independent interval $(0,T)$, and for smooth initial data. This will allow us to construct numerical schemes which do not need to obey any $c$-dependent smallness restriction. In a nutshell, the asymptotic expansion up to the first-order correction term of \eqref{eq:kgr} reads \begin{equation}\label{eq:asympz} z(t,x) = \textstyle \frac12 u_0(t,x) \mathrm{e}^{ic^2t} + \textstyle \frac12 \overline{v_0(t,x)} \mathrm{e}^{-ic^2t} + \mathcal{O}(c^{-2}),\qquad t\in [0,T] \end{equation} where $u_0$ and $v_0$ solve the coupled nonlinear Schr\"odinger equations \begin{equation}\label{eq:nsl0} \begin{aligned} &i \partial_t u_0 -\frac{1}{2}\Delta u_0 = \frac{1}{2\pi}\int_0^{2\pi} f(\textstyle \frac{1}{2}(u_0+\mathrm{e}^{-2i\theta}\overline{v_0}))\,\mathrm{d}\theta,\qquad u_0(0) = \varphi - i \gamma, \\ &i \partial_t v_0 -\frac{1}{2}\Delta v_0 = \frac{1}{2\pi}\int_0^{2\pi} f(\textstyle\frac{1}{2}(v_0+\mathrm{e}^{-2i\theta}\overline{u_0}))\,\mathrm{d}\theta,\qquad v_0(0) = \overline{\varphi + i \gamma}. \end{aligned} \end{equation} The essential point is that the highly-oscillating part in the asymptotic expansion \eqref{eq:asympz} is only contained in the phases $\mathrm{e}^{ic^2t}$ and $\mathrm{e}^{-ic^2t}$, but does not appear in the above Schr\"odinger equations for $u_0$ and $v_0$. In particular the numerical time integration of \eqref{eq:nsl0} can be carried out without any $c$-dependent time step restriction. Thus, with this asymptotic expansion we are able to obtain a well suited numerical approximation to the exact solution $z(t)$ of the Klein-Gordon equation in the non-relativistic limit $c\gg 1$, without any time step restriction as we only need to solve the Schr\"odinger equations \eqref{eq:nsl0} and multiply the numerical approximations of $u_0$ and $v_0$ with the highly-oscillatory phases $\mathrm{e}^{ic^2t}$ and $\mathrm{e}^{-ic^2t}$, respectively. Note that in comparison with \cite{MaNak02} and motivated by numerical implementation, we choose to work here with periodic boundary conditions, that is the space variable $x \in \mathbb{T} ^d$ the $d$-dimensional torus. Moreover, we show that for sufficiently smooth initial values the asymptotic expansion of the exact solution to \eqref{eq:kgr} exists up to an arbitrary order, and can be reconstructed by superposing highly oscillatory terms to cascaded solutions of $c$-independent Schr\"odinger-like systems. After a full discretization of these cascaded terms, we are able to build numerical schemes which approach the exact solution up to errors terms of order $\mathcal{O}(\tau^k+c^{-2N-2} + h^{s})$ by simulating the terms in the asymptotic expansion up to order $N$ with a numerical scheme of order $k$ in time (with step size $\tau$) and Fourier pseudo-spectral discretization on a grid with mesh size $h$. Note that here, $h$ and $\tau$ can be chosen independently of $c$, and that $s = s(N)$ depends on the regularity of the solution and of the degree of the asymptotic expansion. The paper is organized as follows. We commence in Section \ref{sec:boundi} with the essential a priori bounds on the exact solution of the Klein-Gordon equation \eqref{eq:kgr} showing the existence on a time interval independent of the large parameter $c$. In Section \ref{sec:asymp} we will derive an asymptotic expansion of the exact solution in powers of $c^{-2}$. In Section \ref{sec:firstt} we will precisely state the first terms in this expansion, which will allow us to construct numerical schemes which are robust with respect to $c$, see Section \ref{sec:nume}. Furthermore, in Section \ref{sec:enq} we analysis the asymptotic behavior of the conserved physical quantities. \section{A priori bounds on the exact solution}\label{sec:boundi} We consider our model problem \eqref{eq:kgr} set on the $d$-dimensional torus $ \mathbb{T} ^d$. We make the following assumptions: \begin{ass} The initial values $\varphi$ and $\gamma$ are smooth functions independent of c. Furthermore we assume that the nonlinearity $f$ is gauge invariant and polynomial of degree $r>1$ in $z$ and $\overline{z}$. This means that for all number $\alpha \in \mathbb{R} $ and all $z \in \mathbb{C}$, we have $f(e^{i\alpha}z) = e^{i\alpha} f(z)$. Eventually, we assume that $f$ is {\em real} in the sense that for all $z \in \mathbb{C}$, we have $\overline{f(z)} = f(\overline{z})$. In particular, this shows that if $z(0)$ and $\partial_t z(0)$ are in $ \mathbb{R} $, then the solution $z(t)$ is real for all times. Typical examples of such nonlinearity are given by the polynomials $f(z) = \lambda |z|^{2p} z$, $p \geq 1$. \end{ass} For ease of presentation and numerical motivations, we will mainly work on the $d$-dimensional torus $ \mathbb{T} ^d$, $d \in \mathbb{N} \backslash\{0\}$. However, as the only arguments used in the sequel are the local Lipschitz nature of the nonlinearity in Sobolev space and some explicit calculations for diagonal operators in Fourier, the reader should be convinced that all the result of this paper can be easily extended to functions defined on $ \mathbb{R} ^d$. For a function $u(x)$, $x \in \mathbb{T} ^d$ in $H^s := H^s( \mathbb{R} ,\mathbb{C})$ we set $$ \Norm{u}{H^s}^2 = \sum_{k \in \mathbb{Z} ^d} (1 + |k|)^{2s} |\hat u_k|^2, $$ where $$ \hat u_k = \frac{1}{(2\pi)^d} \int_{ \mathbb{T} ^d} e^{-i k \cdot x} u(x) \text{d} x, $$ $k \in \mathbb{Z} ^d$, denote the Fourier coefficients associated with $u$. Here, we have set $$ k\cdot x = k_1 x_1 + \cdots + k_d x_d\quad \mbox{and}\quad |k|^2 = k_1^2 + \cdots+ k_d^2, $$ for $k = (k_1,\ldots,k_d) \in \mathbb{Z} ^d$ and $x = (x_1,\ldots,x_d) \in \mathbb{T} ^d$. For a given $c > 0$, we define the operator $$ \langle \nabla \rangle_c = \sqrt{- \Delta + c^2}. $$ In Fourier, the operator $\langle \nabla \rangle_c$ is represented by the diagonal operator with coefficients $$ (\langle \nabla \rangle_c)_{k\ell} = \delta_{k\ell} \sqrt{|k|^2 + c^2}, \quad k,\ell \in \mathbb{Z} ^d, $$ where $\delta_{k\ell}$ denotes the Kronecker operator. In particular note that the operator $c \langle \nabla \rangle_c^{-1}$ is uniformly bounded with respect to $c$: we have for all $k \in \mathbb{Z} ^d$, $|(c \langle \nabla \rangle_c^{-1})_{kk}| \leq 1$, which implies that for all functions $u$, $\Norm{c \langle \nabla \rangle_c^{-1}u}{H^s} \leq \Norm{u}{H^s}$. With these notations, Equation \eqref{eq:kgr} can be written as \begin{equation}\label{eq:kgrr} \partial_{tt} z + c^2 \langle \nabla \rangle_c^2 z = c^2 f(z). \end{equation} In order to rewrite the above equation as a first-order system in time, we set \begin{equation}\label{eq:uv} u = z - i c^{-1}\langle \nabla \rangle_c^{-1} \partial_t z , \quad v = \overline z - i c^{-1}\langle \nabla \rangle_c^{-1} \partial_t \overline z. \end{equation} Note that we have \begin{equation}\label{eq:zuv} z = \frac12 (u + \overline{v}), \end{equation} and if $z$ is real, then $u = v$. A short calculation shows that in terms of the variables $u$ and $v$ equation \eqref{eq:kgrr} reads \begin{equation} \label{eq:NLSc} \begin{array}{rcl} i \partial_t u &=& -c\langle \nabla \rangle_c u + c\langle \nabla \rangle_c^{-1} f(\textstyle \frac12 (u + \overline v)), \\[2ex] i \partial_t v &=& -c\langle \nabla \rangle_c v + c\langle \nabla \rangle_c^{-1} f(\textstyle \frac12 (\overline u + v)). \end{array} \end{equation} The initial boundary conditions associated with this problem are (see \eqref{eq:kgr}) \begin{equation} \label{eq:BCc} u(0,x) = \varphi(x) - i c \langle \nabla \rangle_c^{-1}\gamma(x), \quad \mbox{and}\quad v(0,x) = \overline{\varphi(x)} - i c \langle \nabla \rangle_c^{-1} \overline{ \gamma(x)}. \end{equation} Eventually, we set \begin{equation} \label{eq:rlb} \boldsymbol{w} = \begin{pmatrix} u \\[1ex] v \end{pmatrix}, \quad \boldsymbol{F}(\boldsymbol{w}) = \begin{pmatrix} f(\textstyle \frac12 (u + \overline v)) \\[1ex] f(\textstyle \frac12 (\overline u + v)) \end{pmatrix}, \quad \mbox{and}\quad \boldsymbol{\psi} = \begin{pmatrix} \varphi - i c \langle \nabla \rangle_c^{-1} \gamma \\[1ex] \overline{\varphi} - i c \langle \nabla \rangle_c^{-1} \overline{\gamma} \end{pmatrix}, \end{equation} so that the problem \eqref{eq:NLSc}-\eqref{eq:BCc} can be written as \begin{equation} \label{eq:NLF} i \partial_t \boldsymbol{w} = - c \langle \nabla \rangle_c \boldsymbol{w} + c \langle \nabla \rangle_c^{-1} \boldsymbol{F}(\boldsymbol{w}), \quad \boldsymbol{w}(0) = \boldsymbol{\psi}. \end{equation} We see that this equation is posed on $ H^s \times H^s$. For an element $\boldsymbol{w} = (u,v)^T \in H^s\times H^s$, we set $$ \Norm{\boldsymbol{w}}{H^s}^2 := \Norm{u}{H^s}^2 + \Norm{v}{H^s}^2. $$ In the previous and in the following, if an operator - like $\langle \nabla \rangle_c$ - acts on $H^s$, we naturally extend diagonally its action on $H^s \times H^s$ by setting for instance for $\boldsymbol{w} = (u,v)^T$, $$ \langle \nabla \rangle_c \boldsymbol{w} = \langle \nabla \rangle_c \begin{pmatrix} u \\ v \end{pmatrix} := \begin{pmatrix} \langle \nabla \rangle_c u \\ \langle \nabla \rangle_c v \end{pmatrix}. $$ The solution of the linear equation ($F = 0$ in the previous equation) can be written as follows: For $\boldsymbol{w} = (u,v) \in H^{s} \times H^s$, $$ \boldsymbol{w}(t) = e^{i t c \langle \nabla \rangle_c} \boldsymbol{\psi}, $$ where the right hand-side is easily defined in Fourier by the formula $$ \forall\, k \in \mathbb{Z} ^d, \quad \widehat{\big(e^{i t c \langle \nabla \rangle_c} \boldsymbol{\psi}\big)}_k = e^{i t c \sqrt{|k|^2 + c^2}} \, \widehat \boldsymbol{\psi}_k, $$ where $\widehat \boldsymbol{\psi}_k$ denote the (vector valued) Fourier coefficients of $\boldsymbol{\psi}$. Note that the linear flow is an isometry in the sense that for $\boldsymbol{\psi} \in H^s\times H^s$, $$ \Norm{e^{it c \langle \nabla \rangle_c} \boldsymbol{\psi}}{H^s} = \Norm{\boldsymbol{\psi}}{H^s}. $$ Note that in the sequel, and because the equation is set on the torus, we will not use any argument involving Strichartz estimates as in \cite{MaNak02} but we will require more regularity on the initial conditions $\gamma$ and $\varphi$. As we assume that the function $f$ is polynomial of degree $r > 2$ in $z$ and $\bar z$ the application $\boldsymbol{F}$ satisfies estimates of the form \begin{equation} \label{eq:lips} \Norm{\boldsymbol{F}(\boldsymbol{w}_1) - \boldsymbol{F}(\boldsymbol{w}_2)}{H^s} \leq C( 1 + \Norm{\boldsymbol{w}_1}{H^s} + \Norm{\boldsymbol{w}_2}{H^s})^{r-1} \Norm{\boldsymbol{w}_1 - \boldsymbol{w}_2}{H^s} \end{equation} for $s > d/2$, with some constants $C$, and for $\boldsymbol{w}_1$ and $\boldsymbol{w}_2$ in $H^s \times H^s$. These estimates are consequences of classical bilinear estimates in $H^s$: \begin{equation} \label{eq:bil1} \Norm{uv}{H^s} \leq C_s \Norm{u}{H^s}\Norm{v}{H^s}\quad\mbox{for}\quad s > d/2, \end{equation} for some constants $C_s$. Moreover the assumption that $f$ is real implies that for all $w \in H^s\times H^s$, $\overline{\boldsymbol{F}(\boldsymbol{w})} = \boldsymbol{F}(\overline{\boldsymbol{w}})$. With these notations, the mild formulation of \eqref{eq:NLF} is given by $$ \boldsymbol{w}(t) = e^{i t c \langle \nabla \rangle_c} \boldsymbol{\psi} + \int_{0}^t c \langle \nabla \rangle_c^{-1} e^{i (t-s) c \langle \nabla \rangle_c} \boldsymbol{F}(\boldsymbol{w}(s)) \text{d} s. $$ Using the fact that the linear flow $e^{i t c \langle \nabla \rangle_c}$ is an isometry in $H^s$ and that $c \langle \nabla \rangle_c^{-1}$ is a bounded operator in $H^s$, in particularly uniformly bounded with respect to $c$, we easily derive the following result by a standard fix point argument combined with \eqref{eq:lips}: \begin{proposition} \label{prop21} Let $s > d/2$ and $M > 0$ a given constant. Then there exist constants $T_s>0$ and $B_s >0$ such that for all $c > 0$ and for all $\boldsymbol{\psi}\in H^s \times H^s$ such that $\Norm{\boldsymbol{\psi}}{H^s} \leq M$, there exists a mild solution $t \mapsto \boldsymbol{w}(t) \in \mathcal{C}^1((0,T_s), H^s \times H^s)$ to the equation \eqref{eq:NLF}, and such that $$ \forall\, t \in (0,T_s), \quad \Norm{\boldsymbol{w}(t)}{H^s} \leq B_s. $$ \end{proposition} In terms of $u$ and $v$, the previous proposition can be reformulated as \begin{corollary} \label{cor21} Let $s > d/2$, and assume that $\varphi \in H^s$ and $\gamma\in H^s$ are given. Then there exist constants $T_s>0$ and $B_s >0$ such that for all $c$ there exists a mild solution $(u(t), v(t)) \in \mathcal{C}^1((0,T_s), H^s \times H^s)$ to the equation \eqref{eq:NLSc} satisfying the initial boundary conditions \eqref{eq:BCc} and such that $$ \forall\, t \in (0,T_s), \quad \Norm{u(t)}{H^s} + \Norm{v(t)}{H^s} \leq B_s. $$ \end{corollary} \begin{proof} Using the expression \eqref{eq:rlb} for $\boldsymbol{\psi}$ in terms of $\varphi$ and $\gamma$, and the fact that $c \langle \nabla \rangle_c^{-1}$ is uniformly bounded in with respect to $c$, we obtain that $\Norm{\boldsymbol{\psi}}{H^s} \leq M$, a constant independent of $c$. This shows that Proposition \ref{prop21} applies. \end{proof} This corollary immediately gives the existence of the exact solution $z(t)$ of \eqref{eq:kgr} on a time interval $(0,T_s)$ and with $H^s$-bounds both independent of $c$. \section{Asymptotic expansion}\label{sec:asymp} On the $c$-independent time interval $(0,T_s)$ given by the previous results, we search $w$ (at least formally) under the form $$ \boldsymbol{w}(t,x) = \boldsymbol{\mathsf{W}}(t,c^2 t,x) = \begin{pmatrix} \mathsf{U}(t,c^2 t,x) \\[1ex] \mathsf{V}(t,c^2t,x) \end{pmatrix}, $$ where we separate the high oscillations from the slow time dependency of the solution. Here, $\boldsymbol{\mathsf{W}}(t,\theta,x)$ is a function defined on $ \mathbb{R} \times \mathbb{T} \times \mathbb{T} ^d$ and taking values in $\mathbb{C}^2$. The equation for $\boldsymbol{\mathsf{W}}$ is the transport equation (see \eqref{eq:NLF}) \begin{equation} \label{eq:trans} i \partial_t \boldsymbol{\mathsf{W}} + i c^2 \partial_\theta \boldsymbol{\mathsf{W}} = -c\langle \nabla \rangle_c \boldsymbol{\mathsf{W}} + c\langle \nabla \rangle_c^{-1} \boldsymbol{F}(\boldsymbol{\mathsf{W}}), \end{equation} with boundary condition \begin{equation} \label{eq:psiinit} \boldsymbol{\mathsf{W}}(0,0,x) = \boldsymbol{\psi}(x). \end{equation} In a first step, we expand the previous equations with respect to the small parameter $c^{-2}$ in the following sense: For a given $k$, we can write \begin{eqnarray} \label{eq:alphn} c\sqrt{k^2 + c^2} = c^2\sqrt{ 1 + \frac{|k|^2}{c^2}} &=& c^2 + \frac12 |k|^2 + \sum_{n \geq 2} \alpha_n c^{2-2n} |k|^{2n}, \nonumber\\ &=& c^2 + \frac12 |k|^2 + \sum_{n \geq 1} \alpha_{n+1} c^{-2n} |k|^{2n +2}, \end{eqnarray} for some coefficients $\alpha_n$ defining a power series of convergent radius $1$. However, as $k$ is an arbitrary integer, the meaning of the previous expansion in terms of operators has to be understood in the sense of asymptotic expansions; this means that for a given sufficiently smooth function $\boldsymbol{\mathsf{W}}$, we can write $$ c \langle \nabla \rangle_c \boldsymbol{\mathsf{W}} = c^2 \boldsymbol{\mathsf{W}} - \frac12 \Delta \boldsymbol{\mathsf{W}} + \sum_{n \geq 1} \alpha_{n+1} c^{-2n} (-\Delta)^{n+1} \boldsymbol{\mathsf{W}}, $$ in the sense that for all $N \in \mathbb{N} $, there exists a constant $C_N$ such that if $\boldsymbol{w} \in H^{s + 2N + 4} \times H^{s + 2N +4}$ for some $s \in \mathbb{N} \backslash\{0\}$, then for all $c > 0$, $$ \Norm{c \langle \nabla \rangle_c \boldsymbol{w} - c^2 \boldsymbol{w} +\frac12 \Delta \boldsymbol{w} - \sum_{n \geq 1}^{N} \alpha_{n+1} c^{-2n} \Delta^{n+1} \boldsymbol{w}}{H^s} \leq C_N c^{-2N -2} \Norm{\boldsymbol{w}}{H^{s + 2N + 4}}. $$ This last equation can be rigorously proved by using Taylor expansions of the function $x \mapsto \sqrt{1 + x}$, $x > 0$ and Fourier decomposition of $\boldsymbol{\mathsf{W}}$. Similarly, we can write $$ c \langle \nabla \rangle_c^{-1} = \Big(1 - \frac{\Delta}{c^2}\Big)^{-1/2} = 1 + \frac{1}{2c^2} \Delta + \sum_{n \geq 2} \beta_n c^{-2n} (-\Delta)^{n}, $$ for some coefficients $\beta_n$. Hence in view of \eqref{eq:rlb} the initial condition $\boldsymbol{\psi}$ can be expanded in the sense of asymptotic expansion as $$ \boldsymbol{\psi} = \sum_{n \geq 0} c^{-2n} \boldsymbol{\psi}_n(x), $$ with \begin{equation} \label{eq:psin} \boldsymbol{\psi}_0 = \begin{pmatrix} \varphi - i \gamma \\[1ex] \overline{\varphi} - i \overline{\gamma} \end{pmatrix}, \quad \boldsymbol{\psi}_1 = -\textstyle\frac{i}{2} \Delta \begin{pmatrix} \gamma \\[1ex] \overline{\gamma} \end{pmatrix}, \quad \mbox{and}\quad \boldsymbol{\psi}_n = \beta_n(-\Delta)^{n} \begin{pmatrix} \gamma \\[1ex] \overline{\gamma} \end{pmatrix}. \end{equation} Similarly, Equation \eqref{eq:trans} can be expanded in the sense of asymptotic expansion for smooth $\boldsymbol{\mathsf{W}}$ as \begin{equation*} \begin{aligned} c^2(i \partial_\theta + 1) \boldsymbol{\mathsf{W}} =& - i \partial_t \boldsymbol{\mathsf{W}} +\textstyle \frac12 \Delta \boldsymbol{\mathsf{W}} + \boldsymbol{F}(\boldsymbol{\mathsf{W}}) +\displaystyle \sum_{n\geq 1} c^{-2n} \boldsymbol{F}_n(\boldsymbol{\mathsf{W}}), \end{aligned} \end{equation*} where $$ \boldsymbol{F}_n(\boldsymbol{\mathsf{W}}) = \alpha_{n+1} (-\Delta)^{n+1} \boldsymbol{\mathsf{W}} + \beta_n (-\Delta)^n \boldsymbol{F}(\boldsymbol{\mathsf{W}}). $$ For later use we will give the first term $\boldsymbol{F}_1(\boldsymbol{\mathsf{W}})$: For $\boldsymbol{\mathsf{W}} = (\mathsf{U},\mathsf{V})^T$, \begin{equation}\label{eq:F1} \boldsymbol{F}_1(\boldsymbol{\mathsf{W}})= \begin{pmatrix} \textstyle \frac18 \Delta ^2 \mathsf{U} + \textstyle \frac12 \Delta f(\textstyle\frac12(\mathsf{U}+\overline{\mathsf{V}}))\\[1ex] \textstyle \frac18 \Delta ^2 \mathsf{V} + \textstyle \frac12 \Delta f(\textstyle\frac12(\mathsf{V}+\overline{\mathsf{U}})) \end{pmatrix}. \end{equation} Note that for $s > d/2$, can prove the following estimates: $$ \forall\, n \geq 1, \quad \Norm{\boldsymbol{F}_n(\boldsymbol{\mathsf{W}})}{H^{s}} \leq C_n \Norm{\boldsymbol{\mathsf{W}}}{H^{s+ 2 + 2n}}^r, $$ where $r$ is the polynomial degree of $f$ (see \eqref{eq:bil1}), and for some constant $C_n$ depending on $n$ and $s$. We now look for an asymptotic expansion for $\boldsymbol{\mathsf{W}}$ as $$ \boldsymbol{\mathsf{W}}(t,\theta,x) = \sum_{n \geq 0} c^{-2n} \boldsymbol{\mathsf{W}}_n (t,\theta,x). $$ With this expression, we can expand the nonlinear term $\boldsymbol{F}(\boldsymbol{\mathsf{W}})$ using the fact that $f$ is polynomial. We obtain an expansion of the form $$ \boldsymbol{F}(\boldsymbol{\mathsf{W}}) = \boldsymbol{F}(\boldsymbol{\mathsf{W}}_0) + \sum_{n \geq 1} c^{-2n} \text{d} \boldsymbol{F}(\boldsymbol{\mathsf{W}}_0) \cdot \boldsymbol{\mathsf{W}}_n + \boldsymbol{S}_n(\boldsymbol{\mathsf{W}}_j \, |\, j \leq n-1), $$ where $\boldsymbol{S}_n(\boldsymbol{\mathsf{W}}_j \, |\, j \leq n-1)$ denotes a nonlinear term depending only on $\boldsymbol{\mathsf{W}}_j$ for $j \leq n-1$. Also note that this term is polynomial in the functions $\boldsymbol{\mathsf{W}}_j$. Here, $ \text{d} \boldsymbol{F}(\boldsymbol{\mathsf{W}}_0)$ is the differential of $\boldsymbol{F}$ at $\boldsymbol{\mathsf{W}}_0$. It is a polynomial term of oder $r-1$, in $\boldsymbol{\mathsf{W}}_0$ if $f$ is polynomial of degree $r$. Hence we get a recurrence relation of the form \begin{equation} \label{eq:induc0} (i \partial_\theta +1)\boldsymbol{\mathsf{W}}_0 = 0, \quad \mbox{and}\quad (i \partial_\theta +1)\boldsymbol{\mathsf{W}}_1 = (- i \partial_t + \textstyle \frac12\Delta) \boldsymbol{\mathsf{W}}_{0} + \boldsymbol{F}(\boldsymbol{\mathsf{W}}_{0}), \end{equation} and for $n \geq 2$, \begin{multline} \label{eq:induc} (i \partial_\theta +1)\boldsymbol{\mathsf{W}}_n = (- i \partial_t + \textstyle \frac12\Delta) \boldsymbol{\mathsf{W}}_{n-1} + \text{d} \boldsymbol{F}(\boldsymbol{\mathsf{W}}_{0}) \cdot \boldsymbol{\mathsf{W}}_{n-1} + \boldsymbol{R}_n(\boldsymbol{\mathsf{W}}_j \, |\, j \leq n-2), \end{multline} where $\boldsymbol{R}_n(\boldsymbol{\mathsf{W}}_j \, |\, j \leq n-2)$ denotes a term depending only on $\boldsymbol{\mathsf{W}}_j$ for $j \leq n-2$ that is polynomial in the functions $\boldsymbol{\mathsf{W}}_j$ and their derivatives. Using the previous estimates we obtain that for all $n$ and all collections $\boldsymbol{\mathsf{W}}_j$, $j = 0,\ldots, n-2$ of sufficiently smooth functions, \begin{equation} \label{eq:derivW} \Norm{\boldsymbol{R}_n(\boldsymbol{\mathsf{W}}_j \, |\, j \leq n-2)}{H^s} \leq C_n \big(\max_{j = 1,\ldots,n-2}\Norm{\boldsymbol{\mathsf{W}}_j}{H^{s + 2(n-j)}}\big)^r , \end{equation} where the constant $C_n$ does not depend on the $\boldsymbol{\mathsf{W}}_j$'s (but depends on $n$ and $s$). Eventually, the initial condition \eqref{eq:psiinit} yields the conditions \begin{equation} \label{eq:psibn} \forall\, n \geq 0, \quad \boldsymbol{\mathsf{W}}_n(0,0,x) = \boldsymbol{\psi}_n(x), \end{equation} where $\boldsymbol{\psi}_n$ is given by the expansion \eqref{eq:psin}. To solve the previous induction relation \eqref{eq:induc} and \eqref{eq:psibn}, we impose the initial condition $\boldsymbol{\mathsf{W}}_j = 0$ for $j < 0$. We also introduce the following space: we say that $\boldsymbol{\mathsf{G}}(t,\theta, x) \in \mathcal{H} _{T}^s$ if $\boldsymbol{\mathsf{G}}$ is a trigonometric polynomial in $\theta$ with coefficients in $\mathcal{C}^{1}((0,T), H^s \times H^s)$. In other words, we can write $G$ as $$ \boldsymbol{\mathsf{G}}(t,\theta,x) = \sum_{a\in \mathbb{Z} , \, |a| \leq p } e^{ia\theta} \boldsymbol{g}^{(a)}(t,x), $$ where $p < + \infty$ and with $\boldsymbol{g}^{(a)}(t,x) \in \mathcal{C}^{1}((0,T), H^s \times H^s)$. By construction, as the nonlinearity is assumed to be polynomial, it is clear that if for $j = 0,\ldots,n-1$, $\boldsymbol{\mathsf{W}}_j$, is an element of $ \mathcal{H} _{T}^{s + 2 (n-j)}$, then we have (see \eqref{eq:derivW}), $$ \boldsymbol{R}_n(\boldsymbol{\mathsf{W}}_j \, |\, j \leq n-2)\in \mathcal{H} _{T}^{s}. $$ We will use repeatedly the following result: \begin{lemma}[Solution of the homological equation] Let $T \in \mathbb{R} $ and $s > 0$. Assume that $\boldsymbol{\mathsf{G}}(t,\theta,x) \in \mathcal{H} _{T}^s$ is given. Then there exists a unique solution $\boldsymbol{\mathsf{W}}_*(t,\theta,x) \in \mathcal{H} _{T}^s$ satisfying the equations \begin{equation} \label{eq:homo} (i \partial_\theta +1)\boldsymbol{\mathsf{W}}_*(t,\theta,x) = \boldsymbol{\mathsf{G}}(t,\theta,x) - \frac{1}{2\pi}\int_{ \mathbb{T} } e^{-i \theta} \boldsymbol{\mathsf{G}}(t,\theta,x) \text{d} \theta , \end{equation} and \begin{equation} \label{eq:train} \forall\, t \in (0,T), \quad \forall x \in \mathbb{T} ^d, \quad \frac{1}{2\pi}\int_{ \mathbb{T} } e^{-i \theta} \boldsymbol{\mathsf{W}}_*(t,\theta,x) \text{d} \theta = 0. \end{equation} Moreover, for all function $\boldsymbol{w}(t,x) \in \mathcal{C}^1((0,T), H^s \times H^s)$, the function $$ \boldsymbol{\mathsf{W}}(t,\theta,x) = e^{i\theta } \boldsymbol{w}(t,x) + \boldsymbol{\mathsf{W}}_*(t,x). $$ satisfies the equation \eqref{eq:homo}. \end{lemma} \begin{proof} We decompose $\boldsymbol{\mathsf{G}}(t,\theta,x)$ in Fourier in $\theta$: we can express this function as $$ \boldsymbol{\mathsf{G}}(t,\theta,x) = \sum_{|a| \leq p} e^{i a\theta} \boldsymbol{g}^{(a)} (t,x), $$ with $p < + \infty$ and where $\boldsymbol{g}^{(a)}(t,x) \in \mathcal{C}^1((0,T),H^s \times H^s)$. Note that $$ \frac{1}{2\pi}\int_{ \mathbb{T} } e^{-i \theta} \boldsymbol{\mathsf{G}}(t,\theta,x) \text{d} \theta = \boldsymbol{g}^{(1)}(t,x). $$ Let us seek $\boldsymbol{\mathsf{W}}_*$ under the form $$ \boldsymbol{\mathsf{W}}_{*}(t,\theta,x) = \sum_{|a| \leq p} e^{i a \theta} \boldsymbol{w}_*^{(a)}(t,x), $$ for some unknown functions $\boldsymbol{w}_*^{(a)}$ in $\mathcal{C}^1((0,T),H^s \times H^s)$. Equation \eqref{eq:homo} can be written as $$ \forall\, |a| \leq p , \quad (1 - a) \boldsymbol{w}_*^{(a)}(t,x) = \boldsymbol{g}^{(a)}(t,x) - \delta_a^1 \boldsymbol{g}^{(1)}(t,x), $$ where $\delta_a^1$ denotes the Kronecker symbol. It is clear that this equation is solvable for the functions $\boldsymbol{w}_*^{(a)}(t,x)$, $a \neq 1$ and that with the condition \eqref{eq:train} we can set $\boldsymbol{w}_*^{(1)}(t,x) = 0$. The last statement is a consequence of the fact that the kernel of the operator $i \partial_\theta + 1$ is the space spanned by $e^{i\theta}$. \end{proof} We are now ready to state the main result of this section: \begin{theorem}\label{thm:main} Let $N \in \mathbb{N} $, $s > d/2$, $c_0 > 0$ and assume that $\varphi\in H^{s + 2N + 4}$ and $\gamma \in H^{s + 2N + 4}$ are given. Then there exists $T = T(s,N) > 0$, and a sequence of functions $\boldsymbol{\mathsf{W}}_n(t,\theta,x) \in \mathcal{H} _{T}^{s + 2(N-n)}$, satisfying the following properties: $\boldsymbol{\mathsf{W}}_0 = \boldsymbol{w}_0 e^{i \theta}$ where $\boldsymbol{w}_0$ satisfies the equation \begin{equation} \label{eq:nlsw0} i \partial_t \boldsymbol{w}_0 =\textstyle \frac12 \Delta \boldsymbol{w}_0 + \langle \boldsymbol{F}\rangle(\boldsymbol{w}_0), \quad \boldsymbol{w}_0(0) = \boldsymbol{\psi}_0, \end{equation} where $\boldsymbol{\psi}_0$ is defined in \eqref{eq:psin} and where by definition, $$ \langle \boldsymbol{F}\rangle(\boldsymbol{w}) = \frac{1}{2\pi} \int_0^{2\pi} \begin{pmatrix} f(u + e^{-2i\theta}\overline {v}) \\�f(v + e^{-2 i \theta} \overline u) \end{pmatrix} \text{d} \theta, \quad \mbox{if} \quad \boldsymbol{w} = \begin{pmatrix} u \\�v \end{pmatrix}, $$ and for all $n$, $\boldsymbol{\mathsf{W}}_n$ solve the equations \eqref{eq:induc}-\eqref{eq:psibn} for $n = 0,\ldots, N$, and can be decomposed into trigonometric polynomials of the form \begin{equation} \label{eq:Wbn} \boldsymbol{\mathsf{W}}_n = \sum_{|a| \leq p_n} e^{ia \theta} \boldsymbol{w}_n^{(a)}(t,x), \end{equation} with $p_n < +\infty$, and where $\boldsymbol{w}_n^{(a)}(t,x) \in \mathcal{C}((0,T), H^{s + 2(n-N)} \times H^{s + 2(n-N)} )$. Moreover, there exists a constant $C_N$ such that for $c > c_0$, if $\boldsymbol{w}(t)$ denotes the solution given by Proposition \ref{prop21} with initial value $\boldsymbol{\psi}$ given by \eqref{eq:rlb} in terms of $\varphi$ and $\gamma$, then \begin{equation} \label{eq:asympt} \forall\, t \leq T, \quad \Norm{\boldsymbol{w}(t,x) - \sum_{n = 0}^N c^{-2n} \boldsymbol{\mathsf{W}}_n (t,c^{-2}t, x)}{H^{s}} \leq C_N c^{-2 N - 2}. \end{equation} \end{theorem} \begin{proof} Let us consider the recursion \eqref{eq:induc} together with the initial condition \eqref{eq:psibn} where the $\boldsymbol{\psi}_n$ are given by \eqref{eq:psin}. For $n = 0$, we obtain the equation $$ i (\partial_\theta + 1) \boldsymbol{\mathsf{W}}_0(t,\theta,x) = 0,\quad \mbox{and}\quad\boldsymbol{\mathsf{W}}(0,0,x) = \boldsymbol{\psi}_0(x) $$ which implies, according to the previous lemma, that \begin{equation} \label{eq:W0} \boldsymbol{\mathsf{W}}_0(t,\theta,x) = e^{i\theta}\boldsymbol{w}_0^{(1)}(t,x), \quad \mbox{and}\quad\boldsymbol{w}_0^{(1)}(0,x) = \boldsymbol{\psi}_0(x) \end{equation} that is we have determined $\boldsymbol{w}_0^{(a)} = 0$ for $a \neq 1$ (and set $p_0 = 1$). For $ n = 1$, the equation \eqref{eq:induc0} is written \begin{equation}\label{eq:W1} i (\partial_\theta + 1) \boldsymbol{\mathsf{W}}_1(t,\theta,x) = (- i \partial_t + \textstyle \frac12\Delta) \boldsymbol{\mathsf{W}}_{0} + \boldsymbol{F}(\boldsymbol{\mathsf{W}}_{0}). \end{equation} According to the result of given in the previous lemma (see \eqref{eq:homo}), this equation will be solvable if the right-hand side satisfies $$ \int_{ \mathbb{T} } (- i \partial_t + \textstyle \frac12\Delta) e^{-i \theta} \boldsymbol{\mathsf{W}}_{0}(t,\theta,x) + e^{-i \theta}\boldsymbol{F}(\boldsymbol{\mathsf{W}}_{0}(t,\theta,x)) \text{d} \theta = 0. $$ In view of \eqref{eq:W0}, this equation can be written as $$ ( i \partial_t - \textstyle \frac12\Delta) \boldsymbol{w}_{0}^{(1)}(t,x) = \displaystyle \frac{1}{2\pi} \int_{ \mathbb{T} } e^{-i \theta}\boldsymbol{F}(e^{i\theta}\boldsymbol{w}_{0}^{(1)}(t,x)) \text{d} \theta, $$ which gives the equation \eqref{eq:nlsw0} with the boundary condition \eqref{eq:psibn}. Now let $T$ be such that $\boldsymbol{w}_0(t,x)$ is a solution to the previous equation with initial value $\boldsymbol{w}(0,x) = \boldsymbol{\psi}_0(x)$ in the space $\mathcal{C}^1((0,T), H^{s + 2N +4} \times H^{s +2N +4})$. Then we can set $\boldsymbol{w}_{0}^{(1)}(t,x) = \boldsymbol{w}_0(t,x)$ and solve for the functions $\boldsymbol{w}_1^{(a)}$, $a \neq 1$ according to the previous lemma. Note that $p_1$ is given by the degree of the polynomial in the nonlinear right-hand side. Now assume that $n \geq 1$ is given, and that $\boldsymbol{w}_k^{(a)}(t,\theta,x)$, $a \leq p_k$ are known functions in $\mathcal{C}^1((0,T, H^{s + 2(N - k) +4}\times H^{s + 2(N-k) +4})$ for $k \leq n-2$ and $p_k < \infty$, and that $\boldsymbol{w}_{n-1}^{(a)}(t,\theta,x)$ is determined for $a \neq 1$. Consider Equation \eqref{eq:induc}. Using \eqref{eq:derivW} and the induction hypothesis, the right-hand side of this equation belongs to $\mathcal{C}^1((0,T, H^{s + 2(N - n)+2}\times H^{s + 2(N-n) +2})$, and is a polynomial in $e^{i\theta}$ of degree $p_n \geq p_{n-1}$, because the nonlinearity is a polynomial. To be solvable, the right-hand side of \eqref{eq:induc} has to be orthogonal to $e^{-i \theta}$ in $L^2( \mathbb{T} )$. By induction hypothesis, only the term $w_{n-1}^{(1)}$ has to determined. In view of \eqref{eq:induc}, we thus see that $w_{n-1}^{(1)}$ has to satisfy a linear equation of the form \begin{equation} \label{eq:eqlin} (- i \partial_t + \textstyle \frac12\Delta) \boldsymbol{w}_{n-1}^{(1)}(t,x) = \boldsymbol{m}_{n-1}(t,x) \cdot \boldsymbol{w}_{n-1}^{(1)} + \boldsymbol{r}_{n-1}(t,x), \end{equation} where $\boldsymbol{m}_{n-1}(t,x)$ is a linear multiplicative operator with coefficients in the space $\mathcal{C}^{1}((0,T), H^{s + 4N+2})$, and where $\boldsymbol{r}(t,x) \in \mathcal{C}^1((0,T, H^{s + 2(N - n)+2}\times H^{s + 2(N-n)+2})$. The initial condition associated with this equation is written, according to \eqref{eq:psibn}, as $$ \boldsymbol{w}_{n-1}^{(1)}(0,x) = \boldsymbol{\psi}_n(x) + \sum_{|a| \leq p_{n-1}, \, \, a \neq 1} \boldsymbol{w}_{n-1}^{(a)}(0,x), $$ where $\boldsymbol{\psi}_n(x)$ is given by \eqref{eq:psin} in terms of $\varphi$ and $\gamma$, and where the other terms in the right-hand side are known by induction hypothesis. As the equation \eqref{eq:eqlin} in linear with coefficients bounded on the time interval $(0,T)$; we deduce that $\boldsymbol{w}_{n-1}^{(1)}(t,x)$ is well defined in $\mathcal{C}^1((0,T, H^{s + 2(N - n)+2}\times H^{s + 2(N-n)+2})$. Using the previous Lemma, we can then solve for the terms $\boldsymbol{w}_{n}^{(a)}$ for $a \neq 1$ in $\mathcal{C}^1((0,T, H^{s + 2(N - n)+2}\times H^{s + 2(N-n)+2})$. To prove the asymptotic expansion, we note that by construction the term $$ \boldsymbol{w}^{N}(t,x) := \sum_{n = 0}^{N} c^{-2n} \boldsymbol{\mathsf{W}}_n (t,c^{-2}t, x) $$ belong to $\mathcal{C}^1((0,T, H^{s +2}\times H^{s +2})$ and satisfies the equation \eqref{eq:NLF} up to an error or order $c^{-2(N+1)}$, that is \begin{multline} \label{eq:approx21} \forall\, t \in (0,T),\\ \begin{array}{rcl} i \partial_t \boldsymbol{w}^N(t,x) &=& - c \langle \nabla \rangle_c \boldsymbol{w}^N(t,x) + c \langle \nabla \rangle_c^{-1} \boldsymbol{F}(\boldsymbol{w}^N(t,x)) + c^{-2N -2}\boldsymbol{r}_c^N(t,x),\\[1ex] \boldsymbol{w}^N(0,x) &=& \boldsymbol{\psi}(x) + c^{-2N - 2} \boldsymbol{\psi}_c^N \end{array} \end{multline} with $$ \boldsymbol{r}^N_c(t,x) \in \mathcal{C}^1((0,T, H^{s}\times H^{s}), \quad \mbox{and}\quad \boldsymbol{\psi}^N_c(x) \in H^s \times H^s, $$ satisfying the following uniform bounds: for a given $c_0 >0$, there exists a constant $C_N$ such that $$ \forall\, c > c_0,\quad \forall t \in (0,T), \quad \Norm{\boldsymbol{w}^N(t,x) }{H^s} + \Norm{\boldsymbol{r}_c^N(t,x)}{H^s} + \Norm{\boldsymbol{\psi}_c^N(x)}{H^s} \leq C_N. $$ Using Proposition \ref{prop21}, we can assume that the exact solution $\boldsymbol{w}(t)$ is also uniformly bounded (with respect to $c$) on $(0,T)$ in $H^s$. We conclude by using \eqref{eq:approx21} and standard comparison arguments based on the Gronwall Lemma on $(0,T)$ and in $H^s$ with $s > d/2$. \end{proof} \section{First terms}\label{sec:firstt} The previous result implies the existence of an asymptotic expansion of $u(t,x)$ and $v(t,x)$ solutions of \eqref{eq:NLSc}-\eqref{eq:BCc} up to any arbitrary order, provided the initial condition is smooth. Note that we will use the following notations: for all $n \geq 0$, $\boldsymbol{\mathsf{W}}_n = (\mathsf{U}_n,\mathsf{V}_n)^T$ with, according to the decomposition \eqref{eq:Wbn} \begin{equation} \label{eq:devU} \mathsf{U}_n = \sum_{|a| \leq p_n} e^{ia \theta} u_n^{(a)}(t,x),\quad \mbox{and}\quad \mathsf{V}_n = \sum_{|a| \leq p_n} e^{ia \theta} v_n^{(a)}(t,x). \end{equation} This implies an asymptotic expansion of the solution $u(t)$, $v(t)$ and $z(t)$ of the form $$ u(t,x) = \sum_{n \geq 0} c^{-2n} u_n(t,c^2t,x) , \quad v(t,x) = \sum_{n \geq 0} c^{-2n} v_n(t,c^2t,x) , $$ and $$ z(t,x) = \sum_{n \geq 0} c^{-2n} z_n(t,c^2t,x), $$ with the relation $z_n = \frac12 (u_n + \overline{v_n})$. Note that here, the equality signs must be understood in the sense of asymptotic expansion. Remark also that in the case where $\varphi$ and $\gamma$ are real, that is when the solution $z(t)$ is real, then we have for all $n$, $u_n = v_n$. Using the result of Theorem \ref{thm:main}, $z_0(t,x)$ is an approximation of the solution $z(t,x)$ over the time interval $(0,T)$ with an error order $\mathcal{O}(c^{-2})$, while \begin{equation} \label{eq:devz} z_0(t,x) + \textstyle\frac{1}{c^2} z_1(t,x) \end{equation} approximates the exact solution $z(t,x)$ over the time interval $(0,T)$ up to terms of order $c^{-4}$. The goal of this section is to express the two first terms $z_0$ and $z_1$. For simplicity we consider the one dimensional case and assume the initial values $\varphi, \gamma$ to be sufficiently smooth so that Theorem \ref{thm:main} can be applied. Note that a generalization to higher dimensions follows the line of argumentation. We will use these computations to derive our numerical schemes in Section \ref{sec:nume}. For reasons of ease we will focus on the linear case $f(z) = \lambda z$, see Section \ref{ex:lin}, and a cubic-nonlinearity $f(z) = \lambda|z|^2 z$, see Section \ref{ex:cub}. Nevertheless we state the equations for $\mathsf{U}_1$ and $\mathsf{V}_1$ in the general case $f(z) = \lambda|z|^{2p} z$, $p\in \mathbb{N} $. The first term $z_0$ approximating the exact solution of Klein-Gordon equation \eqref{eq:kgr} up to terms of order $c^{-2}$ is easily derived from Theorem \ref{thm:main} using \eqref{eq:zuv} and \eqref{eq:rlb}. More precisely, we have $$ z_{0}(t,x) = \textstyle \frac{1}{2} e^{itc^2} u_0(t,x) + \frac12 e^{-itc^2} \overline{v_0}(t,x), $$ where $u_0, v_0$ solve the nonlinear Schr\"odinger system \eqref{eq:nlsw0}. In order to explicit the term $z_1(t,x)$, we expand the non-linearity $f(\frac12 (\mathsf{U} +\overline{\mathsf{V}}))$ in a Taylor-series, up to terms of order $c^{-2}$. With the notation $\mathsf{Z}_k = \frac12 (\mathsf{U}_k +\overline{\mathsf{V}_k})$ and keeping in mind that $f(z) = f(z,\overline{z}) = \lambda \vert z\vert^{2p} z = \lambda z^{p+1} \overline{z}^p$, we obtain \begin{equation*} \begin{aligned} f(\textstyle\frac12 (\mathsf{U} +\overline{\mathsf{V}})) &= f(\mathsf{Z}_0+\textstyle\frac{1}{c^2} \mathsf{Z}_1)+\mathcal{O}(c^{-4}) \\&= f(\mathsf{Z}_0) + \textstyle\frac{1}{c^2}\big(\partial_z f(\mathsf{Z}_0,\overline{\mathsf{Z}_0}) \mathsf{Z}_1 + \partial_{\overline{z}} f(\mathsf{Z}_0,\overline{\mathsf{Z}_0})\overline{\mathsf{Z}_1}\big) +\mathcal{O}(c^{-4})\\ &= f(\mathsf{Z}_0) + \textstyle\frac{\lambda}{c^2} \left[ (p+1) \vert \mathsf{Z}_0\vert^{2p}\mathsf{Z}_1+ p \vert \mathsf{Z}_0\vert^{2p-2}(\mathsf{Z}_0)^2 \overline{\mathsf{Z}_1}\right]+\mathcal{O}(c^{-4}). \end{aligned} \end{equation*} Furthermore, in the above notation we calculate that \begin{equation*} \begin{aligned} & \Delta f = \partial_{zz} f (\partial_x \mathsf{Z})^2 + 2\partial_{z\overline{z}}f(\partial_x \mathsf{Z})(\partial_x \overline{\mathsf{Z}})+\partial_z f (\partial_{xx} \mathsf{Z}) +\partial_{\overline{z}\overline{z}} f (\partial_x \overline{\mathsf{Z}})^2 + \partial_{\overline{z}} f (\partial_{xx} \overline{\mathsf{Z}}), \end{aligned} \end{equation*} where in this formula $f$ is evaluated in $(\mathsf{Z},\overline{\mathsf{Z}})$. Thus, we obtain by \eqref{eq:induc} with $n=2$ (using \eqref{eq:F1} and the above calculations) that the homological equation in terms of $\mathsf{U}_2$ is written \begin{equation}\label{eq:U2} \begin{aligned} (i \partial_\theta +1)\mathsf{U}_2 =& -i\partial_t \mathsf{U}_1 + \textstyle \frac12 \Delta \mathsf{U}_1 + \textstyle \frac18 \Delta ^2 \mathsf{U}_0 \\&+\lambda (p+1) \vert \textstyle \frac12 (\mathsf{U}_0 + \overline{\mathsf{V}_0}) \vert^{2p} \textstyle \frac12(\mathsf{U}_1+\overline{\mathsf{V}_1}) \\ &+ \lambda p\vert \textstyle \frac12(\mathsf{U}_0 + \overline{\mathsf{V}_0}) \vert^{2p-2}( \textstyle \frac12(\mathsf{U}_0+\overline{\mathsf{V}_0}))^{2}\textstyle \frac12(\mathsf{V}_1+\overline{\mathsf{U}_1}) \\&+\lambda (p+1) \vert \textstyle \frac12 (\mathsf{U}_0 + \overline{\mathsf{V}_0}) \vert^{2p}\textstyle \frac14\Delta (\mathsf{U}_0+\overline{\mathsf{V}_0})\\ & + \lambda p\vert \textstyle \frac12(\mathsf{U}_0 + \overline{\mathsf{V}_0}) \vert^{2p-2}( \textstyle \frac12(\mathsf{U}_0+\overline{\mathsf{V}_0}))^{2} \textstyle \frac14\Delta (\mathsf{V}_0+\overline{\mathsf{U}_0}) \\ &+\lambda \textstyle \frac12p(p+1)\vert \textstyle \frac12 (\mathsf{U}_0 + \overline{\mathsf{V}_0}) \vert^{2p-2}\frac12(\overline{\mathsf{U}_0}+\mathsf{V}_0)\big(\partial_x \frac12 (\mathsf{U}_0 + \overline{\mathsf{V}_0}) )^2\\ &+\lambda\textstyle p(p+1)\vert\textstyle \frac12 (\mathsf{U}_0+ \overline{\mathsf{V}_0}) \vert^{2p-2}\textstyle \frac12 (\mathsf{U}_0 + \overline{\mathsf{V}_0}) \vert\partial_x \frac12 (\mathsf{U}_0 + \overline{\mathsf{V}_0})\vert^2\\ &+\lambda \textstyle \frac12p(p-1)\vert\textstyle \frac12 (\mathsf{U}_0 + \overline{\mathsf{V}_0}) \vert^{2p-4}( \frac12 (\mathsf{U}_0 + \overline{\mathsf{V}_0}) )^3\big(\partial_x \frac12 (\overline{\mathsf{U}_0}+\mathsf{V}_0)\big), \end{aligned} \end{equation} and a similar equation for $\mathsf{V}_2$ with $\mathsf{U}_k$ replaced by $\mathsf{V}_k$, $k=0,1,2$, and vice versa. \subsection{Formula for the first terms in the linear case}\label{ex:lin} In the linear case we have $f(z) =\lambda z$. In this case we obtain the following limit equation for $u_0$ and $v_0$ \begin{equation}\label{eq:u0lin} \begin{aligned} \textstyle(- i\partial_t+\frac12 \Delta + \frac{\lambda}{2} )u_0 &= 0,\qquad u_0(0) = \varphi - i \gamma,\\ \textstyle(- i\partial_t+\frac12 \Delta + \frac{\lambda}{2}) v_0 &= 0,\qquad v_0(0) = \overline{\varphi+i\gamma}, \end{aligned} \end{equation} see \eqref{eq:nlsw0}. For $\mathsf{U}_1$ we obtain by \eqref{eq:W1} \begin{equation*} (i\partial_\theta +1)\mathsf{U}_1 = (-i\partial_t +\textstyle \frac12\Delta)u_0\mathrm{e}^{i\theta} + \frac12\lambda(u_0\mathrm{e}^{i\theta}+\overline{v_0}\mathrm{e}^{-i\theta}). \end{equation*} Thus, plugging the Fourier expansion \eqref{eq:devU} of $\mathsf{U}_1$ in $\theta$, we obtain with the orthogonality relation \eqref{eq:u0lin} for $u_0$ that \begin{equation*} \sum_{a\in \mathbb{Z} } (1-a) u_1^{(a)}\mathrm{e}^{ia\theta} = \textstyle \frac{1}{2}\lambda\overline{v_0}\mathrm{e}^{-i\theta}. \end{equation*} Thus, $u_1^{(a)} = 0$ for $a\neq \pm 1$ and $u_{1}^{(-1)} = \textstyle\frac14 \lambda \overline{v_0}$, and we set $\xi_1 := u_1^{(1)}$, so that we can write \begin{equation} \mathsf{U}_1(t,\theta,x)= \textstyle \frac14\lambda \overline{v_0}(t,x) \mathrm{e}^{-i\theta}+\xi_1(t,x)\mathrm{e}^{i\theta}, \end{equation} Note that similarly $\mathsf{V}_1(t,\theta,x)= \frac14\lambda \overline{u_0}(t,x) \mathrm{e}^{-i\theta}+\eta_1(t,x)\mathrm{e}^{i\theta}$ with $\eta_1 := v_1^{(1)}$. In order to determine $\xi_1$ and $\eta_1$ we analyze \eqref{eq:U2} in the linear case, i.e. $p=0$, which reduces to \begin{equation*} (i\partial_\theta +1) \mathsf{U}_2 = -i \partial_t \mathsf{U}_1 +\textstyle \frac12 \Delta \mathsf{U}_1 + \frac18 \Delta^2 \mathsf{U}_0 +\lambda \frac12 (\mathsf{U}_1+\overline{\mathsf{V}_1})+\frac14\lambda \Delta (\mathsf{U}_0+\overline{\mathsf{V}_0}). \end{equation*} To be solvable, the right-hand side of the equation needs to be orthogonal to $\mathrm{e}^{-i\theta}$ which yields the following equation for $\xi_1(t,x)$: \begin{equation*} \textstyle(- i\partial_t+\frac12 \Delta + \frac{\lambda}{2})\xi_1 = \textstyle -\frac18\Delta^2 u_0 -\frac18\lambda^2 u_0 - \frac14 \lambda \Delta u_0. \end{equation*} Using twice the Schr\"odinger equation \eqref{eq:u0lin} for $u_0$, the above equation can be rewritten as \begin{equation*} \textstyle(- i\partial_t+\frac12 \Delta + \frac{\lambda}{2})u_1 = \textstyle \frac12 \partial_{tt} u_0. \end{equation*} The same equation holds true for $\eta_1$ with $\xi_1$ and $u_0$ replaced by $\eta_1$ and $v_0$, respectively. The initial values $\xi_1(0)$ and $\eta_1(0)$ have to be chosen such that $\boldsymbol{\mathsf{W}}_1(0,0,x) = \boldsymbol{\psi}_1(x)$, see \eqref{eq:psibn}. As $\mathsf{U}_1(0) = \frac{1}{4} \lambda\overline{v_0}(0)+\xi_1(0)$ and $\mathsf{V}_1(0) = \frac{1}{4}\lambda \overline{u_0}(0)+\eta_1(0)$ we thus obtain by \eqref{eq:psin} the following initial value problems for $\xi_1$ and $\eta_1$ \begin{equation}\label{eq:u1lin} \begin{aligned} & \left(\textstyle- i\partial_t+\frac12 \Delta + \frac{\lambda}{2}\right) \xi_1 = \textstyle \frac12 \partial_{tt} u_0,\, &\xi_1(0) = \textstyle \frac14 (\Delta u_0(0)-(\Delta+\lambda) \overline{v_0}(0)),\\ & \left(\textstyle- i\partial_t+\frac12 \Delta + \frac{\lambda}{2}\right) \eta_1 = \textstyle \frac12 \partial_{tt} v_0,\, & \eta_1(0) = \textstyle \frac14 (\Delta v_0(0)-(\Delta+\lambda) \overline{u_0}(0)). \end{aligned} \end{equation} We summarize our findings in the following corollary. \begin{corollary}[First terms in the linear case]\label{cor:lin} In the linear case, i.e. $f(z) = \lambda z$, the first and second terms in the asymptotic expansion \eqref{eq:devz} read \begin{equation*} z_0(t,x) = \textstyle \frac{1}{2} e^{itc^2} u_0(t,x) + \frac12 e^{-itc^2} \overline{v_0}(t,x) \end{equation*} and \begin{equation}\label{eq:z2lin} z_1(t,x) =\textstyle \frac\lambda{8} \big(u_0(t,x)\mathrm{e}^{ic^2t}+\overline{v_0}(t,x)\mathrm{e}^{-ic^2t}\big) + \frac{1}{2}\big(\xi_1(t,x)\mathrm{e}^{ic^2t}+\overline{\eta_1}(t,x)\mathrm{e}^{-ic^2t}\big), \end{equation} where $u_0$, $v_0$ and $\xi_1$, $\eta_1$ solve \eqref{eq:u0lin} and \eqref{eq:u1lin}, respectively. Then the functions $z_0(t,x)$ and $z_0(t,x) + \frac{1}{c^2}z_1(t,x)$ approximate the exact solution of \eqref{eq:kgr} in $H^s$ up to terms of order $c^{-2}$ and $c^{-4}$ for initial values in $H^{s+4}$ and $H^{s+6}$, respectively. \end{corollary} Note that both corrections terms \eqref{eq:u0lin} and \eqref{eq:u1lin} can be solved exactly in Fourier. Thus, in the linear case we can state the correction terms $z_0$ and $z_1$ in an exact way. See Section \ref{sec:lin} for further details. \subsection{Formula for the first terms in the cubic case}\label{ex:cub} We consider now the case of cubic nonlinearity, i.e. $p = 1$ with the previous notations. Moreover, for simplicity we consider the case, where $u_0 \equiv v_0$, i.e. $z(0)$ and $\partial_t z(0) \in \mathbb{R} $. Thus, $f(\frac12(\mathsf{U}_0 + \overline{\mathsf{V}_0})) =\textstyle \frac{\lambda}{16}\big (u_0\mathrm{e}^{i\theta}+\overline{u_0}\mathrm{e}^{-i\theta}\big)^3$ and the limit equation for $u_0$ reduces to \begin{equation}\label{eq:u0rel} -i\partial_t u_0 +\textstyle\frac{1}{2}\Delta u_0 + \frac{3}{8} \lambda \vert u_0\vert^2 u_0 = 0,\qquad u_0(0) = \varphi - i \gamma, \end{equation} see \eqref{eq:nlsw0}. Plugging the Fourier expansion \eqref{eq:devU} of $\mathsf{U}_1$ in $\theta$ into \eqref{eq:W1}, we obtain with the orthogonality relation \eqref{eq:u0rel} for $u_0$ that $$ \sum_{a\in \mathbb{Z} } (1-a)u_1^{(a)} \mathrm{e}^{ia\theta} =\textstyle\frac{\lambda}{8} \Big((u_0)^3\mathrm{e}^{3i\theta}+3\vert u_0\vert^2\overline{u_0}\mathrm{e}^{-i\theta} + (\overline{u_0})^3\mathrm{e}^{-3i\theta}\Big). $$ Hence, $u_1^{(a)} = 0$ for $a\neq -3,-1,1,3$ and $$ u_{1}^{(-3)} = \textstyle\frac{\lambda}{32}(\overline{u_0})^3,\quad u_{1}^{(-1)} = \frac{3\lambda}{16}\vert u_0\vert^2 \overline{u_0},\quad\mbox{and}\quad u_1^{(3)} = -\textstyle\frac{\lambda}{16} (u_0)^3. $$ Setting $\xi_1 := u_1^{(1)}$, we obtain \begin{equation}\label{eq:U1cub} \mathsf{U}_1= \textstyle \frac{\lambda}{8}\Big(- \frac12 (u_0)^3 \mathrm{e}^{3i\theta}+\frac32 \vert u_0\vert^2\overline{u_0} \mathrm{e}^{-i\theta}+\frac14 (\overline{u_0})^3\mathrm{e}^{-3i\theta}\Big) + \xi_1 \mathrm{e}^{i\theta} (= \mathsf{V}_1). \end{equation} In order to determine the function $\xi_1$ we need to state the equation for $\mathsf{U}_2 = \sum_{a\in \mathbb{Z} } u_2^{(a)} \mathrm{e}^{i a \theta}$, which reduces in our cubic setting to \begin{equation*} \begin{aligned} \sum_{a\in \mathbb{Z} } (1- & a)u_2^{(a)} \mathrm{e}^{ia\theta} =\sum_{a\in \mathbb{Z} } (-i\partial_t + \textstyle\frac12 \Delta) u_1^{(a)} \mathrm{e}^{ia\theta} +\textstyle\frac18\Delta^2 u_0 \mathrm{e}^{i\theta}\\& +\textstyle\frac{3}{16}\lambda\Big((u_0)^2\mathrm{e}^{2i\theta}+2\vert u_0\vert^2+(\overline{u_0})^2\mathrm{e}^{-2i\theta}\Big)(\Delta u_0\mathrm{e}^{i\theta}+\Delta \overline{u_0}\mathrm{e}^{-i\theta})\\& +\textstyle \frac38 \lambda \Big ((\partial_x u_0)^2\mathrm{e}^{2i\theta}+2\vert \partial_x u_0\vert^2+(\partial_x \overline{u_0})^2\mathrm{e}^{-2i\theta}\Big)(u_0\mathrm{e}^{i\theta}+\overline{u_0}\mathrm{e}^{-i\theta})\\& +\textstyle \frac38\lambda\Big((u_0)^2\mathrm{e}^{2i\theta}+2\vert u_0\vert^2+(\overline{u_0})^2\mathrm{e}^{-2i\theta}\Big)\sum_{a\in \mathbb{Z} } (u_1^{(a)} \mathrm{e}^{ia\theta}+ \overline{u_1^{(a)}}\mathrm{e}^{-ia\theta}), \end{aligned} \end{equation*} see \eqref{eq:U2}. To be solvable, the right-hand side of the equation needs to be orthogonal to $\mathrm{e}^{-i\theta}$ which yields the following equation for $\xi_1(t,x) = u_1^{(1)}(t,x)$: \begin{equation*} \begin{aligned} & (i \partial_t - \textstyle \frac12 \Delta )\xi_1 -\lambda \textstyle \frac34 \vert u_0\vert^2 \xi_1 -\lambda \frac38 (u_0)^2\overline{\xi_1}= \textstyle \frac18 \Delta^2 u_0 +\lambda^2 \frac{51}{256}\vert u_0\vert^4 u_0 +\frac{3}{2}\Delta f(\frac12 u_0). \end{aligned} \end{equation*} The initial value $\xi_1(0)$ needs to be chosen such that $\mathsf{U}_1(0) =- \frac{i}{2}\Delta \gamma$, see \eqref{eq:psibn} as well as the definition \eqref{eq:psin}. Thus, by \eqref{eq:U1cub} we obtain that \begin{equation*} \xi_1(0) = \textstyle \frac{\lambda}{16}u_0(0)^3 - \frac{\lambda}{32}\overline{u_0}(0)^3-\frac{3\lambda}{16}\vert u_0(0)\vert^2 \overline{u_0}(0)+\frac14 \Delta (u_0(0)-\overline{u_0}(0)). \end{equation*} Using \eqref{eq:u0rel} we rewrite the equation for $u_1$ as \begin{equation}\label{eq:u1rel} \begin{aligned} & (i \partial_t - \textstyle \frac12 \Delta )\xi_1 -\lambda \textstyle \frac34 \vert u_0\vert^2 \xi_1 -\lambda \frac38 (u_0)^2\overline{\xi_1}\\& \qquad\qquad\qquad\qquad\qquad= \textstyle \frac14 i \partial_t \Delta u_0 +\lambda^2 \frac{51}{256}\vert u_0\vert^4 u_0 +\textstyle\lambda \frac{3}{32}\Delta (\vert u_0\vert^2 u_0),\\ &\xi_1(0) = \textstyle \frac{\lambda}{16}u_0(0)^3 - \frac{\lambda}{32}\overline{u_0}(0)^3-\frac{3\lambda}{16}\vert u_0(0)\vert^2 \overline{u_0}(0)+\frac14 \Delta (u_0(0)-\overline{u_0}(0)). \end{aligned} \end{equation} Again we summarize our findings in the following corollary. \begin{corollary}[First terms in the cubic case]\label{cor:cub} For a cubic non-linearity, i.e. $f(z) = \lambda \vert z\vert^2 z$, and for real initial values functions $z(0,x) = \varphi(x) \in \mathbb{R} $ and $\partial_t z(0,x) = c^2 \gamma(x) \in \mathbb{R} $, the first and second terms in the expansion \eqref{eq:devz} read \begin{equation*} z_0(t,x) = \textstyle \frac{1}{2} e^{itc^2} u_0(t,x) + \frac12 e^{-itc^2} \overline{u_0}(t,x), \end{equation*} and \begin{equation}\label{eq:corr2} \begin{aligned} z_1(t,x)=& \textstyle \frac{3\lambda}{32}\vert u_0(t,x)\vert^2 (u_0(t,x)\mathrm{e}^{ic^2t}+\overline{u_0}(t,x)\mathrm{e}^{-ic^2t})\\&- \textstyle\frac{\lambda}{64}\big( u_0(t,x)^3\mathrm{e}^{3ic^2t}+\overline{u_0}(t,x)^3\mathrm{e}^{-3ic^2t}\big) \\&+ \textstyle \frac{1}{2} \big( \xi_1(t,x) \mathrm{e}^{ic^2t}+\overline{\xi_1}(t,x)\mathrm{e}^{-ic^2t}\big), \end{aligned} \end{equation} where $u_0$ and $\xi_1$ solve \eqref{eq:u0rel} and \eqref{eq:u1rel}, respectively. Then the functions $z_0(t,x)$ and $z_0(t,x) + \frac{1}{c^2}z_1(t,x)$ approximate the exact solution of \eqref{eq:kgr} in $H^s$ up to terms of order $c^{-2}$ and $c^{-4}$ for initial values in $H^{s+4}$ and $H^{s+6}$, respectively. \end{corollary} \section{Energy and charge density conservation}\label{sec:enq} Analyzing the conserved quantities of the Klein-Gordon equation \eqref{eq:kgr}, we follow Dirac's approach via the ``hole" theory, i.e. antiparticles are related to negative energy eigenstates ($u$ particle with positive charge, $v$ antiparticle with negative charge). This allows us to overcome the problem of \emph{indefinite probability density} (i.e. if we would consider $z(t)$ as a probability amplitude, its probability density $\rho(z) = \frac{i}{2c^2}\left(\overline{z}\partial_t z - z \partial_t\overline{z} \right)$ which is naturally defined by the continuity equation, would allow positive and negative values.) For further details on the quantization of the Klein-Gordon field we refer to \cite{BjoD65}. In this interpretation ($\rho(z)$ being the \emph{charge}-density) the conserved physical quantities of the Klein-Gordon equation are the charge \begin{equation}\label{eq:chargi} Q(z) = \frac{i}{2c^2}\int_ \mathbb{T} z\partial_t \overline{z}-\overline{z} \partial_t z\,\mathrm{d}x = \int_ \mathbb{T} \mathrm{Re}\left( \frac{-i}{c^2} \partial_t z\overline{z}\right) \,\mathrm{d}x, \end{equation} as well as the energy \begin{equation}\label{eq:eni} E(z) = \int_ \mathbb{T} \vert \textstyle \frac{1}{c} \partial_t z\vert^2 + \vert \nabla z\vert^2 + \vert c z\vert^2 - \textstyle \frac{1}{p+1}\vert z\vert^{2p+2}\,\mathrm{d}x. \end{equation} Hence for all times $t$ where the solution exists and is smooth enough, $$ Q(z(t)) = Q(z(0)) = \int_ \mathbb{T} \mathrm{Im}(\overline{\varphi}\gamma)\,\mathrm{d}x, $$ and $$ E(z(t)) = E(z(0)) = c^2\big( \Vert \varphi \Vert_{L^2}^2 + \Vert \gamma \Vert_{L^2}^2\big)+ \int_ \mathbb{T} \vert \nabla \varphi \vert^2 - \textstyle \frac{1}{p+1}\vert \varphi\vert^{2p+2}\,\mathrm{d}x. $$ Using the result given by Theorem \ref{thm:main}, and under smoothness assumptions on $\varphi$ and $\gamma$, we can easily prove that the charge and energy admit asymptotic expansion with respect to $c^{-2}$. In this section, we give the expression of the first terms of these expansions. We commence with the asymptotic expansion of the charge $Q(z)$, i.e. $$ Q(z) = Q_0(\mathsf{U}_0,\mathsf{V}_0) + \sum_{k\geq 1} c^{-2k} Q_k (\mathsf{U}_0,\ldots,\mathsf{U}_{k},\mathsf{V}_0,\ldots,\mathsf{V}_{k}). $$ Using \eqref{eq:uv} we can easily (as we are working on the torus $ \mathbb{T} $ and thus integration by parts yields no boundary-terms) rewrite the charge as \begin{equation} \label{eq:quqv} \begin{aligned} Q(z) &= \frac14 \int_ \mathbb{T} \mathrm{Re}\left( \big(\frac{\langle \nabla \rangle_c}{c} (u-\overline{v})\big)(\overline{u}+v)\right)\,\mathrm{d}x\\ & = \frac14 \int_ \mathbb{T} \vert u\vert^2 - \vert v\vert^2 + \frac{1}{2c^2}( \vert \nabla u\vert^2 - \vert \nabla v\vert^2) \,\mathrm{d}x+ \mathcal{O}(c^{-4}). \end{aligned} \end{equation} Thus, using the asymptotic expansion given in Theorem \ref{thm:main}, we obtain \begin{equation}\label{eq:Q0} Q_0(\mathsf{U}_0,\mathsf{V}_0) = \frac{1}{4}\left(\Vert u_0\Vert_{L^2}^2 -\Vert v_0\Vert_{L^2}^2\right), \end{equation} where $u_0, v_0$ solve the nonlinear Schr\"odinger system \eqref{eq:nlsw0} and hence in particular their $L^2$ norms are conserved. Thus, the first term of the charge reads \begin{equation}\label{eq:Q0} Q_0(\mathsf{U}_0(t),\mathsf{V}_0(t)) = \frac{1}{4}\left(\Vert u_0(0)\Vert_{L^2}^2 -\Vert v_0(0)\Vert^2_{L^2}\right). \end{equation} Using the relation $z = \textstyle \frac12 (u+\overline{v})$ as well as $\partial_t z = \textstyle \frac12 i c \langle \nabla \rangle_c (u-\overline{v})$, see \eqref{eq:NLSc}, we obtain $$ E(z) = \frac12 \int_ \mathbb{T} c^2 (\vert u\vert^2 + \vert v\vert^2) + \vert \nabla u\vert^2 + \vert \nabla v\vert^2 - \textstyle \frac{2}{p+1} \vert \textstyle \frac12(u+\overline{v})\vert^{2p+2}\,\mathrm{d}x + \mathcal{O}(c^{-2}). $$ Hence similarly to $Q(z)$, we obtain an asymptotic expansion of the form $$ E(z) = c^2 E_{-2}(\mathsf{U}_0,\mathsf{V}_0) + \sum_{k\geq 0} c^{-2k} E_k (\mathsf{U}_0,\ldots,\mathsf{U}_{k},\mathsf{V}_0,\ldots,\mathsf{V}_{k}), $$ with $$ E_{-2}(\mathsf{U}_0,\mathsf{V}_0) = \frac{1}{2}\left(\Vert u_0\Vert_{L^2}^2 + \Vert v_0\Vert_{L^2}^2\right). $$ In order to analyze the asymptotic behavior of the energy $E(z)$, we first have to subtract the relativistic rest-energy $c^2Q_{\text{mass}}$ (see below) as from the energetic point of view the non-relativistic limit regime means that $ \mathcal{E} := E - c^2Q_{\text{mass}} $ is small. According to \eqref{eq:quqv}, we can write $$ Q(z) = Q^u(u) + Q^v(v), $$ where in the sense of asymptotic expansion (see \eqref{eq:alphn}), $$ Q^u (u) = \frac14 \int_ \mathbb{T} \vert u\vert^2+ \frac{1}{2c^2} \vert \nabla u\vert^2 + \sum_{j\geq 2} \alpha_j c^{-2j} \vert \nabla^{j} u\vert \,\mathrm{d}x, $$ and where $Q^v(v)$ denotes the part of $Q(z)$ which contains the antiparticle component, i.e. $$ Q^v(v) = Q^u (-v). $$ With this notation, we define the rest energy $c^2Q_{\text{mass}}(z) = 2 c^2(Q^u(u) - Q^v(v))$. Using these notations, we obtain an asymptotic expansion of the form \begin{equation}\label{eq:asympE} \begin{aligned} \mathcal{E} (z) &:= E(z) - 2c^2 (Q^u-Q^v) \\ &\; = \mathcal{E}_0(\mathsf{U}_0,\mathsf{V}_0) + \sum_{k\geq 1} c^{-2k} \mathcal{E} _k (\mathsf{U}_0,\ldots,\mathsf{U}_{k},\mathsf{V}_0,\ldots,\mathsf{V}_{k}). \end{aligned} \end{equation} Using the asymptotic expansion given in Theorem \ref{thm:main}, we calculate that the first-order correction term of the energy $ \mathcal{E} (z)$ in \eqref{eq:asympE} is given by \begin{equation}\label{eq:E0} \mathcal{E} _0(\mathsf{U}_0,\mathsf{V}_0) = \int_ \mathbb{T} \frac14\left (\vert \nabla u_0 \vert^2 + \vert \nabla v_0\vert^2\right) - \frac{1}{p+1} \left \vert \textstyle \frac12(u_0+\overline{v_0}\mathrm{e}^{-2i c^2t})\right\vert^{2p+2}\,\mathrm{d}x, \end{equation} where $u_0,v_0$ solve the nonlinear Schr\"odinger system \eqref{eq:nlsw0}. \section{Construction of robust numerical schemes}\label{sec:nume} The aim of this section is to show that we can use the derived asymptotic expansions to construct numerical schemes for the non-relativistic Klein-Gordon equation \eqref{eq:kgr} which are robust with respect to the speed of light $c$. Firstly we underline our theoretical results in the linear case, where the exact solution to \eqref{eq:kgr} is analytically known, see Section \ref{sec:lin}. Finally, in Section \ref{sec:cub} we construct robust numerical methods for the cubic non-relativistic Klein-Gordon equation and discuss their convergence. Numerical experiments are included. \subsection{Linear case}\label{sec:lin} In the linear case, i.e. $p = 0$, our model problem reads \begin{equation}\label{eq:kgrlin} \begin{aligned} &\frac{1}{c^2} \partial_{tt} z - \Delta z + c^2 z = \lambda z,\\ & z(0,x) = \varphi=\sum_{a\in \mathbb{Z} }\hat{\varphi}_a \mathrm{e}^{iax}, \quad \partial_t z(0,x) =c^2\gamma=c^2\sum_{a\in \mathbb{Z} }\hat{\gamma}_a \mathrm{e}^{iax} \end{aligned} \end{equation} with the exact solution \begin{equation}\label{eq:zlin} \begin{aligned} z(t,x) = \sum_{a\in \mathbb{Z} }\Big (\hat{\varphi}_a\cos(c t&\sqrt{a^2+c^2-\lambda})\\&+\frac{c}{\sqrt{a^2+c^2-\lambda}}\hat{\gamma}_a\sin(ct\sqrt{a^2+c^2-\lambda})\Big)\mathrm{e}^{iax}. \end{aligned} \end{equation} In the linear case we can solve the first- and second-order limit systems exactly in Fourier, see Corollary \ref{cor:lin} and equation \eqref{eq:u0lin} and \eqref{eq:u1lin}, respectively. Thus, we can state the first- and second correction terms $z_0$ and $z_1$ in an exact way. A simple computation shows that \begin{equation}\label{eq:z1lin} z_0(t,x)= \sum_{a\in \mathbb{Z} }\Big(\textstyle \hat{\varphi}_a \cos\Big(\big(c^2+\frac{1}{2}(a^2-\lambda)\big)t\Big)+\hat{\gamma}_a\sin\Big(\big(c^2+\frac{1}{2}(a^2-\lambda)\big)t\Big)\Big)\mathrm{e}^{iax}. \end{equation} Thus, we nicely see that the difference between the exact solution $z$ of the linear Klein-Gordon equation and its first-order correction term $z_0$ in the non-relativistic limit $c\gg 1$ are the formal approximations \begin{equation*} c\sqrt{a^2+c^2-\lambda} \approx c^2+\frac{1}{2}(a^2-\lambda),\quad \frac{c}{\sqrt{a^2+c^2-\lambda}}\approx 1, \end{equation*} which are of order $\mathcal{O}(c^{-2})$ for sufficiently smooth initial values $\varphi$ and $\gamma$. For the second correction term $z_1$ we get after a short computation \begin{equation} \begin{aligned} z_1(t,x) = &\sum_{a\in \mathbb{Z} } \textstyle \Big[\hat{\varphi}_a \frac{(a^2-\lambda)^2t}{8}\sin(\omega_{a,c} t)\\ &+ \textstyle \hat{\gamma}_a\Big( (-\frac{a^2-\lambda}{2})\sin(\omega_{a,c}t)-\frac{(a^2-\lambda)^2t}{8}\cos(\omega_{a,c}t)\Big)\Big]\mathrm{e}^{iax}. \end{aligned} \end{equation} where $$ \omega_{a,c} = c^2+\textstyle\frac{1}{2}(a^2-\lambda). $$ Again this can easily be motivated by formally taking the next approximations in the exact solution $z$, i.e. $$ c\sqrt{a^2+c^2-\lambda} \approx c^2+\frac{1}{2}(a^2-\lambda)-\frac{1}{8c^2}(a^2-\lambda)^2, $$ and $$\frac{c}{\sqrt{a^2+c^2-\lambda}}\approx 1 -\frac{a^2-\lambda}{2c^2}, $$ which are of order $\mathcal{O}(c^{-4})$ for sufficiently smooth initial values $\varphi$ and $\gamma$. The numerical results for the initial values $$ \varphi = \frac{(2+i)}{\sqrt{5}}\cos(x), \quad\gamma = \frac{(1+i)}{\sqrt{2}}\sin(x)+\frac{1}{2}\cos(x) $$ and $\lambda = -1$ on the time interval $[0,1]$ are given in Figure \ref{figure0}. \begin{figure}[h!] \centering \includegraphics[width=0.52\linewidth]{FIGLin.pdf \caption{Numerical simulations for the linear Klein-Gordon equation: Difference between the exact solution $z$ and the first- and second-order approximation terms $y = z_0$ (red, cross) and $y = z_0 + c^{-2}z_1$ (blue, star). The slope of the dash-dotted and dotted line is two and four, respectively.}\label{figure0} \end{figure} \subsection{Cubic non-linearity}\label{sec:cub} In the cubic setting, i.e. $p = 1$, our model problem reads \begin{equation} \label{eq:kgrc} \frac{1}{c^2} \partial_{tt} z - \Delta z + c^2 z = \lambda \vert z\vert^2 z, \quad z(0) = \varphi, \quad \partial_t z(0) =c^2\gamma \end{equation} and the limit Schr\"odinger system \eqref{eq:nsl0} reduces to \begin{equation}\label{eq:nsl0c} \begin{aligned} &i \partial_t u_0 - \frac{1}{2}\Delta u_0 - \frac{\lambda}{8}\big( \vert u_0\vert^2+2\vert v_0\vert^2\big)u_0 =0,\qquad u_0(0) = \varphi - i \gamma, \\ &i \partial_t v_0 - \frac{1}{2}\Delta v_0 -\frac{\lambda}{8}\big( \vert v_0\vert^2+2\vert u_0\vert^2\big)v_0=0,\qquad v_0(0) = \overline{\varphi + i \gamma}, \end{aligned} \end{equation} i.e. the first-order correction term $z_0$ is given by \begin{equation}\label{eq:corr1} z_0(t) = \textstyle \frac{1}{2} u_0(t)\mathrm{e}^{ic^2t} + \frac12 \overline{v_0}(t)\mathrm{e}^{-ic^2t}, \end{equation} where $(u_0, v_0)$ solves the system of nonlinear Schr\"odinger equations \eqref{eq:nsl0c}. Here the enhancement of our Ansatz is clearly seen: the limit Schr\"odinger system can easily be simulated without any $c$-dependent time step restriction, and a well suited approximation to the exact solution in the non-relativistic limit regime is then obtained by simply multiplying the numerical approximations of $u_0$ and $v_0$ with the highly-oscillatory phases $\mathrm{e}^{ic^2t}$ and $\mathrm{e}^{-ic^2t}$, respectively. For the numerical time integration of the first-order limit system, i.e. the Schr\"odinger system \eqref{eq:nsl0c}, we apply an exponential Strang splitting method, where we split the limit system in a natural way into the kinetic \begin{equation}\label{eq:NSLT} i\partial_t \begin{bmatrix} u_0 \\ v_0\end{bmatrix} = \frac{1}{2}\Delta\begin{bmatrix} u_0 \\ v_0\end{bmatrix} \end{equation} and potential part \begin{equation}\label{eq:NSLP} i\partial_t \begin{bmatrix} u_0 \\ v_0\end{bmatrix} = \frac{\lambda}{8}\begin{bmatrix} \vert u_0\vert^2+ 2\vert v_0\vert^2 & 0\\ 0 & \vert v_0\vert^2 +2 \vert u_0\vert^2 \end{bmatrix}\begin{bmatrix} u_0 \\ v_0\end{bmatrix}. \end{equation} The exact flow $\Phi^t(u_0(0),v_0(0))= \Phi^t_{T+P}(u_0(0),v_0(0))$ of the coupled system \eqref{eq:nsl0c} at time $t_n = n\tau$ is then approximated by \begin{equation}\label{eq:Lie} \Phi^{t_n}\approx (\Phi^{\tau/2}_P\circ\Phi^{\tau}_T \circ \Phi^{\tau/2}_P)^n, \end{equation} where $\Phi_T^t(u_0(0),v_0(0))$ and $\Phi_P^t(u_0(0),v_0(0))$ are the exact flows associated to the kinetic Hamiltonian \eqref{eq:NSLT} and potential Hamiltonian \eqref{eq:NSLP}, respectively. Note that the kinetic equation \eqref{eq:NSLT} can be solved exactly in Fourier space. Furthermore note that in the potential equation \eqref{eq:NSLP} the moduli of $u_0$ and $v_0$ are conserved quantities, i.e. $\vert u_0(t)\vert^2 = \vert u_0(0)\vert^2$ an $\vert v_0(t)\vert^2 = \vert v_0(0)\vert^2$ for all $t$ and thus we can solve both split equations in an exact way. For the space discretization, we use pseudo-spectral Fourier with mesh-size $h$ and grid points $x_j = {2jh\pi}$, $j = -K,\ldots,K-1$ where $K = 1/h \in \mathbb{N} $ is the number of frequencies. Note the fully discrete scheme can be easily implement using the Fast Fourier transform algorithm. We refer to \cite{Lubich08,Faou12} for convergence results over finite time of such splitting schemes. \begin{example}[First-order approximation term]\label{ex:firstCorr} In the first example we are interested in the asymptotic expansion up to the first-order correction term $z_0$. The asymptotic expansion derived in the previous section allows us the following convergence result for the natural choice of numerical approximation to $z_0$. \begin{theorem}\label{thm:con0} Let $s > d/2$ and assume that $\varphi$ and $\gamma$ are smooth functions defined on the one-dimensional torus $ \mathbb{T} $. Then there exist constants $T, c_0, C, h_0$ and $\tau_0$ such that the following hold: for $c > c_0$, $\tau < \tau_0$ and $h < h_0$, if we define by \begin{equation}\label{eq:z0num} z^{n,h}_0 := \textstyle \frac12 u_0^{n,h} \mathrm{e}^{ic^2t_n}+\frac12\overline{v_0^{n,h}}\mathrm{e}^{-ic^2t_n} \end{equation} the numerical approximation of the first-order approximation term $z_0(t)$ at time $t_n = n\tau$, where $u_0^{n,h}$ and $v_0^{n,h}$ denote the numerical approximations to the first-order limit system \eqref{eq:nsl0c} obtained by the Strang splitting \eqref{eq:Lie} at time $t_n = n\tau$ with a discrete grid of mesh-size $h$; then if $z(t)$ denotes the exact solution of the equation \eqref{eq:kgrc}, we have \begin{equation*} \Vert z(t_n) -z^{n,h}_0\Vert_{L^2} \leq C( \tau^2 + h^{s} + c^{-2})\text{ for all } t_n \leq T. \end{equation*} \end{theorem} \begin{proof} We have \begin{equation}\label{eq:boundi} \Vert z(t_n) - z^{n,h}_0\Vert_{L^2} \leq \Vert z(t_n) - z_0(t_n)\Vert_{L^2} + \Vert z_0(t_n) - z^{n,h}_0\Vert_{L^2}. \end{equation} By Theorem \ref{thm:main} we have over the time interval $(0,T)$ and $c > c_0$, \begin{equation}\label{eq:boundi1} \Vert z(t_n) - z_0(t_n)\Vert_{L^2}\leq C c^{-2}. \end{equation} To bound the second term we use the following inequality \begin{equation*} \begin{aligned} \Vert z_0(t_n) - z^{n,h}_0\Vert_{L^2} &\leq \Vert\mathrm{e}^{ic^2t_n}( u_0(t_n) - u_0^{n,h})\Vert_{L^2}+\Vert \mathrm{e}^{ic^2t_n}(v_0(t_n) - v_0^{n,h})\Vert_{L^2}\\ &\leq \Vert u_0(t_n) - u_0^{n,h} \Vert_{L^2}+\Vert v_0(t_n) - v_0^{n,h} \Vert_{L^2}. \end{aligned} \end{equation*} Using arguments similar to \cite[Theorem IV.17]{Faou12} (see also \cite{Lubich08}), it is then possible to prove that \begin{equation} \label{eq:umun} \Vert u_0(t_n) - u_0^{n,h}\Vert_{L^2}+\Vert v_0(t_n) - v_0^{n,h}\Vert_{L^2} \leq C( \tau^2 + h^{s} ), \end{equation} if the exact solution is smooth enough (i.e. belongs to some Sobolev space $H^{s'}$ with $s' > s$ sufficiently large). Note that the proof of the fully discrete scheme in \cite{Faou12} is given only for the Lie-splitting, real initial data (that is $\varphi$ and $\gamma$ real valued functions) and uses function spaces based on the Wiener algebra (the $\ell_s^1$ spaces). However, the rigorous proof of \eqref{eq:umun} is essentially a variation on the same theme and does not present any particular difficulty as soon as the solution is smooth enough. Using this bound, we obtain that \begin{equation}\label{eq:boundi2} \Vert z_0(t_n) - z^{n,h}_0\Vert_{L^2} \leq C( \tau^2 + h^{s} ). \end{equation} Plugging the bounds \eqref{eq:boundi1} and \eqref{eq:boundi2} into \eqref{eq:boundi} yields the desired convergence result. \end{proof} The numerical results for the initial data $$ \varphi = \frac{(2+i)}{\sqrt{5}}\cos(x),\,\gamma = \frac{(1+i)}{\sqrt{2}}\sin(x)+\frac{1}{2}\cos(x) $$ and $\lambda = -1$ on the time interval $[0,0.1]$ are given in Figure \ref{figure1}. As a reference solution for $z(t)$ we take the numerical approximation of the Klein-Gordon equation \eqref{eq:kgrc} obtained with an exponential Gautschi method, proposed in \cite{BaoD12}, with a very small time step $\tau$ satisfying the necessary condition $\tau < c^2 h$. Note that the convergence result for the exponential Gautschi method given in \cite[Theorem 9]{BaoD12} shows a convergence rate of order $\mathcal{O} (\tau^2 c^4 + h^{s})$ under some {\em a priori} smoothness assumptions and uniform bounds for the exact solution. \begin{figure}[h!] \centering \includegraphics[width=0.52\linewidth]{figctest1.pdf \caption{Numerical simulations of Example \ref{ex:firstCorr}: Difference between the reference solution $z$ obtained with an exponential Gautschi method with time step $\tau = 10^{-6}$ and the first-order approximation solution $y= z_0^{n,h}$ obtained with $\tau = 10^{-2}$. The slope of the dash-dotted line is two.}\label{figure1} \end{figure} \end{example} \begin{example}[Second-order approximation term]\label{ex:2ndCorr} In this example we simulate the first- and second-order approximation terms $z_0$ and $z_1$ given in \eqref{eq:corr1} and \eqref{eq:corr2}, respectively. For the numerical approximation of the first approximation term we proceed as in Example \ref{ex:firstCorr}. For the numerical approximation of the correction term $z_1$, we solve the equation for $\xi_1$, i.e. equation \eqref{eq:u1rel}, with a symmetric Strang splitting based on the following split-equations: The kinetic equation \begin{equation*} i\partial_t \xi = \textstyle \frac12 \Delta \xi, \end{equation*} which can be solved exactly in Fourier, and the linear potential equation \begin{equation} \label{eq:poteq} i\partial_t \xi = \lambda \textstyle \frac34 \vert u_0\vert^2 \xi +\lambda \frac38 (u_0)^2\overline{\xi}+ \textstyle \frac18 \Delta^2 u_0 +\lambda^2 \frac{51}{256}\vert u_0\vert^4 u_0 +\lambda \frac{3}{16}\Delta(\vert u_0\vert u_0), \end{equation} which we solve as described in the following. We set $\alpha_0 = \mathrm{Re}(u_0)$, $\beta_0 =\mathrm{Im}(u_0)$, $\alpha = \mathrm{Re}(\xi)$, $\beta = \mathrm{Im}(\xi)$, and furthermore $g_0(t) = g(\alpha_0(t),\beta_0(t)) = \textstyle \frac18 \Delta^2 u_0 +\lambda^2 \frac{51}{256}\vert u_0\vert^4 u_0 +\lambda \frac{3}{16}\Delta(\vert u_0\vert u_0)$. Then we can rewrite the potential equation as \begin{equation}\label{eq:poti} \begin{aligned} \partial_t \begin{bmatrix} \alpha \\ \beta\end{bmatrix} = -\lambda \frac38\begin{bmatrix} 2 \alpha_0\beta_0\qquad &( (\alpha_0)^2+3 (\beta_0)^2)\\ - ((\beta_0)^2+3 (\alpha_0)^2)\qquad &- 2 \alpha_0\beta_0\end{bmatrix}\begin{bmatrix} \alpha \\ \beta\end{bmatrix} +\begin{bmatrix}\mathrm{Im}(g_0)\\ -\mathrm{Re}(g_0)\end{bmatrix}. \end{aligned} \end{equation} We set \begin{equation*} A(\alpha_0,\beta_0) = -\lambda \frac38\begin{bmatrix} 2 \alpha_0\beta_0\qquad &( (\alpha_0)^2+3 (\beta_0)^2)\\ - ((\beta_0)^2+3 (\alpha_0)^2)\qquad &- 2 \alpha_0\beta_0\end{bmatrix} \end{equation*} and approximate the potential equation \eqref{eq:poti} with an exponential trapezoidal rule, i.e. for $\alpha^n_k \approx \alpha_k(t_n)$, $\beta^n_k \approx \beta_k(t_n)$, $k=0,1$, $g^n = g(\alpha^n_0,\beta^n_0)$, $t_n = n\tau$, we set \begin{multline} \label{eq:num2corr} \begin{bmatrix} \alpha^{n+1}\\\beta^{n+1}\end{bmatrix}= \mathrm{e}^{\frac{\tau}{2}\big(A(\alpha_0^{n+1},\beta_0^{n+1})+A(\alpha_0^{n},\beta_0^{n})\big)}\textstyle\left(\begin{bmatrix} \alpha^{n}\\ \beta^{n}\end{bmatrix} + \frac{\tau}{2}\begin{bmatrix}\mathrm{Im}(g_0^n)\\ -\mathrm{Re}(g_0^n)\end{bmatrix} \right) \\� +\textstyle\frac{\tau}{2}\begin{bmatrix}\mathrm{Im}(g_0^{n+1})\\ -\mathrm{Re}(g_0^{n+1})\end{bmatrix}. \end{multline} The numerical approximation $\xi^{n+1}$ to the exact solution $\xi(t_{n+1})$ of the potential equation \eqref{eq:poteq} starting in a given point $\xi^n$ is then simply given by $\xi^{n+1} = \alpha^{n+1} + i \beta^{n+1}$. Concerning the space discretization, we use again the Fourier pseudo-spectral method described in the previous Section, and we denote by $h$ the corresponding mesh size. Note that in this setting, the approximation of $g_0$ can be easily made using FFT. Let $u_0^{n,h}$ be defined as in the previous section, and let $\xi^{n,h}_1$ denote the numerical approximation to the full system \eqref{eq:u1rel} at time $t_n$ obtained by the fully discrete Strang splitting described above, where we make half a step with the kinetic, a full step with the potential and again half a step with the kinetic part. Then, as the numerical approximation to the exact second-order approximation term at time $t_n = n\tau$ we take \begin{multline} \label{eq:approxnh} z_0^{n,h} + c^{-2}z_1^{n,h}:= \textstyle \frac12(1+\frac{1}{c^2}\frac{3\lambda}{16}\vert u^{n,h}_0\vert^2) (u_0^n\mathrm{e}^{ic^2t_n}+\overline{u_0^{n,h}}\mathrm{e}^{-ic^2t_n}) \\� - \textstyle\frac{\lambda}{64 c^2}\big( (u_0^{n,h})^3\mathrm{e}^{3ic^2t_n} +(\overline{u_0^{n,h}})^3\mathrm{e}^{-3ic^2t_n}\big) + \textstyle \frac{1}{2c^2} \big( \xi_1^{n,h} \mathrm{e}^{ic^2t_n}+\overline{\xi_1^{n,h}}\mathrm{e}^{-ic^2t_n}\big). \end{multline} For this approximation we have the following convergence result. \begin{theorem} Let $s > d/2$ and assume that $\varphi$ and $\gamma$ are smooth functions defined on the one-dimensional torus $ \mathbb{T} $. Then there exist constants $T, c_0, C, h_0$ and $\tau_0$ such that the following hold: for $c > c_0$, $\tau < \tau_0$ and $h < h_0$, then if $z(t)$ denotes the exact solution of the equation \eqref{eq:kgrc}, we have \begin{equation*} \Vert z(t_n) -(z_0^{n,h} + c^{-2}z_1^{n,h})\Vert_{L^2} \leq C( \tau^2 + h^{s} + c^{-4})\text{ for all } t_n \leq T, \end{equation*} where the approximation $z_0^{n,h} + c^{-2}z_1^{n,h}$ is defined in \eqref{eq:approxnh}. \end{theorem} \begin{proof} Note that by Theorem \ref{thm:main} we have over the time interval $(0,T)$ and $c>c_0$, $$ \Vert z(t_n) - (z_0(t_n) + c^{-2}z_1(t_n))\Vert_{L^2} \leq C c^{-4}.$$ The rest of the proof follows the line of argumentation given for the proof to Theorem \ref{thm:con0}, since the scheme described in \eqref{eq:num2corr} is second-order convergent. We do not give the detail of this proof. Again it is a variation on the same theme as in \cite{Faou12,Lubich08}. \end{proof} The numerical results for the initial data $$ \varphi = \cos(x), \,\gamma =\frac14 \sin(x)+\frac12 \cos(x) $$ and $\lambda = -1$ on the time interval $[0,0.1]$ are given in Figure \ref{figure2b}. \begin{figure}[h!] \centering \includegraphics[width=0.52\linewidth]{figctest2.pdf \caption{Numerical simulations of Example \ref{ex:2ndCorr}: Difference between the reference solution $z$ obtained with an exponential Gautschi method with time step $\tau = 10^{-7}$ and the first- and second-order approximation solutions $y = z_0^{n,h}$ (red , cross) and $y = z_0^{n,h} + c^{-2}z_1^{n,h}$ (blue, star) obtained with $\tau = 10^{-3}$. The slope of the dash-dotted and dotted line is two and four, respectively.}\label{figure2b} \end{figure} \end{example} \begin{example}[Energy and charge-density deviation]\label{ex:EQ} Here we simulate the deviation of the the first energy and charge correction term \eqref{eq:E0} and \eqref{eq:Q0}, respectively. Also, we simulate the energy and charge deviation of the first-correction term $z_0(t)$, i.e. $E(z_0(t))$ and $Q(z_0(t))$, where the energy and charge are defined in \eqref{eq:eni} and \eqref{eq:chargi}, respectively. We set $E(z_0^{n,h})= E(u_0^{n,h}\mathrm{e}^{ic^2t_n},v_0^{n,h}\mathrm{e}^{ic^2t_n})$ the numerical energy of the first correction term $z_0$, for $t_n = n\tau$, and where $u_0^{n,h}$ and $v_0^{n,h}$ are the numerical approximations of the first limit system \eqref{eq:nsl0c} again obtained by the Strang splitting \eqref{eq:Lie}. Similarly, we set $ \mathcal{E} _0^{n,h}= \mathcal{E} _0(u_0^{n,h}\mathrm{e}^{ic^2t_n},v_0^{n,h}\mathrm{e}^{ic^2t_n})$ the approximation of the first term in the asymptotic expansion of the energy. In the same way, we define the numerical charge $Q(z_0^{n,h})= Q(u_0^{n,h}\mathrm{e}^{ic^2t_n},v_0^{n,h}\mathrm{e}^{ic^2t_n})$ of the first correction term, as well as the numerical approximation of the first term in the asymptotic expansion $Q_0^{n,h} = Q_0(u_0^{n,h},v_0^{n,h})$. We carry out the simulations on the time interval $[0,1]$ with the time step $\tau = 10^{-3}$. As the initial values we choose the same as in Example \ref{ex:firstCorr}, but normalize them with respect to the $H^1$ norm. We nicely see that the first energy correction term $ \mathcal{E} _0$ and the first charge correction term $Q_0$ are conserved by the Strang splitting, see Figure \ref{figureE} and Figure \ref{figureQ}, respectively. Furthermore, Figure \ref{figureQz0} and Figure \ref{figureEz0} underline that $\vert Q(z_0^{0,h})-Q(z_0^{n,h})\vert = \mathcal{O}(c^{-2})$ and $\vert E(z_0^{0,h})-E(z_0^{n,h})\vert = \mathcal{O}(1)$. \begin{figure}[h!] \centering \includegraphics[width=0.99\linewidth]{dE2.pdf} \caption{Numerical Simulation of Example \ref{ex:EQ}: Energy deviation of the first correction term $ \mathcal{E} _0^{n,h}$, for different values of oscillations $c$.}\label{figureE} \end{figure} \begin{figure}[h!] \centering \includegraphics[width=0.52\linewidth]{dQ2.pdf} \caption{Numerical Simulation of Example \ref{ex:EQ}: Charge deviation of the first correction term $Q_0^{n,h}$.}\label{figureQ} \end{figure} \end{example} \begin{figure}[h!] \centering \includegraphics[width=0.99\linewidth]{Qz0.pdf} \caption{Numerical Simulation of Example \ref{ex:EQ}: Charge deviation of $Q(z_0^{n,h})$.}\label{figureQz0} \end{figure} \begin{figure}[h!] \centering \includegraphics[width=0.99\linewidth]{Ez0.pdf} \caption{Numerical Simulation of Example \ref{ex:EQ}: Energy deviation of $E(z_0^{n,h})$.}\label{figureEz0} \end{figure}

Introduction {#ece32004-sec-0001} ============ It is well recognized that the Earth\'s biodiversity is undergoing significant change resulting from various types of human activity (Vitousek et al. [1997](#ece32004-bib-0048){ref-type="ref"}; Chapin et al. [2000](#ece32004-bib-0009){ref-type="ref"}; Sala et al. [2000](#ece32004-bib-0035){ref-type="ref"}; Butchart et al. [\[Link\]](#ece32004-bib-0007){ref-type="ref"}; Barnosky et al. [2011](#ece32004-bib-0002){ref-type="ref"}). Studies that explore changes in species richness resulting from anthropogenic drivers of change are essential to understand how human activity affects biodiversity and have important implications for ecosystem management and policy formation. While a number of attempts to synthesize the data from local‐scale biodiversity change studies have been conducted, results have been mixed, with some studies reporting significant decreases in local‐scale species richness (Zvereva et al. [2008](#ece32004-bib-0052){ref-type="ref"}; Gerstner et al. [2014](#ece32004-bib-0019){ref-type="ref"}; Murphy and Romanuk [2014](#ece32004-bib-0030){ref-type="ref"}; Newbold et al. [2015](#ece32004-bib-0031){ref-type="ref"}), while others have found little to no change (Vellend et al. [2013](#ece32004-bib-0045){ref-type="ref"}; Dornelas et al. [2014](#ece32004-bib-0013){ref-type="ref"}; Supp and Ernest [2014](#ece32004-bib-0040){ref-type="ref"}). Interestingly, the syntheses that have found significant declines in species richness are those that have specifically focused on the effects of human drivers of change, while the syntheses that have found little change in species richness included a variety of locations ranging from disturbed to pristine. The significant changes that have been found in local‐scale species richness resulting from human activity highlight the need to further examine where on Earth these changes are occurring and if there are consistent patterns of change in certain biomes resulting from specific drivers. Each of these syntheses has also highlighted similar biases in terms of where local‐scale species richness change studies are being conducted, with the majority of studies taking place in North America and Europe. Unfortunately, these data gaps severely limit ecologist\'s ability to make predictions about where and why local‐scale biodiversity change is occurring around the globe. Several studies have shown that geographic biases exist in the ecological literature (Butchart et al. [\[Link\]](#ece32004-bib-0007){ref-type="ref"}; Pereira et al. [2010](#ece32004-bib-0033){ref-type="ref"}; Ahrends et al. [2011](#ece32004-bib-0001){ref-type="ref"}). A 2012 review of terrestrial ecological observations for a 5‐year period reported an over‐representation of studies in temperate biomes and protected areas (Martin et al. [2012](#ece32004-bib-0027){ref-type="ref"}). While the overall geographic distribution of ecological studies has been reviewed, geographic biases in local‐scale species richness change resulting from anthropogenic drivers have not been analyzed. Given the recent popularity in synthesizing the results of local‐scale species richness change studies to make predictions about how local‐scale biodiversity is changing around the globe, a more detailed analysis of the geographic biases that exist in the literature is necessary. In this study, we examine the geographic biases that exist in terrestrial local‐scale species richness change studies and provide information about where future local‐scale species richness change studies need to be conducted. We specifically focus on species richness change resulting from anthropogenic global change drivers (land‐use change, species invasions, nutrient addition, and warming). We examine how the geographic biases differ for these drivers across the Earth\'s terrestrial biomes and determine the data gaps that exist based on three circumstances: (1) the land area that the biome covers; (2) the future change that the drivers are projected to have in each biome; and (3) the impact that the drivers have on the biodiversity of each biome. The sensitivity of a particular biome to a global change driver will depend on the unique set of features that define the biome, and thus, it is expected that the effects of various global change drivers will differ across the Earth\'s biomes (Sala et al. [2000](#ece32004-bib-0035){ref-type="ref"}). For example, the impacts of climate change on productivity have been shown to differ widely among the Earth\'s biomes (Silva and Anand [2013](#ece32004-bib-0037){ref-type="ref"}; Zhang et al. [2013](#ece32004-bib-0051){ref-type="ref"}). Various studies have assessed the magnitude of global change drivers across the Earth\'s biomes and have provided estimates of which biomes are most vulnerable to certain drivers. The projections of future change across biomes resulting from various global change drivers demonstrate that certain data gaps in local‐scale species richness change studies might be justified based on the level of threat of a particular driver in a specific biome. For example, estimates of land‐use change for the year 2050 predict large proportions of future land‐cover change in temperate broadleaf and mixed forests along with temperate and tropical coniferous forests and project a much lower proportion of change in the tundra and boreal forest biomes (Lee and Jetz [2008](#ece32004-bib-0024){ref-type="ref"}). Therefore, we would expect to find the majority of studies examining the effects of land‐use change on local‐scale species richness to have been conducted in temperate forests and tropical coniferous forests. Projections of net change in the number of 100 of the world\'s worst invasive alien species for the year 2100 (Bellard et al. [2013](#ece32004-bib-0003){ref-type="ref"}) estimate that the highest increase in invasive species will occur in temperate mixed forests. Therefore, an over‐representation of invasion studies in the temperate forest biomes is well justified. By comparing the data gaps in local‐scale species richness change studies with estimates of which biomes will be most and least affected by global change drivers, we aim to identify the most critical data gaps and make suggestions as to where future local‐scale species richness change studies need to be conducted to fill in these gaps. Knowledge of how global change drivers will impact local‐scale species richness and how these impacts compare across biomes is dependent upon available information of these impacts across the Earth\'s biomes, for which the ecological literature is currently lacking. The findings from large‐scale syntheses of local‐scale species richness change will play crucial roles in how human‐impacted ecosystems are managed in the future. Unfortunately, predictions for how human activity will impact local‐scale biodiversity around the globe are seriously hindered by the data gaps that exist in the literature. It is necessary for future biodiversity research to focus on biodiversity change in biomes that are expected to experience significant impacts from a human driver of change and where data on that driver is lacking. Here, we use a dataset of 767 responses of change in species richness taken from 638 human‐mediated species richness change studies to determine where geographic data gaps exist across the various global change drivers, and discuss whether these biases are justified based on the varying levels of threat that the drivers pose to the biomes. Methods {#ece32004-sec-0002} ======= Selection criteria {#ece32004-sec-0003} ------------------ Our dataset was compiled by searching the biological literature for studies that reported the effects of anthropogenic drivers of change on local‐scale species richness in a terrestrial habitat. Four anthropogenic drivers of change that have been identified as major drivers of current biodiversity change were included in the dataset: land‐use change, species invasions, nutrient addition, and warming (Vitousek et al. [1997](#ece32004-bib-0048){ref-type="ref"}). We performed a literature search using the ISI Web of Science database using the following search expressions: "biodiversity "OR "species richness" OR "community change" AND ("invasi\* species" OR "habitat loss" OR "land use change" OR "climate change" OR "warm\*" OR increas\* temperature" OR "eutrophication" OR "nutrient add\*"). We also searched for studies in the references of relevant meta‐analyses and syntheses. A final search of the literature was completed on 17 October 2015. We included studies that experimentally manipulated one of the four drivers of change or observational studies that compared an impacted site with a control (nonimpacted) site. The literature search yielded 638 suitable studies with 767 values of change in species richness that were included in the final dataset. Only studies conducted in terrestrial ecosystems were included in the dataset. This dataset specifically focuses on studies that examine change in species richness in control vs. impacted treatments. Studies that assessed other metrics like evenness and biodiversity indices, those that assessed changes in community composition, and those that examined species richness change using time‐series or before--after studies were not included in the dataset. We chose to focus only on species richness, as it is the most common biodiversity measure used in studies examining the impact of anthropogenic global change drivers. Geographic coordinates were either taken from the papers or in cases where the coordinates were not given they were estimated using Google Earth (Google Inc., California, USA). Studies that reported multiple measures of species richness for the same geographic location (i.e., different species, driver intensities, etc.) were only included in the dataset once and the values were averaged when calculating effect sizes. We only included multiple responses from the same study when they were taken from different countries or biomes. The selection criterion is similar to that in Murphy and Romanuk ([2014](#ece32004-bib-0030){ref-type="ref"}) and a more detailed explanation can be found there. For the purposes of examining data gaps, the habitat loss and land‐use change categories have been combined. Also, for the current study we only included invasion studies that give a measure of species richness change in the entire community rather than just native species. Data analysis {#ece32004-sec-0004} ------------- To investigate the data gaps existing across the Earth\'s biomes, we categorized study responses using the 14 global terrestrial ecoregions identified by Olson et al. ([2001](#ece32004-bib-0032){ref-type="ref"}). We visualized the global distribution of studies by entering the locations of all 767 responses into ArcGIS 10.1 (ESRI [2011](#ece32004-bib-0016){ref-type="ref"}). We used chi‐squared tests to determine whether specific biome--driver combinations are over‐represented or under‐represented. We tested for significant differences between the observed and expected distribution of studies based on three sets of circumstances: (1) the relative land area that the biome covers (e.g., biomes with a larger land area are expected to contain a proportionately greater number of responses); (2) the future change that the drivers are projected to have in each biome (e.g., biomes projected to experience a larger increase in a driver are expected to contain a greater number of responses); and (3) the impact that the drivers have on biodiversity in each biome (e.g., biomes where a driver has a larger impact on biodiversity are expected to contain a greater number of responses). We used projections from various studies to determine the future change of drivers among biomes. For land‐use change, we used the projected proportions of land‐cover transformation due to land‐use change for the year 2050 given in Lee and Jetz ([2008](#ece32004-bib-0024){ref-type="ref"}). Estimates are given for 57 biome--realm combinations and for four socioeconomic scenarios. We averaged the estimates for the biome--realm combinations and across the four socioeconomic scenarios to get one estimate of future land‐use change for each biome. Using these estimates, we determined the proportion of projected land‐use change for each biome for the year 2050 (Table [1](#ece32004-tbl-0001){ref-type="table-wrap"}). For species invasions, we used estimates of the percentage of net change in the numbers of 100 of the world\'s worst invasive species for the year 2080 given in Bellard et al. ([2013](#ece32004-bib-0003){ref-type="ref"}). The biomes categorized in this study slightly differ from our categorization so in some cases the estimates from multiple biomes were averaged. As this study estimates that some biomes will become less suitable for invasive species (i.e., the tropics), the values given for percentage of net change are negative. To calculate the proportion of change in invasive species for each biome and the expected number of studies, we first added 5 to each estimate, thus making all estimates positive (Table [1](#ece32004-tbl-0001){ref-type="table-wrap"}). For nutrient addition, we used the expected changes for the year 2100 given in Sala et al. ([2000](#ece32004-bib-0035){ref-type="ref"}). This study ranks the biomes from 1 to 5 based on future change in nitrogen deposition, and using these ranks, we calculated the proportion of change for each biome (Table [1](#ece32004-tbl-0001){ref-type="table-wrap"}). To determine the expected number of warming studies based on future change, we used the climate anomalies between pre‐industrial conditions and projections for the year 2100 given in Benito‐Garzon et al. ([2014](#ece32004-bib-0004){ref-type="ref"}) (Table [1](#ece32004-tbl-0001){ref-type="table-wrap"}). To calculate the expected number of studies based on the impact that the drivers have on the biodiversity of each biome, we used the estimates given in Sala et al. ([2000](#ece32004-bib-0035){ref-type="ref"}) where the impact of a large change in each driver on biodiversity is ranked from 1 to 5 for each biome. The proportion of impact that each driver has on each biomes biodiversity was calculated from these rankings (Table [1](#ece32004-tbl-0001){ref-type="table-wrap"}). ###### Percentage of anthropogenically driven local‐scale species richness change studies conducted in each biome for four drivers of change, along with the expected percentage of studies based on the land area covered by each biome, the projected future change, and the impact of the driver on biodiversity. Values highlighted indicate that the biome--driver combination is significantly over‐ (blue) or under‐ (red) represented (*P* \< 0.05) based on the highlighted circumstance {#nlm-graphic-3} John Wiley & Sons, Ltd We conducted a separate chi‐squared test for each driver of change and used the total number of studies for each driver to calculate the expected number of studies. For example, the total number of land‐use change studies across all biomes is 342; therefore, to calculate the expected number of studies in each biome for the chi‐squared test based on land area, we multiplied the proportion of land area that each biome covers by 342 to give the expected values. These values were then compared to the observed number of studies in each biome in the chi‐squared test. Biome--driver combinations were considered significantly over‐ or under‐represented if the *P*‐value was equal to or less than 0.05. We used the standard deviation of the average effect sizes for each biome--driver combination as a measure of how variable the magnitude of change in species richness is in each biome. The effect size compares species richness between the control (*c*) and impacted (*i*) treatments in each study and was calculated as the response ratio (Hedges et al. [1999](#ece32004-bib-0500){ref-type="ref"}): $$\text{RR} = \ln\left( \frac{{\overline{X}}_{i}}{{\overline{X}}_{c}} \right)$$ Results {#ece32004-sec-0005} ======= Global study distribution {#ece32004-sec-0006} ------------------------- Our results reveal several biases in the global distribution of anthropogenically driven, local‐scale species richness change studies. Studies included in the dataset were conducted in 79 countries. The five countries with the greatest number of responses were as follows: USA (23.6% of responses), China (8.1% of responses), Brazil (6.5% of responses), Canada (5% of responses), and Australia (3.7% of responses). The Palearctic and Nearctic are the most commonly studied biogeographic realms with 33.8% and 28.6% of studies conducted in these realms, respectively. Across all studies, the Afrotropic and Antarctic realms are significantly under‐represented based on land area (*P* \< 0.001 and *P* = 0.021, df = 7, respectively), while the Nearctic and Oceanic realms are significantly over‐represented (*P* \< 0.001, df = 7). Based on land area, land‐use change studies are significantly over‐represented in the Neotropics (*P* \< 0.001, df = 7) and are significantly under‐represented in the Palearctic (*P* = 0.001, df = 7). Both species invasion and nutrient addition studies are significantly under‐represented in the Afrotropics (*P* = 0.014 and *P* \< 0.001, df = 7, respectively) and significantly over‐represented in the Nearctic and Oceanic (*P* \< 0.001, df = 7) realms. Warming studies are significantly over‐represented in the Nearctic realm (*P* \< 0.001, df = 7). Study distribution by biome {#ece32004-sec-0007} --------------------------- Across all drivers, temperate broadleaf and mixed forests (31%) and tropical moist broadleaf forests (23%) are the most frequently studied biomes (Figs. [1](#ece32004-fig-0001){ref-type="fig"}, [2](#ece32004-fig-0002){ref-type="fig"}). All other biomes contain fewer than 10% of responses. Based on land area, the boreal forest and tropical grassland biomes are significantly understudied (*P* \< 0.001, df = 12), while the temperate broadleaf and mixed forest, temperate coniferous forest, tropical moist broadleaf forest and Mediterranean biomes are significantly over‐represented (*P* \< 0.001, df = 12). {#ece32004-fig-0001} {ref-type="ref"}). Points are semitransparent with darker areas indicating an overlap of studies.](ECE3-6-2938-g002){#ece32004-fig-0002} There is notable disparity in how frequently studied a biome is when the four drivers of change are examined separately (Figs. [1](#ece32004-fig-0001){ref-type="fig"}, [2](#ece32004-fig-0002){ref-type="fig"}). One of the most significant areas of bias in our analysis of human driven local‐scale species richness change studies is the over‐representation of land‐use change studies in tropical and temperate broadleaf forest biomes. Almost half of the land‐use change studies are conducted in tropical moist broadleaf forests and almost a quarter of land‐use change studies are conducted in temperate broadleaf and mixed forests. Based on the proportion of land area that these biomes cover, they are significantly over‐represented, with three times the number of expected land‐use change responses in the tropical moist broadleaf forest (*P* \< 0.001, df = 12), and over two times, the number of expected responses in the temperate broadleaf and mixed forest (*P* \< 0.001, df = 12). All other biomes contain less than 6% of the land‐use change studies (Fig. [1](#ece32004-fig-0001){ref-type="fig"}), and based on land area, the tundra and desert biomes are under‐represented (*P* = 0.007 and *P* \< 0.001, df = 12, respectively). When analyzed based on the projection of future land‐use change and the impact that land‐use change has on the biodiversity of each biome, the tropical moist broadleaf forest remains over‐represented in both categories (*P* \< 0.001, df = 12). The temperate broadleaf and mixed forest biome is over‐represented based on the impact on biodiversity (*P* \< 0.001, df = 12), but not based on projected future land‐use change. Both the temperate and tropical coniferous forest biomes are significantly under‐represented based on the magnitude of land‐use change expected to occur in these biomes in the future (*P* \< 0.001, df = 12), and the tropical coniferous forest is almost significantly under‐represented based on the impact to biodiversity (*P* = 0.055, df = 12). The temperate grassland and tropical dry forest biomes have the lowest variance in the magnitude of species richness change following land‐use change (SD = 0.149, *n* = 7 and SD = 0.2, *n* = 12, respectively), while the desert and tropical coniferous forest biomes have the highest variance (SD = 0.793, *n* = 11 and SD = 0.869, *n* = 3, respectively) (Figure S1). Studies that examine the effect of species invasions on local‐scale species richness change are significantly over‐represented in the temperate broadleaf and mixed forests (p \< 0.001, df = 12; 38% of studies) based on all three categories (Table [1](#ece32004-tbl-0001){ref-type="table-wrap"}). Species invasion studies are also over‐represented in the Mediterranean biome based on land area (*P* \< 0.001, df = 12; 18%), but not based on the projected extent of exotic invasions and the impact of invaders on Mediterranean biodiversity. Invasion studies are relatively well represented in temperate grasslands, deserts, and tropical moist broadleaf forests, yet all other terrestrial biomes contain 5% or fewer of the invasive species responses (Table [1](#ece32004-tbl-0001){ref-type="table-wrap"}). The tundra and boreal forest biomes have the lowest variance in the magnitude of species richness change following species invasions (SD = 0.021, *n* = 2 and SD = 0.085, *n* = 4, respectively), while the temperate broadleaf forest (SD = 0.558, *n* = 60), temperate grassland (SD = 0.606, *n* = 20), and Mediterranean (SD = 0.559, *n* = 28) biomes have the highest variance (Figure S1). Over half of nutrient addition studies are conducted in the temperate region (Fig. [1](#ece32004-fig-0001){ref-type="fig"}, Table [1](#ece32004-tbl-0001){ref-type="table-wrap"}). Based on land area, nutrient addition studies are significantly over‐represented in temperate broadleaf and temperate coniferous forests (*P* \< 0.001, df = 12), but temperate broadleaf and mixed forests are the only biome that is significantly over‐represented based on the projected levels of future nutrient addition and the impact on biodiversity (Table [1](#ece32004-tbl-0001){ref-type="table-wrap"}). Nutrient addition studies are well represented in the tundra, yet fewer are conducted in the boreal forest (Table [1](#ece32004-tbl-0001){ref-type="table-wrap"}). The flooded grassland (SD = 0.046, *n* = 2), temperate grassland (SD = 0.151, *n* = 27), and tundra (SD = 0.159, *n* = 14) biomes have the lowest variance in the magnitude of species richness change following nutrient addition, while the tropical moist forest (SD = 0.928, *n* = 13) and boreal forest (SD = 0.474, *n* = 4) biomes have the highest variance (Figure S1). Similar to the other drivers, warming studies are over‐represented in temperate broadleaf and mixed forests based on land area, projections of future change, and impact on biodiversity (*P* \< 0.001, df = 12; 39% of studies). Warming studies are also well represented in the tundra biome (23%), yet are lacking in the boreal forest (5%) and nonexistent in the tropics with zero warming studies conducted in tropical grasslands and the three tropical forest biomes (Fig. [1](#ece32004-fig-0001){ref-type="fig"}, Table [1](#ece32004-tbl-0001){ref-type="table-wrap"}). The temperate grassland (SD = 0.142, *n* = 8) and temperate broadleaf forest (SD = 0.17, *n* = 32) biomes have the lowest variance in the magnitude of species richness change following warming, while the montane grassland (SD = 0.53, *n* = 6) and tundra (SD = 0.564, *n* = 19) biomes have the highest variance (Figure S1). Discussion {#ece32004-sec-0008} ========== It is evident that research intensity in the various terrestrial biomes is skewed toward different anthropogenic drivers of change. Below we discuss the differences in research intensity across biomes for each of the four human drivers of change and identify the most critical data gaps in terms of where future human driven local‐scale species richness change studies need to be conducted. Land‐use change {#ece32004-sec-0009} --------------- The majority of land‐use change studies are conducted in tropical moist forests and temperate broadleaf forests. It is not surprising that land‐use change studies are most common in these biomes as habitat loss due to human activity and land conversion has been most extensive in tropical and temperate forests (Miles et al. [2006](#ece32004-bib-0029){ref-type="ref"}; Potapov et al. [2009](#ece32004-bib-0034){ref-type="ref"}). Estimates of land‐use change for the year 2050 project an 11.5% change in land cover in tropical moist forests and a 19.5% change in tropical broadleaf and mixed forests worldwide (Lee and Jetz [2008](#ece32004-bib-0024){ref-type="ref"}). While the projections of future land‐cover change in these two biomes are large, there are other biomes with higher proportions of projected land‐cover change that are much less studied. These include temperate coniferous and tropical coniferous forests, which are projected to experience 23% and 28% land‐cover change, respectively, by the year 2050 (Lee and Jetz [2008](#ece32004-bib-0024){ref-type="ref"}), yet very few land‐use change studies that examine changes in local‐scale species richness are conducted in these biomes. This lack of studies is particularly concerning for tropical coniferous forests as not only is this biome significantly under‐represented based on both projections for future change and the impact of change on biodiversity but tropical coniferous forests also have the highest variance in the magnitude of local‐scale species richness change following land‐use change of all the biomes (Figure S1). Compared with the other biomes, the variance in the magnitude of local‐scale species richness change following land‐use change is also relatively high in the tropical moist forests so the large number of land‐use change studies conducted in this biome is justified (Figure S1). However, given the significant impact that land‐use change has on biodiversity (Newbold et al. [2015](#ece32004-bib-0031){ref-type="ref"}), shifting some of the focus of future land‐use change studies from tropical moist and temperate broadleaf forests to the coniferous forest biomes would fill crucial data gaps in how land‐use change impacts biodiversity around the globe. Invasive species {#ece32004-sec-0010} ---------------- Studies that examine the effect of species invasions on local‐scale species richness are most commonly conducted in temperate broadleaf forests (38% of responses) and the Mediterranean (18% of responses). Invasions by exotic species are most likely in regions of high human activity (Didham et al. [2005](#ece32004-bib-0012){ref-type="ref"}; MacDougall and Turkington [2005](#ece32004-bib-0026){ref-type="ref"}) such as the Mediterranean, temperate forest, and grassland biomes, as opposed to remote areas, such as the tundra and boreal forest (Sala et al. [2000](#ece32004-bib-0035){ref-type="ref"}). The over‐representation of invasion studies in the temperate broadleaf and mixed forest biome is well justified as it has been estimated that temperate mixed forests will experience a high increase in invasive alien species for the year 2100 (Bellard et al. [2013](#ece32004-bib-0003){ref-type="ref"}). Our results also show that the temperate broadleaf forest and Mediterranean biomes have high variance in the magnitude of change in local species richness following species invasions so while these biomes are significantly over‐represented further studies in these biomes will be useful in deciphering this variability (Figure S1). We found relatively few species invasion studies conducted in the tropical biomes (8% in tropical moist forests, 5% in tropical grasslands, 1% in tropical dry forests, and none in tropical coniferous forests) compared with the large number of total studies conducted in this biome. Species invasions are not thought to be one of the major threats to biodiversity in the tropical biomes as the high diversity in the tropics minimizes the chance of non‐native species successfully invading (Sala et al. [2000](#ece32004-bib-0035){ref-type="ref"}). Also, as temperatures increase fewer species will be able to successfully inhabit the tropics. Bellard et al. ([2013](#ece32004-bib-0003){ref-type="ref"}) predict that the shift of tropical forests to more extreme climates compared with higher latitude biomes will lead to the tropics becoming less suitable for invasive species in the future. Therefore, the distribution of invasion studies throughout the tropical biomes is appropriate to the level of threat that invasions will have in these areas in the future. Less than 3% of the invasion studies in our dataset are conducted in the tundra and boreal forest biomes. These biomes also have the lowest variation in the magnitude of change in local species richness following invasions compared with the other biomes. Therefore, the few studies conducted in the tundra and boreal forest all show very similar change in species richness from species invasions, suggesting that the low number of studies in these biomes is justified. However, increasing temperatures are resulting in dramatic northward species range shifts (Chen et al. [2011](#ece32004-bib-0010){ref-type="ref"}) and warmer temperatures and longer growing seasons are enabling the survival of more non‐native species to high‐latitude ecosystems (Carlson and Shephard [2007](#ece32004-bib-0008){ref-type="ref"}; Spellman et al. [2014](#ece32004-bib-0038){ref-type="ref"}). The ecological effects of the increasing number of exotic species inhabiting the boreal forest are poorly understood (Sanderson et al. [2012](#ece32004-bib-0036){ref-type="ref"}). The boreal forest plays a key role in various processes that are crucial for global ecosystem functioning, such as biogeochemical cycling (Volney and Fleming [2000](#ece32004-bib-0049){ref-type="ref"}), and given the detrimental effects of invasive species on native biodiversity and ecosystem functioning (Vitousek et al. [1996](#ece32004-bib-0047){ref-type="ref"}; Murphy and Romanuk [2014](#ece32004-bib-0030){ref-type="ref"}), additional studies of the effects of species invasions on boreal forest biodiversity would be useful. Nutrient addition {#ece32004-sec-0011} ----------------- The addition of excess nutrients to ecosystems is expected to be most prevalent in regions of high industrialization and intense agricultural activity (Galloway et al. [2004](#ece32004-bib-0018){ref-type="ref"}). Studies that examine the effects of nutrient addition on local‐scale species richness are most common in temperate broadleaf and mixed forests (34% of responses). Furthermore, while studies of the three other human drivers of change are lacking in the temperate coniferous forest, there are a relatively high number of nutrient addition studies conducted in this biome (13% of responses). The temperate regions of the Earth have large population density and increasing agricultural activity (Sala et al. [2000](#ece32004-bib-0035){ref-type="ref"}). Nutrient addition is expected to be a major driver of future biodiversity change in the temperate forest biomes; therefore, the over‐representation of studies in the temperate forests is not surprising. The variance in the magnitude of change in local species richness following nutrient addition is also relatively high in both temperate forest biomes, which further justifies the over‐representation. Nutrient addition studies are disproportionately conducted in temperate forest and tundra biomes and understudied in the boreal forest biome (2% of responses). Nutrient addition is not expected to have as large an impact on biodiversity in the tropical forest and desert biomes as plant growth in these regions is primarily limited by water availability rather than nitrogen (Vitousek [1984](#ece32004-bib-0046){ref-type="ref"}), yet we found more nutrient addition studies conducted in these biomes (7% and 6% of responses, respectively) than in the boreal forest. Like the tundra biome, nutrient addition in the boreal forest is not generally acknowledged as an immediate threat due to its distance from sources of pollution (Sala et al. [2000](#ece32004-bib-0035){ref-type="ref"}). However, atmospheric nitrogen deposition from agricultural and industrial activity is currently increasing and has the potential to be transported long ranges, resulting in impacts on remote ecosystems (Bergstrom and Jansson [2006](#ece32004-bib-0005){ref-type="ref"}; Holtgrieve et al. [2011](#ece32004-bib-0022){ref-type="ref"}). The boreal forest is an example of a relatively pristine ecosystem that could suffer immense consequences from increased atmospheric nitrogen deposition, as boreal forest plant species are nitrogen limited. Furthermore, the high variance in the magnitude of change in species richness following nutrient addition in the boreal forest suggests that more studies are necessary to determine how boreal forest biodiversity responds to nutrient addition. Warming {#ece32004-sec-0012} ------- Warming studies are most common in temperate broadleaf and mixed forests (39% of responses) and are better represented in the tundra biome compared with the other four drivers of change (23% of responses). Global circulation models predict larger increases in temperature at higher latitudes (Kattenberg et al. [1996](#ece32004-bib-0023){ref-type="ref"}); thus it is not surprising that the majority of studies examining the effects of increasing temperatures on species richness are conducted in these high‐latitude biomes. Interestingly, the magnitude of change in species richness is highly variable in the tundra biome and is much less variable in the temperate broadleaf and mixed forests (Figure S1). This suggests that future studies should focus more on the impacts of warming on tundra biodiversity than on temperate broadleaf forest biodiversity, which is the opposite to what is currently occurring. Only 5% of warming studies are conducted in the boreal forest yet, given its high latitude, this biome is expected to experience significant increases in temperature compared with other biomes. None of the warming studies included in our dataset are conducted in tropical grassland, tropical forests, and desert biomes. While these lower latitude biomes are not expected to experience as large an increase in temperature compared with higher latitude biomes, such as the boreal forest and tundra (Sala et al. [2000](#ece32004-bib-0035){ref-type="ref"}), the ecological communities present in these biomes are expected to be less robust than species at higher latitudes to the effects of warming (Sunday et al. [2011](#ece32004-bib-0039){ref-type="ref"}). It has been suggested that the lower thermal tolerance of tropical species will result in increasing temperatures to have more detrimental effects on tropical communities (Tewksbury et al. [2008](#ece32004-bib-0041){ref-type="ref"}; Tuck and Romanuk [2012](#ece32004-bib-0043){ref-type="ref"}). The difference in the effects of climate change along latitudinal gradients highlights the need to conduct studies and collect data on the effects of warming on species richness across all of the Earth\'s biomes, not just those where temperatures are expected to increase the most. Conclusions {#ece32004-sec-0013} =========== In this study, we identify terrestrial biomes where studies that examine the effects of certain human drivers on local species richness are severely lacking based on the level of threat that the drivers pose to the biomes. The biome--driver combinations that are most critical in terms of where future biodiversity change studies need to be conducted are as follows: land‐use change studies in temperate and tropical coniferous forests, species invasion, and nutrient addition studies in the boreal forest, and warming studies in the boreal forest and tropical biomes. An important caveat to our identification of data gaps is that we only consider studies that examine the effects of human drivers on species richness in control vs. disturbed treatments. There are a variety of other biodiversity change studies that were not included in this analysis. We chose to focus only on species richness, as it is the most common biodiversity measure used in studies examining the effects of human drivers of ecosystem change. However, we expect roughly similar data gaps and patterns across biomes for studies that use biodiversity metrics other than species richness. Conflict of Interest {#ece32004-sec-0015} ==================== None declared. Supporting information ====================== ###### **Figure S1**. Variance (given as standard deviation) in the magnitude of change in species richness following the four human drivers of change. ###### Click here for additional data file. ######   ###### Click here for additional data file. We thank all the authors of the studies included in the dataset for collecting and publishing their data. Thank you to two anonymous reviewers whose suggestions greatly improved this manuscript. G.M. and T.R. were supported by the Natural Sciences and Engineering Research Council, Canada.

I[NTRODUCTION]{.smallcaps} {#sec1-1} ========================== Diabetes mellitus (DM) may be caused by insulin deficiency, insulin resistance, or by a combination of both. Insulin deficiency can be caused by pancreatectomy, pancreatitis, alcoholic chronic pancreatitis, hemochromatosis, cystic fibrosis, mitochondrial DNA mutations, or by drugs/toxins. Insulin deficiency may lead to type 1 diabetes mellitus (T1DM) which may be autoimmune or idiopathic in nature and is present in 9% cases of insulin deficiency. Insulin resistance may also be caused by leprechaunism, autoimmune diseases, lipoatrophy, or endocrinopathies including glucagonoma, pheochromocytoma, acromegaly, Cushing\'s syndrome, and thyroid disease. The combination of both (insulin deficiency and insulin resistance) may lead to gestational diabetes or to T2DM. T2DM may be caused by genetic and environmental factors (diet and exercise) and is present in 85% patients with insulin deficiency and insulin resistance. T1DM is primarily caused by genetic factors, environmental factors, and disorder of the immune regulatory mechanism. A combination of all these three factors causes autoimmune disease. Type 1 diabetes mellitus, in contradiction to T2DM, has an acute presentation. Prediabetes is the phase before the onset of T1DM. This period provides a window of opportunity for early intervention. All available interventions including steroids, immunosuppressants, and cyclosporins can be possibly applied during the prediabetes phase. Genetic predisposition exposes the beta cell mass to putative environmental factors leading to insulitis, beta cell injury and a decline of beta cell mass. This phase subsequently leads to the prediabetes stage which is mediated by cellular T-cell immunity including the humoral autoantibodies (islet cell antibodies, insulin autoantibodies, anti-glutamic acid decarboxylase~65~, IA~2~Ab etc.), and glucose intolerance. T1DM is caused by the destruction of pancreatic beta cells, and thus insulin is no longer produced in these cells, leading to hyperglycemia, ketoacidosis and potentially death if not treated with insulin. Dr. Eugene Opie was the first person to have described the concept of insulitis in 1901. The treatment goals for T1DM are simple and include maintaining near normal blood glucose levels and avoiding long-term complications. Diabetes management is a constant juggle between insulin and maintaining an appropriate lifestyle. Environmental risk factors for T1DM include seasonality at diagnosis at migration to other countries. The risk factors of T1DM recognised in the case-control studies include infant/childhood diet, viruses (exposures as early as in utero), hormones, stress, hygiene, and Vitamin D deficiency. Type 1 diabetes mellitus is also on the increase like T2DM, even though not in the same proportion, but still with a 3--5% increase/year. T1DM is on a constant rise in Finland, Sweden, Colarado, India, and Germany. There has been an increase from an incidence of almost 10/100,000 children to around 60/100,000 children in Finland in the last 50 years. Around 78,000 children under 15 years are estimated to develop T1DM annually worldwide. Of the existing 490,000 children living with this disorder, 24% are in the European region and 23% in the South-East Asian region.\[[@ref1]\] India accounts for most of the children with T1DM in South-East Asia. According to the 6^th^ edition of the International Diabetes Federation diabetes atlas, India has 3 new cases of T1DM/100,000 children of 0--14 years.\[[@ref1]\] The prevalence of diabetes in India is variable, and three sets of data show 17.93 cases/100,000 children in Karnataka, 3.2 cases/100,000 children in Chennai, and 10.2 cases/100,000 children in Karnal (Haryana).\[[@ref2][@ref3][@ref4]\] The bottom line remains that T1DM is quite prevalent and common. A number of trials worldwide are being presently conducted on T1DM. In addition to intervention trials, there are three ongoing studies, in the US, on the natural history of T1DM. These include Diabetes Autoimmunity in the Young Study in Colorado,\[[@ref5]\] Prospective Assessment in New-born of Diabetes Autoimmunity in Florida,\[[@ref6]\] and the Diabetes Evaluation in Washington.\[[@ref7]\] All these are based on newborn genetic screening in the general population, and, therefore, concerns have been raised about proper informed consent. Similar concerns have been raised for the Diabetes Prediction and Prevention study in Finland, which is also screening the general population. Parents of babies who carry high-risk DQB1 alleles receive a letter informing them that their infant is at "high" or "moderate" risk of developing the disease. The likelihood that these children will develop T1DM before they are 35 years of age is only about 6%. Moreover, approximately half of the children who will develop T1DM come from the "low" risk group, which is not eligible to participate in these studies. The Indian Council of Medical Research (ICMR) funded Registry of People with diabetes in India with young age at onset youth dependency ratio was started in the year 2006 with 8 collaborating centres across India which subsequently increased to 10 centres. This registry is by the Government of India and called the ICMR registry. This registry is not just focusing at T1DM; but giving an overview of diabetes in the young. The registry has collected data on age of onset of diabetes, presentation, treatment patterns, insulin regimens prescribed, complications (neuropathy, retinopathy, nephropathy), infection, and modes of mortality. Results of the registry will soon be published. From various epidemiological studies, it has been found that T1DM is encountered majorly in the under-privileged children. The Changing Diabetes in Children program has its reason of existence because of this affliction of people from low socioeconomic strata, along with the higher socioeconomic strata, but the former is very important for this program. There is also evidence that T1DM is, in part, a genetic disorder. Identical twins (i.e., monozygous twins) are more likely to both have T1DM than nonidentical twins (i.e., dizygous twins). But concordance rates in identical twins are \<50%, supporting the hypothesis that environmental factors are also important in the development of T1DM. The T1DM risk is increased 15-fold for 1^st^ degree relatives, and the risk is \~ 6% till 30 years of age. The risk increases in the presence of susceptibility genes. DQB1 is the best single genetic marker for T1DM, and it is the gene most often used to identify individuals with a high risk of developing disease. However, risk estimates based on DQB1 alone are less precise than those based on the combination of alleles at both the DQA1 and DQB1 loci. These combinations are called haplotypes, and reflect the specific alleles on each of the two chromosomes. Not all haplotypes (i.e., chromosomes) with DQB1 \*0201 or \*0302 have high-risk DQA1 alleles. Thus, DQB1 and DQA1 typing provide more accurate risk estimates than those based on DQB1 alone. However, it is also more expensive. For that reason, most screening is based only on DQB1. The two DQA1-DQB1 haplotypes that are most strongly associated with T1DM are DQA1 \*0501-DQB1 \*0201 and DQA1 \*0301-DQB1 \*0302. That is, chromosomes with DQB1 \*0201 and DQA1 \*0501 confer a higher risk for T1DM than chromosomes with DQB1 \*0201 with some other DQA1 allele (not \* 0501). Similarly, chromosomes with DQB1 \*0302 and DQA1\*0301 confer a higher risk for T1DM than chromosomes with DQB1 \*0302 and some other DQA1 allele (not \* 0301). The relative increase in T1DM risk by the number of high-risk haplotypes varies with a difference in ethnicity. The relative risk in African-Americans with two high-risk DQA1-DQB1 haplotypes is 45% while it is 16% in Caucasians and only 11% in Asians. The absolute risk (or actual likelihood) of developing T1DM in Caucasians and African Americans with two high-risk haplotypes is about 3% before 30 years of age, depending on the population. The risk for Asians with two high-risk haplotypes is much lower (\<1%). Therefore, even if a person carries high-risk haplotypes, their chances of developing T1DM are quite low irrespective of their ethnicity. In families where there is already one person with T1DM, the risk to unaffected relatives is much higher than that for the general population. For example, in Caucasian families, siblings of an individual with T1DM are about 15 times more likely to develop the disease then a person from the general population (i.e., without a family history of the disease). Since siblings share half their genes with each other, researchers often look at whether the human leukocyte antigen (HLA) haplotypes carried by an unaffected sibling are the same as those carried by the affected individual. Unaffected siblings can share two, one or no HLA haplotypes with their affected brother or sister. As can be seen, the T1DM risk for a sibling who has the same two HLA haplotypes as their T1DM sibling is quite high (about 25%). If they share one or no HLA haplotypes with their affected sibling, their risk is much lower, about 8% and 1%, respectively. The fact that the risk for individuals who have no shared HLA haplotype is still increased compared to that for the general population suggests that genes other than those in the HLA region must also increase susceptibility for the disease. Future diabetes studies should evaluate the targets for young children, the various diabetes regimens, risk of hypoglycemia and techniques to avoid it, developmental issues and education. T1DM is the actual diabetes mellitus and is the virgin diabetes while other diabetes types are from mixed etiology. In a recent South-East Asian meeting, it was concluded that insulin is a birthright for children, and it is the duty of the government, organisation, and citizen to make it available for all T1DM patients. **Source of Support:** Nil **Conflict of Interest:** None declared.

1.. Introduction   {#sec1} ================== Microcrystalline materials, with crystallite sizes in the micron to submicron range, are prevalent in many areas of science, including biology, chemistry, materials science and mineralogy. In each of these fields, structure elucidation is central to the understanding of a particular material's properties. When single crystals large enough for conventional structure analysis are not available, powder diffraction techniques are usually applied, and indeed over the last two decades methods for data collection and structure determination have become increasingly powerful, allowing quite complex structures to be solved (David & Shankland, 2008[@bb11]). Nonetheless, powder diffraction techniques reach their limits in the case of complex structures with many atoms in the asymmetric unit and disordered materials. With the introduction of dedicated synchrotron radiation sources in the 1980s, the potential for monochromatic single-crystal microdiffraction was recognized and explored immediately (Eisenberger *et al.*, 1984[@bb16]; Rieck *et al.*, 1988[@bb39]; Andrews *et al.*, 1988[@bb2]; Harding, 1996[@bb19]). Nowadays, most synchrotron facilities have at least one beamline dedicated to monochromatic microcrystal diffraction. However, the size of the microcrystals used for most measurements is still larger than 500 µm^3^. This is because the handling and mounting of tiny crystals are delicate operations, the data collection requires that the crystal be rotated within the very small X-ray beam in a controlled manner, and the risk of radiation damage may increase as the crystal size decreases. The white beam, Laue diffraction technique is an attractive alternative to the monochromatic one, because it takes full advantage of the X-ray energy spectrum at a synchrotron and requires no rotation of the crystal. With the white beam provided at a synchrotron source, a reasonably large amount of data can be collected in a single shot. The acquisition time is fast, and only a few patterns recorded on stationary randomly oriented crystallites are necessary to get a full dataset (Cornaby *et al.*, 2010[@bb10]; Dejoie, McCusker, Baerlocher, Kunz & Tamura, 2013[@bb14]). This makes the method attractive for the structural characterization of beam sensitive crystals (Cornaby *et al.*, 2010[@bb10]) and for *in situ* (Perry *et al.*, 2014[@bb38]) or time-resolved studies (Yorke *et al.*, 2014[@bb51]). One of the drawbacks of the method is the complexity of the data analysis, caused by the energy dependence of correction factors, and the overlap of harmonic reflections (Helliwell *et al.*, 1989[@bb21]). Furthermore, even though more data can be collected in a single shot without rotating the crystal, the duration of the exposure to a white synchrotron beam is sometimes still long enough for radiation damage to occur. Another way of collecting microcrystal diffraction data is by using a suitably equipped electron microscope. This allows extremely small crystallites (\~100 nm on an edge) to be examined, because electrons interact with matter much more strongly than do X-rays. Unfortunately, precisely because of this strong interaction, the electron diffraction intensities can be severely distorted by the effects of dynamical and multiple scattering, and this complicates their use in structure analysis. This problem has been reduced with the introduction of the precession electron diffraction technique (Vincent & Midgley, 1994[@bb47]; Midgley & Eggeman, 2015[@bb28]), and three-dimensional electron diffraction datasets can now be collected, using either the method of automated diffraction tomography (Kolb *et al.*, 2007[@bb26]) or that of rotation electron diffraction (Zhang *et al.*, 2010[@bb53]). Over the last few years, an impressive array of structures from nanometer-sized crystals has been investigated successfully (Gorelik *et al.*, 2012[@bb17]; Willhammar *et al.*, 2014[@bb50]). However, these methods are not applicable to all materials, such as beam sensitive materials or materials requiring a controlled atmosphere. The non-kinematical intensities also remain an issue, because the intensities are not reliable enough to allow the structure to be fully refined, although progress has also been made in this area (Palatinus *et al.*, 2015[@bb35]). With the development of X-ray free-electron laser (XFEL) sources, which create ultra-fast X-ray pulses of unprecedented brilliance, a new option for the structural characterization of microcrystalline materials is arising. By exposing a small crystallite (from nm to a few µm) to a single monochromatic pulse of tens of femtoseconds in duration, a diffraction pattern can be obtained before the crystal is damaged. If such single-pulse diffraction patterns, collected sequentially on many randomly oriented crystallites, are combined, it is possible to determine the crystal structure (Chapman *et al.*, 2011[@bb8]; Boutet *et al.*, 2012[@bb5]). In a typical serial femtosecond crystallography (SFC) experiment, hundreds of thousands of individual snapshots are collected from a flowing suspension of nanocrystals in a liquid jet. Several analytical tools and dedicated algorithms have been developed to filter the hundreds of thousands of frames collected, to index individual snapshots and to reconstruct the reflection intensities using a Monte Carlo integration method (White *et al.*, 2012[@bb49], 2013[@bb48]; Hattne *et al.*, 2014[@bb20]; Brewster *et al.*, 2015[@bb6]). The latter is crucial for further use of the data, because only a single slice of the Ewald sphere is accessed in each diffraction pattern, so the reflections are only partially measured. To date most activity in the realm of structure analysis using XFEL radiation has focused on monochromatic beams and macromolecular systems (Barends *et al*., 2014[@bb3]). Our prime interest is in the area of organic and inorganic materials, where the diffraction patterns are much sparser. One way of increasing the number of reflections in diffracting condition is to use a larger energy bandpass, as in a Laue diffraction experiment. This capability will be available to a limited extent at the SwissFEL facility (PSI, Switzerland), a new XFEL source scheduled to come online in 2017 (Patterson *et al.*, 2010[@bb36]; Patterson *et al.*, 2014[@bb37]). Indeed, one of the unique features of the SwissFEL beam will be that the bandpass of the X-ray beam can be adjusted to give as much as a 4% energy spread. With the 4%-energy-bandpass mode, not only can more reflections be recorded per shot, but the intensities will also be measured more reliably (Dejoie, McCusker, Baerlocher, Abela *et al.*, 2013[@bb13]; White *et al.*, 2013[@bb48]). The broader range of energies puts more reflections in diffracting condition and simultaneously allows the full width of most reflections to be measured, making integration easier. This provokes the question as to whether or not the SFC approach using a broad bandpass beam could be applied to materials with smaller unit cells (up to *ca* 25000 Å^3^). Computer simulations indicate that with such a setup, a full dataset can be obtained with just a few hundred crystals, instead of tens or hundreds of thousands (Dejoie, McCusker, Baerlocher, Abela *et al.*, 2013[@bb13]). Here we describe the application of this approach to experimental data collected on three inorganic materials at a synchrotron source under conditions mimicking the SwissFEL beam. We have developed new algorithms to index the experimentally measured single-shot patterns of randomly oriented crystals of such materials, and show that the measured intensities are accurate and suitable for structure analysis. To improve the efficiency of the serial snapshot crystallography data collection for materials with small unit cells, we also investigated the possibility of collecting and analyzing data from more than one randomly oriented crystal in a single snapshot. 2.. Mimicking the SwissFEL beam on SNBL at ESRF   {#sec2} ================================================= At present, there is no experimental facility that can produce an intense X-ray beam with a 4% energy bandpass directly, so the single-crystal diffractometer on the Swiss--Norwegian Beamlines (SNBL/BM01A) at the European Synchrotron Radiation Facility (ESRF) was used to mimic the SwissFEL setup experimentally. The broad bandpass mode was achieved by collecting a diffraction pattern while the monochromator was scanned over a 4% energy range (average energy 17.34 keV, or 0.7153 Å). Because the SNBL setup was not suited for snapshot data collection on a series of individual crystals, the 4% bandpass patterns were collected by rotating single crystals mounted on MiTeGen MicroMeshes (Fig. 1[▶](#fig1){ref-type="fig"}). A first series was obtained by taking a 'snapshot' after each 1° rotation around the *ϕ* axis of the goniometer (*i.e.* 360 different frames or crystal orientations). In order to simulate the case of crystals with platelike morphology which would probably display a preferential orientation on a flat surface, a second series of measurements was carried out by rotating around a single face of one of the crystals (*ψ* scan). A conventional monochromatic *ϕ* scan (0.25° steps) was also performed on the same test crystals to obtain reference datasets. A two-dimensional Dectris Pilatus 2M detector was positioned at a distance of 144 mm from the sample for the *ϕ* scans, and at 224 mm for the *ψ* scans. Geometry calibration (sample-detector distance, normal incidence position of the detector, tilt angle of the detector) was carried out with the *XMAS* program (Tamura, 2014[@bb45]), using an LaB~6~ reference powder pattern. Three test samples were used: the zeolite ZSM-5 (Olson *et al.*, 1981[@bb29]), a hydrated caesium cyanoplatinate (Cs~2~\[Pt(CN)~4~\]·H~2~O) (Johnson *et al.*, 1977[@bb22]), and the mineral sanidine (Ackermann *et al.*, 2004[@bb1]). As with all cyanoplatinates, Cs~2~\[Pt(CN)~4~\]·H~2~O is toxic if swallowed. Appropriate laboratory safety procedures apply when handling this family of compounds. These materials have unit cells typical of small-molecule and inorganic structures and cover three different crystal systems (Table 1[▶](#table1){ref-type="table"}). To compensate for the much lower flux of the bending-magnet synchrotron beam (at least four orders of magnitude lower than that expected for SwissFEL), relatively large single crystals (*ca* 15000 µm^3^) were used. In order to test the possibility of indexing multiple patterns, two grids with multiple randomly dispersed crystals were also prepared, one with three ZSM-5 crystals on a grid and one with 15 (Fig. 1[▶](#fig1){ref-type="fig"}). 3.. Indexing single snapshot patterns   {#sec3} ======================================= Indexing the single snapshots involved overcoming four problems: (1) an accurate orientation had to be retrieved from a single frame, (2) the data were collected using a 4% energy bandpass, so the wavelength associated with each diffraction spot is indeterminate, (3) organic/inorganic materials typically have small unit cells, so the number of observations per frame is limited, and (4) the orientations of multiple crystals should be determined from a single frame. It is normally possible to obtain unit-cell parameters from a powder diffraction pattern, so this information was used as *a priori* knowledge to index the 4% bandpass data. Two ways of dealing with the data were investigated. One approaches the problem from a polychromatic point of view using a modification of an algorithm developed at the Advanced Light Source to index Laue microdiffraction data (Tamura, 2014[@bb45]), and the other from a more conventional monochromatic point of view, assuming an average wavelength (Fig. 2[▶](#fig2){ref-type="fig"}). 3.1.. Laue approach   {#sec3.1} --------------------- Our first intuition was to use the tools developed by the Laue microdiffraction community, where finding orientations of crystals with known cell parameters from a single frame is common practice (Chung & Ice, 1999[@bb9]; Tamura, 2014[@bb45]). Typically, the indexing procedure consists of matching the experimental pattern with the corresponding portion of the reciprocal lattice calculated from the unit-cell parameters. Since the exact energy associated with each reflection is not known, only the directions of the reciprocal lattice vectors can be used. The angles between these vectors, which are unaffected by the energy of the incident radiation, are then compared with a calculated list of angles generated from the known unit cell. This approach, implemented in the *XMAS* software, was developed for indexing Laue patterns (5--24 keV range) of small-unit-cell samples of known structure (*i.e.* reflection intensities can be calculated). These assumptions made the indexing of 4% bandpass data of samples of unknown structure slow and highly unreliable. To overcome the problems, some new features were introduced into the conventional angle computation method. Instead of classifying the reciprocal lattice points generated from the unit-cell parameters according to their associated structure factors, the points are now sorted by decreasing *d* spacing, so no *a priori* knowledge of the structure is required. The second major modification is the introduction of a three-dimensional pattern-matching algorithm to recover the orientation matrix. The indexing process is started by selecting a set of measured reflections with the highest intensities and a set of calculated reciprocal lattice points with the largest *d* spacings. The number depends on the sample, the quality of the data, and the number of crystals measured. The selected reflections are converted into normalized reciprocal space vectors **q** ~norm~ (*i.e.* the magnitudes are set to 1) and the reciprocal space points to vectors **q**. A three-dimensional pattern-matching algorithm, adapted from the one described by Van Wamelen *et al.* (2004[@bb46]) in two-dimensional space, is then applied. First, a pair of **q** ~norm~ vectors defining an angle is selected from the measured data (starting from the strongest reflections), and the observed vectors in the immediate vicinities of these two are identified. The 'distance criteria' defining the local environment of the selected pair of **q** ~norm~ vectors has to be chosen carefully to keep the computation time low without decreasing the indexing success rate. This 'first neighborhood' is then compared with calculated vector pair environments to find local matches. These local matches are classified according to their goodness-of-fit, and then the global match including all vectors is examined. The algorithm stops when a sufficiently good global match (number of indexed peaks) is found. The orientation matrix obtained is refined, a check for reflections not indexed in the first step is performed, and then the orientation matrix is refined again. If the parameters are properly defined (in particular the number of reflections used and the distance criteria), this nearest-neighbors approach is considerably more efficient and faster than the conventional alignment-type approach used in most Laue indexing algorithms. In the case of multiple crystals per frame, a sequential approach is used. As soon as the first set is identified, the indexed peaks are removed from the original list of measured reflections, and the indexing process is repeated. This Laue-based indexing algorithm is implemented in Fortran 90, as part of the *XMAS* software package. 3.2.. Monochromatic approach   {#sec3.2} ------------------------------ The 4% bandpass data are quite close to monochromatic, and can be interpreted as such by assuming an average wavelength for each reflection. We examined several methods used to index conventional monochromatic data, some of which are also being used through the *CrystFEL* or the *cctbx.xfel* programs to index single frames collected on macromolecular materials in XFEL experiments (White *et al.*, 2012[@bb49]; Hattne *et al.*, 2014[@bb20]). These indexing methods typically convert a set of positions of recorded reflections into reciprocal-space vectors and analyze them for periodicity in order to determine the basis vectors and to find the metric and orientation of the unit cell. In *XDS* (Kabsch, 1988[@bb23], 1993[@bb24], 2010[@bb25]), differences between reciprocal-lattice vectors are calculated and accumulated in a histogram. The resulting clusters allow a unique basis to be determined. In the case of 4% bandpass data, this method was quickly abandoned, because it yielded streaks of reflections, rather than clusters -- an effect of averaging the wavelength. Another approach was developed by Duisenberg (1992[@bb15]) in the *DIRAX* program, where row periodicity is sought by projecting all observed reciprocal lattice vectors onto the normal to the plane given by three randomly selected points. Later methods implemented in *MOSFLM* (Steller *et al.*, 1997[@bb44]; Leslie, 2006[@bb27]), *DENZO* (Otwinoswki & Minor, 1996[@bb31]; Otwinoswki *et al.*, 2012[@bb32]) and *LABELIT* (Sauter *et al.*, 2004[@bb41]) followed a similar approach, but probed all possible directions and used a fast Fourier transform (FFT) to look for periodicity. These methods gave some useful results for 4% bandpass data for frames obtained from a single-crystal, but the majority of the frames simply did not contain enough reflections to determine at least two independent base vectors, and there was no guarantee that the method would scale well to multiple crystals. It became clear that these monochromatic indexing methods were ill-suited for indexing single snapshots of organic/inorganic materials taken with broad bandpass radiation. Therefore, a strategy to deal specifically with such data was devised. The indexing procedure was split into two parts. The first part entails a search for suitable orientation matrices, and the second the evaluation of each orientation matrix. For the first part, two different approaches were developed (designated *mono1* and *mono2* for convenience). The *mono1* algorithm resembles a classic approach reported by Busing & Levy (1967[@bb7]) and Sparks (1976[@bb43]) some 40 years ago, in which the crystal orientation is determined by assigning indices to two low-resolution non-collinear reflections manually. In our case, the initial indices are not known, but the lattice spacings for all reflections are. Therefore, each reflection can be assigned a set of pre-calculated indices (those with similar *d* spacings). A semi-exhaustive search for the orientation matrix is then performed by trying all combinations of non-collinear spots and their corresponding indices. The angle between the **q** vectors is independent of the incident radiation, so only pairs of reflections where the angle of the observed reflections closely matches those of their tentative indices are kept. For each pair, the orientation of the crystal can then be determined. Although the algorithm is fast and reliable for one crystal per frame, it was found to scale poorly with increasing number of diffracting crystals. Therefore, a second routine, *mono2*, was developed. This is a 'brute-force' approach that simply samples a fine grid of all possible crystal orientations. Steller *et al.* (1997[@bb44]) described an algorithm for generating all possible directions on a hemisphere. With this algorithm and a spacing of approximately 1.7° (0.03 radian) between vectors, a sufficiently dense coverage is achieved, and by adding a stepwise rotation (again *ca* 1.7° increments) of 360° around each of these vectors, roughly 1.5 million equally distributed rotation matrices that cover all possible crystal orientations can be generated efficiently. Both the *mono1* and *mono2* approaches yield a large number of candidate orientation matrices, and these are evaluated in the second part of the routine. First, each candidate orientation matrix is used to transform the **q** vectors of the measured reflections to indices, and this will normally yield non-integer values. A reflection is considered indexed if all three indices *hkl* are close to integer values. *nfit* is the number of indexed reflections in the pattern, and *score* gives the residual sum of the squared differences of the distances between the observed and calculated indices. To keep computation times low, each orientation matrix with *nfit* below a certain threshold value *nmin1* is discarded. For each remaining solution, the rotation matrix is optimized against *score*/*nfit* ^2^ using a least-squares routine. This refinement step significantly improves the number of indexed peaks, despite the fact that the average (monochromatic) wavelength is still used. At this point, the wavelength for each reflection is recovered by minimizing the distance between the indices of a reflection and its nearest integer equivalent. The *score* and *nfit* values for each optimized orientation matrix are recalculated, and again, orientation matrices with *nfit* below a second threshold value *nmin2* (typically higher than *nmin1*) are discarded. Equivalent solutions are merged according to the Laue symmetry of the space group. All remaining independent solutions are then ranked by *score*/*nfit* ^2^. For one crystal in the beam, simply the highest ranked orientation matrix is taken. To handle measurements of multiple crystals, a greedy weighted set-optimization algorithm (Young, 2008[@bb52]) was implemented to identify the smallest subset of orientation matrices that indexes the largest number of reflections. This last algorithm had to be modified to take into account that not all crystal orientations are necessarily found, so a number of reflections may be left unindexed. The main difficulty in the monochromatic approach is to define meaningful values for the thresholds *nmin1* and *nmin2*, especially when dealing with a large number of frames with a fluctuating number of crystals, incident flux, and scattering power of the crystals. In order to estimate an appropriate starting value, the brute force algorithm can be run with large angle increments (17°, 0.3 rad) to probe a representative distribution of approximately 1500 orientations. The *mean* and standard deviation (*sd*) of all corresponding values of *nfit* are calculated. In this way, the approximate number of reflections that can be indexed with incorrect orientations can be estimated for a particular frame (*i.e.* its 'noise' level). A meaningful formula for *nmin1* was found empirically to be *mean* + 3 × *sd*. By applying this formula, the indexing algorithm can adjust to arbitrary variations in the data automatically, and this improves the effectiveness of the routine considerably. The *mono1* and *mono2* algorithms are implemented in Python2.7, using the pandas, numpy and scipy libraries. Crystallographic computations are performed using the *cctbx* software (Grosse-Kunstleve *et al.*, 2002[@bb18]). 4.. Application to test data   {#sec4} ============================== 4.1.. Indexing results   {#sec4.1} ------------------------ The three algorithms were evaluated using data collected on three test samples. Indexing results are reported in Table 2[▶](#table2){ref-type="table"}. For ZSM-5, with the largest unit cell, all algorithms work well, but the monochromatic approaches are faster. The *a* and *b* axes of ZSM-5 differ by only 0.5%, but all three algorithms could assign the axes correctly in almost all cases, despite the fact that there is a 4% range of energies. The unit cell of Cs~2~\[Pt(CN)~4~\]·H~2~O is much smaller, so the number of reflections per frame decreased. This made very little difference to the *laue* approach, except to make it faster, but did affect the *mono1* algorithm, where only 205 of the 360 frames could be indexed. Of these, however, 99% were correct. For sanidine, with the smallest unit cell, only 29% of the frames could be indexed using the *mono1* method, but all of these were indexed correctly. The *mono2* algorithm, however, had no problem dealing with the smaller unit cells, and correctly indexed all frames for both materials. With the *laue* approach, 79% of the sanidine frames could be indexed correctly in a short period of time, but the increasing number of false positive results may be an indication that the indexing procedure approaches its limits with such small unit cells. The three algorithms were also applied to multi-crystal data (Table 2[▶](#table2){ref-type="table"}, Fig. 1[▶](#fig1){ref-type="fig"}) collected on the ZSM-5 sample. Both monochromatic approaches could index the multicrystal frames successfully, but the *mono2* algorithm worked considerably better and was easier to set up. With the *mono2* algorithm, it was possible to retrieve a large number of crystal orientations while maintaining reasonable computation times with a desktop computer. With three crystals on the grid, on average 2.8 crystal orientations per frame could be recovered. With 15 crystals, 8.5 crystal orientations per frame could be retrieved for 12 independent crystals. Although 15 crystals can be seen on the grid (Fig. 1[▶](#fig1){ref-type="fig"} *a*), 3 of them do not diffract strongly enough. It was also possible to use the *laue* approach for multiple crystals, but the computation time became a limiting factor, so the tests with 15 crystals were not pursued. As before, the main source of error was the similarity between the *a* and *b* parameters. It was noted, however, that as the number of crystals increased, so did the potential for incorrect assignment of indices, because the chances of reflections from different crystals accidentally overlapping increased. We estimate that with the present setup between 5 and 10 crystals per shot would be optimal, but by using larger detectors with higher resolution, this number will certainly increase. It is perhaps important to note that fine tuning is possible in all algorithms, but we were interested in having a fully automated robust procedure that would be able to process a large amount of data without intervention. The main point is that reflections on most of the single snapshot images can indeed be indexed, despite the differences in unit cell, symmetry and number of crystals, and this means that a useful dataset can be extracted reliably for different materials. We also tested the monochromatic based algorithms on pure monochromatic data (*ϕ* scan) collected on all 3 samples. Although the algorithms were optimized to deal with a 4% bandpass, they still performed satisfactorily in indexing these data. We therefore reason that single frames collected with any bandpass between 0 and 5% can be indexed successfully. For energy bandpass higher than 5%, we expect a Laue approach to be better. 4.2.. Limitations   {#sec4.2} ------------------- As mentioned earlier, increasing the number of crystals diffracting simultaneously increases the probability of accidental overlap of reflections from different crystals, and this will affect both the indexing process and the intensity extraction. In particular, in the case of the monochromatic indexing approach, the least-squares refinement of the rotation matrix may be affected if alien peaks are close to the reciprocal lattice points of the crystal being refined. This effect is exacerbated with plate-like crystals dispersed on a grid, where most will lie on the same face and then only be randomly oriented around the normal of the face. This is the case for the two multi-crystal ZSM-5 preparations (Fig. 1[▶](#fig1){ref-type="fig"}), for which 2 of the 3 crystals and at least 9 of the 15 crystals share a similar orientation of the *b* axis. To investigate the effect of flat or needle-shaped crystals dispersed on a grid, 4% bandpass data were collected on a single-crystal of ZSM-5 in a *ψ*-scan mode (rotation around one face of the crystal), and results were compared with data collected with a more conventional *ϕ* scan. The geometry of the experimental setup (*ψ*-scan mode) only allowed a 90° rotation. Nonetheless, the effect on the completeness of the data (number of unique reflections measured *versus* total number expected), calculated in the 1.0--5.0 Å resolution range (where it is maximal), was not as dramatic as might be expected. The values of 54% and 60% were obtained for the *ψ*-scan and *ϕ*-scan datasets, respectively. For a real SFC experiment, of course, a sample mount that reduces preferred orientation of the crystals should be used. Because a broad bandpass beam was used, specific problems inherent to Laue diffraction have to be checked. One of them is the possible presence of harmonics (*e.g.* the reflections 100, 200, and 300 will have identical diffraction angles for wavelengths *λ, λ*/2, and *λ*/3, respectively). The formula gives the relationship between the minimum distance between two reciprocal lattice points Δ*d\**, the diffraction angle *θ*, and the wavelengths of the two delimiting Ewald spheres *λ* ~min~ and *λ* ~max~. The maximum diffraction angle *θ* that can be used to suppress all harmonics can then be calculated for any given unit cell and wavelength range. For ZSM-5, the minimum distance between two reciprocal lattice points is given by \|*a*\*\|, where *a* = 20.0022 Å (Table 1[▶](#table1){ref-type="table"}). This gives a maximum 2*θ* angle of 46.2° for *λ* ~average~ = 0.7153 Å (*d* spacing = 0.9116 Å). With the current setup (sample--detector distance 144 mm), 93% of all reflections fall below this threshold, and only the *h*00 and 0*k*0 reflection classes are affected at the higher angles. When SwissFEL comes online, the minimum wavelength will be close to 1.000 Å, so the probability of recording harmonic reflections will be reduced even further. 4.3.. Structure solution and refinement   {#sec4.3} ----------------------------------------- To test the data analysis procedures beyond the indexing step and to evaluate the quality of the measured data, reflection intensities for the indexed patterns were integrated using the seed-skewness method (Bolotovsky *et al.*, 1995[@bb4]). The patterns indexed with the *laue* approach were used for this purpose. Reflection intensities were corrected for various factors using techniques similar to those used for Laue microdiffraction data (Dejoie *et al.*, 2011[@bb12]). In particular, each frame was corrected for the decay of the intensity of the incident flux with time, and for the variation of the incident flux with energy. All frames with less than 20 indexed reflections for ZSM-5 and 6 indexed reflections for sanidine and Cs~2~\[Pt(CN)~4~\]·H~2~O were discarded. Six incorrectly indexed patterns remained in the final dataset for ZSM-5, as a result of the *a* and *b* axes being reversed. On the other hand, all frames for sanidine and Cs~2~\[Pt(CN)~4~\]·H~2~O in the final dataset were correctly indexed. As the volume of the crystal was constant in this experiment, no scaling between the individual frames was applied. The situation will be different at SwissFEL, where each crystal can have a different volume and where the variation of the incident flux from shot to shot will be significant. These problems will be addressed in a future publication. In order to identify the fully measured reflections in each dataset, the wavelength associated with each reflection (determined during the indexing process) was examined. Those lying close to the wavelength boundaries are most likely to be partially measured (Dejoie, McCusker, Baerlocher, Abela *et al.*, 2013[@bb13]). A simple schematic drawing is shown in Fig. 3[▶](#fig3){ref-type="fig"}, where the width of the white border between the shaded region and the delimiting Ewald spheres corresponds to the radius of a reflection. If the center of a reflection is within the shaded area it will be fully measured. Of course, not only do reflection widths vary from sample to sample, they can also vary with *hkl* and they may be asymmetric, so the simple picture given in Fig. 3[▶](#fig3){ref-type="fig"} may need to be modified for real data. Note that reflections close to the origin of the reciprocal lattice will never be fully measured in a single shot. On the other hand, an increasing number of reflections are fully measured as the diffraction angle increases. Datasets with increasingly larger borders were constructed. The internal *R* value (*R* ~int~) and completeness provided by the *cctbx* software (Grosse-Kunstleve *et al.*, 2002[@bb18]) were used to monitor the datasets up to a resolution of 1.0 Å (Fig. 4[▶](#fig4){ref-type="fig"}). For all samples, a clear decrease in *R* ~int~ is observed. An *R* ~int~ of \~10% (or less) is achieved for all three samples with a border of 2×10^−3^ Å^−1^. This level is directly related to the elimination of partially measured reflections near the two boundaries of the energy range and close to the origin of the reciprocal lattice. The corresponding completeness values are still more than 80% for these reduced datasets (Fig. 4[▶](#fig4){ref-type="fig"} *b*). On the other hand, the resolution range gets smaller, because the lower resolution reflections are only partially measured and are therefore removed. As a result, the lower resolution limits for sanidine, Cs~2~\[Pt(CN)~4~\]·H~2~O, and ZSM-5 decrease from 6.648 to 2.234 Å, 7.777 to 2.344 Å, and 11.126 to 2.653 Å, respectively. Full datasets (partially measured reflections included) and reduced datasets were used for structure solution using the charge-flipping algorithm (Oszlányi & Sütő, 2004[@bb30]) in *Superflip* (Palatinus & Chapuis, 2007[@bb34]). The structures of the three samples could be solved in all cases, despite the presence of partially measured reflections in the full datasets. The electron density map obtained with the ZSM-5 data is shown in Fig. 5[▶](#fig5){ref-type="fig"} to give a visual impression of the quality of the data. After structure solution, an agreement valuewhere is the corrected integrated intensity, *s* is the scale factor and is calculated from the structural model (obtained in the structure solution step), was calculated for each frame, in order to identify any outlying frames (*i.e.* frames with one or more partially measured reflections). Individual agreement values calculated for the full dataset and for a reduced dataset are shown graphically for all three samples in Fig. 6[▶](#fig6){ref-type="fig"}. For each sample, the optimal reduced dataset was chosen as a compromise between completeness (≥ 80%) and quality (*R* ~int~ ≤ 10%). With the full datasets, the agreement values are highly dispersed (*R* up to 50%), and many outliers (*R* \> 50%) are present. However, with the reduced datasets, most of the agreement values lie below 25%, and only a few outliers (*R* \> 30%) can be seen. This is consistent with the presence of partially measured reflections at the borders of the energy range and close to the origin of the reciprocal lattice in the full datasets. For ZSM-5, the six frames in which *a* and *b* were wrongly assigned in the indexing process are the most obvious outliers. The other two with agreement values slightly above 30% result from the presence of a single partially measured reflection still remaining in each frame. For Cs~2~\[Pt(CN)~4~\]·H~2~O, no outliers can be spotted in the reduced dataset. For sanidine, the number of outliers in the reduced dataset is larger than for the other two samples. This is related to the presence of one or two partially measured reflections in these particular frames. The reason for the partial measurement of these reflections is mainly related to the intrinsic breadths of the sanidine reflections. From the conventional monochromatic datasets, the average widths of the Bragg reflections were extracted for each of the three test samples using the *CrysAlis* software (Oxford Diffraction, 2008[@bb33]) (Table 3[▶](#table3){ref-type="table"}). The widths of the sanidine reflections are clearly larger than those of ZSM-5 and Cs~2~\[Pt(CN)~4~\]·H~2~O, so for sanidine the border at the boundaries needs to be increased even further, if only fully measured reflections are to be included. This can also be seen in Fig. 4[▶](#fig4){ref-type="fig"}, where the *R* ~int~ value reaches a plateau when all reflections on all frames within a dataset are fully measured. This happens at a much larger border size for sanidine than for the other two samples. Structure refinements were then performed with the *SHELX* software (Sheldrick, 2008[@bb42]) using the reduced datasets (2×10^−3^ Å^−1^ border for ZSM-5, and 2.5×10^−3^ Å^−1^ for sanidine and Cs~2~\[Pt(CN)~4~\]·H~2~O). Frames with an *R* value above 0.3 for ZSM-5 and sanidine were discarded. Atomic coordinates, displacement parameters, and an overall scale factor were allowed to vary. In all three structures, anisotropic displacement parameters were used for the cations and isotropic ones for the oxygen atoms. Refinement results are given in Table 4[▶](#table4){ref-type="table"}. Agreement values obtained from a conventional monochromatic refinement are also given for comparison. The *R* factors obtained after refinement with the reduced datasets are still slightly higher than those obtained with the monochromatic data. The main reasons for these differences are probably the presence of a few remaining partially measured reflections, the fact that the correction factors used after the extraction of the intensities still require further improvement, and/or the differences between the intensity extraction procedure used in *CrysAlis* and in our method. 5.. Implications for the SwissFEL experiment   {#sec5} ============================================== In view of these results, an appropriate setup for a SwissFEL experiment for micron- or even submicron-sized crystals can be devised. Unlike proteins, most small-unit-cell materials are stable without a mother liquor, so they can be mounted on a grid similar to those used in an electron microscopy or a Laue microdiffraction experiment (Dejoie, McCusker, Baerlocher, Kunz & Tamura, 2013[@bb14]), or on some other low-background solid support (Ring *et al.*, 2011[@bb40]). For crystals with a needle- or plate-like morphology, which tend to orient preferentially, a support with a roughened surface can be used to obtain more orientations and/or the sample can be tilted with respect to the beam from frame to frame to increase the number of different orientations and thereby access a larger portion of reciprocal space. No special precautions need to be taken to ensure that only one crystal is irradiated per shot. On the contrary, it is actually an advantage to have more than one crystal in the beam, and our indexing algorithms can successfully deal with such data. Thus, simply by spreading the microcrystals on the sample support and translating the support between shots, useable data will be recorded on every frame, saving both beamtime and sample. 6.. Conclusion   {#sec6} ================ The fact that single snapshot diffraction patterns from materials with small unit cells can indeed be indexed opens up new horizons for the structure analysis of polycrystalline materials. It means that high-quality single-crystal X-ray diffraction data can be obtained from tiny crystals in a very short period of time, and this has implications for all areas of materials research, particularly for beam-sensitive samples. Furthermore, because several crystals can be measured simultaneously, agglomerates do not pose a problem and a useful subset of a full dataset can be measured with a single 10--50 femtosecond pulse. The latter offers exciting possibilities in the area of time-resolved experiments. The unique 4%-energy-bandpass mode that is planned for the SwissFEL facility is particularly well suited for this kind of experiment, because it allows more reflections to be measured in a single shot, and they are measured completely. While our algorithms were developed with SwissFEL in mind, they can be applied to any data collected in single snapshot mode with a broad bandpass beam. We expect future studies using this technique to demonstrate just how diverse its application can be. The indexing programs will be made available in the public domain, once they have been properly tested. We thank the SNX Council of SNBL for granting us beamtime to perform the experiment. Funding from the Swiss National Science Foundation and Chevron ETC is gratefully acknowledged. The Advanced Light Source is supported by the Director, Office of Science, Office of Basic Energy Sciences, Materials Sciences Division, of the US Department of Energy under Contract No. DE-AC02-05CH11231 at Lawrence Berkeley National Laboratory and University of California, Berkeley, California. {#fig1} {#fig2} {#fig3} {ref-type="fig"}). Both are calculated up to a resolution of 1 Å.](m-02-00361-fig4){#fig4} {#fig5} ![*R* values for the 360 individual frames for (*a*) sanidine, (*b*) Cs~2~\[Pt(CN)~4~\]·H~2~O, and (c) ZSM-5. In each case, the results for all data are shown on the left and for the reduced dataset (2×10^−3^ Å^−1^ border for ZSM-5, and 2.5×10^−3^ Å^−1^ for sanidine and Cs~2~\[Pt(CN)~4~\]·H~2~O) on the right. For sanidine, the outliers that are removed by increasing the border further are shown as filled squares. For ZSM-5, the incorrectly indexed patterns (*a* and *b* reversed) are indicated with filled triangles.](m-02-00361-fig6){#fig6} ###### Unit cells of the three test samples   Space group *a* () *b* () *c* () () Volume (^3^) ------------------------- ------------- --------- --------- --------- --------- -------------- Sanidine *C*2*/m* 8.5832 13.0076 7.1943 116.023 722 Cs~2~\[Pt(CN)~4~\]H~2~O *P*6~5~ 9.7910   19.5100   1620 ZSM-5 *Pnma* 20.0022 19.8990 13.3830   5327 ###### Results of the indexing tests (360 frames per data collection) Sample Average No. of reflections per crystal Method No. of patterns indexed No. of correct solutions Time[†](#tfn1){ref-type="table-fn"} (per frame) -------------------------------------------------- ---------------------------------------- --------- ------------------------- ------------------------------------ ------------------------------------------------- Sanidine 11 *laue* 332 285[‡](#tfn2){ref-type="table-fn"} 1min     *mono1* 105 105 4 s     *mono2* 359 359 40 s             Cs~2~\[Pt(CN)~4~\]H~2~O 15 *laue* 360 357[‡](#tfn2){ref-type="table-fn"} 1min     *mono1* 205 203 5 s     *mono2* 360 360 12 s             ZSM-5[§](#tfn3){ref-type="table-fn"} 1 crystal 44 *laue* 360 347 110min     *mono1* 353 343 0.5 s     *mono2* 360 351 12 s             ZSM-5[§](#tfn3){ref-type="table-fn"} 3 crystals 23 *laue* 835 791 1030min     *mono1* 964 826 22 s     *mono2* 1004 902 12 s             ZSM-5[§](#tfn3){ref-type="table-fn"} 15 crystals 57 *mono1* 2190 1965 55 s     *mono2* 3027 2854 30 s Indexing trials were performed on a single core of a 2010 Mac Pro equipped with a dual 2.4GHz Quad-Core Intel Xeon processor. Solutions with less than 6 indexed peaks were discarded. Incorrect solutions for ZSM-5 correspond to *a*/*b* flipping. ###### Average reflection profile () extracted from monochromatic data Profiles were extracted using the program *CrysAlis* (Oxford Diffraction, 2008[@bb33]) (e1 and e2 are in the plane of the detector, and e3 is the scanning direction). Direction Sanidine Cs~2~\[Pt(CN)~4~\]H~2~O ZSM-5 ----------- ---------- ------------------------- ------- e1 0.135 0.078 0.095 e2 0.166 0.109 0.118 e3 0.513 0.220 0.314 ###### Refinement of the three structures using reduced datasets *R* values from a refinement using conventional monochromatic data are included at the bottom of the table for comparison.   Sanidine Cs~2~\[Pt(CN)~4~\]H~2~O ZSM-5 ------------------------------------------------------------------- ------------ ------------------------- ------------ No. of reflections 1900 3093 9492 No. of unique reflections 675 1047 2615 Resolution range () 0.7452.234 0.9162.344 0.9072.653 *R* ~int~ [†](#tfn4){ref-type="table-fn"} 0.053 0.097 0.075 *R* ~sigma~ [‡](#tfn5){ref-type="table-fn"} 0.031 0.049 0.040 Completeness (%)[§](#tfn6){ref-type="table-fn"} 80 84 80 No. of parameters 64 65 216 *R*1[¶](#tfn7){ref-type="table-fn"} \[*F* ~obs~ \> 4(*F* ~obs~)\] 0.069 0.037 0.068 *R*1[¶](#tfn7){ref-type="table-fn"} (all) 0.074 0.037 0.157 *With monochromatic data:*       *R* ~int~ [†](#tfn4){ref-type="table-fn"} 0.029 0.053 0.049 *R*1[¶](#tfn7){ref-type="table-fn"} \[*F* ~obs~ \> 4(*F* ~obs~)\] 0.024 0.020 0.053 *R*1[¶](#tfn7){ref-type="table-fn"} (all) 0.026 0.020 0.057 *R* ~int~ = \| (mean)\|/. *R* ~sigma~ = \[()\]/. Calculated for the resolution range 15. *R*1 = \|\|\| \|\|\|/\|\|.

\section{Introduction and background}\label{sectionintro} \color{black} Let ${\mathbb X}$ be an infinite dimensional separable Banach space over the real or complex field $\mathbb{K}$. A sequence ${\mathcal B}=({\bm x}_i)_{i\in\mathbb{N}}$ is \emph{fundamental} if it generates the entire space, that is ${\mathbb X}=\overline{[{\bm x}_i: i\in \mathbb{N}]}$, and it is \emph{minimal} or a \emph{minimal system} if there is a (unique) sequence ${\mathcal B}^*=({\bm x}_i^*)_{i\in\mathbb{N}}$ in the dual space ${\mathbb X}^*$ (which we call the \emph{biorthogonal functionals}) such that ${\bm x}_i^*({\bm x}_j)=\delta_{i,j}$ for all $i,j\in \mathbb{N}$. If in addition ${\mathcal B}^*$ is total, that is if $$ {\mathbb X}^*=\overline{[{\bm x}_i^*: i\in \mathbb{N}]}^{w^*}, $$ the sequence ${\mathcal B}$ is a \emph{Markushevich basis} for ${\mathbb X}$. When there is $\mathbf{C}>0$ such that $$ \|\sum_{j=1}^{n}{\bm x}_i^{*}(x){\bm x}_i\|\le \mathbf{C}\|x\|,\qquad\forall x\in {\mathbb X},\ \forall n\in \mathbb{N}, $$ the sequence ${\mathcal B}$ is a \emph{Schauder basis} and the minimum $\mathbf{C}>0$ for which the above inequality holds is the \emph{basis constant} of ${\mathcal B}$. A \emph{basic sequence} is a Schauder basis for the closure of its span. A Schauder basis ${\mathcal B}$ is $\mathbf{C}$-\emph{unconditional} if $$ \|\sum_{j=1}^{\infty}b_j{\bm x}_j^*(x){\bm x}_j\|\le \mathbf{C}\|x\|,\qquad\forall x\in {\mathbb X},\ \forall (b_j)_{j\in \mathbb{N}}\subset\mathbb{K}\colon\ |b_j|\le 1\ \forall j\in \mathbb{N}, $$ and it is $\mathbf{C}$-suppression unconditional if $$ \|P_A(x)\|\le \mathbf{C}\|x\|,\qquad\forall x\in {\mathbb X},\ \forall A\in \mathbb{N}^{<\infty} $$ where $P_A(x)$ denotes the projection of $x$ on $A$ (with respect to ${\mathcal B}$), that is \begin{align*} P_A(x)=\sum_{j\in A}{\bm x}_j^*(x){\bm x}_j, \end{align*} with the convention that the sum is zero if $A$ is empty. \\ It is well-known that $\mathbf{C}$-unconditionality entails $\mathbf{C}$-suppression unconditionality, whereas $\mathbf{C}$-suppression unconditionality entails $2\mathbf{C}$-unconditionality if $\mathbb{K}=\mathbb{R}$ and $4\mathbf{C}$-unconditionality if $\mathbb{K}=\mathbb{C}$. \\ In this paper, unless otherwise stated by a \emph{basis} ${\mathcal B}$ we mean a Markushevich basis with biorthogonal functionals ${\mathcal B}^*$. We will refer to this sequence as the \emph{dual basis} of ${\mathcal B}$. Except in Section~\ref{sectionseparation}, we will assume that both ${\mathcal B}$ and ${\mathcal B}^*$ are bounded and we set \begin{align} \lambda :=\sup_{i\in \mathbb{N}}\|{\bm x}_i\|, \qquad \lambda ':=\sup_{i\in \mathbb{N}}\|{\bm x}_i^*\|, \qquad\lambda'' :=\sup_{i\in \mathbb{N}}\|{\bm x}_i\|\|{\bm x}_i^*\|, \label{constants} \end{align} a notation that we will use for all such bases. As usual, we use $\supp(x)$ to denote the support of $x\in {\mathbb X}$, that is the set $\{i\in \mathbb{N}: {\bm x}^*_i(x)\ne 0\}$ and we set $\mathbb{N}^{<\infty}:=\{A\subset \mathbb{N}: |A|<\infty\}$.\\ Given a basis ${\mathcal B}$ for ${\mathbb X}$, the \emph{Thresholding Greedy Algorithm} (TGA for short) gives approximations to vectors in ${\mathbb X}$ in terms of greedy sums, or equivalently, in terms of projections on greedy sets. We will consider the more general concept of $t$-greedy sets, which are involved in approximations given by \emph{Weak Thresholding Greedy Algorithms} (WTGA for short). \begin{definition}\label{definiciongreedyset} Let $0<t\le 1$. For each $m\in\mathbb{N}$, a set $A\subset \mathbb{N}$ is an \emph{$m$-$t$-greedy set for} $x\in {\mathbb X}$ if $|A|=m$ and \begin{equation} \min_{j\in A}\left|{\bm x}_j^*\left(x\right)\right|\ge t\max_{j\in \mathbb{N}\setminus A}\left|{\bm x}_j^*\left(x\right)\right| \end{equation} If $t=1$, $A$ is called an \emph{$m$-greedy set for} $x$. By ${\mathcal G}(x,m,t)$ we denote the set of all $m$-$t$-greedy sets for $x$, with ${\mathcal G}(x,0,t)$ consisting only of the empty set, and we let ${\mathcal G}(x,t):=\bigcup_{m\in \mathbb{N}}{\mathcal G}(x,m,t)$. Also, by $\Lambda_m(x)$ we denote the element of ${\mathcal G}(x,m,1)$ with the property that for any $B\in {\mathcal G}(x,m,1)$ such that $B\not=\Lambda_m(x)$, we have $$ \max(\Lambda_m(x)\setminus B)<\min( B\setminus \Lambda_m(x)). $$ \end{definition} The TGA was introduced by Temlyakov in~\cite{T1998} in the context of the trigonometric system, and extended by Konyagin and Temlyakov to general Banach spaces in~\cite{KT1999}, where the authors defined the concepts of \emph{greedy} and \emph{quasi-greedy} Schauder bases. A Schauder basis ${\mathcal B}$ for ${\mathbb X}$ is greedy with constant $\mathbf{C}>0$ (or \emph{$\mathbf{C}$-greedy}) if \begin{equation} \|x-P_{\Lambda_m}(x)\|\le \mathbf{C}\sigma_{m}(x),\qquad\forall x\in {\mathbb X},\ \forall m\in \mathbb{N}, \label{greedydefi1} \end{equation} where $\sigma_m(x)$ is the best $m$-term approximation error (with respect to ${\mathcal B}$) given by \begin{equation} \sigma_m(x)=\sigma_m(x)[{\mathcal B},{\mathbb X}]=\inf_{\substack{y\in {\mathbb X}\\|\supp(y)|\le m}}\|x-y\|.\label{errorapprox} \end{equation} \begin{remark}\label{remarkallthesame} Note that, due to the continuity of the norm, it is equivalent to take $|\supp(y)|=m$ in \eqref{errorapprox}. Also, a standard small perturbation argument gives that \eqref{greedydefi1} is equivalent to \begin{equation} \|x-P_A(x)\|\le \mathbf{C}\sigma_{m}(x),\qquad\forall x\in {\mathbb X},\ \forall m\in \mathbb{N},\ \forall A \in {\mathcal G}(x,m,1) \nonumber \end{equation} and also to \begin{equation} \|x-P_A(x)\|\le \mathbf{C}\sigma_{m}(x),\qquad \forall x\in {\mathbb X},\ \forall m\in \mathbb{N},\ \text{for some}\ A\in {\mathcal G}(x,m,1). \nonumber \end{equation} \end{remark} Since their inception, greedy bases have been widely studied; see for example the book by Temlyakov \cite{T2011}, the more recent articles \cite{AA2016}, \cite{AABW2021}, \cite{AAW2019}, \cite{BB2017} and the references therein. Greedy bases are a subclass of quasi-greedy ones, defined as follows: ${\mathcal B}$ is $\mathbf{C}$-quasi-greedy if \begin{equation} \|P_{\Lambda_m}(x)\|\le \mathbf{C}\|x\|,\qquad\forall x\in {\mathbb X},\ \forall m\in \mathbb{N}.\nonumbe \end{equation} Following \cite{AA2017}, we say that ${\mathcal B}$ is \textit{$\mathbf{C}$-suppression quasi-greedy} if \begin{equation} \|x-P_{\Lambda_m}(x)\|\le \mathbf{C}\|x\|,\qquad\forall x\in {\mathbb X},\ \forall m\in \mathbb{N}.\nonumbe \end{equation} Intermediate structures between these two are almost greedy bases -introduced by Dilworth, Kalton, Kutzarova and Temlyakov in~\cite{DKKT2003}- and semi-greedy bases -~defined by Dilworth, Kalton and Kutzarova in \cite{DKK2003}-. A basis is almost greedy with constant $\mathbf{C}>0$ (or \textit{$\mathbf{C}$-almost greedy}) if \begin{equation} \|x-P_{\Lambda_m}(x)\|\le \mathbf{C}\widetilde{\sigma}_{m}(x),\qquad\forall x\in {\mathbb X},\ \forall m\in \mathbb{N},\nonumbe \end{equation} where $\widetilde{\sigma}_m(x)$ is the best $m$-term approximation error to $x$ via projections (with respect to ${\mathcal B}$), given by \begin{equation} \widetilde{\sigma}_m(x)=\widetilde{\sigma}_m[{\mathcal B},{\mathbb X}](x)=\inf_{\substack{B\subset \mathbb{N}\\|B|=m}}\|x-P_B(x)\|.\label{projectionerror} \end{equation} Notice that, since ${\mathcal B}^*$ is weak star null, in order to compute \eqref{projectionerror}, it is equivalent to take $|B|=m$ or $|B|\le m$. On the other hand, ${\mathcal B}$ is semi-greedy with constant $\mathbf{C}>0$ (or \emph{$\mathbf{C}$-semi-greedy}) if \begin{equation} \inf_{|\supp(y)|\subset \Lambda_m(x)}\|x-y\|\le \mathbf{C}\sigma_{m}(x),\qquad\forall x\in {\mathbb X},\ \forall m\in \mathbb{N}. \nonumbe \end{equation} The algorithm associated with a semi-greedy basis is called a \emph{Chebyshev Greedy Algorithm} (CGA for short). \begin{remark}\label{remarkallequal3} As in the case of greedy bases (Remark~\ref{remarkallthesame}), one may replace $\Lambda_m(x)$ by all $A\in {\mathcal G}(x,m,1)$ or at least one $A\in {\mathcal G}(x,m,1)$ in the definitions of quasi-greedy, almost greedy and semi-greedy bases, obtaining equivalent notions, with the same constant $\mathbf{C}$. \end{remark} Almost greedy bases have been studied, among other papers, in \cite{AA2017, AADK2019, B2019,DHK2006, DKK2003}, while semi-greedy bases have been studied, for example, in \cite{BL2021, B2019, BBGHO2018, T2008}. \\ Originally defined for Schauder bases, the concepts of quasi-, almost and semi-greediness were extended to and studied in the context of seminormalized minimal systems with seminormalized biorthogonal functionals (see for example \cite{AABW2021, BBGHO2018, DKO2015, W2000}). It is known that quasi-greedy systems are Markushevich bases \cite{AABW2021}, and that almost greediness and semi-greediness are equivalent concepts for Markushevich bases (see \cite[Theorem~4.2]{BL2021}, \cite[Theorem~1.10]{B2019} and \cite[Theorem~5.3]{DKK2003}), but not for general minimal systems \cite[Example~4.5]{BL2021}. \\ Weaker versions of the TGA and the CGA have also been studied. In the case of the WTGA, one may mention for example \cite{DKO2015, DKSW2012, G2009, KT2002, KT2003}, and for the WCGA see \cite{DGHKT2021, DKO2015, DKSW2012, T2015} among others. These algorithms consider approximations involving $t$-greedy sets, for some $0<t\le 1$. \\ Recently, Dilworth, Kutzarova, Temlyakov and Wallis extended the concepts almost greedy and semi-greedy Schauder bases to the context of sequences of weights \cite{DKTW2018}. In their work, the authors follow a similar extension for greedy Schauder bases previously introduced and studied by Kerkyacharian, Picard and Temlyakov \cite{KPT2006}. In \cite{DKTW2018} it is shown that weight-almost greedy bases are weight-semi-greedy and that the converse holds for Schauder bases if the Banach space has finite cotype. The cotype condition is removed in \cite{B2020}, where the author raises the question of extending the results to a more general class of Markushevich bases. \\ In this context, we focus on weight-semi-greedy bases as well as a weak variant of them involving weak algorithms, and in particular the implication from (weak) weight-semi-greediness to weight-almost greediness, under different hypothesis. We prove that, for a large class of weights, (weak) weight-semi-greedy Markushevich bases are weight-almost greedy. Also, we prove that if we impose some mild conditions on the basis, the above holds for any weight. Finally, in this context we show that for any weight $\mathbf{w}$, a weight semi-greedy basis with weight $\mathbf{w}$ is truncation quasi-greedy and $\mathbf{w}$-superdemocratic. In the course of our study, some new properties arise naturaly, namely \emph{almost semi-greedy} Markushevich bases, and its weighted and weak counterparts. We show that this property turns out to be equivalent to squeeze symmetry, and can be used to characterize bidemocracy. \\ Additionally, we also study parameters involving the WTGA in the classical context, that is for constant weights. \\ The paper is structured as follows: In Section~\ref{sectionseparation}, we recall results from \cite{BL2021} about the finite dimensional separation property (FDSP) which is useful to replace arguments involving the Schauder basis constant when working with Markushevich bases. We also study a variant of this property for bounded uniformly discrete sequences that allows us to improve some of the upper bounds that would be obtained by using the FDSP. In Section~\ref{sectionweightweaksemigreedy} we study weight-semi-greedy bases. We introduce the weak notions of weight-almost and semi-greedy bases and prove our main results: Theorem~\ref{theoremswemigreedydisjointgreedynotc0}, and Theorem~\ref{theoremc0notl1norming}. Our proofs lead us to consider an intermediate notion that we call almost semi-greedy bases and the formally weaker versions thereof, as well as their weighted counterparts. These properties show a tight connection between the CTGA and two properties generally studied in connection to the TGA: bidemocracy and squeeze symmetry. This can be found in Section~\ref{sectionalmostsemigreedy}. To complete the framework for our study of weighted bases, in Section~\ref{sectionweightweakalmostgreedy}, we deal with the formally weaker variant of weight-almost greedy bases. We show that, as it is the case of weak almost greedy bases (see \cite[Proposition 2.3]{BL2021}, \cite[Theorem 6.4]{DKSW2012}) they are equivalent to the weight almost greedy property. In Section~\ref{sectionquasi-greedyalmostgreedy}, we extend some of the results on the WTGA and CTGA from \cite{BL2021}. In particular, we study parameters that allow us to estimate how the algorithms involved in said results perform with respect to bases that are not necessarily semi-greedy. Additionally, for bases that are (weak)-semi-greedy, we improve the known estimate for the quasi-greedy constant, and give a new estimate for the almost greedy one. Finally, in Section~\ref{sectionexamples}, we give some examples of bases with the properties we study. \\ Our general notation is standard. In addition to what was set before, unless otherwise stated, ${\mathbb X}$, ${\mathbb Y}$ and ${\mathbb Z}$ denote infinite-dimensional Banach spaces, whereas $\mathbb{E}$ and ${\mathbb F}$ denote finite dimensional spaces and ${\mathbb V}$ stands for a Banach space without any restrictions on the dimension. Given a Banach space ${\mathbb V}$ over $\mathbb{K}$, $S_{{\mathbb V}}$ denotes its unit sphere and ${\mathbb V}^*$ denotes its dual space. For $x\in {\mathbb V}$, $\widehat{x}$ denotes the image of $x$ in the bidual space ${\mathbb V}^{**}$, via the canonical inclusion. The same notation will be used for subsets of ${\mathbb V}$. The constant $\kappa$ is set as $\kappa=1$ if $\mathbb{K}=\mathbb{R}$, and $\kappa=2$ if $\mathbb{K}=\mathbb{C}$. \\ Given any set $A\subset \mathbb{N}$, we define $$ \ensuremath{\mathcal{E}}_A:=\{\varepsilon=(\varepsilon_i)_{i\in A}\colon\ |\varepsilon_i|=1\quad \forall i\in A\} $$ with the convention that $\ensuremath{\mathcal{E}}_A=\emptyset$ if $A=\emptyset$. When $A\subset \mathbb{N}$ is finite and ${\mathcal B}=({\bm x}_i)_{i\in\mathbb{N}}$ is a basis, for $\varepsilon\in\ensuremath{\mathcal{E}}_A$, we denote $$ {\mathbf 1}_{\varepsilon, A}:={\mathbf 1}_{\varepsilon, A,{\mathcal B}}=\sum_{i\in A}\varepsilon_i{\bm x}_i $$ with the convention that any sum over the empty set is zero. Also, if $\varepsilon\in \ensuremath{\mathcal{E}}_A$ and $B\subset A$, we write ${\mathbf 1}_{\varepsilon,B}$ considering the natural restriction of $\varepsilon$ to $B$. If $\varepsilon_j=1$ for all $j$, we write ${\mathbf 1}_A$. If ${\mathcal B}^*=({\bm x}_i^*)_{i\in\mathbb{N}}$ is the dual basis of a basis ${\mathcal B}$, the supremum norm of $x\in {\mathbb V}$ is $\|x\|_{\infty}:=\displaystyle \sup_{i\in \mathbb{N}}|{\bm x}_i^*(x)|$, whereas for $1\le p<\infty$, $\|x\|_p$ denotes the usual $\ell_p$ norm, when it is defined. Finally, for each $x\in {\mathbb X}$ we define $\varepsilon(x)=(\sgn({\bm x}_i^*(x)))_{i\in \mathbb{N}}$, where $\sgn(0)=1$ and $\sgn(t)t=|t|$ for all $t\in \mathbb{K}$. \\ All the remaining relevant terminology and preliminaries will be given in corresponding sections. \section{Separation properties}\label{sectionseparation} In this section, we recall the definition of the finite dimensional separation property (FDSP) and some results from \cite{BL2021} used to prove the implication from semi-greedy to almost greedy for Markushevich bases. Also, we study a related property that allows us to improve some of the upper bounds that would be obtained by using the FDSP. \begin{definition} (\cite[Definition 3.1]{BL2021}) \label{definitionseparation} A sequence $(u_i)_{i\in\mathbb{N}}\subseteq {\mathbb X}$ has the \emph{finite dimensional separation property} (FDSP for short) if there is a positive constant $\mathbf{M}$ such that for every separable subspace ${\mathbb Z}\subset {\mathbb X}$ and every $\epsilon>0$, there is a basic subsequence $(u_{i_k})_{k\in\mathbb{N}}$ with basis constant no greater than $\mathbf{M}+\epsilon$ satisfying the following: For every finite dimensional subspace ${\mathbb F} \subset {\mathbb Z}$ there is $j_{{\mathbb F},\epsilon}=j_{{\mathbb F},((u_{i_k})_{k\in \mathbb{N}},\epsilon)}\in \mathbb{N}$ such that \begin{equation} \|x\|\le (\mathbf{M}+\epsilon)\|x+z\|,\label{separation} \end{equation} for all $x\in {\mathbb F}$ and all $z\in \overline{[u_{i_k}:k> j_{{\mathbb F}, \epsilon}]}$. We call any such subsequence a \emph{finite dimensional separating sequence} for $({\mathbb Z}, \mathbf{M}, \epsilon)$, and the minimum $\mathbf{M}$ for which this property holds will be called the \emph{finite dimensional separation constant} $\mathbf{M}_{fs}[ (u_i)_{i\in \mathbb{N}}, {\mathbb X}]$, leaving the sequence and the space implicit when it is clear. \end{definition} \begin{remark}Note that a subsequence $(u_{i_k})_{k\in\mathbb{N}}$ is finite dimensional separating for $({\mathbb Z}, \mathbf{M}, \epsilon)$ if and only if \eqref{separation} holds for any $x\in S_{{\mathbb F}}$ and all $z\in [u_{i_k}:k> j_{{\mathbb F},\epsilon}]$. \end{remark} Recall that a subspace ${\mathbb Y}\subset {\mathbb X}^*$ is said to be $r$-norming for ${\mathbb X}$, $0<r\le 1$, if $$ r\|x\|\le \sup_{\substack{x^*\in S_{{\mathbb Y}}}}|x^*(x)|. $$ Additionally, we will say that a set $R\subset {\mathbb X}^*$ is $r$-norming if the subspace it spans in ${\mathbb X}^*$ is $r$-norming. \\ Also recall that a sequence $(v_i)_{i\in \mathbb{N}}$ is a \emph{block basis} of a Markushevich basis $({\bm x}_k)_{k\in \mathbb{N}}$ if there is a sequence of scalars $(b_k)_{k\in\mathbb{N}}$ and sequences of positive integers $(n_i)_{i\in\mathbb{N}}$, $(m_i)_{i\in\mathbb{N}}$ with $n_i\le m_i<n_{i+1}$ for all $i$ such that $$ v_i=\sum\limits_{k=n_i}^{m_i}b_k{\bm x}_k, $$ with at least one nonzero $b_k$ for each $i\in \mathbb{N}$. In particular, any subsequence of a Markushevich basis is a block basis of it. \begin{proposition} {\rm (\cite[{Proposition 3.11}]{BL2021})} \label{propositionseparation} Let $(v_i)_{i\in\mathbb{N}}\subset {\mathbb X}$ be a block basis of a Markushevich basis $({\bm y}_k)_{k\in\mathbb{N}}$ for a subspace ${\mathbb Y}\subset {\mathbb X}$ with biorthogonal functionals $({\bm y}_k^{*})_{k\in\mathbb{N}}$. Let $(a_i)_{i\in\mathbb{N}}$ be a scalar sequence such that $(z_i:=a_i v_i)_{i\in\mathbb{N}}$ is seminormalized. The following hold: \begin{enumerate}[\upshape (i)] \item \label{Markushevich} $(z_i)_{i\in\mathbb{N}}$ and $(v_i)_{i\in\mathbb{N}}$ have the finite dimensional separation property with the same constant, that is $\mathbf{M}_{fs}[(v_i)_{i\in\mathbb{N}}, {\mathbb X}]=\mathbf{M}_{fs}[(z_i)_{i\in\mathbb{N}}, {\mathbb X}]$. \item \label{weakzero} If either $0\in \overline{\{z_i \}}^{w}_{i\in\mathbb{N}}$ or ${\mathbb X}$ is a dual space and $0\in \overline{\{z_i \}}^{w^{*}}_{i\in\mathbb{N}}$, then $\mathbf{M}_{fs}=1$. \item \label{noncompact} If $\overline{\{z_i \}}^{w}_{i\in\mathbb{N}}$ is not weakly compact, then \begin{equation} \mathbf{M}_{fs} \le \Big( 2+\inf \Big\{ \frac{\|x^{**}\|}{\dist{(x^{**},\widehat{{\mathbb X}})}}\colon\ \ x^{**}\in \overline{\{\widehat{z}_{i\in \mathbb{N}}\}}^{w^{*}}_{i\in\mathbb{N}}\setminus \widehat{{\mathbb X}}\Big\}\Big)^2.\nonumber \end{equation} \item \label{norming} If ${\mathbb Y}={\mathbb X}$ and $\overline{[{\bm y}^*_k \colon k\in \mathbb{N}]}$ is $r$-norming, then $\mathbf{M}_{fs}\le r^{-1}$. \item \label{schauder} If ${\mathbb Y}={\mathbb X}$ and $({\bm y}_k)_{k\in\mathbb{N}}$ is a Schauder basis for ${\mathbb X}$ with constant $\mathbf{K}_b$, then $\mathbf{M}_{fs}\le \mathbf{K}_{b}$. \end{enumerate} \end{proposition} \begin{remark}\label{remarkdisjointlysupported} Proposition~\ref{propositionseparation} remains valid if we replace a block basis with any sequence of nonzero vectors, that are pairwise disjointly supported. \end{remark} Next, we prove further results involving a similar property. Given $\delta>0$, we say that a set $S\subset {\mathbb X}$ is $\delta$-uniformly discrete if $\|x-y\|\ge \delta$ for all $x, y\in S$, $x\ne y$. For a sequence $(x_i)_{i\in \mathbb{N}}$ by $\delta$-uniformly discrete we mean that $\|x_i-x_j\|\ge \delta$ for all $i\not=j$. \begin{lemma}\label{lemmaseparated1} Let $S\subset {\mathbb X}$ be a bounded uniformly discrete set and ${\mathbb F}\subset {\mathbb X}$ a finite-dimensional subspace. Given $\epsilon>0$, there are $x \ne y\in S$ such that for every $b \in \mathbb{K}$ and every $z\in {\mathbb F}$, $$ \|z\|\le (1+\epsilon)\|z+b (x -y)\|. $$ \end{lemma} \begin{proof} Choose $\delta>0$ so that $S$ is $\delta$-uniformly discrete, and $0<\epsilon'<\epsilon$ so that \begin{equation} 0<\frac{1}{1-4\delta^{-1}\epsilon'-\epsilon'}<1+\epsilon.\label{epsilon'} \end{equation} Let $\{z_1,\dots,z_n\}$ be an $\epsilon'$-net in $S_{{\mathbb F}}$, and $\{z_1^*,\dots,z_n^*\}\subset S_{{\mathbb X}^*}$ so that $z_j^*(z_j)=1$ for all $1\le j\le n$. Since $S$ is bounded, there exists $z_0^{**}\in {\mathbb X}^{**}$ a $w^*$-accumulation point of $\widehat{S}\subset {\mathbb X}^{**}$. Hence, there are $x\ne y\in S$, such that for $1\le k\le n$, $$ |z_0^{**}(z_k^{*})-z_k^{*}(x)|\le \epsilon' \quad \text{and}\quad |z_0^{**}(z_k^{*})-z_k^{*}(y)|\le \epsilon' $$ Fix $z\in S_{{\mathbb F}}$, and choose $1\le k\le n$ so that $$ \|z_k-z\|\le \epsilon'. $$ Now pick $b \in \mathbb{K}$. If $|b|\le 2\delta^{-1}$, then \begin{align*} \|z+b(x - y)\|\ge& \|z_k+b(x - y)\|-\epsilon'\ge |z_{k}^{*}(z_k+b(x - y)|-\epsilon'\\ \ge& 1-|b||z_k^{*}(x - y)|-\epsilon'\\ \ge& 1-2\delta^{-1}|z_0^{**}(z_k)-z_k^{*}(x) + z_{k}^{*}(y) - z_0^{**}(z_k)|-\epsilon'\\ \ge& 1-4\delta^{-1}\epsilon'-\epsilon'. \end{align*} Hence, by \eqref{epsilon'}, $$ \|z\|=1\le (1+\epsilon)\|z+b(x - y)\|. $$ On the other hand, if $|b|>2\delta^{-1}$, then $$ \|z+b(x - y)\|\ge |b|\|x - y\|-\|z\|\ge 1=\|z\|. $$ This completes the proof for $z\in S_{{\mathbb F}}$, and hence by scaling for all $z\in {\mathbb F}$. \end{proof} \begin{lemma}\label{lemmaxi-x2} Let ${\mathbb X}$ be a Banach space, and $(u_j)_{j\in\mathbb{N}}\subseteq {\mathbb X}$ a bounded uniformly discrete sequence. Then, for any separable subspace ${\mathbb Z}\subset {\mathbb X}$ and $\epsilon>0$ there is a subsequence $(u_{j_n} )_{n\in\mathbb{N}}$ such that the sequence $\{u_{j_{2n-1}}-u_{j_{2n}}\}_{n\in\mathbb{N}}$ is basic with basis constant no greater than $(1+\epsilon)$ and satisfies the following: For any finite dimensional subspace ${\mathbb F}\subset {\mathbb Z}$ and every $\xi>0$, there is $r_{{\mathbb F},\xi}\in \mathbb{N}$ such that for all $y\in {\mathbb F}$ and all $v \in \overline{[u_{j_{2n-1}}-u_{j_{2n}}: n> r_{{\mathbb F},\xi}]}$, $$ \|y\|\le (1+\xi)\|y+v \|. $$ \end{lemma} \begin{proof} The argument is very similar to that of \cite[Lemma 3.5]{BL2021}; we give a proof for the sake of completeness. Fix ${\mathbb Z}\subset {\mathbb X}$ a separable subspace and $\epsilon>0$. Choose a sequence $(v_j)_{j\in\mathbb{N}}\subset {\mathbb Z}$, dense in ${\mathbb Z}$, and a sequence of positive scalars $(\epsilon_j)_{j\in\mathbb{N}}$ so that $\prod\limits_{j=1}^{\infty}(1+\epsilon_j)\le (1+\epsilon)$. Let $j_0:=1$. Applying Lemma~\ref{lemmaseparated1} to the set $\{u_j\}_{j>j_0}$, we can find $j_0<j_1<j_2$ so that for all $y\in [v_k, u_k: 1\le k\le j_0 ]$ and all $b\in \mathbb{K}$, $$ \|y\|\le (1+\epsilon_1)\|y+b (u_{j_1}-u_{j_2}) \|. $$ Similarly, we can find $j_2<j_3<j_4$ so that for all $y\in [v_k, u_k: 1\le k \le j_2]$ and all $b\in \mathbb{K}$, $$ \|y\|\le (1+\epsilon_2)\|y+b (u_{j_3}-u_{j_4}) \|. $$ By an inductive argument, we obtain a strictly increasing sequence of positive integers $\{j_n\}_{n\in\mathbb{N}}$ such that for all $y\in [v_k, u_k\colon 1\le k\le j_{2n-2}]$, $b\in \mathbb{K}$ and $n\in\mathbb{N}$, $$ \|y\|\le (1+\epsilon_n)\|y+b (u_{j_{2n-1}}-u_{j_{2n}})\|. $$ Then, for any positive integers $m\le l$, any $y\in [v_k, u_k\colon 1\le k\le j_{2m-2}]$ and any scalars $(a_n)_{m \le n\le l}$, \begin{align*} \|y\|\le& \prod\limits_{n=m}^{l}(1+\epsilon_{n})\|y+\sum\limits_{n=m}^{l}a_n (u_{j_{2n-1}}-u_{j_{2n}})\|\\ \le& \prod\limits_{n=m}^{\infty}(1+\epsilon_{n})\|y+\sum\limits_{n=m}^{l}a_n (u_{j_{2n-1}}-u_{j_{2n}})\|. \end{align*} In particular, $(u_{j_{2n-1}}-u_{j_{2n}})_{n\in \mathbb{N}}$ is basic with basis constant no greater than $\prod\limits_{n=1}^{\infty}(1+\epsilon_n)\le 1+\epsilon$, and, given ${\mathbb F} \subset [ v_j\colon 1\le j\le n]$ for some $n\in \mathbb{N}$, we can pick $r_{{\mathbb F},\xi}$ using the above computation. Now, standard density arguments allow us to obtain the result for any finite dimensional subspace of ${\mathbb Z}=\overline{[v_j \colon j\in\mathbb{N}]}$. \end{proof} \medskip \begin{corollary}\label{corollaryboth} Let ${\mathcal B}=({\bm x}_i)_{i\in\mathbb{N}} \subset {\mathbb X}$ be a seminormalized Markushevich basis for ${\mathbb Y}\subset {\mathbb X}$, with finite dimensional separation constant no greater than $\mathbf{M}$. Then for every separable subspace ${\mathbb Z}\subset {\mathbb X}$ and every $\epsilon>0$, there is a basic subsequence $({\bm x}_{i_k})_{k\in\mathbb{N}}$ satisfying the following condition: For every finite dimensional subspace ${\mathbb F}\subset {\mathbb Z}$, there is $s_{{\mathbb F},\epsilon}\in\mathbb{N}$ such that for every $x\in {\mathbb F}$, every $y \in \overline{[{\bm x}_{i_k}: k\ge s_{{\mathbb F},\epsilon}]}$ and every $z\in \overline{[{\bm x}_{i_{2k-1}}-{\bm x}_{i_{2k}}: k\ge s_{{\mathbb F},\epsilon}]}$, \begin{equation} \|x\|\le \min{\{(\mathbf{M}+\epsilon)\|x+y\|, (1+\epsilon)\|x+z\|\}}.\nonumber \end{equation} \end{corollary} \begin{proof} Fix ${\mathbb Z}\subset {\mathbb X}$ a separable subspace and $\epsilon>0$. An application of Proposition~\ref{propositionseparation} gives a separating subsequence $({\bm x}_{i_l})_{l\in\mathbb{N}}$ for $({\mathbb Z},\mathbf{M},\epsilon)$. Thus, for any finite dimensional ${\mathbb F}\subset {\mathbb Z}$, every $x\in {\mathbb F}$ and every $y \in \overline{\left[{\bm x}_{i_l}: l > j_{{\mathbb F},\epsilon}\right]}$, $$ \|x\|\le (\mathbf{M}+\epsilon)\|x+y\|. $$ Let ${\mathcal B}^*=({\bm x}_i^*)_{i\in\mathbb{N}}$ be the dual basis of ${\mathcal B}$. Since $({\bm x}^*_i)_{i\in \mathbb{N}}$ is bounded, $({\bm x}_{i_l})_{l\in\mathbb{N}}$ is uniformly discrete, then by Lemma~\ref{lemmaxi-x2} we obtain a further subsequence $({\bm x}_{i_{l_k}})_{k\in\mathbb{N}}$ such that for any finite dimensional ${\mathbb F}\subset {\mathbb Z}$, every $x\in {\mathbb F}$ and every $z \in \overline{\left[{\bm x}_{i_{l_{2k-1}}}-{\bm x}_{i_{l_{2k}}}: k> r_{{\mathbb F},\epsilon}\right]}$, $$ \|x\|\le (1+\epsilon)\|x+z\|. $$ Taking for each ${\mathbb F}$, $s_{{\mathbb F},\epsilon}:=1+\max\{r_{{\mathbb F},\epsilon}, j_{{\mathbb F},\epsilon}\}$, it is immediate from the above that $({\bm x}_{i_{l_k}})_{k\in\mathbb{N}}$ has the desired properties. \end{proof} \medskip \section{Weak weight-semi-greedy bases}\label{sectionweightweaksemigreedy} Let $\mathbf{w}=(w_i)_{i\in \mathbb{N}}$ be a sequence of positive numbers, and for each set $A\subset \mathbb{N}$, let $$ w(A):=\sum_{i\in A}w_i. $$ The sequence $\mathbf{w}$ is called a \emph{weight} and, for $A\subset\mathbb{N}$, $w(A)$ is the $\mathbf{w}$-measure of $A$ (which might be infinite if $|A|=\infty$). In this section we introduce the weak weight-semi-greedy property, and study its relation with other notions studied in this context, in particular the weight-almost greedy property, introduced in \cite{DKTW2018}, which is a weaker version of the weight greedy property introduced and studied in \cite{KPT2006}. \begin{definition}\label{definitionweightsemigreedyetc}Let ${\mathcal B}$ be a basis for ${\mathbb X}$, $\mathbf{w}$ a weight, and $\mathbf{C}>0$. Then: \begin{itemize} \item ${\mathcal B}$ is \emph{weight-almost greedy with weight $\mathbf{w}$ and constant $\mathbf{C}$} (or $\mathbf{C}$-$\mathbf{w}$-almost greedy) if \begin{equation*} \|x-P_{\Lambda_m(x)}\|\le \mathbf{C} \inf_{\substack{B\in \mathbb{N}^{<\infty}\\ w(B)\le w(\Lambda_m)}}\|x-P_B(x)\|, \qquad\forall x\in {\mathbb X},\ \forall m\in \mathbb{N}. \end{equation*} \item ${\mathcal B}$ is weight-semi-greedy with weight $\mathbf{w}$ and constant $\mathbf{C}$ (or $\mathbf{C}$-$\mathbf{w}$-semi-greedy) if \begin{equation*} \inf_{\supp(y)\subset \Lambda_m(x)}\|x-y\|\le \mathbf{C} \inf_{\substack{z\in {\mathbb X}\\|\supp(z)|<\infty\\ w(\supp(z))\le w(\Lambda_m)}}\|x-z\|, \qquad\forall x\in {\mathbb X},\ \forall m\in \mathbb{N}. \end{equation*} \end{itemize} \end{definition} As in the case of their regular counterparts, these weighted properties can be defined considering $\Lambda_m$, or all greedy sets, or at least one, obtaining equivalent notions. Also, for $\mathbf{w}$-greedy and $\mathbf{w}$-semi-greedy bases, it is equivalent to consider approximations using vectors with only finite support or any support, provided that the weight condition is kept. For the sake of completion, we give the proof for weight-semi-greedy bases; the proof for weight-greedy bases is similar. We will use a result that follows at once from the proof of \cite[Proposition 2.3]{B2020}, which is an extension of \cite[Proposition 4.5]{DKTW2018} without the Schauder hypothesis, and which uses the same definition of $\mathbf{w}$-semi-greedy bases given above. \begin{lemma}\label{lemmal1thenc0}Let ${\mathcal B} =({\bm x}_i)_{i\in \mathbb{N}}$ be a basis for ${\mathbb X}$ and $\mathbf{w} =(w_i)_{i\in \mathbb{N}}$ a weight. If ${\mathcal B}$ is $\mathbf{w}$-semi-greedy and $(w_{i_k})_{k\in \mathbb{N}}$ is a subsequence such that $\sum_{k}w_{i_k}<\infty$, then $({\bm x}_{i_k})_{k\in \mathbb{N}}$ is a basic sequence equivalent to the canonical unit vector basis of $\mathtt{c}_0$. \end{lemma} \begin{remark}\label{remarkprojectionc0}\rm Note that under the conditions of Lemma~\ref{lemmal1thenc0}, if $A\subset \mathbb{N}$ is an infinite set such that $w(A)<\infty$, then the projections $P_A(x)$ are defined for each $x\in {\mathbb X}$, with unconditional convergence of the sums. Indeed, this follows at once from the fact that ${\mathcal B}^*$ is weak star null. Also, due to the totality condition, if $x$ has infinite support and finite $\mathbf{w}$-measure, then $$ x=\sum_{i\in \mathbb{N}}{\bm x}_i^*(x){\bm x}_i, $$ again with unconditional convergence. In particular, by \cite[Theorem 4.3]{DKTW2018}, this holds for $\mathbf{w}$-almost greedy bases. \end{remark} \begin{lemma}\label{lemmaallthesameweightsemigreedy} Let ${\mathcal B}$ be a basis for ${\mathbb X}$, $\mathbf{w}$ a weight and $\mathbf{C}>0$. The following are equivalent: \begin{enumerate}[\upshape (i)] \item \label{sgnallgredyset} For all $x\in {\mathbb X}$, $m\in \mathbb{N}$ and $A\in {\mathcal G}(x,m,1)$, there is $y\in {\mathbb X}$ with $\supp(y)\subset A$ such that $$ \|x-y\|\le \mathbf{C} \inf_{\substack{z\in {\mathbb X}\\ w(\supp(z))\le w(A)}}\|x-z\|. $$ \item \label{sgnallgredysetfinite} For all $x\in {\mathbb X}$, $m\in \mathbb{N}$ and $A\in {\mathcal G}(x,m,1)$, there is $y\in {\mathbb X}$ with $\supp(y)\subset A$ such that $$ \|x-y\|\le \mathbf{C} \inf_{\substack{z\in {\mathbb X}\\|\supp(z)|<\infty\\ w(\supp(z))\le w(A)}}\|x-z\|. $$ \item \label{sgnatgredyset} For all $x\in {\mathbb X}$ and $m\in \mathbb{N}$, there is $y\in {\mathbb X}$ with $\supp(y)\subset \Lambda_m(x)$ such that $$ \|x-y\|\le \mathbf{C} \inf_{\substack{z\in {\mathbb X}\\ w(\supp(z))\le w(\Lambda_m)}}\|x-z\|. $$ \item \label{sgnatgredysetfinite} For all $x\in {\mathbb X}$ and $m\in \mathbb{N}$, there is $y\in {\mathbb X}$ with $\supp(y)\subset \Lambda_m(x)$ such that $$ \|x-y\|\le \mathbf{C} \inf_{\substack{z\in {\mathbb X}\\|\supp(z)|<\infty\\ w(\supp(z))\le w(\Lambda_m)}}\|x-z\|. $$ \item \label{sgnsomegredyset} For all $x\in {\mathbb X}$ and $m\in \mathbb{N}$, there is $A\in {\mathcal G}(x,m,1)$ and $y\in {\mathbb X}$ with $\supp(y)\subset A$ such that $$ \|x-y\|\le \mathbf{C} \inf_{\substack{z\in {\mathbb X}\\ w(\supp(z))\le w(A)}}\|x-z\|. $$ \item \label{sgnsomegredysetfinite} For all $x\in {\mathbb X}$ and $m\in \mathbb{N}$, there is $A\in {\mathcal G}(x,m,1)$ and $y\in {\mathbb X}$ with $\supp(y)\subset A$ such that $$ \|x-y\|\le \mathbf{C} \inf_{\substack{z\in {\mathbb X}\\|\supp(z)|<\infty\\ w(\supp(z))\le w(A)}}\|x-z\|. $$ \end{enumerate} \end{lemma} \begin{proof}The implications \ref{sgnallgredyset}$ \Longrightarrow$ \ref{sgnatgredyset} $\Longrightarrow$ \ref{sgnsomegredyset} and \ref{sgnallgredysetfinite} $\Longrightarrow$ \ref{sgnatgredysetfinite} $\Longrightarrow$ \ref{sgnsomegredysetfinite}, as well as \ref{sgnallgredyset} $\Longrightarrow$ \ref{sgnallgredysetfinite}, together with \ref{sgnatgredyset} $\Longrightarrow$ \ref{sgnatgredysetfinite} and \ref{sgnsomegredyset} $\Longrightarrow$ \ref{sgnsomegredysetfinite} are immediate. \\ Let us prove \ref{sgnsomegredyset} $\Longrightarrow$ \ref{sgnallgredyset}: Fix $A\in {\mathcal G}(x,m,1)$. We may assume $x\not=P_A(x)$ (else, we take $y=x$), so $|\supp(x)|>m$. For each $l\in \mathbb{N}$ choose $z_l\in {\mathbb X}$ with $w(\supp(z_l))\le w(A)$ so that $$ \|x-z_l\|\le (1+\frac{1}{l}) \inf_{\substack{z\in {\mathbb X}\\ w(\supp(z))\le w(A)}}\|x-z\|, $$ which is possible because $A\subsetneq \supp(x)$ so the infimum above is not zero. \\ For each $n \in \mathbb{N}$, set $x_{n}:=x+ \frac{1}{n}P_A(x)$. As ${\mathcal G}(x_n,m,1)=\{A\}$ for each $n\in \mathbb{N}$, for each $l\in \mathbb{N}$ there is $y_{n,l}$ with $\supp(y_{n,l})\subset A$ such that $$ \|x_n-y_{n,l}\|\le \mathbf{C}\|x_n-z_l\|. $$ Given that $A\in \mathbb{N}^{<\infty}$, for fixed $l\in \mathbb{N}$ there is $y_l$ with $\supp(y_l)\subset A$ and a subsequence $(y_{n_{k,l},l})_{k\in \mathbb{N}}$ convergent to $y_l$. Letting $k\to \infty$ in $$ \|x_{n_{k,l}}-y_{n_{k,l},l}\|\le \mathbf{C}\|x_{n_{k,l}}-z_l\|, $$ we obtain $$ \|x-y_l\|\le \mathbf{C}\|x-z_l\|\le \mathbf{C} (1+\frac{1}{l}) \inf_{\substack{z\in {\mathbb X}\\ w(\supp(z))\le w(A)}}\|x-z\|. $$ Reasoning as before and taking a subsequence if necessary, we may assume that $(y_l)_{l\in \mathbb{N}}$ is convergent to some $y$ with $\supp(y)\subset A$, so we complete the step letting $l\rightarrow \infty$. \\ The implication \ref{sgnsomegredysetfinite} $\Longrightarrow$ \ref{sgnallgredysetfinite} is proven by the same argument as that given above to prove \ref{sgnsomegredyset} $\Longrightarrow$ \ref{sgnallgredyset}. \\ Finally, we show that \ref{sgnatgredysetfinite} $\Longrightarrow$ \ref{sgnatgredyset}. Fix $x\in {\mathbb X}$ and $m\in \mathbb{N}$. Suppose there is $z\in {\mathbb X}$ with $|\supp(z)|=\infty$ and $w(\supp(z))\le w(\Lambda_m(x))$. Then by Remark~\ref{remarkprojectionc0}, given $\epsilon>0$ there is a finite set $B\subset \supp(z)$ such that $$ \|z-P_B(z)\|\le \epsilon. $$ It follows that $$ \inf_{\substack{z\in {\mathbb X}\\ w(\supp(z))\le w(\Lambda_m)}}\|x-z\|=\inf_{\substack{z\in {\mathbb X}\\|\supp(z)|<\infty\\ w(\supp(z))\le w(\Lambda_m)}}\|x-z\|, $$ so the proof is complete. \end{proof} It was shown in \cite[Theorem 4.3]{DKTW2018} that every $\mathbf{w}$-almost greedy Schauder basis is $\mathbf{w}$-semi greedy (see also \cite[Theorem 1.11]{B2020}, which improves the bound for the $\mathbf{w}$-semi-greedy constant). Both proofs are valid for Markushevich basis. The implication from $\mathbf{w}$-semi-greedy to $\mathbf{w}$-almost greedy was first proven for Schauder bases in spaces with finite cotype in \cite[Theorem 4.15]{DKTW2018}. The cotype condition in \cite[Theorem 4.15]{DKTW2018} was later removed in \cite[Theorem 1.11]{B2020}. In these papers, $\mathbf{w}$-almost greedy bases were characterized as those being quasi-greedy and $\mathbf{w}$-superdemocratic or $\mathbf{w}$-disjoint-superdemocratic. Below, we find hypothesis weaker than the Schauder condition under which $\mathbf{w}$-semi-greedy bases are $\mathbf{w}$-almost greedy. Additionally, we prove that all $\mathbf{w}$-semi-greedy Markushevich bases are $\mathbf{w}$-superdemocratic and that they have a property, called ``Property (C)'', with roots in \cite[Lemma 2.2]{DKKT2003} that was studied, for instance, in \cite{B2020} and \cite{BDKOW2019}. Before we go on, we give the relevant definitions, as well as some related notions. \begin{definition} Let ${\mathcal B}$ be a basis for ${\mathbb X}$, $\mathbf{w}$ a weight and $\mathbf{C}>0$. Then: \begin{itemize} \item ${\mathcal B}$ is \emph{weight superdemocratic with weight $\mathbf{w}$ and constant $\mathbf{C}$} (or $\mathbf{C}$-$\mathbf{w}$-superdemocratic) if \begin{align} \|{\mathbf 1}_{\varepsilon,A}\|\le \mathbf{C} \|{\mathbf 1}_{\varepsilon',B}\|,\qquad \forall A, B \in \mathbb{N}^{<\infty}, w(A)\le w(B), \varepsilon\in \ensuremath{\mathcal{E}}_A, \varepsilon'\in \ensuremath{\mathcal{E}}_B. \nonumber \end{align} \item ${\mathcal B}$ is \emph{weight disjoint superdemocratic with weight $\mathbf{w}$ and constant $\mathbf{C}$} (or $\mathbf{C}$-$\mathbf{w}$-disjoint superdemocratic) if the above holds for $A$ and $B$ disjoint sets. \end{itemize} When taking ${\mathbf 1}_A$ and ${\mathbf 1}_B$ instead of ${\mathbf 1}_{\varepsilon,A}$ and ${\mathbf 1}_{\varepsilon',B}$, the basis ${\mathcal B}$ is $\mathbf{w}$-democratic (see \cite{DKTW2018, KPT2006}) and $\mathbf{w}$-disjoint democratic, respectively. \end{definition} \begin{definition}Let ${\mathcal B}$ be a basis for ${\mathbb X}$ with dual basis ${\mathcal B}^*=({\bm x}_i^*)_{i\in \mathbb{N}}$. We say that ${\mathcal B}$ \emph{has Property (C) with constant $\mathbf{K}>0$} if \begin{align} \min_{i\in A}|{\bm x}_i^*(x)|\|{\mathbf 1}_{\varepsilon,A}\|\le \mathbf{K}\|x\|,\qquad\forall x\in {\mathbb X},\, \forall A\in \mathbb{N}^{<\infty},\, \forall \varepsilon\in \ensuremath{\mathcal{E}}_{A}.\label{propertyC} \end{align} \end{definition} When \eqref{propertyC} holds only for $\varepsilon=\varepsilon(x)$, this property has been studied under the name ``truncation-quasi-greediness'' (see \cite{AABBL2021} and \cite{AABBL2021b}), which reflects the fact that the restricted truncation operator is bounded (see \cite{AABW2021}). These two definitions are equivalent (see \cite[Proposition 4.16]{AABW2021}). We will keep the latter terminology, though some of our proofs establish also upper bounds for the constant in \eqref{propertyC}. \begin{remark}\rm \label{remarkpropc=tqg} Note that if ${\mathcal B}$ is $\mathbf{C}$-truncation quasi-greedy, it has the $2\kappa \mathbf{C}^2$-property (C). Indeed, given $x\in {\mathbb X}$ with ${\bm x}_i^*(x)\in \mathbb{R}_{\ge 0}$ for all $i$, $A\in {\mathcal G}(x,m,1)$ and $\varepsilon \in \ensuremath{\mathcal{E}}_A$, then \begin{align*} \min_{i\in A}|{\bm x}_i^*(x)|\|{\mathbf 1}_{\varepsilon, A}\|\le& \min_{i\in A}|{\bm x}_i^*(x)|\kappa \max_{\substack{\varepsilon'\in \ensuremath{\mathcal{E}}_A\\\varepsilon'_j\in \{-1,1\} \forall j}}\|{\mathbf 1}_{\varepsilon', A}\|\le 2\kappa \min_{i\in A}|{\bm x}_i^*(x)|\max_{B\subset A}\|{\mathbf 1}_{B}\|\\ \le& 2\kappa \min_{i\in A}|{\bm x}_i^*(x)| \mathbf{C}\|{\mathbf 1}_{A}\|\le 2\kappa\mathbf{C}^2\|x\|. \end{align*} The general case follows because for every $x\in {\mathbb X}$, the basis ${\mathcal B}_x=({\bm y}_i:=\varepsilon_i {\bm x}_i)$ where the $\varepsilon_i$ are chosen so that ${\bm y}_i^*(x)\in \mathbb{R}_{\ge 0}$ for all $i$ is also $\mathbf{C}$-truncation quasi-greedy. \\ Note that the above argument also shows that if ${\mathcal B}$ is $\mathbf{C}$-truncation quasi-greedy, $\|{\mathbf 1}_{\varepsilon, A}\|\le 2\kappa \mathbf{C}\|{\mathbf 1}_{\varepsilon', A}\|$ for any $A\in \mathbb{N}^{<\infty}$, $\varepsilon,\varepsilon'\in \ensuremath{\mathcal{E}}_A$. \end{remark} To prove the implication from $\mathbf{w}$-semi-greedy to $\mathbf{w}$-almost greedy bases, we use the concept of disjoint $\mathbf{w}$-almost greedy bases, which we define in terms of all greedy sets for convenience. \begin{definition}\label{definitiondisjointweightalmostgreedyetc} Let ${\mathcal B}$ be a basis for ${\mathbb X}$, $\mathbf{w}$ a weight, and $\mathbf{C}>0$. We say that ${\mathcal B}$ is \emph{weight-disjoint almost greedy with weight $\mathbf{w}$ and constant $\mathbf{C}$} (or $\mathbf{C}$-$\mathbf{w}$-disjoint almost greedy) if \begin{equation*} \|x-P_A(x)\|\le \mathbf{C} \inf_{\substack{B\in \mathbb{N}^{<\infty}\\ w(B)\le w(A)\\ B\cap A=\emptyset}}\|x-P_B(x)\|, \qquad\forall x\in {\mathbb X},\, \forall A \in {\mathcal G}(x,1). \end{equation*} \end{definition} We will use the following elementary result, which is a weighted variant of \cite[Lemma 6.2]{AABW2021}; we give a proof for the sake of completion. \begin{lemma}\label{lemmadisjoint=}Let ${\mathcal B}$ be a basis for ${\mathbb X}$, $\mathbf{w}$ a weight and $\mathbf{C}>0$. Then ${\mathcal B} $ is $\mathbf{C}$-$\mathbf{w}$-disjoint almost greedy if and only if it is $\mathbf{C}$-$\mathbf{w}$-almost greedy. \end{lemma} \begin{proof} Suppose ${\mathcal B}$ is $\mathbf{C}$-$\mathbf{w}$-disjoint almost greedy, fix $x\in {\mathbb X}$, $A\in {\mathcal G}(x,1)$, and $B\subset\mathbb{N}$ with $w(B)\le w(A)$. If $B\cap A=\emptyset$ or $A=B$, there is nothing to prove. Else, since $A\setminus B\in {\mathcal G}(x-P_{A\cap B}(x),1)$ and $w(B\setminus A)\le w(A\setminus B)$, we have \begin{align*} \|x-P_A(x)\|=&\|x-P_{A\cap B}(x)- P_{A\setminus B}(x)\|\le \mathbf{C} \|x-P_{A\cap B}(x)- P_{B\setminus A}(x)\|\\ =&\mathbf{C}\|x-P_B(x)\|. \end{align*} \end{proof} \begin{remark}\rm In the definition of $\mathbf{w}$-almost greedy bases, the projections are taken over finite sets, which guarantees that they are well defined. But if a basis is $\mathbf{w}$-almost greedy, by Remark~\ref{remarkprojectionc0}, we have $$ \inf_{\substack{B\in \mathbb{N}^{<\infty}\\ w(B)\le w(\Lambda_m)}}\|x-P_B(x)\| = \inf_{\substack{B\subset \mathbb{N}\\ w(B)\le w(\Lambda_m)}}\|x-P_B(x)\|, \qquad\forall x\in {\mathbb X},\, \forall m\in \mathbb{N}. $$ Similar considerations hold for the infima taken over all sets $A\in {\mathcal G}(x,m,1)$. \end{remark} Next, we define a propety that can be seen as an extension of the weight-semi-greedy property as well as an extension, to the context of weights, of the weak semi-greedy property given in \cite[Definition 1.7]{BL2021} . \begin{definition}\label{definitionweakweightsemigreedy} Let ${\mathcal B}$ be a basis for ${\mathbb X}$, $\mathbf{w}$ a weight, $\mathbf{C}>0$ and $0<s\le 1$. We say that ${\mathcal B}$ is \emph{weak weight-semi-greedy with parameter $s$, weight $\mathbf{w}$ and constant $\mathbf{C}$} (or $\mathbf{C}$-$s$-$\mathbf{w}$-semi-greedy) if, for every $x\in {\mathbb X}$ and $m\in \mathbb{N}$, there is $A\in {\mathcal G}(x,m,s)$ and $y\in {\mathbb X}$ with $\supp(y)\subset A$ such that \begin{equation} \|x-y\|\le \mathbf{C}\inf_{\substack{z\in {\mathbb X}\\|\supp(z)|<\infty\\mathbf{w}(\supp(z))\le w(A)}}\|x-z\|.\label{weakweightsemigreedy} \end{equation} We denote by ${\mathcal {GS}}(x,m,s)$ the subset of ${\mathcal G}(x,m,s)$ for which the above holds. \end{definition} \begin{remark}\rm Note that the set ${\mathcal {GS}}(x,m,s)$ depends on $\mathbf{C}$, but we will leave that implicit if is no risk of ambiguity; else, we will write ${\mathcal {GS}}(x,m,s,\mathbf{C})$. The same consideration applies to later definitions as is the case with Definition~\ref{definitionweakweightalmostsemigreedy} and Definition~\ref{definitionweakweightalmostgreedy}. \end{remark} It is known that a basis is greedy, quasi-greedy, or almost greedy if and only if the relevant definition holds for elements with finite support (see, for example, \cite[Lemma 2.3]{Oikhberg2017} for quasi-greedy bases; similar arguments hold for almost greedy or greedy ones). We will show that this is also true for weak weight-semi-greedy bases. First, we need some auxiliary lemmas. \begin{lemma}\label{lemmaGxmtfinite} Let ${\mathcal B}$ be a basis for ${\mathbb X}$ with dual basis ${\mathcal B}^*=({\bm x}_i)_{i\in \mathbb{N}}$ and $0<t\le 1$. Fix $x\in {\mathbb X}$ and $m\in \mathbb{N}$ such that $|\supp(x)|\ge m$, then ${\mathcal G}(x,m,t)$ is a finite set and ${\bm x}_i^*(x)\not=0$ for all $i\in A$ with $A\in {\mathcal G}(x,m,t)$. \end{lemma} \begin{proof} Let $$ B:=\bigcup_{A\in {\mathcal G}(x,m,t)}A\qquad\text{and}\qquad b:=\inf_{i\in B}|{\bm x}_i^*(x)|, $$ and pick any $j\in \supp(x)$. Since ${\mathcal B}^*$ is weak star null, there is $i_0\in \mathbb{N}$ such that $$ |{\bm x}_i^*(x)|<t |{\bm x}_j^*(x)|,\quad \forall i\ge i_0. $$ Hence, $B\subset \{1,\dots, i_0\}$ and ${\mathcal G}(x,m,t)$ is a finite set, so $b$ is a minimum. Now choose $A\in {\mathcal G}(x,m,t)$ and $n\in A$ with $|{\bm x}_n^*(x)|=b$. If $b=0$, choose $j\in \supp(x)\setminus A$. Then $0=b\ge t |{\bm x}_j^*(x)|>0$, a contradiction. \end{proof} Next, we strengthen \cite[Lemma 2.2]{Oikhberg2017} for Markushevich bases. \begin{lemma}\label{lemmaallfinite}Let ${\mathcal B}$ be a basis for ${\mathbb X}$, $x\in {\mathbb X}$, and $\epsilon>0$. The following hold: \begin{enumerate}[ \rm (i)] \item \label{lemmaallfinitenogreedy}Given $D\in \mathbb{N}^{<\infty}$, there is $y\in {\mathbb X}$ with finite support such that $\|x-y\|<\epsilon$ and $P_D(x)=P_D(y)$. \item \label{lemmaallfinitegreedy} Given $m_0\in \mathbb{N}$ and $0<t\le 1$, there is $y\in {\mathbb X}$ with finite support such that $\|x-y\|<\epsilon$, and, for each $1\le m\le m_0$, ${\mathcal G}(x,m,t)={\mathcal G}(y,m,t)$ and $P_A(y)=P_A(x)$ for each $A\in {\mathcal G}(x,m,t)$. \end{enumerate} \end{lemma} \begin{proof} To prove \ref{lemmaallfinitenogreedy} choose a finitely supported $z$ so that $\|x-P_D(x)-z\|< (1+\|P_D\|)^{-1}\epsilon$, and define $y:=z-P_D(z)+P_D(x)$. We have $$ \|x-y\|\le \|x-P_D(x)-z\|+\|P_D(z-x+P_D(x)\|\le (1+\|P_D\|)\|x-P_D(x)-z\|<\epsilon. $$ To prove \ref{lemmaallfinitegreedy} set ${\mathcal B}^*=({\bm x}^*_i)_{i\in \mathbb{N}}$ the dual basis of ${\mathcal B}$. We may assume that $x$ has infinite support, so the hypotheses of Lemma~\ref{lemmaGxmtfinite} hold for $m=m_0$. Let \begin{align*} B:=&\bigcup_{A\in {\mathcal G}(x,m_0,t)}A, &&b:=\min_{i\in B}|{\bm x}_i^*(x)|. \end{align*} First note that if $m_0>1$, $1\le m< m_0$ and $A\in {\mathcal G}(x,m,t)$, then $A\cup A_1\in {\mathcal G}(x,m_0,t)$ for every $A_1\in {\mathcal G}(x-P_A(x),m_0-m,1)$. Hence, $A\subset B$. \\ Since ${\mathcal B}^*$ is weak star null, there is $i_0\in \mathbb{N}$ such that $|{\bm x}_i^*(x)|< 2^{-1}tb$ for all $i\ge i_0$. Set \begin{align*} D:=&\{1,\dots,i_0\}, &&\epsilon_1:=\frac{t \min\{\epsilon, 1, b\}}{2(1+\lambda')}, \end{align*} and let $y$ be obtained by an application of \ref{lemmaallfinitenogreedy} to $x$, $D$ and $\epsilon_1$. To see that $y$ has the desired properties, first note that for each $n\in D$ and each $k\not\in D$, \begin{align} |{\bm x}_k^*(y)|\le |{\bm x}_k^*(x-y)|+|{\bm x}_k^*(x)|\le \lambda'\|x-y\|+2^{-1}tb< tb\le t|{\bm x}_{n}^*(x)|=t|{\bm x}_{n}^*(y)|.\label{lemmaallfinitegreedyoutofD} \end{align} Since $i_0>m_0$, it follows that ${\mathcal G}(y,m,t)\subset D$ for all $1\le m\le m_0$. Now fix $1\le m\le m_0$, and choose $A\in{\mathcal G}(x,m,t)$, $n\in A$ and $k\not\in A$. If $k\in D$, then $$ |{\bm x}_n^*(y)|=|{\bm x}_n^*(x)|\ge t |{\bm x}_k^*(x)|=t|{\bm x}_k^*(y)|, $$ whereas if $k\not \in D$, then $t |{\bm x}_k^*(y)|\le |{\bm x}_{n}^*(y)$ by \eqref{lemmaallfinitegreedyoutofD}. Therefore, $A\in {\mathcal G}(y,m,t)$. \\ Similarly, choose $A\in {\mathcal G}(y,m,t)$, $n\in A$ and $k\not\in A$. Since $n\in D$, the case $k\in D$ is handled as before but changing the roles of $x$ and $y$, whereas if $k\not \in D$, then $t|{\bm x}_k^*(x)|<b\le |{\bm x}_n^*(y)|=|{\bm x}_n^*(x)|$. We conclude that $A \in {\mathcal G}(x,m,t)$. \end{proof} \begin{remark}\label{remarkpbanach}Note that the proof of Lemma~\ref{lemmaGxmtfinite} holds without changes for $p$-Banach spaces, for $0<p<1$, whereas that of Lemma~\ref{lemmaallfinite} holds as well, with only straightforward modifications: just choose $z$ so that $\|x-P_D(x)-z\|^p< (1+\|P_D\|^p)^{-1}\epsilon^p$, and use $p$-convexity. Thus, for Markushevich bases, our result strengthens \cite[Lemma 7.2]{BB2020} in addition to \cite[Lemma 2.2]{Oikhberg2017}. \end{remark} \begin{remark}\label{remarkfinitesupport} Note also that the totality hypothesis does not play a role in the proofs of Lemmas~\ref{lemmaGxmtfinite} and Lemma~\ref{lemmaallfinite}. Moreover, in the proof of Lemma~\ref{lemmaallfinite}~\ref{lemmaallfinitenogreedy}, even if ${\mathcal B}^*$ is not total we can pick $z\in [{\mathcal B}]$ and obtain $y\in [{\mathcal B}]$, in other words we obtain $y$ as a finite linear combination of the ${\bm x}_i$'s. Thus, if $|\supp(x)|\ge m_0$, we can also get $y\in [{\mathcal B}]$ in Lemma~\ref{lemmaallfinite}~\ref{lemmaallfinitegreedy}. However, if $x\not\in [{\mathcal B}]$ and $|\supp(x)|<m_0$, without totality we cannot obtain $y\in [{\mathcal B}]$ in Lemma~\ref{lemmaallfinite}~\ref{lemmaallfinitegreedy}, because in that case, for every $A\in {\mathcal G}(x,m_0-1,t)$ and every $n\not\in A$, ${\bm x}_n^*(x)=0$ and $A\cup \{n\}\in {\mathcal G}(x,m_0,t)$. On the other hand, with only straightforward modifications the proofs of \cite[Lemma 7.2]{BB2020} and \cite[Lemma 2.2]{Oikhberg2017} yield $y\in [{\mathcal B}]$ even when ${\mathcal B}^*$ is not total. \end{remark} Now we can prove the aforementioned equivalence. \begin{lemma}\label{lemmafinitesupportisenough}Let ${\mathcal B}$ be a basis for ${\mathbb X}$, $\mathbf{w}$ a weight, $\mathbf{C}>0$ and $0<s\le 1$. Suppose that the conditions of Definition~\ref{definitionweakweightsemigreedy} hold for $x$ with finite support. Then ${\mathcal B}$ is $\mathbf{C}$-$s$-$\mathbf{w}$-semi-greedy. Moreover, the conditions of Definition~\ref{definitionweakweightsemigreedy} hold even if the infimum is taken without the restriction $|\supp(z)|<\infty$. \end{lemma} \begin{proof} Pick $x\in {\mathbb X}\setminus \{0\}$ and $m\in \mathbb{N}$. If $|\supp(x)|<m$, then $\supp(x) \subset A$ for every $A \in {\mathcal G}(x,m,t)$, and there is nothing to prove. Otherwise, for every $n\in \mathbb{N}$, by Lemma~\ref{lemmaallfinite} there is $x_n\in {\mathbb X}$ with finite support such that $\|x-x_n\|\le n^{-1}$ and ${\mathcal G}(x_n,m,t)={\mathcal G}(x,m,t)$. By hypothesis, for each $n$ there are $A_n\in {\mathcal G}(x,m,t)$ and $y_n\in {\mathbb X}$ with $\supp(y_n)\subset A_n$ such that $$ \|x_n-y_n\|\le \mathbf{C}\inf_{\substack{z\in {\mathbb X}\\|\supp(z)|<\infty\\mathbf{w}(\supp(z))\le w(A_n)}}\|x_n-z\|. $$ By Lemma~\ref{lemmaGxmtfinite}, ${\mathcal G}(x,m,t)$ is finite. Thus, passing to a subsequence, we may assume $A_n=A$, a fixed set. Passing to a further subsequence, we may also assume that there is $y$ supported in $A$ such that $\|y_n-y\|\le n^{-1}$ for all $n$. Now pick $z\in {\mathbb X}$ with $w(\supp(z))\le w(A)$. By Remark~\ref{remarkprojectionc0}, for each $n\in \mathbb{N}$ there is $z_n\in {\mathbb X}$ with finite support contained in $\supp(z)$ such that $\|z_n-z\|\le n^{-1}$. Hence, \begin{align*} \|x-y\|\le&2n^{-1}+\|x_n-y_n\|\le 2n^{-1}+\mathbf{C}\|x_n-z_n\|\le \mathbf{C}\|x-z\|+2n^{-1}(1+\mathbf{C}), \forall n\in \mathbb{N}. \end{align*} As this holds for every $n\in \mathbb{N}$, we get $\|x-y\|\le \mathbf{C}\|x-z\|$. Now the proof is completed by taking infimum over all such $z$. \end{proof} Our next result collects some general facts about weak $\mathbf{w}$-semi-greedy, $\mathbf{w}$-democratic and $\mathbf{w}$-superdemocratic bases. In particular, we give upper bounds for the norms of vectors of the form $\|{\mathbf 1}_{\varepsilon,A}\|$ in terms of the $\mathbf{w}$-measure of $A$, and for the unconditionality parameter ${\mathbf k}_m={\mathbf k}_m[{\mathcal B},{\mathbb X}]$ (also known as \emph{conditionality constant} or parameter), defined by $$ {\mathbf k}_m:=\sup_{\substack{A\subset \mathbb{N}\\|A|\le m}}\|P_A\|, $$ which is used to measure how far a basis is from being unconditional (see for example \cite{AAB2021}, \cite{AABBL2021}, \cite{AAGHR2015}, \cite{AAW2020}, \cite{BBG2017}, \cite{BBGHO2018} and \cite{DKO2015}). Below, we appeal to the constants $\lambda$ and $\lambda'$ of \eqref{constants}. \begin{proposition}\label{propositionupperbounds}Let ${\mathcal B}=({\bm x}_i)_{i\in \mathbb{N}}$ be a $\mathbf{C}$-$s$-$\mathbf{w}$-semi-greedy basis for ${\mathbb X}$, $\mathbf{w}=(w_i)_{i\in \mathbb{N}}$ a weight, $\mathbf{C}>0$ and $0<s\le 1$, and let $$ \mathbf{C}_1:=3\mathbf{C} s^{-1}(1+\lambda \lambda ')\lambda \max\{2\inf_{j\in \mathbb{N}} w_j^{-1},1\}. $$ The following hold: \begin{enumerate}[\upshape (i)] \item \label{boundwithweight} For every finite set $A\subset \mathbb{N}$ and $\varepsilon \in \ensuremath{\mathcal{E}}_{A}$, $$ \|{\mathbf 1}_{\varepsilon,A}\|\le \mathbf{C}_1\max\{w(A),1\}. $$ \item \label{functionals} For every $x^*\in S_{{\mathbb X}^*}$, $$ w(\{j\in \mathbb{N}: |x^*({\bm x}_j)|>\mathbf{C}_1 w_j\})\le 1. $$ \item \label{condparameter} For each $m\in \mathbb{N}$, $$ {\mathbf k}_m\le \mathbf{C}_1\lambda '\max\{\sup_{\substack{A\subset \mathbb{N}\\|A|\le m}}w(A),1\}. $$ \end{enumerate} If ${\mathcal B}$ is $\mathbf{K}$-$\mathbf{w}$-disjoint superdemocratic (in particular, if it is $\mathbf{K}$-$\mathbf{w}$-almost greedy), the above results hold if we replace $\mathbf{C}_1$ with $$ \mathbf{K}_1:=4\mathbf{K}\lambda \max\{2\inf_{j\in \mathbb{N}}w_j^{-1},1\}. $$ If ${\mathcal B}$ is $\mathbf{K}$-$\mathbf{w}$-disjoint democratic, then the above holds with $\mathbf{C}_1$ replaced by $2\kappa \mathbf{K}_1$. \end{proposition} \begin{proof} To prove \ref{boundwithweight}, fix $0<\epsilon<1$, and choose $j_0\in \mathbb{N}$ so that $$ w_{j_0}^{-1}\le \inf_{j\in \mathbb{N}}w_j^{-1}+\epsilon. $$ Given $A$ and $\varepsilon$ as in the statement, define the possibly empty set $$ A_1:=\{i\in A: w_i<2^{-1}w_{j_0} \}. $$ We first estimate the norm $\|{\mathbf 1}_{\varepsilon,A\setminus A_1}\|$: \begin{align} \|{\mathbf 1}_{\varepsilon,A\setminus A_1}\|\le& \lambda |A\setminus A_1|\le 2\lambda w_{j_0}^{-1}w(A)\le 2\lambda (1+\epsilon)\max\{\inf_{j\in \mathbb{N}}w_j^{-1},1\}\max\{w(A),1\}\nonumber\\ \le& \frac{1}{3}(1+\epsilon)\mathbf{C}_1\max\{w(A),1\}. \label{heavierelements} \end{align} If $A_1=\emptyset$, as $\epsilon$ is arbitrary there is nothing else to prove. Else, to estimate the norm $\|{\mathbf 1}_{\varepsilon,A_1}\|$ choose a partition of $A_1$ as follows: First, pick a set $A_{1,1}\subset A_1$ of maximum cardinality such that $w(A_{1,1})\le w_{j_0}$. If $A_{1,1}\not=A_1$, then pick $A_{1,2}\subset A_1\setminus A_{1,1}$ of maximum cardinality such that $w(A_{1,2})\le w_{j_0}$, and so on. By this procedure, we get a partition of $A_1$ into finitely many sets $\{A_{1,k}\}_{1\le k\le k_1}$ with $w(A_{1,k})\le w_{j_0}$ for all $1\le k\le k_1$. If $k_1>1$, then by construction, for every $1\le k\le k_1-1$, there is $i\in A_1\setminus A_{1,k}$ such that $w(A_{1,k})+w_i>w_{j_0}$, which implies that $w(A_{1,k})>2^{-1}w_{j_0}$. Thus, $$ w(A_1)> \sum\limits_{k=1}^{k_1-1}w(A_{1,k})\ge 2^{-1}(k_1-1)w_{j_0}, $$ so \begin{equation} k_1\le 2w(A_1)w_{j_0}^{-1}+1\le 2w(A)w_{j_0}^{-1}+1.\label{upperboundk1} \end{equation} For each $1\le k\le k_1$, define $$ z_k:=(1+\epsilon)s^{-1}{\bm x}_{j_0}+{\mathbf 1}_{\varepsilon,A_{1,k}}. $$ As ${\mathcal G}(z_k,1,s)=\{\{j_0\}\}$, there is $b_k\in \mathbb{K}$ such that $$ \|z_k-b_k{\bm x}_{j_0}\|\le \mathbf{C}\inf_{\substack{|\supp(z)|<\infty\\mathbf{w}(\supp(z))\le w_{j_0}}}\|z_k-z\|\le \mathbf{C}(1+\epsilon)s^{-1}\|{\bm x}_{j_0}\|. $$ Hence, by the triangle inequality, $$ \|{\mathbf 1}_{\varepsilon,A_{1,k}}\|\le \|z_k-b_k{\bm x}_{j_0}\|+\|((1+\epsilon)s^{-1}-b_k){\bm x}_{j_0}\|\le (1+\epsilon)s^{-1}\mathbf{C}\|{\bm x}_{j_0}\|+ |(1+\epsilon)s^{-1}-b_k|\|{\bm x}_{j_0}\|. $$ Let ${\mathcal B}^*=({\bm x}^*_i)_{i\in \mathbb{N}}$ be the dual basis of ${\mathcal B}$. Since $$ |(1+\epsilon)s^{-1}-b_k|=|{\bm x}_{j_0}^{*}(z_k-b_k{\bm x}_{j_0})|\le \|{\bm x}_{j_0}^{*}\|\|z_k-b_k{\bm x}_{j_0}\|\le \mathbf{C}\|{\bm x}_{j_0}^{*}\|(1+\epsilon)s^{-1}\|{\bm x}_{j_0}\| $$ we obtain $$ \|{\mathbf 1}_{\varepsilon,A_{1,k}}\|\le \mathbf{C}(1+\epsilon)s^{-1}(1+\|{\bm x}_{j_0}\|\|{\bm x}_{j_0}^{*}\|)\|{\bm x}_{j_0}\|. $$ Using again the triangle inequality and \eqref{upperboundk1}, we get \begin{align*} \|{\mathbf 1}_{\varepsilon,A_1}\|\le& (2w(A)w_{j_0}^{-1}+1) \mathbf{C}(1+\epsilon)s^{-1}(1+\|{\bm x}_{j_0}\|\|{\bm x}_{j_0}^{*}\|)\|{\bm x}_{j_0}\|\\ \le& 2\mathbf{C}(1+\epsilon)s^{-1}(1+\lambda \lambda ')\lambda \max\{2w_{j_0}^{-1}w(A),1\}\\ =&\frac{2}{3} \mathbf{C}_1 (1+\epsilon)(\max\{2\inf_{j\in \mathbb{N}} w_j^{-1},1\})^{-1}\max\{2w_{j_0}^{-1}w(A),1\}\\ \le& \frac{2}{3}\mathbf{C}_1(1+\epsilon)\max\{w(A),1\}. \end{align*} Given that $\epsilon$ is arbitrary, the proof of \ref{boundwithweight} is completed combining the above inequality with \eqref{heavierelements}. \\ Now suppose \ref{functionals} is false, and choose $x^*\in S_{{\mathbb X}^*}$ for which the result does not hold. Then, there is $A\subset \mathbb{N}$ finite with $w(A)>1$ such that $$ |x^*({\bm x}_j)|>\mathbf{C}_1 w_j,\qquad\forall j\in A. $$ Define $\varepsilon\in \ensuremath{\mathcal{E}}_A$ by $$ \varepsilon_j:=\frac{|x^*({\bm x}_j)|}{x^*({\bm x}_j)}, \qquad \forall j\in A. $$ As $w(A)>1$, using \ref{boundwithweight} we get $$ \mathbf{C}_1w(A)\ge \|{\mathbf 1}_{\varepsilon,A}\|\ge |x^*({\mathbf 1}_{\varepsilon,A})|=\sum_{j\in A}|x^*({\bm x}_j)|>\sum_{j\in A}\mathbf{C}_1 w_j=\mathbf{C}_1 w(A), $$ a contradiction.\\ To prove \ref{condparameter}, fix $x\in {\mathbb X}$, $m\in \mathbb{N}$, and $A\subset \mathbb{N}$ with $|A|\le m$. By \ref{boundwithweight}, \begin{align*} \|P_A(x)\|\le& \|x\|_{\infty}\max_{\varepsilon\in \ensuremath{\mathcal{E}}_A}\|{\mathbf 1}_{\varepsilon,A}\|\le \lambda '\mathbf{C}_1\max\{w(A),1\}\|x\|. \end{align*} so the proof is completed by taking supremum.\\ Now suppose that ${\mathcal B}$ is $\mathbf{K}$-$\mathbf{w}$-disjoint superdemocratic. Then all of the steps of the above proof hold with the only modification consisting in replacing $\mathbf{C}_1$ with $\mathbf{K}_1$, except for the bounds for $\|{\mathbf 1}_{\varepsilon, A\setminus A_{1}}\|$ and $\|{\mathbf 1}_{\varepsilon, A_{1}}\|$; we give bounds for these norms as follows: First, $$ \|{\mathbf 1}_{\varepsilon, A_{1,k}}\|\le \mathbf{K}\|{\bm x}_{j_0}\|\le \mathbf{K}\lambda, \qquad \forall 1\le k\le k_1. $$ Thus, using \eqref{upperboundk1}, $$ \|{\mathbf 1}_{\varepsilon, A_{1}}\|\le k_1\mathbf{K}\lambda \le \mathbf{K}\lambda (1+2w(A)w_{j_0}^{-1})\le \frac{1}{2}\mathbf{K}_1(1+2\epsilon)\max\{w(A),1\}. $$ On the other hand, arguing as in the proof of \eqref{heavierelements} we obtain $$ \|{\mathbf 1}_{\varepsilon,A\setminus A_1}\|\le 2\lambda (1+\epsilon)\max\{1,\inf_{j\in \mathbb{N}}w_j^{-1}\}\max\{w(A),1\}\le \frac{1}{2}(1+\epsilon)\mathbf{K}_1\max\{w(A),1\}, $$ and the result follows by the above inequalities. \\ If ${\mathcal B}$ is $\mathbf{K}$-$\mathbf{w}$-disjoint almost greedy, it is $\mathbf{K}$-$\mathbf{w}$-disjoint superdemocratic by Lemma~\ref{lemmadisjoint=} and \cite[Theorem 1.5]{B2020}. \\ Finally, if ${\mathcal B}$ is $\mathbf{K}$-$\mathbf{w}$-disjoint democratic, by convexity we have $$ \|{\mathbf 1}_{\varepsilon, A_{1,k}}\|\le 2\kappa \sup_{B\subset A_{1,k}}\|{\mathbf 1}_{B}\| \le 2\kappa \mathbf{K}\|{\bm x}_{j_0}\|\le 2\kappa \mathbf{K}\lambda, \qquad 1\le k\le k_1, $$ and the rest of the proof is the same as that of the $\mathbf{K}$-$\mathbf{w}$-disjoint superdemocratic case. \end{proof} \begin{remark}\label{remarkquantitativec0}\rm It is known that if ${\mathcal B}$-is $\mathbf{w}$-semi-greedy or $\mathbf{w}$-disjoint superdemocratic and $\mathbf{w}\in \ell_1$, ${\mathcal B}$ is equivalent to the canonical unit vector basis of $\mathtt{c}_{0}$ (see \cite[Proposition 2.3]{B2020}, \cite[Proposition 3.10]{BDKOW2019}, \cite[Proposition 4.5]{DKTW2018}), whereas the result for $\mathbf{w}$-disjoint democratic bases can be obtained via a straightforward modification of the proof of \cite[Proposition 3.10]{BDKOW2019}. Using Proposition~\ref{propositionupperbounds} we can obtain some quantitative variants of these results, as well as a similar result for weak weight-semi-greedy bases. More precisely, if $x$ is finitely supported, then \begin{align} \|x\|\le& \|x\|_{\infty}\max_{\varepsilon\in \ensuremath{\mathcal{E}}_{\supp(x)}}\|{\mathbf 1}_{\varepsilon,\supp(x)}\|\le \lambda '\mathbf{C}_2\max\{1,\|\mathbf{w}\|_{1}\}\|x\|\label{newupperbound} \end{align} where, using the notation of Proposition~\ref{propositionupperbounds}, \begin{equation*} \mathbf{C}_2=\begin{cases} \mathbf{C}_1 & \text{ if ${\mathcal B}$ is $\mathbf{C}$-$s$-$\mathbf{w}$-semi-greedy};\\ \mathbf{K}_1 & \text{ if ${\mathcal B}$ is $\mathbf{K}$-$\mathbf{w}$-disjoint-superdemocratic};\\ 2\kappa \mathbf{K}_1 & \text{ if ${\mathcal B}$ is $\mathbf{K}$-$\mathbf{w}$-disjoint-democratic}. \end{cases} \end{equation*} Since the set of finitely supported elements is dense in ${\mathbb X}$, the bound on the right-hand side of \eqref{newupperbound} holds for any $x\in {\mathbb X}$. \end{remark} \begin{remark}\label{remarkl1wdem}\rm It follows from Remark~\ref{remarkquantitativec0} that Lemma~\ref{lemmal1thenc0} and Remark~\ref{remarkprojectionc0} also hold if we replace the $\mathbf{w}$-semi-greedy property by $\mathbf{w}$-democracy. \end{remark} Next, we prove that when $\mathbf{w}\not \in \mathtt{c}_0$, every $\mathbf{w}$-semi-greedy basis is $\mathbf{w}$-almost greedy. While $\mathbf{w}$-almost greediness entails $\mathbf{w}$-superdemocracy (see \cite{DKTW2018}, \cite{B2020}), we give a direct proof of this result, as the upper bounds for the $\mathbf{w}$-superdemocracy constant might be of interest as well. We also give an upper bound for the truncation quasi-greedy constant. First we prove an auxiliary result which will allow us to pick adequate $s$-greedy sets. \begin{lemma}\label{lemmasgset}Let ${\mathcal B}$ be a $\mathbf{C}$-$s$-$\mathbf{w}$-semi-greedy basis for ${\mathbb X}$, $\mathbf{w}$ a weight, $\mathbf{C}>0$ and $0<s\le 1$, and let ${\mathcal B}^*=({\bm x}^*_i)_{i\in \mathbb{N}}$ be the dual basis of ${\mathcal B}$. For every $x\in {\mathbb X}$ and every nonempty finite set $A\subset \mathbb{N}$, there is $m\in \mathbb{N}$ and $B\in {\mathcal {GS}}(x,m,s)$ such that $A\subset B$ and for all $j\in B$, $$ |{\bm x}_j^*(x)|\ge s^{2} \min_{i\in A}|{\bm x}_i^*(x)|. $$ \end{lemma} \begin{proof} Let $c:=\min_{i\in A}|{\bm x}_i^*(x)|$. Clearly we may assume $c>0$. Since ${\mathcal B}^*$ is weak star null, there is $n_0\in \mathbb{N}$ such that for each $n\ge n_0$, every set in ${\mathcal G}(x,n,s)$ contains $A$. Let $$ n_1:=\min\{{n\in \mathbb{N}}: \exists B\in {\mathcal {GS}}(x,n,s): B\supset A\}, $$ and choose $B \in {\mathcal {GS}}(x,n_1,s)$ containing $A$. If $B=A$, there is nothing to prove. Otherwise, since $n_1>1$, we can choose $D \in {\mathcal {GS}}(x,n_1-1,s)$. By the minimality of $n_1$, it follows that $$ A\not\subset D. $$ Hence, for all $j\in D$, \begin{equation} |{\bm x}_j^*(x)|\ge s c.\label{boundforthesmallerset} \end{equation} Thus, if there exists $j_0\in D\setminus B$, it follows that for all $j\in B$, \begin{equation} |{\bm x}_j^*(x)|\ge s |{\bm x}_{j_0}^*(x)|\ge s^{2} c.\nonumber \end{equation} On the other hand, if $D\subseteq B$, given that $ A\not\subset D$ and $A\subseteq B$, there is $i_1\in A$ such that $$ B=D\cup \{ i_1\}, $$ which implies that \eqref{boundforthesmallerset} also holds for all $j\in B$. \end{proof} Before going on, we note that the weak $\mathbf{w}$-semi-greedy and $\mathbf{w}$-disjoint and $\mathbf{w}$-super-democracy properties can be extended from the context of Markushevich bases to that of general minimal systems. Some minor and straightforward modifications are necessary to account for the fact that if the system is not a Markushevich basis, having finite support is not the same as being a finite linear combination of the elements of the system. In the more general context, the proofs of Proposition~\ref{propositionupperbounds} and Lemma~\ref{lemmasgset} are valid as well. However, we continue working with Markushevich bases as the hypothesis of being a total system is used in the remaining proofs of this section, in particular to guarantee that we can appeal to the separation properties studied in Section~\ref{sectionseparation}.\\ Next, we prove our first case of the implication from weak weight-semi-greedy to weight-almost greedy bases. We will use the following notation: Given finite sets $A,B\subset \mathbb{N}$, we write $A<B$ to mean that $\max(A)<\min(B)$, and for $j\in \mathbb{N}$, we write $j<A$ to mean that $j<\min(A)$. We use similar conventions for ``$>$'', ``$\ge$'' and ``$\le$''. \begin{theorem}\label{theoremswemigreedydisjointgreedynotc0} Let ${\mathcal B}=({\bm x}_i)_{i\in \mathbb{N}}$ be a $\mathbf{C}$-$s$-$\mathbf{w}$-semi-greedy basis for ${\mathbb X}$, $\mathbf{w}=(w_i)_{i\in \mathbb{N}}$ a weight, $\mathbf{C}>0$ and $0<s\le 1$. Suppose that $\mathbf{w}$ has a subsequence $(w_{i_k})_{k\in \mathbb{N}}$ that is bounded below, that is $$ \inf_{k\in \mathbb{N}}w_{i_k}>0. $$ Let $\mathbf{M}:=\mathbf{M}_{fs}(({\bm x}_{i_k})_{k\in \mathbb{N}}, {\mathbb X})$ and let ${\mathcal B}^*=({\bm x}^*_i)_{i\in \mathbb{N}}$ be the dual basis of ${\mathcal B}$. The following hold: \begin{enumerate}[\upshape (i)] \item \label{wdisjointgreedyest1} For every $x\in {\mathbb X}$, $m\in \mathbb{N}$, $0<t\le 1$, $A\in {\mathcal G}(x,m,t)$, and $y$ with $w(\supp(y))\le w(A)$ and $\supp(y)\cap A=\emptyset$, $$ \|x-P_A(x)\|\le \mathbf{C}\mathbf{M} \max\{1+8 t^{-1}s^{-1} \lambda \lambda ',1+6\mathbf{C} t^{-1}s^{-3} \} \|x-y\|. $$ Thus, ${\mathcal B}$ is $\mathbf{w}$-almost greedy with constant as above taking $t=1$. \item \label{wdisjointgreedyest2a} For every $x\in {\mathbb X}$, every $A\in \mathbb{N}^{<\infty}$ and every $\varepsilon\in \ensuremath{\mathcal{E}}_A$, if $$ w(A) \le w(\{i\in \mathbb{N}: |{\bm x}_i^*(x)|\ge 1\}) $$ then $$ \|{\mathbf 1}_{\varepsilon,A}\|\le 2s^{-1}\mathbf{C}\mathbf{M}\max\{\lambda \lambda ',2s^{-2}\mathbf{C} \} \|x\|. $$ Thus, ${\mathcal B}$ is $\mathbf{w}$-superdemocratic and truncation quasi-greedy, in each case with constant as above. \end{enumerate} \end{theorem} \begin{proof} To prove \ref{wdisjointgreedyest1}, choose $0<\epsilon<1$ and let $({\bm x}_{i_{k_j}})_{j\in \mathbb{N}}$ be a subsequence given by an application of Corollary~\ref{corollaryboth} to $({\bm x}_{i_k})_{k\in \mathbb{N}}$ and $({\mathbb X}, \mathbf{M}, \epsilon)$. Fix $x$, $m$, $t$ and $A$ as in the statement, and $y$ such that $w(\supp(y))\le w(A)$ and $\supp(y)\cap A=\emptyset$. We assume first that both $x$ and $y$ have finite support, and we may also assume $x\not=P_A(x)$, so $$ a:=\min_{i\in A}|{\bm x}_i^*(x)| $$ is positive. Pick $i_0>\supp(x)\cup \supp(y)$, and set $ {\mathbb F}:=[ {\bm x}_i: 1\le i\le i_0]$. We will consider two cases: \\ \paragraph{Case 1} Suppose that there is a set $E\subset \{i_{k_j}\}_{j\ge s_{{\mathbb F},\epsilon}}$ such that $|E|\le 8$ and $2w(A)\le w(E)$. Define $$ z_1:=x-P_A(x)+a(1+\epsilon) t^{-1}s^{-1}{\mathbf 1}_{E }. $$ Notice that $|{\bm x}_i^*(z_1)|=|{\bm x}_i^*(x)|\le t^{-1}a$ for all $i\not \in E$, so ${\mathcal G}(z_1,|E|,s)=\{E\}$. Hence, there is $z_2\in {\mathbb X}$ with $\supp(z_2) \subset E$ such that $$ \|z_1-z_2\|\le \mathbf{C} \inf_{\substack{|\supp(z)|<\infty\\mathbf{w}(\supp(z))\le w(E)}}\|z_1-z\|. $$ Given that $w(\supp(y))+w(A)\le w(E)$, we have \begin{align*} \|z_1-z_2\|\le& \mathbf{C}\|z_1+P_A(x)-y\|\le \mathbf{C}\|x-y\|+\mathbf{C} a(1+\epsilon)t^{-1}s^{-1}\|{\mathbf 1}_E\|\\ \le& \mathbf{C}\|x-y\|+8\mathbf{C} a(1+\epsilon)t^{-1}s^{-1} \lambda . \end{align*} Pick any $i\in A$. Since $A\cap \supp(y)=\emptyset$, we have $$ a\le |{\bm x}_i^*(x-y)|\le \lambda '\|x-y\|. $$ Thus, $$ \|z_1-z_2\|\le (\mathbf{C}+8\mathbf{C}(1+\epsilon)t^{-1}s^{-1} \lambda \lambda ')\|x-y\|. $$ Given that $$ \|x-P_A(x)\|\le (\mathbf{M}+\epsilon)\|z_1-z_2\|, $$ it follows that \begin{equation} \|x-P_A(x)\|\le \mathbf{C} (\mathbf{M}+\epsilon) (1+8(1+\epsilon) t^{-1}s^{-1} \lambda \lambda ')\|x-y\|.\label{lightsgreedyset} \end{equation} \paragraph{Case 2} Suppose there is ${j_0}>s_{{\mathbb F},\epsilon}$ such that \begin{equation} w(A)>4w_{i_{k_{j}}},\qquad \forall j\ge {j_0}.\label{heavysgreedyset1} \end{equation} Choose $l_1, l_2\ge 3$ so that \begin{align*} w\left(\{i_{k_{2({j_0}+d)-1}},i_{k_{2({j_0}+d)}}: 1\le d< l_1\}\right)\le& w(A)\le w\left(\{i_{k_{2({j_0}+d)-1}},i_{k_{2({j_0}+d)}}: 1\le d\le l_1\}\right);\\ w\left(\{i_{k_{2({j_0}+l_1+d)-1}},i_{k_{2({j_0}+l_1+d)}}: 1\le d< l_2\}\right)\le& w(A)\le w\left(\{i_{k_{2({j_0}+l_1+d)-1}},i_{k_{2({j_0}+l_1+d)}}: 1\le d\le l_2\}\right). \end{align*} Set \begin{align*} &E_{1,1}:=\{i_{k_{2({j_0}+d)-1}}: 1\le d<l_1\}; &&E_{1,2}:=\{i_{k_{2({j_0}+d)}}: 1\le d<l_1\};\\ &E_{2,1}:=\{i_{k_{2({j_0}+l_1+d)-1}}: 1\le d<l_2\}; &&E_{2,2}:=\{i_{k_{2({j_0}+l_1+d)}}: 1\le d<l_2\};\\ &E_{3,1}:=\{i_{k_{2({j_0}+l_1)-1}},i_{k_{2({j_0}+l_1+l_2)-1}}\}; &&E_{3,2}:=\{i_{k_{2({j_0}+l_1)}},i_{k_{2({j_0}+l_1+l_2)}}\};\\ &E:=\bigcup_{b=1}^{3}\bigcup_{d=1}^{2}E_{b,d}. \end{align*} It follows from our choices and \eqref{heavysgreedyset1} that \begin{align} \max_{1\le b\le 3}w(E_{b,1}\cup E_{b,2})\le w(A)\le\frac{w(E)}{2}. \label{weightcloseabovebelow} \end{align} Now define $$ z_3:=x-P_A(x)-as^{-1}t^{-1}(1+\epsilon)(\sum_{b=1}^{3}{\mathbf 1}_{E_{b,1}} -{\mathbf 1}_{E_{b,2}}). $$ As before, we have $|{\bm x}_i^*(z_3)|=|{\bm x}_i^*(x-P_A(x))|\le a t^{-1}$ for all $i\not\in E$, so ${\mathcal G}(z_3,|E|,s)=\{E\}$. Hence, there is $z_4$ with $\supp(z_4)\subset E$ such that $$ \|z_3-z_4\|\le \mathbf{C} \inf_{\substack{z\in {\mathbb X}\\mathbf{w}(\supp(z))\le w(E)\\|\supp(z)|<\infty}}\|z_3-z\|. $$ Since $w(\supp(y))+w(A)\le 2w(A) \le w(E)$, we have \begin{align} \|x-P_A(x)\|\le& (\mathbf{M}+\epsilon)\|z_3-z_4\|\le \mathbf{C}(\mathbf{M}+\epsilon) \|z_3+P_A(x)-y\|\nonumber\\ \le& \mathbf{C}(\mathbf{M}+\epsilon)\|x-y\|+ \mathbf{C}(\mathbf{M}+\epsilon)as^{-1}t^{-1}(1+\epsilon)\|\sum_{b=1}^{3}{\mathbf 1}_{E_{b,1}} -{\mathbf 1}_{E_{b,2}}\|.\label{leftsidebound} \end{align} To estimate $\|\sum_{b=1}^{3}{\mathbf 1}_{E_{b,1}} -{\mathbf 1}_{E_{b,2}}\|$, set $$ z_5:=x-y+a s^2(1-\epsilon)({\mathbf 1}_{E_{1,1}}-{\mathbf 1}_{E_{1,2}}). $$ By Lemma~\ref{lemmasgset} there is a set $D\supset A$ with $D\in {\mathcal {GS}}(z_5,|D|,s)$ such that $$ \min_{j\in D}|{\bm x}_j^*(z_5)|\ge s^2\min_{i\in A}|{\bm x}_i^*(z_5)|=s^2\min_{i\in A}|{\bm x}_i^*(x)|=s^2a, $$ which implies that $D\subset \{1,\dots,i_0\}$. Choose $z_6$ with $\supp(z_6)\subset D$ so that \begin{equation} \|z_5-z_6\|\le \mathbf{C} \inf_{\substack{z\in {\mathbb X}\\mathbf{w}(\supp(z))\le w(D)\\|\supp(z)|<\infty}}\|z_5-z\|. \nonumbe \end{equation} Given that $z_6\in {\mathbb F}$, using \eqref{weightcloseabovebelow} and the fact that $w(A)\le w(D)$ we infer that \begin{align*} \|a s^2(1-\epsilon)({\mathbf 1}_{E_{1,1}}-{\mathbf 1}_{E_{1,2}})\|\le& \|z_5-z_6\|+\|x-y-z_6\|\le (2+\epsilon)\|z_5-z_6\|\\ \le& (2+\epsilon)\mathbf{C}\|z_5-a s^2(1-\epsilon)({\mathbf 1}_{E_{1,1}}-{\mathbf 1}_{E_{1,2}})\|\\ =& (2+\epsilon)\mathbf{C}\|x-y\|. \end{align*} The same argument gives $$ \|a s^2(1-\epsilon)({\mathbf 1}_{E_{b,1}}-{\mathbf 1}_{E_{b,2}})\|\le (2+\epsilon)\mathbf{C}\|x-y\,|\qquad\forall 2\le b\le 3. $$ Combining these inequalities with \eqref{leftsidebound}, by the triangle inequality we get \begin{equation} \|x-P_A(x)\|\le \mathbf{C}(\mathbf{M}+\epsilon)(1+3(2+\epsilon)\mathbf{C}(1+\epsilon)(1-\epsilon)^{-1}t^{-1}s^{-3})\|x-y\|. \label{heavygreedyset2} \end{equation} As $\epsilon$ is arbitrary, a combination of \eqref{lightsgreedyset} and \eqref{heavygreedyset2} gives \ref{wdisjointgreedyest1} for $x, y\in {\mathbb X}$ with finite support. Now choose again $x,m, t, A$ and $y$ as in the statement, $x\not=P_A(x)$, fix $\delta>0$, and let $\mathbf{K}:=\mathbf{C}\mathbf{M} \max\{1+8 t^{-1}s^{-1} \lambda \lambda ',1+6\mathbf{C} t^{-1}s^{-3} \}$. By Lemma~\ref{lemmaallfinite}, there is $x_1\in {\mathbb X}$ with finite support such that $A\in {\mathcal G}(x_1,m,t)$, $P_A(x_1)=P_A(x)$, and $\|x-x_1\|\le \delta$. Also, Remark~\ref{remarkprojectionc0} gives $y_1\in {\mathbb X}$ with finite support such that $\supp(y_1)\subset \supp(y)$ and $\|y-y_1\|\le \delta$. Applying the result for vectors with finite support, we obtain \begin{align*} \|x-P_A(x)\|\le&\|x_1-x\|+\|x_1-P_A(x_1)\|\le \delta+\mathbf{K}\|x_1-y_1\|\\ \le& \delta+\mathbf{K}\|x-y\|+\mathbf{K}\|x_1-x\|+\mathbf{K}\|y-y_1\|\le 3\delta+\mathbf{K}\|x-y\|. \end{align*} Since $\delta$ is arbitrary, the proof of \ref{wdisjointgreedyest1} is complete. \\ To prove \ref{wdisjointgreedyest2a}, choose $0<\epsilon<1$, let $({\bm x}_{i_{k_j}})_{j\in \mathbb{N}}$ be as in the proof of \ref{wdisjointgreedyest1}, and let $A$, $\varepsilon$ and $x$ be as in the statement, and suppose $x$ has finite support. Choose $i_0>A\cup \supp(x)$, and set $$ {\mathbb F}:=[{\bm x}_i: 1\le i\le i_0]. $$ We will consider again two cases: \\ \paragraph{Case 1} Suppose there are $j_2>j_1> s_{{\mathbb F},\epsilon}$ such that $$ w(A)\le w_{i_{k_{j_1}}}+w_{i_{k_{j_2}}}. $$ Set $$ E:=\{i_{k_{j_1}}, i_{k_{j_2}}\} \qquad \text{and} \qquad z_1:={\mathbf 1}_{\varepsilon, A} +(1+\epsilon) s^{-1}{\mathbf 1}_{E }. $$ Since ${\mathcal G}(z_1,2,s)=\{E\}$, there is $z_2\in \mathbb{N}$ with $\supp(z_2)\subset E$ such that $$ \|z_1-z_2\|\le \mathbf{C} \inf_{\substack{|\supp(z)|<\infty\\mathbf{w}(\supp(z))\le w(E)}}\|z_1-z\|. $$ Hence, \begin{equation} \|{\mathbf 1}_{\varepsilon,A}\|\le (\mathbf{M}+\epsilon)\|z_1-z_2\|\le (\mathbf{M}+\epsilon) \mathbf{C} (1+\epsilon) s^{-1}\|{\mathbf 1}_{E}\|.\nonumbe \end{equation} Given that $$ \|{\mathbf 1}_E\|\le 2\lambda \le 2\lambda \lambda '\|x\|, $$ we obtain \begin{equation} \|{\mathbf 1}_{\varepsilon,A}\|\le 2(\mathbf{M}+\epsilon)\mathbf{C}(1+\epsilon)s^{-1}\lambda \lambda '\|x\|. \label{lightsupport7} \end{equation} \paragraph{Case 2} Suppose that there is ${j_0}>s_{{\mathbb F},\epsilon}$ such that $$ w(A)\ge 2 w(i_{k_{j}}),\qquad \forall j\ge {j_0}, $$ and choose $l_1\ge 2$ so that $$ w\left(\{i_{k_{2({j_0}+d)-1}},i_{k_{2({j_0}+d)}}: 1\le d< l_1\}\right)\le w(A)\le w\left(\{i_{k_{2({j_0}+d)-1}},i_{k_{2({j_0}+d)}}: 1\le d\le l_1\right\}). $$ Define \begin{align*} &E_{1,1}:=\{i_{k_{2({j_0}+d)-1}}: 1\le d<l_1\}; &&E_{1,2}:=\{i_{k_{2({j_0}+d)}}: 1\le d<l_1\};\\ &E_{2,1}:=\{i_{k_{2({j_0}+l_1)-1}} \}; &&E_{2,2}:=\{i_{k_{2({j_0}+l_1)}}\};\\ &E:=\bigcup_{b=1}^{2}\bigcup_{d=1}^{2}E_{b,d}. \end{align*} Note that $\max\{w(E_{1,1}\cup E_{1,2}),w(E_{2,1}\cup E_{2,2})\}\le w(A)\le w(E)$. Let $$ B:=\{i\in \mathbb{N}: |{\bm x}_i^*(x)|\ge 1\}\qquad\text{and}\qquad z_1:=x + s^2(1-\epsilon)({\mathbf 1}_{E_{1,1}}-{\mathbf 1}_{E_{1,2}}). $$ By Lemma~\ref{lemmasgset}, there is a set $D\supset B$ with $D\in {\mathcal {GS}}(z_1,|D|,s)$ and a vector $z_2$ with$\supp(z_2)\subset D$ such that \begin{equation} \min_{j\in D}|{\bm x}_j^*(x)|\ge s^2,\nonumber \end{equation} and \begin{equation} \|z_1-z_2\|\le \mathbf{C} \inf_{\substack{z\in {\mathbb X}\\mathbf{w}(\supp(z))\le w(D)\\|\supp(z)|<\infty}}\|z_1-z\|. \nonumbe \end{equation} Since $D\subset\{1,\dots,i_0\}$, considering that $w(D)\ge w(B)\ge w(A)\ge w(E_{1,1})+w(E_{1,2})$, we have \begin{align*} \| s^2(1-\epsilon)({\mathbf 1}_{E_{1,1}}-{\mathbf 1}_{E_{1,2}})\|\le& \|z_1-z_2\|+\|x-z_2\|\le (2+\epsilon)\|z_1-z_2\|\\ \le& (2+\epsilon)\mathbf{C}\|x\|. \end{align*} The same argument gives $$ \| s^2(1-\epsilon)({\mathbf 1}_{E_{2,1}}-{\mathbf 1}_{E_{2,2}})\|\le (2+\epsilon)\mathbf{C}\|x\|. $$ To finish the proof, define $$ z_3:= {\mathbf 1}_{\varepsilon, A}+ s^{-1}(1+\epsilon)(\sum_{b=1}^{2}{\mathbf 1}_{E_{b,1}} -{\mathbf 1}_{E_{b,2}}). $$ Since ${\mathcal G}(z_3,|E|,s)=\{E\}$, there is $z_4$ with $\supp(z_4)\subset E$ such that $$ \|z_3-z_4\|\le \mathbf{C} \inf_{\substack{z\in {\mathbb X}\\mathbf{w}(\supp(z))\le w(E)\\|\supp(z)|<\infty}}\|z_3-z\|. $$ Therefore, \begin{align*} \|{\mathbf 1}_{\varepsilon, A}\|\le& (\mathbf{M}+\epsilon)\|z_3-z_4\|\le \mathbf{C}(\mathbf{M}+\epsilon) s^{-1}(1+\epsilon)\|\sum_{b=1}^{2}{\mathbf 1}_{E_{b,1}} -{\mathbf 1}_{E_{b,2}}\|\\ \le& 2\mathbf{C}^2(\mathbf{M}+\epsilon) s^{-3}(1+\epsilon)(1-\epsilon)^{-1}(2+\epsilon)\|x\|. \end{align*} Now the proof is completed combining the above result with \eqref{lightsupport7}, and letting $\epsilon$ tend to zero. \end{proof} We have yet to consider the case of $\mathbf{w}\in \mathtt{c}_0\setminus\ell_1$, which (to us) is more intricate. The main obstacle is that we are not able to take a separating sequence of ${\mathcal B}$ given by an application of Proposition~\ref{propositionseparation} or Corollary~\ref{corollaryboth} to prove that (weak) $\mathbf{w}$-semi-greedy bases are $\mathbf{w}$-almost greedy, because we cannot guarantee that there is one such sequence $({\bm x}_{i_k})_{k\in\mathbb{N}}$ with $\mathbf{w}\left(\{i_k\}_{k\in\mathbb{N}}\right)=\infty$, and if the $\mathbf{w}$-measure were finite, there would be greedy sets with arbitrarily greater $\mathbf{w}$-measure than the entire sequence, precluding the kind of approximation we have used in our proofs so far. Even so, if ${\mathcal B}^*$ is $r$-norming for some $0<r\le 1$, we can still prove that weak $\mathbf{w}$-semi-greedy bases are $\mathbf{w}$-almost greedy. Our next result handles this case. \begin{theorem}\label{theoremc0notl1norming} Let ${\mathcal B}$ be $\mathbf{C}$-$s$-$\mathbf{w}$-semi-greedy a basis for ${\mathbb X}$, $\mathbf{w}=(w_i)_{i\in \mathbb{N}}$ a weight, $\mathbf{C}>0$, and $0<s\le 1$. Suppose that $\mathbf{w}$ has a subsequence $(w_{i_k})_{k\in \mathbb{N}}\in \mathtt{c_0}\setminus \ell_1$ and that there is $\mathbf{M}>0$ such that $$ \mathbf{M}_{fs}((v_k)_{k\in \mathbb{N}},{\mathbb X})\le \mathbf{M}, $$ for every block basis $(v_k)_{k\in \mathbb{N}}$ of ${\mathcal B}$ with the property that $(w(\supp(v_k)))_{k\in\mathbb{N}}$ is bounded. Then, the following hold: \begin{enumerate}[\upshape (i)] \item \label{wdisjointgreedyest3} For all $x\in {\mathbb X}$, $m\in \mathbb{N}$, $0<t\le 1$, $A\in {\mathcal G}(x,m,t)$, and $y\in {\mathbb X}$ with $w(\supp(y))\le w(A)$, $$ \|x-P_A(x)\|\le \mathbf{C} \mathbf{M}(1+6\mathbf{C} t^{-1}s^{-3}) \|x-y\|. $$ In particular, ${\mathcal B}$ is $\mathbf{w}$-almost greedy with constant as above taking $t=1$. \item \label{wdisjointgreedyest4} {Let ${\mathcal B}^*=( {\bm x}^*_i)_{i\in\mathbb{N}}$ be the dual basis of ${\mathcal B}$.} For every nonempty finite set $A\subset \mathbb{N}$, every $\varepsilon\in \ensuremath{\mathcal{E}}_A$ and every $x\in {\mathbb X}$, if $$ w(A) \le w(\{i\in \mathbb{N}: |{\bm x}_i^*(x)|\ge 1\}) $$ then $$ \|{\mathbf 1}_{\varepsilon,A}\|\le 4s^{-3}\mathbf{C}^2\mathbf{M} \|x\|. $$ \end{enumerate} Thus, ${\mathcal B}$ is $\mathbf{w}$-superdemocratic and truncation quasi-greedy, in each case with constant as above. \\ In addition, if ${\mathcal B}^*$ is $r$-norming for some $0<r\le 1$, the above conditions hold with $\mathbf{M}=r^{-1}$. \end{theorem} \begin{proof} The statement for the case of an $r$-norming dual basis follows from Proposition~\ref{propositionseparation}, so we need to prove \ref{wdisjointgreedyest3} and \ref{wdisjointgreedyest4}. To prove the former, fix $x$, $m$, $t$ and $A$ as in the statement, and $y$ with $w(\supp(y))\le w(A)$. We may assume $x\not=P_A(x)$, and we will also assume that both $x$ and $y$ have finite support. Set $$ a:=\min_{i\in A}|{\bm x}_i^*(x)|, $$ choose $i_0>\supp(x)\cup \supp(y)$, and set $ {\mathbb F}:=[ {\bm x}_i: 1\le i\le i_0]$. Given that $(w_{i_k})_{k\in \mathbb{N}}\in \mathtt{c_0}\setminus \ell_1$, there is a sequence of finite sets of positive integers $(A_k)_{k\in \mathbb{N}}$ such that for all $k\in \mathbb{N}$, $i_0<A_k<A_{k+1}$ and \begin{equation} \frac{w(A)}{3}\le w(A_k)\le \frac{w(A)}{2}. \label{part1sets} \end{equation} By hypothesis $\mathbf{M}_{fs}(({\mathbf 1}_{A_k})_{k\in \mathbb{N}},{\mathbb X})\le \mathbf{M}$. To simplify our notation, we may assume that $({\mathbf 1}_{A_k})_{k\in \mathbb{N}}$ is already a separating sequence for $({\mathbb X},\mathbf{M},\epsilon)$. By Proposition~\ref{propositionupperbounds}, $({\mathbf 1}_{A_k})_{k\in \mathbb{N}}$ is bounded. Since it is also $\lambda '^{-1}$-uniformly discrete, we may choose $0<\epsilon<1$ and apply Lemma~\ref{lemmaxi-x2} to $({\mathbf 1}_{A_k})_{k\in \mathbb{N}}$. Again, we assume that $({\mathbf 1}_{A_k})_{k\in \mathbb{N}}$ is already the subsequence given by the lemma. Set $$ \mathbb{L}:=\{k\in 5\mathbb{N}: k> \max\{r_{{\mathbb F},({\mathbf 1}_{A_n})_{n\in \mathbb{N}},\epsilon}, j_{{\mathbb F},({\mathbf 1}_{A_n})_{n\in \mathbb{N}},\epsilon}, \}\}, $$ and for every $k\in \mathbb{L}$, define \begin{align*} z_{1,k}:=&x-P_A(x)+s^{-1}t^{-1}a(1+\epsilon)\sum_{l=1}^{3}({\mathbf 1}_{A_{2k+2l-1}}-{\mathbf 1}_{A_{2k+2l}});\\ z_{2,k,l}:=&x-y-s^2(1-\epsilon)a({\mathbf 1}_{A_{2k+2l-1}}-{\mathbf 1}_{A_{2k+2l}}), \qquad\forall 1\le l\le 3. \end{align*} Fix $k$ as above and $1\le l\le 3$. By Lemma~\ref{lemmasgset}, there is a set $D\supset A$ such that $D\in {\mathcal {GS}}(z_{2,k,l},|D|,s)$ and $$ \min_{j\in D}|{\bm x}_j^*(z_{2,k,l})|\ge s^2\min_{j\in A}|{\bm x}_j^*(z_{2,k,l})|=s^2\min_{j\in A}|{\bm x}_j^*(x)|=s^2 a. $$ It follows that $D\subset \{1,\dots,i_0\}$, so there is $z_{3,k,l}\in {\mathbb F}$ such that $$ \|z_{2,k,l}-z_{3,k,l}\|\le \mathbf{C} \inf_{\substack{z\in {\mathbb X}\\|\supp(z)|<\infty\\mathbf{w}(\supp(z))\le w(D)}}\|z_{2,k,l}-z\|. $$ As $D\supset A$, using the above inequality and \eqref{part1sets} we get $$ \|z_{2,k,l}-z_{3,k,l}\|\le \mathbf{C}\|x-y\|. $$ Hence, \begin{align} \|s^2(1-\epsilon)a({\mathbf 1}_{A_{2k+2l-1}}-{\mathbf 1}_{A_{2k+2l}})\|\le& \|z_{2,k,l}-z_{3,k,l}\|+\|x-y-z_{3,k,l}\|\nonumber\\ \le& (2+\epsilon) \|z_{2,k,l}-z_{3,k,l}\|\le \mathbf{C}(2+\epsilon)\|x-y\|.\label{firstleftboundikl} \end{align} Now we consider $z_{1,k}$: set $$ B_k:=\bigcup_{l=1}^{3}A_{2k+2l-1}\cup A_{2k+2l}. $$ Notice that ${\mathcal G}(z_{1,k},|B_k|,s)=\{B_k\}$. Hence, there is $z_{4,k}$ with $\supp(z_{4,k})\subset B_k$ such that $$ \|z_{1,k}-z_{4,k} \|\le \mathbf{C} \inf_{\substack{z\in {\mathbb X}\\|\supp(z)|<\infty\\mathbf{w}(\supp(z))\le w(B_k)}}\|z_{1,k}-z\|. $$ By \eqref{part1sets}, $2w(A)\le w(B_k)$. Since $w(\supp(y))\le w(A)$, this gives \begin{align} \|z_{1,k}-z_{4,k} \|& \le \mathbf{C} \|z_{1,k}+P_A(x)-y\| \nonumber \\ & \le \mathbf{C}\|x-y\|+\mathbf{C} a(1+\epsilon)s^{-1}t^{-1}\|\sum_{l=1}^{3}({\mathbf 1}_{A_{2k+2l-1}}-{\mathbf 1}_{A_{2k+2l}})\|.\label{firstrightboundikl} \end{align} Note that the sequence $$ \big(u_k:=a(1+\epsilon)s^{-1}t^{-1}\sum_{l=1}^{3}({\mathbf 1}_{A_{2k+2l-1}}-{\mathbf 1}_{A_{2k+2l}})-z_{4,k}\big)_{k\in \mathbb{L}} $$ has the finite dimensional separation property with constant $\le \mathbf{M}$. Indeed, if $u_k=0$ for infinitely many values of $k$, the sequence has this property with constant $1$, whereas if this is not the case, there is $k_0\in \mathbb{L}$ such that the subsequence beginning in $k_0$ is a block basis of ${\mathcal B}$, so we have this bound by hypothesis. In particular, it follows that there is $k\in \mathbb{L}$ such that \begin{align*} \|x-P_A(x)\|\le& (\mathbf{M}+\epsilon)\|x-P_A(x) +a(1+\epsilon)s^{-1}t^{-1}\sum_{l=1}^{3}({\mathbf 1}_{A_{2k+2l-1}}-{\mathbf 1}_{A_{2k+2l}})-z_{4,k}\|\\ =&(\mathbf{M}+\epsilon)\|z_{1,k}-z_{4,k}\|, \end{align*} which, when combined with \eqref{firstleftboundikl}, \eqref{firstrightboundikl} and the triangle inequality gives $$ \|x-P_A(x)\|\le \mathbf{C}(\mathbf{M}+\epsilon)(1+3\mathbf{C}(1+\epsilon)(1-\epsilon)^{-1}(2+\epsilon)t^{-1}s^{-3})\|x-y\|. $$ As $\epsilon$ is arbitrary, this completes the proof of \ref{wdisjointgreedyest3} for $x$ and $y$ with finite support, and the general case is proven by the argument given in the proof of Theorem~\ref{theoremswemigreedydisjointgreedynotc0}. \\ The proof of \ref{wdisjointgreedyest4} is similar: Fix $A$, $\varepsilon$ and $x$ as in the statement, set $$ B:=\{i\in \mathbb{N}: |{\bm x}_i^*(x)|\ge 1\}, $$ choose $0<\epsilon<1$ and $i_0>A\cup \supp(x)$. Now choose a sequence of sets of positive integers $(A_k)_{k\in \mathbb{N}}$ so that for all $k$, $i_0<A_k<A_{k+1}$ and \begin{equation} \frac{w(A)}{4}\le w(A_k) \le \frac{w(A)}{2}.\label{part2sets} \end{equation} As before, we assume that $({\mathbf 1}_{A_k})_{k\in\mathbb{N}}$ is already a separating sequence for $({\mathbb X},\mathbf{M},\epsilon)$, and that we have applied Lemma~\ref{lemmaxi-x2}. Set \begin{align*} {\mathbb F}:=&[ {\bm x}_i: 1\le i\le i_0];\\ \mathbb{L}:=&\{k\in 5\mathbb{N}: k >\max\{r_{{\mathbb F},({\mathbf 1}_{A_n})_{n\in \mathbb{N}},\epsilon}, j_{{\mathbb F},({\mathbf 1}_{A_n})_{n\in \mathbb{N}},\epsilon},\} \}. \end{align*} For each $k\in \mathbb{L}$, define \begin{align*} z_{1,k}:=&{\mathbf 1}_{\varepsilon,A}+s^{-1} (1+\epsilon)\sum_{l=1}^{2}({\mathbf 1}_{A_{2k+2l-1}}-{\mathbf 1}_{A_{2k+2l}});\\ z_{2,k,l}:=&x-s^2(1-\epsilon) ({\mathbf 1}_{A_{2k+2l-1}}-{\mathbf 1}_{A_{2k+2l}}),\qquad\forall 1\le l\le 2. \end{align*} Fix $k\in \mathbb{L}$ and $1\le l\le 2$. By Lemma~\ref{lemmasgset}, there is a set $D\supset B$ such that $D\in {\mathcal {GS}}(z_{2,k,l},|D|,s)$ and $$ \min_{j\in D}|{\bm x}_j^*(z_{2,k,l})|\ge s^2\min_{j\in B}|{\bm x}_j^*(z_{2,k,l})|=s^2\min_{j\in B}|{\bm x}_j^*(x)|\ge s^2, $$ which implies that $D\subset \{1,\dots,i_0\}$. Hence, there is $z_{3,k,l}\in {\mathbb F}$ such that $$ \|z_{2,k,l}-z_{3,k,l}\|\le \mathbf{C} \inf_{\substack{z\in {\mathbb X}\\|\supp(z)|<\infty\\mathbf{w}(\supp(z))\le w(D)}}\|z_{2,k,l}-z\|. $$ As $D\supset B$, using the above inequality, \eqref{part2sets} and the fact that $w(A)\le w(B)$ we get $$ \|z_{2,k,l}-z_{3,k,l}\|\le \mathbf{C}\|x\|. $$ Thus, the property of Lemma~\ref{lemmaxi-x2} gives \begin{align} \|s^2(1-\epsilon)({\mathbf 1}_{A_{2k+2l-1}}-{\mathbf 1}_{A_{2k+2l}})\|\le& \|z_{2,k,l}-z_{3,k,l}\|+\|x-z_{3,k,l}\|\nonumber\\ \le& (2+\epsilon) \|z_{2,k,l}-z_{3,k,l}\|\le \mathbf{C}(2+\epsilon)\|x\|.\label{firstleftboundikl2} \end{align} For fixed $k\in \mathbb{L}$, set $$ B_k:=\bigcup_{l=1}^{2}A_{2k+2l-1}\cup A_{2k+2l}. $$ Note that ${\mathcal G}(z_{1,k},|B_k|,s)=\{B_k\}$. Thus, there is $z_{4,k}$ with support contained in $B_k$ such that $$ \|z_{1,k}-z_{4,k} \|\le \mathbf{C} \inf_{\substack{z\in {\mathbb X}\\|\supp(z)|<\infty\\mathbf{w}(\supp(z))\le w(B_k)}}\|z_{1,k}-z\|. $$ By \eqref{part2sets}, $w(A)\le w(B_k)$, so \begin{equation} \|z_{1,k}-z_{4,k} \|\le \mathbf{C} \|z_{1,k} -{\mathbf 1}_{\varepsilon,A} \|=\mathbf{C} (1+\epsilon)s^{-1}\|\sum_{l=1}^{2}({\mathbf 1}_{A_{2k+2l-1}}-{\mathbf 1}_{A_{2k+2l}})\|.\label{firstrightboundikl2} \end{equation} As before, the sequence $$ \big((1+\epsilon)s^{-1}\sum_{l=1}^{2}({\mathbf 1}_{A_{2k+2l-1}}-{\mathbf 1}_{A_{2k+2l}})-z_{4,k}\big)_{k\in \mathbb{L}} $$ has the finite dimensional separation property with constant $\le \mathbf{M}$. In particular, there is $k\in \mathbb{L}$ such that \begin{align*} \|{\mathbf 1}_{\varepsilon,A}\|\le& (\mathbf{M}+\epsilon)\|z_{1,k}-z_{4,k}\| \end{align*} which, when combined with \eqref{firstleftboundikl2}, \eqref{firstrightboundikl2} and the triangle inequality gives $$ \|{\mathbf 1}_{\varepsilon,A}\|\le 2\mathbf{C}^2(\mathbf{M}+\epsilon)(1+\epsilon)(1-\epsilon)^{-1}(2+\epsilon)s^{-3}\|x\|. $$ As $\epsilon$ is arbitrary, the proof is complete. \end{proof} It remains to study the general case $\mathbf{w}\in \mathtt{c}_0\setminus \ell_1$. We do not know whether all (weak) $\mathbf{w}$-semi-greedy bases are quasi-greedy, but we can prove that they are $\mathbf{w}$-superdemocratic and truncation quasi-greedy. To do so, we only need to address the cases that do not meet the conditions of Theorem~\ref{theoremswemigreedydisjointgreedynotc0} or Theorem~\ref{theoremc0notl1norming}. The following result covers all such cases. \begin{proposition}\label{propositionc0notl1} Let ${\mathcal B}$ be a $\mathbf{C}$-$s$-$\mathbf{w}$-semi-greedy basis for ${\mathbb X}$, $\mathbf{w}$ a weight, $\mathbf{C}>0$ and $0<s\le 1$. Let ${\mathcal B}^*=({\bm x}^*_i)_{i\in \mathbb{N}}$ be the dual basis of ${\mathcal B}$ and suppose that ${\mathcal B}$ ha a seminormalized block basis $(v_k)_{k\in \mathbb{N}}$ with $(w(\supp(v_k))_{k\in \mathbb{N}}$ bounded that is not weakly null. Then, there is $\mathbf{K}>0$ such that \begin{enumerate}[\upshape (i)] \item \label{anotherone}For all $x\in {\mathbb X}$ and all $A\in \mathbb{N}^{<\infty}$, $$ \min_{i\in A}|{\bm x}_i^*(x)| w(A)\le \mathbf{K} \|x\|. $$ \item \label{normlikeweight}For all $A\in \mathbb{N}^{<\infty}$ and all $\varepsilon\in \ensuremath{\mathcal{E}}_{A}$, $$ \max\{w(A),1\}\mathbf{K}^{-1}\le \|{\mathbf 1}_{\varepsilon,A}\|\le \mathbf{K}\max\{w(A),1\}. $$ \item \label{propC}For all $x\in {\mathbb X}$, all $A\in \mathbb{N}^{<\infty}$ and all $\varepsilon\in \ensuremath{\mathcal{E}}_{A}$, $$ \min_{i\in A}|{\bm x}_i^*(x)|\|{\mathbf 1}_{\varepsilon,A}\|\le \mathbf{K}\|x\|. $$ \item \label{superdem}For all $A, B\in \mathbb{N}^{<\infty}$ with $w(A)\le w(B)$, and all $\varepsilon\in \ensuremath{\mathcal{E}}_{A}$, $\varepsilon'\in \ensuremath{\mathcal{E}}_{B}$, $$ \|{\mathbf 1}_{\varepsilon,A}\|\le \mathbf{K}\|{\mathbf 1}_{\varepsilon',B}\|. $$ \end{enumerate} \end{proposition} \begin{proof} As this is not a quantitative result, we will not keep track of the constants (even so, note that the right-hand side of the inequality in \ref{normlikeweight} was estimated in Proposition~\ref{propositionupperbounds}). \\ Note that \ref{normlikeweight} follows from \ref{anotherone}, Proposition~\ref{propositionupperbounds}, and the fact that $\|x\|_{\infty}\le \lambda '\|x\|$ for all $x\in {\mathbb X}$; \ref{propC} follows from that fact together with \ref{anotherone} and \ref{normlikeweight}, whereas \ref{superdem} follows from \ref{normlikeweight}. Thus, we only need to prove \ref{anotherone}. \\ Let ${\mathcal B}=( {\bm x}_i)_{i\in\mathbb{N}}$. Since $(v_k)_{k\in \mathbb{N}}$ is not weakly null, passing to a subsequence we may assume there is $\epsilon>0$ and $x^*\in S_{X^*}$ such that $$ |x^*(v_k)|\ge \epsilon, \qquad\forall k\in\mathbb{N}. $$ For each $k\in \mathbb{N}$, set $A_k:=\supp(v_k)$, and define $c_1:=\max\{1,\| (\|v_k\|_{\infty})_{k\in\mathbb{N}}\|_{\infty}\}$.\\ Note that $c_1$ is a well-defined positive number because ${\mathcal B}^*$ and $(v_k)_{k\in \mathbb{N}}$ are both bounded. For each $k\in \mathbb{N}$, choose $\varepsilon^{(k)}\in \ensuremath{\mathcal{E}}_{A_k}$ so that $$ \varepsilon_j^{(k)} x^*({\bm x}_j)\ge 0, \qquad \forall j\in A_k. $$ Note that $({\mathbf 1}_{\varepsilon^{(k)},A_k})_{k\in \mathbb{N}}$ is bounded by Proposition~\ref{propositionupperbounds}. Let $\epsilon_1:=\epsilon c_1^{-1}$. For each $k\in \mathbb{N}$, we have \begin{align} x^*({\mathbf 1}_{\varepsilon^{(k)},A_k})=&\sum_{j\in A_k}|x^*({\bm x}_j)|\ge c_1^{-1}\sum_{j\in A_k}|{\bm x}_j^*(v_k)| |x^*({\bm x}_j)|\ge c_1^{-1}|x^*(\sum_{j\in A_k}{\bm x}_j^*(v_j){\bm x}_j)|\nonumber\\ =& c_1^{-1}|x^*(v_k)| \ge c_1^{-1}\epsilon =\epsilon_1>0. \label{notweaklynull} \end{align} Set $$ a=\liminf_{k\to \infty}w(A_k). $$ By hypothesis, $a$ is a nonnegative real number. We claim that $a>0$. Otherwise, there would be a subsequence $({\mathbf 1}_{\varepsilon^{(k_j)},A_{k_j}})_{j\in \mathbb{N}}$ such that $(w(A_{k_j}))_{j\in \mathbb{N}}\in \ell_1$. By Lemma~\ref{lemmal1thenc0}, $(({\bm x}_i)_{i\in A_{k_j}})_{j\in \mathbb{N}}$ would be equivalent to the canonical unit vector basis of $\mathtt{c}_{0}$, so all of its bounded block bases would be weakly null, contradicting \eqref{notweaklynull}. Thus, passing to a subsequence if necessary we may assume that \begin{equation} \frac{a}{2}\le w(A_k)\le 2a, \qquad\forall k\in\mathbb{N}. \label{similarweights5} \end{equation} Let $\mathbf{M}:=\mathbf{M}_{fs}(({\mathbf 1}_{\varepsilon^{(k)},A_k})_{k\in \mathbb{N}},{\mathbb X})$. Applying Proposition~\ref{propositionseparation}, again we may assume that $({\mathbf 1}_{\varepsilon^{(k)},A_k})_{k\in \mathbb{N}}$ is already a separating sequence with the properties of Definition~\ref{definitionseparation} for $({\mathbb X},\mathbf{M},1)$. \\ Now fix $x\in {\mathbb X}$, $A$ a finite nonempty subset of $\mathbb{N}$, and $\varepsilon\in \ensuremath{\mathcal{E}}_{A}$. We may assume that $A\subset \supp(x)$. If $w(A)\le 3a$, pick any $i\in A$. We have \begin{equation} \min_{i\in A}|{\bm x}_i^*(x)| w(A)\le 3a |{\bm x}_{i}^*(x)|\le 3a \lambda '\|x\|. \label{lightset5} \end{equation} On the other hand, if $w(A)>3a$, define $b:=\min_{i\in A}|{\bm x}_i^*(x)|$, choose $i_0\in \mathbb{N}$ so that \begin{equation} bs^2>2|{\bm x}_i^*(x)|,\qquad\forall i\ge i_0, \label{smallenough5} \end{equation} and set $$ {\mathbb F}:=[x,{\bm x}_i:1\le i\le i_0]. $$ By \eqref{similarweights5}, there is $B>j_{{\mathbb F},1}$ such that \begin{equation} \sum_{k\in B}w(A_k)\le w(A)\le 2 \sum_{k\in B}w(A_k). \label{approaxingAinweight} \end{equation} Set $$ z_1:=x+ \frac{bs^{2}}{2}\sum_{k\in B}{\mathbf 1}_{\varepsilon^{(k)},A_k}. $$ By Lemma~\ref{lemmasgset}, there is $D\supset A$ such that $D\in {\mathcal {GS}}(z_1,|D|,s)$ and $$ \min_{j\in D}|{\bm x}_j^*(z_1)|\ge s^2 \min_{j\in A}|{\bm x}_j^*(z_1)|=s^2 \min_{j\in A}|{\bm x}_j^*(x)|=bs^2. $$ It follows from \eqref{smallenough5} that $D\subset \{1,\dots,i_0\}$, so there is $z_2\in {\mathbb F}$ such that $$ \|z_1-z_2\|\le \mathbf{C} \inf_{\substack{z\in {\mathbb X}\\|\supp(z)|<\infty\\mathbf{w}(\supp(z))\le w(D)}}\|z_{1}-z\|. $$ Thus, using \eqref{approaxingAinweight} and the separating condition on $({\mathbf 1}_{\varepsilon^{(k)},A_k})_{k\in \mathbb{N}}$ we deduce that \begin{align*} \|\frac{b s^{2}}{2} \sum_{k\in B}{\mathbf 1}_{\varepsilon^{(k)},A_k}\|\le& \|z_1-z_2\|+\|x-z_2\|\le (\mathbf{M}+2)\|z_1-z_2\|\le (\mathbf{M}+2)\mathbf{C}\|x\|. \end{align*} On the other hand, by \eqref{notweaklynull}, \eqref{similarweights5} and \eqref{approaxingAinweight}, \begin{align*} \|\sum_{k\in B}{\mathbf 1}_{\varepsilon^{(k)},A_k}\|\ge& |x^*(\sum_{k\in B}{\mathbf 1}_{\varepsilon^{(k)},A_k})|\ge \epsilon_1 |B|\ge \epsilon_1 \sum_{k\in B}\frac{w(A_k)}{2a}\ge \frac{\epsilon_1}{4a} w(A). \end{align*} Hence, $$ \min_{i\in A}|{\bm x}_i^*(x)|w(A)=bw(A) \le 8 a \epsilon_1^{-1} s^{-2}\mathbf{C}(\mathbf{M}+2)\|x\|. $$ The proof is completed combining the above inequality and \eqref{lightset5}. \end{proof} In \cite{BDKOW2019}, the authors introduced and studied the \emph{weighted Property (A)}, extending to the weight setting a property that arises naturally in the context of the TGA. Property (A) has been studied for example in \cite{AA2017}, \cite{AW2006} and \cite{BB2017}. \begin{definition}(\cite[Definition~1.3]{BDKOW2019}) Let ${\mathcal B}$ be a basis for ${\mathbb X}$, $\mathbf{w}$ a weight and $\mathbf{C}>0$. We say that ${\mathcal B}$ \emph{has the $\mathbf{C}$-$\mathbf{w}$-Property (A)} if $$ \|x+{\mathbf 1}_{\varepsilon,A}\|\le \mathbf{C} \|x+{\mathbf 1}_{\varepsilon',B}\|, $$ for any $ x\in {\mathbb X}, \|x\|_{\infty}\le 1$, for any $A,B\in \mathbb{N}^{<\infty}$ such that $w(A)\le w(B)$ with $A\cap B=\emptyset$ and $\supp(x)\cap (A\cup B)=\emptyset,$ and for any $\varepsilon\in \ensuremath{\mathcal{E}}_A, \varepsilon'\in \ensuremath{\mathcal{E}}_B$. \end{definition} From our previous results, we have the following corollary, which extends \cite[Theorem 5.2]{BDKOW2019}. \begin{corollary}\label{corollaryPropCwA} Let ${\mathcal B}$ be a basis for ${\mathbb X}$, $\mathbf{w}$ a weight and $0<s\le1$. If ${\mathcal B}$ is $s$-$\mathbf{w}$-semi-greedy, it is truncation quasi-greedy and $\mathbf{w}$-superdemocratic. Hence, it has the $\mathbf{w}$-Property (A). \end{corollary} \begin{proof} Note that if $\mathbf{w}\in \mathtt{c}_0\setminus \ell_1$ and the hypotheses of Proposition~\ref{propositionc0notl1} do not hold, then by Proposition~\ref{propositionseparation} the hypotheses of Theorem~\ref{theoremc0notl1norming} do, with $\mathbf{M}=1$. Hence, it follows from Theorems~\ref{theoremswemigreedydisjointgreedynotc0}, ~\ref{theoremc0notl1norming}, Proposition~\ref{propositionc0notl1} and Lemma~\ref{lemmal1thenc0} that if ${\mathcal B}$ is $s$-$\mathbf{w}$-semi-greedy, it is truncation quasi-greedy and $\mathbf{w}$-superdemocratic. Then, by \cite[Proposition 3.13]{BDKOW2019}, it has the $\mathbf{w}$-Property (A). \end{proof} Before we end this section, there are two questions about weak weight-semi-greedy bases that need to be addressed. First, in Definition~\ref{definitionweakweightsemigreedy} we require the existence of \emph{one} $s$-greedy set for which \eqref{weakweightsemigreedy} holds; this is in line with similar definitions in \cite{BL2021} and \cite{DKSW2012}. An alternative would have been to require that \eqref{weakweightsemigreedy} holds for \emph{all} $s$-greedy sets; this would have been similar to the definition of $\mathbf{n}$-$s$-quasi greedy bases in \cite{Oikhberg2017}. So, one question is whether the two approaches are equivalent. The second is whether a basis that is $s$-$\mathbf{w}$-semi-greedy for some $0<s\le 1$ and weight $\mathbf{w}$, is also $t$-$\mathbf{w}$-semi-greedy for all $0<t\le 1$. After proving that weak weight-semi-greedy bases are truncation quasi-greedy, we are able to tackle and answer both questions in the affimative. In order to do so, we will use the following definition (see e.g. \cite{AABW2021}). \begin{definition}A basis ${\mathcal B}$ is \emph{suppression unconditional for constant coefficients} with constant $\mathbf{C}$ ($\mathbf{C}$-SUCC) if $$ \|{\mathbf 1}_{\varepsilon, B}\|\le \mathbf{C}\|{\mathbf 1}_{\varepsilon,A}\| $$ for all $B\subset A\in \mathbb{N}^{<\infty}$ and all $\varepsilon\in \ensuremath{\mathcal{E}}_A$. \end{definition} Clearly a basis that is either $\mathbf{C}$-truncation quasi-greedy or $\mathbf{C}$-$\mathbf{w}$-superdemocratic for some weight $\mathbf{w}$ is also $\mathbf{C}$-SUCC. Hence, by Corollary~\ref{corollaryPropCwA}, any weak weight-semi-greedy basis is SUCC. \begin{theorem}\label{theoremalltgreedysetsallt}Let ${\mathcal B}$ be a basis for ${\mathbb X}$, $\mathbf{w}$ a weight and $0<s\le 1$. If ${\mathcal B}$ is $s$-$\mathbf{w}$-semi-greedy, then for every $0<t\le 1$ there is $\mathbf{C}(s,t)>0$ such that for every $x\in {\mathbb X}$, $m\in \mathbb{N}$, and $A\in{\mathcal G}(x,m,t)$, there is $y\in {\mathbb X}$ with $\supp(y)\subset A$ such that \begin{align} \|x-y\|\le \mathbf{C}(s,t)\inf_{\substack{z\in {\mathbb X}\\mathbf{w}(\supp(z))\le w(A)}}\|x-z\|.\label{theoremalltgreedysetsalltallsets} \end{align} In particular, ${\mathcal B}$ is $t$-$\mathbf{w}$-semi-greedy for all $0<t\le 1$. \end{theorem} \begin{proof} Set ${\mathcal B}=({\bm x}_i)_{i\in \mathbb{N}}$ and ${\mathcal B}^*=({\bm x}^*_i)_{i\in \mathbb{N}}$ its dual basis. Suppose ${\mathcal B}$ is $\mathbf{C}_1$-$s$-$\mathbf{w}$-semi-greedy. By the previous remarks and Corollary~\ref{corollaryPropCwA}, ${\mathcal B}$ is $\mathbf{C}_2$-SUCC and $\mathbf{C}_3$-truncation quasi-greedy for some positive constants $\mathbf{C}_2\le \mathbf{C}_3$. Set $x,m,t, A$ as in the statement. If $x=P_A(x)$, there is nothing to prove. Otherwise, take $a:=\min_{i\in A}|{\bm x}_i^*(x)|>0$ because $A\in {\mathcal G}(x,m,t)$. Choose $y\in [{\bm x}_i: i\in A]$ so that \begin{align} \|x-y\|=\min_{\substack{z\in {\mathbb X}\\ \supp(z)\subset A }} \|x-z\|, \label{theoremalltgreedysetsalltmbestapprox} \end{align} pick $\epsilon>0$, and set $$u:=x+at^{-1}s^{-1}(1+\epsilon){\mathbf 1}_{\varepsilon(x), A}.$$ For every $i\in A$ and $k\not \in A$, \begin{align*} |{\bm x}_k^*(u)|=&|{\bm x}_k^*(x)|\le t^{-1}a<s |{\bm x}_i^*(u)|. \end{align*} Hence, ${\mathcal G}(u,m,s)=\{A\}$. By Lemma~\ref{lemmafinitesupportisenough}, there is $y_1\in [{\bm x}_i:i\in A]$ such that \begin{align*} \|u-y_1\|\le \mathbf{C}_1\inf_{\substack{z\in {\mathbb X}\\mathbf{w}(\supp(z))\le w(A)}}\|u-z\|. \end{align*} Thus, setting $$y_2:=-at^{-1}s^{-1}(1+\epsilon){\mathbf 1}_{\varepsilon(x), A}+y_1,$$ we get \begin{align} \|x-y_2\|=&\|u-y_1\|\le \mathbf{C}_1 \inf_{\substack{z\in {\mathbb X}\\mathbf{w}(\supp(z))\le w(A)}}\|x-z +at^{-1}s^{-1}(1+\epsilon){\mathbf 1}_{\varepsilon(x), A}\|.\label{theoremalltgreedysetsalltmbestapprox2} \end{align} Now choose $v\in {\mathbb X}$ with $w(\supp(v))\le w(A)$. If $\supp(v)\subset A$, then $\|x-y\|\le \|x-v\|$. Otherwise, set \begin{align*} &A_1:=A\cap \supp(v), &&A_2:=A\setminus \supp(v). \end{align*} By \eqref{theoremalltgreedysetsalltmbestapprox2}, \begin{align} \|x-y_2\|\le& \mathbf{C}_1\|x-v-at^{-1}s^{-1}(1+\epsilon){\mathbf 1}_{\varepsilon(x), A_1}+at^{-1}s^{-1}(1+\epsilon){\mathbf 1}_{\varepsilon(x), A} \|\nonumber\\ \le&\mathbf{C}_1\|x-v\|+a\mathbf{C}_1t^{-1}s^{-1}(1+\epsilon)\|{\mathbf 1}_{\varepsilon(x), A_2}\|.\label{theoremalltgreedysetsalltmbestapprox3} \end{align} Note that $A_2\not=\emptyset$. Let $A_3$ be a greedy set for $x-v$ of minimum cardinality containing $A_2$. We have \begin{align*} \min_{i\in A_3}|{\bm x}_i^*(x-v)|=&\min_{i\in A_2}|{\bm x}_i^*(x-v)|=\min_{i\in A_2}|{\bm x}_i^*(x)|\ge \min_{i\in A}|{\bm x}_i^*(x)|=a. \end{align*} Since $A_2\subset A_3$, \begin{align*} a\|{\mathbf 1}_{\varepsilon(x), A_2}\|=& a\|{\mathbf 1}_{\varepsilon(x-v), A_2}\|\le \mathbf{C}_2a\|{\mathbf 1}_{\varepsilon(x-v),A_3}\|\le \mathbf{C}_2\mathbf{C}_3\|x-v\|. \end{align*} Combining the above with \eqref{theoremalltgreedysetsalltmbestapprox} and \eqref{theoremalltgreedysetsalltmbestapprox3} it follows that $$ \|x-y\|\le \mathbf{C}_1(1+\mathbf{C}_2\mathbf{C}_3t^{-1}s^{-1}(1+\epsilon))\|x-v\|. $$ Taking infimum over all such $v$ and letting $\epsilon\rightarrow 0$, we conclude that \eqref{theoremalltgreedysetsalltallsets} holds for $\mathbf{C}(s,t)=\mathbf{C}_1(1+\mathbf{C}_2\mathbf{C}_3 t^{-1}s^{-1})$. \end{proof} \bigskip \section{Weak weight-almost semi-greedy bases.}\label{sectionalmostsemigreedy} An examination of the proofs of Proposition~\ref{propositionupperbounds}, Theorem~\ref{theoremswemigreedydisjointgreedynotc0}\ref{wdisjointgreedyest2a}, Theorem~\ref{theoremc0notl1norming}\ref{wdisjointgreedyest4}, Proposition~\ref{propositionc0notl1} and Corollary~\ref{corollaryPropCwA} shows that these results do not need the full strength of the $\mathbf{C}$-$s$-$\mathbf{w}$-semi-greedy property, but can be obtained using approximations by projections. This suggests an ``almost semi-greedy'' property, and its corresponding weak and weighted versions. In this section we study the position of such bases with respect to the already known ones. \begin{definition}\label{definitionweakweightalmostsemigreedy} Let ${\mathcal B}$ be a basis for ${\mathbb X}$, $\mathbf{w}$ a weight, $\mathbf{C}>0$, and $0<s\le 1$. We say that ${\mathcal B}$ is \emph{weak weight-almost semi-greedy with parameter $s$, weight $\mathbf{w}$ and constant $\mathbf{C}$} (or $\mathbf{C}$-$s$-$\mathbf{w}$-almost semi-greedy) if, for every $x\in {\mathbb X}$ and $m\in \mathbb{N}$, there is $A\in {\mathcal G}(x,m,s)$ and $y\in {\mathbb X}$ with $\supp(y)\subset A$ such that \begin{equation} \|x-y\|\le \mathbf{C}\inf_{\substack{B\in \mathbb{N}^{<\infty}\\ w(B)\le w(A)}}\|x-P_B(x)\|.\label{definitionweakweightalmostsemigreedyb} \end{equation} We denote by $\mathcal{GAS}(x,m,s)$ the subset of ${\mathcal G}(x,m,s)$ for which the above holds. In case $s=1$ and $w_n=1$ for all $n\in \mathbb{N}$, we say that ${\mathcal B}$ is \emph{$\mathbf{C}$-almost semi-greedy}. \end{definition} As pointed out, the same proofs of the aforementioned results hold verbatim under this weaker hypothesis, obtaining the same estimates when we replace the $s$-$\mathbf{w}$-semi-greedy constant with the $s$-$\mathbf{w}$-almost semi-greedy one. Thus, in particular, any $s$-$\mathbf{w}$-almost semi-greedy basis is truncation quasi-greedy and $\mathbf{w}$-superdemocratic. It turns out that truncation quasi-greediness and $\mathbf{w}$-superdemocracy characterize $\mathbf{w}$-almost semi-greediness and its weak variant, as we prove next. \begin{proposition}\label{propositionequivalencesalmostsemigreedy} Let ${\mathcal B}$ be a basis for ${\mathbb X}$ and $\mathbf{w}$ a weight. The following are equivalent: \begin{enumerate}[ \rm (i)] \item \label{propositionequivalencesalmostsemigreedytqg+wsd} ${\mathcal B}$ is truncation quasi-greedy and $\mathbf{w}$-superdemocratic. \item \label{propositionequivalencesalmostsemigreedytqg+wd}${\mathcal B}$ is truncation quasi-greedy and $\mathbf{w}$-democratic. \item \label{propositionequivalencesalmostsemigreedyalmostsemiall+} For every $0<s\le 1$, ${\mathcal B}$ is $s$-$\mathbf{w}$-almost semi-greedy.\\ Moreover, there is $\mathbf{C}(s)>0$ such that for each $x\in {\mathbb X}$ and $m\in \mathbb{N}$, $\mathcal{GAS}(x,m,s, \mathbf{C}(s))={\mathcal G}(x,m,s)$ and, for each $A\in {\mathcal G}(x,m,s)$, there is $D\subset A$ such that \eqref{definitionweakweightalmostsemigreedyb} holds for $\mathbf{C}=\mathbf{C}(s)$ and $y=P_D(x)$. \item \label{propositionequivalencesalmostsemigreedyalmostsemis}There is $0<s\le 1$ such that ${\mathcal B}$ is $s$-$\mathbf{w}$-almost semi-greedy. \item \label{propositionequivalencesalmostsemigreedydisjoint}For each $0<s\le 1$ there is $\mathbf{C}(s)$ such that \begin{align*} \|x\|\le& \mathbf{C}(s)\|x-P_B(x)\|, \end{align*} for all $x\in {\mathbb X}$ and $B\in \mathbb{N}^{<\infty}$ such that there is $m\in \mathbb{N}$ and $A\in {\mathcal G}(x,m,s)$ disjoint from $B$ with $w(B)\le w(A)$. \end{enumerate} \end{proposition} \begin{proof} The implications \ref{propositionequivalencesalmostsemigreedytqg+wsd} $\Longrightarrow$ \ref{propositionequivalencesalmostsemigreedytqg+wd} and \ref{propositionequivalencesalmostsemigreedyalmostsemiall+} $\Longrightarrow$ \ref{propositionequivalencesalmostsemigreedyalmostsemis} are immediate, whereas \ref{propositionequivalencesalmostsemigreedyalmostsemis} $\Longrightarrow$ \ref{propositionequivalencesalmostsemigreedytqg+wsd} is the counterpart of Corollary~\ref{corollaryPropCwA}.\\ To prove \ref{propositionequivalencesalmostsemigreedydisjoint} $\Longrightarrow$ \ref{propositionequivalencesalmostsemigreedyalmostsemiall+}, fix $0<s\le 1$, $x\in {\mathbb X}$, $m\in \mathbb{N}$, $A\in {\mathcal G}(x,m,s)$, and $B\in \mathbb{N}^{<\infty}$ so that $w(B)\le w(A)$. Let $E:=A\cap B$, $A_1:=A\setminus B$, $B_1:=B\setminus A$ and $D\subset A$ so that $\|x-P_D(x)\|\le \|x-P_S(x)\|$ for all $S\subset A$. Since $A_1\in {\mathcal G}(x-P_E(x),|A_1|,s)$ and $w(B_1)\le w(A_1)$, we have $$ \|x-P_{D}(x)\|\le \|x-P_{E}(x)\|\le \mathbf{C}(s)\|x-P_{E}(x)-P_{B_1}(x-P_E(x))\|=\|x-P_B(x)\|. $$ Now the result follows by taking infimum. \\ To prove \ref{propositionequivalencesalmostsemigreedytqg+wd} $\Longrightarrow$ \ref{propositionequivalencesalmostsemigreedytqg+wsd}, suppose that ${\mathcal B}$ is $\mathbf{C}_1$-truncation quasi-greedy and $\mathbf{C}_2$-$\mathbf{w}$-democratic, and fix $A,B\in \mathbb{N}^{<\infty}$ with $w(A)\le w(B)$, $\varepsilon\in \ensuremath{\mathcal{E}}_A$, $\varepsilon'\in \ensuremath{\mathcal{E}}_B$. By Remark~\ref{remarkpropc=tqg}, $$ \|{\mathbf 1}_{\varepsilon,A}\|\le 2\kappa \mathbf{C}_1\|{\mathbf 1}_{A}\|\le 2\kappa \mathbf{C}_1\mathbf{C}_2\|{\mathbf 1}_{B}\|\le 4\kappa^2\mathbf{C}_1^2\mathbf{C}_2\|{\mathbf 1}_{\varepsilon', B}\|. $$ Finally, set ${\mathcal B}^*=({\bm x}^*_i)_{i\in \mathbb{N}}$ the dual basis of ${\mathcal B}$ to prove \ref{propositionequivalencesalmostsemigreedytqg+wsd}$\Longrightarrow$\ref{propositionequivalencesalmostsemigreedydisjoint}. Suppose ${\mathcal B}$ is $\mathbf{C}_1$-truncation quasi-greedy and $\mathbf{C}_2$-$\mathbf{w}$-superdemocratic, and choose $s,x,m, B$ and $A$ as in the statement. Let $A_1$ be a greedy set for $x-P_B(x)$ containing $A$ of minimum cardinality. Given that $w(B)\le w(A_1)$, $A\in {\mathcal G}(x-P_B(x),m,s)$ and $$ \min_{i\in A}|{\bm x}_i^*(x)|=\min_{i\in A}|{\bm x}_i^*(x-P_B(x))|=\min_{i\in A_1}|{\bm x}_i^*(x-P_B(x))|, $$ we have \begin{align*} \|x\|\le& \|x-P_B(x)\|+\|P_B(x)\|\le \|x-P_B(x)\|+\max_{i\in B}|{\bm x}_i^*(x)|\max_{\varepsilon\in \ensuremath{\mathcal{E}}_{B}}\|{\mathbf 1}_{\varepsilon,B}\|\\ \le& \|x-P_B(x)\|+s^{-1}\min_{i\in A_1}|{\bm x}_i^*(x-P_B(x))|\mathbf{C}_2\|{\mathbf 1}_{\varepsilon(x-P_B(x)), A_1}\|\\ \le&(1+s^{-1}\mathbf{C}_1\mathbf{C}_2)\|x-P_B(x)\|. \end{align*} \end{proof} \begin{remark}\label{remarkonlymarkushevich}\rm Note that the proof of the implication \ref{propositionequivalencesalmostsemigreedyalmostsemis} $\Longrightarrow$ \ref{propositionequivalencesalmostsemigreedytqg+wsd} of Proposition~\ref{propositionequivalencesalmostsemigreedy} is the only one in which the totality condition on ${\mathcal B}^*$ is used. For that implication, the condition is essential: indeed, \cite[Example 4.5]{BL2021} shows that even semi-greedy systems need not be democratic. In fact, it is easily shown that the system of \cite[Example 4.5]{BL2021} is not truncation quasi-greedy, either, or even unconditional for constant coefficients (see \cite[Definition 3]{W2000}). \end{remark} Next, we give an analogue of Lemma~\ref{lemmafinitesupportisenough} for weak weight-almost semi-greedy bases, which is proved in a similar manner. \begin{lemma}\label{lemmafinitesupportisenoughasg}Let ${\mathcal B}$ be a basis for ${\mathbb X}$, $\mathbf{w}$ a weight and $0<s\le 1$. Suppose that there is $\mathbf{C}>0$ such that the conditions of Definition~\ref{definitionweakweightalmostsemigreedy} hold for $x$ with finite support. Then ${\mathcal B}$ is $\mathbf{C}$-$s$-$\mathbf{w}$-almost semi-greedy. Moreover, the conditions of Definition~\ref{definitionweakweightalmostsemigreedy} hold even if the infimum is taken without the restriction $|B|<\infty$. \end{lemma} \begin{proof} Pick $x\in {\mathbb X}\setminus \{0\}$ and $m\in \mathbb{N}$. As in the proof of Lemma~\ref{lemmafinitesupportisenough}, we assume $|\supp(x)|\ge m$ and, for every $n\in \mathbb{N}$, we choose $x_n\in {\mathbb X}$ with finite support such that $\|x-x_n\|\le n^{-1}$ and ${\mathcal G}(x_n,m,t)={\mathcal G}(x,m,t)$. By hypothesis, for each $n$ there are $A_n\in {\mathcal G}(x,m,t)$ and $y_n\in {\mathbb X}$ with $\supp(y_n)\subset A_n$ such that $$ \|x_n-y_n\|\le \mathbf{C}\inf_{\substack{D\in \mathbb{N}^{<\infty}\\mathbf{w}(D)\le w(A_n)}}\|x_n-P_D(x_n)\|. $$ Passing to a subsequence, we assume $A_n=A$ and there is $y$ supported in $A$ such that $\|y_n-y\|\le n^{-1}$ for all $n$. Given $B\in \mathbb{N}^{<\infty}$ with $w(B)\le w(A)$, \begin{align*} \|x-y\|\le&2n^{-1}+\|x_n-y_n\|\le 2n^{-1}+\mathbf{C}\|x_n-P_B(x_n)\|\\ \le& n^{-1}(2+\mathbf{C}+\mathbf{C}\|P_B\|)+\mathbf{C}\|x-P_B(x)\|. \end{align*} As this holds for every $n\in \mathbb{N}$, it follows that $\|x-y\|\le \mathbf{C}\|x-P_B(x)\|$. Thus, taking infimum over all such $B$, we obtain that ${\mathcal B}$ is $\mathbf{C}$-$\mathbf{w}$-almost semi-greedy. \\ Now set $x, m$ as before, choose $A\in {\mathcal G}(x,m,t)$ and $y\in {\mathbb X}$ with $\supp(y)\subset A$ so that \eqref{definitionweakweightalmostsemigreedyb} holds, and suppose there is $B\subset \mathbb{N}$ with $|B|=\infty$ and $w(B)\le w(A)$. By Proposition~\ref{propositionequivalencesalmostsemigreedy}, ${\mathcal B}$ is $\mathbf{w}$-superdemocratic. Hence, by Remarks~\ref{remarkprojectionc0} and~\ref{remarkl1wdem}, $P_B(x)$ is well-defined and, for each $n\in \mathbb{N}$, there is $B_n\subset B$ such that $B_n\in \mathbb{N}^{<\infty}$ and $\|P_B(x)-P_{B_n}(x)\|\le n^{-1}$. Thus, \begin{align*} \|x-y\|\le \mathbf{C}\|x-P_{B_n}(x)\|\le \mathbf{C}\|x-P_{B}(x)\|+\mathbf{C} n^{-1}. \end{align*} As $n$ is arbitrary, we conclude that $\|x-y\|\le \mathbf{C}\|x-P_{B}(x)\|$, and the proof is completed by taking supremum. \end{proof} In many cases equivalent weights $\mathbf{w} \approx \mathbf{w}'$ (those satisfying $\mathbf{C}_1 w_n\le w_n'\le \mathbf{C}_2 w_n$ for all $n\in \mathbb{N}$, for some $0<\mathbf{C}_1 \le \mathbf{C}_2<\infty$) produce the same weighted bases. This happens for the $\mathbf{w}$-Property (A), $\mathbf{w}$-superdemocracy and $\mathbf{w}$-democracy which are equivalent to their respective $\mathbf{w}'$-counterparts (see \cite[Proposition 3.5, Remark 3.6]{BDKOW2019}). This extends to $\mathbf{w}$-almost greedy bases characterized as quasi-greedy and $\mathbf{w}$-democratic \cite[Theorem 2.6]{DKTW2018}. Proposition~\ref{propositionequivalencesalmostsemigreedy} combined with the aforementioned result for $\mathbf{w}$-democracy gives the following natural result. \begin{corollary}\label{corollaryequivalentweightsalmostsemi}Let ${\mathcal B}$ be a basis for ${\mathbb X}$, $0<s\le 1$. If $\mathbf{w}$ and $\mathbf{w}'$ are equivalent weights, then ${\mathcal B}$ is $s$-$\mathbf{w}$-almost semi-greedy if and only if it is $s$-$\mathbf{w}'$-almost semi-greedy. \end{corollary} In the case of constant weights, Proposition~\ref{propositionequivalencesalmostsemigreedy} sheds light on some of the connections between the Chebyshevian Thresholding Greedy algorithm with bidemocratic and squeeze symmetric bases, which are defined below. Recall that squeeze symmetric bases are those that can be sandwiched between two symmetric bases of quasi-Banach spaces with equivalent fundamental functions (see \cite{AAB2021} and \cite{AABW2021} for further definitions and results). \begin{definition}\label{definitionsqueezesymmetric}Let ${\mathcal B}$ be a basis for ${\mathbb X}$. We say that ${\mathcal B}$ is \emph{squeeze symmetric} if there are quasi-Banach spaces ${\mathbb Y}$, ${\mathbb Z}$ with symmetric bases ${\mathcal B}_1=({\bm y}_n)_{n\in\mathbb{N}}$ and ${\mathcal B}_2=({\bm z}_n)_{n\in\mathbb{N}}$ and bounded linear operators $T: {\mathbb Y}\rightarrow {\mathbb X}$ and $S: {\mathbb X}\rightarrow {\mathbb Z}$ such that by $T({\bm y}_n)={\bm x}_n$ and $S({\bm x}_n)={\bm z}_n$, and there is $\mathbf{C}>0$ such that $$ \|{\mathbf 1}_{\varepsilon,A}[{\mathcal B}_1]\|\le \mathbf{C}\|{\mathbf 1}_{\varepsilon,A}[{\mathcal B}_2]\|,\qquad\forall A\in \mathbb{N}^{<\infty},\ \forall \varepsilon\in \ensuremath{\mathcal{E}}_A. $$ \end{definition} It was proven in \cite[Proposition 9.4, Corollary 9.15]{AABW2021} that a basis is squeeze symmetric if and only if it is truncation quasi-greedy and superdemocratic (or democratic, \cite[Proposition 4.16]{AABW2021}). In our context, that is when ${\mathcal B}$ is Markushevich basis for a Banach space ${\mathbb X}$, a combination of the aforementioned results with Proposition~\ref{propositionequivalencesalmostsemigreedy} gives the following. \begin{corollary}Let ${\mathcal B}$ be a basis for ${\mathbb X}$. The following are equivalent: \begin{itemize} \item ${\mathcal B}$ is almost semi-greedy. \item ${\mathcal B}$ is truncation quasi-greedy and superdemocratic. \item ${\mathcal B}$ is truncation quasi-greedy and democratic. \item ${\mathcal B}$ is squeeze symmetric. \end{itemize} \end{corollary} Bidemocratic basis were introduced in \cite{DKKT2003} to study duality in connection to the TGA. A central result is that a quasi-greedy basis ${\mathcal B}$ is bidemocratic if and only if both ${\mathcal B}$ and ${\mathcal B}^*$ are almost greedy, or partially greedy \cite[Theorem 5.4]{DKKT2003}. \begin{definition}\label{definitionbidemocratic}A basis ${\mathcal B}$ is \emph{bidemocratic} if there is $\mathbf{C}>0$ such that, for each $m\in \mathbb{N}$, \begin{align*} \sup_{\substack{A\subset \mathbb{N}\\|A|\le m}}\sup_{\substack{B\subset \mathbb{N}\\|B|\le m}}\|{\mathbf 1}_{A}\|\|{\mathbf 1}_{B}^{*}\| \le& \mathbf{C} m. \end{align*} \end{definition} Bidemocratic bases have also been studied for their own sake in \cite{AABBL2021}, where it was proven that a basis ${\mathcal B}$ for a Banach space ${\mathbb X}$ is bidemocratic if and only if both ${\mathcal B}$ and ${\mathcal B}^*$ are truncation quasi-greedy and (super)democratic \cite[Corollary 2.6]{AABBL2021}. In the case of Markushevich bases, we can combine this result with Proposition~\ref{propositionequivalencesalmostsemigreedy} to obtain the following. \begin{corollary}\label{corollarybidem}A basis ${\mathcal B}$ is bidemocratic if and only if both ${\mathcal B}$ and ${\mathcal B}^*$ are almost semi-greedy. \end{corollary} \section{Weak weight-almost greedy bases} \label{sectionweightweakalmostgreedy} It is known that in the standard case -equivalently, in the case of constant weights-, almost greediness with respect to the weak algorithm is equivalent to almost greediness \cite[Proposition 2.3]{BL2021} and \cite[Theorem 6.4]{DKSW2012}. Thus, it is natural to ask whether this equivalence has an extension to general weights. In this section, we answer this question in the affirmative. To do so, we introduce and briefly study a weighted extension of the notion of weak almost greedy bases from \cite{BL2021} - which can also be seen as a weak-algorithm extension of the concept of weight-almost greedy bases from \cite{DKTW2018} (see remarks after Definition~\ref{definitionweightsemigreedyetc}). \begin{definition}\label{definitionweakweightalmostgreedy} Let ${\mathcal B}$ be a basis for ${\mathbb X}$, $\mathbf{w}$ a weight, $\mathbf{C}>0$, and $0<s\le 1$. We say that ${\mathcal B}$ is \emph{weak weight almost greedy with parameter $s$, weight $\mathbf{w}$ and constant $\mathbf{C}$} (or $\mathbf{C}$-$s$-$\mathbf{w}$-almost-greedy) if, for every $x\in {\mathbb X}$ and $m\in \mathbb{N}$, there is $A\in {\mathcal G}(x,m,s)$ such that \begin{equation} \|x-P_A(x)\|\le \mathbf{C}\inf_{\substack{B\in \mathbb{N}^{<\infty}\\mathbf{w}(B)\le w(A)}}\|x-P_B(x)\|.\nonumbe \end{equation} We denote by ${\mathcal {GA}}(x,m,s)$ the subset of ${\mathcal G}(x,m,s)$ for which the above bound holds. \end{definition} \smallskip \begin{remark}\rm \label{remarkdisjointsuperdem} First, note that every $\mathbf{C}$-$s$-$\mathbf{w}$-almost-greedy is $\mathbf{C} s^{-1}$-$\mathbf{w}$-disjoint superdemocratic. Indeed, given disjoint nonempty sets $A,B\in \mathbb{N}^{<\infty}$ with $w(A)\le w(B)$, $\varepsilon\in \ensuremath{\mathcal{E}}_A, \varepsilon'\in \ensuremath{\mathcal{E}}_B$, for every $\epsilon>0$ we have $$ \|{\mathbf 1}_{\varepsilon, A}\|=\|{\mathbf 1}_{\varepsilon, A}+(1+\epsilon)s^{-1}{\mathbf 1}_{\varepsilon',B}-(1+\epsilon)s^{-1}{\mathbf 1}_{\varepsilon',B}\|\le \mathbf{C}(1+\epsilon)s^{-1}\|{\mathbf 1}_{\varepsilon',B}\|, $$ where we used that ${\mathcal G}({\mathbf 1}_{\varepsilon, A}+(1+\epsilon)s^{-1}{\mathbf 1}_{\varepsilon',B}, |B|,s)=\{B\}$. Finally, as in addition ${\mathcal B}$ is $\mathbf{w}$- semi greedy, Proposition~\ref{propositionupperbounds} applies to these bases. In particular, if $\mathbf{w}\in \ell_1$, then ${\mathcal B}$ is equivalent to the canonical unit vector basis of $\mathtt{c}_{0}$. \end{remark} To prove that $\mathbf{w}$-$s$-almost greedy bases are quasi-greedy, we need an auxiliary result analogous to Lemma~\ref{lemmasgset}. \begin{lemma}\label{lemmaagset}Let ${\mathcal B}$ be a $\mathbf{C}$-$s$-$\mathbf{w}$-almost greedy basis for ${\mathbb X}$ and let ${\mathcal B}^*=( {\bm x}^*_i)_{i\in\mathbb{N}}$ be the dual basis of ${\mathcal B}$. For every $x\in {\mathbb X}$ and every nonempty set $A\in \mathbb{N}^{<\infty}$, there is $m\in \mathbb{N}$ and $E\in {\mathcal {GA}}(x,m,s)$ such that $A\subset E$ and for all $j\in E$, $$ |{\bm x}_j^*(x)|\ge s^{2} \min_{i\in A}|{\bm x}_i^*(x)|. $$ \end{lemma} \begin{proof} This is proven by the same argument as Lemma~\ref{lemmasgset}. \end{proof} Now we can prove the main result of this section. The proof is based on that of \cite[Proposition 2.3]{BL2021}, which in turn is based on the arguments of the proof of \cite[Proposition 4.4]{DKSW2012}. \begin{theorem}\label{theoremswalmostgreedy}Let ${\mathcal B}=({\bm x}_i)_{i\in\mathbb{N}}$ be a $\mathbf{C}$-$s$-$\mathbf{w}$-almost greedy basis for ${\mathbb X}$, with $\mathbf{C}>0$, $0<s\le 1$ and $\mathbf{w}$ a weight such that $\mathbf{w}\not \in \ell_1$, and set $$ \mathbf{K}:=\mathbf{C} s^{-1} \max\{2\mathbf{C} s^{-1},\lambda \lambda '\}. $$ The following hold: \begin{enumerate}[\upshape (i)] \item \label{whyperdem} For all $A,B \in \mathbb{N}^{<\infty}$ with $w(A)\le w(B)$, if $(a_i)_{i\in A}$, $(b_i)_{i\in B}$ are scalars such that $$ \max_{i\in A}|a_i|\le \min_{i\in B}|b_i|, $$ then $$ \|\sum_{i\in A}a_i{\bm x}_i\|\le \mathbf{K}\|\sum_{i\in B}b_i{\bm x}_i\|. $$ \item \label{qg} For every $0<t\le 1$, $x\in {\mathbb X}$, $m\in \mathbb{N}$ and $A\in {\mathcal G}(x,m,t)$, $$ \|P_A(x)\|\le (1+ t^{-1}s^{-2}\mathbf{K})(1+\mathbf{C}) \|x\|. $$ In particular, ${\mathcal B}$ is quasi-greedy with constant as above taking $t=1$. \end{enumerate} \end{theorem} \begin{proof} To prove \ref{whyperdem}, fix $A,B, (a_i)_{i\in A}$, $(b_i)_{i\in B}$ as in the statement with $$ a:=\max_{i\in A}|a_i|>0, $$ and consider two cases: \\ First, suppose there is $j>A$ such that $ w_{j}\ge w(A)$. As $a\le \lambda'\|\sum_{i\in B}b_i{\bm x}_i\|$, by Remark~\ref{remarkdisjointsuperdem} and convexity \begin{equation} \|\sum_{i\in A}a_i{\bm x}_i\|\le a \sup_{\varepsilon\in \ensuremath{\mathcal{E}}_A} \|{\mathbf 1}_{\varepsilon,A}\|\le a s^{-1}\mathbf{C}\|{\bm x}_{j}\|\le a s^{-1}\mathbf{C}\lambda \le s^{-1}\mathbf{C}\lambda \lambda '\|\sum_{i\in B}b_i{\bm x}_i\|.\label{lightweight9v2} \end{equation} On the other hand, if there is no such $j$, since $\mathbf{w}\not\in \ell_1$ we can choose two finite sets of positive integers $A_2>A_1>A \cup B$ so that $$ \max\{w(A_1),w(A_2)\}\le w(A)\le w(A_1)+w(A_2). $$ As before, by Remark~\ref{remarkdisjointsuperdem}, convexity and the triangle inequality, \begin{equation} \|\sum_{i\in A}a_i{\bm x}_i\|\le a s^{-1}\mathbf{C}\|{\mathbf 1}_{A_1}+{\mathbf 1}_{A_2}\|\le a s^{-1}\mathbf{C}\|{\mathbf 1}_{A_1}\|+ a s^{-1}\mathbf{C}\|{\mathbf 1}_{A_2}\|.\label{oneside6v2} \end{equation} Fix $0<\epsilon<1$, and set $$ z_1:=\sum_{i\in B}b_i{\bm x}_i +(1-\epsilon)a s {\mathbf 1}_{A_1}. $$ Given that ${\mathcal G}(z_1,|B|,s)=\{B\}$ and $w(A_1)\le w(B)$, we have $$ \|(1-\epsilon)a s {\mathbf 1}_{A_1}\|=\|z_1-P_B(z_1)\|\le \mathbf{C}\|z_1-P_{A_1}(z_1)\|=\mathbf{C}\|\sum_{i\in B}b_i{\bm x}_i\|. $$ By the same argument, the above bound also holds for $\|(1-\epsilon)a s {\mathbf 1}_{A_2}\|$. From this and \eqref{oneside6v2} we deduce that $$ \|\sum_{i\in A}a_i{\bm x}_i\|\le 2\mathbf{C}^2s^{-2}(1-\epsilon)^{-1}\|\sum_{i\in B}b_i{\bm x}_i\|. $$ As $\epsilon$ is arbitrary, the proof of \ref{whyperdem} is completed combining the above estimate and \eqref{lightweight9v2}. \\ Now, set ${\mathcal B}^*=({\bm x}^*_i)_{i\in\mathbb{N}}$ the dual basis of ${\mathcal B}$. Fix $t$, $x$, $m$ and $A$ as in \ref{qg}, assuming $x\not=P_A(x)$, and set $$ a:=\min_{i\in A}|{\bm x}_i^*(x)|. $$ By Lemma~\ref{lemmaagset}, there is $A\subset D\in {\mathcal {GA}}(x,|D|,s)$ such that $$ \min_{i\in D}|{\bm x}_i^*(x)|\ge s^2a. $$ Let $B:=D\setminus A$. If $B=\emptyset$, then $\|P_A(x)\|\le \|x\|+\|x-P_D(x)\|\le (\mathbf{C}+1)\|x\|$. Otherwise, $$ \max_{i\in B}|{\bm x}_i^*(x)|\le t^{-1}a\le t^{-1}s^{-2}\min_{i\in D}\|{\bm x}_i^*(x)\|. $$ Hence, by \ref{whyperdem} $$ \|P_B(x)\|\le t^{-1}s^{-2}\mathbf{K}\|P_D(x)\|. $$ Thus, by the triangle inequality, \begin{align*} \|P_A(x)\|\le& \|P_B(x)\|+\|P_D(x)\|\le (1+ t^{-1}s^{-2}\mathbf{K})\|P_D(x)\|\\ \le& (1+ t^{-1}s^{-2}\mathbf{K})(\|x\|+\|x-P_D(x)\|)\le (1+ t^{-1}s^{-2}\mathbf{K})(1+\mathbf{C})\|x\|. \end{align*} \end{proof} Finally, we have: \begin{remark}\label{remarkequivalentweights}\rm If $\mathbf{w}\approx \mathbf{w}'$ and $0<s\le 1$, then a basis ${\mathcal B}$ is $s$-$\mathbf{w}$-almost greedy if and only if it is $s$-$\mathbf{w}'$-almost greedy. This follows by Theorem~\ref{theoremswalmostgreedy} and the remarks before Corollary~\ref{corollaryequivalentweightsalmostsemi}. \end{remark} \section{Lebesgue and Chebyshevian Lebesgue parameters}\label{sectionquasi-greedyalmostgreedy} In this section, we study some parameters that involve the (weak) TGA and CGA, and also improve the known bounds for the quasi-greedy and almost greedy constants of weak semi-greedy bases. We will use the following auxiliary definitions. \begin{definition}\label{definitiontquasigreedyparameter}Let ${\mathcal B}$ be a basis for ${\mathbb X}$, $0<t\le 1$ and $m\in\mathbb{N}$. \begin{itemize} \item The \emph{$t$-quasi-greedy parameter} $\overline{{\mathbf g}}(m,t)$ is defined by \begin{equation} \overline{{\mathbf g}}(m,t):=\inf_{\mathbf{C}>0} \Big\{ \|P_A(x)\|\le \mathbf{C}\|x\|, \qquad\forall x\in {\mathbb X},\; \forall A \in {\mathcal G}(x,m,t)\Big\}. \nonumber \end{equation} \item The \emph{suppression $t$-quasi-greedy parameter} $\widehat{{\mathbf g}}(m,t)$ is defined by \begin{equation} \widehat{{\mathbf g}}(m,t):=\inf_{\mathbf{C}>0}\ \Big\{ \|x-P_A(x)\|\le \mathbf{C}\|x\|, \qquad\forall x\in {\mathbb X},\; \forall A \in {\mathcal G}(x,m,t)\Big\}. \nonumber \end{equation} \end{itemize} \end{definition} \noindent For $t=1$, the parameter $\overline{{\mathbf g}}(m,1)$ has been considered in \cite{AABBL2021}, and the quasi-greedy parameter given by ${\mathbf g}_m:=\max_{n\le m}\overline{{\mathbf g}}(n,1)$ has been studied as well, for example in \cite{BBGHO2018}. Next we define the Chebyshevian Lebesgue parameters associated to weak tresholding greedy algorithms. \begin{definition}\label{definitionChebyshevianlebesguetype} Let ${\mathcal B}$ be a basis for ${\mathbb X}$, $0<t\le 1$ and $m\in \mathbb{N}$. \begin{itemize} \item The \emph{upper Chebyshevian Lebesgue parameter}, $\mathbf{L}_{ch}^u(m,t)=\mathbf{L}_{ch}^u(m,t)[{\mathcal B},{\mathbb X}]$, is given by \begin{eqnarray*} \mathbf{L}_{ch}^u(m,t):=\inf_{\mathbf{C}>0} \left\lbrace \|x-y\|\le \mathbf{C}\sigma_m(x): \begin{array}{ll} \forall x\in {\mathbb X},\ \forall A\in {\mathcal G}(x,m,t),\\ \exists y\in {\mathbb X}: \supp(y)\subset A \end{array} \right\rbrace. \end{eqnarray*} \item The \emph{lower Chebyshevian Lebesgue parameter}, $\mathbf{L}_{ch}^l(m,t)=\mathbf{L}_{ch}^l(m,t)[{\mathcal B},{\mathbb X}]$, is given by \begin{eqnarray*} \mathbf{L}_{ch}^l(m,t):=\inf_{\mathbf{C}>0} \left\lbrace \|x-y\|\le \mathbf{C}\sigma_m(x): \begin{array}{ll} \forall x\in {\mathbb X},\ \exists A\in {\mathcal G}(x,m,t),\\ \exists y\in {\mathbb X}: \supp(y)\subset A \end{array} \right\rbrace. \end{eqnarray*} \end{itemize} As usual, to simplify our notation we leave the space and basis implicit when clear. \end{definition} \begin{remark}\rm Under the conditions of Definition~\ref{definitionChebyshevianlebesguetype}, it is clear that there is a smallest $\mathbf{C}>0$ in the definition of $\mathbf{L}_{ch}^{u}(m,t)$, so the parameter is attained. The same holds for $\mathbf{L}_{ch}^{l}(m,t)$, since for each $x\in {\mathbb X}$ with $|\supp(x)|>m$, the set ${\mathcal G}(x,m,t)$ is finite. \end{remark} \begin{remark}\rm The upper Chebyshevian Lebesgue parameter is the Chebyshevian Lebesgue constant introduced in \cite{DKO2015} and also studied in \cite{BBGHO2018}. \end{remark} Next, we study a relation between $\widehat{{\mathbf g}}$, $\mathbf{M}_{fs}$ and $\mathbf{L}_{ch}^l$ which, in particular gives a slight improvement of the known bound for the suppression quasi-greedy constant of semi-greedy bases. More precisely, \cite[Theorem 4.2]{BL2021} gives that if ${\mathcal B}$ is $\mathbf{C}$-$s$-weak semi-greedy, it is $\mathbf{K}$-suppression quasi-greedy with $\mathbf{K}\le \mathbf{M}_{fs}\mathbf{C}(1+(\mathbf{M}_{fs}+1)\mathbf{C} s^{-2})$, whereas Theorem~\ref{theoremswemigreedydisjointgreedynotc0} avoids the quadratic factor $\mathbf{M}_{fs}^2$ at the cost of involving $\lambda$ and $\lambda'$, and gives $\mathbf{K}\lesssim \mathbf{M}_{fs}\mathbf{C} s^{-1} \max\{\mathbf{C} s^{-2}, \lambda \lambda'\}$. Our next result gives $\mathbf{K}\le \mathbf{M}_{fs}\mathbf{C}(1+ 2\mathbf{C} s^{-2})$. \begin{proposition}\label{propositionchevyshevianqg}Let ${\mathcal B}$ be a basis for ${\mathbb X}$. For all $0<s\le 1$, $0<t\le 1$ and $m\in \mathbb{N}$, we have the following estimates: \begin{equation} \widehat{{\mathbf g}}(m,t)\le \mathbf{M}_{fs} \mathbf{L}_{ch}^l(2\floor{\frac{m+1}{2}},s) (1+ 2\mathbf{L}_{ch}^l(2\floor{\frac{m+1}{2}},s) t^{-1}s^{-2}). \label{semigreedytoquasigreedy} \end{equation} Hence, if ${\mathcal B}$ is $\mathbf{C}$-$s$-weak semi-greedy, it is $\mathbf{M}_{fs}\mathbf{C}(1+2\mathbf{C} s^{-2})$-suppression quasi-greedy. \end{proposition} \begin{proof} Set ${\mathcal B}=( {\bm x}_i)_{i\in\mathbb{N}}$ and ${\mathcal B}^*=( {\bm x}^*_i)_{i\in\mathbb{N}}$. Choose $0<\epsilon<1$, and let $({\bm x}_{i_k})_{k\in \mathbb{N}}$ be a subsequence given by an application of Corollary~\ref{corollaryboth} to ${\mathcal B}$ and $({\mathbb X}, \mathbf{M}_{fs}, \epsilon)$. \\ Fix $x\in {\mathbb X}$ with finite support, and $A \in {\mathcal G}(x,m,t)$. We may assume $x\not=P_A(x)$. Set $$ a:=\max_{i\not\in A}|{\bm x}_i^{*}(x)|, $$ and pick $i_0>\supp(x)$. Set \begin{align*} n:=&\floor{\frac{m+1}{2}};\\ {\mathbb F}:=&[{\bm x}_i: 1\le i\le i_0];\\ E_1:=&\{i_{2(s_{{\mathbb F},\epsilon}+j)-1}:1\le j\le n\};\\ E_2:=&\{i_{2(s_{{\mathbb F},\epsilon}+j)}:1\le j\le n\};\\ z:=&x+ ats(1-\epsilon)\left({\mathbf 1}_{E_1}-{\mathbf 1}_{E_2}\right). \end{align*} Note that $s_{{\mathbb F},\epsilon}> i_0$ (othewise, ${\bm x}_{s_{{\mathbb F},\epsilon}}\in {\mathbb F}$ and $\|{\bm x}_{s_{{\mathbb F},\epsilon}}\|\le (\mathbf{M}_{fs}+\epsilon)\|{\bm x}_{s_{{\mathbb F},\epsilon}}-{\bm x}_{s_{{\mathbb F},\epsilon}}\|=0)$. So, the sets $E_1$, $E_2$ and $A$ are pairwise disjoint. Since $A\in {\mathcal G}(x,m,t)$, we have $$ |\{1\le i\le i_0 : |{\bm x}_i^{*}(z)|\ge at \}|\ge |A|+1=m+1\ge 2n. $$ Thus, $$ D \subset \{1,\dots,i_0\},\qquad \forall D \in {\mathcal G}(z,2n,s). $$ It follows that there is $u\in {\mathbb F}$ such that $$ \|z-u\|\le \mathbf{L}_{ch}^l(2n,s) \sigma_{2n}(z)\le \mathbf{L}_{ch}^l(2n,s)\|x\|. $$ Hence, by Corollary~\ref{corollaryboth}, \begin{align} ats(1-\epsilon)\|{\mathbf 1}_{E_1}-{\mathbf 1}_{E_2}\|\le& \|z-u\|+\|x-u\|\le (2+\epsilon)\|z-u\|\nonumber\\ \le&(2+\epsilon)\mathbf{L}_{ch}^l(2n,s)\|x\|. \label{newpart} \end{align} Now set $$ y:=x-P_A(x)-s^{-1}(1+\epsilon)a \left({\mathbf 1}_{E_1}-{\mathbf 1}_{E_2}\right). $$ Given that $$|{\bm x}_i^*(x-P_A(x)) |\le a,\qquad\forall i \in \mathbb{N},$$ we have ${\mathcal G}(y,2n,s)=\left\lbrace E_1\cup E_2\right\rbrace$. Thus, there is $v\in {\mathbb X}$ with $\supp(v)\subset E_1\cup E_2$ such that \begin{align} \|y-v\|\le& \mathbf{L}_{ch}^l(2n,s) \sigma_{2n}(y)\nonumber\\ \le& \mathbf{L}_{ch}^l(2n,s) \|x\|+ \mathbf{L}_{ch}^l(2n,s) s^{-1}(1+\epsilon)a \| {\mathbf 1}_{E_1}-{\mathbf 1}_{E_2}\|.\label{oldpart} \end{align} Again, by Corollary~\ref{corollaryboth}, combining \eqref{newpart} and \eqref{oldpart} we obtain \begin{align*} \|x-P_A(x)\|&\le (\mathbf{M}_{fs} +\epsilon)\|y-v\|\\ &\le (\mathbf{M}_{fs} +\epsilon) \mathbf{L}_{ch}^l(2n,s) (1+\mathbf{L}_{ch}^l(2n,s) t^{-1}s^{-2}(1+\epsilon)(2+\epsilon)(1-\epsilon)^{-1})\|x\|. \end{align*} As $\epsilon$ is arbitrary, the proof of \eqref{semigreedytoquasigreedy} is complete for $x$ with finite support. The general case is handled as in the proof of Theorem~\ref{theoremswemigreedydisjointgreedynotc0}. \\Now if ${\mathcal B}$ is $\mathbf{C}$-$s$-weak semi-greedy, then \begin{align*} \sup_{m\in \mathbb{N}}\widehat{{\mathbf g}}(m,1)\le& \mathbf{M}_{fs}\sup_{\substack{n\in \mathbb{N}\\n\text{ is even}}}\mathbf{L}_{ch}^l(n,s)(1+2\sup_{\substack{n\in \mathbb{N}\\n\text{ is even}}}\mathbf{L}_{ch}^l(n,s) s^{-2}) \le \mathbf{M}_{fs}\mathbf{C}(1+2\mathbf{C} s^{-2}). \end{align*} \end{proof} \begin{remark}\label{remarklinearqg}\rm Note that the linear factor $\mathbf{M}_{fs}$ in \eqref{semigreedytoquasigreedy} is necessary. Indeed, \cite[Example 4.4]{BL2021} shows that for each $\alpha>0$, there is a basis ${\mathcal B}$ for a subspace ${\mathbb X}$ of $\ell_1$ that has quasi-greedy constant $\mathbf{C}_{\alpha}> \alpha$ and semi-greedy constant $\mathbf{K}_{\alpha}\le 4$; the first computation of the proof implies that ${\mathcal B}$ is a Schauder basis equivalent to the canonical unit vector basis of $\ell_1$ with basis constant no greater than $3+2\alpha$, and the constant $\mathbf{M}_{fs}$ is no greater than the basis constant by Proposition~\ref{propositionseparation}\ref{schauder}. \end{remark} Next, we consider the following Lebesgue-type parameters. \begin{definition}\label{definitionlebesguedisjoint}Let ${\mathcal B}$ be a basis for ${\mathbb X}$, $0<t\le 1$ and $m\in \mathbb{N}$. \begin{itemize} \item The \emph{Lebesgue parameter}, $\mathbf{L}(m,t)=\mathbf{L}(m,t)[{\mathcal B},{\mathbb X}]$, is given by \begin{eqnarray*} \mathbf{L}(m,t):=\inf_{\mathbf{C}>0} \left\lbrace \|x-P_A(x)\|\le \mathbf{C} \|x-y\|: \begin{array}{ll} \forall x\in {\mathbb X},\ \forall A\in {\mathcal G}(x,m,t),\\ \forall y\in {\mathbb X}: |\supp(y)|\le m \end{array} \right\rbrace. \end{eqnarray*} \item The \emph{disjoint Lebesgue parameter}, $\mathbf{L}_d(m,t) =\mathbf{L}_d(m,t)[{\mathcal B},{\mathbb X}]$, is given by $$ \mathbf{L}_d(m,t):=\displaystyle \inf_{\mathbf{C}>0} \left\lbrace \|x-P_A(x)\|\le \mathbf{C} \|x-y\|: \begin{array}{ll} \forall x\in {\mathbb X},\ \forall A\in {\mathcal G}(x,m,t), \forall y\in {\mathbb X}:\\ |\supp(y)|\le m,\ \supp(y)\cap A=\emptyset \end{array} \right\rbrace. $$ \end{itemize} \end{definition} The Lebesgue parameter for $t=1$ has been widely been studied, in the literature, for example in \cite{BBG2017}, \cite{BBGHO2018}, \cite{GHO2013}, \cite{Oswald2001}, among others. The parameter involving the WTGA for any $0<t\le 1$ has been studied for example in \cite{DKO2015} and \cite{BBGHO2018}. The disjoint parameters are weaker variants suggested by Theorem~\ref{theoremswemigreedydisjointgreedynotc0}. Our next result gives upper bounds for the disjoint Lebesgue parameter in terms of the lower Chebyshevian Lebesgue parameter. \begin{proposition}\label{propositionsemigreedydisjointgreedy}Let ${\mathcal B}$ be a basis for ${\mathbb X}$ and $\mathbf{M}:=\mathbf{M}_{fs}({\mathcal B}, {\mathbb X})$. For all $0<s\le 1$, $0<t\le 1$ and $m\in \mathbb{N}$, we have the following estimates: \begin{enumerate}[\rm (i)] \item \label{lebesgue1} $\mathbf{L}_d(m,t)\le \mathbf{M}\mathbf{L}_{ch}^l(2m,s)(1+2(\mathbf{M}+1)\mathbf{L}_{ch}^l(m,s)t^{-1}s^{-2})$. \item \label{lebesgue1v} If $m$ is even, in addition we have: $$ \mathbf{L}_d(m,t)\le \mathbf{M}\mathbf{L}_{ch}^l(2m,s)(1+4\mathbf{L}_{ch}^l(m,s)t^{-1}s^{-2}). $$ \item \label{lebesgue1v2} If $m$ is odd, $m>1$, in addition we have: $$ \mathbf{L}_d(m,t)\le \mathbf{M}\mathbf{L}_{ch}^l(2m,s)(1+(4\mathbf{L}_{ch}^l(m-1,s)+2\mathbf{L}_{ch}^{l}(2,s))t^{-1}s^{-2}). $$ \end{enumerate} In particular, if ${\mathcal B}$ is $\mathbf{C}$-$s$-weak semi-greedy, it is $\mathbf{K}$-almost greedy with $$ \mathbf{K}\le\min\left\lbrace \mathbf{M}\mathbf{C}(1+2\mathbf{C}(\mathbf{M}+1) s^{-2}, \max\{ \mathbf{M}\mathbf{C}(1+6\mathbf{C} s^{-2}), 1+\lambda\lambda'+\lambda'' \}\right\rbrace. $$ \end{proposition} \begin{proof} Set ${\mathcal B}=( {\bm x}_i)_{i\in\mathbb{N}}$ and ${\mathcal B}^*=( {\bm x}^*_i)_{i\in\mathbb{N}}$. Choose $0<\epsilon<1$, and let $({\bm x}_{i_k})_{k\in \mathbb{N}}$ be a subsequence given by Corollary~\ref{corollaryboth} for $({\mathbb X}, \mathbf{M}, \epsilon)$. To prove \ref{lebesgue1}, fix $x\in {\mathbb X}$ with finite support and $A\in {\mathcal G}(x,m,t)$. We may assume that $P_A(x)\not=x$, so $|\supp(x)|>m$. Pick $y\in {\mathbb X}$ so that $$ |\supp(y)|\le m\qquad\text{and} \qquad \supp(y)\cap A=\emptyset, $$ set $a:=\min_{i \in A}|{\bm x}_i^*(x)|$ and choose $i_0>\supp(x)\cup \supp(y)$. Now define \begin{align*} {\mathbb F}:=&[ {\bm x}_i: 1\le i\le i_0];\\ E_l:=&\{i_{2s_{{\mathbb F},\epsilon}+j+lm }: 1\le j\le m\}, \quad \forall l\in \{0,1\} ;\\ z:=&x-P_A(x)+a(1+\epsilon)s^{-1}t^{-1}({\mathbf 1}_{E_0}+{\mathbf 1}_{E_1}). \end{align*} We have $$ |{\bm x}_i^*(z)|=|{\bm x}_i^*(x-P_A(x))|\le t^{-1}a,\qquad \forall i\not\in E_0\cup E_1. \label{tgreedycondition0} $$ Thus, $$ {\mathcal G}(z,2m,s)=\{E_0\cup E_1\}. $$ It follows that there is $v\in {\mathbb X}$ with $\supp(v)\subset E_0\cup E_1$ such that \begin{equation*} \|z-v\|\le \mathbf{L}_{ch}^l(2m,s)\sigma_{2m}(z). \end{equation*} Hence, \begin{align} \|x-P_A(x)\| & \le (\mathbf{M}+\epsilon)\|z-v\|\le (\mathbf{M}+\epsilon)\mathbf{L}_{ch}^l(2m,s)\|z+P_A(x)-y\|\nonumber\\ &\le (\mathbf{M}+\epsilon)\mathbf{L}_{ch}^l(2m,s)(\|x-y\|+a(1+\epsilon)s^{-1}t^{-1}\|{\mathbf 1}_{E_0}+{\mathbf 1}_{E_1}\|).\label{leftbound} \end{align} Set $$ u_l:=x-y-a(1-\epsilon)s{\mathbf 1}_{E_l},\qquad \forall l\in \{0,1\}. $$ Given that $\supp(y)\cap A=\emptyset$, we have $$ |\{1\le i\le i_0: |{\bm x}_i^*(x-y)|\ge a\}|\ge |A|= m. $$ It follows that for $l\in \{0,1\}$, every element of ${\mathcal G}(u_l,m,s)$ is contained in $\{1,\dots,i_0\}$, so there is $v_l\in {\mathbb F}$ such that \begin{equation*} \|u_l-v_l\|\le \mathbf{L}_{ch}^l(m,s) \sigma_m(u_l),\qquad \forall l\in \{0,1\}. \end{equation*} This entails that \begin{align} a(1-\epsilon)s\|{\mathbf 1}_{E_l}\|\le& \|u_l-v_l\|+\|x-y-v_l\|\le (1+\mathbf{M}+\epsilon)\|u_l-v_l\|\nonumber\\ \le& (1+\mathbf{M}+\epsilon)\mathbf{L}_{ch}^l(m,s) \sigma_m(u_l)\nonumber\\ \le& (1+\mathbf{M}+\epsilon)\mathbf{L}_{ch}^l(m,s) \|x-y\|,\qquad \forall l\in \{0,1\}. \label{foreven} \end{align} Combining the above estimate with \eqref{leftbound}, we get \begin{align*} & \|x-P_A(x)\|\nonumber\\ &\le (\mathbf{M}+\epsilon)\mathbf{L}_{ch}^l(2m,s)(1+2(1+\mathbf{M}+\epsilon)\mathbf{L}_{ch}^l(m,s)(1+\epsilon)(1-\epsilon)^{-1}s^{-2}t^{-1})\|x-y\|. \end{align*} As $\epsilon$ is arbitrary, the proof of \ref{lebesgue1} is complete.\\ To prove \ref{lebesgue1v}, we use the same argument, with only the following modifications: For $l\in \{0,1\}$ and $n\in \mathbb{N}$, define $\varepsilon_n:=(-1)^{n+1}$, and substitute ${\mathbf 1}_{\varepsilon, E_l}$ for ${\mathbf 1}_{E_l}$ in the definitions of $z$ and $u_l$. \\ When $|E_l|=m$ is even, ${\mathbf 1}_{\varepsilon,E_l}$ is a sum that meets the conditions of Corollary~\ref{corollaryboth}, so instead of \eqref{foreven} we obtain \begin{align*} a(1-\epsilon)s\|{\mathbf 1}_{\varepsilon, E_l}\|\le& \|u_l-v_l\|+\|x-y-v_l\|\le (1+1+\epsilon)\|u_l-v_l\|\nonumber\\ \le& (2+\epsilon)\mathbf{L}_{ch}^l(m,s) \sigma_m(u_l)\le (2+\epsilon)\mathbf{L}_{ch}^l(m,s) \|x-y\|\quad \forall l\in \{0,1\}, \nonumbe \end{align*} and the result follows as before. \\ Finally, suppose $m$ is odd, $m>1$. The proof is as that for $m$ even, with the following modifications: For $l\in \{0,1\}$, let \begin{align*} E_l:=&\{i_{2s_{{\mathbb F},\epsilon}+j+l(m-1) }: 1\le j\le m-1\}, \end{align*} and let $E_2:=\{i_{2s_{{\mathbb F},\epsilon}+2m-1}, i_{2s_{{\mathbb F},\epsilon}+2m}\}$, with $\varepsilon_n:=(-1)^{n+1}$ as before. Set $$ z:=x-P_A(x)+a(1+\epsilon)s^{-1}t^{-1}({\mathbf 1}_{\varepsilon,E_0}+{\mathbf 1}_{\varepsilon, E_1}+{\mathbf 1}_{\varepsilon,E_2}). $$ Since $|E_{0}|=|E_{1}|=m-1$, for $l\in \{0,1\}$ we get \begin{align} a(1-\epsilon)s\|{\mathbf 1}_{\varepsilon, E_l}\|\le& \|u_l-v_l\|+\|x-y-v_l\|\le (1+1+\epsilon)\|u_l-v_l\|\nonumber\\ \le& (2+\epsilon)\mathbf{L}_{ch}^l(m-1,s) \sigma_{m-1}(u_l)\le (2+\epsilon)\mathbf{L}_{ch}^l(m-1,s) \|x-y\|,\nonumber \end{align} whereas $|E_2|=2$, so \begin{align} a(1-\epsilon)s\|{\mathbf 1}_{\varepsilon, E_2}\|\le& \|u_2-v_2\|+\|x-y-v_2\|\le (1+1+\epsilon)\|u_2-v_2\|\nonumber\\ \le& (2+\epsilon)\mathbf{L}_{ch}^l(2,s) \sigma_2(u_2)\le (2+\epsilon)\mathbf{L}_{ch}^l(2,s) \|x-y\|. \nonumber \end{align} and the proof is completed as in the even case. \\ Finally, suppose that ${\mathcal B}$ is $\mathbf{C}$-$s$-weak semi-greedy. Combining \ref{lebesgue1}, \ref{lebesgue1v}, \ref{lebesgue1v2} we get that $$ \mathbf{L}_d(m,1)\le \mathbf{M}\mathbf{C} \min\{ 1+2\mathbf{C}(\mathbf{M}+1) s^{-2}, 1+6\mathbf{C} s^{-2} \} $$ for all $m\ge 2$. Hence, using Lemma~\ref{lemmadisjoint=}, to complete the proof of the remaining inequality, we only need to prove that if $A=\{k\}\in {\mathcal G}(x,1,1)$ and $j\not =k$, then $$ \|x-P_A(x)\|\le (1+\lambda\lambda'+\lambda'') \|x-{\bm x}_{j}^*(x){\bm x}_j\|. $$ But this follows easily from the triangle inequality and the fact that $A$ is a greedy set for $x$: \begin{align*} \|x-P_A(x)\|\le& \|x-{\bm x}_{j}^*(x){\bm x}_j\|+\|P_A(x)\|+\|{\bm x}_{j}^*(x){\bm x}_j\|\\ =&\|x-{\bm x}_{j}^*(x){\bm x}_j\|+\|P_A(x-{\bm x}_{j}^*(x){\bm x}_j)\|+|{\bm x}_{j}^*(x)|\|{\bm x}_j\|\\ \le& (1+\lambda'')\|x-{\bm x}_{j}^*(x){\bm x}_j\|+|{\bm x}_{k}^*(x-{\bm x}_{j}^*(x){\bm x}_j)|\lambda\\ \le& (1+\lambda''+\lambda\lambda')\|x-{\bm x}_{j}^*(x){\bm x}_j\|. \end{align*} \end{proof} Next, we consider parameters involving only projections rather than arbitrary vectors. \begin{definition}\label{definitionlebesguedisjointalmost}Let ${\mathcal B}$ be a basis for ${\mathbb X}$, $0<t\le 1$ and $m\in \mathbb{N}$. \begin{itemize} \item The \emph{almost greedy parameter}, $\mathbf{L}_{a}(m,t) =\mathbf{L}_{a}(m,t)[{\mathcal B},{\mathbb X}]$, is given by $$ \mathbf{L}_{a}(m,t) :=\displaystyle \inf_{\mathbf{C}>0} \left\lbrace \|x-P_A(x)\|\le \mathbf{C} \|x-P_B(x)\|: \begin{array}{ll} \forall x\in {\mathbb X},\ \forall A\in {\mathcal G}(x,m,t),\\ \forall B\in \subset \mathbb{N}, |B|= m \end{array} \right\rbrace. $$ \item The \emph{disjoint almost greedy parameter}, $\mathbf{L}_{a,d}(m, t)=\mathbf{L}_{a,d}(m,t)[{\mathcal B},{\mathbb X}]$, is given by \begin{align*} \mathbf{L}_{a,d}(m, t):=\inf_{\mathbf{C}>0} \left\lbrace \|x-P_A(x)\|\le \mathbf{C} \|x-P_B(x)\|:\begin{array}{ll} \forall x\in {\mathbb X},\ \forall A\in {\mathcal G}(x,m,t),\\ B\subset \mathbb{N}, |B|= m: B\cap A=\emptyset \end{array} \right\rbrace. \end{align*} \end{itemize} \end{definition} In the case $t=1$, the almost greedy parameter has been studied for example in \cite{AAB2021}, \cite{BBG2017} and \cite{GHO2013}. The disjoint variants are weaker versions naturally suggested by Lemma~\ref{lemmadisjoint=}. \\ It is immediate that $\mathbf{L}_{a,d}(m, t)\le \min\{\mathbf{L}_{a}(m, t),\mathbf{L}_d(m,t)\}$, and the arguments of Lemma~\ref{lemmadisjoint=} show that $$ \mathbf{L}_{a}(m, t)\le \max_{1\le j\le m}\mathbf{L}_{a,d}(m, t). $$ Indeed, if $x\in {\mathbb X}$, $A\in {\mathcal G}(x,m,t)$ and $B\subset \mathbb{N}$, with $|B|=m$, are such that $0<\min\{|A\cap B|,|A\setminus B|\}$, then $A\setminus B\in {\mathcal G}(x-P_{A\cap B}(x), |A\setminus B|,t)$, so \begin{align*} \|x-P_A(x)\|=&\|x-P_{A\cap B}(x)-P_{A\setminus B}(x-P_{A\cap B}(x))\|\\ \le& \mathbf{L}_{a,d}(|A\setminus B|, t)\|x-P_{A\cap B}(x)-P_{B\setminus A}(x-P_{A\cap B}(x))\|\\ \le& \max_{1\le j\le m}\mathbf{L}_{a,d}(j, t)\|x-P_B(x)\|. \end{align*} Thus, Proposition~\ref{propositionsemigreedydisjointgreedy} also can be used to obtain bounds for $\mathbf{L}_{a}(m,t)$. In the case $t=1$, another route to find such upper bounds is to combine Proposition~\ref{propositionchevyshevianqg} and the results of Section~\ref{sectionseparation} with \cite[Theorem 3.3]{AAB2021}. First, we find estimates for the squeeze symmetry parameter, defined as follows. \begin{definition}\cite[Lemma~1.4(iii)]{AAB2021} \label{definitionsqueezeparameter} Let ${\mathcal B}$ be a basis for ${\mathbb X}$ with dual basis ${\mathcal B}^*=( {\bm x}^*_i)_{i\in\mathbb{N}}$ and $m\in \mathbb{N}$. The $m$-\emph{squeeze symmetry parameter}, $\bm \lambda_m=\bm \lambda_m[{\mathcal B},{\mathbb X}]$, is given by \begin{align*} \bm \lambda_m:=\inf_{\mathbf{C}>0} \left\lbrace \min_{n\in A}|{\bm x}_n^*(x)|\|1_{\varepsilon,B}\|\le \mathbf{C}\|x\|: \begin{array}{ll} \forall x\in {\mathbb X},\ \forall A\in {\mathcal G}(x,m,1),\\ \forall B\subset \mathbb{N}: |B|=m, \varepsilon\in \ensuremath{\mathcal{E}}_B \end{array} \right\rbrace. \end{align*} \end{definition} \begin{proposition}\label{propositionsqueezeparameter}Let ${\mathcal B}$ be a basis for ${\mathbb X}$ and $\mathbf{M}:=\mathbf{M}_{fs}({\mathcal B}, {\mathbb X})$. For all $0<s\le 1$ and $m\in \mathbb{N}$ we have the following estimates: \begin{enumerate}[\rm (i)] \item $\bm \lambda_m\le (\mathbf{L}_{ch}^l(m,s))^2\mathbf{M}(1+\mathbf{M})s^{-2},\qquad\forall m\in \mathbb{N}$. \label{propositionsqueezeparameter0} \item $\bm \lambda_m\le 2(\mathbf{L}_{ch}^l(m,s))^2\mathbf{M} s^{-2},\quad \qquad \qquad\forall m\in 2\mathbb{N}$.\label{propositionsqueezeparameter2} \item $\bm \lambda_m\le 2(\mathbf{L}_{ch}^l(m-1,s))^2\mathbf{M} s^{-2}+\lambda\lambda',\quad\forall m\in 2\mathbb{N}-1,$ where $\mathbf{L}_{ch}(0,s):=0$. \label{propositionsqueezeparameter1} \end{enumerate} \end{proposition} \begin{proof} The proof is very similar to those of Proposition~\ref{propositionsemigreedydisjointgreedy} and Theorem~\ref{theoremswemigreedydisjointgreedynotc0}\ref{wdisjointgreedyest2a} but simpler, so we shall be brief. Set ${\mathcal B}=( {\bm x}_i)_{i\in\mathbb{N}}$ and ${\mathcal B}^*=( {\bm x}^*_i)_{i\in\mathbb{N}}$. Choose $0<\epsilon<1$, and let $({\bm x}_{i_k})_{k\in \mathbb{N}}$ be a subsequence given by Corollary~\ref{corollaryboth} for $({\mathbb X}, \mathbf{M}, \epsilon)$. Now fix $x\in {\mathbb X}$ with finite support, $A\in {\mathcal G}(x,m,1)$, and $B\subset \mathbb{N}, |B|=m$. We may assume $a:=\min_{i \in A}|{\bm x}_i^*(x)|>0$. Choose $i_0>B\cup \supp(x)$, and set \begin{align*} {\mathbb F}:=&[{\bm x}_i: 1\le i\le i_0];\quad E_0:=\{i_{2s_{{\mathbb F},\epsilon}+j}: 1\le j\le m\};\quad z:=a{\mathbf 1}_{\varepsilon,B} +a(1+\epsilon)s^{-1}{\mathbf 1}_{E_0}. \end{align*} Since ${\mathcal G}(z,m,s)=\{E_0\}$, there is $v\in {\mathbb X}$ with $\supp(v)\subset E_0$ such that $\|z-v\|\le \mathbf{L}_{ch}^l(m,s)\sigma_{m}(z)$. Hence, \begin{align*} \|a{\mathbf 1}_{\varepsilon,B}\|\le& (\mathbf{M}+\epsilon)\|z-v\|\le (\mathbf{M}+\epsilon)\mathbf{L}_{ch}^l(m,s)\|z-a{\mathbf 1}_{\varepsilon,B}\|\\ =&(\mathbf{M}+\epsilon)\mathbf{L}_{ch}^l(m,s)a(1+\epsilon)s^{-1}\|{\mathbf 1}_{E_0}\|.\nonumber \end{align*} Let $u:= x-a(1-\epsilon)s{\mathbf 1}_{E_0}$. Since every element of ${\mathcal G}(u,m,s)$ is contained in $\{1,\dots,i_0\}$, there is $v\in {\mathbb F}$ such that \begin{equation*} \|u-v\|\le \mathbf{L}_{ch}^l(m,s) \sigma_m(u). \end{equation*} This entails that \begin{align} a(1-\epsilon)s\|{\mathbf 1}_{E_0}\|\le& \|u-v\|+\|x-v\|\le (1+\mathbf{M}+\epsilon)\|u-v\|\nonumber\\ \le& (1+\mathbf{M}+\epsilon)\mathbf{L}_{ch}^l(m,s) \|x\|.\label{propositionsqueezeparameter0generalcase} \end{align} Combining the above estimates and taking into account that $\epsilon$ is arbitrary, we get \begin{align*} \|a{\mathbf 1}_{\varepsilon,B}\|\le (\mathbf{L}_{ch}^l(m,s))^2\mathbf{M}(1+\mathbf{M})s^{-2} \|x\|, \end{align*} so the proof of \ref{propositionsqueezeparameter0} for finitely supported vectors is complete. The general case follows using some of the arguments of Theorem~\ref{theoremswemigreedydisjointgreedynotc0}.\\ The proof of \ref{propositionsqueezeparameter2} is like that of \ref{propositionsqueezeparameter0}, with only the following differences: the set $E_0$ has a partition $E_1\cupdot E_2$ with $|E_1|=|E_2|=\frac{m}{2}$ resulting from an application of Corollary~\ref{corollaryboth}, and instead of ${\mathbf 1}_{E_0}$ we define $z$ and $u$ using ${\mathbf 1}_{E_1}-{\mathbf 1}_{E_2}$, so that instead of \eqref{propositionsqueezeparameter0generalcase} we get \begin{align*} a(1-\epsilon)s\|{\mathbf 1}_{E_1}-{\mathbf 1}_{E_2}\|\le& \|u-v\|+\|x-v\|\le (1+1+\epsilon)\|u-v\|\nonumber\\ \le& (2+\epsilon)\mathbf{L}_{ch}^l(m,s) \|x\|. \end{align*} Finally, to prove \ref{propositionsqueezeparameter1}, write $B=B_0\cupdot \{j_0\}$ and $A=A_{0}\cup \{k_0\}$. Then $$ a\|{\bm x}_{j_0}\|\le \lambda\lambda'\|x\| $$ and, if $m>1$, by \ref{propositionsqueezeparameter2} we have $$ a\|{\mathbf 1}_{\varepsilon, B_0}\|\le \min_{n\in A_0}|{\bm x}_n^*(x)|\|{\mathbf 1}_{\varepsilon, _0}\|\le 2(\mathbf{L}_{ch}^l(m-1,s))^2\mathbf{M} s^{-2}\|x\|. $$ Applying the triangle inequality we get the desired result. \end{proof} Combining Propositions~\ref{propositionchevyshevianqg} and~\ref{propositionsqueezeparameter} with \cite[Theorem 3.3]{AAB2021}, we obtain the following. \begin{corollary}\label{corollaryalternative} Let ${\mathcal B}$ be a basis for ${\mathbb X}$ and $\mathbf{M}:=\mathbf{M}_{fs}({\mathcal B}, {\mathbb X})$. For $0<s\le 1$ and $m\in \mathbb{N}$ let $$ p_m:=\max_{1\le j\le m} \mathbf{L}_{ch}^l(2\floor{\frac{j+1}{2}},s) (1+ 2\mathbf{L}_{ch}^l(2\floor{\frac{j+1}{2}},s) s^{-2}). $$ The following hold: \begin{align*} \mathbf{L}_{a}(m,1)\le& \mathbf{M} p_{m}+ (\mathbf{L}_{ch}^l(m,s))^2\mathbf{M}(1+\mathbf{M})s^{-2}&&\forall m\in \mathbb{N};\\ \mathbf{L}_{a}(m,1)\le& \mathbf{M} p_{m}+ 2(\mathbf{L}_{ch}^l(m,s))^2\mathbf{M} s^{-2}&&\forall m\in 2\mathbb{N};\\ \mathbf{L}_{a}(m,1)\le& \mathbf{M} p_{m}+2(\mathbf{L}_{ch}^l(m-1,s))^2\mathbf{M} s^{-2}+\lambda\lambda'&&\forall m\in 2\mathbb{N}-1. \end{align*} In particular, if ${\mathcal B}$ is $\mathbf{C}$-$s$-weak semi-greedy, it is $\mathbf{K}$-almost greedy with $$ \mathbf{K}\le \mathbf{M} \mathbf{C}(1+2\mathbf{C} s^{-2})+\min\{\mathbf{C}^2\mathbf{M}(1+\mathbf{M})s^{-2}, 2\mathbf{C}^2\mathbf{M} s^{-2}+\lambda\lambda'\}. $$ \end{corollary} \section{Examples}\label{sectionexamples} In this section, we construct bases with some of the properties we have studied. We leave aside the case $\mathbf{w}\in \ell_1$ because in that case any $\mathbf{w}$-democratic basis is equivalent to the canonical unit vector basis of $\mathtt{c}_0$ (see Remark~\ref{remarkquantitativec0}). Similar considerations apply to the case $\mathbf{w}\not \in \ell_{\infty}$ (see \cite{B2020} and \cite{DKTW2018}). Also, it is known that if $\mathbf{w}\in\ell_{\infty}$ is a weight, there is a $\mathbf{w}$-greedy basis which is not equivalent to the canonical basis of $\mathtt{c}_0$. Indeed, for $1<p<\infty$, if ${\mathbb X}_p$ is the completion of $\mathtt{c}_{00}$ with the norm \begin{align} \|(a_n)_{n\in\mathbb{N}}\|=&\max\{\|(a_n)_{n\in\mathbb{N}}\|_{\infty},(\sum_{n\in \mathbb{N}}w_n|a_n|^p)^{\frac{1}{p}}\},\label{Xp} \end{align} the canonical vector basis ${\mathcal B}_p$ of ${\mathbb X}_p$ is $1$-unconditional, and $\|{\mathbf 1}_{\varepsilon, A}\|=\max\{1,(w(A))^{\frac{1}{p}}\}$ for all $A\in \mathbb{N}^{<\infty}$ (see \cite[Remark 4.10]{DKTW2018}) for $p=2$). For this reason, we will focus on constructing conditional bases. When $\mathbf{w}$ is seminormalized, the weighted properties are equivalent to their standard counterparts - that is, those involving constant weights -, and there are many examples of conditional bases showing a broad overview of the particularities of the different greedy-type bases (see e.g., \cite{AABBL2021}, \cite{AABW2021}, \cite{AADK2019}, \cite{DHK2006}, \cite{DKK2003}, \cite{KT1999} and \cite{W2000}, among others). On the other hand, we are not aware of any examples in the literature of conditional $\mathbf{w}$-almost greedy bases in either of the following cases: $\mathbf{w}\in \mathtt{c}_{0}\setminus \ell_1$, and $\mathbf{w}\in \ell_{\infty}\setminus \mathtt{c}_{0}$ with $(w_{n}^{-1})_{n\in\mathbb{N}}\not\in \ell_{\infty}$. Our purpose, in this section, is to give examples of such bases, as well as examples of $\mathbf{w}$-almost semi-greedy bases that are not $\mathbf{w}$-almost greedy. We begin with the weight in $\mathtt{c}_{0}\setminus \ell_1$ defined by $$ \mathbf{w}^{(1)}:=(n^{-\frac{1}{2}}\log(n+1))_{n\in\mathbb{N}}. $$ Our first task is to construct a conditional $\mathbf{w}^{(1)}$-almost greedy basis or, equivalently by \cite[Theorem 2.6]{DKTW2018}, quasi-greedy and $\mathbf{w}^{(1)}$-democratic. We will need an elementary lemma. \begin{lemma}\label{lemmalogcomputation}For all $A\in \mathbb{N}^{<\infty}$, \begin{align*} \sum_{n\in A}n^{-\frac{3}{4}}\le& 4 (\sum_{n\in A}n^{-\frac{1}{2}}\log(n+1))^{\frac{1}{2}}. \end{align*} \end{lemma} \begin{proof} By induction on $|A|$. If $|A|=1$, the result follows by an immediate computation. Fix $n_0\in \mathbb{N}$, and suppose the result holds for $1\le |A|\le n_0$. If $|A|=n_0+1$, let $n_1:=\max A$, and $A_1:=A\setminus\{n_1\}$. Using the inductive hypothesis, we obtain \begin{align*} &\big(\sum_{n\in A}n^{-\frac{3}{4}}\big)^2=\big(\sum_{n\in A_1}n^{-\frac{3}{4}}\big)^2\ +\ n_1^{-\frac{3}{2}} + 2 n_1^{-\frac{3}{4}}(\sum_{n\in A_1}n^{-\frac{3}{4}})\\ \le & 16 \sum_{n\in A_1}n^{-\frac{1}{2}}\log(n+1)\ +\ n_1^{-\frac{3}{2}}\ +\ 8 n_1^{-\frac{3}{4}}\Big(\sum_{n\in A_1}n^{-\frac{1}{2}}\log(n+1)\Big)^{\frac{1}{2}}\\ = & 16\sum_{n\in A}n^{-\frac{1}{2}}\log(n+1)\ +\ 8 n_1^{-\frac{3}{4}}\Big(\sum_{n\in A_1}n^{-\frac{1}{2}}\log(n+1)\Big)^{\frac{1}{2}}\ + n_1^{-\frac{3}{2}}-16n_1^{-\frac{1}{2}}\log(n_1+1) \\ \le &16 \sum_{n\in A}n^{-\frac{1}{2}}\log(n+1)\ +\ 8 n_1^{-\frac{3}{4}}\Big(\sum_{n\in A_1}n^{-\frac{1}{2}}\log(n+1)\Big)^{\frac{1}{2}}-15n_1^{-\frac{1}{2}}\log(n_1+1)\\ \le & 16 \sum_{n\in A}n^{-\frac{1}{2}}\log(n+1)\ +\ n_1^{-\frac{1}{2}}\log^{\frac{1}{2}}(n_1+1)\Big(8 n_1^{-\frac{1}{4}}\big(\sum_{n\in A_1}n^{-\frac{1}{2}}\big)^{\frac{1}{2}}-15\log^{\frac{1}{2}}(n_1+1)\Big)\\ \le& 16 \sum_{n\in A}n^{-\frac{1}{2}}\log(n+1)\ +\ n_1^{-\frac{1}{2}}\log^{\frac{1}{2}}(n_1+1)\big(8\sqrt{2}-15 \log^{\frac{1}{2}}(3)\big)\\ <& 16 \sum_{n\in A}n^{-\frac{1}{2}}\log(n+1), \end{align*} and the proof is complete. \end{proof} Now we can construct the example, which is a variant of one given in \cite{KT1999} (see also, \cite[página 35]{T2011} or \cite[página 266]{T2008}). \begin{example}\label{examplewalmostc0}For each $(a_n)_{n\in\mathbb{N}}\in \mathtt{c}_{00}$, define \begin{align*} &\|(a_n)_{n\in\mathbb{N}}\|_{\diamond}:=\|( (w^{(1)}_n)^{\frac{1}{2}} a_n)_{n\in\mathbb{N}}\|_2 &\rm and \qquad &\|(a_n)_{n\in\mathbb{N}}\|_{\circ}:=\sup_{m\in \mathbb{N}}|\sum_{n=1}^{m}n^{-\frac{3}{4}}a_n|. \end{align*} Let ${\mathbb X}$ be the completion of $\mathtt{c}_{00}$ with the norm $$ \|(a_n)_{n\in\mathbb{N}}\|:=\max\{\|(a_n)_{n\in\mathbb{N}}\|_{\diamond}, \|(a_n)_{n\in\mathbb{N}}\|_{\circ},\|(a_n)_{n\in\mathbb{N}}\|_{\infty}\}, $$ and let ${\mathcal B}$ be the canonical vector basis and ${\mathcal B}^*$ its dual basis. Then ${\mathcal B}$ is a normalized monotone conditional $\mathbf{w}^{(1)}$-almost greedy Schauder basis. \end{example} \begin{proof} It is clear that ${\mathcal B}$ is a normalized monotone Schauder basis. Now fix $A\in \mathbb{N}^{<\infty}$ and $\varepsilon\in \ensuremath{\mathcal{E}}_A$. We have $$ \|{\mathbf 1}_{\varepsilon, A}\|_{\diamond}=(w^{(1)}(A))^{\frac{1}{2}} $$ and, by Lemma~\ref{lemmalogcomputation}, $$ \|{\mathbf 1}_{\varepsilon, A}\|_{\circ}\le \sum_{n\in A}n^{-\frac{3}{4}}\le 4 (w^{(1)}(A))^{\frac{1}{2}}. $$ Thus, $$ \max\{1,(w^{(1)}(A))^{\frac{1}{2}}\}\le \|{\mathbf 1}_{\varepsilon, A}\|\le \max\{1, 4(w^{(1)}(A))^{\frac{1}{2}}\}. $$ It follows that ${\mathcal B}$ is $4$-$\mathbf{w}^{(1)}$-superdemocratic. To prove that it is quasi-greedy, fix $x\in {\mathbb X}$ with $\|x\|=1$, $m\in \mathbb{N}$, $A\in {\mathcal G}(x,m,1)$, and define $a:=\min_{n\in A}|{\bm x}_n^*(x)|$. We may assume $a>0$. Let $k_0:=\floor{ a^{-4}}$ and $B:=\{1,\dots, k_0\}$. We have \begin{align*} \|P_{A\cap B}(x)\|_{\circ}\le& \|P_B(x)\|_{\circ}+\|P_{B\setminus A}(x)\|_{\circ}\le 1+a\sum_{n\in B}n^{-\frac{3}{4}}\le 1+4a k_0^{\frac{1}{4}}\le 5, \end{align*} whereas, for $n>k_0$, $n\in A$, $|{\bm x}_n^{*}(x)|>n^{-\frac{1}{4}}$, and we have \begin{align*} \|P_{A\setminus B}(x)\|_{\circ}\le& \sum_{\substack{n\in A\\ n> k_0}} n^{-\frac{3}{4}}|{\bm x}_n^*(x)| \le \sum_{\substack{n\in A\\ n> k_0}} n^{-\frac{1}{2}}|{\bm x}_n^*(x)|^2 \le \sum_{\substack{n\in A}} n^{-\frac{1}{2}}\log(n+1)|{\bm x}_n^*(x)|^2\le 1. \end{align*} As the respective inequalities are immediate for $\|\cdot\|_{\diamond}$ and $\|\cdot\|_{\infty}$, this proves that ${\mathcal B}$ is $6$-quasi-greedy. It only remains to prove that it is conditional. For each $m\in \mathbb{N}$, define \begin{align*} &y_m:=\sum_{n=1}^{m}n^{-\frac{1}{4}}\log^{-1}(n+1){\bm x}_n & \text{and}\qquad &z_m:=\sum_{n=1}^{m}(-1)^{n}n^{-\frac{1}{4}}\log^{-1}(n+1){\bm x}_n. \end{align*} Then $\|y_m\|_{\infty}=\|z_m\|_{\infty}= \log^{-1}(2)$ and \begin{align*} \|y_m\|_{\diamond}=&\|z_m\|_{\diamond}=(\sum_{n=1}^{m}n^{-1}\log^{-1}(n+1))^{\frac{1}{2}}. \end{align*} On the other hand, \begin{align*} \|y_m\|_{\circ}=&\sum_{n=1}^{m}n^{-1}\log^{-1}(n+1), \end{align*} whereas \begin{align*} \|z_m\|_{\circ}=\max_{1\le k\le m}|\sum_{n=1}^{k}(-1)^{n}n^{-1}\log^{-1}(n+1)|\le 2\|z_m\|_{\infty}=2 \log^{-1}(2). \end{align*} Thus, $$\|z_m\|^{-1}\|y_m\|\xrightarrow[m\to \infty] {}\infty, $$ and the proof is complete. \end{proof} Next, we modify Example~\ref{examplewalmostc0} to obtain a $\mathbf{w}^{(1)}$-almost semi-greedy basis that is neither quasi-greedy nor, in any order, a Schauder basis. For the construction, we also adapt some of the arguments from \cite[Theorem 3.13 and Proposition 4.17]{AABBL2021}. \begin{example}\label{examplewalmostsemic0} Let $(A_m)_{m\in\mathbb{N}}\subset \mathbb{N}^{<\infty}$ be a sequence of nonempty integer intervals such that $A_m<A_{m+1}$ for all $m$, and $$ \big(\sum_{n\in A_m}n^{-1}\log^{-1}(n+1)\big)_{m\in\mathbb{N}} \subset (1,\infty) $$ is an unbounded sequence. For each $(a_n)_{n\in\mathbb{N}}\in \mathtt{c}_{00}$, define \begin{align*} &\|(a_n)_{n\in\mathbb{N}}\|_{\triangleleft}:=\sup_{m\in \mathbb{N}}|\sum_{n\in A_m}n^{-\frac{3}{4}}a_n|, \end{align*} and let $\|\cdot\|_{\diamond}$ be as in Example~\ref{examplewalmostc0}. Let ${\mathbb X}$ be the completion of $\mathtt{c}_{00}$ with the norm $$ \|(a_n)_{n\in\mathbb{N}}\|:=\max\{\|(a_n)_{n\in\mathbb{N}}\|_{\diamond}, \|(a_n)_{n\in\mathbb{N}}\|_{\triangleleft},\|(a_n)_{n\in\mathbb{N}}\|_{\infty}\}, $$ and let ${\mathcal B}$ be the canonical vector basis with ${\mathcal B}^*$ its dual basis. Then ${\mathcal B}$ is a normalized $\mathbf{w}^{(1)}$-almost semi-greedy basis that is not quasi-greedy nor, in any order, a Schauder basis. \end{example} \begin{proof} It is clear that ${\mathcal B}$ is normalized. First, we prove that ${\mathcal B}$ is a Markushevich basis: suppose otherwise, and fix $x\in {\mathbb X}\setminus \{0\}$ with ${\bm x}_k^*(x)=0$ for all $k\in\mathbb{N}$. Pick a sequence $(x_n)_{n\in\mathbb{N}}\subset [{\mathcal B}]\setminus \{0\}$ so that $\|x-x_n\|\le n^{-1}$ for each $n\in \mathbb{N}$. Now choose a strictly increasing sequence $(m(n))_{n\in\mathbb{N}}\subset \mathbb{N}$ so that for each $n\in \mathbb{N}$, $A_{m(n)}>\max_{1\le k\le n}\supp(x_k)$, and set $B_n:=\{1,\dots, c_n:=\max(A_{m(n)})\}$. Let $n_1:=1$ and $y_1=x_1$. Given that for each $k\in\mathbb{N}$, $$ |{\bm x}_k^*(x_n)|\xrightarrow[n\to\infty]{}0<\inf_{n\in \mathbb{N}}\|x_n\|, $$ there is $n_2>c_{n_1}$ such that $\|P_{B_{n_1}}(x_{n_2})\|\le 2^{-1}$ and $x_{n_2}\not=P_{B_{n_1}}(x_{n_2})$. Hence, if $y_{2}:=x_{n_2}-P_{B_{n_1}}(x_{n_2})$, then $$ y_2\not=0,\quad \supp(y_2)>B_{n_1}\supset \supp(y_1), \quad \|y_2-x\|\le \|x_{n_2}-x\|+\|P_{B_{n_1}}(x_{n_2})\| \le 2^{-1}+2^{-1}. $$ In addition, for each $m\in \mathbb{N}$, if $A_m\cap \supp(y_2)\not=\emptyset$, then $A_m>\supp(y_1)$. Similarly, we can find $n_3>c_{n_2} $ so that $\|P_{B_{n_2}}(x_{n_3})\|\le 3^{-1}$ and $x_{n_3}\not=P_{B_{n_2}}(x_{n_3})$. Hence, if $y_3:=x_{n_3}-P_{B_{n_2}}(x_{n_3})$, $$y_3\not=0,\quad \supp(y_3)>B_{n_2}\supset \supp(y_2), \quad \|y_3-x\|\le\|x_{n_3}-x\|+\|P_{B_{n_2}}(x_{n_3})\|\le \frac{2}{3},$$ and for each $m\in \mathbb{N}$, if $A_m\cap \supp(y_3)\not=\emptyset$, then $A_m>\supp(y_2)$. In this manner, we find an increasing sequence $(n_k)_{k\in \mathbb{N}}$ so that $(y_k:=x_{n_k}-P_{B_{n_k}}(x_{n_k}))_{k\in \mathbb{N}}$ has the following properties: for each $k\in\mathbb{N}$, $$y_k\in [{\mathcal B}]\setminus\{0\},\qquad \supp(y_{k+1})>B_{n_k}\supset \supp(y_{k}), \qquad \|y_k-x\|\le 2 k^{-1}$$ and, for every $m, k\in \mathbb{N}$, if $A_m\cap \supp(y_{k})\not=\emptyset$ then $A_m\cap \supp(y_j)=\emptyset$ for all $j\ne k$.\\ For a contradiction, let us show that $(y_k)_{k\in\mathbb{N}}$ converges to $0$ and consequently $x=0$. First, notice that as $ \supp(y_{k+1})\cap \supp(y_{k}) =\emptyset$, $\|y_k\|_{\infty} \le \|y_k-y_{k+1}\|_{\infty}$ and $\|y_k\|_{\diamond} \le \|y_k-y_{k+1}\|_{\diamond}$. Also, as each $A_m$ intersects with only one $\supp(y_j)$, $\|y_k\|_{\triangleleft} \le \|y_k-y_{k+1}\|_{\triangleleft}$. Thus, $$ \|y_k\|\le \|y_k-y_{k+1}\|\le 4k^{-1}\xrightarrow[k\to \infty]{}0. $$ Therefore, ${\mathcal B}$ is a Markushevich basis. \\ Next, we prove that ${\mathcal B}$ is $\mathbf{w}^{(1)}$-almost semi-greedy: by Proposition~\ref{propositionequivalencesalmostsemigreedy}, it suffices to show that ${\mathcal B}$ is truncation quasi-greedy and $\mathbf{w}^{(1)}$-superdemocratic. To prove the former, fix $x\in {\mathbb X}$, $m\in \mathbb{N}$ and $A\in {\mathcal G}(x,m,1)$. Let $a:=\min_{n\in A}|{\bm x}_n^*(x)|$. For each $m\in \mathbb{N}$ for which $A\cap A_m\not=\emptyset$ and every $\varepsilon\in \ensuremath{\mathcal{E}}_A$, by Lemma~\ref{lemmalogcomputation} we have \begin{align*} a|\sum_{n\in A_m\cap A}n^{-\frac{3}{4}}\varepsilon_n|\le& a \sum_{n\in A\cap A_m}n^{-\frac{3}{4}}\le 4 a(\sum_{n\in A\cap A_m}n^{-\frac{1}{2}}\log(n+1))^{\frac{1}{2}}\\ \le&4 (\sum_{n\in A\cap A_m}n^{-\frac{1}{2}}\log(n+1) |{\bm x}_n^*(x)|^2)^{\frac{1}{2}}\le 4\|x\|. \end{align*} We conclude that ${\mathcal B}$ is $4$-truncation quasi-greedy. Now fix $A\in \mathbb{N}^{<\infty}$ and $\varepsilon\in \ensuremath{\mathcal{E}}_A$, and let $\|\cdot\|_{e2}$ be the norm on $\mathtt{c}_{00}$ defined in Example~\ref{examplewalmostc0}. By the computations in the aforementioned example, we have $$ (w^{(1)}(A))^{\frac{1}{2}}=\|{\mathbf 1}_{\varepsilon, A}\|_{\diamond}\le \|{\mathbf 1}_{\varepsilon, A}\|\le \|{\mathbf 1}_{\varepsilon, A}\|_{e2}\le 4(w^{(1)}(A))^{\frac{1}{2}}. $$ Thus, ${\mathcal B}$ is $4$-$\mathbf{w}^{(1)}$-superdemocratic. \\ Finally, we prove that ${\mathcal B}$ is neither quasi-greedy nor, in any order, a Schauder basis. Fix $\pi:\mathbb{N}\rightarrow\mathbb{N}$ a biyection, and let ${\mathcal B}_{\pi}:=({\bm x}_{\pi(n)})_{n\in\mathbb{N}}$ be the reordered basis. For each $m\in \mathbb{N}_{\ge 2}$, let $d_m:=|A_m|$, and write $D_m:=\pi^{-1}(A_m)=\{l_{m,1}<\cdots<l_{m,d_m}\}$. Now consider $1\le j_m<d_m$ defined by $$ j_m:=\min_{1\le k < d_m}\left\lbrace\sum_{j=k+1}^{d_m}\pi(l_{m,j})^{-1}\log^{-1}(\pi(l_{m,j})+1)\le \sum_{j=1}^k\pi(l_{m,j})^{-1}\log^{-1}(\pi(l_{m,j})+1)\right\rbrace, $$ Then, we have \begin{align} \sum_{j=j_m+1}^{d_m}\pi(l_{m,j})^{-1}\log^{-1}(\pi(l_{m,j})+1)\le& \sum_{j=1}^{j_m}\pi(l_{m,j})^{-1}\log^{-1}(\pi(l_{m,j})+1)\nonumber\\ \le& 1+\sum_{j=j_m+1}^{d_m}\pi(l_{m,j})^{-1}\log^{-1}(\pi(l_{m,j})+1). \label{examplewalmostsemic0dif<1} \end{align} For each $m\ge 2$, define \begin{align} z_{m}:&=\sum_{j=1}^{j_m}\pi(l_{m,j})^{-\frac{1}{4}}\log^{-1}(\pi(l_{m,j})+1){\bm x}_{\pi(l_{m,j})}-\sum_{j=j_m+1}^{d_m} \pi(l_{m,j})^{-\frac{1}{4}}\log^{-1}(\pi(l_{m,j})+1){\bm x}_{\pi(l_{m,j})},\label{examplewalmostsemic0-zm} \end{align} that satisfies \begin{align} \|z_m\|_{\infty}\le& 1,\qquad \|z_m\|_{\triangleleft}\le 1,\qquad \|z_m\|_{\diamond}=\big(\sum_{n\in A_m}n^{-1}\log^{-1}(n+1)\big)^{\frac{1}{2}}\label{examplewalmostsemic0forlater0};\\ \|\sum_{j=1}^{j_m}{\bm x}_{\pi(l_{m,j})}^*(z_m){\bm x}_{\pi(l_{m,j})}\|_{\triangleleft}=&\sum_{j=1}^{j_m}\pi(l_{m,j})^{-1}\log^{-1}(\pi(l_{m,j})+1)\ge 2^{-1}\sum_{n\in A_m}n^{-1}\log^{-1}(n+1).\label{examplewalmostsemic0forlater} \end{align} Hence, $$ \liminf_{m\to \infty} \frac{\|z_m\|}{\|\sum_{j=1}^{j_m}{\bm x}_{\pi(l_{m,j})}^*(z_m){\bm x}_{\pi(l_{m,j})}\|}=0, $$ so ${\mathcal B}_{\pi}$ is not a Schauder basis. Moreover, when $\pi$ is the identity mapping on $\mathbb{N}$, we have $D_m=A_m$ and $E_m:=\{l_{m,1},\dots, l_{m,j_m}\}\in {\mathcal G}(z_m,j_m,1)$, which proves that ${\mathcal B}$ is not quasi-greedy. \end{proof} In our next example, we use the previous one to construct a $\mathbf{w}^{(1)}$-almost semi-greedy conditional Schauder basis that is not quasi-greedy. In the construction, we will use the following lemma. \begin{lemma}\label{lemmagoesSchauder} Let ${\mathcal B}=({\bm x}_n)_{n\in\mathbb{N}}$ be a basis for ${\mathbb X}$ and $\mathbf{w}$ a weight. Define ${\mathbb Y}$ as the completion of $\mathtt{c}_{00}$ with the norm $$ \|(a_n)_{n\in\mathbb{N}}\|_{{\mathbb Y}}=\sup_{m\in \mathbb{N}}\sup_{1\le k\le m}\|\sum_{n=k}^{m}a_n{\bm x}_n\|_{{\mathbb X}}, $$ and let ${\mathcal B}_1:=({\bm y}_n)_{n\in\mathbb{N}}$ be the canonical basis of ${\mathbb Y}$. The following hold: \begin{enumerate}[\rm (i)] \item \label{lemmagoesSchauderbasisnorm}${\mathcal B}$ is a monotone Schauder basis and $\|{\bm y}_n\|_{{\mathbb Y}}=\|{\bm x}_n\|_{{\mathbb X}}$ for all $n\in \mathbb{N}$. \item \label{lemmagoesSchauder1A} For every $A\in \mathbb{N}^{<\infty}$ and every $\varepsilon\in \ensuremath{\mathcal{E}}_A$, $$ \|{\mathbf 1}_{\varepsilon, A}\|_{{\mathbb Y}}\le \max_{B\subset A}\|{\mathbf 1}_{\varepsilon, B}\|_{{\mathbb X}}. $$ \item \label{lemmagoesSchauderdem} If ${\mathcal B}$ is $\mathbf{C}$-$\mathbf{w}$-(super)democratic, so is ${\mathcal B}_1$. \item \label{lemmagoesSchaudertqg}If ${\mathcal B}$ is $\mathbf{C}$-truncation quasi-greedy, ${\mathcal B}_1$ is $2\kappa\mathbf{C}^2$-truncation quasi-greedy. \end{enumerate} \end{lemma} \begin{proof} \ref{lemmagoesSchauderbasisnorm} and \ref{lemmagoesSchauder1A} are immediate from the definitions. To prove \ref{lemmagoesSchauderdem}, we consider the superdemocracy case, as the democracy one is proven in the same manner. Fix $A,B\in \mathbb{N}^{<\infty}$ with $w(A)\le w(B)$, $\varepsilon\in \ensuremath{\mathcal{E}}_A$ and $\varepsilon'\in \ensuremath{\mathcal{E}}_B$. By \ref{lemmagoesSchauder1A}, $$ \|{\mathbf 1}_{\varepsilon, A}\|_{{\mathbb Y}}\le \max_{D\subset A}\|{\mathbf 1}_{\varepsilon, D}\|_{{\mathbb X}}\le \mathbf{C} \|{\mathbf 1}_{\varepsilon',B}\|_{{\mathbb X}}\le \mathbf{C} \|{\mathbf 1}_{\varepsilon',B}\|_{{\mathbb Y}}. $$ Now suppose ${\mathcal B}$ is $\mathbf{C}$-truncation quasi-greedy. Let $({\bm x}^*_i)_{i\in\mathbb{N}}$ and $({\bm y}^*_i)_{i\in\mathbb{N}}$ be the dual bases of ${\mathcal B}$ and ${\mathcal B}_1$, respectively. Given $y\in {\mathbb Y}$ with finite support and $A\in {\mathcal G}(y,m,1)$ for some $m\in \mathbb{N}$, set $x:=\sum_{n\in \supp(y)}{\bm y}_n^*(y){\bm x}_n$. It follows from \ref{lemmagoesSchauder1A} and Remark~\ref{remarkpropc=tqg} that \begin{align*} \min_{n\in A}|{\bm y}_n^*(y)|\|{\mathbf 1}_{\varepsilon(y), A}\|_{{\mathbb Y}}\le& \min_{n\in A}|{\bm x}_n^*(x)|\ \max_{B\subset A}\|{\mathbf 1}_{\varepsilon(x), B}\|_{{\mathbb X}} \le 2\kappa \mathbf{C}^2 \|x\|_{{\mathbb X}} \le 2\kappa \mathbf{C}^2\|y\|_{{\mathbb Y}}. \end{align*} This completes the proof of \ref{lemmagoesSchaudertqg} and of the lemma. \end{proof} \begin{remark}\label{remarkgoesschauder}Note that Lemma~\ref{lemmagoesSchauder} does not require a totality hypothesis on ${\mathcal B}^*$. Also, one could note that the basis ${\mathcal B}_1$ obtained by this method inherits from ${\mathcal B}$ several properties studied in the context of greedy approximation in addition to those stated in the lemma, such as quasi-greediness, quasi-greediness for largest coefficients, unconditionality for constant coefficients, and bidemocracy. \end{remark} \begin{example}\label{examplewalmostsemic0schauder}There is a $\mathbf{w}^{(1)}$-almost semi-greedy Schauder basis that is not quasi-greedy. \end{example} \begin{proof} We will use the construction and notation of Example~\ref{examplewalmostsemic0} unless otherwise specified, with ${\mathbb X}$ the space, ${\mathcal B}$ the $\mathbf{w}^{(1)}$-almost semi-greedy basis of that example respectively, and $\pi:\mathbb{N}\rightarrow \mathbb{N}$ the identity mapping. Take $(z_m)_{m\in \mathbb{N}}$ as in \eqref{examplewalmostsemic0-zm}. Then for each $m\in \mathbb{N}$, \begin{align*} \|z_m\|_{\triangleleft}\le 1\qquad\text{and}\qquad |{\bm x}_n^*(z_m)|\le 1\;\ \forall n\in A_m=\supp(z_m). \end{align*} Also, it follows from \eqref{examplewalmostsemic0dif<1} and \eqref{examplewalmostsemic0-zm} that \begin{align*} 0\le& \sum_{n\in A_m}n^{-\frac{3}{4}}{\bm x}_n^*(z_m)\le 1. \end{align*} From the above inequalities we deduce that there is a bijection $\rho_m:A_m\rightarrow A_m$ such that \begin{align} \max_{\substack{k\in A_m}}|\sum_{n\le k} \rho_m(n)^{-\frac{3}{4}}{\bm x}_{\rho_m(n)}^*(z_m)|\le 1,\label{new} \end{align} which entails that \begin{align} \max_{\substack{j,k\in A_m\\j\le k}}|\sum_{n=j}^{k} \rho_m(n)^{-\frac{3}{4}}{\bm x}_{\rho_m(n)}^*(z_m)|\le 2.\label{examplewalmostsemic0schauderlessthan2} \end{align} Define $\rho:\mathbb{N}\rightarrow \mathbb{N}$ by \begin{align*} \rho(n):= \begin{cases} \rho_m(n) & \text{if } n\in A_m;\\ n & \text{if } n\not\in \bigcup_{m\in\mathbb{N}}A_m. \end{cases} \end{align*} Then $\rho:\mathbb{N}\rightarrow \mathbb{N}$ is a bijection. For each $n\in \mathbb{N}$, let ${\bm z}_n:={\bm x}_{\rho(n)}$, and let ${\mathcal B}_0:=({\bm z}_n)_{n\in\mathbb{N}}$. As $\supp_{{\mathcal B}_0}(z_m)=\supp_{{\mathcal B}}(z_m)=A_m$ for each $m\in \mathbb{N}$, it follows from \eqref{examplewalmostsemic0schauderlessthan2} that \begin{align} \sup_{m\in \mathbb{N}} \sup_{{\substack{j,k\in A_m\\j\le k}}}\color{black}|\sum_{n=j}^{k} \rho(n)^{-\frac{3}{4}}{\bm z}_{n}^*(z_m)|\le 2.\label{examplewalmostsemic0schauderlessthan2allm} \end{align} On the other hand, for each $m\in \mathbb{N}$, it follows from \eqref{examplewalmostsemic0forlater} that \begin{align} \|P_{\rho^{-1}(E_m),{\mathcal B}_0}(z_m)\|=\|P_{E_m,{\mathcal B}}(z_m)\|\ge& 2^{-1}\sum_{n\in A_m}n^{-1}\log^{-1}(n+1).\nonumbe \end{align} Let ${\mathcal B}_1=({\bm y}_n)_{n\in\mathbb{N}}$ be the Schauder basis obtained from ${\mathcal B}_0$ by an application of Lemma~\ref{lemmagoesSchauder}. Then ${\mathcal B}_1$ is truncation quasi-greedy and $\mathbf{w}^{(1)}$-superdemocratic since, ${\mathcal B}_0$ (which is a reordering of ${\mathcal B}$) has these properties. Hence, by Proposition~\ref{propositionequivalencesalmostsemigreedy}, ${\mathcal B}_1$ is $\mathbf{w}^{(1)}$-almost semi-greedy. \\ For each $m\in \mathbb{N}$, let $$ y_m:=\sum_{n\in \mathbb{N}}{\bm z}_n^*(z_m){\bm y}_m. $$ We have $$ \|P_{\rho^{-1}(E_m),{\mathcal B}_1}(y_m)\|_{{\mathbb Y}}\ge \|P_{\rho^{-1}(E_m),{\mathcal B}_0}(z_m)\|\ge 2^{-1}\sum_{n\in A_m}n^{-1}\log^{-1}(n+1),\quad \forall m\in \mathbb{N}. $$ On the other hand, it follows from \eqref{examplewalmostsemic0forlater0} and \eqref{examplewalmostsemic0schauderlessthan2allm} that \begin{align*} \|y_m\|_{{\mathbb Y}}=& \sup_{{\substack{j,k\in A_m\\j\le k}}} \|\sum_{n=j}^{k}{\bm z}_n^*(z_m){\bm z}_n\|_{{\mathbb X}}=\max\{ \|z_m\|_{\infty},\|z_m\|_{\diamond},\sup_{{\substack{j,k\in A_m\\j\le k}}}|\sum_{n=j}^{k} \rho(n)^{-\frac{3}{4}}{\bm z}_{n}^*(z_m)|\}\\ \le& 2+ (\sum_{n\in A_m}n^{-1}\log^{-1}(n+1))^{\frac{1}{2}}. \end{align*} It follows that $$ \liminf_{m\to \infty}\frac{\|y_m\|_{{\mathbb Y}}}{\|P_{\rho^{-1}(E_m),{\mathcal B}_1}(y_m)\|_{{\mathbb Y}}}=0. $$ Since $E_m\in {\mathcal G}_{{\mathcal B}}(z_m, d_m,1)$, we have $\rho^{-1}(E_m)\in {\mathcal G}_{{\mathcal B}_0}(z_m, d_m,1)= {\mathcal G}_{{\mathcal B}_1}(y_m, d_m,1)$. Therefore, ${\mathcal B}_1$ is not quasi-greedy. \end{proof} \begin{remark}\label{remarkschauderbidem}While not the focus of this paper, we point out that Remark~\ref{remarkgoesschauder} combined with the proof of Example~\ref{examplewalmostsemic0schauder} suggests a way of using the constructions of \cite[Theorems 3.6, 3.13]{AABBL2021} to obtain bidemocratic Schauder bases that are not quasi-greedy, for a wide range of fundamental functions. \end{remark} Now we turn to the case of $\mathbf{w}\in \ell_{\infty}\setminus \mathtt{c}_{0}$ with $(w_{n}^{-1})_{n\in\mathbb{N}}\not\in \ell_{\infty}$. Here, we have more flexibility constructing conditional bases: given two weights $\mathbf{w}$, $\mathbf{w}'$, define their combined weight ${\mathbf{W}}(\mathbf{w},\mathbf{w}')$ by \begin{align*} W_{2n-1}(\mathbf{w},\mathbf{w}'):=w_n\qquad\text{and}\qquad W_{2n}(\mathbf{w},\mathbf{w}'):= w'_n. \end{align*} We will show that given $\mathbf{w}\in \mathtt{c}_0$ and $\mathbf{w}'$ seminormalized, there is a ${\mathbf{W}}(\mathbf{w},\mathbf{w}')$-almost greedy conditional Schauder basis, and obtain similar results for the weighted almost semi-greedy property. To that end, we will combine the $\mathbf{w}$-greedy basis ${\mathcal B}_p$ of the space ${\mathbb X}_p$ which we defined using \eqref{Xp}, with suitable conditional bases. First we need a technical lemma. \begin{lemma}\label{lemmawdemdouble}Let $\mathbf{w}, \mathbf{w}'$ be weights, $\mathbf{C}>0$, and ${\mathcal B}_1=({\bm x}_n)_{n\in\mathbb{N}}$, ${\mathcal B}_2=({\bm y}_n)_{n\in\mathbb{N}}$ bases for ${\mathbb X}$ and ${\mathbb Y}$ respectively. Suppose that for all $A,B\in \mathbb{N}^{<\infty}$, $\|{\mathbf 1}_{A}\|_{{\mathbb X}}\le \mathbf{C} \|{\mathbf 1}_{B}\|_{{\mathbb Y}}$ whenever $w(A)\le w'(B)$. Then, for all $A,B\in \mathbb{N}^{<\infty}$ such that $w(A)\le 2w'(B)$, $$ \|{\mathbf 1}_{A}\|_{{\mathbb X}}\le (2\mathbf{C}+\lambda_0 \lambda'_0)\|{\mathbf 1}_{B}\|_{{\mathbb Y}}, $$ where $\lambda_0$ and $\lambda'_0$ are the maxima between the constants $\lambda$ and $\lambda'$ in \eqref{constants} for ${\mathcal B}_1$ and ${\mathcal B}_2$. \end{lemma} \begin{proof} Fix $A, B\in \mathbb{N}^{<\infty}$ with $w(A)\le 2w'(B)$. First choose two (possibly empty) sets $A_1,A_2$ as follows: $A_1$ is a subset of $A$ with maximum $w$-measure such that $w(A_1)\le w'(B)$, and $A_2$ is a subset of $A\setminus A_1$ with the same property. Now let $A_3:=A\setminus (A_1\cup A_2)$. We claim that $|A_3|\le 1$. Indeed, if this is false, choose $m_1,m_2\in A\setminus (A_1\cup A_2)$, with $m_1\not=m_2$. Then by our choice of $A_1$ and $A_2$, $w(A_1\cup \{m_1\})>w'(B)$ and $w(A_2\cup \{m_2\})>w'(B)$. Hence, $w(A)>2w'(B)$, a contradiction. We conclude that $$ \|{\mathbf 1}_{A}\|_{{\mathbb X}}\le \|{\mathbf 1}_{A_1}\|_{{\mathbb X}}+\|{\mathbf 1}_{A_2}\|_{{\mathbb X}}+\|{\mathbf 1}_{A_3}\|_{{\mathbb X}}\le (2\mathbf{C}+\lambda_0\lambda_0')\|{\mathbf 1}_{B}\|_{{\mathbb Y}}. $$ \end{proof} \begin{remark}Note that Lemma~\ref{lemmawdemdouble} can be applied to a weight $\mathbf{w}$ and a $\mathbf{C}$-$\mathbf{w}$-democratic basis ${\mathcal B}$, taking ${\mathcal B}_1={\mathcal B}_2={\mathcal B}$ and ${\mathbb X}={\mathbb Y}$. \end{remark} The next lemma forms the basis of our final construction. \begin{lemma}\label{lemmasumofspaces}Let ${\mathcal B}_1=({\bm x}_n)_{n\in\mathbb{N}}$ be a basis for ${\mathbb X}$ and ${\mathcal B}_2=({\bm y}_n)_{n\in\mathbb{N}}$ a basis for ${\mathbb Y}$. Define ${\mathbb Z}:={\mathbb X}\oplus {\mathbb Y}$ with the norm $\|(x,y)\|_{{\mathbb Z}}:=\max\{\|x\|_{{\mathbb X}},\|y\|_{{\mathbb Y}}\}$, and let ${\mathcal B}=({\bm z}_n)_{n\in\mathbb{N}}$ be given by \begin{align*} {\bm z}_{2n-1}:=({\bm x}_n,0)\qquad\text{and}\qquad {\bm z}_{2n}:=(0,{\bm y}_n). \end{align*} Then ${\mathcal B}$ is a basis for ${\mathbb Z}$, and the following hold: \begin{enumerate}[ \rm (i)] \item \label{lemmasumofspacesschauder} ${\mathcal B}$ is a Schauder basis if and only if ${\mathcal B}_1$ and ${\mathcal B}_2$ are Schauder bases. \item \label{lemmasumofspacesqg}${\mathcal B}$ is unconditional (quasi-greedy) if and only if ${\mathcal B}_1$ and ${\mathcal B}_2$ are unconditional (quasi-greedy). \item \label{lemmasumofspacestcg} ${\mathcal B}$ is truncation quasi-greedy if and only if ${\mathcal B}_1$ and ${\mathcal B}_2$ are truncation-quasi-greedy. \item \label{lemmasumofspaceswdem} ${\mathcal B}$ is ${\mathbf{W}}(\mathbf{w},\mathbf{w}')$-democratic if and only if ${\mathcal B}_1$ is $\mathbf{w}$-democratic, ${\mathcal B}_2$ is $\mathbf{w}'$-democratic, and there is $\mathbf{C}>0$ such that for all $A,B\in \mathbb{N}^{<\infty}$, we have $$ \|{\mathbf 1}_{A}\|_{{\mathbb X}}\le \mathbf{C} \|{\mathbf 1}_{B}\|_{{\mathbb Y}} $$ if $w(A)\le w'(B)$, and $$ \|{\mathbf 1}_{A}\|_{{\mathbb Y}}\le \mathbf{C} \|{\mathbf 1}_{B}\|_{{\mathbb X}} $$ if $w'(A)\le w(B)$. \end{enumerate} \end{lemma} \begin{proof} \ref{lemmasumofspacesschauder} and \ref{lemmasumofspacestcg} are clear. To prove \ref{lemmasumofspacestcg}, suppose that ${\mathcal B}_j$ is $\mathbf{C}_j$-truncation quasi-greedy, $j=1,2$, and fix $z\in {\mathbb Z}$, $m\in \mathbb{N}$, and $A\in {\mathcal G}(z,m,1)$. Let $a:=\min_{i\in A}|{\bm z}_{i}^*(z)|$ with $({\bm z}^*_i)_{i\in\mathbb{N}}$ the dual basis of ${\mathcal B}$. Write $z=(x,y)$, and set \begin{align*} &A_1:=\{n\in \mathbb{N}: 2n-1\in A\}, &&A_2:=\{n\in \mathbb{N}: 2n\in A\}. \end{align*} Note that $A_1\in {\mathcal G}(x,|A_1|,1)$ and $A_2\in {\mathcal G}(y,|A_2|,1)$. Hence, taking $\varepsilon(z)$, $\varepsilon(x)$ and $\varepsilon(y)$ with respect to the respective bases, we have \begin{align*} a\|{\mathbf 1}_{\varepsilon(z),A}\|_{{\mathbb Z}}\le& \max \{a\|{\mathbf 1}_{\varepsilon(x),A_1}\|_{{\mathbb X}}, a\|{\mathbf 1}_{\varepsilon(y),A_2}\|_{{\mathbb Y}}\} \le\max\{\mathbf{C}_1\|x\|_{{\mathbb X}},\mathbf{C}_2\|y\|_{{\mathbb Y}}\}\\ \le& \max\{\mathbf{C}_1,\mathbf{C}_2\}\|z\|_{{\mathbb Z}}. \end{align*} This proves that ${\mathcal B}$ is $\max\{\mathbf{C}_1,\mathbf{C}_2\}$-truncation quasi-greedy. \\ On the other hand, if ${\mathcal B}$ is $\mathbf{C}$-truncation quasi-greedy, it is immediate that so are ${\mathcal B}_1$ and ${\mathcal B}_2$.\\ It only remains to prove \ref{lemmasumofspaceswdem}. Suppose first that ${\mathcal B}$ is $\mathbf{C}$-${\mathbf{W}}(\mathbf{w},\mathbf{w}')$-democratic, and fix $A, B\in \mathbb{N}^{\infty}$ with $w(A)\le w(B)$. Given that $w(D)=W(\mathbf{w},\mathbf{w}')(2D-1)$ for all $D\in \mathbb{N}^{<\infty}$, we have \begin{align*} \|{\mathbf 1}_{A}\|_{{\mathbb X}}=&\|{\mathbf 1}_{2A-1}\|_{{\mathbb Z}}\le \mathbf{C}\|{\mathbf 1}_{2B-1}\|_{{\mathbb Z}}=\mathbf{C}\|{\mathbf 1}_{B}\|_{{\mathbb X}}. \end{align*} Hence, ${\mathcal B}_1$ is $\mathbf{C}$-$\mathbf{w}$-democratic. Similarly, ${\mathcal B}_2$ is $\mathbf{C}$-$\mathbf{w}$-democratic. Now fix $A,B\in \mathbb{N}^{\infty}$ with $w(A)\le w'(B)$. We have \begin{align*} \|{\mathbf 1}_{A}\|_{{\mathbb X}}=&\|{\mathbf 1}_{ 2A-1}\|_{{\mathbb Z}} \le \mathbf{C}\|{\mathbf 1}_{2B}\|_{{\mathbb Z}}=\mathbf{C}\|{\mathbf 1}_{B}\|_{{\mathbb Y}}. \end{align*} The case $w'(A)\le w(B)$ is proven in the same manner.\\ Now suppose that ${\mathcal B}_1$ is $\mathbf{C}_1$-$\mathbf{w}$-democratic and ${\mathcal B}_2$ is $\mathbf{C}_2$-$\mathbf{w}$-democratic, and let $\mathbf{C}$ be as in the statement. Fix $A,B\in \mathbb{N}$ with $W(\mathbf{w},\mathbf{w}')(A)\le W(\mathbf{w},\mathbf{w}')(B)$, and define \begin{align*} &A_1:=\{n\in \mathbb{N}: 2n-1\in A\}, &&A_2:=\{n\in \mathbb{N}: 2n\in A\};\\ &B_1:=\{n\in \mathbb{N}: 2n-1\in B\}, &&B_2:=\{n\in \mathbb{N}: 2n\in B\}. \end{align*} Since $W(\mathbf{w},\mathbf{w}')(A)=w(A_1)+w'(A_2)$ and the same holds for $B$, we have \begin{align*} w(A_1)+w'(A_2)\le& w(B_1)+w'(B_2)\le 2\max\{w(B_1),w'(B_2)\}. \end{align*} Suppose first that the maximum above is $w(B_1)$. Then $\max\{w(A_1),w'(A_2)\}\le 2w(B_1)$. Taking $\lambda$ and $\lambda'$ with respect to the basis ${\mathcal B}$ of ${\mathbb Z}$, an application of Lemma~\ref{lemmawdemdouble} yields \begin{align*} \|{\mathbf 1}_{A}\|_{{\mathbb Z}}=&\max\{\|{\mathbf 1}_{A_1}\|_{{\mathbb X}},\|{\mathbf 1}_{A_2}\|_{{\mathbb Y}}\} \le \max\{(2\mathbf{C}_1+\lambda\lambda')\|{\mathbf 1}_{B_1}\|_{{\mathbb X}},(2\mathbf{C}+\lambda\lambda')\|{\mathbf 1}_{B_1}\|_{{\mathbb X}}\}\\ \le& (\lambda\lambda'+2\max\{\mathbf{C},\mathbf{C}_1\})\|{\mathbf 1}_{B}\|_{{\mathbb Y}}. \end{align*} The same argument holds if the maximum is $w'(B_2)$ - we just get $\mathbf{C}_2$ instead of $\mathbf{C}_1$ in the upper bound. Thus, we conclude that ${\mathcal B}$ is $(\lambda \lambda'+2\max\{\mathbf{C},\mathbf{C}_1,\mathbf{C}_2\})$-${\mathbf{W}}(\mathbf{w},\mathbf{w}')$-democratic. \end{proof} \begin{corollary}\label{corollaryexamplescombined} Let $\mathbf{w},\mathbf{w}'$ be weights, with $\mathbf{w}\in \mathtt{c}_{0}$ and $\mathbf{w}'$ seminormalized. The following hold: \begin{enumerate}[\rm (i)] \item \label{corollaryexamplescombinedwalmost}There exists a conditional ${\mathbf{W}}(\mathbf{w},\mathbf{w}')$-almost greedy Schauder basis. \item \label{corollaryexamplescombinedwalmostsemiSch}There exists a conditional ${\mathbf{W}}(\mathbf{w},\mathbf{w}')$-almost semi-greedy Schauder basis that is not quasi-greedy. \item \label{corollaryexamplescombinedwalmostsemi}There exists a conditional ${\mathbf{W}}(\mathbf{w},\mathbf{w}')$-almost semi-greedy basis that is not quasi-greedy nor, in any order, a Schauder basis. \end{enumerate} \end{corollary} \begin{proof} Due to the equivalence of the above properties for equivalent weights, we may assume that $\mathbf{w}'$ is constant. To prove \ref{corollaryexamplescombinedwalmost}, pick $1<p<\infty$, and let ${\mathcal B}_p$ be the basis of the space ${\mathbb X}_p$ defined using \eqref{Xp}. Now choose a conditional almost greedy Schauder basis with $\|{\mathbf 1}_{A}\|\approx |A|^{\frac{1}{p}}$ for all $A\in \mathbb{N}^{<\infty}$ (for example, apply \cite[Theorem 4.9]{AABBL2021}), and then apply Lemma~\ref{lemmasumofspaces}. Now \ref{corollaryexamplescombinedwalmostsemiSch} is proven by the same argument as \ref{corollaryexamplescombinedwalmost}, with the only difference that, as the second basis in our construction, instead of a conditional almost greedy Schauder basis we choose an almost semi-greedy Schauder basis that is not quasi-greedy (in light of Corollary~\ref{corollarybidem}, bases with such properties can be found in \cite[Proposition 3.17]{AABBL2021}). Finally, \ref{corollaryexamplescombinedwalmostsemi} is again proven using the above construction, but in this case, our second basis is almost semi-greedy but neither quasi-greedy nor, in any order, a Schauder basis (for example, we can find such bases in \cite[Theorem 3.13]{AABBL2021}). \end{proof} \section{Open Questions} In Theorem~\ref{theoremc0notl1norming}, we proved that if $\mathbf{w}\in \mathtt{c}_{0}\setminus \ell_1$, ${\mathcal B}^*$ is norming and ${\mathcal B}$ is (weak) $\mathbf{w}$-semi-greedy, ${\mathcal B}$ is also $\mathbf{w}$-almost greedy. However, we do not know whether this holds in general, so a salient question is whether it does. In light of Corollary~\ref{corollaryPropCwA} and \cite[Theorem 4.3]{DKTW2018}, this is equivalent to ask whether such bases are quasi-greedy. \begin{question}\label{questionremainingcase} Let $\mathbf{w}\in \mathtt{c}_{0}\setminus \ell_1$. Is every (weak) $\mathbf{w}$-semi-greedy Markushevich basis quasi-greedy, and thus $\mathbf{w}$-almost greedy? \end{question} There are further questions about the case $\mathbf{w}\in \mathtt{c}_{0}\setminus \ell_1$ that arise from our research: We proved that in that case, if ${\mathcal B}$ is a (weak) $\mathbf{w}$-(almost) semi-greedy basis, it is $\mathbf{w}$-superdemocratic and truncation quasi-greedy, but in the proof we used a non-constructive argument and we were not able to obtain quantitative results, that is upper bounds for the $\mathbf{w}$-superdemocracy and truncation quasi-greedy constants depending on the (weak) $\mathbf{w}$-(almost)-semi-greedy constant and other known and simple properties of ${\mathcal B}$. \begin{question}\label{questionquantitative} Let $\mathbf{w}\in \mathtt{c}_{0}\setminus \ell_1$ and suppose ${\mathcal B}$ is $t$-$\mathbf{w}$-(almost) semi-greedy. Is there an upper bound for the $\mathbf{w}$-superdemocracy constant and the truncation quasi-greedy constant depending on the $t$-$\mathbf{w}$-(almost) semi-greedy constant and perhaps other, simple properties of the basis? \end{question} The weighted variants of (weak) almost (semi) greediness, democracy and superdemocracy are all preserved under equivalent weights. On the other hand, given Question~\ref{questionremainingcase}, the following problem remains open. \begin{question}\label{questionequivalentweights}Let $\mathbf{w}$ and $\mathbf{w}'$ be equivalent weights, with $\mathbf{w}, \mathbf{w}'\in \mathtt{c}_{0}\setminus \ell_1$, and let $0<s\le 1$. If ${\mathcal B}$ is $s$-$\mathbf{w}$-semi-greedy, is it $s$-$\mathbf{w}'$-semi-greedy? \end{question}

Background ========== Diabetic neuropathy is the most common complication of diabetes. It usually has a distal symmetric pattern that slowly progresses proximally with major implications to quality of life. It affects the function, mood and sleep patterns of the majority of affected patients. Symptoms include neuropathic pain, burning, tingling, decreased sensation, and loss of temperature perception \[[@B1]-[@B4]\]. Neuropathic pain is present in 30% to 50% of patients with diabetes \[[@B5],[@B6]\]. Glycemic control has been shown to be effective in slowing the progression of diabetic neuropathy \[[@B7]\]. Patients with painful neuropathy often need other agents to palliate their symptoms. Agents used include tricyclic antidepressants, anticonvulsants, serotonin-norepinephrine reuptake inhibitors, opiates, opiate-like substances and topical medications. However, these medications usually at best provide only partial pain relief \[[@B2],[@B5],[@B6]\]. There is no clear evidence-based guidance about treatment selection of analgesic agents for painful diabetic neuropathy. Stepwise approach and algorithms may be used, but comparative effectiveness of treatments to control pain in patients with diabetic neuropathy is unclear and there is no evidence of relative superiority across the different drug classes or the individual agents. We therefore decided to conduct a systematic review to appraise and summarize the totality of evidence regarding the efficacy of the most common non-parenteral (oral and topical) analgesic therapeutic options for the treatment of diabetic neuropathy. Methods ======= Objectives ---------- Considering the large number of pharmacologic therapies and the availability of multiple systematic reviews that identified the efficacy trials of these drugs, we decided to follow an 'umbrella' \[[@B8],[@B9]\] approach to identify eligible randomized controlled trials (RCTs). In brief, we will identify systematic reviews that compare available therapeutic options for painful diabetic neuropathy to placebo or any other active comparators. Eligible systematic reviews will be retrieved and used to identify relevant RCTs. We will subsequently build a dataset comprising eligible RCTs to obtain, critically appraise for quality, and extract data from each RCT to be used for the meta-analyses of relevant outcomes. Extracted data will be combined using a network design to draw inferences about the absolute and relative efficacy (from direct and indirect comparisons) of individual therapeutic options (Figure  [1](#F1){ref-type="fig"}). {#F1} Umbrella systematic review -------------------------- ### Interventions Two study investigators with expertise in diabetes (VMM and AT) provided input about the most common drugs used for diabetic neuropathy in USA and Europe. A list describing the most commonly used therapies deemed eligible for this review, with their minimum effective dose and therapeutic range based on American Diabetes Association (ADA) recommendations or expert opinion is presented in Table  [1](#T1){ref-type="table"}. For a trial to be included, the intervention dose evaluated must be at least the minimum effective dose cited in Table  [1](#T1){ref-type="table"}. This was set to reduce the risk of bias when an agent is compared to an ineffective dose of a competitor. If more than one dose is evaluated within the same RCT, we will include the results of patients treated with the highest dose in the efficacy comparisons, again ensuring that this dose is within the usual or recommended use of the agent. ###### Most commonly used therapies for treatment of diabetic neuropathy **Class** **Medication** **Minimum effective dose (total daily dose)** ------------------------------------------ ----------------------- ----------------------------------------------- Tricyclics Amitriptyline 50 mg   Desipramine 100 mg   Imipramine 100 mg   Nortriptyline 50 mg Selective serotonin  reuptake inhibitors Duloxetine 60 mg   Paroxetine 40 mg   Venlafaxine^a^ 75 mg Anticonvulsants Carbamazepine 600 mg   Gabapentin 900 mg   Lamotrigine^a^ 100 mg   Oxcarbazepine^a^ 600 mg   Pregabalin 150 mg   Sodium valproate^a^ 400 mg   Topiramate^a^ 100 mg Topical analgesics Capsaicin 0.0075%     Doxepin^a^     Lidocaine 5% patch^a^     Mexiletine^a^     Pentoxifyline^a^   Analgesic opiates Morphine^a^ 15 mg   Oxycodone 20 mg   Tapentadol^a^ 100 mg   Tramadol 100 mg Aldose reductase  inhibitors^a^ Epalrestat 150 mg   Ranirestat 20 mg   Fidarestat 1 mg   Ponalrestat 600 mg   Sorbinil 250 mg Various Lacosamide^a^ 100 mg ^a^No dose recommended by the American Diabetes Association. Study design ------------ During the umbrella stage of this systematic review, we will perform a literature search for all systematic reviews (any review with a clear and relevant clinical question, an explicit strategy and list of results) summarizing evidence from randomized trials testing treatments for painful diabetic neuropathy or neuralgia (Additional file [1](#S1){ref-type="supplementary-material"}). From these reviews, we will identify eligible trials. These should enroll patients with painful distal symmetrical polyneuropathy and randomly allocate these patients to treatment and control while keeping patients unaware of the therapy to which they have been allocated. There will be no restriction based on language of publication, number of patients or type of diabetes. Figure  [2](#F2){ref-type="fig"} illustrates the process of study selection. {#F2} Information sources and search methods -------------------------------------- We will conduct a comprehensive literature search for systematic reviews published in the last 5 years (2007 to 2011) for diabetic neuropathy or neuralgia in major electronic databases (MEDLINE™ via Ovid, EMBASE via Ovid, and the Cochrane Database of Systematic Reviews). Two study investigators with experience in systematic reviews (VMM and MHM) and an expert reference librarian developed the search strategy (Additional file [1](#S1){ref-type="supplementary-material"}) using a combination of controlled vocabulary (MeSH terms) and key words for the concepts of treatment of neuralgia or diabetic neuropathy. Two reviewers working independently will identify by title and abstract screening systematic reviews for the treatment of diabetic neuropathy. We will retrieve systematic reviews and meta-analyses deemed to be relevant and review them in full text to extract all eligible RCTs. These will be retrieved in full text for further assessment. Two reviewers will independently determine RCT eligibility. Disagreement will be solved by consensus, or if consensus is not achieved a third reviewer will serve as an arbitrator. We will assess chance-adjusted agreement (κ statistic) for each step. We will extract descriptive, methodological and analgesic efficacy data in duplicate. Outcomes -------- The main outcome of this study is improvement of pain. We are going to assess this outcome both as a dichotomous (for example, percentage of patients experiencing an improvement of pain of at least ≥30%) and a continuous variable (for example, standardized mean difference on a pain scale). If both dichotomous and continuous outcome data are available from the same trial, data for each variable will be analyzed separately. If pain is reported both in upper and lower extremities, we will analyze data only for the lower extremities. We understand that several outcomes describing pain may be reported and we have therefore decided to build a hierarchy of the pain outcomes we will consider. We will prefer, in decreasing order of relevance, intensity, overall pain, quality, duration, and timing. If pain is reported at multiple time points, we will assess efficacy of treatment at subacute (≤3 months) and chronic timeframes (anything \>3 months). If multiple assessments are reported, we will use the longest time point within the aforementioned categories. We expect a significant variety of scales measuring our outcomes of interest. We will standardize all scales, hence creating a unitless metric of pain, by expressing differences in number of standard deviations (standardized difference), which will allow comparison among the different scales (Figure  [3](#F3){ref-type="fig"}). {#F3} For parallel-arm placebo-controlled trials and comparative effectiveness head-to-head trials, differences will be calculated between arms. Crossover trials are going to be analyzed according to recommendations from the Cochrane Handbook for Systematic Reviews of Interventions \[[@B8]\]. The network approach will generate a ranking list based on efficacy. According to the results we will rank all the included medications from most to least effective based on their effect on pain scores. Data collection and extraction ------------------------------ Data from RCTs will be extracted in duplicate, by independent reviewers, using a standardized, piloted, web-based data extraction form. We will re-extract data from each randomized clinical trial on patient demographics, baseline characteristics (duration of diabetes, hemoglobin A~1c~ (HbA~1c~)), study design, sample size, intervention type (class, type and dosages), scale type, reporting of functional status outcomes or impaired sleep pattern. Risk of bias assessment ----------------------- We will use the Cochrane Collaboration's risk of bias tool to evaluate the methodological quality of RCTs. Two reviewers will independently assess risk of bias using the following domains: random sequence generation (adequate if based on computer generated random numbers, tables of random numbers or similar), allocation concealment (adequate if based on central randomization, sealed envelopes or similar), blinding of patients care givers or outcome assessors, incomplete outcome data (due to high rate of discontinuation, loss to follow-up, type of analysis, or imputation of missing data), selective reporting and other bias (source of funding, whether by not-for-profit or for-profit sources.) Disagreements between reviewers will be resolved by discussion or will be arbitrated by a third reviewer (MHM). We will summarize the risk of bias at all six domains to produce an overall risk of bias for every trial. This will be deemed high if there is presence of high bias in any key domain (allocation concealment or blinding of patients), low if risk of bias is low for all key domains, and unclear in all other cases. We chose a priori to consider allocation concealment and blinding as key quality domains, due to their relative importance for assessment of subjective outcomes such as pain, and have formulated an a priori rule to summarize risk of bias for every class of outcomes across domains, based on scoring of risk of bias in every domain of the Cochrane Risk of Bias tool \[[@B8]\]. Grading of quality of evidence ------------------------------ We will use the GRADE ('Grading of Recommendation, Assessment, Development and Evaluation') framework to evaluate the quality of evidence. Evidence derived from the RCTs will be initially considered high (at low risk of bias) and subsequently downgraded based on methodological limitations, imprecision, inconsistency, indirectness and publication bias \[[@B10]\]. A final generated grade will reflect the confidence we have in the reported estimate of effect. Statistical analysis -------------------- We will pool odds ratios (ORs) for pain improvement reported as a dichotomous outcome and standardized mean differences (SMDs) using Hedges' adjusted g for pain scores based on different scales from included studies, accompanying 95% confidence intervals (CI). Then, we will conduct pairwise meta-analyses and estimate the combined efficacy of treatment using a random effects model using the method of DerSimonian and Laird, with the estimate of heterogeneity being taken from the Mantel-Haenszel model \[[@B11]\]. Network meta-analyses will be adopted to combine direct and indirect evidence using Lumley's generalized linear mixed models based on Frequentists' method. The Bayesian Markov Chain Monte Carlo method will also be used as an alternative statistical method to validate the results. Generalized linear mixed models ------------------------------- We will construct linear mixed models to combine direct and indirect evidence. First introduced by Lumley \[[@B12]\], linear mixed models can be used to estimate the effects of an indirect comparison with two or more linking treatments (paths). For example, to pool the indirect evidence between treatment A and treatment B, we can combine estimates from treatment A to treatment C, treatment B to treatment C, treatment A to treatment D, and treatment B to treatment D. Another advantage of linear mixed models is it can evaluate the agreement of indirect comparisons through different linking treatments, also called 'incoherence', and incorporate the incoherence in the calculation of confidence interval of the pooled estimate. If there is direct evidence, linear mixed models also allow us to pool direct and indirect evidence and get more efficient estimate. R statistical software V. 2.15.1 will be used to fit generalized linear mixed models. Dealing with missing data ------------------------- It is likely that some eligible RCTs will not report all relevant data such as the standard deviation or other important variability measures \[[@B13]\]. In order to include these studies in the analysis we will try first to calculate them through algebraic manipulation of the available information such as confidence intervals, p or t values. If that was not possible, we will attempt to contact study authors for these data. If that approach also failed we will borrow these measures from other similar studies as Furukawa *et al*. suggested. When assumptions are made in estimating these measures, such assumptions will be tested in sensitivity analysis \[[@B14]\]. Subgroup analysis ----------------- To investigate potential heterogeneity across studies, we will conduct subgroup analysis based on class of medicine, quality of evidence (low and unclear risk of bias vs high risk of bias), oral versus topical agents, HbA~1c~ level, duration of diabetes, and patients' age. Random effects meta-regression models will be used to quantify the difference between subgroups and test for statistical significant interactions among subgroups. Sensitivity analysis -------------------- We will conduct sensitivity analysis to assess the effects of individual studies on the combined estimates and determine whether certain studies dominate the pooled event (particularly if these studies were at high risk of bias). Any imputations made for measures of dispersion (that is, standard deviation) of the difference in pain outcome will be tested in sensitivity analysis. If correlation coefficients are needed to estimate standard deviations, different coefficients will be tested. If certain studies have borderline eligibility status for any reason, analysis will be conducted with and without such studies. Finally, a major sensitivity analysis would be to determine whether Bayesian analysis results are different from the Frequentist method ones. Robustness of analysis (consistency throughout sensitivity analyses and assumptions) would increase our confidence in the comparative effectiveness analysis. We also plan to use the Bayesian Markov Chain Monte Carlo method to combine direct and indirect evidence and fit the model in WinBUGS V. 1.4.3 \[[@B15],[@B16]\]. This method allows for the pooling of indirect and direct comparisons of effect between the different classes of therapy for diabetic neuropathy, while preserving within-trial randomization, and allowing comparisons of agents not addressed within any of the individual trials \[[@B15],[@B16]\]. A random-effects model will be fitted due to potential for heterogeneity among included trials. The posterior distribution of all parameters will be estimated using non-informative priors, in order for results to be represented solely by the included data. Results will be based from 100,000 iterations, after a 50,000-iteration burn-in. Inconsistency will be evaluated by comparing the estimates from direct comparisons and those from the indirect comparisons for the magnitude and direction of effect. Appropriateness of model fit will be evaluated using the residual deviance, where good model fit is represented by the residual deviance value approximating the number of unconstrained data points \[[@B17]\]. The Bayesian Markov Chain Monte Carlo method will provide us an alternative statistical method to validate our results. We will use the *I*^2^ statistic and Cochran's Q test to assess heterogeneity across studies. We will explore heterogeneity for potential intraclass differences. Finally, publication bias and small study effect will be assessed, whenever possible, using the Egger regression asymmetry test and the Begg adjusted rank correlation test \[[@B18],[@B19]\]. All statistical analyses, except generalized linear mixed models, and the Bayesian Markov Chain Monte Carlo method, will be conducted using STATA V. 12.0. Reporting --------- The study will be reported in accordance with the recommendations set forth by the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) work groups \[[@B20]\]. We will present network graphs depicting the effect size for direct, indirect and combined analyses highlighting the quality and risk of bias in each network loop. Discussion ========== This network meta-analysis aims to synthesize available direct and indirect evidence of effectiveness of analgesics in the treatment of painful diabetic neuropathy. We expect to use evidence of inconsistency between direct and indirect comparisons to identify publication bias affecting the literature on head-to-head comparisons. The network approach will offer the opportunity to generate a ranking based on efficacy. This, along with known side effects, costs, and administration burdens will enable patients and clinicians to make choices that best reflect this evidence, their context, and the informed preferences of patients with painful diabetic neuropathy. Competing interests =================== The authors declare they have no conflicts of interest. Authors' contributions ====================== MLG, AT and JPB developed the search criteria, selection criteria, specified outcome measures determined the data to be abstracted and drafted the manuscript. ZW and OJP determined the analyses and interpretation of data to be conducted. VMM and MHM drafted the manuscript and did critical revisions. All authors read, provided feedback and approved the final manuscript. Supplementary Material ====================== ###### Additional file 1 EMBASE/MEDLINE. ###### Click here for file Acknowledgements ================ This protocol was funded by the CTSA and the Mayo Foundation for Medical Education and Research.

\section{Introduction} Among the fascinating flow phenomena occurring on small scales \cite{stone_review,squiresquake}, super-hydrophobicity offers a unique bridge between microscopic features and macroscopic behavior \cite{feng02,quere05,sun05,quere08}. Super-hydrophobic surfaces are non-wetting surfaces which possess sufficiently-large geometrical roughness that a liquid droplet {deposited on the surface would not fill the grooves of the surface roughness, but instead remain in a a fakir-like state where the droplet only touches the surface at the edge of the roughness }(Fig.~\ref{general}a). As a result, super-hydrophobic surfaces possess very high effective contact angles, and display remarkable macro-scale wetting properties \cite{feng02,quere05,sun05,quere08}. One particularly interesting characteristic of super-hydrophobic surfaces is their low viscous friction. Since fluid in contact with the surface only barely touches it, but is instead mostly in contact with the surrounding air, small droplets can roll very easily, a phenomenon known as the lotus-leaf effect. In general, super-hydrophobic surfaces are expected to provide opportunities for significant drag reduction in the laminar regime, as has been confirmed by experiments cited below. The effective viscous friction of a solid surface is usually quantified by a so-called slip length, denoted here by $\lambda$, which is the distance below the solid surface where the no-slip boundary would be satisfied if the flow field was linearly extrapolated, and the no-slip boundary condition corresponds to $\lambda =0$ \cite{neto05,laugareview,bocquet07}. At low Reynolds number, the only characteristics affecting the friction of super-hydrophobic surfaces arise from their geometry, specifically (a) the distribution of liquid/solid and liquid/air contact at the edge of the roughness elements of the surface, and (b) the shape of the liquid/air free surface. For {one-dimensional} surfaces, {\it i.e.} surfaces which have one homogeneous direction (Fig.~\ref{general}b), significant drag reduction can be obtained when the homogeneous direction is parallel or perpendicular to the direction of flow, as demonstrated experimentally \cite{ou04,ou05,gogte05,choi06,truesdell06,maynes07,lee08} and theoretically \cite{philip72,philip72_2,lauga03,cottin-bizonne04,davies06,maynes07,ybert07,ng09}. \begin{figure}[t] \begin{center} \includegraphics[width=0.25\textwidth]{fig1.eps} \end{center} \caption{(color online) Typical topology of super-hydrophobic surfaces: (a) Side view of a random super-hydrophobic surface, illustrating the {contact of the liquid (gray) at the edge of the surface roughness (blue)}; (b) Top view of a {one-dimensional} surface; (c) Top view of regular {two-dimensional} surfaces where the air (white) is the continuous phase; (d) Top view of regular {two-dimensional} surfaces where the solid (blue) is the continuous phase. The case illustrated on Fig.~\ref{general}d (top) is the focus of the current paper.} \label{general} \end{figure} For {two-dimensional} surfaces, an additional free parameter is the topology of the air/solid partition on the planar surface, and two general types can be distinguished. In the first type, the air is the continuous phase, and the liquid is in contact with the solid only at isolated, unconnected, locations (Fig.~\ref{general}c). This is the most commonly-studied type of super-hydrophobic surface, and arises for surface roughness in the shape of bumps or posts on an otherwise flat solid surface \cite{ou04,ou05,gogte05,choi06_PRL,joseph06,ybert07,lee08,ng09_2}. The second type of surface, less studied, is one where the solid is the continuous phase, and the liquid/air contact occurs on isolated domains (Fig.~\ref{general}d). These surfaces can be obtained by making holes in an otherwise flat material \cite{ng09_2}, and have been used to study the influence of the geometry of the liquid/air free interface on the viscous friction of the surface \cite{steinberger07,hyvaluoma08,legendre08_2,davis09}. A super-hydrophobic surface with a continuous solid phase can also be obtained by using an intertwined solid mesh. Recently this method has been exploited experimentally, using coated steel \cite{feng04,sun05}, and coated copper \cite{wang07}, as reproduced in Fig.~\ref{grid}a and b. The resulting surface is also the relevant geometrical configuration for flow past textile and fabric material. In this paper, we present an analytical method to calculate the friction of such a mesh-like (or grid-like) surface in the case where the mesh is thin in the plane (mesh aspect ratio $\epsilon \ll 1$). {The analysis, based on a distribution of flow singularities, leads to an infinite system of linear equations for the approximate Fourier coefficients of the flow around the mesh, and gives results with relative error of order $\epsilon \ln (1/\epsilon)$}. The results for the effective slip length of the surface are computed, compared to simple estimates, and used to propose a practical fit displaying the expected logarithmic dependance on the solid {area} fraction \cite{philip72,philip72_2,lauga03,ybert07}. \begin{figure}[t] \begin{center} \includegraphics[width=0.45\textwidth]{fig2.eps} \end{center} \caption{(color online) Mesh-like super-hydrophobic surface. (a): Coated copper mesh at the micron scale \cite{wang07}; (b): The mesh from Figs.~\ref{grid}a is super-hydrophobic (water droplet, volume about 4 $\mu$l) \cite{wang07}; (c): For our calculation we consider a planar square mesh of periodicity $b$ and width $\epsilon b$; (d): A shear flow is set up in one of the principal directions of the mesh, with shear rate $\dot{\gamma}$; the far-field flow is given by ${\bf v}=\dot\gamma (z + \lambda) {\bf e}_x$, where $\lambda$ is the effective slip length of the mesh. Figs.~\ref{grid}a and \ref{grid}b reprinted from S. Wang, Y. Song and L. Jiang (2007) ``Microscale and nanoscale hierarchical structured mesh films with super-hydrophobic and superoleophilic properties induced by long-chain fatty acids'', {\it Nanotechnology}, {\bf 18}, 015103, courtesy of Shutao Wang and with permission from IOP Publishing. } \label{grid} \end{figure} \section{Shear flow past a rectangular grid} \subsection{Calculation background} \label{history} Our paper follows classical work on the broadside motion of rigid bodies at low Reynolds number in the asymptotic limit for narrow cross-sections \cite{happel,kimbook}. Here, we build on the method of Leppington and Levine who studied axisymmetric potential problems involving an annular disk by using a distribution of singularities, and obtained an efficient method in the limit where the radii difference of the disk approaches zero \cite{leppington72}. Roger and Hussey studied the flat annular ring problem both experimentally and theoretically using a beads-on-a-shell model to represent a distribution of point forces \cite{roger82}. Subsequently, Stewartson showed that viscous fluid exhibits a marked reluctance to flow through a thin torus and the flux function has an essential singularity in the limit in which the hollow boundary approaches a circle and disappears \cite{stewartson83}. This `blockage' feature was further demonstrated by Davis and James in considering flow through a array of narrow annular disks placed in planes normal to the flow in order to model a fibrous medium \cite{davis96}. Their analysis was made tractable by retaining only the inverse square root term in the force density function. A similar asymptotic estimate was used by Davis in modeling the broadside oscillations of a thin grid of the type discussed below \cite{davis93}. Since edgewise motions induce the same edge singularity in the force density, {we follow in this paper a similar analytical treatment}. \subsection{Model problem} We consider a thin stationary planar square mesh in the $(x,y)$ plane subject to a shear flow, with shear rate $\dot\gamma$, along the $x$ direction, which is one of the principal directions of the mesh, as illustrated in Figs.~\ref{grid}c and d. The mesh has periodicity $b$ and a small width $\epsilon b$. We denote by ${\bf e}_x$ and ${\bf e}_y$ the unit vectors parallel to the mesh axes, and ${\bf e}_z$ is the third Cartesian vector. We further assume that the shape of the interface between the liquid and the air located underneath the mesh is planar, with no protrusion of the mesh inside the fluid. Since the mesh is a super-hydrophobic surface and presents to the flow a combination of no-slip domains (the mesh) and perfectly-slipping domains (the free surfaces), the effective flow past the surface can be described as a slipping flow, and is given in the far field by ${\bf v}=\dot\gamma (z + \lambda) {\bf e}_x$, where $\lambda$ is the effective surface slip length ($\lambda > 0$). The goal of the calculation below is to present an accurate analytic calculation for $\lambda$ in the limit where $\epsilon\ll 1$. In order to perform the calculation, we move in the frame translating at the steady velocity $U{\bf e}_x$, with $ U=\dot\gamma \lambda$. In that case, the problem is equivalent to the uniform translation of an thin mesh in its own plane at velocity $-U{\bf e}_x$, and the goal is to derive the value of the shear rate $\dot\gamma$ in the far-field, $|z| \gg b$. The slip length will then be given by $\lambda = U/\dot\gamma$. Edge effects are ignored by assuming an infinite mesh as then periodic point force singularities can be used to describe the fluid flow field. The rectangular grid naturally introduces Fourier series with respect to Cartesian coordinates, and the analysis establishes a set of integral equations with the same logarithmic dependence on the geometrical parameters of the grid as in the flows described {in the references discussed in \S\ref{history}.} The Reynolds number of the viscous incompressible flow is assumed to be sufficiently small for the velocity field ${\bf v}$ to satisfy the creeping flow (Stokes) equations \cite{happel,kimbook} \begin{equation}\label{stokes} \mu\nabla^2{\bf v}=\nabla p, \qquad \nabla\cdot{\bf v}=0, \end{equation} where $\mu$ is the coefficient of viscosity and $p$ is the dynamic pressure. The prescribed grid velocity is given by \begin{equation}\label{Gvel} {\bf v}=-U{\bf e}_x \hbox{ at }z=0, \qquad (x,y)\in G. \end{equation} In a typical grid element $0\leq x,y\leq b$, $G$ is complementary to the square hole bounded by the lines $x$ or $y=\frac{1}{2}\epsilon b$ or $(1-\frac{1}{2}\epsilon) b$ (shown in black in Fig.~\ref{zoom}a). \begin{figure}[t] \begin{center} \includegraphics[width=0.3\textwidth]{fig3.eps} \end{center} \caption{(color online) A typical grid element (see text for details).} \label{zoom} \end{figure} \subsection{Solution using a superposition of singularities} {Following the calculation for broadside oscillations of the grid in Ref.~\cite{davis93},} the fluid motion can be represented as due to a distribution of tangentially directed Stokeslets over the flat mesh $G$ and the density functions must be both periodic in two dimensions and symmetric with respect to the sides of each square. The field due to a two-dimensional square array, period $b$, of point forces of strength $4\pi\mu Ub$ directed parallel to the mesh's motion, is governed by \begin{equation}\label{div} \nabla\cdot{\bf v}_A=0 \end{equation} and \begin{eqnarray}\label{forcearray} \mu\nabla^2{\bf v}_A-\nabla p_A&=&4\pi\mu Ub {\bf e}_x\delta(z) \sum_{n_1=-\infty}^\infty\sum_{n_2=-\infty}^\infty\delta(x-n_1b) \delta(y-n_2b)\nonumber\\ &=&\frac{4\pi\mu U}{b} {\bf e}_x \delta(z)\sum_{m_1=-\infty}^\infty \sum_{m_2=-\infty}^\infty\exp\left[\frac{2\pi i}{b}(m_1x+m_2y)\right]. \end{eqnarray} With ${\bf m}=m_1{\bf e}_x+m_2{\bf e}_y$, the ${\bf m}={\bf 0}$ term in Eqs.~\eqref{div} and (\ref{forcearray}) yields \begin{equation} \mu\nabla^2{\bf v}_0-\nabla p_0=\frac{4\pi\mu U}{b}{\bf e}_x \delta(z), \qquad \nabla\cdot{\bf v}_0=0, \end{equation} whose solution, \begin{equation} {\bf v}_0=\frac{2\pi U|z|}{b}{\bf e}_x, \qquad p_0=0, \end{equation} exhibits the anticipated shear at infinity. Note that the suppression of the immaterial arbitrary multiple of $U{\bf e}_x$ in ${\bf v}_0$ ensures uniqueness below. The solution of \begin{equation} \mu\nabla^2{\bf v}-\nabla p=\frac{2\mu U}{b}{\bf e}_x \int_{-\infty}^\infty e^{ikz}dk\sum_{{\bf m}}{}'\exp\left[\frac {2\pi i}{b}{\bf m}\cdot{\bf r}\right], \qquad \nabla\cdot{\bf v}=0, \end{equation} with a prime denoting that the ${\bf m}={\bf 0}$ term is omitted, is readily found by Fourier transform techniques (see Ref.~\cite{hasimoto59}). Thus the flow governed by Eqs.~\eqref{div} and (\ref{forcearray}) is compactly expressed as \begin{equation}\label{arrayflow} {\bf v}_A=U\left[\left(\frac{2\pi|z|}{b}-S_1\right){\bf e}_x+ \nabla\frac{\partial S_2}{\partial x}\right], \qquad p_A=\mu U\frac{\partial S_1}{\partial x}, \end{equation} where \begin{equation}\label{ess1} S_1=\sum_{{\bf m}}{}'\frac{1}{|{\bf m}|}\exp\left[\frac {2\pi}{b}(i{\bf m}\cdot{\bf r}-|{\bf m}||z|)\right]=\nabla^2 S_2, \end{equation} with \begin{eqnarray}\label{ess2} S_2&=&-\frac{2}{b}\int_{-\infty}^\infty e^{ikz}dk\sum_{{\bf m}}{}' \exp\left[\frac{2\pi i}{b}{\bf m}\cdot{\bf r}\right] \left[\left(\frac{2\pi i}{b}{\bf m}\right)^2+k^2\right]^{-2}\nonumber\\ &=&-\frac{b^2}{8\pi^2}\sum_{{\bf m}}{}'\frac{1}{|{\bf m}|^2}\left( \frac{1}{|{\bf m}|}+\frac{2\pi}{b}|z|\right)\exp\left[\frac {2\pi}{b}(i{\bf m}\cdot{\bf r}-|{\bf m}||z|)\right]\cdot \end{eqnarray} Only the velocities at the mesh and at infinity are needed for the subsequent analysis. The solution Eq.~(\ref{arrayflow}) shows that \begin{equation}\label{veeAinf} {\bf v}_A\sim\frac{2\pi U|z|}{b}{\bf e}_x { ={\bf v}_0} \hbox{ as }|z|\to\infty, \end{equation} {since all terms in Eq.~\eqref{ess1} exhibit exponential decay,} and \begin{equation}\label{veeAgrid} [{\bf v}_A]_{z=0}=U{\bf e}_x\left(-S_1+\frac{\partial^2S_2} {\partial x^2}\right)_{z=0}=-U{\bf e}_x\sum_{{\bf m}}{}' \frac{1}{|{\bf m}|} \exp\left[\frac{2\pi i}{b}{\bf m}\cdot{\bf r}\right]C_{{\bf m}}, \end{equation} where, after substitution of Eqs.~(\ref{ess1}) and (\ref{ess2}), we have \begin{equation}\label{ceem} C_{{\bf m}}=1-\frac{m_1^2}{2|{\bf m}|^2}\cdot \end{equation} In particular, we see that \begin{subeqnarray}\label{ceeminf} C_{(m_1,0)}&=&\frac{1}{2}, \qquad C_{{\bf m}}\to\frac{1}{2} \hbox{ as } m_1\to\infty, \\ C_{(0,m_2)}&=&1, \qquad C_{{\bf m}}\to 1 \hbox{ as }m_2\to\infty. \end{subeqnarray} The shaded region in Fig.~\ref{zoom}b suffices for the distribution of periodic point forces and the enforcement of the prescribed mesh velocity (Eq.~\ref{Gvel}). The hexagon in Fig.~\ref{zoom}b lying along the $x$-axis is given by $|y|\leq\frac{1}{2}\epsilon b,|y|\leq x\leq b-|y|$, or $-1\leq w,s\leq 1$ in terms of new variables $(w,s)$ defined by \begin{equation}\label{newvars} w=\frac{\frac{1}{2}b-x}{\frac{1}{2}b-y}, \qquad y=\frac{1}{2}\epsilon bs, \qquad x=\frac{1}{2}b[1-(1-\epsilon s)w]. \end{equation} Here $w$ may be identified as the tangent of an angle subtended at the center of the square. The hexagon in Fig.~\ref{zoom}b lying along the $y$-axis is given similarly by interchanging $x$ and $y$ in Eq.~(\ref{newvars}). The flow generated by the translating mesh {can be written as} \begin{eqnarray} \label{vee} {\bf v}&=&\frac{1}{4}\int_{-1}^1\int_{-1}^1\left\{f_x(w,s){\bf v}_A \left[x-\frac{1}{2}b(1-w+\epsilon sw),y-\frac{1}{2}\epsilon bs,z\right] \right.\nonumber\\ && \left. +f_y(w,s){\bf v}_A\left[x-\frac{1}{2}\epsilon bs, y-\frac{1}{2}b(1-w+\epsilon sw),z\right]\right\}\,{\rm d}}\def\c{\color{red} w{\rm d}}\def\c{\color{red} s, \end{eqnarray} {where $f_x$ and $f_y$ are dimensionless force densities.} The prescribed mesh velocity is then obtained by enforcing Eq.~(\ref{Gvel}) at each point of the two hexagons. Thus Eq.~(\ref{vee}) gives \begin{subeqnarray}\label{veegrid} -U{\bf e}_x &=& \frac{1}{4}\int_{-1}^1\int_{-1}^1\left\{f_x(w,s) {\bf v}_A\left[\frac{1}{2}b(w-W+\epsilon SW-\epsilon sw), \frac{1}{2}\epsilon b(S-s),0\right]\right. \nonumber\\ && \left. + f_y(w,s){\bf v}_A\left[\frac{1}{2}b(1-W+\epsilon SW- \epsilon s),\frac{1}{2}b(\epsilon S-1+w-\epsilon sw),0\right]\right\} \,{\rm d}}\def\c{\color{red} w{\rm d}}\def\c{\color{red} s, \quad \\ -U{\bf e}_x &=& \frac{1}{4}\int_{-1}^1\int_{-1}^1\left\{f_x(w,s) {\bf v}_A\left[\frac{1}{2}b(\epsilon S-1+w-\epsilon sw), \frac{1}{2}b(1-W+\epsilon SW-\epsilon s),0\right]\right. \nonumber \\ && \left. + f_y(w,s){\bf v}_A\left[\frac{1}{2}\epsilon b(S-s), \frac{1}{2}b(w-W+\epsilon SW-\epsilon sw),0\right]\right\}\,{\rm d}}\def\c{\color{red} w{\rm d}}\def\c{\color{red} s \end{subeqnarray} for all $ -1\leq W,S\leq 1$. We further note that the summation in Eq.~(\ref{veeAgrid}) takes the simpler form \begin{eqnarray} \sum_{{\bf m}}{}'\frac{1}{|{\bf m}|} \exp\left[\frac{2\pi i}{b}{\bf m}\cdot{\bf r}\right]C_{{\bf m}}= \sum_{m_1=1}^\infty \frac{1}{m_1}\cos\frac{2\pi}{b}m_1x + 2 \sum_{m_2=1}^\infty \frac{1}{m_2}\cos\frac{2\pi}{b}m_2y \nonumber\\ +4\sum_{m_1=1}^\infty\sum_{m_2=1}^\infty \frac{1}{\sqrt{m_1^2+m_2^2}} \left[1-\frac{m_1^2}{2(m_1^2+m_2^2)}\right] \cos\frac{2\pi}{b}m_1x \cos\frac{2\pi}{b}m_2y, \end{eqnarray} from which we note that large contributions to Eqs.~(\ref{veegrid}) arise from summations of type \begin{equation} \sum_{m=1}^\infty \frac{1}{m}\cos\pi m\epsilon(S-s){ = \ln\left[\frac{1}{2}\csc \frac{\pi \epsilon}{2}|S-s|\right]} \sim -\ln(\pi\epsilon|S-s|). \end{equation} Substitution of Eq.~(\ref{veeAgrid}) into Eqs.~(\ref{veegrid}) and use of Eqs.~(\ref{ceeminf}) now yields, when terms that tend to zero as $\epsilon\to 0$ are neglected, \begin{subeqnarray}\label{inteq2D} 1 &=& \frac{1}{4}\int_{-1}^1\int_{-1}^1\left\{f_x(w,s)\left[ \sum_{m_1=1}^\infty\frac{1}{m_1}\cos\pi m_1(w-W)-2\ln(\pi\epsilon|S-s|) \right.\right. \nonumber\\ &&\left. +4\sum_{m_1=1}^\infty\cos\pi m_1(w-W)\left\langle\sum_{m_2=1} ^\infty\left(\frac{1}{|{\bf m}|}C_{{\bf m}}-\frac{1}{m_2}\right)- \ln(\pi\epsilon|S-s|)\right\rangle\right] \nonumber \\ &&+f_y(w,s)\left[\sum_{m_1=1}^\infty \frac{(-1)^{m_1}}{m_1}\cos\pi m_1W +2\sum_{m_2=1}^\infty \frac{(-1)^{m_2}}{m_2}\cos\pi m_2w\right. \nonumber \\ && \left.\left.+4\sum_{m_1=1}^\infty\sum_{m_2=1}^\infty \frac{(-1)^{m_1+m_2}}{|{\bf m}|}C_{{\bf m}}\cos\pi m_1W\cos\pi m_2w \right]\right\}\,{\rm d}}\def\c{\color{red} w{\rm d}}\def\c{\color{red} s,\\ 1 &=& \frac{1}{4}\int_{-1}^1\int_{-1}^1\left\{f_x(w,s)\left[ \sum_{m_1=1}^\infty \frac{(-1)^{m_1}}{m_1}\cos\pi m_1w+ 2\sum_{m_2=1}^\infty\frac{(-1)^{m_2}}{m_2}\cos\pi m_2W\right.\right. \nonumber \\ &&\left.+4\sum_{m_1=1}^\infty\sum_{m_2=1}^\infty \frac{(-1)^{m_1+m_2}} {|{\bf m}|}C_{{\bf m}}\cos\pi m_1w\cos\pi m_2W\right] \nonumber\\ &&+f_y(w,s)\left[-\ln(\pi\epsilon|S-s|)+2\sum_{m_2=1}^\infty \frac{1}{m_2}\cos\pi m_2(w-W)\right. \nonumber \\ &&\left.\left. +4\sum_{m_2=1}^\infty\cos\pi m_2(w-W)\left\langle \sum_{m_1=1}^\infty\left(\frac{1}{|{\bf m}|}C_{{\bf m}}-\frac{1}{2m_1} \right)-\frac{1}{2}\ln(\pi\epsilon|S-s|)\right\rangle\right]\right\} \,{\rm d}}\def\c{\color{red} w{\rm d}}\def\c{\color{red} s\quad\quad\,\, \end{subeqnarray} for all $-1\leq W,S\leq 1$. All the $w$-integrals in these integral equations can be expressed in terms of Fourier coefficients defined as \begin{equation}\label{Fcoeffts} [f_{xn}(s),f_{yn}(s)]=\frac{1}{2}\int_{-1}^1 [f_x(w,s),f_y(w,s)] \cos n\pi w {\rm d}}\def\c{\color{red} w \qquad (n\geq 0), \end{equation} because the symmetric forcing ensures that the corresponding sine coefficients are all zero, {and where we use brackets to define simultaneously two sets of Fourier coefficients.} We then identify Eqs.~(\ref{inteq2D}) as a pair of Fourier cosine series in $W$, for each $S$ in $[-1,1]$, and, by considering coefficients of $\cos\pi m_1W$ and $\cos\pi m_2W$ in the respective series, we obtain \begin{subeqnarray}\label{inteqs} 1=-\int_{-1}^1 f_{x0}(s)\ln(\pi\epsilon|S-s|)ds+\sum_{m_2=1}^\infty \frac{(-1)^{m_2}}{m_2}\int_{-1}^1 f_{ym_2}(s){\rm d}}\def\c{\color{red} s, \\ 0=\int_{-1}^1 f_{xm_1}(s)\left[\frac{1}{2m_1}+2\sum_{m_2=1}^\infty \left(\frac{1}{|{\bf m}|}C_{{\bf m}}-\frac{1}{m_2}\right)-2 \ln(\pi\epsilon|S-s|)\right]{\rm d}}\def\c{\color{red} s\nonumber\\ +\frac{(-1)^{m_1}}{2m_1}\int_{-1}^1f_{y0}(s)ds+2\sum_{m_2=1}^\infty \frac{(-1)^{m_1+m_2}}{|{\bf m}|}C_{{\bf m}}\int_{-1}^1f_{ym_2}(s) {\rm d}}\def\c{\color{red} s, \qquad (m_1\geq 1), \\ 2=-\int_{-1}^1 f_{y0}(s)\ln(\pi\epsilon|S-s|)ds+\sum_{m_1=1}^\infty \frac{(-1)^{m_1}}{m_1}\int_{-1}^1 f_{xm_1}(s){\rm d}}\def\c{\color{red} s, \\ 0=\int_{-1}^1 f_{ym_2}(s)\left[\frac{1}{m_2}+\sum_{m_1=1}^\infty \left(\frac{2}{|{\bf m}|}C_{{\bf m}}-\frac{1}{m_1}\right)- \ln(\pi\epsilon|S-s|)\right]{\rm d}}\def\c{\color{red} s \nonumber\\ +\frac{(-1)^{m_2}}{m_2}\int_{-1}^1f_{x0}(s)ds+2\sum_{m_1=1}^\infty \frac{(-1)^{m_1+m_2}}{|{\bf m}|}C_{{\bf m}}\int_{-1}^1f_{xm_1}(s){\rm d}}\def\c{\color{red} s, \qquad (m_2\geq 1), \end{subeqnarray} for all $-1\leq S\leq 1$. These two related sets of equations have the form \begin{equation}\label{ABeqn1n} \int_{-1}^1 [f_{xn}(s),f_{yn}(s)]\ln\left(\frac{1}{|S-s|}\right){\rm d}}\def\c{\color{red} s =\hbox{ constant }=[A_n,B_n]\ln 2 \qquad (-1\leq S\leq 1), \end{equation} whose solution is \begin{equation} [f_{xn}(s),f_{yn}(s)]=\frac{[A_n,B_n]}{\pi(1-s^2)^{1/2}} \qquad (-1\leq s\leq 1), \end{equation} with the particular property \begin{equation}\label{ABeqn2n} \int_{-1}^1 [f_{xn}(s),f_{yn}(s)]{\rm d}}\def\c{\color{red} s=[A_n,B_n]. \end{equation} Thus, as demonstrated more rigorously by Leppington and Levine \cite{leppington72} and exploited frequently elsewhere, the logarithmic kernel obtained as an asymptotic estimate yields the inverse square root function as the associated asymptotic solution. The substitution of Eqs.~(\ref{ABeqn1n}) and (\ref{ABeqn2n}) into Eqs.~(\ref{inteqs}) finally yields an infinite system of linear equations \begin{subeqnarray}\label{ABeqn3n} 1&=&A_0\ln\left(\frac{2}{\pi\epsilon}\right)+\sum_{m_2=1}^\infty \frac{(-1)^{m_2}}{m_2}B_{m_2}, \slabel{a}\\ 0&=&A_{m_1}\left[\ln\left(\frac{2}{\pi\epsilon}\right)+\frac{1}{4m_1}+ \sum_{m_2=1}^\infty\left(\frac{1}{|{\bf m}|}C_{{\bf m}}-\frac{1}{m_2} \right)\right] \nonumber\\ && +\frac{(-1)^{m_1}}{4m_1}B_0+\sum_{m_2=1}^\infty\frac{(-1)^{m_1+m_2}} {|{\bf m}|}C_{{\bf m}}B_{m_2}, \quad (m_1\geq 1),\\ 2 & = & B_0\ln\left(\frac{2}{\pi\epsilon}\right)+\sum_{m_1=1}^\infty \frac{(-1)^{m_1}}{m_1}A_{m_1}, \slabel{b}\\ 0&=&B_{m_2}\left[\ln\left(\frac{2}{\pi\epsilon}\right)+\frac{1}{m_2}+ \sum_{m_1=1}^\infty\left(\frac{2}{|{\bf m}|}C_{{\bf m}}-\frac{1}{m_1} \right)\right] \nonumber\\ &&+\frac{(-1)^{m_2}}{m_2}A_0+2\sum_{m_1=1}^\infty\frac{(-1)^{m_1+m_2}} {|{\bf m}|}C_{{\bf m}}A_{m_1}, \quad (m_2\geq 1). \end{subeqnarray} \subsection{Determination of the slip length} The flow at infinity is determined by substitution of Eq.~(\ref{veeAinf}) into Eq.~(\ref{vee}), which gives \begin{equation}\label{flowinf} {\bf v}\sim\frac{2\pi U|z|}{b}{\bf e}_x\frac{1}{4}\int_{-1}^1 \int_{-1}^1[f_x(w,s)+f_y(w,s)]{\rm d}}\def\c{\color{red} w{\rm d}}\def\c{\color{red} s\quad \hbox{as } |z|\to\infty, \end{equation} {whose simple form is due to the exponential decay of all Fourier modes except ${\bf m}={\bf 0}$.} Then Eqs.~(\ref{Fcoeffts}) and (\ref{ABeqn2n}) show that the shear rate at infinity is given by \begin{equation} \dot\gamma = \frac{\pi U (A_0+B_0)}{b}, \end{equation} and therefore the slip length, $\lambda $, is found to be \begin{equation}\label{slip} \lambda \equiv \frac{U}{\dot\gamma}=\frac{b}{\pi(A_0+B_0)}, \end{equation} which is proportional to the only other length scale in the problem, namely $b$. {Notably, only the zeroth-order coefficients of Eqs.~(\ref{ABeqn3n}) contribute to the slip length}. The system given by Eqs.~(\ref{ABeqn3n}), independent of $b$, is solved in truncated form for various values of $\epsilon$ in \S\ref{numerics}. \subsection{Error estimate} An error estimate can be gleaned by retaining all terms in proceeding from Eqs.~(\ref{veegrid}) to Eqs.~(\ref{inteqs}). Although the former suggests that the symmetry of each grid element implies a mathematically even dependence on $\epsilon$, convergence considerations prevent the error bound from being $O(\epsilon^2)$. For example, the Fourier cosine series for $(\pi-x)^2$ in $[0,2\pi]$ shows that \begin{equation} \sum_{m=1}^\infty\frac{\sin(m\pi\epsilon S)\sin(m\pi\epsilon s)} {\pi^2m^2}=\frac{\epsilon}{2}\min(s,S)[1-\epsilon\max(s,S)] =O(\epsilon). \end{equation} Hence the (relative) error estimate is $O(\epsilon\ln(1/\epsilon))$, as in various sample calculations given by Leppington and Levine \cite{leppington72}. Mathematically, its presence is due to an elliptic integral generated by the fundamental singularity. \section{Numerical calculation and analytical estimate of the mesh slip length} \label{numerics} \begin{figure}[t] \begin{center} \includegraphics[width=0.8\textwidth]{fig4.eps} \end{center} \caption{(color online) (a): Dimensionless effective slip length for the mesh-like super-hydrophobic surface, $\lambda / b$, as a function of the aspect ratio of the mesh, $\epsilon$. (b): Ratio between the slip length given by solving Eqs.~\eqref{ABeqn3n} and the simple estimates provided by Eq.~\eqref{Ldef} ($\lambda/\lambda_1$, blue, solid line) and Eq.~\eqref{2nd_order} ($\lambda/\lambda_2$, black, dash-dotted line), as a function of the aspect ratio.} \label{results} \end{figure} \subsection{Numerical results} \label{fit} We now solve numerically the infinite series given by Eqs.~\eqref{ABeqn3n} by truncating it at a finite value $m_1=m_2=N$. The numerical results are displayed in Fig.~\ref{results}a, where we plot the dimensionless slip length, $\lambda/b$, as a function of the aspect ratio of the mesh, $\epsilon$, for a truncation size of $N=10^2$. The exact size of the truncation has little influence on the final computed results; for example, when $\epsilon=10^{-3}$, the relative change in the computed slip length is less than 0.01 \% between $N=10^2$ and $N=10^3$. Our results show that the surface does show a reduction in friction ({\it i.e.} $\lambda > 0$), and the effective slip length, always on the order of ({but typically smaller than}) the mesh size, increases when the solid {area} fraction decreases. Since the data on Fig.~\ref{results}a are plotted on a semilog scale, we see that we recover the expected logarithmic relationship between the effective slip length of the mesh and its aspect ratio \cite{philip72,philip72_2,lauga03,ybert07}. A least squares fit to the data shown on Fig.~\ref{results}a leads to the empirical formula \begin{equation}\label{fit1} {\frac{\lambda}{b}=\Lambda \ln { \epsilon} + \Pi,} \end{equation} where the values {$\Lambda=-0.107$ and $\Pi= -0.069$, gives result with maximum relative error of 0.4\%. Alternatively, if we define the solid area fraction of the surface $\phi_s$, it is easy to see that $\phi_s=\epsilon(2-\epsilon)$. A least-square fit of the type \begin{equation} \frac{\lambda}{b}=\Lambda_2 \ln {\phi_s} + \Pi_2, \end{equation} with the parameters $\Lambda_2= -0.107$ and $\Pi_2= 0.003$, leads to a maximum relative error of $0.6\%$. As a matter of comparison, the fit proposed in Ref.~\cite{ng09_2}, based on numerical simulations for $\phi_s \gtrsim 0.05$, leads to a relative error of 5\% with the small-$\phi_s$ results obtained here; the asymptotic calculations presented in this paper are therefore able to quantitatively agree with the results of Ref.~\cite{ng09_2} which are valid for much larger mesh sizes.} \subsection{Comparison with simple estimates} An estimate of the mesh slip length can be obtained analytically by performing the truncation by hand. If only $A_0,B_0$ --- that is the mean values of the density functions with respect to $w$ --- are retained in Eqs.~(\ref{ABeqn3n}), the surviving two equations give the lowest-order estimate \begin{equation}\label{Ldef} \frac{\lambda_1}{b}=\frac{1}{\pi (A_0+B_0)}=\frac{1}{3\pi }\ln\left(\frac{2}{\pi\epsilon}\right)\cdot \end{equation} We define here $L=\ln (2/\pi \epsilon)$ for subsequent convenience. {Note that the slope of the relationship given by Eq.~\eqref{Ldef} is $1/3\pi\approx 0.106$, which is very close to the fitted value $|\Lambda|=0.107$ in Eq.~\eqref{fit1}.} Further, on writing Eqs.~(\ref{a}) and (\ref{b}) as \begin{subeqnarray} 3&=&(A_0+B_0)\ln\left(\frac{2}{\pi\epsilon}\right)+\sum_{m=1}^\infty \frac{(-1)^{m}}{m}(A_m+B_m), \\ -1&=&(A_0-B_0)\ln\left(\frac{2}{\pi\epsilon}\right)-\sum_{m=1}^\infty \frac{(-1)^{m}}{m}(A_m-B_m), \end{subeqnarray} and observing that the matrix elements in Eqs.~(\ref{ABeqn3n}) have $L$ only on the diagonal, it may be deduced that \begin{equation}\label{slipcoefft} \frac{\lambda_1}{b}-\frac{L}{3\pi}=O(L^{-1}). \end{equation} As an illustration of this result, the truncation of Eqs.~(\ref{ABeqn3n}) at $m_1=m_2=1$ shows that the leading-order correction to Eq.~(\ref{Ldef}) leads to the improved estimate \begin{equation}\label{2nd_order} \frac{\lambda_2}{b}=\frac{L}{3\pi}-\frac{L+\frac{1}{2}}{6\pi\left[L^2+\frac{1} {4}L-\frac{1}{2}+\frac{9}{4\sqrt{2}}(L+\frac{1}{2})\right]}\cdot \end{equation} On Fig.~\ref{results}b, we show the ratio between the actual slip length (as calculated in Fig.~\ref{results}a), and the simple estimates from Eq.~\eqref{Ldef} (blue, solid line) and Eq.~\eqref{2nd_order} (black, dash-dotted line). We see that these simple formulae over-estimate the slip length (by up to 10\%), but asymptotically converge to the correct result when $\epsilon \to 0$. As expected, the second-order solution, Eq.~\eqref{2nd_order}, is quantitatively better than the first-order one, Eq.~\eqref{Ldef}. \section{Conclusion} In this work we have considered super-hydrophobic surface in the shape of a square mesh, and have presented an analytical method to calculate their effective slip length. {Our analysis, which is valid for asymptotically small aspect ratio of the mesh, agrees also quantitatively with numerical calculations valid for larger mesh width~\cite{ng09_2}. The results we obtain show that, for thin meshes, such surfaces can provide slip lengths on the order of the mesh size. However, for realistic mesh aspect ratio, say 10\% or larger, the slip length is only a fraction of the typical mesh length (see Fig.~\ref{results}).} The increase of the slip length with the solid-to-air ratio of the overall surface is logarithmic, a universal feature of super-hydrophobic coatings with long and thin no-slip domains \cite{philip72,philip72_2,lauga03,ybert07}. Beyond such physical scaling, our results {provide a first analytical prediction} for the resistance to shear flow of the recently devised super-hydrophobic surfaces made of metal wires \cite{feng04,sun05}, and more generally to geometries with grid-like features such as fabric and textiles composed of interwoven threads. Although we have considered here an idealized model system, the analytical approach developed in the paper has the advantage that it provides the effective slip length of the surface with relative error of order $\epsilon \log (1/\epsilon)$, which is significantly better than the more traditional but only logarithmically correct slender-body approach. \section*{Acknowledgments} We thank Shutao Wang for allowing us to reproduce in Fig.~\ref{grid} the picture from Ref.~\cite{wang07}. This work was supported in part by the National Science Foundation (Grants No. CTS-0624830 and CBET-0746285).

*Dear Editor,* Chronic patellar tendinopathy is a common clinical condition that is managed by physical therapists and is common among athletes and non-athletes alike. Patellar tendinopathy is characterized by the absence of inflammatory cells and prostaglandins and an increased presence of fibroblasts and disorganized collagen ([@A15301R1]). Therefore, this condition is not inflammatory as physicians had thought but is a degenerative condition. Extrinsic factors such as inappropriate footwear, sport technique, training errors, and intrinsic factors such as muscle weakness and/or inflexibility, and misalignment are the main factors that lead to patellar tendinopathy ([@A15301R1]-[@A15301R3]). Functional activities such as squat or hop can cause pain in this condition ([@A15301R2], [@A15301R3]). The ideal term for clinical diagnosis is patellar tendinopathy because this term refers to the painful tendon without implying the pathology ([@A15301R1]). Jumper's knee was the first diagnostic term used for this condition. However, this condition can occur in people who are not athletes and athletes who do not perform jumping in their sports. Thus, jumper's knee is not an appropriate term for clinical diagnosis ([@A15301R2]). Patellar tendonitis is also an incorrect term for clinical diagnosis because the condition is not inflammatory but is degenerative as mentioned. The best diagnostic term may be the term patellar tendinosis because this term refers to the pathology of the tendon ([@A15301R1]-[@A15301R3]). The most common site of pain in patellar tendinopathy is the inferior pole of the patella. Patellar tendinopathy pain can also be reported at the tibial tuberosity, at the superior pole of the patella and at the midportion of the tendon ([@A15301R1]-[@A15301R3]). There are four sites of pain in this condition, and an appropriate diagnostic term is needed for each specific pain site. Therefore, the term patellar tendinopathy is a general diagnostic term and cannot be used for all the above-mentioned sites of pain. Moreover, it is unknown if the pathology of the above-mentioned sites are the same. Many physiotherapy techniques have been recommended for the treatment/rehabilitation of patellar tendinopathy such as electrotherapeutic (ultrasound, ESWT, laser, iontophoresis) and non-electrotherapeutic modalities (exercise programs, soft tissue manipulation, and acupuncture). All the above-mentioned treatments intend to improve symptoms (pain and function) of patellar tendinopathy but have totally different mechanisms of action. However, a treatment is effective when it reverses the pathology of the tendinopathy and not only improve the symptoms. Nowadays, eccentric training is the most common physiotherapy treatment for patients with patellar tendinopathy ([@A15301R1]-[@A15301R3]). The question that arises is whether eccentric training is effective for all sites of patellar tendinopathy? For example, eccentric training with dorsiflexion is effective for patients with mid-portion Achilles tendinopathy ([@A15301R4]-[@A15301R9]), but eccentric training without dorsiflexion has positive effects on patients with insertional Achilles tendinopathy ([@A15301R10]). Therefore, different sites of Achilles tendinopathy are managed with different protocols of eccentric exercises. Eccentric training consisting of squats is an effective treatment approach when the patellar tendinopathy is at the inferior pole of the patella; however, no studies have investigated the effectiveness of eccentric training on other sites of patellar tendinopathy. Thus, studies determining the effectiveness of eccentric exercises at other sites of patellar tendinopathy are needed. Future well-designed studies are needed to find firstly, the most appropriate diagnostic term for each site of patellar tendinopathy and secondly, the appropriate protocol of eccentric training for each site of pain and the pathology of this disorder. **Financial Disclosure:**The author declared no conflict of interest. **Funding/Support:**This work received no support.

1. Introduction {#sec0001} =============== The mathematical modelling of engineered biological systems has entered a new era in recent years with the expansion and standardisation of existing models aimed at collating disparate components of these processes and provide scientists, engineers and practitioners with the tools to better predict, control and optimise them [@bib0023]. In engineered biological systems, mechanistic modelling reached consensus with the development of the Activated Sludge Models [@bib0009], [@bib0010] for wastewater treatment processes, followed by the Anaerobic Digestion Model No. 1 (ADM1) [@bib0012] a few years later. The development of ADM1 was enabled largely due to the possibilities for better identification and characterisation of functional microbial groups responsible for the chemical transformations within anaerobic digesters. It describes a set of stoichiometric and kinetic functions representing the standard anaerobic processes, remaining the scientific benchmark to the present day. However, there has been a growing awareness that the model should take advantage of improved empirical understanding and extension of biochemical processes included in its structure, to acquire a better trade-off between model realism and complexity [@bib0013]. The full ADM1 model is highly parameterised with a large number of physical, chemical and biological processes described by numerous state variables and algebraic expressions. Whilst suitable for dynamic simulation, more rigorous mathematical analysis of the model is difficult. To the authors knowledge, only numerical investigations are available [@bib0003]. Due to the analytical intractability of the full ADM1, work has been made towards the construction of simpler models that preserve biological meaning whilst reducing the computational effort required to find mathematical solutions of the model equations [@bib0007], [@bib0008]. The most common models used to describe microbial systems are two-tiered models, which take the form of a cascade of two biological reactions where one substrate is consumed by one microorganism to produce a product that serves as the main limiting substrate for a second microorganism. When the second organism has no feedback on the first organism, the system is known as commensalistic [@bib0016], [@bib0020]. The system has a cascade structure and the number of steady-states and their (mathematical) stability as a function of model inputs and parameters may be investigated [@bib0001], [@bib0002], [@bib0019]. When the growth of the first organism is affected by the substrate produced by the second organism the system is known as syntrophic. For instance, if the first organism is inhibited by high concentrations of the product, the extent to which the first substrate is degraded by the first organism depends on the efficiency of the removal of the product by the second organism. The mathematical analysis of such a model is more delicate than for commensalistic models, (see for instance early work by [@bib0004], [@bib0014], [@bib0015], [@bib0027] and the more recent papers [@bib0005], [@bib0011], [@bib0017], [@bib0018], [@bib0022]). An important and interesting extension should be mentioned here: [@bib0025], [@bib0026] analysed an 8-dimensional mathematical model, which includes syntrophy and inhibition, both mechanisms considered by [@bib0002] and [@bib0006]. As an example of this for anaerobic digestion, a previous study investigated the effect of maintenance on the stability of a two-tiered 'food-chain' comprising two species and two substrates [@bib0028]. Maintenance is defined as the energy consumed by an organism that is used for all biological processes other than growth. In [@bib0028] and here, it is analogous to a first-order decay rate constant, or biomass death term. Although the authors were not able to determine the general conditions under which this four dimensional syntrophic consortium was stable, further work has shown that a model with generality can be used to answer the question posed, determining that the two-tiered food-chain is always stable when maintenance is included [@bib0018]. More recently, the model described by [@bib0028] was extended by the addition of a third organism and substrate to create a three-tiered 'food-web' [@bib0024], as shown in [Fig. 1](#fig0001){ref-type="fig"}. In this paper, the existence and stability of the steady-states were determined only numerically. Although the results were important in revealing emergent properties of this extended model, the motivation of this work is to give an analytical study of the model. Moreover, our analysis does not require that growth functions are of the specific form considered and are valid for a large class of growth functions. This is critical as it provides the means by which microbiologists can theoretically test the influence of the growth characteristics of organisms on the properties of the system and the interactions between multiple species.Fig. 1Schematic of the three-tier chlorophenol mineralising food-web indicating the flow and conversion of chemical substrates and products in the system.Fig. 1 Here, we pursue a generalised description and analysis of the model given by [@bib0024]. Chlorophenols are chemicals of importance due to their impact on the environment and to public health, their recalcitrance in food-webs and resistance to aerobic biodegradation via the oxygenase enzyme [@bib0021]. Although we consider the monochlorophenol isomer here, extension to multiple isomeric chlorophenols would be straightforward. It is important to note that, although the particular biological transformation provided is for chlorophenol mineralisation, the structure and methods employed are much more general and apply to any theoretical ecological interactions that may be hypothesised or observed in a microbial community. We therefore stress that this work can provide a good approach for analytically investigating the behaviour of microbial food-webs where numerical parameters are difficult to obtain or uncertain. Ultimately, we demonstrate here the advantage of a generalised approach for mathematical analysis of microbial ecological interactions from both a theoretical perspective and its potential if providing knowledge in applied research where these communities and processes are studied empirically. The paper is organised as follows. In [Section 2](#sec0002){ref-type="sec"}, we present a description of the model to be investigated, and its reduction in [Section 3](#sec0003){ref-type="sec"}. Model assumptions and notations are provided in [Section 4](#sec0004){ref-type="sec"}. In [Section 5](#sec0005){ref-type="sec"} we demonstrate the existence of the three steady-states and define four interesting cases for specific parameter values that are investigated using the solutions, whilst also indicating the regions of existence of the steady-states for the operating parameter values. We present results on the behaviour of the system whilst varying two control parameters in [Section 6](#sec0006){ref-type="sec"}. In [Section 7](#sec0011){ref-type="sec"} we perform local stability analysis of the steady-states without maintenance and, in [Section 8](#sec0016){ref-type="sec"}, we undertake a comprehensive numerical stability analysis of the four cases for both the model with and without maintenance. We show that our approach leads to the discovery of five operating regions in which one leads to the possibility of instability of the positive steady-state, where all three organisms exist, a fact that has not been reported by [@bib0024]. Indeed, we suggest that a stable limit-cycle can occur at the boundary with this region. Finally, in [Section 9](#sec0021){ref-type="sec"}, we make comment on the role of the kinetic parameters used in the four example cases, in maintaining stability, which points to the importance of the relative aptitude of the two hydrogen consumers in sustaining a viable chlorophenol mineralising community. In the Appendix we describe the numerical method used in [Section 8](#sec0016){ref-type="sec"}, give the assumptions on the growth functions we used and the proofs of the results. 2. The model {#sec0002} ============ The model developed in [@bib0024] has six components, three substrate (chlorophenol, phenol and hydrogen) and three biomass (chlorophenol and phenol degraders, and a hydrogenotrophic methanogen) variables. The substrate and biomass concentrations evolve according to the six-dimensional dynamical of ODEs: $$\begin{aligned} \frac{dX_{\text{ch}}}{dt} & {= - DX_{\text{ch}} + Y_{\text{ch}}f_{0}\left( S_{\text{ch}},S_{H_{2}} \right)X_{\text{ch}} - k_{\text{dec},\text{ch}}X_{\text{ch}}} \\ \end{aligned}$$$$\begin{aligned} \frac{dX_{\text{ph}}}{dt} & {= - DX_{\text{ph}} + Y_{\text{ph}}f_{1}\left( S_{\text{ph}},S_{H_{2}} \right)X_{\text{ph}} - k_{\text{dec},\text{ph}}X_{\text{ph}}} \\ \end{aligned}$$$$\begin{aligned} \frac{dX_{H_{2}}}{dt} & {= - DX_{H_{2}} + Y_{H_{2}}f_{2}\left( S_{H_{2}} \right)X_{H_{2}} - k_{\text{dec},H_{2}}X_{H_{2}}} \\ \end{aligned}$$$$\begin{aligned} \frac{dS_{\text{ch}}}{dt} & {= D\left( S_{\text{ch},\text{in}} - S_{\text{ch}} \right) - f_{0}\left( S_{\text{ch}},S_{H_{2}} \right)X_{\text{ch}}} \\ \end{aligned}$$$$\begin{array}{ccl} \frac{dS_{\text{ph}}}{dt} & = & {D\left( S_{\text{ph},\text{in}} - S_{\text{ph}} \right) + \frac{224}{208}\left( 1 - Y_{\text{ch}} \right)f_{0}\left( S_{\text{ch}},S_{H_{2}} \right)X_{\text{ch}}} \\ & & {- f_{1}\left( S_{\text{ph}},S_{H_{2}} \right)X_{\text{ph}}} \\ \end{array}$$$$\begin{array}{ccl} \frac{dS_{H_{2}}}{dt} & = & {D\left( S_{H_{2},\text{in}} - S_{H_{2}} \right) + \frac{32}{224}\left( 1 - Y_{\text{ph}} \right)f_{1}\left( S_{\text{ph}},S_{H_{2}} \right)X_{\text{ph}}} \\ & & {- \frac{16}{208}f_{0}\left( S_{\text{ch}},S_{H_{2}} \right)X_{\text{ch}} - f_{2}\left( S_{H_{2}} \right)X_{H_{2}}} \\ \end{array}$$where *S*~ch~ and *X*~ch~ are the chlorophenol substrate and degrader concentrations, *S*~ph~ and *X*~ph~ for phenol and $S_{H_{2}}$ and $X_{H_{2}}$ for hydrogen; *Y*~ch~, *Y*~ph~ and *Y*~H2~ are the yield coefficients, $224/208\left( 1 - Y_{\text{ch}} \right)$ represents the part of chlorophenol degraded to phenol, and $32/224\left( 1 - Y_{\text{ph}} \right)$ represents the part of phenol that is transformed to hydrogen. Growth functions take Monod form with hydrogen inhibition acting on the phenol degrader and represented in *f*~1~ (see ([Eq. 7](#eq0007){ref-type="disp-formula"})) as a product inhibition term. In [@bib0024], this model is given in dimensionless form that significantly reduces the number of parameters describing the dynamics. In the analysis of the generalised model ([Sections 5](#sec0005){ref-type="sec"} and [7](#sec0011){ref-type="sec"}), we do not assume that the growth functions *f*~0~, *f*~1~ and *f*~2~ have the specific analytical expression shown. We will only assume that the growth functions satisfy properties that are listed in [Section 4](#sec0004){ref-type="sec"}. Therefore, we cannot benefit from the dimensionless rescaling used by [@bib0024], because it uses some kinetics parameters of the specific growth functions ([Eq. 7](#eq0007){ref-type="disp-formula"}), while we work with general *unspecified* growth functions. In [Section 3](#sec0003){ref-type="sec"} we consider another rescaling that does not use the kinetics parameters. For the numerical analyses given in [Section 8](#sec0016){ref-type="sec"}, we return to an assumed set of Monod growth functions given by ([Eq. 7](#eq0007){ref-type="disp-formula"}), in order to directly compare with the results found by [@bib0024]. $$\begin{array}{rcl} {f_{0}\left( S_{\text{ch}},S_{H_{2}} \right)} & = & {\frac{k_{m,\text{ch}}S_{\text{ch}}}{K_{S,ch} + S_{\text{ch}}}\frac{S_{H_{2}}}{K_{S,H_{2},c} + S_{H_{2}}}} \\ {f_{1}\left( S_{\text{ph}},S_{H_{2}} \right)} & = & {\frac{k_{m,\text{ph}}S_{\text{ph}}}{K_{S,\text{ph}} + S_{\text{ph}}}\frac{1}{1 + \frac{S_{H_{2}}}{K_{i,H_{2}}}}} \\ {f_{2}\left( S_{H_{2}} \right)} & = & \frac{k_{m,H_{2}}S_{H_{2}}}{K_{S,H_{2}} + S_{H_{2}}} \\ \end{array}$$ Here, apart from the four operating (or control) parameters, which are the inflowing concentrations *S*~ch,\ in~, *S*~ph,\ in~, $S_{H_{2},\text{in}}$ and the dilution rate *D*, that can vary, all others have biological meaning and are fixed depending on the organisms and substrate considered. We use the following notations in ([(1)](#eq0001){ref-type="disp-formula"}, [(2)](#eq0002){ref-type="disp-formula"}, [(3)](#eq0003){ref-type="disp-formula"}, [(4)](#eq0004){ref-type="disp-formula"}, [(5)](#eq0005){ref-type="disp-formula"}, [(6)](#eq0006){ref-type="disp-formula"}): $$\begin{array}{cl} & {X_{0} = X_{\text{ch}},\quad X_{1} = X_{\text{ph}},\quad X_{2} = X_{H_{2}}} \\ & {S_{0} = S_{\text{ch}},\quad S_{1} = S_{\text{ph}},\quad S_{2} = S_{H_{2}}} \\ & {S_{0}^{\text{in}} = S_{\text{ch},\text{in}},\quad S_{1}^{\text{in}} = S_{\text{ph},\text{in}},\quad S_{2}^{\text{in}} = S_{H_{2},\text{in}}} \\ & {Y_{0} = Y_{\text{ch}},\quad Y_{1} = Y_{\text{ph}},\quad Y_{2} = Y_{H_{2}}} \\ & {Y_{3} = \frac{224}{208}\left( 1 - Y_{\text{ch}} \right),\quad Y_{4} = \frac{32}{224}\left( 1 - Y_{\text{ph}} \right),\quad Y_{5} = \frac{16}{208}} \\ & {a_{0} = k_{\text{dec},\text{ch}},\quad a_{1} = k_{\text{dec},\text{ph}},\quad a_{2} = k_{\text{dec},H_{2}}} \\ \end{array}$$With these notations ([(1)](#eq0001){ref-type="disp-formula"}, [(2)](#eq0002){ref-type="disp-formula"}, [(3)](#eq0003){ref-type="disp-formula"}, [(4)](#eq0004){ref-type="disp-formula"}, [(5)](#eq0005){ref-type="disp-formula"}, [(6)](#eq0006){ref-type="disp-formula"}) can be written as follows: $$\begin{aligned} \frac{dX_{0}}{dt} & {= - DX_{0} + Y_{0}f_{0}\left( S_{0},S_{2} \right)X_{0} - a_{0}X_{0}} \\ \end{aligned}$$$$\begin{aligned} \frac{dX_{1}}{dt} & {= - DX_{1} + Y_{1}f_{1}\left( S_{1},S_{2} \right)X_{1} - a_{1}X_{1}} \\ \end{aligned}$$$$\begin{aligned} \frac{dX_{2}}{dt} & {= - DX_{2} + Y_{2}f_{2}\left( S_{2} \right)X_{2} - a_{2}X_{2}} \\ \end{aligned}$$$$\begin{aligned} \frac{dS_{0}}{dt} & {= D\left( S_{0}^{\text{in}} - S_{0} \right) - f_{0}\left( S_{0},S_{2} \right)X_{0}} \\ \end{aligned}$$$$\begin{aligned} \frac{dS_{1}}{dt} & {= D\left( S_{1}^{\text{in}} - S_{1} \right) + Y_{3}f_{0}\left( S_{0},S_{2} \right)X_{0} - f_{1}\left( S_{1},S_{2} \right)X_{1}} \\ \end{aligned}$$$$\begin{aligned} \frac{dS_{2}}{dt} & {= D\left( S_{2}^{\text{in}} - S_{2} \right) + Y_{4}f_{1}\left( S_{1},S_{2} \right)X_{1} - Y_{5}f_{0}\left( S_{0},S_{2} \right)X_{0} - f_{2}\left( S_{2} \right)X_{2}} \\ \end{aligned}$$Furthermore, we restrict our analysis to the case where we only have one substrate added to the system, such that: $S_{0}^{in} > 0,$$S_{1}^{in} = 0,$ and $S_{2}^{in} = 0$. As shown in [@bib0024], the general case with $S_{1}^{in} > 0,$ and $S_{2}^{in} > 0$ present many important and interesting behaviours and deserves future work. 3. Model reduction {#sec0003} ================== To ease the mathematical analysis, we can rescale the system ([(8)](#eq0008){ref-type="disp-formula"}, [(9)](#eq0009){ref-type="disp-formula"}, [(10)](#eq0010){ref-type="disp-formula"}, [(11)](#eq0011){ref-type="disp-formula"}, [(12)](#eq0012){ref-type="disp-formula"}, [(13)](#eq0013){ref-type="disp-formula"}) using the following change of variables adapted from [@bib0018]: $$\begin{array}{cl} & {x_{0} = \frac{Y_{3}Y_{4}}{Y_{0}}X_{0},\quad x_{1} = \frac{Y_{4}}{Y_{1}}X_{1},\quad x_{2} = \frac{1}{Y_{2}}X_{2}} \\ & {s_{0} = Y_{3}Y_{4}S_{0},\quad s_{1} = Y_{4}S_{1},\quad s_{2} = S_{2}} \\ \end{array}$$We obtain the following system: $$\begin{aligned} \frac{dx_{0}}{dt} & {= - Dx_{0} + \mu_{0}\left( s_{0},s_{2} \right)x_{0} - a_{0}x_{0}} \\ \end{aligned}$$$$\begin{aligned} \frac{dx_{1}}{dt} & {= - Dx_{1} + \mu_{1}\left( s_{1},s_{2} \right)x_{1} - a_{1}x_{1}} \\ \end{aligned}$$$$\begin{aligned} \frac{dx_{2}}{dt} & {= - Dx_{2} + \mu_{2}\left( s_{2} \right)x_{2} - a_{2}x_{2}} \\ \end{aligned}$$$$\begin{aligned} \frac{ds_{0}}{dt} & {= D\left( s_{0}^{\text{in}} - s_{0} \right) - \mu_{0}\left( s_{0},s_{2} \right)x_{0}} \\ \end{aligned}$$$$\begin{aligned} \frac{ds_{1}}{dt} & {= - Ds_{1} + \mu_{0}\left( s_{0},s_{2} \right)x_{0} - \mu_{1}\left( s_{1},s_{2} \right)x_{1}} \\ \end{aligned}$$$$\begin{aligned} \frac{ds_{2}}{dt} & {= - Ds_{2} + \mu_{1}\left( s_{1},s_{2} \right)x_{1} - \omega\mu_{0}\left( s_{0},s_{2} \right)x_{0} - \mu_{2}\left( s_{2} \right)x_{2}} \\ \end{aligned}$$where the inflowing concentration is: $$s_{0}^{\text{in}} = Y_{3}Y_{4}S_{0}^{\text{in}},$$the growth functions are: $$\begin{array}{ccl} & & {\mu_{0}\left( s_{0},s_{2} \right) = Y_{0}f_{0}\left( \frac{s_{0}}{Y_{3}Y_{4}},s_{2} \right)} \\ & & {\mu_{1}\left( s_{1},s_{2} \right) = Y_{1}f_{1}\left( \frac{s_{1}}{Y_{4}},s_{2} \right)} \\ & & {\mu_{2}\left( s_{2} \right) = Y_{2}f_{2}\left( s_{2} \right)} \\ \end{array}$$and $$\omega = \frac{Y_{5}}{Y_{3}Y_{4}} = \frac{1}{2\left( 1 - Y_{0} \right)\left( 1 - Y_{1} \right)}$$The benefit of our rescaling is that it permits to fix in ([(14)](#eq0014){ref-type="disp-formula"}, [(15)](#eq0015){ref-type="disp-formula"}, [(16)](#eq0016){ref-type="disp-formula"}, [(17)](#eq0017){ref-type="disp-formula"}, [(18)](#eq0018){ref-type="disp-formula"}, [(19)](#eq0019){ref-type="disp-formula"}) all yield coefficients to one except that denoted by *ω* and defined by ([Eq. 22](#eq0022){ref-type="disp-formula"}), and to discuss the existence and stability with respect to this sole parameter. Using ([Eq. 21](#eq0021){ref-type="disp-formula"}) and the growth functions ([Eq. 7](#eq0007){ref-type="disp-formula"}), we obtain the model ([(14)](#eq0014){ref-type="disp-formula"}, [(15)](#eq0015){ref-type="disp-formula"}, [(16)](#eq0016){ref-type="disp-formula"}, [(17)](#eq0017){ref-type="disp-formula"}, [(18)](#eq0018){ref-type="disp-formula"}, [(19)](#eq0019){ref-type="disp-formula"}) with the following Monod-type growth functions: $$\begin{array}{ccl} & & {\mu_{0}\left( s_{0},s_{2} \right) = \frac{m_{0}s_{0}}{K_{0} + s_{0}}\frac{s_{2}}{L_{0} + s_{2}}} \\ & & {\mu_{1}\left( s_{1},s_{2} \right) = \frac{m_{1}s_{1}}{K_{1} + s_{1}}\frac{1}{1 + s_{2}/K_{i}}} \\ & & {\mu_{2}\left( s_{2} \right) = \frac{m_{2}s_{2}}{K_{2} + s_{2}}} \\ \end{array}$$where: $$\begin{array}{cl} & {m_{0} = Y_{0}k_{m,\text{ch}},\quad K_{0} = Y_{3}Y_{4}K_{s,\text{ch}},\quad L_{0} = K_{S,H_{2},c}} \\ & {m_{1} = Y_{1}k_{m,\text{ph}},\quad K_{1} = Y_{4}K_{s,\text{ph}},\quad K_{i} = K_{i,H_{2}}} \\ & {m_{2} = Y_{2}k_{m,H_{2}},\quad K_{2} = K_{S,H_{2}}} \\ \end{array}$$For the numerical simulations we will use the nominal values in [Table 1](#tbl0001){ref-type="table"} given in [@bib0024].Table 1Nominal parameter values. We use units expressed in Chemical Oxygen Demand (COD) as used by [@bib0012], [@bib0024].Table 1ParametersNominal valuesUnits*k*~*m*,\ ch~29kgCOD~S~/kgCOD~X~/d*K*~*S*,\ ch~0.053kgCOD/m^3^*Y*~ch~0.019kgCOD~X~/kgCOD~S~*k*~*m*,\ ph~26kgCOD~S~/kgCOD~X~/d*K*~*S*,\ ph~0.302kgCOD/m^3^*Y*~ph~0.04kgCOD~X~/kgCOD~S~$k_{m,H_{2}}$35kgCOD~S~/kgCOD~X~/d$K_{S,H_{2}}$2.5$\times 10^{- 5}$kgCOD/m^3^$K_{S,{H_{2},c}}$1.0$\times 10^{- 6}$kgCOD/m^3^$Y_{H_{2}}$0.06kgCOD~X~/kgCOD~S~*k*~dec,\ i~0.02$d^{- 1}$$K_{I,H_{2}}$3.5$\times 10^{- 6}$kgCOD/m^3^ 4. Assumptions on the model and notations {#sec0004} ========================================= Our study does not require that growth functions are of Monod type ([Eq. 23](#eq0023){ref-type="disp-formula"}). Actually, the results are valid for a more general class of growth functions satisfying the following conditions, which concur with those given by ([Eq. 23](#eq0023){ref-type="disp-formula"}): H1For all *s*~0~ \> 0 and *s*~2~ \> 0 then $0 < \mu_{0}\left( s_{0},s_{2} \right) < + \infty$ and $\mu_{0}\left( 0,s_{2} \right) = 0,$$\mu_{0}\left( s_{0},0 \right) = 0$.H2For all *s*~1~ \> 0 and *s*~2~ ≥ 0 then $0 < \mu_{1}\left( s_{1},s_{2} \right) < + \infty$ and $\mu_{1}\left( 0,s_{2} \right) = 0$.H3For all *s*~2~ \> 0 then $0 < \mu_{2}\left( s_{2} \right) < + \infty$ and $\mu_{2}\left( 0 \right) = 0$.H4For all *s*~0~ \> 0 and *s*~2~ \> 0, $$\frac{\partial\mu_{0}}{\partial s_{0}}\left( s_{0},s_{2} \right) > 0,\quad\frac{\partial\mu_{0}}{\partial s_{2}}\left( s_{0},s_{2} \right) > 0.$$H5For all *s*~1~ \> 0 and *s*~2~ \> 0, $$\frac{\partial\mu_{1}}{\partial s_{1}}\left( s_{1},s_{2} \right) > 0,\quad\frac{\partial\mu_{1}}{\partial s_{2}}\left( s_{1},s_{2} \right) < 0.$$H6For all *s*~2~ \> 0, $\frac{d\mu_{2}}{ds_{2}}\left( s_{2} \right) > 0$.H7The function $\left. s_{2}\mapsto\mu_{0}\left( + \infty,s_{2} \right) \right.$ is monotonically increasing and the function $\left. s_{2}\mapsto\mu_{1}\left( + \infty,s_{2} \right) \right.$ is monotonically decreasing. The proof of the following result is standard and hence omitted. Proposition 1*For non-negative initial conditions, all solutions of the system (Eqs.* [14](#eq0014){ref-type="disp-formula"}*--*[19](#eq0019){ref-type="disp-formula"}*) are bounded and remain non-negative for all t* \> 0*.* In the following lemma we define the functions *M*~0~(*y, s*~2~), *M*~1~(*y, s*~2~) and *M*~2~(*y*), which are the inverse functions of the functions *s*~0~↦*μ*~0~(*s*~0~, *s*~2~), *s*~1~↦*μ*~1~(*s*~1~, *s*~2~) and *s*~2~↦*μ*~2~(*s*~2~), respectively. Lemma 1*Let s*~2~ ≥ 0 *be fixed. There exists a unique function* $$\left. y \in \left\lbrack 0,\mu_{0}\left( + \infty,s_{2} \right) \right)\mapsto M_{0}\left( y,s_{2} \right) \in \left\lbrack 0, + \infty \right), \right.$$*such that for s*~0~ ≥ 0, *s*~2~ ≥ 0 *and*$y \in \left\lbrack 0,\mu_{0}\left( + \infty,s_{2} \right) \right),$ *we have:* $$\left. s_{0} = M_{0}\left( y,s_{2} \right)\Longleftrightarrow y = \mu_{0}\left( s_{0},s_{2} \right) \right.$$*Let s*~2~ ≥ 0 *be fixed. There exists a unique function* $$\left. y \in \left\lbrack 0,\mu_{1}\left( + \infty,s_{2} \right) \right)\mapsto M_{1}\left( y,s_{2} \right) \in \left\lbrack 0, + \infty \right), \right.$$*such that for s*~1~ ≥ 0, *s*~2~ ≥ 0 *and*$y \in \left\lbrack 0,\mu_{1}\left( + \infty,s_{2} \right) \right),$ *we have:* $$\left. s_{1} = M_{1}\left( y,s_{2} \right)\Longleftrightarrow y = \mu_{1}\left( s_{1},s_{2} \right) \right.$$*There exists a unique function* $$\left. y \in \left\lbrack 0,\mu_{2}\left( + \infty \right) \right)\mapsto M_{2}\left( y \right) \in \left\lbrack 0, + \infty \right), \right.$$*such that, for s*~2~ ≥ 0 *and*$y \in \left\lbrack 0,\mu_{2}\left( + \infty \right) \right)$ *we have:* $$\left. s_{2} = M_{2}\left( y \right)\Longleftrightarrow y = \mu_{2}\left( s_{2} \right) \right.$$ Using the functions *M*~0~, *M*~1~ and *M*~2~ we define now the function *ψ*(*s*~2~, *D*), *F*~1~(*D*), *F*~2~(*D*) and *F*~3~(*D*). Let us observe first that by **H7**, for $D + a_{0} < \mu_{0}\left( + \infty, + \infty \right)$ and $D + a_{1} < \mu_{1}\left( + \infty,0 \right)$ there exist unique values $s_{2}^{0} = s_{2}^{0}\left( D \right)$ and $s_{2}^{1} = s_{2}^{1}\left( D \right),$ see [Fig. 2](#fig0002){ref-type="fig"}(a): $$\mu_{0}\left( + \infty,s_{2}^{0} \right) = D + a_{0},\quad\mu_{1}\left( + \infty,s_{2}^{1} \right) = D + a_{1}$$Assume that *ω* \< 1. Let *D* be fixed such that $s_{2}^{0}\left( D \right)$ and $s_{2}^{1}\left( D \right)$ exist. We consider the function defined on $\left( s_{2}^{0}\left( D \right),s_{2}^{1}\left( D \right) \right)$ by: $$\begin{array}{ccl} & & \left. s_{2} \in \left( s_{2}^{0}\left( D \right),s_{2}^{1}\left( D \right) \right)\mapsto\psi\left( s_{2},D \right) \right. \\ & & {\quad = M_{0}\left( D + a_{0},s_{2} \right) + \frac{M_{1}\left( D + a_{1},s_{2} \right) + s_{2}}{1 - \omega},} \\ \end{array}$$It should be noted that *ψ*(*s*~2~, *D*) \> 0 for $s_{2}^{0}\left( D \right) < s_{2} < s_{2}^{1}\left( D \right)$. The function *ψ* together with the values $s_{2}^{0},$$s_{2}^{1}$ all depend on *D*. However, to avoid cumbersome notation we will use the more precise form *ψ*(*s*~2~, *D*), $s_{2}^{0}\left( D \right)$ and $s_{2}^{1}\left( D \right)$ only if necessary. From ([Eq. 25](#eq0025){ref-type="disp-formula"}), ([Eq. 26](#eq0026){ref-type="disp-formula"}) and ([Eq. 28](#eq0028){ref-type="disp-formula"}) we deduce that: $$M_{0}\left( D + a_{0},s_{2}^{0} \right) = + \infty,\quad M_{1}\left( D + a_{1},s_{2}^{1} \right) = + \infty$$Therefore, we have, see [Fig. 2](#fig0002){ref-type="fig"}(b): $$\lim\limits_{s_{2}\rightarrow s_{2}^{0}}\psi\left( s_{2} \right) = \lim\limits_{s_{2}\rightarrow s_{2}^{1}}\psi\left( s_{2} \right) = + \infty$$Hence, the function *ψ*(*s*~2~), which is positive and tends to $+ \infty$ at the extremities of the interval $\left( s_{2}^{0},s_{2}^{1} \right),$ has a minimum value on this interval. We add the following assumption: H8The function *ψ* has a unique minimum ${\overline{s}}_{2}\left( D \right)$ on the interval $\left( s_{2}^{0}\left( D \right),s_{2}^{1}\left( D \right) \right)$ and $\frac{\partial\psi}{\partial s_{2}}\left( s_{2},D \right)$ is negative on $\left( s_{2}^{0}\left( D \right),{\overline{s}}_{2}\left( D \right) \right)$ and positive on $\left( {\overline{s}}_{2}\left( D \right),s_{2}^{1}\left( D \right) \right),$ respectively.Fig. 2Graphical definitions. (a): $s_{2}^{0}$ and $s_{2}^{1}$. (b) : *ψ*(*s*~2~), ${\overline{s}}_{2},$*F*~1~(*D*) and *F*~2~(*D*). (c): $s_{2}^{\flat}$ and $s_{2}^{\sharp}$.Fig. 2Fig. 3Graphs of $s_{2}^{0}\left( D \right)$ and $s_{2}^{1}\left( D \right)$ (in black) and $M_{2}\left( D + a_{2} \right)$ (in red) and graphical depiction of $I_{1} = \left\lbrack 0,D_{1} \right),$ where *D*~1~ is the solution of $s_{2}^{0}\left( D \right) = s_{2}^{1}\left( D \right),$ and *I*~2~. (a): $I_{2} = \left\lbrack 0,D_{2} \right)$ where *D*~2~ is the solution of $M_{2}\left( D + a_{2} \right) = s_{2}^{1}\left( D \right)$. (b) : $I_{2} = \left\lbrack 0,D_{2} \right)$ where *D*~2~ is the solution of $M_{2}\left( D + a_{2} \right) = s_{2}^{0}\left( D \right)$. (c): *I*~2~ is empty. (d) : $I_{2} = \left( D_{2min},D_{2max} \right)$ where *D*~2*min*~ and *D*~2*max*~ are the solutions of $M_{2}\left( D + a_{2} \right) = s_{2}^{0}\left( D \right)$ and $M_{2}\left( D + a_{2} \right) = s_{2}^{1}\left( D \right),$ respectively. Cases (a)--(d) are obtained with the numerical parameter values listed in [Tables 2](#tbl0002){ref-type="table"} and [3](#tbl0003){ref-type="table"}. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)Fig. 3 The value ${\overline{s}}_{2}$ depends on *D*. However, to avoid cumbersome notation we will use the more precise form ${\overline{s}}_{2}\left( D \right)$ only if necessary. We consider the function *F*~1~(*D*), which is the infimum of *ψ*(*s*~2~, *D*), with respect to the variable *s*~2~, see [Fig. 2](#fig0002){ref-type="fig"}(b): $$\begin{aligned} {F_{1}\left( D \right)} & {= \inf\limits_{s_{2} \in {(s_{2}^{0}{(D)},s_{2}^{1}{(D)})}}\psi\left( s_{2},D \right) = \psi\left( {\overline{s}}_{2},D \right)} \\ \end{aligned}$$We also define the functions *F*~2~(*D*) and *F*~3~(*D*): $$\begin{aligned} {F_{2}\left( D \right)} & {= \psi\left( M_{2}\left( D + a_{2} \right),D \right)} \\ \end{aligned}$$$$\begin{aligned} {F_{3}\left( D \right)} & {= \frac{\partial\psi}{\partial s_{2}}\left( M_{2}\left( D + a_{2} \right),D \right)} \\ \end{aligned}$$The function *F*~1~(*D*) is defined for *D* ∈ *I*~1~ where: $$I_{1} = \left\{ D \geq 0:s_{2}^{0}\left( D \right) < s_{2}^{1}\left( D \right) \right\}$$The functions *F*~2~(*D*) and *F*~3~(*D*) are defined for *D* ∈ *I*~2~ where: $$I_{2} = \left\{ D \in I_{1}:s_{2}^{0}\left( D \right) < M_{2}\left( D + a_{2} \right) < s_{2}^{1}\left( D \right) \right\}$$For all for *D* ∈ *I*~2~, *F*~1~(*D*) ≤ *F*~2~(*D*). The equality $F_{1}\left( D \right) = F_{2}\left( D \right)$ holds if, and only if, $M_{2}\left( D + a_{2} \right) = {\overline{s}}_{2}\left( D \right)$ that is, $\frac{\partial\psi}{\partial s_{2}}\left( M_{2}\left( D + a_{2} \right) \right) = 0,$ that is if, and only if, $F_{3}\left( D \right) = 0$. We define $$I_{3} = \left\{ D \in I_{2}:F_{3}\left( D \right) < 0 \right\}$$For the Monod-type growth functions ([Eq. 23](#eq0023){ref-type="disp-formula"}), straightforward computations show that the functions *M*~0~, *M*~1~ and *M*~2~ are given respectively by: For $y \in \left\lbrack 0,\mu_{0}\left( + \infty,s_{2} \right) = \frac{m_{0}s_{2}}{L_{0} + s_{2}} \right),$$$M_{0}\left( y,s_{2} \right) = \frac{K_{0}y}{\frac{m_{0}s_{2}}{L_{0} + s_{2}} - y}$$For $y \in \left\lbrack 0,\mu_{1}\left( + \infty,s_{2} \right) = \frac{m_{1}}{1 + s_{2}/K_{i}} \right),$$$M_{1}\left( y,s_{2} \right) = \frac{K_{1}y}{\frac{m_{1}}{1 + s_{2}/K_{i}} - y}$$For $y \in \left\lbrack 0,\mu_{2}\left( + \infty \right) = m_{2} \right),$$$M_{2}\left( y \right) = \frac{K_{2}y}{m_{2} - y}$$Moreover, we have: $$s_{2}^{0}\left( D \right) = \frac{L_{0}\left( D + a_{0} \right)}{m_{0} - D - a_{0}},\quad s_{2}^{1}\left( D \right) = \frac{K_{i}\left( m_{1} - D - a_{1} \right)}{D + a_{1}}$$$$\psi\left( s_{2},D \right) = \frac{K_{0}\left( D + a_{0} \right)}{m_{0} - D - a_{0}}\frac{L_{0} + s_{2}}{s_{2} - s_{2}^{0}\left( D \right)} + \frac{\frac{K_{1}\left( K_{i} + s_{2} \right)}{s_{2}^{1}\left( D \right) - s_{2}} + s_{2}}{1 - \omega}$$and $$\frac{\partial^{2}\psi}{\partial s_{2}^{2}}\left( s_{2},D \right) = \frac{2K_{0}\left( D + a_{0} \right)}{m_{0} - D - a_{0}}\frac{L_{0} + s_{2}^{0}\left( D \right)}{\left( s_{2} - s_{2}^{0}\left( D \right) \right)^{3}} + \frac{2K_{1}\left( K_{i} + s_{2}^{1}\left( D \right) \right)}{\left( 1 - \omega \right)\left( s_{2}^{1}\left( D \right) - s_{2} \right)^{3}}$$Hence, $\frac{\partial^{2}\psi}{\partial s_{2}^{2}} > 0$ for all $s_{2} \in \left( s_{2}^{0}\left( D \right),s_{2}^{1}\left( D \right) \right),$ so that the function *s*~2~↦*ψ*(*s*~2~, *D*) is convex and, thus, it satisfies assumption **H8**, see [Fig. 2](#fig0002){ref-type="fig"}(b). The minimum ${\overline{s}}_{2}\left( D \right)$ is a solution of an algebraic equation of degree 4 in *s*~2~. Although mathematical software, such as *Maple*, cannot give solutions explicitly with respect to the parameters, ${\overline{s}}_{2}\left( D \right)$ could be obtained analytically since algebraic equations of degree 4 can theoretically be solved by quadratures. We do not try to obtain such an explicit formula. However, if the biological parameters are fixed, the function ${\overline{s}}_{2}\left( D \right)$ and, hence, $F_{1}\left( D \right) = \psi\left( {\overline{s}}_{2}\left( D \right),D \right),$ can be obtained numerically. Since *M*~2~ and *ψ* are given explicitly the functions *F*~2~(*D*) and *F*~3~(*D*) are given explicitly with respect to the biological parameters in ([Eq. 23](#eq0023){ref-type="disp-formula"}). Since $\left. D\mapsto s_{2}^{0}\left( D \right) \right.$ is increasing and $\left. D\mapsto s_{2}^{1}\left( D \right) \right.$ is decreasing, and assuming $s_{2}^{0}\left( 0 \right) < s_{2}^{1}\left( 0 \right),$ the domain of definition *I*~1~ of *F*~1~(*D*) is an interval $I_{1} = \left\lbrack 0,D_{1} \right),$ where *D*~1~ is the solution of $s_{2}^{0}\left( D \right) = s_{2}^{1}\left( D \right),$ see [Fig. 3](#fig0003){ref-type="fig"}. For the domain of definition *I*~2~ of *F*~2~(*D*), several cases can be distinguished. *I*~2~ is an interval $I_{2} = \left\lbrack 0,D_{2} \right),$ where *D*~2~ is the solution of $M_{2}\left( D + a_{2} \right) = s_{2}^{1}\left( D \right),$ see [Fig. 3](#fig0003){ref-type="fig"}(a), or the solution of equation $M_{2}\left( D + a_{2} \right) = s_{2}^{0}\left( D \right),$ see [Fig. 3](#fig0003){ref-type="fig"}(b). *I*~2~ is empty, see [Fig. 3](#fig0003){ref-type="fig"}(c). *I*~2~ is an interval $I_{2} = \left( D_{2min},D_{2max} \right)$ where *D*~2*min*~ and *D*~2*max*~ are the solutions of $M_{2}\left( D + a_{2} \right) = s_{2}^{0}\left( D \right)$ and $M_{2}\left( D + a_{2} \right) = s_{2}^{1}\left( D \right),$ respectively, see [Fig. 3](#fig0003){ref-type="fig"}(d). 5. Existence of steady-states {#sec0005} ============================= In this paper, as we are restricted only to local stability analysis, any reference to steady-state stability should be considered as local stability. A steady-state of ([(14)](#eq0014){ref-type="disp-formula"}, [(15)](#eq0015){ref-type="disp-formula"}, [(16)](#eq0016){ref-type="disp-formula"}, [(17)](#eq0017){ref-type="disp-formula"}, [(18)](#eq0018){ref-type="disp-formula"}, [(19)](#eq0019){ref-type="disp-formula"}) is obtained by setting the right-hand sides equal to zero: $$\begin{aligned} {\left\lbrack \mu_{0}\left( s_{0},s_{2} \right) - D - a_{0} \right\rbrack x_{0}} & {= 0} \\ \end{aligned}$$$$\begin{aligned} {\left\lbrack \mu_{1}\left( s_{1},s_{2} \right) - D - a_{1} \right\rbrack x_{1}} & {= 0} \\ \end{aligned}$$$$\begin{aligned} {\left\lbrack \mu_{2}\left( s_{2} \right) - D - a_{2} \right\rbrack x_{2}} & {= 0} \\ \end{aligned}$$$$\begin{aligned} {D\left( s_{0}^{\text{in}} - s_{0} \right) - \mu_{0}\left( s_{0},s_{2} \right)x_{0}} & {= 0} \\ \end{aligned}$$$$\begin{aligned} {- Ds_{1} + \mu_{0}\left( s_{0},s_{2} \right)x_{0} - \mu_{1}\left( s_{1},s_{2} \right)x_{1}} & {= 0} \\ \end{aligned}$$$$\begin{aligned} {- Ds_{2} + \mu_{1}\left( s_{1},s_{2} \right)x_{1} - \omega\mu_{0}\left( s_{0},s_{2} \right)x_{0} - \mu_{2}\left( s_{2} \right)x_{2}} & {= 0} \\ \end{aligned}$$A steady-state exists (or is said to be 'meaningful') if, and only if, all its components are non-negative. Lemma 2*The only steady-state of* ([Eqs. 14](#eq0014){ref-type="disp-formula"}*--*[19](#eq0019){ref-type="disp-formula"}*), for which*$x_{0} = 0$ *or*$x_{1} = 0,$ *is the washout steady-state*${\text{SS}1} = \left( x_{0},x_{1},x_{2},s_{0},s_{1},s_{2} \right),$ *where*$x_{0} = 0,$$x_{1} = 0,$$x_{2} = 0,$$s_{0} = s_{0}^{in},$$s_{1} = 0$ *and*$s_{2} = 0$*. This steady-state always exists and is always locally stable.* Besides the steady-state SS1, the system can have at most two other type of steady-states. •SS2: *x*~0~ \> 0, *x*~1~ \> 0 and $x_{2} = 0,$ where species *x*~2~ is washed out, while species *x*~0~ and and *x*~1~ exist. We show below that, generally, the system can have two steady-states SS2^♭^ and SS2^♯^.•SS3: *x*~0~ \> 0, *x*~1~ \> 0, and *x*~2~ \> 0, where all populations are maintained. We show below that the steady-state SS3 is unique if it exists. We can state now the necessary and sufficient conditions of existence of SS2 and SS3. Lemma 3*If ω* ≥ 1 *then SS2 does not exist. If ω* \< 1 *then SS2 exists if, and only if,*$s_{0}^{\text{in}} \geq F_{1}\left( D \right)$*. Therefore, a necessary condition for the existence of SS2 is that D* ∈ *I*~1~, *where I*~1~ *is defined by (Eq.* [33](#eq0033){ref-type="disp-formula"}*). If*$s_{0}^{\text{in}} \geq F_{1}\left( D \right)$ *then each solution s*~2~ *of equation* $$\psi\left( s_{2} \right) = s_{0}^{\text{in}},\quad s_{2} \in \left( s_{2}^{0},s_{2}^{1} \right)$$*gives a steady-state*${\text{SS}2} = \left( x_{0},x_{1},x_{2},s_{0},s_{1},s_{2} \right)$ *where* $$\begin{array}{cl} & {s_{0} = M_{0}\left( D + a_{0},s_{2} \right),\quad s_{1} = M_{1}\left( D + a_{1},s_{2} \right)} \\ & {x_{0} = \frac{D}{D + a_{0}}\left( s_{0}^{\text{in}} - s_{0} \right),\quad x_{1} = \frac{D}{D + a_{1}}\left( s_{0}^{\text{in}} - s_{0} - s_{1} \right),\quad x_{2} = 0} \\ \end{array}$$ Lemma 4*If ω* ≥ 1 *then SS3 does not exist. If ω* \< 1 *then SS3 exists if, and only if,*$s_{0}^{\text{in}} > F_{2}\left( D \right)$*. Therefore, a necessary condition of existence of SS3 is that D* ∈ *I*~2~, *where I*~2~ *is defined by (Eq.* [34](#eq0034){ref-type="disp-formula"}*). If*$s_{0}^{\text{in}} > F_{2}\left( D \right)$ *then the steady-state*${\text{SS}3} = \left( x_{0},x_{1},x_{2},s_{0},s_{1},s_{2} \right)$ *is given by* $$\begin{array}{ccl} & & {s_{0} = M_{0}\left( D + a_{0},M_{2}\left( D + a_{2} \right) \right),\quad s_{1} = M_{1}\left( D + a_{1},M_{2}\left( D + a_{2} \right) \right),} \\ & & {s_{2} = M_{2}\left( D + a_{2} \right)} \\ \end{array}$$*called the break-even concentrations, and* $$\begin{array}{ccl} & & {x_{0} = \frac{D}{D + a_{0}}\left( s_{0}^{\text{in}} - s_{0} \right),\quad x_{1} = \frac{D}{D + a_{1}}\left( s_{0}^{\text{in}} - s_{0} - s_{1} \right),} \\ & & {x_{2} = \frac{D}{D + a_{2}}\left( \left( 1 - \omega \right)\left( s_{0}^{\text{in}} - s_{0} \right) - s_{1} - s_{2} \right)} \\ \end{array}$$ Remark 1If $s_{0}^{\text{in}} > F_{1}\left( D \right)$ then ([Eq. 41](#eq0041){ref-type="disp-formula"}) has exactly two solutions denoted by $s_{2}^{\flat}$ and $s_{2}^{\sharp}$ and such that, see [Fig. 2](#fig0002){ref-type="fig"}(c), $$s_{2}^{0} < s_{2}^{\flat} < {\overline{s}}_{2} < s_{2}^{\sharp} < s_{2}^{1}$$If $s_{0}^{\text{in}} = F_{1}\left( D \right)$ then $s_{2}^{0} < s_{2}^{\flat} = {\overline{s}}_{2} = s_{2}^{\sharp} < s_{2}^{1}$. To these solutions, $s_{2}^{\flat}$ and $s_{2}^{\sharp},$ correspond two steady-states of SS2, which are denoted by SS2^♭^ and SS2^♯^. These steady-states coalesce when $s_{0}^{\text{in}} = F_{1}\left( D \right)$. Since *F*~1~(*D*) ≤ *F*~2~(*D*), the condition $s_{0}^{\text{in}} > F_{2}\left( D \right)$ for the existence of the positive steady-state SS3 implies that the condition $s_{0}^{\text{in}} > F_{1}\left( D \right)$ for the existence of the two steady-states SS2^♭^ and SS2^♯^ is satisfied. Therefore, if SS3 exists then SS2^♭^ and SS2^♯^ exist and are distinct. If $s_{0}^{\text{in}} = F_{2}\left( D \right)$ then SS3 coalesces with SS2^♭^ if *F*~3~(*D*) \< 0, and with SS2^♯^ if *F*~3~(*D*) \> 0, respectively. Remark 2Using ([Eq. 20](#eq0020){ref-type="disp-formula"}), the conditions $s_{0}^{\text{in}} > F_{1}\left( D \right)$ and $s_{0}^{\text{in}} > F_{2}\left( D \right)$ of existence of the steady-state SS2 and SS3, respectively are equivalent to the conditions $$S_{\text{ch},\text{in}} > \frac{F_{1}\left( D \right)}{Y_{3}Y_{4}}\text{and} S_{\text{ch},\text{in}} > \frac{F_{2}\left( D \right)}{Y_{3}Y_{4}}$$respectively expressed with respect to the inflowing concentration *S*~ch,\ in~. Moreover, for the nominal parameter values given in [Table 1](#tbl0001){ref-type="table"}, we have *ω* ≈ 0.53. Therefore, the steady-states SS2 and SS3 exist if the conditions on the operating parameters stated in [Lemmas 3](#enun0004){ref-type="statement"} and [4](#enun0005){ref-type="statement"}, respectively are satisfied, see [Table 7](#tbl0007){ref-type="table"}. 6. Operating diagram {#sec0006} ==================== The *Operating diagrams* show how the system behaves when we vary the two control parameters *S*~ch,\ in~ and *D* (i.e., substrate inflow rate and medium flow-rate per culture volume) in ([(1)](#eq0001){ref-type="disp-formula"}, [(2)](#eq0002){ref-type="disp-formula"}, [(3)](#eq0003){ref-type="disp-formula"}, [(4)](#eq0004){ref-type="disp-formula"}, [(5)](#eq0005){ref-type="disp-formula"}, [(6)](#eq0006){ref-type="disp-formula"}). This conventional bifurcation plot is used to visualise the existence and stability of steady-states with respect to these operating parameters, as they are the parameters most easily manipulated in a chemostat by which the system under scrutiny can be examined. According to [Remark 2](#enun0007){ref-type="statement"}, the curve *Γ*~1~ of equation $$S_{\text{ch},\text{in}} = \frac{1}{Y_{3}Y_{4}}F_{1}\left( D \right)$$is the border to which SS2 exists, and the curve *Γ*~2~ of equation $$S_{\text{ch},\text{in}} = \frac{1}{Y_{3}Y_{4}}F_{2}\left( D \right)$$is the border to which SS3 exists, see [Fig. 4](#fig0004){ref-type="fig"}. If we want to plot the operating diagram we must fix the values of the biological parameters. In the remainder of the Section we plot the operating diagrams corresponding to cases (a)--(d) depicted in [Fig. 3](#fig0003){ref-type="fig"}.Fig. 4The curves *Γ*~1~ (black), *Γ*~2~ (red) and *Γ*~3~ (green) for case (a). (i) : regions of steady-state existence, with maintenance. On the right, a magnification for $0 < D < D_{3} = 0.058$ showing the region $\mathcal{J}_{4}$. (ii) : regions of steady-state existence and their stability, without maintenance. On the right, a magnification for $0 < D < D_{3} = 0.078$ showing the regions $\mathcal{J}_{4}$ and $\mathcal{J}_{5}$. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)Fig. 4 6.1. Operating diagram: case (a) {#sec0007} -------------------------------- This case corresponds to the parameter values used by [@bib0024]. We have seen in [Table 2](#tbl0002){ref-type="table"} that the curves *Γ*~1~ and *Γ*~2~ are defined for *D* \< *D*~1~ and *D* \< *D*~2~, respectively and that they are tangent for $D = D_{3},$ where $D_{1} = 0.432,$$D_{2} = 0.373$ and $D_{3} = 0.058$. Therefore, they separate the operating plane (*S*~ch,\ in~, *D*) into four regions, as shown in [Fig. 4](#fig0004){ref-type="fig"}(i), labelled $\mathcal{J}_{1},$$\mathcal{J}_{2}$ and $\mathcal{J}_{3}$ and $\mathcal{J}_{4}$.Table 2Parameter values of the maintenance terms *a~i~*, $i = 0,1,2,$ for cases (a), (b) and (c) of [Fig. 3](#fig0003){ref-type="fig"}. Unspecified parameter values are as in [Table 1](#tbl0001){ref-type="table"}. The table gives the values of *D*~1~, *D*~2~ and *D*~3~ where $I_{1} = \left\lbrack 0,D_{1} \right),$$I_{2} = \left\lbrack 0,D_{2} \right)$ and $I_{3} = \left\lbrack 0,D_{3} \right)$.Table 2$K_{S,{H_{2},c}}$*a~i~D*~1~*D*~2~*D*~3~(a)$1.0 \times 10^{- 6}$0.020.4320.3730.05800.4520.3930.078(b)$4.0 \times 10^{- 6}$0.020.3290.236$I_{3} = I_{2}$00.3490.256$I_{3} = I_{2}$(c)$7.0 \times 10^{- 6}$0.020.287$I_{2} = \varnothing$00.303$I_{2} = \varnothing$ The results are summarised in [Table 4](#tbl0004){ref-type="table"}, which shows the existence of the steady-states SS1, SS2 and SS3 in the regions of the operating diagram in [Fig. 4](#fig0004){ref-type="fig"}(i). 6.2. Operating diagram: case (b) {#sec0008} -------------------------------- This case corresponds to the parameter values used by [@bib0024], except that $K_{S,{H_{2},c}}$ is changed from $1.0 \times 10^{- 6}$ to $4.0 \times 10^{- 6}$. We have seen in [Table 2](#tbl0002){ref-type="table"} that the curves *Γ*~1~ and *Γ*~2~ are defined for *D* \< *D*~1~ and *D* \< *D*~2~, respectively and *F*~1~(*D*) \< *F*~2~(*D*) for all *D* \< *D*~2~, where $D_{1} = 0.329$ and $D_{2} = 0.236$. Therefore, they separate the operating plane (*S*~ch,\ in~, *D*) in three regions, as shown in [Fig. 5](#fig0005){ref-type="fig"}(i), labelled $\mathcal{J}_{1},$$\mathcal{J}_{3}$ and $\mathcal{J}_{4}$.Fig. 5The curves *Γ*~1~ (black), *Γ*~2~ (red) and *Γ*~3~ (green) for case (b). (i) : regions of steady-state existence, with maintenance. (ii) : regions of steady-state existence and their stability, without maintenance. On the right, a magnification for 0 \< *D* \< 0.1. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)Fig. 5 The results are summarised in [Table 5](#tbl0005){ref-type="table"}, which shows the existence of the steady-states SS1, SS2 and SS3 in the regions of the operating diagram in [Fig. 5](#fig0005){ref-type="fig"}(i). Note that the region $\mathcal{J}_{2}$ has disappeared. 6.3. Operating diagram: case (c) {#sec0009} -------------------------------- This case corresponds to the parameter values used by [@bib0024], except that $K_{S,{H_{2},c}}$ is changed from $1.0 \times 10^{- 6}$ to $7.0 \times 10^{- 6}$. We have seen in [Table 2](#tbl0002){ref-type="table"} that the curve *Γ*~1~ is defined for $D < D_{1} = 0.287$ and that *I*~2~ is empty so that SS3 does not exist. Therefore, *Γ*~1~ separates the operating plane (*S*~ch,\ in~, *D*) in two regions, as shown in [Fig. 6](#fig0006){ref-type="fig"}(i), labelled $\mathcal{J}_{1}$ and $\mathcal{J}_{4}$.Fig. 6The curve *Γ*~1~ for case (c). (i) : regions of steady-state existence, with maintenance. (ii) : regions of steady-state existence and their stability, without maintenance. On the right, a magnification for 0 \< *D* \< 0.1.Fig. 6 The results are summarised in [Table 6](#tbl0006){ref-type="table"}, which shows the existence of the steady-states SS1 and SS2 in the regions of the operating diagram in [Fig. 6](#fig0006){ref-type="fig"}(i). Note that the region $\mathcal{J}_{3}$ of existence of SS3 has disappeared. 6.4. Operating diagram: case (d) {#sec0010} -------------------------------- We end this discussion on the role of kinetic parameters by the presentation of this case, which presents a new behaviour that did not occur in the preceding cases: there exists *D*~2*min*~ such that for *D* \< *D*~2*min*~ the system cannot have a positive steady-state SS3. This case corresponds to the parameter values used by [@bib0024], except that three of them are changed as indicated in [Table 3](#tbl0003){ref-type="table"}. This table shows that the curves *Γ*~1~ and *Γ*~2~ are defined for *D* \< *D*~1~ and *D*~2*min*~ \< *D* \< *D*~2*max*~ and that they are tangent for $D = D_{3},$ where $D_{1} = 0.238,$$D_{2min} = 0.101,$$D_{2max} = 0.198$ and $D_{3} = 0.161$. Therefore, *Γ*~1~ and *Γ*~2~ separate the operating plane (*S*~ch,\ in~, *D*) in four regions, as shown in [Fig. 7](#fig0007){ref-type="fig"}(i), labelled $\mathcal{J}_{1},$$\mathcal{J}_{2},$$\mathcal{J}_{3}$ and $\mathcal{J}_{4}$. The results are summarised in [Table 4](#tbl0004){ref-type="table"}, which shows the existence of the steady-states SS1, SS2 and SS3 in the regions of the operating diagram in [Fig. 7](#fig0007){ref-type="fig"}(i).Fig. 7The curves *Γ*~1~ (black), *Γ*~2~ (red) and *Γ*~3~ (green) for case (d). (i) : regions of steady-state existence, with maintenance. (ii) : regions of steady-state existence and their stability, without maintenance. On the right, a magnification for 0 \< *D* \< 0.1. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)Fig. 7Table 3Parameter values of the maintenance terms *a~i~*, $i = 0,1,2,$ for case (d) of [Fig. 3](#fig0003){ref-type="fig"}: $K_{S,{H_{2},c}} = 1.2 \times 10^{- 5},$$K_{S,H_{2}} = 0.5 \times 10^{- 5}$ and $k_{m,H_{2}} = 5$. Unspecified parameter values are as in [Table 1](#tbl0001){ref-type="table"}. The table gives the values of *D*~1~, *D*~2*min*~, *D*~2*max*~ and *D*~3~ where $I_{1} = \left\lbrack 0,D_{1} \right),$$I_{2} = \left( D_{2min},D_{2max} \right)$ and $I_{3} = \left( D_{2min},D_{3} \right)$.Table 3*a~i~D*~1~*D*~2*min*~*D*~2*max*~*D*~3~(d)0.020.2380.1010.1980.16100.2580.1210.2180.181Table 4Existence of steady-states in the regions of the operating diagrams of [Fig. 4](#fig0004){ref-type="fig"}(i) and [Fig. 7](#fig0007){ref-type="fig"}(i).Table 4RegionSteady-states$\mathcal{J}_{1}$SS1$\mathcal{J}_{2} \cup \mathcal{J}_{4}$SS1, SS2^♭^, SS2^♯^$\mathcal{J}_{3}$SS1, SS2^♭^, SS2^♯^, SS3Table 5Existence of steady-states in the regions of the operating diagram of [Fig 5](#fig0005){ref-type="fig"}(i).Table 5RegionSteady-states$\mathcal{J}_{1}$SS1$\mathcal{J}_{4}$SS1, SS2^♭^, SS2^♯^$\mathcal{J}_{3}$SS1, SS2^♭^, SS2^♯^, SS3Table 6Existence of steady-states in the regions of the operating diagram of [Fig 6](#fig0006){ref-type="fig"}(i).Table 6RegionSteady-states$\mathcal{J}_{1}$SS1$\mathcal{J}_{4}$SS1, SS2^♭^, SS2^♯^Table 7Existence (with or without maintenance) and stability (without maintenance) of steady-states. The conditions for existence and stability are satisfied given that the inequalities are observed.Table 7ExistenceStabilitySS1Always existsAlways stableSS2^♭^$s_{0}^{\text{in}} > F_{1}\left( D \right)$Always unstableSS2^♯^$s_{0}^{\text{in}} > F_{1}\left( D \right)$*F*~3~(*D*) \> 0 and $s_{0}^{in} < F_{2}\left( D \right)$SS3$s_{0}^{\text{in}} > F_{2}\left( D \right)$*F*~3~(*D*) ≥ 0 or*F*~3~(*D*) \< 0 and $F_{4}\left( D,s_{0}^{\text{in}} \right) > 0$ 7. Local stability without maintenance {#sec0011} ====================================== We know that SS1 is always stable. However, the analytical study of the stability of SS2 and SS3 is very difficult because the conditions for Routh-Hurwitz in the 6-dimensional case are intractable. For this reason we will consider in this section the question of the stability only for the case without maintenance ($a_{0} = a_{1} = a_{2} = 0$), since the system reduces to a three-dimensional system. The general case will be considered only numerically in [Section 8](#sec0016){ref-type="sec"}. When maintenance is not considered in the model, the steady-states SS1, SS2 and SS3 are given by: 1.${\text{SS}1} = \left( 0,0,0,s_{0}^{\text{in}},0,0 \right)$2.${\text{SS}2} = \left( x_{0},x_{1},0,s_{0},s_{1},s_{2} \right)$ where *s*~2~ a solution of equation $$s_{0}^{\text{in}} = \psi\left( s_{2} \right) = M_{0}\left( D,s_{2} \right) + \frac{M_{1}\left( D,s_{2} \right) + s_{2}}{1 - \omega}$$and $$s_{0} = M_{0}\left( D,s_{2} \right),\;\;\, s_{1} = M_{1}\left( D,s_{2} \right)\quad x_{0} = s_{0}^{\text{in}} - s_{0},\;\;\, x_{1} = s_{0}^{\text{in}} - s_{0} - s_{1}$$3.${\text{SS}3} = \left( x_{0},x_{1},x_{2},s_{0},s_{1},s_{2} \right)$ where $$\begin{aligned} s_{2} & {= M_{2}\left( D \right),\quad s_{0} = M_{0}\left( D,s_{2} \right),\quad s_{1} = M_{1}\left( D,s_{2} \right)} \\ x_{0} & {= s_{0}^{\text{in}} - s_{0},} \\ x_{1} & {= s_{0}^{\text{in}} - s_{0} - s_{1},_{q}uadx_{2} = \left( 1 - \omega \right)\left( s_{0}^{\text{in}} - s_{0} \right) - s_{1} - s_{2}} \\ \end{aligned}$$ Proposition 2*Let*${\text{SS}2} = \left( x_{0},x_{1},0,s_{0},s_{1},s_{2} \right)$ *be a steady-state. Then SS2 is stable if, and only if, μ*~2~(*s*~2~) \< *D and*$\frac{\partial\psi}{\partial s_{2}} > 0$*.* Therefore, SS2^♭^ is always unstable and SS2^♯^ is stable if, and only if, *μ*~2~(*s*~2~) \< *D*. This last condition is equivalent to $M_{2}\left( D \right) > s_{2}^{\sharp},$ which implies that *F*~3~(*D*) \> 0. Hence, if SS3 exists then SS2^♯^ is necessarily unstable. Therefore, SS2^♯^ is stable if, and only if, *F*~3~(*D*) \> 0 and SS3 does not exist. Proposition 3*Let*${\text{SS}3} = \left( x_{0},x_{1},x_{2},s_{0},s_{1},s_{2} \right)$ *be a steady-state. If F*~3~(*D*) ≥ 0 *then SS3 is stable as long as it exists. If F*~3~(*D*) \< 0 *then SS3 can be unstable. The instability of SS3 occurs inparticular when s*~2~ *is sufficiently close to*$s_{2}^{\flat},$ *i.e. when SS3 is sufficiently close to* SS2^♭^*.* The condition *F*~3~(*D*) ≥ 0 is equivalent to $\frac{\partial\psi}{\partial s_{2}}\left( M_{2}\left( D \right) \right) \geq 0,$ namely $s_{2} = M_{2}\left( D \right) \in \left\lbrack {\overline{s}}_{2},s_{2}^{\sharp} \right)$. If $\frac{\partial\psi}{\partial s_{2}} < 0,$ viz. $s_{2} \in \left( s_{2}^{\flat},{\overline{s}}_{2} \right),$ then SS3 can be unstable. When *D* is such that *F*~3~(*D*) \< 0, the determination of the boundary between the regions of stability and instability of SS3 needs to examine the Routh-Hurwitz condition of stability for SS3. For this purpose we define the following functions. Let ${\text{SS}3} = \left( x_{0},x_{1},x_{2},s_{0},s_{1},s_{2} \right)$ be a steady-state and $$E = \frac{\partial\mu_{0}}{\partial s_{0}},\quad F = \frac{\partial\mu_{0}}{\partial s_{2}},\quad G = \frac{\partial\mu_{1}}{\partial s_{1}},\quad H = - \frac{\partial\mu_{1}}{\partial s_{2}},\quad I = \frac{d\mu_{2}}{ds_{2}}$$evaluated at the steady-state SS3 defined by [(48)](#eq0048){ref-type="disp-formula"}, i.e. for $$s_{2} = M_{2}\left( D \right),\quad s_{0} = M_{0}\left( D,s_{2} \right),\quad s_{1} = M_{1}\left( D,s_{2} \right)$$For *D* ∈ *I*~3~ and $s_{0}^{\text{in}} > F_{2}\left( D \right),$ we define: $$\begin{array}{ccl} {F_{4}\left( D,s_{0}^{\text{in}} \right)} & = & {\left( EIx_{0}x_{2} + \left\lbrack E\left( G + H \right) - \left( 1 - \omega \right)FG \right\rbrack x_{0}x_{1} \right)f_{2}} \\ & & {+ \left( Ix_{2} + \left( G + H \right)x_{1} + \omega Fx_{0} \right)GIx_{1}x_{2}} \\ \end{array}$$where $f_{2} = Ix_{2} + \left( G + H \right)x_{1} + \left( E + \omega F \right)x_{0}$. Notice that to compute $F_{4}\left( D,s_{0}^{\text{in}} \right)$ we must replace *x*~0~, *x*~1~, *x*~2~, *s*~0~, *s*~1~ and *s*~2~ by their values at SS3, given by [(48)](#eq0048){ref-type="disp-formula"}. Hence, this function depends on the operating parameters *D* and $s_{0}^{\text{in}}$. For each fixed *D* ∈ *I*~3~, $F_{4}\left( D,s_{0}^{\text{in}} \right)$ is polynomial in $s_{0}^{\text{in}}$ of degree 3 and tends to $+ \infty$ when $s_{0}^{\text{in}}$ tends to $+ \infty$. Therefore, it is necessarily positive for large enough $s_{0}^{\text{in}}$. The values of the operating parameters *D* and $s_{0}^{\text{in}}$ for which $F_{4}\left( D,s_{0}^{\text{in}} \right)$ is positive correspond to the stability of SS3 as shown in the following proposition. Proposition 4*Let*${\text{SS}3} = \left( x_{0},x_{1},x_{2},s_{0},s_{1},s_{2} \right)$ *be a steady-state. If F*~3~(*D*) \< 0 *then SS3 is stable if, and only if,*$F_{4}\left( D,s_{0}^{\text{in}} \right) > 0$*.* The results on the existence of steady-states (with or without maintenance) of [Lemmas 2](#enun0003){ref-type="statement"}--[4](#enun0005){ref-type="statement"}, and their stability (without maintenance) of [Props 2](#enun0008){ref-type="statement"}--[4](#enun0010){ref-type="statement"}, are summarised in [Table 7](#tbl0007){ref-type="table"}. 7.1. Operating diagram: case (a) {#sec0012} -------------------------------- This case corresponds to the parameter values used by [@bib0024] but without maintenance. We see from [Table 2](#tbl0002){ref-type="table"} that the curves *Γ*~1~ and *Γ*~2~ of the operating diagram, given by ([Eq. 45](#eq0045){ref-type="disp-formula"}) and ([Eq. 46](#eq0046){ref-type="disp-formula"}), respectively are defined now for *D* \< *D*~1~ and *D* \< *D*~2~, respectively and that they are tangent for $D = D_{3},$ where $D_{1} = 0.452,$$D_{2} = 0.393$ and $D_{3} = 0.078$. Beside these curves, we plot also on the operating diagram of [Fig. 4](#fig0004){ref-type="fig"}(ii), the curve *Γ*~3~ of equation $$F_{4}\left( D,Y_{3}Y_{4}S_{\text{ch},\text{in}} \right) = 0$$According to [Prop. 4](#enun0010){ref-type="statement"}, this curve is defined for $D < D_{3} = 0.078$ and it separates the region of existence of SS3 into two subregions labelled $\mathcal{J}_{3}$ and $\mathcal{J}_{5},$ such that SS3 is stable in $\mathcal{J}_{3}$ and unstable in $\mathcal{J}_{5}$. The other regions $\mathcal{J}_{1},$$\mathcal{J}_{2}$ and $\mathcal{J}_{4}$ are defined as in the previous section. The operating diagram is shown [Fig. 4](#fig0004){ref-type="fig"}(ii). It looks very similar to [Fig. 4](#fig0004){ref-type="fig"}(i), except near the origin, as it is indicated in the magnification for $0 < D < D_{3} = 0.078$. From [Table 7](#tbl0007){ref-type="table"}, we deduce the following result Proposition 5[Table 8](#tbl0008){ref-type="table"} *shows the existence and stability of the steady-states SS1, SS2 and SS3 in the regions of the operating diagram in* [Fig. 4](#fig0004){ref-type="fig"}*(ii).*Table 8Existence and stability of steady-states in the regions of the operating diagrams of [Fig. 4](#fig0004){ref-type="fig"}(ii) and [Fig. 7](#fig0007){ref-type="fig"}(ii)., where S and U indicate stable and unstable steady-states, respectively.Table 8RegionSS1SS2^♭^SS2^♯^SS3$\mathcal{J}_{1}$S$\mathcal{J}_{2}$SUS$\mathcal{J}_{3}$SUUS$\mathcal{J}_{4}$SUU$\mathcal{J}_{5}$SUUU 7.2. Operating diagram: case (b) {#sec0013} -------------------------------- We see from [Table 2](#tbl0002){ref-type="table"} that the curves *Γ*~1~ and *Γ*~2~ are defined now for $D < D_{1} = 0.349$ and $D < D_{2} = 0.256,$ respectively and that *F*~1~(*D*) \< *F*~2~(*D*) for all *D*. Beside these curves, we plot also on the operating diagram of [Fig. 5](#fig0005){ref-type="fig"}(ii), the curve *Γ*~3~ of equation ([Eq. 50](#eq0050){ref-type="disp-formula"}) which separates the region of existence of SS3 into two subregions labelled $\mathcal{J}_{3}$ and $\mathcal{J}_{5},$ such that SS3 is stable in $\mathcal{J}_{3}$ and unstable in $\mathcal{J}_{5}$. Therefore, the curves *Γ*~1~, *Γ*~2~ and *Γ*~3~ separate the operating plane (*S*~ch,\ in~, *D*) into four regions, as shown in [Fig. 5](#fig0005){ref-type="fig"}(ii), labelled $\mathcal{J}_{1},$$\mathcal{J}_{3},$$\mathcal{J}_{4}$ and $\mathcal{J}_{5}$. Proposition 6[Table 9](#tbl0009){ref-type="table"} *shows the existence and stability of the steady-states SS1, SS2 and SS3 in the regions of the operating diagram in* [Fig. 5](#fig0005){ref-type="fig"}*(ii)*.Table 9Existence and stability of steady-states in the regions of the operating diagram of [Fig. 5](#fig0005){ref-type="fig"}(ii).Table 9RegionSS1SS2^♭^SS2^♯^SS3$\mathcal{J}_{1}$S$\mathcal{J}_{3}$SUUS$\mathcal{J}_{4}$SUU$\mathcal{J}_{5}$SUUU 7.3. Operating diagram: case (c) {#sec0014} -------------------------------- We see from [Table 2](#tbl0002){ref-type="table"} that *Γ*~1~ is defined for $D < D_{1} = 0.303$ and that *I*~2~ is empty so that SS3 does not exist. Therefore, *Γ*~1~ separates the operating plane (*S*~ch,\ in~, *D*) into two regions, as shown in [Fig. 6](#fig0006){ref-type="fig"}(ii), labelled $\mathcal{J}_{1}$ and $\mathcal{J}_{4}$. Proposition 7[Table 10](#tbl0010){ref-type="table"} *shows the existence and stability of the steady-states SS1, SS2 and SS3 in the regions of the operating diagram in* [Fig. 6](#fig0006){ref-type="fig"}*(ii).*Table 10Existence and stability of steady-states in the regions of the operating diagram of [Fig. 6](#fig0006){ref-type="fig"}(ii).Table 10RegionSS1SS2^♭^SS2^♯^SS3$\mathcal{J}_{1}$S$\mathcal{J}_{4}$SUU 7.4. Operating diagram: case (d) {#sec0015} -------------------------------- We see in [Table 3](#tbl0003){ref-type="table"} that the curves *Γ*~1~ and *Γ*~2~ are defined for *D* \< *D*~1~ and *D*~2*min*~ \< *D* \< *D*~2*max*~ and that they are tangent for $D = D_{3},$ where $D_{1} = 0.258$ and $D_{2min} = 0.121,$$D_{2max} = 0.218$ and $D_{3} = 0.181$. Beside these curves, we plot also on the operating diagram of [Fig. 7](#fig0007){ref-type="fig"}(ii), the curve *Γ*~3~ defined by ([Eq. 50](#eq0050){ref-type="disp-formula"}), which separates the region of existence of SS3 into two subregions labelled $\mathcal{J}_{3}$ and $\mathcal{J}_{5},$ such that SS3 is stable in $\mathcal{J}_{3}$ and unstable in $\mathcal{J}_{5}$. Therefore, the curves *Γ*~1~, *Γ*~2~ and *Γ*~3~ separate the operating plane (*S*~ch,\ in~, *D*) into five regions, as shown in [Fig. 7](#fig0007){ref-type="fig"}(ii), labelled $\mathcal{J}_{1},$$\mathcal{J}_{2},$$\mathcal{J}_{3},$$\mathcal{J}_{4}$ and $\mathcal{J}_{5}$. Proposition 8[Table 8](#tbl0008){ref-type="table"} *shows the existence and stability of the steady-states SS1, SS2 and SS3 in the regions of the operating diagram in* [Fig. 7](#fig0007){ref-type="fig"}*(ii).* 8. Numerical analysis to illustrate the analytical results {#sec0016} ========================================================== The aim of this section is to study numerically (the method is explained in [Appendix A](#sec0023){ref-type="sec"}) the existence and stability of the steady-states SS2 and SS3. We obtain numerically the operating diagrams that were described in [Sections 5](#sec0005){ref-type="sec"} and [7](#sec0011){ref-type="sec"}. The results in this section confirm the results on existence of the steady-states obtained in [Section 5](#sec0005){ref-type="sec"} in the cases with or without maintenance and the results of stability obtained in [Section 7](#sec0011){ref-type="sec"} in the case without maintenance. These results permit also to extend the analytical results and elucidate the problem of the local stability of SS2 and SS3, which was left open in [Section 7](#sec0011){ref-type="sec"}. 8.1. Operating diagram: case (a) {#sec0017} -------------------------------- We endeavoured to find numerically the operating conditions under which SS3 is unstable, previously unreported by [@bib0024]. Given that we have determined analytically in [Proposition 3](#enun0009){ref-type="statement"} that when SS3 is close to SS2^♭^ it becomes unstable, we performed numerical simulations with the parameters defined in [Table 1](#tbl0001){ref-type="table"} over an operating region similar to that shown in [Fig. 2](#fig0002){ref-type="fig"} from [@bib0024] whilst also satisfying our conditions. In [Fig. 8](#fig0008){ref-type="fig"} we show the case when maintenance is excluded. When magnified, we observe more clearly that region $\mathcal{J}_{5}$ does exist for the conditions described above, and also note that the region $\mathcal{J}_{4}$ occurs in a small area between $\mathcal{J}_{1}$ and $\mathcal{J}_{5},$ which corresponds to the results shown in [Fig. 4](#fig0004){ref-type="fig"}(ii), and is in agreement with [Proposition 5](#enun0011){ref-type="statement"}. In [Fig. 9](#fig0009){ref-type="fig"} we confirm that region $\mathcal{J}_{5}$ does exist for the conditions described above, when maintenance is included, but could not be determined analytically, the curve *Γ*~3~ is absent in [Fig. 4](#fig0004){ref-type="fig"}(i). Furthermore, we demonstrate that there is evidence of Hopf bifurcation, which occurs along the boundary of *F*~3~(*D*) for values of *D* \< *D*~3~ by selecting values of *S*~ch,\ in~ (indicated by $\left( \alpha \right) - \left( \delta \right)$ in [Fig. 9](#fig0009){ref-type="fig"}) at a fixed dilution rate of $0.01\mspace{6mu} d^{- 1},$ and running dynamic simulations for 10000 *d*. The three-dimensional phase plots, with the axes representing biomass concentrations, are shown in [Fig. 10](#fig0010){ref-type="fig"}, and show that as *S*~ch,\ in~ approaches $\mathcal{J}_{3}$ from $\mathcal{J}_{5},$ emergent periodic orbits are shown to diminish to a stable limit cycle at the boundary (see  [Appendix B](#sec0024){ref-type="sec"} for details). Subsequently, increasing *S*~ch,\ in~ to $\mathcal{J}_{3}$ results in the orbit reducing to a fixed point equilibrium at SS3.Fig. 8Numerical analysis for the existence and stability of steady-states for case (a), without maintenance. On the right, a magnification for 0 \< *D* \< 0.16.Fig. 8Fig. 9Numerical analysis for the existence and stability of steady-states for case (a), with maintenance. This is a magnification for 0 \< *D* \< 0.1, showing the presence and extent of region $\mathcal{J}_{5}$ undetectable by the analytical method. The coordinates labelled $\left( \alpha \right) - \left( \delta \right)$ are subsequently used to simulate the system dynamics, as shown in the proceeding [Fig. 10](#fig0010){ref-type="fig"}.Fig. 9Fig. 10Three-dimensional phase plane diagrams of the biomass dynamics for $t = 10000\mspace{6mu} d,$ showing initial (green dot) and final (red dot) conditions for a dilution rate, $D = 0.01\mspace{6mu} d^{- 1}$ and chlorophenol input, *S*~ch,\ *in*~ (*kgCOD*/*m*^3^), of (*α*): $S_{\text{ch},in} = 0.01$ - the system converges to SS1, (*β*): $S_{\text{ch},in} = 0.097$ - the system enters a periodic orbit of increasing amplitude, ultimately converging to SS1, (*γ*): $S_{\text{ch},in} = 0.10052$ - the system is close to a stable limit cycle, (*δ*): $S_{\text{ch},in} = 0.16$ - the system undergoes damped oscillations and converges to SS3. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)Fig. 10 8.2. Operating diagram: case (b) {#sec0018} -------------------------------- Whilst the numerical parameters chosen for this work are taken from the original study [@bib0024], there somewhat arbitrary nature leaves room to explore the impact of the parameters on the existence and stability of the steady-states. Case (b), discussed in [Sections 6.2](#sec0008){ref-type="sec"} and [7.2](#sec0013){ref-type="sec"}, involves a small increase to the half-saturation constant (or inverse of substrate affinity), $K_{S,{H_{2},c}},$ of the chlorophenol degrader on hydrogen. Following the same approach as with the preceding case, we confirm in [Fig. 11](#fig0011){ref-type="fig"}(i) the [Proposition 6](#enun0012){ref-type="statement"} in the scenario without maintenance. Furthermore, the extension of this proposition with maintenance included, corresponding to the existence and stability of all three steady-states given in [Table 9](#tbl0009){ref-type="table"}, is shown in Fig [11](#fig0011){ref-type="fig"}(ii). It shows the region $\mathcal{J}_{5}$ that cannot be obtained analytically (cf. [Fig. 5](#fig0005){ref-type="fig"}(i)). In both cases, region $\mathcal{J}_{2}$ has disappeared, as observed analytically. Additionally, the ideal region $\mathcal{J}_{3},$ where all organisms are present and stable, diminishes.Fig. 11Numerical analysis for the existence and stability of steady-states for case (b). (i) : without maintenance. (ii) : with maintenance.Fig. 11 8.3. Operating diagram: case (c) {#sec0019} -------------------------------- Here, $K_{S,H_{2},c},$ was further increased and confirm the [Proposition 7](#enun0013){ref-type="statement"}, where the function SS3 never exist and SS2 never stable for the case without maintenance. The extension of this proposition to the case with maintenance, shown in [Table 10](#tbl0010){ref-type="table"}, produce similar results as shown in the comparison of [Fig. 12](#fig0012){ref-type="fig"}(i) and (ii).Fig. 12Numerical analysis for the existence and stability of steady-states for case (c). (i) : without maintenance. (ii) : with maintenance. On the right, a magnification for 0 \< *D* \< 0.1.Fig. 12 8.4. Operating diagram: case (d) {#sec0020} -------------------------------- With the final investigated scenario, where $k_{m,H_{2}} < k_{m,\text{ch}}$ and $K_{S,H_{2}} < K_{S,{H_{2},c}},$ we observe once again the presence of all operating regions, $\mathcal{J}_{1} - \mathcal{J}_{5},$ without and with maintenance, as shown in [Fig. 13](#fig0013){ref-type="fig"}. It can be seen that regions $\mathcal{J}_{4}$ and $\mathcal{J}_{5}$ increase at low dilution rates across a much larger range of *S*~ch,\ in~ than in the default case (a), and the desired condition (stable SS3) is restricted to a much narrower set of *D*.Fig. 13Numerical analysis for the existence and stability of steady-states for case (d). (i) : without maintenance. (ii) : with maintenance.Fig. 13 As with the previous cases, the numerical analysis for case (d) confirms the [Proposition 8](#enun0014){ref-type="statement"} without maintenance and its extension to the case with maintenance, indicated in [Table 8](#tbl0008){ref-type="table"}. 9. The role of kinetic parameters {#sec0021} ================================= Finally, we give brief consideration to the characterisation of the four cases discussed in the preceding sections. The main difference between cases (a) or (b) and cases (c) or (d) is that, for small values of *D*, the coexistence steady-state SS3 can exist for cases (a) and (b), but cannot exist for cases (c) or (d). The cases (a) or (b) occur if and only if $s_{2}^{0}\left( 0 \right) < M_{2}\left( 0 \right)$ holds or $s_{2}^{0}\left( 0 \right) = M_{2}\left( 0 \right)$ and $\frac{ds_{2}^{0}}{dD}\left( 0 \right) < \frac{dM_{2}}{dD}\left( 0 \right)$ hold, viz. $$\begin{array}{r} {\frac{L_{0}a_{0}}{m_{0} - a_{0}} < \frac{K_{2}a_{2}}{m_{2} - a_{2}}\text{or}} \\ \end{array}$$$$\begin{array}{r} {\frac{L_{0}a_{0}}{m_{0} - a_{0}} = \frac{K_{2}a_{2}}{m_{2} - a_{2}}\text{and}\frac{L_{0}m_{0}}{\left( m_{0} - a_{0} \right)^{2}} < \frac{K_{2}m_{2}}{\left( m_{2} - a_{2} \right)^{2}}} \\ \end{array}$$The cases (c) or (d) occur if and only if $s_{2}^{0}\left( 0 \right) > M_{2}\left( 0 \right)$ holds or $s_{2}^{0}\left( 0 \right) = M_{2}\left( 0 \right)$ and $\frac{ds_{2}^{0}}{dD}\left( 0 \right) > \frac{dM_{2}}{dD}\left( 0 \right)$ hold, viz. $$\begin{array}{r} {\frac{L_{0}a_{0}}{m_{0} - a_{0}} > \frac{K_{2}a_{2}}{m_{2} - a_{2}}\text{or}} \\ \end{array}$$$$\begin{array}{r} {\frac{L_{0}a_{0}}{m_{0} - a_{0}} = \frac{K_{2}a_{2}}{m_{2} - a_{2}}\text{and}\frac{L_{0}m_{0}}{\left( m_{0} - a_{0} \right)^{2}} > \frac{K_{2}m_{2}}{\left( m_{2} - a_{2} \right)^{2}}} \\ \end{array}$$ Notice that it is easy to make the difference between case (c) and case (d): the first occurs when $M_{2}\left( D_{1} \right) < s_{2}^{0}\left( D_{1} \right)$ and the second when $M_{2}\left( D_{1} \right) > s_{2}^{0}\left( D_{1} \right)$. Since *D*~1~ is the positive solution of the algebraic quadratic equation $s_{2}^{0}\left( D \right) = s_{2}^{1}\left( D \right),$ it is possible to have an expression for *D*~1~ with respect to the biological parameters. However, this is a complicated expression involving many parameters and the preceding conditions $M_{2}\left( D_{1} \right) < s_{2}^{0}\left( D_{1} \right)$ or $M_{2}\left( D_{1} \right) > s_{2}^{0}\left( D_{1} \right)$ have no biological interpretation. We simply remark here that the function $s_{2}^{0}\left( D \right)$ has a vertical asymptote for $D = m_{0} - a_{0}$ and the function *M*~2~(*D*) has a vertical asymptote for $D = m_{2} - a_{2}$. Therefore, if $m_{0} - a_{0} < m_{2} - a_{2}$ then case (c) occurs, so that a necessary (but not sufficient) condition for case (d) to occur is $m_{0} - a_{0} > m_{2} - a_{2}$. If *m*~2~ is sufficiently small then case (d) can occur. The observations from the numerical analysis suggest that the role of the chlorophenol degrader as a secondary hydrogen scavenger is critical in maintaining full chlorophenol mineralisation and system stability, particularly at higher dilution rates, as shown by comparing cases (c) and (d) . More significantly, the results coupled with the parameter relationships shown in ([(51)](#eq0051){ref-type="disp-formula"}, [(52)](#eq0052){ref-type="disp-formula"}, [(53)](#eq0053){ref-type="disp-formula"}, [(54)](#eq0054){ref-type="disp-formula"}), highlight the necessary conditions under which the ideal case (SS3 stable) is achieved and, in general, this is a coupling of the two key parameters describing the half-saturation constant and maximum specific growth rates between the two hydrogen competitors. 10. Conclusions {#sec0022} =============== In this work we have generalised a simplified mechanistic model describing the anaerobic mineralisation of chlorophenol in a two-step food-web. We give conditions for the existence and stability of the steady-states in the case that maintenance is excluded from the system. However, with a decay term present, purely analytical determination of stability was not achievable. We confirm the findings of previous numerical analysis by [@bib0024] that with chlorophenol as the sole input substrate, three steady-states are possible. However, the analysis goes further and we determine that under certain operating conditions, two of these steady-states (SS2 and SS3) can become stable, whilst SS1 always exists and is always stable. Furthermore, without maintenance we can explicitly determine the stability of the system, and form analytical expressions of the boundaries between the different stability regions. As the boundary of $\mathcal{J}_{3}$ is not open to analytical determination in the case with maintenance, we determined numerically (substituting the general growth function with the classical Monod-type growth kinetics) the existence and stability of the system over a range of practical operating conditions (dilution rate and chlorophenol input). For comparison and confirmation, we also performed this for the case without maintenance and found the same regions in both cases, with variations only in their shape and extent. For example, whilst the boundary between $\mathcal{J}_{1}$ and $\mathcal{J}_{4}$ terminates at the origin without maintenance, with maintenance it is located at *F*~1~(0)/*Y*~3~*Y*~4~ ≈ 0.0195. More interestingly, the addition of a decay term results in an extension of the SS3 unstable steady-state, reducing the potential for successful chlorophenol demineralisation at relatively low dilution rates and substrate input concentrations. Additionally, we show that at the boundary between $\mathcal{J}_{3}$ and $\mathcal{J}_{5},$ there is numerical evidence of Hopf bifurcation occurring and that a limit cycle in SS3 emerges. Finally, we gave an example of how the model could be used to probe the system to answer specific questions regarding model parameterisation. Here we have indicated that a switch in dominance between two organisms competing for hydrogen results in the system becoming unstable and a loss in viability. This is perhaps intuitive to microbiologists, but here it has been proven using mathematical analysis and could be used to determine critical limits of the theoretical parameter values in shifting between a stable and unstable system. Whilst parameters are not arbitrary in reality, the potential for microbial engineering or synthetic biology to manipulate the properties of organisms makes this observation all the more pertinent. Appendix A. Numerical methods {#sec0023} ============================= We consider sets of operating parameters (*D* and *S*~ch,\ in~) for each of the three steady-states, and using Matlab, the complex polynomials for each steady-state can be solved by substitution of parameter values (see [Table 1](#tbl0001){ref-type="table"}) into the explicit solution. By investigating the signs of the solutions and the eigenvalues, respectively we determine which steady-states are meaningful and stable. By exploring a localised region of suitable operating parameters, we then generate a phase plot showing where each steady-state is stable, bistable or unstable. Appendix B. Numerical evidence for Hopf Bifurcation {#sec0024} =================================================== In [Section 8](#sec0016){ref-type="sec"}, we show the operating diagrams with the parameters given in [Table 1](#tbl0001){ref-type="table"}, and determine numerically that as the parameter *S*~ch,\ in~ increases at a fixed dilution rate ($D = 0.01\mspace{6mu} d^{- 1}$), the system bifurcates through several stability domains. We claim that as we cross the boundary between regions $\mathcal{J}_{5}$ and $\mathcal{J}_{3},$ we observe a Hopf bifurcation, and, in $\mathcal{J}_{5},$ close to the boundary with $\mathcal{J}_{3},$ a limit cycle appears. In order to test this numerically, we checked the real parts of the six eigenvalues at each point along the transect shown in Fig [9](#fig0009){ref-type="fig"} (10000 points in total), and plotted their values. [Fig B.14](#fig0014){ref-type="fig"} indicates the conditions for a Hopf bifurcation are satisfied as eigenvalues 2 and 3 both change their sign when passing through the coordinate (0, 0.1034) and the real part of all eigenvalues 1, 4 and 6 remain negative.Fig. B.14Real parts of the eigenvalues determined at $D = 0.01$ and $S_{\text{ch},\text{in}} = \left\lbrack 0.08,0.12 \right\rbrack,$ in the case with maintenance. The red vertical lines indicate the location where the eigenvalue crosses zero. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)Fig. B.14 Appendix C. Proofs {#sec0025} ================== In this Section we give the proofs of the results. C1. General case {#sec0026} ---------------- In these proofs, we do not assume that the growth function are of Monod type ([Eq. 23](#eq0023){ref-type="disp-formula"}). We only assume that the growth functions satisfy **H1**--**H8**. ### Proof of [Lemma 1](#enun0002){ref-type="statement"} {#sec0030} Let *s*~2~ ≥ 0 be fixed. By **H4**, the function $$\left. s_{0} \in \left\lbrack 0, + \infty \right)\mapsto\mu_{0}\left( s_{0},s_{2} \right) \in \left\lbrack 0,\mu_{0}\left( + \infty,s_{2} \right) \right) \right.$$is monotonically increasing. Hence, it has an inverse function denoted by: $$\left. y \in \left\lbrack 0,\mu_{0}\left( + \infty,s_{2} \right) \right)\mapsto M_{0}\left( y,s_{2} \right) \in \left\lbrack 0, + \infty \right), \right.$$such that for all *s*~0~ ≥ 0, *s*~2~ ≥ 0 and $y \in \left\lbrack 0,\mu_{0}\left( + \infty,s_{2} \right) \right)$ ([Eq. 25](#eq0025){ref-type="disp-formula"}) holds. The proofs for the existence of *M*~1~ and *M*~2~ are similar and are left to the reader. C2. Existence of steady-states {#sec0027} ------------------------------ ### Proof of [Lemma 2](#enun0003){ref-type="statement"} {#sec0031} Assume first that $x_{0} = 0$. Then, as a consequence of ([Eq. 38](#eq0038){ref-type="disp-formula"}), $s_{0} = s_{0}^{\text{in}}$ and, as a consequence of ([Eq. 39](#eq0039){ref-type="disp-formula"}), we have: $$Ds_{1} + \mu_{1}\left( s_{1},s_{2} \right)x_{1} = 0$$which implies $s_{1} = 0$ and $\mu_{1}\left( s_{1},s_{2} \right)x_{1} = 0$. Therefore, as a consequence of ([Eq. 36](#eq0036){ref-type="disp-formula"}) we have $x_{1} = 0$. Replacing $x_{0} = 0$ and $x_{1} = 0$ in ([Eq. 40](#eq0040){ref-type="disp-formula"}), we have: $$Ds_{2} + \mu_{2}\left( s_{2} \right)x_{2} = 0$$which implies $s_{2} = 0$ and $\mu_{2}\left( s_{2} \right)x_{2} = 0$. Therefore, as a consequence of ([Eq. 37](#eq0037){ref-type="disp-formula"}), $x_{2} = 0$. Hence, the steady-state is SS1. Assume now that $x_{1} = 0$. Then, as a consequence of ([Eq. 40](#eq0040){ref-type="disp-formula"}), we have: $$Ds_{2} + \omega\mu_{0}\left( s_{0},s_{2} \right)x_{0} + \mu_{2}\left( s_{2} \right)x_{2} = 0$$which implies $s_{2} = 0,$$\mu_{0}\left( s_{0},s_{2} \right)x_{0} = 0$ and $\mu_{2}\left( s_{2} \right)x_{2} = 0$. Therefore, as a consequence of ([Eq. 35](#eq0035){ref-type="disp-formula"}), $x_{0} = 0$. As shown previously this implies that the steady-state is SS1. Evaluated at SS1 the Jacobian matrix of ([(14)](#eq0014){ref-type="disp-formula"}, [(15)](#eq0015){ref-type="disp-formula"}, [(16)](#eq0016){ref-type="disp-formula"}, [(17)](#eq0017){ref-type="disp-formula"}, [(18)](#eq0018){ref-type="disp-formula"}, [(19)](#eq0019){ref-type="disp-formula"}) is: $$\begin{bmatrix} {- D - a_{0}} & 0 & 0 & 0 & 0 & 0 \\ 0 & {- D - a_{1}} & 0 & 0 & 0 & 0 \\ 0 & 0 & {- D - a_{2}} & 0 & 0 & 0 \\ 0 & 0 & 0 & {- D} & 0 & 0 \\ 0 & 0 & 0 & 0 & {- D} & 0 \\ 0 & 0 & 0 & 0 & 0 & {- D} \\ \end{bmatrix}$$Thus, SS1 is stable. ### Proof of [Lemma 3](#enun0004){ref-type="statement"} {#sec0032} Since *x*~0~ \> 0 and *x*~1~ \> 0, then, as a consequence of ([Eqs. 35](#eq0035){ref-type="disp-formula"}, [36](#eq0036){ref-type="disp-formula"}), we have: $$\mu_{0}\left( s_{0},s_{2} \right) = D + a_{0},\quad\mu_{1}\left( s_{1},s_{2} \right) = D + a_{1}$$Hence, using ([Eqs. 25](#eq0025){ref-type="disp-formula"}, [26](#eq0026){ref-type="disp-formula"}), we have: $$s_{0} = M_{0}\left( D + a_{0},s_{2} \right),\quad s_{1} = M_{1}\left( D + a_{1},s_{2} \right)$$Using ([Eqs. 38](#eq0038){ref-type="disp-formula"}, [39](#eq0039){ref-type="disp-formula"}), we have ([Eq. 42](#eq0042){ref-type="disp-formula"}). Using ([Eq. 40](#eq0040){ref-type="disp-formula"}) we obtain: $$- s_{2} + \left( s_{0}^{\text{in}} - s_{0} - s_{1} \right) - \omega\left( s_{0}^{\text{in}} - s_{0} \right) = 0$$If *ω* ≥ 1 this equation has no solution. If *ω* \< 1 this equation is equivalent to: $$s_{0}^{\text{in}} = s_{0} + \frac{s_{1} + s_{2}}{1 - \omega}.$$Using ([Eq. C.1](#eq0055){ref-type="disp-formula"}) we see that *s*~2~ must be a solution of ([Eq. 41](#eq0041){ref-type="disp-formula"}). Since *s*~1~ \> 0 and *s*~2~ \> 0 then, from ([Eq. C.2](#eq0056){ref-type="disp-formula"}) we have necessarily: $$s_{1} + s_{2} = \left( 1 - \omega \right)\left( s_{0}^{\text{in}} - s_{0} \right) > 0$$so that $s_{0}^{\text{in}} - s_{0} > 0$. From ([Eq. 42](#eq0042){ref-type="disp-formula"}) we deduce that *x*~0~ \> 0. Since $s_{0}^{\text{in}} - s_{0} > 0$ and *s*~2~ \> 0 then, from ([Eq. C.2](#eq0056){ref-type="disp-formula"}) we have necessarily: $$\omega\left( s_{0}^{\text{in}} - s_{0} \right) + s_{2} = s_{0}^{\text{in}} - s_{0} - s_{1} > 0$$so that $s_{0}^{\text{in}} - s_{0} - s_{1} > 0$ From ([Eq. 42](#eq0042){ref-type="disp-formula"}) we deduce that *x*~1~ \> 0. ### Proof of [Lemma 4](#enun0005){ref-type="statement"} {#sec0033} Since *x*~0~ \> 0, *x*~1~ \> 0 and *x*~2~ \> 0, then, as a consequence of ([(35)](#eq0035){ref-type="disp-formula"}, [(36)](#eq0036){ref-type="disp-formula"}, [(37)](#eq0037){ref-type="disp-formula"}), we obtain: $$\mu_{0}\left( s_{0},s_{2} \right) = D + a_{0},\quad\mu_{1}\left( s_{1},s_{2} \right) = D + a_{1},\quad\mu_{2}\left( s_{2} \right) = D + a_{2}$$Hence, using ([(25)](#eq0025){ref-type="disp-formula"}, [(26)](#eq0026){ref-type="disp-formula"}, [(27)](#eq0027){ref-type="disp-formula"}), *s*~0~, *s*~1~ and *s*~2~ are given by ([Eq. 43](#eq0043){ref-type="disp-formula"}). Using ([(38)](#eq0038){ref-type="disp-formula"}, [(39)](#eq0039){ref-type="disp-formula"}, [(40)](#eq0040){ref-type="disp-formula"}) we obtain ([Eq. 44](#eq0044){ref-type="disp-formula"}). For *x*~2~ to be positive it is necessary that *s*~0~, *s*~1~ and *s*~2~ satisfy the condition: $$\left( 1 - \omega \right)\left( s_{0}^{\text{in}} - s_{0} \right) > s_{1} + s_{2},$$If *ω* ≥ 1 this equation has no solution. If *ω* \< 1 this equation is equivalent to the condition: $$s_{0}^{\text{in}} > s_{0} + \frac{s_{1} + s_{2}}{1 - \omega}.$$Using ([Eq. 43](#eq0043){ref-type="disp-formula"}), this condition is the same as $$s_{0}^{\text{in}} > \psi\left( M_{2}\left( D + a_{2} \right) \right) = F_{2}\left( D \right)$$Therefore, from ([Eq. C.3](#eq0057){ref-type="disp-formula"}) we have $s_{0}^{\text{in}} - s_{0} > 0$ and $s_{0}^{\text{in}} - s_{0} - s_{1} > 0,$ so that *x*~0~ \> 0 and *x*~1~ \> 0. C3. Stability of steady-states {#sec0028} ------------------------------ We use the change of variables: $$z_{0} = s_{0} + x_{0},\quad z_{1} = s_{1} + x_{1} - x_{0},\quad z_{2} = s_{2} + x_{2} + \omega x_{0} - x_{1}$$Therefore, ([Eqs. 14](#eq0014){ref-type="disp-formula"}--[19](#eq0019){ref-type="disp-formula"}), with $a_{0} = a_{1} = a_{2} = 0,$ become: $$\begin{aligned} \frac{dx_{0}}{dt} & {= - Dx_{0} + \mu_{0}\left( z_{0} - x_{0},z_{2} - \omega x_{0} + x_{1} - x_{2} \right)x_{0}} \\ \end{aligned}$$$$\begin{aligned} \frac{dx_{1}}{dt} & {= - Dx_{1} + \mu_{1}\left( z_{1} + x_{0} - x_{1},z_{2} - \omega x_{0} + x_{1} - x_{2} \right)x_{1}} \\ \end{aligned}$$$$\begin{aligned} \frac{dx_{2}}{dt} & {= - Dx_{2} + \mu_{2}\left( z_{2} - \omega x_{0} + x_{1} - x_{2} \right)x_{2}} \\ \end{aligned}$$$$\begin{aligned} \frac{dz_{0}}{dt} & {= D\left( s_{0}^{\text{in}} - z_{0} \right)} \\ \end{aligned}$$$$\begin{aligned} \frac{dz_{1}}{dt} & {= - Dz_{1}} \\ \end{aligned}$$$$\begin{aligned} \frac{dz_{2}}{dt} & {= - Dz_{2}} \\ \end{aligned}$$In the variables (*x*~0~, *x*~1~, *x*~2~, *z*~0~, *z*~1~, *z*~2~) where *z*~0~, *z*~1~ and *z*~2~ are defined by ([Eq. C.4](#eq0058){ref-type="disp-formula"}), the steady-states SS1, SS2 and SS3 are given by: 1.${\text{SS}1} = \left( 0,0,0,s_{0}^{\text{in}},0,0 \right)$2.${\text{SS}2} = \left( x_{0},x_{1},0,s_{0}^{\text{in}},0,0 \right),$ where *x*~0~ and *x*~1~ are defined by ([Eq. 47](#eq0047){ref-type="disp-formula"}).3.${\text{SS}3} = \left( x_{0},x_{1},x_{2},s_{0}^{\text{in}},0,0 \right),$ where *x*~0~, *x*~1~ and *x*~2~ are defined by ([Eq. 48](#eq0048){ref-type="disp-formula"}). Let $\left( x_{0},x_{1},x_{2},s_{0}^{\text{in}},0,0 \right)$ be a steady-state. The Jacobian matrix of ([Eqs. C.5](#eq0059){ref-type="disp-formula"}--[C.10](#eq0064){ref-type="disp-formula"}) has the block triangular form: $$\mathbf{J} = \begin{bmatrix} \mathbf{J}_{1} & \mathbf{J}_{2} \\ 0 & \mathbf{J}_{3} \\ \end{bmatrix}$$where $$\mathbf{J}_{1} = \left\lbrack \begin{array}{rrr} {\mu_{0} - D - \left( E + \omega F \right)x_{0}} & {Fx_{0}} & {- Fx_{0}} \\ {\left( G + \omega H \right)x_{1}} & {\mu_{1} - D - \left( G + H \right)x_{1}} & {Hx_{1}} \\ {- \omega Ix_{2}} & {Ix_{2}} & {\mu_{2} - D - Ix_{2}} \\ \end{array} \right\rbrack$$$$\mathbf{J}_{2} = \begin{bmatrix} {Ex_{0}} & 0 & {Fx_{0}} \\ 0 & {Gx_{1}} & {- Hx_{1}} \\ 0 & 0 & {Ix_{2}} \\ \end{bmatrix},\quad\mathbf{J}_{3} = \begin{bmatrix} {- D} & 0 & 0 \\ 0 & {- D} & 0 \\ 0 & 0 & {- D} \\ \end{bmatrix}$$and $$E = \frac{\partial\mu_{0}}{\partial s_{0}},\quad F = \frac{\partial\mu_{0}}{\partial s_{2}},\quad G = \frac{\partial\mu_{1}}{\partial s_{1}},\quad H = - \frac{\partial\mu_{1}}{\partial s_{2}},\quad I = \frac{d\mu_{2}}{ds_{2}}$$are evaluated at the steady-state. Since **J** is a block triangular matrix, its eigenvalues are $- D$ (with multiplicity 3) together with the eigenvalues of the 3 × 3 upper-left matrix **J~1~**. Note that we have used the opposite sign for the partial derivative $H = - \partial\mu_{1}/\partial s_{2},$ so that all constants involved in the computation become positive, which will simplify the analysis of the characteristic polynomial of **J~1~**. ### Proof of [Proposition 2](#enun0008){ref-type="statement"} {#sec0034} Evaluated at SS2, the matrix **J~1~** is: $$\mathbf{J}_{1} = \begin{bmatrix} {- \left( E + \omega F \right)x_{0}} & {Fx_{0}} & {- Fx_{0}} \\ {\left( G + \omega H \right)x_{1}} & {- \left( G + H \right)x_{1}} & {Hx_{1}} \\ 0 & 0 & {\mu_{2} - D} \\ \end{bmatrix}$$Since **J~1~** is a block triangular matrix, its eigenvalues are simply $\mu_{2} - D,$ together with the eigenvalues of the 2 × 2 upper-left matrix. Note that the trace of this 2 × 2 matrix is negative. Hence, its eigenvalues are of negative real part if, and only if, its determinant is positive, that is if, and only if, $$\begin{array}{r} {E\left( G + H \right) - \left( 1 - \omega \right)FG > 0} \\ \end{array}$$Using $$\frac{\partial M_{0}}{\partial s_{2}} = - \frac{\partial\mu_{0}}{\partial s_{2}}\left\lbrack \frac{\partial\mu_{0}}{\partial s_{0}} \right\rbrack^{- 1} = - F/E$$$$\frac{\partial M_{1}}{\partial s_{2}} = - \frac{\partial\mu_{1}}{\partial s_{2}}\left\lbrack \frac{\partial\mu_{1}}{\partial s_{1}} \right\rbrack^{- 1} = H/G$$we deduce from $$\psi\left( s_{2} \right) = M_{0}\left( D,s_{2} \right) + \frac{M_{1}\left( D,s_{2} \right) + s_{2}}{1 - \omega}$$that $$\frac{\partial\psi}{\partial s_{2}} = \frac{\partial M_{0}}{\partial s_{2}} + \frac{\frac{\partial M_{1}}{\partial s_{2}} + 1}{1 - \omega} = - \frac{F}{E} + \frac{\frac{H}{G} + 1}{1 - \omega}$$Hence, $$\frac{\partial\psi}{\partial s_{2}} = \frac{E\left( G + H \right) - \left( 1 - \omega \right)FG}{\left( 1 - \omega \right)EG}$$Therefore, the condition of stability, ([Eq. C.11](#eq0065){ref-type="disp-formula"}), is equivalent to $\frac{\partial\psi}{\partial s_{2}} > 0$. Hence, we have proven that SS2 is stable if, and only if, *μ*~2~(*s*~2~) \< *D* and $\frac{\partial\psi}{\partial s_{2}} > 0$. ### Proof of [Proposition 3](#enun0009){ref-type="statement"} {#sec0035} Evaluated at SS3, the matrix **J~1~** is: $$\mathbf{J}_{1} = \begin{bmatrix} {- \left( E + \omega F \right)x_{0}} & {Fx_{0}} & {- Fx_{0}} \\ {\left( G + \omega H \right)x_{1}} & {- \left( G + H \right)x_{1}} & {Hx_{1}} \\ {- \omega Ix_{2}} & {Ix_{2}} & {- Ix_{2}} \\ \end{bmatrix}$$The characteristic polynomial is given by: $$\begin{array}{r} {\lambda^{3} + f_{2}\lambda^{2} + f_{1}\lambda + f_{0} = 0} \\ \end{array}$$where $$\begin{aligned} f_{2} & {= Ix_{2} + \left( G + H \right)x_{1} + \left( E + \omega F \right)x_{0}} \\ \end{aligned}$$$$\begin{aligned} f_{1} & {= \Delta x_{0}x_{1} + EIx_{0}x_{2} + GIx_{1}x_{2}} \\ \end{aligned}$$$$\begin{aligned} f_{0} & {= EGIx_{0}x_{1}x_{2}} \\ \end{aligned}$$and $\Delta = E\left( G + H \right) - \left( 1 - \omega \right)FG$. To satisfy the Routh-Hurwitz criteria, we require *f~i~* \> 0, for $i = 0,1,2$ and $f_{1}f_{2} - f_{0} > 0$. Notice that: $$f_{1}f_{2} - f_{0} = \left( EIx_{0}x_{2} + \Delta x_{0}x_{1} \right)f_{2} + \left( Ix_{2} + \left( G + H \right)x_{1} + \omega Fx_{0} \right)GIx_{1}x_{2}$$We always have *f*~0~ \> 0 and *f*~2~ \> 0. From ([Eq. C.12](#eq0066){ref-type="disp-formula"}) we deduce that $\Delta = \left( 1 - \omega \right)EG\frac{\partial\psi}{\partial s_{2}}$ . Therefore, if *F*~3~(*D*) ≥ 0, viz. $\frac{\partial\psi}{\partial s_{2}} \geq 0,$ then *Δ* ≥ 0. Hence, *f*~1~ \> 0 and $f_{1}f_{2} - f_{0} > 0,$ so that SS3 is stable. On the other hand, if $\frac{\partial\psi}{\partial s_{2}} < 0$ and *x*~2~ is very small, which occurs when SS3 is very close to SS2^♭^, then *f*~2~ has the sign of *Δ* since the term with *x*~2~ is negligible compared to the term *Δx*~0~*x*~1~: $$f_{1} = \Delta x_{0}x_{1} + \left( EIx_{0} + GIx_{1} \right)x_{2} < 0$$Thus, SS3 is unstable. ### Proof of [Proposition 4](#enun0010){ref-type="statement"} {#sec0036} Since we always have *f*~0~ \> 0 and *f*~2~ \> 0, from the previous proof it follows that SS3 is stable if, and only if, $f_{1}f_{2} - f_{0} > 0$. Indeed, this condition implies that we have also *f*~1~ \> *f*~0~/*f*~2~ \> 0. Using ([Eq. 49](#eq0049){ref-type="disp-formula"}) and ([Eq. C.17](#eq0071){ref-type="disp-formula"}), we see that: $$f_{1}f_{2} - f_{0} = F_{4}\left( D,s_{0}^{\text{in}} \right)$$Therefore, the condition $f_{1}f_{2} - f_{0} > 0$ is equivalent to $F_{4}\left( D,s_{0}^{\text{in}} \right) > 0$. C4. Operating diagrams {#sec0029} ---------------------- ### Proof of [Proposition 5](#enun0011){ref-type="statement"} {#sec0037} We know that SS1 always exists and is stable. We know that SS2^♭^ is unstable if it exists. The curves *Γ*~1~ and *Γ*~2~ of the operating diagram, given by ([Eq. 45](#eq0045){ref-type="disp-formula"}) and ([Eq. 46](#eq0046){ref-type="disp-formula"}), respectively are defined for *D* \< *D*~1~ and *D* \< *D*~2~, respectively and they are tangent for $D = D_{3},$ where $D_{1} = 0.452,$$D_{2} = 0.393$ and $D_{3} = 0.078$. Therefore *F*~3~(*D*) \< 0 for *D* \< *D*~3~ and *F*~3~(*D*) \> 0 for *D* \> *D*~3~. Using [Table 7](#tbl0007){ref-type="table"} and [Remark 2](#enun0007){ref-type="statement"}, we obtain the following results: •$\mathcal{J}_{1}$ is defined by *D* ≥ *D*~1~ or 0 \< *D* \< *D*~1~ and *S*~ch,\ in~ \< *F*~1~(*D*)/*Y*~3~*Y*~4~. Therefore, SS1 is the only existing steady-state in this region.•$\mathcal{J}_{2}$ is defined by *D*~3~ \< *D* \< *D*~1~ and *F*~1~(*D*)/*Y*~3~*Y*~4~ \< *S*~ch,\ in~ \< *F*~2~(*D*)/*Y*~3~*Y*~4~. Therefore, both steady-state SS2 exist and SS2^♯^ is stable since *F*~3~(*D*) \> 0.•$\mathcal{J}_{3}$ is defined by 0 \< *D* \< *D*~2~ and *F*~2~(*D*)/*Y*~3~*Y*~4~ \< *S*~ch,\ in~, and *F*~4~(*D, Y*~3~*Y*~4~*S*~ch,\ in~) \> 0 when 0 \< *D* \< *D*~3~. Therefore, SS3 exists and is stable, both steady-state SS2 exist and SS2^♯^ is unstable since *F*~3~(*D*) \< 0.•$\mathcal{J}_{4}$ is defined by 0 \< *D* \< *D*~3~ and *F*~1~(*D*)/*Y*~3~*Y*~4~ \< *S*~ch,\ in~ \< *F*~2~(*D*)/*Y*~3~*Y*~4~. Therefore, both steady-state SS2 exist and SS2^♯^ is unstable since *F*~3~(*D*) \< 0.•$\mathcal{J}_{5}$ is defined by 0 \< *D* \< *D*~3~, *F*~2~(*D*)/*Y*~3~*Y*~4~ \< *S*~ch,\ in~, and *F*~4~(*D, Y*~3~*Y*~4~*S*~ch,\ in~) \< 0. Therefore, SS3 exists and is unstable and both steady-state SS2 exist and SS2^♯^ is unstable since *F*~3~(*D*) \< 0. ### Proof of [Propositions 6](#enun0012){ref-type="statement"}--[8](#enun0014){ref-type="statement"} {#sec0038} The result follows from [Table 7](#tbl0007){ref-type="table"} and [Remark 2](#enun0007){ref-type="statement"}. The details are as in the proof of [Proposition 5](#enun0011){ref-type="statement"}. Matthew J. Wade was funded by the Biotechnology and Biological Sciences Research Council UK (BB/K003240/2 Engineering synthetic microbial communities for biomethane production) and by the Institute of Sustainability, Newcastle University. The authors would like to thank the anonymous reviewers for their valuable and detailed comments.

1. Introduction {#sec1} =============== Presbycusis or age-related hearing loss is a frequent, degenerative neurosensory disorder in ageing societies. Presbycusis affects more than a third of the elderly population at retirement age \[[@bib1]\]. Hearing loss in this vulnerable population substantially contributes to cognitive decline, depression and social isolation \[[@bib2],[@bib3]\] and represents an important socio-economic burden. Hearing aids and cochlear implants alleviate symptoms, but are not a causal treatment. Finding new treatments to reverse or slow down the progression of age-related hearing loss will have major consequences for the affected individuals and for society as a whole. The etiology of presbycusis is multifactorial and involves a complex interaction of genes. It can be triggered by a variety of external and internal factors such as noise exposure, ototoxic molecules and medical conditions, among others \[[@bib4]\]. The histopathological correlates of age-related hearing loss in the inner ear are the loss of mechanosensitive hair cells, spiral ganglion neurons, supporting cells and cells in the stria vascularis with a gradient from base to apex \[[@bib5]\]. Therefore, the typical age-related hearing loss is predominantly affecting the higher frequencies, which are coded at the cochlear base. Oxidative stress, i.e. high levels of reactive oxygen species (ROS), has traditionally been thought to lead to tissue damage and eventually cell death through non-specific protein, lipid and DNA oxidation. However more recent concepts rather favor a role of physiological ROS levels in cellular signaling \[[@bib6]\], and a dysregulation of signaling through excessive ROS levels \[[@bib7]\]. Abundant evidence demonstrates changes in the cochlear redox environment with age \[[@bib8]\]. For instance, several antioxidant genes, including *SOD2*, *GST* or *UCP*, have been associated to age-related hearing loss \[[@bib9]\] and genetic mouse models deficient for antioxidant genes, including *Nrf2* \[[@bib10],[@bib11]\], *Sod1* \[[@bib12]\] or genes regulating the mTOR pathway \[[@bib13]\], show accelerated age-related hearing loss. Furthermore, age-related hearing loss (ARHL) has been shown to be slowed by supplementation with antioxidants in laboratory animals, and a few studies have investigated the effect of antioxidants supplements against ARHL in humans \[[@bib14]\]. While a role of excessive oxidant generation has been widely described as causative of hearing loss, the exact sources of oxidants are unclear. NADPH oxidases (NOX) are a family of enzymes whose main biochemical function is the production of ROS -- namely superoxide radical anion O~2~^•-^ and H~2~O~2~. In mammals, the NOX family consists of seven isoforms (NOX1-5, DUOX1, 2). NOX have several subunits and the p22^phox^ subunit is crucial for the function of several NOX isoforms, namely NOX1 to NOX4. In fact, p22^phox^ is essential for NOX stabilisation and activity: the *CYBA* CRISPR knockout is devoid of ROS generation in NOX1, NOX2, NOX3 and NOX4 expressing cells \[[@bib15]\]. Therefore, p22^phox^ is a master regulator of ROS generation as it regulates most NOX-derived oxidants. The biological function of NOX-derived ROS is broad, from host defense, to cellular signaling and hormone biosynthesis. NOX-derived ROS and in particular H~2~O~2~, resulting from superoxide dismutation are important second messengers in cell signaling. Through a reversible reaction with H~2~O~2~, cysteine residues are oxidized, changing the function of the respective protein \[[@bib16]\]. Thereby, ROS regulates the activity of protein tyrosine phosphatases \[[@bib17]\], permeability of ion channels \[[@bib18]\] or affinity of transcription factors for their target DNA sequence \[[@bib19]\] and thus regulate important physiological function in the cell (i.e. proliferation, differentiation, survival metabolism or motility). While NOX have important physiological functions in virtually all organ systems, an over-activation of these enzyme systems can lead to oxidative stress through overproduction of ROS and ultimately to oxidative stress-driven disease (e.g. fibrosis, cardiovascular disease, neurodegeneration). We hypothesize that in the inner ear, NOX activation results in hearing loss. NOX3 is highly and exclusively expressed in the inner ear \[[@bib20],[@bib21]\], and its function has been mainly attributed to otoconia formation in the developing vestibular system, whereas its function remains unknown in the cochlear tissues \[[@bib20],[@bib22]\]. Nox3 mutant mice develop a vestibular deficiency, leading to a head-tilt phenotype. Mice with a loss of function mutation of the p22^phox^ subunit show a similar vestibular phenotype as NOX3 mutant mice with, in addition, a defect of innate immunity due to the absence of NOX2 \[[@bib23]\]. Although growing evidence suggests a primary role of NOX and in particular NOX3 in different cochlear pathologies \[[@bib24]\]; the contribution of NOX isoforms in age-related hearing loss has never been investigated. The fact that NOX3 is exclusively expressed in the inner ear makes it a prime target for interventions aiming at slowing down ROS production and consecutive cell-damage in the inner ear. While NOX3 is only found in the inner ear, more broadly expressed NOX family members are also present in the inner ear and may contribute to hearing and balance disorders. Indeed, NOX2 is highly expressed in microglia cells, which are abundant in spiral ganglia \[[@bib25]\], and inhibition of microglia activation was protective in a mouse model of neomycin-induced hearing loss \[[@bib26]\]. NOX4 is strongly expressed in vascular endothelium and could play a role in the stria vascularis. And indeed, NOX4 overexpressing transgenic mice show an increased sensitivity to noise-induced hearing loss \[[@bib27]\]. Given the central role of p22^phox^ in NOX activity, p22^phox^ deficient mice are a particularly useful tool for studying the role of NOX in inner ear \[[@bib23]\]. Indeed, considering the absence of Nox5 in the mouse, p22^phox^ deficient mice can be defined as pan-Nox-deficient mouse model \[[@bib28]\]. The nmf333 mouse model (A.B6 Tyr + nmf333 -- jax; A/J genetic background) harbors a missense mutation in *CYBA* leading to a functional inactivation of NOX1-4 \[[@bib29]\]. The A/J strain shows a very early age-related hearing loss and is therefore an interesting model to study oxidative-related pathologies of the inner ear and to evaluate the effects of potential otoprotective drugs \[[@bib30]\]. In the present study, the A/J mouse model (nmf333) was used to investigate the impact of NOX in the progression of age-related hearing loss. Our data demonstrate that p22^phox^ is a key regulator of age-dependent hearing loss. Mechanistically, we show that p22^phox^ regulates the calcium release from intracellular stores in auditory neurons and that deletion of p22^phox^ protects from excitotoxic insult in auditory neurons. Since, p22^phox^ is a key regulator of redox pathways due to its main function as controlling NOX activity, our data suggest a role of oxidant-derived NOX in age-related auditory neuropathy and show promises for novel therapeutic approaches for future treatment of age-related hearing pathologies. 2. Results {#sec2} ========== **A/J mouse as model for age-related hearing loss.** In order to characterize the early onset hearing loss in A/J mice, we investigated hearing thresholds starting over an age range from 4 to 26 weeks ([Fig. 1](#fig1){ref-type="fig"}A and B). Click-evoked auditory brain stem response (ABR) hearing thresholds were close to 45 dB SPL in young adult mice (4 weeks old) and progressed up to 75 dB SPL in old mice ([Fig. 1](#fig1){ref-type="fig"}A). The pure tone frequencies measurement (2--32 kHz) was more informative: while young mice audiogram was comprised between 2 and 32 kHz (with a threshold of 30 dB SPL at 11.3 kHz) ([Fig. 1](#fig1){ref-type="fig"}A; empty squares) we observed a progressive loss of hearing at high frequencies narrowing the hearing spectrum (from 32  kHz at 4 weeks to 5.7  kHz at 26 weeks old). Accordingly, maximal amplitude of the ABR decreased with age ([Fig. 1](#fig1){ref-type="fig"}B). The most important hearing loss was observed at around 11.3--16 kHz ([Fig. 1](#fig1){ref-type="fig"}A). Consistent with the high frequency hearing loss phenotype, we observed a progressive degeneration of the sensory epithelium from the base to the apical turn ([Supplementary Fig. 1](#appsec1){ref-type="sec"}). While the apical turn was relatively well conserved in old mice, there was an important cellular degeneration in the basal region ([Fig. 1](#fig1){ref-type="fig"}C--F). This was accompanied by a dramatic decrease in the number of synaptic ribbons per inner hair cell (IHC) ([Fig. 1](#fig1){ref-type="fig"}G--J) and a marked decrease in the neuronal density in the spiral ganglion with age ([Fig. 1](#fig1){ref-type="fig"}K-N). Interestingly, the kinetics of degeneration, as observed in the basal turn of the cochlea, showed synaptic and post synaptic morphological alteration (decrease in number of ribbons per IHC, auditory neuron density and in the ABR wave I) occurring prior to the sensory epithelium degeneration ([Fig. 1](#fig1){ref-type="fig"}O; the dotted line represent control values obtained with mice at 4 weeks of age). Therefore, we conclude that A/J mice have an early onset age-related hearing loss which is mainly caused by neural degeneration and subsequent loss of the sensory epithelium.Fig. 1**Age-related hearing loss in wild type A/J mice: functional and morphological alterations in auditory neurons precede the loss of sensory epithelium.** A) Auditory brainstem response hearing thresholds of A/J mice were assessed from 4 to 26 weeks old animals, on a group of 11 WT A/J mice following clicks and pure tone frequencies stimulation. B) Elevation of hearing threshold is accompanied with a decrease of the ABR maximal amplitude (peak to peak amplitude, (p--p)), assessed at 75 dB SPL. C--F) Basal turn representative sensory epithelium immunostainings (Myo7a in green) of the WT A/J cochlea at 4 weeks old (C), 6 weeks old (D), 18 weeks old (E) and 26 weeks old (F). G-J) On the same samples, the number of synaptic ribbons between inner hair cells (nucleus in white) and spiral ganglion neurons was also determined using CtBP2 immunostaining (red dots), using 4 weeks old (G), 6 weeks old (H), 18 weeks old (I) and 26 weeks old (J) mice. K--N) Representative mid-modiolar hematoxilin-eosin staining of A/J mouse basal cochlear turns at 4 weeks old (K), 6 weeks old (L), 18 weeks old (M) and 26 weeks old (N). O) Bar graph showing the quantification of outer hair cells (green), synaptic ribbons (red), SGN density (pink), ABR wave I amplitude (light grey) and ABR threshold (black) in the 6 and 18 weeks old mice basal turn, relatively to the 4 weeks old group that serve as reference (100%). n = 11 animals. Scale bar (C--F) = 25 μm, (G--J) = 10 μm and (K--N) = 50 μm. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)Fig. 1 **Pattern of NOX expression in the cochlea.** While NOX3 is known to be highly enriched in the inner ear, the presence of other NOX isoforms was not studied \[[@bib24],[@bib31]\]. Therefore, we addressed the expression of NOX isoforms in the cochlea of the human fetus ([Fig. 2](#fig2){ref-type="fig"}A) and the adult A/J mouse ([Fig. 2](#fig2){ref-type="fig"}B--K). By qPCR, we found relatively high levels of *NOX2*, *NOX3* and *NOX4* mRNA (CT values below 30) in both human ([Fig. 2](#fig2){ref-type="fig"}A) and mouse cochlea ([Fig. 2](#fig2){ref-type="fig"}B). In both species, *NOX1* and *DUOX1/2* levels in the cochlea were below detection threshold. *NOX5* levels in the human cochlea were low (Ct values 33--35). Note that *NOX5* gene is absent in mice and rats \[[@bib28]\]. The fetal cochlear tissues used for real time qPCR is heterogenous and contained spiral ganglion neurons, as well as parts of the stria vascularis. We therefore aimed to obtain more precise information about the localization of the *NOX* mRNA. There are major issues with antibodies for the different NOX isoforms \[[@bib32]\]. We therefore rather performed in situ hybridization (using the RNAscope technology) in order to localize the mRNA of the different isoforms ([Fig. 2](#fig2){ref-type="fig"}C--K). We compared expression of *Nox2*, *Nox3* and *Nox4* in three different subparts of the cochlea, namely the organ of Corti, the stria vascularis and the spiral ganglia. Signals with the *Nox2* and *Nox4* probes ([Supplementary Fig. 2](#appsec1){ref-type="sec"}) were detected throughout the cochlea without a strong signal in specific parts. In contrast, the NOX3 probe gave a very strong signal in the spiral ganglia ([Fig. 2](#fig2){ref-type="fig"}H), comparable or even stronger than the control gene (*Ppib*, [Fig. 2](#fig2){ref-type="fig"}F; [Supplementary Fig. 3](#appsec1){ref-type="sec"}). The *Nox3* signal also appeared above background in the stria vascularis ([Fig. 2](#fig2){ref-type="fig"}K), but not in hair cells ([Fig. 2](#fig2){ref-type="fig"}E).Fig. 2**Pattern of NOX mRNA expression in the mouse and human cochlea.** Expression level of Nox isoforms assessed by qPCR in human fetal cochlea (A) and 6 weeks old A/J mouse cochlea (B). Each graph shows the Ct values as observed for *NOX1* to *NOX5* and *CYBA* (p22^phox^). To note the absence of *Nox5* in mice. Level of expression of *NOX1* and *DUOX1/2* (not shown) was below the detection threshold. *EEF1* was used as housekeeping gene. n = 3 to 6 cochlea/group. (C--K) mRNA expression level of *Nox3* in the organ of Corti (OC) (E), the spiral ganglia (SGN) (H) and the stria vascularis (SV) (K) of a 6 weeks old WT A/J mouse as assessed by in situ hybridization (RNAscope®) (red dots). Peptidylprolyl isomerase B gene (*Ppib*) was used as positive control (C, F, I) and dihydrodipicolinate reductase (*DapB*) expressed in the bacteria E. Coli was used as negative control (D, G, J). Nuclei were stained with DAPI (grey). Pictures are representative from 3 independent experiments. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)Fig. 2 **p22**^**phox**^ **deficient A/J mice are protected from early onset hearing loss**. As several NOX isoforms are present in the inner ear, we next investigated hearing of A/J mice, harboring a loss of function mutation on the NOX1-4 subunit p22^phox^ (nmf333) ([Fig. 3](#fig3){ref-type="fig"}). For comparison, the results of wild type mice, as already described in [Fig. 1](#fig1){ref-type="fig"}, are added. Hearing thresholds of mice were investigated monthly from 4 to 22 weeks old ([Fig. 3](#fig3){ref-type="fig"}A--F). Four weeks old p22^phox^ mutant mice, exhibited an audiogram profile similar to wild type littermates, indicating that there is no basal difference in hearing between p22^phox^ mutant and WT mice ([Fig. 3](#fig3){ref-type="fig"}A). As already described in [Fig. 1](#fig1){ref-type="fig"}, from an age of 6 weeks, high frequencies hearing loss was observed in WT mice ([Fig. 3](#fig3){ref-type="fig"}B--F, [Fig. 3](#fig3){ref-type="fig"}K). In contrast, only minor hearing impairment was observed in mutant mice. In particular the high frequency hearing loss was almost completely prevented in p22^phox^ deficient animals up to 18 weeks old ([Fig. 3](#fig3){ref-type="fig"}K). In fact, the pattern of hearing loss was distinct in p22^phox^ deficient mice. At 22 weeks old, p22^phox^ mutant showed only a minor (statistically not significant) hearing loss. This minor hearing loss appeared to be equally distributed along the tonotopic axis ([Fig. 2](#fig2){ref-type="fig"}F). This is in contrast to the WT animal where the amplitude of the ABR recorded at 75 dB SPL consistently showed age-dependent decrease at high frequencies ([Fig. 3](#fig3){ref-type="fig"}H; [Supplementary Fig. 4](#appsec1){ref-type="sec"}) and following click stimuli ([Fig. 3](#fig3){ref-type="fig"}G). In addition to the hearing threshold elevation, WT mice exhibited significant delay in the ABR wave I with age ([Fig. 3](#fig3){ref-type="fig"}I and J). Together, these experiments clearly demonstrate the role of p22^phox^-dependent NOX in age-related hearing loss in A/J mice.Fig. 3**p22**^**phox**^**-deficient mice are protected from early onset hearing loss.** A-F) Audiogram of A/J WT (black triangles) versus p22^phox^ deficient littermates (white squares) as determined by auditory brainstem response following clicks and pure tone frequencies stimulation, from 4 to 22 weeks old mice. A) 4 weeks old, B) 6 weeks old, C) 10 weeks old, D) 14 weeks old, E) 18 weeks old and F) 22 weeks old. G-H) Bar graph showing ABR wave I amplitudes of WT and p22^phox^−/− littermates at the indicated ages following G) click or H) 16 kHz pure tone stimuli at 75 dB SPL. I-J) Bar graph showing ABR wave I latencies of WT and p22^phox^ −/− littermates at the indicated ages following I) click or J) 16 kHz pure tone stimuli at 75 dB SPL. Note that at 16 kHz, the sound stimulus was below hearing thresholds of 22 and 26 weeks old WT (Below threshold). K) Bar graph showing hearing threshold shifts at 18 weeks old. n = 11 WT; 12 p22^phox^+/-; 8 p22^phox^−/−.Fig. 3 **Conserved architecture of the sensory epithelium in p22**^**phox**^ **deficient mice**. Mice were sacrificed at different ages (6, 18 and 26 weeks old) for evaluation of cochlear histomorphology ([Fig. 4](#fig4){ref-type="fig"}, [Fig. 5](#fig5){ref-type="fig"} and [Supplementary Fig. 5](#appsec1){ref-type="sec"}). In young animals (6 weeks old), approximately at the onset of hearing loss, the sensory epithelium was overall well conserved and comparable between WT and p22^phox^ animals ([Fig. 4](#fig4){ref-type="fig"}A--E). However, in contrast to p22^phox^ deficient mice, we observed a significant decrease in the presynaptic ribbons number ([Fig. 4](#fig4){ref-type="fig"}F--J) and auditory neurons density ([Fig. 4](#fig4){ref-type="fig"}K-M) in the basal turn of WT mice cochlea. On the functional level, this was accompanied by significant decrease in the ABR wave I amplitude ([Fig. 4](#fig4){ref-type="fig"}N) which was compensated by increased gain (wave IV amplitude) and modified ABR wave IV/wave I ratio ([Fig. 4](#fig4){ref-type="fig"}O--P). Aged WT mice (18 weeks old; [Supplementary Figs. 5](#appsec1){ref-type="sec"} and 26 weeks old; [Fig. 5](#fig5){ref-type="fig"}) exhibited significant loss of OHC in the basal region of the cochlea, consistent with the high frequency hearing loss ([Supplementary Figs. 5A--E](#appsec1){ref-type="sec"} and [Fig. 5](#fig5){ref-type="fig"}A--D). Remarkably, architecture of the sensory epithelium of p22^phox^ deficient mice was fully preserved. In the apical turn, few missing OHC were observed in both WT and p22^phox^ deficient mice but without significant differences. Accordingly in the basal turn of the cochlea, synaptic ribbons were virtually absent in wild type mice, while they were well conserved in p22^phox^ deficient mice ([Supplementary Figs. 5F--J](#appsec1){ref-type="sec"} and [Fig. 5](#fig5){ref-type="fig"}E--J). No significant differences were observed in the apex. Overall, we observed that cochlear morphological and functional changes observed in A/J mice appeared first at the synaptic/postsynaptic level, anteceding the alteration of the sensory epithelium. These damages were prevented by p22^phox^ (*Cyba*) loss of function.Fig. 4**p22**^**phox**^**deficiency prevents age-related auditory neuropathy.** A-D) Representative sensory epithelium immunostainings (Myo7a in green) of the basal (A--B) and apical turn (C--D) of the A/J WT (A and C) and p22^phox^ deficient (p22^phox^−/−) (B and D) cochleae, at the age of 6 weeks old. E) Bar graph showing the number of outer hair cells (OHC)/section of 100 μm. F--I) On the same samples, the number of synaptic ribbons between inner hair cells and spiral ganglion neurons was determined using CtBP2 immunostainings (red): F) and H) show respectively the WT mice basal and apical part of the cochlea. G) and I) show respectively the p22^phox^ deficient mice basal and apical turns of the cochlea. J) Bar graph showing the number of ribbons/inner hair cell (IHC). K-L) Representative mid-modiolar hematoxylin and eosin staining showing the medial turn of 6 weeks old WT (K) and p22^phox^ −/− littermate (L). M) Bar graph showing the density of spiral ganglion neurons (SGN)/mm^2^ in the different parts of the cochlea of WT and p22^phox^ −/− animals. n = 8 WT; n = 7 p22^phox^ −/−. N--O) ABR wave I (N) and ABR wave IV (O) amplitudes generated following clicks stimulation from 20 to 120 dB SPL (3 dB SPL increments) determined in 6 weeks old animals (WT in black; p22^phox^−/− in grey). P) Graph showing the ratio value of waves IV/I amplitudes as recorded following click stimuli at increasing intensities. n = 11 WT; n = 8 p22^phox^ −/−. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)Fig. 4Fig. 5**p22**^**phox**^**deficiency prevents age-related histological alterations of the cochlea.** A-D) Representative sensory epithelium immunostaining (Myo7a in green) of (A--B) the basal and (C--D) the apical turn of the A/J WT (A and C) and p22^phox^ deficient (p22^phox−/-^) (B and D) cochleae, at the age of 26 weeks old. E-H) On the same samples, the number of synaptic ribbons between inner hair cells and spiral ganglion neurons was determined using CtBP2 immunostaining (red). E) and G) respectively, represent the basal and apical part of the cochlea of WT mice. F) and H) are respectively, the basal and apical turns of p22^phox^ deficient littermates. I-J) Bar graph showing the number of ribbons/inner hair cell (IHC). n = 3 WT; n = 3 p22phox −/−. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)Fig. 5 **Gene expression profiling in the cochlea from wild type and p22**^**phox**^ **deficient mice.** To further explore molecular pathways underlying the protective effect of p22^phox^ deletion in age-related hearing loss, we performed cochlear transcriptome analysis of p22^phox^ deficient mice and WT littermates ([Fig. 6](#fig6){ref-type="fig"}). We chose young animals (6 weeks) to identify early transcriptome alterations, anteceding the onset of hearing loss. Our results revealed genes associated with several molecular functions that were statistically significantly downregulated in p22^phox^ mutants ([Fig. 6](#fig6){ref-type="fig"}A). The most redundant molecular function which was lower in *CYBA* mutant mice, is related to Ca^2+^ homeostasis, and in particular ryanodine dependent Ca^2+^ release (FDR\<0.05). Relevant genes include all 3 isoforms of the ryanodine receptors (*Ryr1*, *2*, *3*), involved in Ca^2+^ release from intracellular organelles, *Otoferlin*, *Vamp1* and *Snap25*, involved in Ca^2+^ dependent pre-synaptic vesicle fusion and Slc17a6, involved in glutamate transport ([Fig. 6](#fig6){ref-type="fig"}B). We also identified other genes involved in the auditory synapse function that showed a trend towards lower expression albeit not significant in the RNAseq analysis: the inner hair cells glutamate transporter (*Slc17a8*), glutamate ionotropic receptor AMPA type subunit 2 (Gria2) and Adenosine A1 receptor (*Adora1*) (data not shown). qPCR results confirmed the significantly lower expression of ryanodine receptor 1, 2, 3, otoferlin, *Slc17a6*, *Slc17a8* and *Gria2* ([Fig. 6](#fig6){ref-type="fig"}C). Except for genes associated with ribbons (presynaptic: *Vamp1, Snap25* and *Otof*), all genes were significantly enriched in the SGN (postsynaptic *Ryr1-3 Gria2, Adora1*) ([Supplementary Fig. 6](#appsec1){ref-type="sec"}). Note that *Ryr3* was poorly expressed in both pre and post-synapse (Ct value of 32.6). Interestingly, downregulated genes were all centered on Ca^2+^ and glutamate signaling and might have a major role in the excitatory pathway of auditory neurons, a well-known molecular pathway of hearing loss \[[@bib8]\].Fig. 6**NADPH oxidase-dependent upregulation of genes involved in the glutamatergic pathway.** RNAseq analysis was performed on 6 weeks old WT and p22^phox^ deficient cochleae. n = 3 WT; n = 3 p22^phox^−/−. A) Bar graph shows gene ontologies in the molecular function category that were downregulated in p22^phox^ deficient mice. B) RNAseq data visualized on a volcano plot reveals several downregulated genes involved in synapse and post synapse function, particularly in glutamate transport and signaling (green), Ca2+ dependent fusion of presynaptic vesicle (red) and Ca^2+^ induced Ca^2+^ release (blue). C) Bar graph showing differentially expressed genes as assessed by qPCR and expressed relatively to WT cochleae. The Ct value for the WT cochlea is indicated in white. n = 5 WT; n = 5 p22^phox^ −/− D) *Ex vivo* culture and stimulation of auditory neurons from WT A/J mice with increasing concentration of glutamate (0.1--50 μM). Before glutamate treatment, auditory neurons were loaded with Ca^2+^ sensitive ratiometric probe (FLUO-8). Traces show the kinetics of cytosolic Ca^2+^ release upon glutamate stimulation (black trace; glutamate 50 μM). E) Dose response curve of glutamate induced Ca^2+^ release allowing determination of glutamate EC50. F) Impact of pre-treatment with ryanodine or adenosine (1--100 μM) on glutamate induced Ca^2+^ release. SGN were stimulated with 5 μM glutamate, corresponding to EC90. Data represent the average±SEM of 3 independent experiments. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)Fig. 6 **Analysis of the impact of adenosine and ryanodine on glutamate-induced excitation of auditory neurons**. To address the functional relevance of these genes in the excitatory response, we measured Ca^2+^ mobilization in the cytosol of auditory neurons in response to glutamate ([Fig. 6](#fig6){ref-type="fig"}D and E). Auditory neurons were obtained from spiral ganglia sphere forming stem cells as previously described \[[@bib33]\]. While the genesis of spiral ganglia from this system has been described previously, so far no studies addressing the glutamatergic function have been performed on such a model. Upon glutamate stimulation, robust and immediate increase in cytosolic Ca^2+^ concentration was observed in the auditory neurons ([Fig. 6](#fig6){ref-type="fig"}D). This excitatory response was dose dependent and showed an EC~50~ in the micromolar range ([Fig. 6](#fig6){ref-type="fig"}E), similar to glutamate dose response curves reported in other systems \[[@bib34],[@bib35]\]. In cortical neurons, adenosine, through adenosine receptor type I (*Adora1*), was shown to antagonize the glutamatergic pathway \[[@bib36]\]. Interestingly, *Adora1* tended to be upregulated in mutant mice cochlea (FDR = 0.1443). To investigate the functional relevance of ryanodine receptors, we pre-incubated auditory neurons for 30 min with ryanodine, which under these conditions inhibits glutamate-induced Ca^2+^ release from the ryanodine-sensitive Ca^2+^ stores. We observed a dose-dependent inhibition of Ca^2+^ release by ryanodine with a maximal inhibition of 60%, suggesting that a substantial portion of glutamate-induced Ca^2+^ elevations in auditory neurons originates from ryanodine-sensitive Ca^2+^ stores ([Fig. 6](#fig6){ref-type="fig"}F, blue line and symbols). We next investigated the functional impact of activation of adenosine receptors. We observed a dose-dependent inhibition of cytosolic Ca^2+^ elevation by adenosine with a maximal inhibition of 40%. Thus, similar as seen in other types of neurons, adenosine receptors are inhibitors of glutamatergic cell activation stores ([Fig. 6](#fig6){ref-type="fig"}F, magenta line and symbols). **Decreased Ca**^**2+**^ **induced Ca**^**2+**^ **release (CICR) in p22**^**phox**^ **deficient auditory neurons confers a protection against glutamate-mediated excitotoxicity.** To further characterize the excitatory function of p22^phox^ −/− auditory neurons, we compared the expression of key genes involved in glutamate response, namely ryanodine receptors and glutamate ionotropic receptor *Gria2* ([Fig. 7](#fig7){ref-type="fig"}A). The data demonstrate significant downregulation of ryanodine receptor 1 and 2, as well as Gria2 in p22^phox^-deficient auditory neurons ([Fig. 7](#fig7){ref-type="fig"}A). To understand the physiological relevance of this differential expression, we compared glutamate-induced Ca^2+^ release in sphere-derived auditory neurons from WT and p22^phox^−/− mice ([Fig. 7](#fig7){ref-type="fig"}B). Upon glutamate stimulation, we observed a dose-dependent increase of Ca^2+^ cytosolic concentration with a comparable EC~50~ in both WT (5.635 μM) and p22^phox^ −/− (5.183 μM) auditory neurons. However, the overall amplitude of the signal was significantly lower in p22^phox^ −/− auditory neurons ([Fig. 7](#fig7){ref-type="fig"}B; [Supplementary Fig. 7](#appsec1){ref-type="sec"}). Furthermore, in absence of p22^phox^, the ryanodine-sensitive Ca^2+^ release was significantly decreased ([Fig. 7](#fig7){ref-type="fig"}C and D), consistent with the lower expression of *Ryr1* and *2* ([Fig. 7](#fig7){ref-type="fig"}A). Thus, we identified a functional role for p22^phox^ in auditory neurons, namely the control of Ca^2+^ release from ryanodine-sensitive stores. Based on the results of the gene expression analysis, as well as the functional characterization of the p22^phox^−/− and WT spiral ganglia cells, we hypothesized that the lower amplitude of the CICR in p22^phox^ deficient auditory neurons is protective against glutamate induced excitotoxicity. We therefore compared the impact of glutamate on level of cellular integrity (ATP levels) and morphology of WT and p22^phox^ −/− spiral ganglia cells ([Fig. 7](#fig7){ref-type="fig"}E--G). Upon glutamate exposure, WT auditory neurons showed dramatic morphological changes, with, in particular loss of neurites ([Fig. 7](#fig7){ref-type="fig"}E). Following glutamate exposure, neurite length in the WT cells was decreased by 50% ([Fig. 7](#fig7){ref-type="fig"}F). Remarkably, p22^phox^ deficient auditory neurons were almost fully protected from glutamate induced morphological changes with only minor decrease in average neurite length (not statistically significant). The data also showed significant decrease in ATP concentration in WT SGN in a glutamate concentration dependent manner ([Fig. 7](#fig7){ref-type="fig"}G). By contrast, the ATP content did not significantly decrease after glutamate exposure in p22^phox^-deficient cells. Together, the data demonstrate that downregulation of the excitatory pathway confers a protection against glutamate mediated toxicity in p22^phox^ deficient auditory neurons.Fig. 7**p22**^**phox**^**deficiency mitigates glutamatergic response of auditory neurons and protects from glutamate-induced toxicity.** A) mRNA levels of Ryanodine receptor 1 (Ryr1), Ryanodine receptor 2 (Ryr2) and Glutamate ionotropic receptor AMPA type subunit 2 (Gria2) in spiral ganglion neurons from p22^phox^ deficient (grey bars) expressed relatively to WT mice (black bars, normalized to 1). n = 3 WT; n = 3 p22^phox^ −/−. B) Dose response curve of glutamate induced Ca^2+^ release in WT (black trace) and p22^phox^ −/− (grey trace) auditory neurons. p \< 0.0001 with two way ANOVA. C) Effect of increasing concentrations of Ryanodine on glutamate-induced Ca^2+^ mobilization. Auditory neurons were pre-treated with ryanodine (0.3--100 μM) and stimulated with EC90 glutamate (5 μM). The bar graph shows the relative effect of ryanodine on Ca^2+^ mobilization, induced by the glutamate treatment (WT, black; p22^phox^ −/−, grey). D) The bar graph shows the proportion of glutamate stimulated Ca^2+^ mobilization inhibited by the different concentrations of ryanodine. E) High glutamate concentration impact on auditory neuron morphology. Auditory neurons from WT and p22phox-deficient A/J mice were cultured ex vivo and treated with glutamate 1 mM for 6 h. Cells were then fixed with PFA 4% and subject to BIII tubulin immunostaining (red). Nuclei were colored with DAPI (blue). Pictures are representative of 3 independent experiments. F) Bar graph showing the relative neurite length of auditory neurons treated or not with glutamate, normalized to the WT values. G) Impact of increasing glutamate concentrations on auditory neuron viability. Auditory neurons from WT or p22^phox^-deficient A/J mice were cultured ex vivo and treated with increasing concentrations of glutamate (0.1--3 mM). Viability was then assessed using an ATP measurement kit (ATP lite). Data represent the average±SEM of 3 independent experiments. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)Fig. 7 3. Discussion {#sec3} ============= Our results provide the new evidence that the absence of functional p22^phox^ in the inner ear robustly protects from age-related hearing loss and related morphological changes in mice. Mechanistically, we found that the protection is mediated by decrease of the excitatory calcium/glutamate signaling pathway in auditory neurons. Our data provide an in-depth description of the early and progressive hearing loss in A/J mice over the first 6 months of life, beyond and in line with earlier reports \[[@bib37],[@bib38]\]. In this background, onset of hearing loss occurs at 6 weeks of age together with predominant signs of synaptic and post-synaptic damages, including loss of auditory neurons and synaptic ribbons, as well as a decreased amplitude of ABR wave 1. Consistent with recent report in human \[[@bib39]\], the auditory neuropathy precedes loss of hair cells, which occurs at a later stage in our model. This model mimics the progression of human presbycusis as the hearing loss in A/J mice starts in the high frequencies while lower frequencies are only affected later in life. Three mutations known to affect hearing are present in A/J mice, thereby providing structural and metabolic explanations for the severe and rapid age-related hearing loss. These include: i) a mutation in the age-related hearing loss-1 (*ahl1*) locus, which codes for cadherin 23, a structural protein of the stereocilia; ii) a mutation in *ahl4*, coding for citrate synthase, a mitochondrial enzyme of the Krebs cycle, important for ATP and NADPH synthesis and iii) mitochondrially encoded tRNA arginine, leading to a decreased efficiency in amino acid incorporation. Interestingly, a recent *in vitro* study demonstrated that decreased citrate synthase expression leads to increased production of ROS \[[@bib37]\]. This increase of ROS was proposed to be due to a decrease in the cellular antioxidant system. However, the fact that p22^phox^ mutation can almost fully prevent the age-related hearing loss in spite of the presence of the above-mentioned mutations suggests that p22^phox^ and/or its downstream pathways are major driver of age-related hearing loss. The A/J mouse model is a powerful tool for the study of age-related hearing loss and otoprotection. As the absence of p22^phox^ abolishes NOX1--4 catalytic activity, our data strongly support a role of NOX-derived ROS in presbycusis. However, the NOX isoform involved in hearing loss remains to be identified. Nevertheless as of today, the differential distribution of NOX isoforms in the inner ear has not been clearly elucidated. Given the problems of specificity of commercially available NOX antibodies \[[@bib32]\], we focused our efforts on detection of NOX mRNA by qPCR and in situ hybridization providing a striking similarity of NOX isoform expression between human and mouse. The qPCR experiments confirmed high expression levels of NOX3 and p22^phox^, but also provided evidence of high levels of NOX2 and NOX4 in the cochlea. The *in-situ* hybridization experiments using the high-resolution technology RNAscope® yielded relevant new information. *Nox2* and *Nox4* mRNA appeared equally distributed throughout the cochlea. We had rather expected some strong Nox2-positive microglia cells in the region of the spiral ganglion. Note however that quiescent microglia express only low levels of Nox2, which most likely explains our observations. The most striking result was the high level of *Nox3* mRNA in the spiral ganglion cells and its virtual absence in hair cells. This finding supports a role of Nox3 in the morphological changes observed in the cochleae of 6 weeks old animals, where at the first onset of hearing loss, ribbons and SGN density are affected but not hair cells. At this point, it is not clear whether the p22^phox^-dependent degeneration of hair cells is secondary to degeneration of SGN or whether it is an independent process that occurs later in time. Concerning the first option, it is unlikely that degeneration of SGN by itself leads to loss of outer hair cells, as at least in the case of peripherin knock-out mice, SGN, but not OHCs are lost \[[@bib40]\]. However we cannot exclude some NOX3-dependent efferent signaling mechanism that leads to a secondary OHC death. Yet, the second option, namely a delayed SGN-independent OHC death, certainly appears possible. While under physiological conditions in 6 week old mice, expression levels of NOX3 was low, it might be upregulated with aging. Indeed, there are indications that upon stress, Nox3 might be upregulated in hair cells \[[@bib31]\]. Alternatively other p22^phox^-dependent NOX enzymes might be upregulated in the aging mice. Cell type-specific NOX knock-out will be necessary to definitively clarify this issue. The magnitude of the protective effect of p22^phox^ deficiency is remarkable. Indeed, up to 6 month of age, there was only a minor hearing loss in p22^phox^ mutant mice, while wild type A/J mice were virtually deaf at this age. This raises the question how the lack of p22^phox^ can have such a profound effect. Indeed, NADPH oxidase are not the only source of ROS in the cochlea (e.g. mitochondria are probably important), and mechanisms other than ROS are also likely to play a role in cochlear damage \[[@bib8]\]. Thus, p22^phox^ may represent "the straw that breaks the camel\'s back", i.e. the cumulative effect of many insults leads to hearing loss and by removing one of them (i.e. p22^phox^), the function of the cochlea is maintained. Our results demonstrate that p22^phox^ loss of function is strikingly important in sensory neurons as it leads to a lower expression of genes of the neuronal excitatory pathway in the cochlea. This supports a key role of NOX-derived ROS in this mechanism. Downregulated genes harbor functions in Ca^2+^ signaling (ryanodine receptors 1,2,3), in synaptic transmission (Otoferlin, *Snap25* and *Vamp1*), as well as in the glutamatergic pathway (*Slc17a6*, *Slc17a8* and *Gria2*). The cytosolic free Ca^2+^ concentration plays a key role in pre- and post-synaptic auditory signal transmission. On the presynaptic side, Ca^2+^ influx through Ca^2+^ channels activates synaptic vesicle fusion and subsequent glutamate release \[[@bib41], [@bib42], [@bib43]\]. On the post-synaptic side, glutamatergic transmission involve the glutamate ionotropic AMPA type subunit 2 (*Gria2*) leading to Ca^2+^ increase in auditory neurons \[[@bib44]\]. The lower expression of neuronal excitatory genes may be explained by a delay in development as previously described \[[@bib45]\] or to a downregulation of these genes in p22^phox^ mutant mice. Interestingly, lower expression levels of ryanodine receptors in p22^phox^ mutant mice are sufficient for normal hearing. However, our data also suggest that enhanced expression of ryanodine receptors leads to glutamate-induced toxicity and subsequent hearing loss, which was mitigated in p22^phox^ loss of function mice. In our *in vitro* auditory neuron model, we provide direct evidence that ryanodine receptors are involved in the post-synaptic Ca^2+^ signal: glutamate-induced Ca^2+^ elevations are inhibited by ryanodine in a dose-dependent manner. From the relative inhibition of the Ca^2+^ signal, we can conclude that ryanodine receptors are important mediators of the excitatory signal in the cochlea at the post synaptic level. This observation is consistent with previous findings that AMPA-type glutamate receptors and intracellular Ca^2+^ stores are coupled via RyR-dependent CICR in primary auditory neurons \[[@bib46]\]. Furthermore, intracochlear perfusion of ryanodine in guinea pigs in vivo has been shown to decrease auditory nerve compound action potentials, indicating the importance of RyRs in cochlear transduction \[[@bib47]\]. Direct activation of Ryr2 by ROS is well-documented \[[@bib48],[@bib49]\], however, in the present study, transcriptomic data suggest that NOX-derived ROS rather act at the transcriptional level. Further characterization of the promoters and the transcriptional machinery will be required to understand whether specific redox-sensitive transcription factors are involved. To explore key finding of our study, namely a protection of p22^phox^-deficient animals from age-related hearing loss, other findings need to be discussed. The A1AR adenosine receptor has been shown to be linked to hearing loss and NOX signaling: mice deficient for the A1AR exhibit greater susceptibility to noise exposure \[[@bib50]\] and A1AR activation confers protection against cisplatin ototoxicity through downregulation of NOX3 \[[@bib51]\]. Our results show only a trend towards increased levels of A1AR in p22^phox^-deficient mice (p = 0.0064; FDR = 0.1443). Interestingly, A1AR is an inhibitor of the excitatory pathway \[[@bib52],[@bib53]\], and in our hands, adenosine significantly decreased the excitatory signal in auditory neurons ([Fig. 6](#fig6){ref-type="fig"}). This might provide an additional element for the understanding why p22^phox^-deficient mice are protected. To the best of our knowledge, this is the first study investigating age-related hearing loss in p22^phox^ deficient mice. However, the impact of Nox3 deletion has respectively been addressed in the context of cisplatin \[[@bib31]\] and noise \[[@bib38]\] induced hearing loss. The first study reports a protective impact in cisplatin-induced hearing loss following injection of Nox3 siRNA in the middle ear while the second publication by Lavinsky and collaborators using Nox3 mutant mouse concluded that Nox3 might be protective against noise-induced inner ear damage. In terms of pattern of hearing loss, the Lavinsky study did not show an impact of Nox3 deficiency on noise-induced hearing loss at high frequencies, but a small protective effect of Nox3 at low frequencies, statistically significant only at 8 kHz. Our results studying age-related hearing loss showed a protective effect of p22^phox^ deficiency at frequencies above 8 kHz. The reason for these apparently conflicting data may be due to i) differences in the type of hearing loss (noise-induced, ototoxicity and age-related), ii) mice in a different genetic background were studied (black6 vs A/J), and iii) importantly, a different type of NOX-deficient mouse was investigated (NOX3-deficiency vs. p22^phox^ deficiency, corresponding to a functional NOX1-4 knock-out). Future studies addressing each isoform regulated by p22^phox^ may provide a better understanding of such differences. Our study provides first conclusive evidence that p22^phox^ is key regulator of age-related hearing loss, most likely through regulation of the activity of ROS-generating NADPH oxidases. The extent of the protection of p22^phox^-deficient animals is astonishingly strong. Our data suggest that p22^phox^ -- master regulator of the NOX pathway -- is not only regulating the levels of non-specific ROS-induced damage in the cochlea. In fact, we demonstrate that level of expression of genes of the excitatory pathway is strikingly diminished in p22^phox^ mutant mice. Thus, our data suggest that NOX-enhanced excitotoxicity contributes to age-related hearing loss. We believe that targeting NOX catalytic activity is a promising strategy to prevent age-related hearing loss. Small molecules NOX inhibitors would be interesting tools, but might be difficult to target to the inner ear. Local delivery of molecular therapies (i.e. RNA knock-down of p22^phox^) may provide long term protection against presbycusis and other related inner ear damage, but will equally have to be precisely targeted \[[@bib54]\]. 4. Methods {#sec4} ========== 4.1. Animal procedures {#sec4.1} ---------------------- A.B6 *Tyr*^+^-*Cyba*^*nmf333*^/J mice were purchased from Jackson lab (strain 005445) and colonies were maintained at the animal facility of the University of Geneva through heterozygous x heterozygous mating, ensuring equal proportion of WT and p22^phox^ deficient littermates. Both male and female, were subjected to hearing threshold determination using the auditory brainstem response at different time points from 4 weeks old to 26 weeks old. At the end of the experiments, animals were sacrificed for cochlea histology or mRNA extraction. For the auditory tests, or before sacrifice, mice were anesthetized with a mix of Ketamine (100 mg/kg) and Xylazine (10 mg/kg). 4.2. Auditory brainstem response {#sec4.2} -------------------------------- Animals were anesthetized by intraperitoneal injection Ketamine 100 mg/kg and of Xylazine 5 mg/kg, and placed upon a heating pad to maintain body temperature. Depth of anesthesia was tested every 30 min by testing the pedal withdrawal reflex. If necessary, additional injections of about 50% of the initial dose were given. ABR recordings were performed in a sound proof chamber (IAC Acoustics, Illinois IL, USA). For stimulus generation and recording of responses, a multi-function IO-Card (National Instruments, Austin TX, USA) was used, housed in an IBM compatible computer. An integrated software package for stimulus generation and recording (Audiology_lab; Otoconsult, Frankfurt, Germany) was used. Sound pressure level was controlled with an attenuator and amplifier (Otoconsult, Frankfurt, Germany). Stimuli were delivered to the ear in a calibrated open system by a loudspeaker (AS04004PR-R, PUI Audio, Inc., Dayton, USA) placed 3 cm lateral to the animals' pinna. Sound pressure was calibrated on-line prior to each measurement with a microphone probe system (Bruel&Kjaer 4191) placed near the animals' ear. Recorded signals were amplified and bandpass filtered (80 dB; 0.2--3.0 kHz) using a filter/amplifier unit (Otoconsult, Frankfurt, Germany). For the recordings, silver electrodes were placed subcutaneously on the mouse forehead (+), on the mastoid of the recorded ear (−) and a reference electrode on the back. ABR were recorded, following stimulation with 100μs clicks or 3 ms tone pipes (2.0--45.2  kHz at a resolution of 2 steps per octave). For all frequencies, ABRs were recorded from 0 to 90 dB SPL in 3 dB steps. Electrical signals were averaged over 256 repetitions of stimulus pairs with alternating phase. Hearing thresholds were defined as the sound pressure level where a stimulus-correlated response was clearly identified by visual inspection of the averaged signal. 4.3. Cochlea histology (mid modiolar cut, cytocochleograms and ribbons staining) {#sec4.3} -------------------------------------------------------------------------------- At the end of the last ABR measurement, anaesthetized mice were sacrificed by cervical dislocation followed by decapitation. Temporal bones were isolated from the skull in order to extract the inner ears, and after proper dissection cochlea were placed in 4% paraformaldehyde overnight at room temperature. Cochleae were decalcified using USEDECALC solution (Medite commercial solution) under sonication for 48 h (following ultrasonic technique of USE 33 decalcification machine from Medite, Cat. No. 03-3300-00). Finally, decalcified cochlea were microdissected to perform cytocochleograms and immunohistochemistry, or mid modiolar cuts (embedded in paraffin) with Hematoxilin-Eosin staining. As general rule, one ear from each animal was used for cytocochleograms and the other ear for mid-modiolar staining. ### 4.3.1. Immunohistochemistry and confocal microscopy (cytocochleograms) {#sec4.3.1} Decalcified cochleae were microdissected with thin forceps under binocular microscope. The bony shell was removed to expose the Organ of Corti (OC), followed by the removal of the stria vascularis and separation of the sensory epithelium from the spiral ganglion. The basal and middle turns of the OC were cut in two pieces each one and the apical turn was preserved as one piece, resulting in 5 pieces transferred into 400 μL PBS solution in a 48 well plate. Samples were then permeabilized (3% Triton-X 100 in PBS 1X) for 30 min at room temperature and immersed in a blocking buffer, containing 2% bovine serum albumin (BSA) and 0.01% Triton-X 100 in PBS, for 1 h at room temperature. Explants were incubated with anti-MyoVIIa (1:200, rabbit; Proteus, USA) and anti-CtBP2 (1:200, monoclonal mouse, BD Bioscience) antibodies in blocking buffer overnight at 4 °C. On the following day, tissues were rinsed three times with PBS and incubated, for 2 h at room temperature, with the secondary antibodies anti-rabbit Alexa Fluor 488 (1:500; Invitrogen, USA) and goat anti-mouse Alexa 555 (1:500; Life Technologies) diluted in blocking buffer. Explants were washed 3 times with PBS and mounted on a glass slide with Fluoroshield containing DAPI (Sigmaaldrich, USA). The labelled cells were visualized with a confocal laser-scanning microscope (Zeiss LSM700) equipped with a CCD camera (Leica Microsystems) employing the Plan-Neofluar 20X/0.50 and Plan-Apochromat 63X/1.4 (Oil) objectives. Hair cell and ribbons counting was performed using image J software. ### 4.3.2. Mid modiolar hematoxylin eosin staining {#sec4.3.2} Decalcified cochlea were dehydrated and embedded in paraffin. Mid-modiolar cuts of 5 μm were processed and loaded onto gelatin-coated slides. Harris' Hemalun/eosin staining was performed on the slides. The samples were qualitatively analyzed for density of spiral ganglia using image J software. 4.4. Quantitative analysis of cochlea morphology {#sec4.4} ------------------------------------------------ The SGN density was determined by nuclear counting, employing the hematoxylin-eosin stained slides, of the Rosenthal\'s canal at each different cochlear turn (apical, medial and basal). Five non-consecutive sections were evaluated. To normalize the data, the total neuronal amount was divided between each Rosenthal\'s area (in mm^2^), and the measured densities were averaged. For the synaptic ribbons count, 10-15 μm Z stack (0.7 μm steps) were performed with the 63X objective and merged in a single picture, to ensure a proper counting of the ribbons distributed along the Z axis. For each section of the OC, ribbons belonging to two different segments of 10--15 inner hair cells were counted to obtain an average of each cochlear turn. Additionally, 20X pictures were employed to evaluate hair cells (HC) survival (cytocochleogram). Inner hair cells (IHC) and outer hair cells (OHC) population were recorded in two representative areas of 100 μm from each cochlear section, in order to average the HC survival rate of each turn (apical, medial and basal). The open source image J program was used for the image analysis. 4.5. Gene expression analysis {#sec4.5} ----------------------------- ### 4.5.1. Mouse cochlea RNA extraction {#sec4.5.1} Cochlea were quickly dissected in cold PBS, carefully removing remaining blood and surrounding tissues, immediately frozen in liquid nitrogen and kept at −80 °C for further RNA extraction. Cochlea were physically homogenized with clean steel balls and tissueLyser (Qiagen) for 30 s at 30 rpm, and the DNA and RNA precipitated with 750 μl Trizol. After adding 150 μl chloroform, (shaking, resting and centrifuging step) we collected the aqueous phase and performed column purification with Qiagen RNeasy Micro kit, ignoring the lysis and homogenization steps described in the kit protocol (adding directly 70% ethanol) (adapted from Ref. \[[@bib55]\]). ### 4.5.2. Human foetal cochlea RNA extraction {#sec4.5.2} The inner ear was isolated from aborted human fetuses ranging from W8 to W12 post conception. Procurement and procedures were performed with full approval by the Ethics Committee of the State Geneva, Switzerland and following signed informed consent of the donors. The postmenstrual date was used for the calculation of the fetal stage. Tissue dissection was performed in ice-cold Hanks' balanced salt solution (HBSS). The rest of the procedure was similar as for mouse cochlea RNA extraction (see paragraph above). ### 4.5.3. Real time quantitative polymerase chain reaction {#sec4.5.3} Following the cochlear extraction, RNA concentration was determined using a Nanodrop. 500 ng of RNA was used for cDNA synthesis using the Takara PrimeScript RT reagent Kit, following manufacturer\'s instruction. Real-time PCR was performed using SYBR green assay on a 7900HT SDS system from ABI. The efficiency of each primer was verified with serial dilutions of cDNA. Relative expression levels were calculated by normalization to the geometric mean of the three house-keeping genes (*Eef1a*, *Tubb* and *Actb* for mouse and *EEF1a*, *B2M* and *GAPDH* for human samples). The highest normalized relative quantity was arbitrarily designated as a value of 1.0. Fold changes were calculated from the quotient of means of these RNA normalized quantities and reported as ±SEM. Sequences of the primers used are provided in [Supplementary Table 1](#appsec1){ref-type="sec"}. ### 4.5.4. RNAscope® (ACDbio, bio-techne, minneapolis, Minnesota, United States) {#sec4.5.4} Cochleae were fixed, decalcified and embedded in paraffin (see above - mid modiolar cuts). 5 μm mid modiolar slices were loaded onto gelatin-coated slides and subjected to RNAscope probe hybridization. RNAscope Multiplex Fluorescent V2 assay (Bio-techne, Cat. No. 323110) or RNAscope 2.5 HD Assay -- BROWN assay (Bio-techne, Cat. No. 322310) was performed according to manufacturer\'s protocol, paraffin sections were hybridized with the probes Mm-Nox3-C1 (Bio-techne, Cat. No. 481989), Mm-Ppib-C1 (Bio-techne, Cat. No. 313911) as positive control and DapB-C1 (Bio-techne, Cat. No. 310043) as negative control at 40 °C for 2 h and revealed with TSA Opal570 (PerkinElmer, Cat. No. FP1488001KT). For Fluorescent assay, tissues were counterstained with DAPI and mounted with FluoromountG (Bioconcept, Cat. No. 100.01). For BROWN assay, Tissues were counterstained with Mayer hematoxyline for 30 s, dehydrated, cleared and mounted with Tissue-Tek® Glas™ (Sakura, Cat. No. 1408). Samples were visualized with a confocal laser-scanning microscope (Zeiss LSM700) equipped with a CCD camera (Leica Microsystems) employing the Plan-Neofluar 20X/0.50 and Plan-Apochromat 63X/1.4 (Oil) objectives. Pictures were analyzed using image J software. ### 4.5.5. RNA seq {#sec4.5.5} TruSeq ribodepleted stranded mRNA from 6 weeks old A/J mice (3 WT vs. 3 p22^phox^ deficient) was applied to eliminate ribosomic RNA and sequenced using Illumina TruSeq protocol. The sequencing quality control was done with FastQC v.0.11.5. The sequences were mapped with the TopHat v.2.0.11 software to the UCSC mm10 mouse reference. The biological quality control and summarization were performed using the PicardTools v.1.141. The table of counts with the number of reads mapping each gene feature of UCSC mm10 was prepared with HTSeq v0.6p1. The differential expression analysis was performed with the statistical analysis R/Bioconductor package edgeR v. 3.14.0. Briefly, the counts were normalized according to the library size and filtered. The genes having a count above 1 count per million reads (cpm) in at least 4 samples were kept for the analysis. The initial number of genes in the set was 23′420 and after the poorly or non-expressed genes were filtered out, 15′126 genes were left. The p values of differentially expressed gene analysis were corrected for multiple testing problems using the Benjamini-Hochberg (BH) procedure. All RNA-sequencing data files were submitted to the arrayexpress database at EMBL-EBI ([www.ebi.ac.uk/arrayexpress](https://www.ebi.ac.uk/arrayexpress){#intref0010}) under the accession number E-MTAB-8668. 4.6. In vitro culture of mouse SGN {#sec4.6} ---------------------------------- Mouse spiral ganglion neurons were isolated from p2-p5 A/J pups, cultured and passaged as previously described \[[@bib33]\]. Briefly, neurospheres from spiral ganglia were maintained in DMEM:F12 with 1x N2 and B27 supplement in presence of bFGF, IGF1, Heparan sulfate and EGF. For passages or differentiation, cells were dissociated with acutase and mechanically disintegrated. Following filtration with a 70 μm filter, retaining non dissociated spheres, SGN were counted and plated in a 384 or 96 well plate coated with matrigel (hESC qualified, Corning, New York, USA) with medium without growth factors, in order to induce differentiation. After 5 days of differentiation, cells were employed for cytosolic Ca2+ measurement and excitotoxicity assay. 4.7. Measurement of Ca^2+^ cytosolic release {#sec4.7} -------------------------------------------- 16000 cells/well were differentiated to SGN on a 96 well plate coated with matrigel, as described above (*in vitro* culture of spiral ganglion neurons). Differentiated cells were loaded with FLUO-8 (Interchim, Montluçon, France) according to the manufacturer protocol. After 45min incubation at 37 °C, the glutamate-induced cytosolic calcium release was assessed in a FDSS/μCELL Functional Drug Screening System (Hamamatsu, Yokohama, Japan). The neuronal kinetics of calcium release was followed over 10min following glutamate addition, with 1 measure every 0.5 s. The impact of ryanodine or adenosine (0-100 μM), respectively added 30min or 24 h prior glutamate, was also investigated. 4.8. Impact of high glutamate concentration on auditory neuron ATP content {#sec4.8} -------------------------------------------------------------------------- 16000 cells/well were differentiated to SGN on a 96 well plate coated with matrigel, as described above (*in vitro* culture of spiral ganglions). Differentiated cells were then treated with increasing concentrations of glutamate for 6 h and the SGN viability was assessed using the ATPLite kit (PerkinElmer, Wellesley, USA), following manufacturer\'s instructions using spectra L reader. 4.9. Impact of high glutamate concentration on auditory neuron morphology {#sec4.9} ------------------------------------------------------------------------- Auditory neuron progenitors were seeded on matrigel coated coverslips and differentiated into auditory neurons for 5 days. Cells were then exposed to 1 mM glutamate for 6 h and fixed for 10 min with PFA 4%. Following the fixation, neurons were permeabilized (0.2% Triton-X100 in PBS 1X) for 30 min at room temperature and immersed in a blocking buffer containing 2% BSA and 0.01% Triton-X 100 for 1 h at room temperature. SGNs were incubated with the primary antibody anti-BIII tubulin (1:1000, mouse; Biolegend, USA) in blocking buffer overnight at 4 °C. On the following day, tissues were rinsed three times with PBS and incubated with the secondary antibody anti-mouse Alexa Fluor 555 (1:500; Invitrogen, USA) in blocking buffer for 2 h at room temperature. Coverslips were washed 3 times with PBS and mounted on a glass slide with Fluoroshield containing DAPI (Sigmaaldrich, USA). The labelled cells were visualized with a confocal laser-scanning microscope (Zeiss LSM700) equipped with a CCD camera (Leica Microsystems) with Plan-Neofluar 10X/0.30 NA objective. Neurite length evaluation was performed using image J software. 4.10. Statistics {#sec4.10} ---------------- All data were analyzed using Two-way ANOVA followed by Bonferroni multiple comparison test using GraphPad Prism software unless where stated otherwise in the Figure legend. Values with p \< 0.05 was considered as statistically significant. \*; p \< 0.05, \*\*p \< 0.01, \*\*\*; p \< 0.005, \*\*\*\*p \< 0.0005. 4.11. Study approval {#sec4.11} -------------------- All procedures were approved by the local veterinary office and the Commission for Animal experimentation of the Canton of Geneva, Switzerland, authorization number GE/106/18. Author contributions {#sec5} ==================== FR, KHK and PS designed the study; FR, MC, GNS, AM, VBCK, MP and YC acquired the data; FR, GNS, YC, VJ, KHK and PS analyzed the data; FR, VJ, KHK and PS wrote the manuscript. Declaration of competing interest ================================= The authors have declared that no conflict of interest exists. Appendix A. Supplementary data {#appsec1} ============================== The following are the Supplementary data to this article:Multimedia component 1Multimedia component 1Multimedia component 2Multimedia component 2 This work was supported by funding from the Fondation Louis Jeantet"recherche translationnelle" (to K.H.K. and P.S.) and Swiss National Science Foundation grant (SNF number 31003A-179478), Bärguf Foundation and Bodifée Foundation. The authors would like to thank Dr. Jessica Sordet from the Histology platform of the Ecole Polytechnique Fédérale de Lausanne (EPFL) for RNAscope® experiments, Dr. Marcus Muller for the implementation of the auditory station, Mrs. Heifa Alsheikh for the help with ABR analysis and Mrs. Rebecca Sipione for the help with the graphical abstract design. The authors also would like to thank Prof. Marie Cohen for the help with the human samples collection. Supplementary data to this article can be found online at <https://doi.org/10.1016/j.redox.2020.101434>. [^1]: PS and KHK share last authorship.

Liver cancer is the second most common cause of cancer-related death, and its incidence is increasing worldwide. Hepatocellular carcinoma (HCC) accounts for 85--90% of cancers arising in the liver. Although many efforts have been made to improve the prognosis of HCC, the mortality from HCC almost equals its incidence, suggesting a dismal prognosis of HCC. The etiologies of HCC are evident, and viral hepatitis contributes the most. Thus, the primary goal in reducing mortality from HCC is the prevention of viral hepatitis. However, for the population already infected with viral hepatitis, the best strategy would be secondary prevention and early detection of HCC with active surveillance. The first essential step in the surveillance strategy is the identification of the population at risk of HCC. The current guidelines simply define high risk as those with cirrhosis from any etiology, and those with chronic hepatitis B virus (HBV) infection irrespective of cirrhosis, based on the estimated HCC risk per year. However, the risk of HCC varies greatly in the target population. Therefore, there is a need for new assessments that can accurately stratify the risk of HCC, enabling more precise surveillance strategies. Recent advances in sequencing technologies have allowed for the identification of genetic variants, i.e., single nucleotide polymorphisms (SNPs), using genome-wide association studies (GWAS). Unlike the previously-used candidate-gene approach, GWAS enables us to identify genes without any previously suspected role in the pathophysiology of a given disease ([@bb0025]). To date, numerous SNPs associated with the development of HCC have been identified in case-control studies ([@bb0025]). MHC class I polypeptide-related chain A (MICA) is a membrane protein that acts as a ligand for NKG2D to activate anti-tumor effects through natural killer cells and CD8^+^ T cells ([@bb0005]). [@bb0015] demonstrated that MICA polymorphism and serum levels of MICA were associated with a risk for hepatitis C virus (HCV)-induced HCC. In the same group, MICA variants and serum MICA levels were suggested as prognostic biomarkers for HBV-induced HCC, showing a poorer prognosis in the group with high levels of serum MICA ([@bb0020]). In this issue of *EBioMedicine*, Huang and colleagues examined four candidate SNPs in a large longitudinal cohort ([@bb0010]). They reported that the MICA A allele and a high serum level of MICA were associated with HCV-related HCC development in cirrhotic patients who did not achieve sustained virological response (SVR), but not in non-cirrhotic patients and/or patients with SVR. The authors suggest that combining the genetic variants of the MICA gene and serum levels of MICA proteins may enhance the predictive power in the high-risk population for HCC. They stressed the impact of the results of their study on the association detected in cirrhotic patients without achieving SVR. With the introduction of direct acting antivirals (DAAs), it is anticipated that SVR will be achieved in most patients with hepatitis C. However, the availability and affordability of DAAs vary by region, and the role of DAAs in the immune system remains unanswered. Taken together, the authors\' conclusion that combining MICA gene variants and serum level of MICA protein may provide insight for closer follow-up strategies and re-treatment priority seems reasonable. The main contribution of this work lies in the use of longitudinal follow-up data and the adjustment of potential confounders, such as preexisting cirrhosis and viral eradication, in their analysis. Once the association between certain SNPs and the risk of cancers in GWAS has been found in the setting of a case-control study, it should ideally be validated by prospective studies, taking into account non-genetic baseline features (known risk factors) and competing risks (liver disease-related death) ([@bb0025]) before being used in clinical practice. However, care is needed in assessing the utility of genetic tests in real clinical settings, and there are several issues to be considered. Racial differences in the frequency of SNPs have been suggested and should be explored thoroughly. When interpreting associated SNPs, clinical features that might influence the risk for HCC (liver inflammation, fibrosis, or degree of cirrhosis) should be reflected through robust adjustment to avoid misinterpretation of the impact of given SNPs. The role of these genetic variants in the carcinogenic process of HCC and plausible mechanisms should also be elucidated. Lastly, before implementing the use of genetic tests to define population at high-risk for HCC, its cost-effectiveness has to be examined. Declaration of interests {#s0005} ======================== J Choi has nothing to disclose. Y-SL is an advisory board member of Bayer Healthcare, Bristol-Myers Squibb, and Gilead Sciences, and receives research funding from Bayer Healthcare, Bristol-Myers Squibb, Gilead Sciences, and Novartis. We have no other disclosures to report.

Related literature {#sec1} ================== For background to phase transition materials see: Li *et al.* (2008[@bb2]); Zhang *et al.* (2009[@bb5]). Experimental {#sec2} ============ {#sec2.1} ### Crystal data {#sec2.1.1} C~8~H~12~N^+^·Cl^−^*M* *~r~* = 157.64Monoclinic,*a* = 9.4739 (19) Å*b* = 9.894 (2) Å*c* = 9.6709 (19) Åβ = 96.31 (3)°*V* = 901.0 (3) Å^3^*Z* = 4Mo *K*α radiationμ = 0.35 mm^−1^*T* = 293 K0.4 × 0.3 × 0.2 mm ### Data collection {#sec2.1.2} Rigaku Mercury2 diffractometerAbsorption correction: multi-scan (*CrystalClear*; Rigaku, 2005[@bb3]) *T* ~min~ = 0.880, *T* ~max~ = 0.9329081 measured reflections2068 independent reflections1585 reflections with *I* \> 2σ(*I*)*R* ~int~ = 0.038 ### Refinement {#sec2.1.3} *R*\[*F* ^2^ \> 2σ(*F* ^2^)\] = 0.047*wR*(*F* ^2^) = 0.150*S* = 1.012068 reflections91 parametersH-atom parameters constrainedΔρ~max~ = 0.38 e Å^−3^Δρ~min~ = −0.28 e Å^−3^ {#d5e413} Data collection: *CrystalClear* (Rigaku, 2005[@bb3]); cell refinement: *CrystalClear*; data reduction: *CrystalClear*; program(s) used to solve structure: *SHELXS97* (Sheldrick, 2008[@bb4]); program(s) used to refine structure: *SHELXL97* (Sheldrick, 2008[@bb4]); molecular graphics: *SHELXTL* (Sheldrick, 2008[@bb4]); software used to prepare material for publication: *PRPKAPPA* (Ferguson, 1999[@bb1]). Supplementary Material ====================== Crystal structure: contains datablocks I, global. DOI: [10.1107/S1600536810020568/pv2285sup1.cif](http://dx.doi.org/10.1107/S1600536810020568/pv2285sup1.cif) Structure factors: contains datablocks I. DOI: [10.1107/S1600536810020568/pv2285Isup2.hkl](http://dx.doi.org/10.1107/S1600536810020568/pv2285Isup2.hkl) Additional supplementary materials: [crystallographic information](http://scripts.iucr.org/cgi-bin/sendsupfiles?pv2285&file=pv2285sup0.html&mime=text/html); [3D view](http://scripts.iucr.org/cgi-bin/sendcif?pv2285sup1&Qmime=cif); [checkCIF report](http://scripts.iucr.org/cgi-bin/paper?pv2285&checkcif=yes) Supplementary data and figures for this paper are available from the IUCr electronic archives (Reference: [PV2285](http://scripts.iucr.org/cgi-bin/sendsup?pv2285)). The author is grateful to the starter fund of Southeast University for financial support to purchase the diffractometer. Comment ======= There has been interest in the study of phase transition materials, including organic ligands, metal-organic coordination compounds, organic-inorganic hybrids, etc. (Li *et al.*, 2008; Zhang *et al.*, 2009). Exploring the phase transition materials, the dielectric properties of the title compound, have been investigated in my laboratory. Unfortunately, there was no distinct anomaly observed from 93 K to 380 K, (m.p. 408 K-410 K). In this article, the crystal structure of the title compound has been presented. The asymmetric unit of the title compound contains a 2,4-dimethylanilinium cation and a chloride anion (Fig. 1). The non-H atoms of the 2,4-dimethylanilinium cation are essentially coplanar. In the crystal structure, all hydrogen atoms bonded to the ammonium nitrogen (N1) are hydrogen bonded to the chloride ions with N···Cl distances in the range 3.080 (2) - 3.136 (2) Å; four formula units of the title compound are hydrogen bonded to form sixteen membered macrocyclic rings in the *bc*-plane (Tab. 1, Fig. 2). Experimental {#experimental} ============ The title compound was prepared by the reaction of 2,4-dimethylbenzenamine (1.21 g, 10 mmol) and hydrochloric acid solution (1.01 g, 10 mmol) in 30 ml methanol. The reaction mixture was filtered and left at room temperature for 4 days. Colorless crystals were obtained by slow evaporation. Refinement {#refinement} ========== Positional parameters of all H atoms were calculated geometrically and were allowed to ride on the atoms to which they are bonded, with N---H = 0.89 Å and C---H = 0.93 and 0.96 Å, for aryl and methyl type H-atoms, respectively, *U*~iso~(H) = 1.2 to 1.5*U*~eq~(C/N). Figures ======= {#Fap1} {#Fap2} Crystal data {#tablewrapcrystaldatalong} ============ ------------------------- --------------------------------------- C~8~H~12~N^+^·Cl^−^ *F*(000) = 336 *M~r~* = 157.64 *D*~x~ = 1.162 Mg m^−3^ Monoclinic, *P*2~1~/*c* Mo *K*α radiation, λ = 0.71073 Å Hall symbol: -P 2ybc Cell parameters from 7851 reflections *a* = 9.4739 (19) Å θ = 3.2--27.5° *b* = 9.894 (2) Å µ = 0.35 mm^−1^ *c* = 9.6709 (19) Å *T* = 293 K β = 96.31 (3)° Prism, colourless *V* = 901.0 (3) Å^3^ 0.4 × 0.3 × 0.2 mm *Z* = 4 ------------------------- --------------------------------------- Data collection {#tablewrapdatacollectionlong} =============== ------------------------------------------------------------------ -------------------------------------- Rigaku Mercury2 diffractometer 2068 independent reflections Radiation source: fine-focus sealed tube 1585 reflections with *I* \> 2σ(*I*) graphite *R*~int~ = 0.038 Detector resolution: 13.6612 pixels mm^-1^ θ~max~ = 27.5°, θ~min~ = 3.5° CCD_Profile_fitting scans *h* = −12→12 Absorption correction: multi-scan (*CrystalClear*; Rigaku, 2005) *k* = −12→12 *T*~min~ = 0.880, *T*~max~ = 0.932 *l* = −12→12 9081 measured reflections ------------------------------------------------------------------ -------------------------------------- Refinement {#tablewraprefinementdatalong} ========== ------------------------------------- ------------------------------------------------------------------------------------------------ Refinement on *F*^2^ Primary atom site location: structure-invariant direct methods Least-squares matrix: full Secondary atom site location: difference Fourier map *R*\[*F*^2^ \> 2σ(*F*^2^)\] = 0.047 Hydrogen site location: inferred from neighbouring sites *wR*(*F*^2^) = 0.150 H-atom parameters constrained *S* = 1.01 *w* = 1/\[σ^2^(*F*~o~^2^) + (0.0863*P*)^2^ + 0.190*P*\] where *P* = (*F*~o~^2^ + 2*F*~c~^2^)/3 2068 reflections (Δ/σ)~max~ \< 0.001 91 parameters Δρ~max~ = 0.38 e Å^−3^ 0 restraints Δρ~min~ = −0.28 e Å^−3^ ------------------------------------- ------------------------------------------------------------------------------------------------ Special details {#specialdetails} =============== ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Geometry. All e.s.d.\'s (except the e.s.d. in the dihedral angle between two l.s. planes) are estimated using the full covariance matrix. The cell e.s.d.\'s are taken into account individually in the estimation of e.s.d.\'s in distances, angles and torsion angles; correlations between e.s.d.\'s in cell parameters are only used when they are defined by crystal symmetry. An approximate (isotropic) treatment of cell e.s.d.\'s is used for estimating e.s.d.\'s involving l.s. planes. Refinement. Refinement of *F*^2^ against ALL reflections. The weighted *R*-factor *wR* and goodness of fit *S* are based on *F*^2^, conventional *R*-factors *R* are based on *F*, with *F* set to zero for negative *F*^2^. The threshold expression of *F*^2^ \> σ(*F*^2^) is used only for calculating *R*-factors(gt) *etc*. and is not relevant to the choice of reflections for refinement. *R*-factors based on *F*^2^ are statistically about twice as large as those based on *F*, and *R*- factors based on ALL data will be even larger. ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Fractional atomic coordinates and isotropic or equivalent isotropic displacement parameters (Å^2^) {#tablewrapcoords} ================================================================================================== ----- -------------- -------------- -------------- -------------------- -- *x* *y* *z* *U*~iso~\*/*U*~eq~ N1 0.61300 (18) 0.39383 (17) 0.84555 (17) 0.0480 (4) H1A 0.5798 0.4779 0.8370 0.072\* H1B 0.5968 0.3608 0.9279 0.072\* H1C 0.5694 0.3426 0.7783 0.072\* C1 0.7662 (2) 0.39422 (18) 0.8348 (2) 0.0436 (5) C3 0.9610 (2) 0.4527 (2) 0.7141 (2) 0.0538 (5) H3A 0.9972 0.4933 0.6388 0.065\* C2 0.8145 (2) 0.4537 (2) 0.7185 (2) 0.0473 (5) C6 0.8563 (2) 0.3347 (2) 0.9380 (2) 0.0533 (5) H6A 0.8205 0.2946 1.0138 0.064\* C4 1.0551 (2) 0.3944 (2) 0.8157 (2) 0.0562 (5) C5 1.0013 (3) 0.3352 (2) 0.9280 (3) 0.0609 (6) H5A 1.0628 0.2951 0.9977 0.073\* C8 0.7157 (3) 0.5163 (3) 0.6043 (2) 0.0658 (7) H8A 0.6591 0.5843 0.6427 0.099\* H8B 0.6547 0.4478 0.5600 0.099\* H8C 0.7700 0.5567 0.5370 0.099\* C7 1.2133 (3) 0.3985 (3) 0.8049 (3) 0.0834 (9) H7A 1.2629 0.3545 0.8843 0.125\* H7B 1.2440 0.4908 0.8018 0.125\* H7C 1.2332 0.3529 0.7216 0.125\* Cl1 0.49251 (6) 0.19132 (5) 0.62110 (6) 0.0603 (2) ----- -------------- -------------- -------------- -------------------- -- Atomic displacement parameters (Å^2^) {#tablewrapadps} ===================================== ----- ------------- ------------- ------------- -------------- -------------- -------------- *U*^11^ *U*^22^ *U*^33^ *U*^12^ *U*^13^ *U*^23^ N1 0.0525 (10) 0.0481 (9) 0.0456 (9) −0.0007 (7) 0.0146 (7) −0.0017 (7) C1 0.0499 (11) 0.0389 (10) 0.0434 (10) 0.0000 (8) 0.0116 (8) −0.0035 (7) C3 0.0564 (12) 0.0585 (13) 0.0494 (11) −0.0015 (10) 0.0185 (9) −0.0012 (9) C2 0.0545 (12) 0.0456 (10) 0.0436 (10) 0.0024 (8) 0.0126 (8) 0.0004 (8) C6 0.0638 (14) 0.0515 (12) 0.0454 (11) 0.0031 (9) 0.0090 (10) 0.0047 (9) C4 0.0512 (12) 0.0599 (13) 0.0583 (12) 0.0030 (10) 0.0098 (10) −0.0126 (10) C5 0.0619 (14) 0.0652 (14) 0.0542 (12) 0.0134 (11) −0.0002 (10) −0.0005 (11) C8 0.0661 (14) 0.0782 (16) 0.0543 (13) 0.0066 (12) 0.0122 (11) 0.0231 (12) C7 0.0516 (14) 0.111 (2) 0.0883 (19) 0.0052 (14) 0.0084 (13) −0.0130 (17) Cl1 0.0705 (4) 0.0570 (4) 0.0571 (4) −0.0126 (2) 0.0235 (3) −0.0140 (2) ----- ------------- ------------- ------------- -------------- -------------- -------------- Geometric parameters (Å, °) {#tablewrapgeomlong} =========================== ---------------- ------------- ---------------- ----------- N1---C1 1.466 (3) C6---H6A 0.9300 N1---H1A 0.8900 C4---C5 1.380 (3) N1---H1B 0.8900 C4---C7 1.514 (3) N1---H1C 0.8900 C5---H5A 0.9300 C1---C6 1.372 (3) C8---H8A 0.9600 C1---C2 1.391 (3) C8---H8B 0.9600 C3---C4 1.379 (3) C8---H8C 0.9600 C3---C2 1.394 (3) C7---H7A 0.9600 C3---H3A 0.9300 C7---H7B 0.9600 C2---C8 1.500 (3) C7---H7C 0.9600 C6---C5 1.388 (3) C1---N1---H1A 109.5 C3---C4---C5 118.2 (2) C1---N1---H1B 109.5 C3---C4---C7 120.4 (2) H1A---N1---H1B 109.5 C5---C4---C7 121.4 (2) C1---N1---H1C 109.5 C4---C5---C6 120.7 (2) H1A---N1---H1C 109.5 C4---C5---H5A 119.7 H1B---N1---H1C 109.5 C6---C5---H5A 119.7 C6---C1---C2 122.39 (19) C2---C8---H8A 109.5 C6---C1---N1 119.27 (17) C2---C8---H8B 109.5 C2---C1---N1 118.33 (18) H8A---C8---H8B 109.5 C4---C3---C2 123.4 (2) C2---C8---H8C 109.5 C4---C3---H3A 118.3 H8A---C8---H8C 109.5 C2---C3---H3A 118.3 H8B---C8---H8C 109.5 C1---C2---C3 116.01 (19) C4---C7---H7A 109.5 C1---C2---C8 122.39 (19) C4---C7---H7B 109.5 C3---C2---C8 121.60 (19) H7A---C7---H7B 109.5 C1---C6---C5 119.3 (2) C4---C7---H7C 109.5 C1---C6---H6A 120.3 H7A---C7---H7C 109.5 C5---C6---H6A 120.3 H7B---C7---H7C 109.5 ---------------- ------------- ---------------- ----------- Hydrogen-bond geometry (Å, °) {#tablewraphbondslong} ============================= -------------------- --------- --------- ----------- --------------- *D*---H···*A* *D*---H H···*A* *D*···*A* *D*---H···*A* N1---H1A···Cl1^i^ 0.89 2.27 3.136 (2) 164 N1---H1B···Cl1^ii^ 0.89 2.27 3.128 (2) 163 N1---H1C···Cl1 0.89 2.20 3.080 (2) 170 -------------------- --------- --------- ----------- --------------- Symmetry codes: (i) −*x*+1, *y*+1/2, −*z*+3/2; (ii) *x*, −*y*+1/2, *z*+1/2. ###### Hydrogen-bond geometry (Å, °) *D*---H⋯*A* *D*---H H⋯*A* *D*⋯*A* *D*---H⋯*A* -------------------- --------- ------- ----------- ------------- N1---H1*A*⋯Cl1^i^ 0.89 2.27 3.136 (2) 164 N1---H1*B*⋯Cl1^ii^ 0.89 2.27 3.128 (2) 163 N1---H1*C*⋯Cl1 0.89 2.20 3.080 (2) 170 Symmetry codes: (i) ; (ii) .

Background {#Sec1} ========== The need for an alternative to fossil fuels, with rising prices and declining reserves, has led to renewed interest in microalgae as a potential source for biofuel production \[[@CR1]-[@CR3]\]. Nutrient availability, temperature, and light intensity are the major environmental determinants of algal growth, reproduction, and morphology, including the accumulation of lipids in the form of triacylglycerols (TAGs) \[[@CR4]-[@CR6]\]. Photoautotrophically grown algae offer better solar energy conversion efficiency and a range of technical and ethical advances compared to traditional oil crops \[[@CR7]-[@CR9]\]. Considering their increasing importance as bioproducers and the need to achieve an optimized balance between lipid production and growth, thorough analyses of the underlying molecular mechanisms that mediate stress-induced accumulation of lipids in microalgae are necessary. Nevertheless, these analyses are still in a very early stage \[[@CR10]\]. Growth-inhibiting conditions and an imbalance between carbon and some macro- and micro- elements such as Fe, S, Zn, or N \[[@CR11]\] lead to metabolic rearrangements modulating cell division, morphology, and photosynthetic capacity \[[@CR12]\] and the accumulation of starch \[[@CR13]\] and/or lipids \[[@CR14]\] to protect cellular structures and increase the microalgae survival probability under adverse circumstances. The accumulation of lipid bodies in *Chlamydomonas reinhardtii* under N deficiency has been recently documented in detail \[[@CR4],[@CR14],[@CR15]\], establishing a well-known environment in which changes in morphology and some key genes are defined. These studies together with the availability of a sequenced genome \[[@CR16]\], proteomics and metabolomics protocols and databases \[[@CR17]-[@CR27]\], pathway annotations \[[@CR28]-[@CR30]\], and a wide range of molecular biology \[[@CR31]\] and transcriptomics tools \[[@CR32]\] make *C. reinhardtii* the premier molecular model for research in microalgae. The employment of recent advances in high throughput profiling methodologies has allowed the system-level characterization of *C. reinhardtii* at transcriptomic \[[@CR11],[@CR15],[@CR33]\], proteomic, and metabolomic levels \[[@CR19],[@CR21],[@CR34]\]. In the present study, we have added a further layer of investigation, in particular distinguishing short- and long-term adaptive mechanisms and the recovery phase of the cells from N starvation to normal vegetative growth. In contrast to previous studies on differential gene expression \[[@CR33],[@CR35]\] or metabolomics analyses \[[@CR34]\], we have studied N starvation and the following recovery process after N readdition during a four-day experiment. Using classical physiological measurements, mass spectrometry for quantitative proteomics (GeLC-LTQ-Orbitrap-MS) and metabolomic (GC-MS) changes, and mining available datasets we depicted the responses of *C. reinhardtii* to available N, showing the dynamic behavior of the biochemical pathways and metabolism to the N availability and providing new potential bioengineering targets for increased lipid accumulation. Results {#Sec2} ======= Physiological responses to nitrogen starvation and recovery in *Chlamydomonas reinhardtii* cells {#Sec3} ------------------------------------------------------------------------------------------------ *Chlamydomonas reinhardtii* cells show a high ability to adapt dynamically to environmental conditions. The stress adaption process is based in short- and long-term changes in metabolism affecting the morphological phenotype. Therefore, we have selected controls (0 h), three sampling times under N starvation (5 h-N5h, 24 h-N24h, 72 h-N72h), and two further samplings after N replenishment (77 h + N5h, 96 h + N24h), aiming to cover both short- and long-term responses for acclimation and recovery. N starvation leads to a block of growth (Figure [1](#Fig1){ref-type="fig"}), which is significantly slower than that in control cultures (Additional file [1](#MOESM1){ref-type="media"}: Figure S1). The fresh weight (FW) of the cultures was sustained during initial N starvation, and reduced after 72 h. Also, chlorophylls were affected, with a turn in the culture color from green to yellow under N starvation with a 30% decrease in Fv/Fm (Figures [1](#Fig1){ref-type="fig"} and [2](#Fig2){ref-type="fig"}). As expected, N starvation induced a quick accumulation of lipids, a 1.75-fold increase in 72 h (Figure [1](#Fig1){ref-type="fig"}), most of them in the form of lipid bodies (Figure [2](#Fig2){ref-type="fig"}). N starvation is affecting the normal physiological behavior of the cells, blocking cell growth and reducing photosynthesis. Cells stop dividing probably because there is no available N for sustaining protein and nucleotide biosynthesis. Under these circumstances intracellular N should be recycled to support critical life-supporting pathways. Photosynthesis and antennas are reduced to avoid oxidative cellular damages and changes in cellular pools. The excess of available energy and carbon is then channeled into an increased production of lipids. Lipids act not only as an energy and carbon sink, but also in membrane stabilization.Figure 1**Physiological measurements.** Variations in the photosynthetic rate (Fv/Fm) and in the accumulation of lipids, fatty acids, and fresh weight during the experimental time course. Different letters within series indicate significant differences (ANOVA followed by a TukeyHSD, *P* \< 0.05). Fresh weight and fatty acid values were normalized as a relative abundance considering the average contents in T0. Fv/Fm values were plotted without transformation.Figure 2**Lipid bodies accumulation during the time course experiment.** Nitrogen starvation causes a quick accumulation of lipid bodies (yellow dots) with a maximum density obtained after 72 h of culture. Notice that the autofluorescence of the cells is dramatically reduced at 72 to 77 h, indicating a sharp decrease in the pigment content. Lipid bodies disappear completely 24 h after the addition of ammonia (96 h). Cells were fixed in 1% formaldehyde and stored at 4°C until analysis. Cells were stained with Nile red to identify the lipid bodies. N replenishment quickly reverted these physiological adaptations, with total lipids, Fv/Fm, and chlorophylls returning to normal levels in 24 h (Figures [1](#Fig1){ref-type="fig"} and [2](#Fig2){ref-type="fig"}). The FW of cell cultures was increased 1.5- and 2.5-fold at 5 h or 24 h after N replenishment. The degradation of TAGs and also the recovered photosynthetetic activity probably provided the energy needed for this fast recovery of vegetative cell growth. These results support the hypothesis that there are sensing mechanisms that regulate not only N uptake and N assimilation, but also many other metabolic pathways (ranging from glycolysis to secondary metabolism) defining physiology and cell growth. An initial functional overview of the changes in the proteome (Additional file [2](#MOESM2){ref-type="media"}: Tables S1, S2) and metabolome (Additional file [2](#MOESM2){ref-type="media"}: Table S3) showed that N starvation affects most of the cellular metabolism (Figure [3](#Fig3){ref-type="fig"}, Additional file [3](#MOESM3){ref-type="media"}: Figure S2, Additional file [4](#MOESM4){ref-type="media"}: Figure S3, Additional file [5](#MOESM5){ref-type="media"}: Figure S4, Additional file [6](#MOESM6){ref-type="media"}: Figure S5, Additional file [7](#MOESM7){ref-type="media"}: Figure S6), that is, ammonia transport and fixation proteins, photorespiration, or oxidative pentose phosphate pathways. Differential pathways are depicted below.Figure 3**Representation of N starvation- and recovery-induced changes in major pathways and processes using MapMan visualization platform and our** ***Chlamydomonas reinhardtii*** **mapping.** Lines in gray represent individual proteins that were differentially accumulated during the experiment (*P* \< 0.05) at each category, thick red lines represent the average value within all of the clustered proteins at each time point, and thin red lines represent the average ± one standard deviation. Protein abundances were normalized as a percentage of the maximal value in the time series. We applied multivariate analyses for integrating physiological measurements with protein and metabolite datasets, reducing the complexity of the data, and performing untargeted correlation network analysis \[[@CR22],[@CR36]-[@CR38]\]. Principal component analysis (PCA) and partial least square discriminant analysis (PLS-DA) resulted in a similar classification of the samples (Figure [4](#Fig4){ref-type="fig"}A; Additional file [8](#MOESM8){ref-type="media"}: Figure S7a, b; Additional file [2](#MOESM2){ref-type="media"}: Tables S4,S5) while sparse partial least square (sPLS) analysis gave a slightly different picture (Additional file [8](#MOESM8){ref-type="media"}: Figure S7c; Additional file [2](#MOESM2){ref-type="media"}: Table S6). PC1 seems to explain the adaption to N starvation: ammonia transport and assimilation enzymes are positively correlated to this component, while N demanding activities, like polyamine and protein biosynthesis, and amino acid degradation are negatively correlated. Furthermore, the high correlation of central metabolism enzymes and cGMP-dependent kinases showed that N starvation also affects normal respiration. The high correlation of glycerol pathways and C18:2 indicated the importance of lipid biosynthesis during N starvation. PC2, on the other hand, explains the N-starvation recovery and growth on the basis of an increased energy production through lipid degradation, and the recovery of the pigments biosynthesis. These findings support our previous hypothesis, based on physiological parameters, of how N starvation affects the cell function since the previously described effects can be explained at a biochemical level based on PC1 and PC2 (plotting other components does not improve the separation between groups (Additional file [8](#MOESM8){ref-type="media"}: Figure S7d, e, f)). Biclustering of functional categories of protein changes (Figure [4](#Fig4){ref-type="fig"}B) against the time points of N starvation and N readdition demonstrated impressively that the recovered cells are more similar to the control cells and that there is a complete remodeling of cellular processes during starvation and recovery. Clusters of short- and long-term responsive functional processes can be identified. Furthermore, we have used sPLS networks in which protein and metabolites were used as predictive variables to explain the physiological parameters aiming to establish interactions between proteins, metabolites, and phenotype (Figure [4](#Fig4){ref-type="fig"}C, Additional file [6](#MOESM6){ref-type="media"}: Figure S5D, Additional file [9](#MOESM9){ref-type="media"}: Figure S8, Additional file [10](#MOESM10){ref-type="media"}: Figure S9).Figure 4**Classification of the different samples according to multivariate methods.** **(A)** Principal componet analysis (PCA) of the integrated proteomic, metabolomic, and physiological datasets. Glutamate family enzymes (ASS, ASL, OCT, NAG), ammonia metabolism (Cre13.g592200.t1.2), purine biosynthesis proteins (Cre07.g318750.t1.2, Cre08.g364800.t1.2), NADH:ubiquinone oxidoreductase, cGMP-dependent kinases (Cre03.g199050.t1.2), glycolysis enzymes (PK, GAP-DH, PEPC), glyceraldehyde-3P-DH, glycerol, and C18:3 showed high correlations to PC1. Calvin cycle proteins (SBPase, PPE), chloroplastic ATPase, amino acid degradation, polyamine synthase, fatty acid elongation, catalases, and aspartic acid showed a negative correlation to PC1. Fresh weight, alanine, beta oxidation-related proteins (Acyl-CoA oxidases, HADH), oxidoreductases (Cre16.g677950.t1.3, g13806.t1, g4488.t1, g9426.t1), signal peptide and protein peptidases, and tetrapyrrole biosynthesis proteins showed a high correlation to PC2, while organic acids (fumaric and glyceric), phosphate, and photosynthesis-related enzymes (light reaction and carbon fixation), showed a negative correlation. These variables were used to infer the biological meaning of the principal components 1 and 2. Loading matrix is available in Additional file [2](#MOESM2){ref-type="media"}: Table S4. **(B)** Hierarchical clustering and heatmap of the analyzed proteins grouped by functional category according to MapMan. Three different clusters (0, 5 h, and 96 h, 24 h, and 72 and 77 h) can be distinguished, showing the different degrees of response to N starvation. The aggrupation of the 96-h and 5-h samples remarks the effect of the N repletion over the cultures. **(C)** sPLS-based network, showing the significant interaction between proteins and metabolites. Three major clusters can be distinguished, corresponding to fresh weight, glycerol/C18:2/C16, and N metabolism, which are mixed in the image. Fv/Fm and C16 are outside these groups acting as a link. For further details see [Results](#Sec2){ref-type="sec"} and [Discussion](#Sec11){ref-type="sec"}. This complex network is depicted in Additional file [6](#MOESM6){ref-type="media"}: Figure S5D, Additional file [10](#MOESM10){ref-type="media"}: Figures S9 and Additional file [11](#MOESM11){ref-type="media"}: Figure S10. Dynamics of 1,658 proteins and 52 metabolites during nitrogen starvation and recovery {#Sec4} ------------------------------------------------------------------------------------- The application of mass-spectrometry-based profiling of proteins and metabolites allowed a comprehensive analysis of the responses to N starvation and N readdition. More than 15,000 peptides and 3,200 proteins were detected in 552,529 spectra obtained from whole cell and nuclei protein extracts. 1,658 proteins were above the minimum abundance threshold for confident quantitation (Additional file [2](#MOESM2){ref-type="media"}: Tables S1 and S2). This number represents about 15% of the genes in the *C. reinhardtii* genome \[[@CR16]\]. GC-MS allowed the unequivocal identification of 52 primary metabolites after comparison to reference standards (Additional file [2](#MOESM2){ref-type="media"}: Table S3) and quantified 60% of them with differential accumulation, at least in one of the sampling times. The use of proteomics- and metabolomics-based approaches provides a direct readout of the metabolic and physiological adaptive mechanisms that are present in the cell and links molecular dynamics to genome-based theoretical metabolic networks \[[@CR39],[@CR40]\]. The reconstruction of the *C. reinhardtii* network based on KEGG Orthology (KO) annotation consists in 7,330 reactions belonging to 263 pathways, which are catalyzed by 713 enzyme classes \[[@CR23]\]. Of the total proteins described in this study, 845 were annotated to enzyme commission (EC) numbers using Biomart, the KEGG enzyme database, and manual annotation during the curation process. These enzymes defined 447 classes of reactions belonging to 157 pathways, constituting more than 50% of the modeled class reactions and pathways in *C. reinhardtii* \[[@CR23],[@CR41]\]. We have further analyzed the total proteome using MapMan \[[@CR30],[@CR42]\] and a custom bin map for the release 5.3 of the *Chlamydomonas* genome. A total of 1,233 proteins were assigned to functional bins. Employing this classification tool allowed us to plot the dynamics of the differentially accumulated proteins (*P* \<0.05) to illustrate the overall changes in the metabolome (Figure [3](#Fig3){ref-type="fig"}), and also to focus specifically on photosynthetic (Additional file [3](#MOESM3){ref-type="media"}: Figure S2), N (Additional file [4](#MOESM4){ref-type="media"}: Figure S3, Additional file [5](#MOESM5){ref-type="media"}: Figure S4), lipid (Additional file [6](#MOESM6){ref-type="media"}: Figure S5), and nucleotide metabolism (Additional file [7](#MOESM7){ref-type="media"}: Figure S6). The integration of the generated data with the available datasets with transcriptomic \[[@CR33]\] and metabolomic \[[@CR34]\] data in the overlapping experimental points (0 and 72 h) allowed a robust and comprehensive analysis of the molecular changes induced by N starvation. However, 166 differentially accumulated proteins of the total protein dataset were not mapped to any pathway, indicating the need for a continuous improvement of functional annotation. Response of *Chlamydomonas reinhardtii* primary metabolism to N depletion and recovery {#Sec5} -------------------------------------------------------------------------------------- N starvation induces quick and profound changes in primary metabolism (Figure [3](#Fig3){ref-type="fig"}, Additional file [3](#MOESM3){ref-type="media"}: Figure S2, Additional file [4](#MOESM4){ref-type="media"}: Figure S3, Additional file [5](#MOESM5){ref-type="media"}: Figure S4, Additional file [6](#MOESM6){ref-type="media"}: Figure S5, Additional file [7](#MOESM7){ref-type="media"}: Figure S6). Glyceraldehyde-3-P-dehydrogenase (GAPDH) showed maximum peaks at 72 h. Interestingly, one pyruvate kinase isoform (Cre06.g280950.t1.2) was only detectable after N depletion, showing a drop in abundance 24 h after N repletion. At the same time pyruvate increased after N repletion. Fructose bisphosphate aldolase (FBA2) and glucose-6-P-isomerase (PGIC) showed a peak after 5 h and were then repressed until N repletion. This repression might lead to the accumulation of the glucose-6-P-pool which could be channeled into starch synthesis. Indeed, starch accumulates during N starvation \[[@CR4],[@CR13]\]. Starch synthases (SS1, SS2, SS4, GBSS1) were upregulated during starvation, and quickly decreased after N repletion. The sugars fructose, glucose, and trehalose are accumulated during starvation (+15-, +5-, and +9-fold, respectively), being reduced after N repletion. Trehalose has been described as a substitute of N-containing compatible solutes \[[@CR43]\] and a positive regulator of starch biosynthesis \[[@CR44]\]. More recently a key role of trehalose in stress responses by regulating the SNF/bZIP system was postulated \[[@CR45]\]. Glucose and fructose can also act as regulators of this system involving hexokinase (HXK) and potentially sucrose and trehalose metabolism \[[@CR46]\]. However, sucrose metabolism is not clarified completely in *Chlamydomonas reinhardtii*. One gene which is at least assigned to sucrose metabolism is sucrose phosphate phosphatase (SPP). HXK and SPP abundances were already increased 24 h after N starvation, showing a maximum peak after 72 h, being quickly downregulated after N repletion. On the other hand, the pool of glucose-6-P can also enter the oxidative pentose phosphate pathway. The rate-controlling enzyme of this pathway glucose-6-P-dehydrogenase (G6PDH) and 6-phosphogluconate dehydrogenase (6PGDH) showed an increased abundance under N starvation with a peak 5 h after N repletion (+3 to 5fold). In contrast, nonoxidative enzymes, transaldolase and cytosolic ribose-5-P-isomerase, showed an opposite trend with a minimum presence after 24 h of stress, potentially reducing the synthesis of ribulose-5-P (see discussion below). To extend the classical correlation analysis, we applied Granger causality analysis (see also \[[@CR22],[@CR38],[@CR47],[@CR48]\]). Here, time-lagged correlations are exploited with the potential that pairs of precusors and products can be identified. The Granger causality function is implemented in the COVAIN toolbox \[[@CR22],[@CR49]\]. The COVAIN function also applies a Benjamini-Hochberg correction to reveal only significant correlations (for further information see the COVAIN manual; <http://www.univie.ac.at/mosys/software.html>). Because of the low number of time points, only a few Granger causalities have been identified and they have to be treated carefully. For example, in Figure [5](#Fig5){ref-type="fig"}A Granger causality showed a steady increase of NADH:ubiquinone reductase (Cre09.g415850.t1.2) that was negatively correlated with one of the precursors of NAD-biosynthesis nicotinic acid (Figure [5](#Fig5){ref-type="fig"}A), pointing to an increase in mitochondrial activity. The citric acid cycle (tricarboxylic acid, TCA) enzymes were in general upregulated, with a twofold increase for most of the proteins after 72 h, and a quick return to control levels after N repletion. Both succinate dehydrogenase (SDH2) and pyruvate dehydrogenase (PDH1) behaved the same. The abundance of TCA intermediates was not consistent with a unified response of this pathway. Pyruvate concentration is reduced during N starvation while malate is accumulated. Succinate and oxalate levels are constant during the whole experiment. The malic enzyme (ME3) also remains at constant levels during N starvation, but after repletion its abundance is increased sevenfold, which is correlated to a fivefold decrease of malate abundance. The upregulation of the TCA cycle is also functional for the production of carbon backbones for amino acid biosynthesis by anaplerotic reactions.Figure 5**Examples for identified Granger causalities.** Metabolite, protein, and physiological data were combined and analyzed with respect to time-shifted correlations using the toolbox COVAIN \[[@CR38]\]. This procedure, called Granger causality, allows for the identification of directed correlations or nonlinear processes. Therefore, these associations must be carefully interpreted. **(A)** The metabolite nicotinic acid (NA) shows a time-shift behavior with respect to an NAD-reductase. Because NA is an intermediate in NAD biosynthesis, the increased levels of NAD-reductase may be involved in the consumption of NA. **(B)** A clathrin assembly protein involved in vesicle formation (COP II, see also [Discussion](#Sec11){ref-type="sec"}) follows with a time-shift the accumulation of total lipids. After readdition of N, the protein concentration declines. This protein is also highly correlated to a major lipid droplet protein, glycerol, and various fatty acids, indicating its potential role in the formation of lipid bodies during N starvation (for further details see [Results](#Sec2){ref-type="sec"} and [Discussion](#Sec11){ref-type="sec"} sections). Most of the proteins related to oxidative phosphorylation are accumulated during N starvation. NADH-DH isoforms, ATP synthases, and cytochrome c oxidase (COX) increased up to seven-, two-, or eightfold, respectively (Additional file [2](#MOESM2){ref-type="media"}: Table S1). This response is not shared with other nutrient depletions such as sulfur \[[@CR50]\] or temperature stress \[[@CR51]\]. N starvation induces a decrease in abundance of proteins involved in photosynthesis and increased carbon concentrating mechanism (CCM) {#Sec6} -------------------------------------------------------------------------------------------------------------------------------------- A sharp decrease of light-dependent reactions proteins (Additional file [3](#MOESM3){ref-type="media"}: Figure S2) followed N starvation. The reduction of cytochrome b6f complex (negative fivefold), ATP synthases (negative twofold), and photosystem II (PSII) proteins may explain at least partially the observed reduction in photosynthesis, confirming the observations of Majeran *et al.* \[[@CR52]\], which showed a degradation of cytochrome b6f and light harvesting complexes (LHCs) after N starvation. Interestingly, although only two PSII proteins reached control levels, the Fv/Fm rate was fully recovered 24 h after replenishment. These results showed that, as proposed by Plumley and Schmidt \[[@CR53]\], the efficiency of PSII during N starvation is not mainly limited by protein content but by other compounds such as chlorophylls. This was supported phenotypically and molecularly with a sharp recovery of green cells in the culture and the upregulation of the tetrapyrrole-related biosynthetic pathways (Figure [3](#Fig3){ref-type="fig"}). Other pathways such as the cyclic electron flow have been recently described as essential mechanisms for photosynthesis \[[@CR54]\] to provide ATP in situations of stress \[[@CR55]\]. Cyclic electron flow-related enzymes showed a peak after 72 h (up to threefold), being then reduced after repletion, confirming its essential role as initially suggested by Peltier and Schmidt \[[@CR56]\]. Most of the enzymes belonging to the CO~2~ fixation pathway, like RuBisCO large subunit (RBCL), FBPase, transketolase, or phosphoglycerate kinase, were reduced under N starvation (Additional file [3](#MOESM3){ref-type="media"}: Figure S2) as previously pointed out in \[[@CR34]\]. Other proteins such as RuBisCO activase (RBCA) or phosphoribulokinase (PRKA) were increased with time, with a maximum abundance peak measured after 72 h in the case of PRKA. Carbonic anhydrases (CAH1, CAH3, CAH7, and CAH8) belonging to CCMs are differentially regulated during N starvation. CAH3 decreased in abundance; conversely, CAH7 and CAH8 increased in abundance, peaking at 72 or 77 h. Phosphoenolpyruvate carboxylases (PEPC1, PEPC2) were not detectable when N was present in the media (0 and 96 h) and its maximum abundance was detected at around 72 h. Apparently, *C. reinhardtii* is highly flexible to maintain CO~2~ fixation, triggering specific enzymes depending on the type of stress, as cold stress was characterized by increased PEPC2 and CAH3 \[[@CR22]\]. We have not detected a good correlation between changes of photosynthesis-related proteins and its transcripts (r^2^ = 0.166), showing that the specific regulation of protein species abundance does occur at the translational level, but also by posttranslational and degradation mechanisms. Nitrogen metabolism quickly adapts to an environment without ammonia {#Sec7} -------------------------------------------------------------------- N starvation leads to an increased abundance of specific mRNAs, proteins, and metabolites \[[@CR15],[@CR33],[@CR34],[@CR57]\]. In our study we used the strain CC503, a *nit1 nit2* mutant lacking an active nitrate signaling pathway (*nit2*). As expected, no differences within nitrate transporters, nitrate, and nitrite reductases (Additional file [4](#MOESM4){ref-type="media"}: Figure S3) were detected. By contrast, proteins involved in NH~4~^+^ assimilation were overaccumulated: two ammonia transporters (g3261.t2, Cre13.g569850.t1.2) were induced by N starvation. Interestingly, Cre13.g569850.t1.2 accumulation continued with a maximum abundance 24 h after N repletion (a threefold increase compared to 24 h of N starvation), indicating a long responsive time in this element. Enzymes of the Glu-Gln (GS/GOGAT) system were also upregulated, and the abundance of all glutamine synthase (GS) isoforms increased five- to tenfold, just to decrease quickly to control levels after N repletion. On the other hand, glutamate synthase (GSN) isoforms behave differentially: GSN1 abundance is stable, while GSN2 was induced by N starvation. Glutamate dehydrogenases (GDH1, GDH2), involved in N degradation, showed a tenfold decrease starting to respond to N repletion after 24 h (slower than GS and GSN enzymes). The correlation of fold change at transcript and protein levels (r^2^ \> 0.8) points to the transcriptional regulation of this system. The central amino acid metabolism and some degradation pathways were decreased (two- to threefold) after N starvation (Figure [3](#Fig3){ref-type="fig"}). Conversely, the glutamate family pathway quickly responded to the availability of N, being strongly upregulated by N starvation with a fourfold increase in N-acetyl-gamma-glutamyl-phosphate reductase (AGPR), arginosuccinate lyase (AGS1) and synthase (ARG7). Furthermore, three enzymes of the branched amino acid biosynthesis (acetolactate synthase 1 and 2, ALS1, ALS2; acetohydroxy acid isomeroreductase, AAI1) showed abundance peaks at 5 h and 77 h. Most of the amino acids were tightly connected in the sPLS correlation network (Additional file [11](#MOESM11){ref-type="media"}: Figure S10) with glutamate and its derivative pyroglutamic acid among the central nodes. These nodes also linked proteins involved in a wide range of activities and functions, showing the diversity of the responses triggered by N starvation. Cytosolic ribosomes are accumulated during N starvation, while chloroplastidic ribosomes are degraded {#Sec8} ----------------------------------------------------------------------------------------------------- A general ribosomal degradation and resynthesis is classically associated to N starvation \[[@CR58]\]. These observations are partially supported, at least for chloroplast ribosomes, by a recent high-throughput transcriptomic study \[[@CR33]\]. We detected a differential behavior between chloroplast and cytoplasmic ribosomes (Additional file [5](#MOESM5){ref-type="media"}: Figure S4). Evidence of protein accumulation was observed for cytosolic ribosomes, with a two- to tenfold increase in the proteins corresponding to 40S. In the case of chloroplastidic ribosomes, the abundance of both subunits decreased after N starvation, being downregulated by more than 40% of the quantified proteins. However, these proteins quickly responded to N replenishment, recovering its initial abundance values in 5 h. Our data refine the current model of ribosome recycling following N starvation, in contrast to previous theories pointing to an untargeted ribosomal degradation-resynthesis. Here, we rather observe a targeted accumulation of cytosolic and a degradation of chloroplastidic ribosomes. Furthermore, the differential behavior of cytosolic, increased, and chloroplast, decreased, ribosomes suggests a more active role of the nuclear encoded proteins for adapting to N starvation. Response of lipid metabolism to N starvation {#Sec9} -------------------------------------------- N starvation led to the accumulation of lipids (Figures [1](#Fig1){ref-type="fig"} and [2](#Fig2){ref-type="fig"}) mainly in the form of TAGs \[[@CR10]\]. In consequence we expected to detect upregulated lipid biosynthetic pathways but, surprisingly, most of the enzymes of the fatty acid biosynthesis were not changed after starvation (Additional file [6](#MOESM6){ref-type="media"}: Figure S5A). The key enzymes 3-ketoacyl-CoA-synthase 1 (KAS1), which catalyzes the reaction of acetyl-ACP with chain-extending malonyl-ACP, and one isoform of enoyl-ACP-reductase (Cre06.g294950.t1.3), which catalyzes the last step of the fatty acid biosynthesis, are downregulated. This behavior was previously shown at protein level \[[@CR59]\]. The downregulation of two triglyceride lipases (Cre01.g002400.t1.3, g9707.t1) suggests that this decrease compensates the decrease in lipogenic enzymes allowing an effective lipid accumulation. At transcript level a modest increment (less than +2-fold) of biosynthetic genes of fatty acids under N deprivation was observed \[[@CR33]\]. An undescribed monogalactosyl diacylglycerol synthase annotated to locus g14367.t1 showed a +3-fold change following N starvation. The product of this reaction could serve as a substrate for the desaturation of oleic and linoleic acid \[[@CR60]\] while also controlling the cell proliferation by blocking replicative DNA polymerase as described in animals \[[@CR61]\]. The pool of glycerol, essential for the formation of TAG, was significantly increased after N starvation (+4-fold; Additional file [2](#MOESM2){ref-type="media"}: Table S3), and can be explained by the accumulation of glycerol-3-phosphate dehydrogenase/dihydroxyacetone-3-phosphate reductases (GPD2, GPD4) (Additional file [6](#MOESM6){ref-type="media"}: Figure S5B). These enzymes showed a high correlation to glycerol in the interaction network (Additional file [6](#MOESM6){ref-type="media"}: Figure S5C). This network also links acetate, glycerol, and C18:2, oxidoreductases, ribosomal proteins, and a major lipid droplet protein (Cre09.g405500.t1.3) \[[@CR62],[@CR63]\]. This protein has a high similarity to a major protein in lipid droplets in *Haemotococcus pluvialis* and *Dunaliella* (BLAST e-values of 10^−84^ and 10^−46^, respectively, NCBI-nr database). Cre09.g405500.t1.3 was not detected at 0 h, but from 5 h to 72 h it increased its abundance up to 14-fold. Proteins for NH~4~^+^ transport and assimilation and specific signaling proteins (BSU1) were also part of this network. These results demonstrate the power of the integrative networks based on sPLS-correlation for discovering new interactions between proteins and metabolites, allowing the capability to also accurately associate uncharacterized proteins to functional clusters. N starvation leads not only to the accumulation of lipids, but also to the change of the total lipid composition of the cells. We have studied 10 long chain fatty acids (Additional file [2](#MOESM2){ref-type="media"}: Table S3, Additional file [6](#MOESM6){ref-type="media"}: Figure S5D). Four of them, C16:3, C16:0, C18:2, and C18:3 (9,12,15), were significantly different in our experiment. C16:0, C16:3, and C18:2 were accumulated, while C18:3 (9,12,15) was reduced. Chloroplastidic desaturase ∆12 (CDD12, Cre13.g590500.t1.2) is eightfold reduced at transcript level, and may explain the reduction in C18:3. Stearoyl-CoA desaturase ∆9 (SCD, Cre09.g397250.t1.2) transcripts are twofold increased during N starvation and may also explain the increase of C18:2. How this system is regulated is not clear, and cannot be elucidated alone at the protein or transcript level. N replenishment causes a quick activation of beta-oxidation pathways, with increased levels of five acyl-CoA-oxidase isoforms and also the recovery of the levels of the lipase Cre01.g002400.t1.3. Nitrogen starvation induces significant changes in the nuclear proteome {#Sec10} ----------------------------------------------------------------------- To further investigate changes at the nuclear regulatory level we first studied nuclei-enriched fractions of *C. reinhardtii* (Additional file [2](#MOESM2){ref-type="media"}: Table S2). Secondly, we performed an *in silico* analysis of total cell and nuclear fractions focusing on nuclear proteins related to signaling and transcriptional regulation and finding 268 proteins (Additional file [2](#MOESM2){ref-type="media"}: Table S8), from which 136 were differentially expressed (*P* ≤0.01). As expected, the levels of histones and other core proteins did not significantly change with time, whereas other chromatin-interacting proteins followed different dynamics. N starvation induced an initial decrease in RNA polymerase I and RNA helicases, which were partially recovered after repletion. On the other hand, N starvation triggered the accumulation of DNA binding proteins, such as Cre06.g252000.t1.2, a leucine/zipper transcription factor, which was quickly silenced again after N replenishment (a fivefold reduction in 5 h). An Argonaute-like protein (AGO; Cre04.g214250.t1.3) was also induced by N repletion, suggesting a role of the siRNA system \[[@CR64]\]. The DEAD/DEAH box helicase Cre01.g021600.t1.2 is also accumulated in the absence of N in the medium, indicating that some SWI/SNF regulation pathway may be occurring. This class of enzymes and regulation has been described to be part of the abiotic stress response mechanism \[[@CR65]\]. We have used the Plant Transcription Factor database \[[@CR66]\] to mine the *Chlamydomonas* genome release 5.3 for finding transcription factors. However, many of these postulated transcription factors were not detected in the nuclear-enriched fractions, pointing to its low abundance and the technical challenges for high-throughput identification. To date most of these proteins are uncharacterized, and even for the known families their members can have different regulatory roles \[[@CR67]\]. N starvation is known for triggering gametogenesis and sexual reproduction in *C. reinhardtii* \[[@CR31]\], but we have not detected related proteins, probably because they are expressed at very low levels or they are not functionally annotated. Discussion {#Sec11} ========== *C. reinhardtii* cells respond to changes in the availability of ammonia by drastically changing their metabolism and normal development. In this work we analyzed the changes in the metabolome and proteome, also integrating other available datasets, thus presenting the most complete and comprehensive overview of the response to total available nitrogen and its recovery. The design of this study, together with the employed analytical techniques, allowed us to update current knowledge, providing new insights for the understanding of the complex metabolic dynamics that follow N starvation and, for the first time, recovery of vegetative growth after N replenishment, by using high-throughput omics analyses. Nitrogen/carbon balance defines the metabolic switch under N starvation {#Sec12} ----------------------------------------------------------------------- The sensing of the balance between N and C is transduced into a change in N- and C-responsive pathways, leading to a metabolic adjustment that dominates many pathways and defines the observed physiological and phenotypical changes. However, it has to be considered that the presence of acetate in the medium will change the cellular sensitivity towards the maintenance of favorable C/N ratios \[[@CR68]\], causing increased responses to the oxidative stress as a side effect of the downregulation of photosynthesis and carbon assimilation \[[@CR69]\]. N starvation also leads to increased ammonia uptake and assimilation proteins and a decreased nucleic acid and protein biosynthesis, which may be the cause for the reduced cell growth. The reduced photosynthesis is proposed to be mediated by the degradation of LHCs and cytochromes, rather as an adaptive mechanism to stress and energy contents than as a way to recycle nitrogen. The reduction of PSII activity would lead to the production of less reactive oxygen species, but at a cost of reducing ATP and NADPH production. But even in a situation of reduced growth, as observed under N starvation, energy is still necessary to sustain metabolism. We detected an increase of cyclic electron flow proteins which pump H^+^ to the thylakoid lumen in order to increase ATP production. Furthermore, and based on the metabolic reconstruction, an increased pool of glucose-6-P was predicted. Part of this glucose-6-P can join the initial reactions of the oxidative pentose pathway. The corresponding enzymes were accumulated, rather than glycolytic enzymes. In contrast, nonoxidative enzymes, transaldolase, and cytosolic ribose-5-P-isomerase showed an opposite trend with a minimum presence after 24 h of stress, potentially reducing the synthesis of ribulose-5-P. The pool of ribose-5-P could enter the chloroplastic reductive pentose phosphate pathway to produce 3-phosphoglycerate, which can continue with the glycolytic pathway. The result of this reprogramming could be the production of enough pyruvate and NADPH + H^+^ to support biosynthetic processes at a cost of half ATP payback compared to glycolysis. This hypothesis was supported by the accumulation of phosphoribulokinase (PRK1), which followed the same abundance trend as G6PDH and 6PGDH. These changes in protein abundance are coincident with the changes in the mRNA expression levels given by \[[@CR33]\] in an N-depletion study; furthermore, \[[@CR34]\] recently described the increased accumulation in the metabolites 3-phosphoglycerate and ribose-5-P during N starvation. This remodeling, although energetically unfavorable, can be used by the cell to maintain adequate levels of NADPH + H^+^, which are required to maintain an increased lipid production. However, future studies applying flux analyses are necessary to investigate these processes in more detail. The functional role of cyclic electron flow and the quick recovery of PSII efficiency (correlated rather to pigment biosynthesis than to the recovery of PSII proteins) indicate that photosystems and electron transfer chain were fully functional during N starvation, even though the abundance of some subunits and cytochromes was reduced. Athough CO~2~ fixation enzymes were in general reduced, the abundance of CCM proteins CAH7, CAH8, and PEPC1 increased. CAH7 was only detected after a long exposure to N depletion (72 h) and may explain why this enzyme was not detected in previous studies limited in time to not more than 48 h \[[@CR33],[@CR34]\]. CAH7 and CAH8 are similar and both of them contain a hydrophobic chain which localizes them in periplasmic space associated to the plasma membrane, probably covering decreased CAH1 and 2 activities and transferring inorganic CO~2~ to a transport system or pore \[[@CR70]\]. The specific role of these CAHs under nutrient deprivation remains undetermined and cannot be easily explained at this time. By contrast, the kinases belonging to the Snrk2-SNF2 family have an essential role in the control of gene expression through the activation of bZIP transcription factors and SWI/SNF chromatin remodeling complexes \[[@CR71],[@CR72]\]. These were quickly responsive to N starvation in our experimental system. Kinases of this family control the lipid accumulation in yeast \[[@CR73]\], and have been linked to increased tolerance to N deficiency in *Arabidopsis* \[[@CR74]\] or S in *C. reinhardtii* \[[@CR50]\]. SNF families, also implied in energy sensing and gene regulation \[[@CR75]\], control the metabolic response to other stresses like cold acclimation, inducing the accumulation of sugars and starch \[[@CR51]\]. The accumulation of the signaling molecules glucose-6-P and trehalose \[[@CR45],[@CR76]\] and the specific dynamics of HEX and SPP found in the N starvation, provide an extra support to this hypothesis. Branched chain amino acids linked N metabolism and cell growth {#Sec13} -------------------------------------------------------------- The unavailability of N caused a rearrangement of amino acid metabolism at the same time. The pool of glutamate, essential for ammonia assimilation, was drastically reduced after starvation, triggering aspartate aminotransferases and citrate synthases from the oxoglutarate pathway. Enzymes belonging to branched chain amino acids (BCAs) were negatively correlated to the abundance of Val, Ile, and Leu, which indicates that enzyme abundance is increased to compensate the lower abundance of BCAs. Leu and Ile were significantly correlated (r^2^ \> 0.9) to Cre16.g687350.t1.2, an acyl-CoA oxidase and mediator of amino acid degradation, and a monofunctional catalase which may be controlling the oxidative damage. Interestingly, the same interaction cluster was found when analyzing the sPLS-based network; it was the bridge that linked FW and amino acid-nitrogen clusters. The increased abundance of carbamoyl phosphate synthetase subunits (CMP1, CMP2) and the dynamics shown by arginine and ornithine suggest the importance of the urea cycle for recycling the ammonia originated in the protein breakdown. Are glycerol metabolism, MLDP, and COP II proteins key elements for lipid droplet formation and TAG accumulation? {#Sec14} ----------------------------------------------------------------------------------------------------------------- The increased abundance of total lipids and lipid droplets following N-deprivation could indicate that the TAG of the droplets is synthesized de novo. The expected increase of the fatty acid biosynthetic pathway was, in general, not present but two Acetyl-CoA-Synthetases (g1224.t1, Cre12.g507400.t1.2) required for the activation of fatty acids and forming part of lipid droplets \[[@CR62]\] were significantly accumulated. Glycerol metabolism, needed to provide the backbone for the glycerolipid biosynthesis, was significantly responsive, with a combined accumulation of the proteins of this pathway of more than 50-fold. Cre09.g405500, a major lipid droplet protein (MLDP) \[[@CR62],[@CR63]\], increased 14-fold under N starvation and was linked to C18:2 and glycerol, showing further correlation (\>0.93) with clathrin and COP II, which forms vesicle coats and allows liposome fusion \[[@CR77]\]. The MLDP protein was identified by Moellering and Benning and by James *et al.* after analyzing lipid bodies associated proteins \[[@CR62],[@CR78]\]. Recently Wase *et al.* described a similar behavior of this protein under N deprivation, showing a 4.1-fold change after six days of culture without nitrogen \[[@CR59]\]. MLDP is highly correlated with lipid accumulation during N starvation and dropped completely after N replenishment. MLDP and other proteins linked by sPLS and Granger causality analysis such as COP II provide hub proteins in the adaptation process to N starvation and revovery (see Figure [5](#Fig5){ref-type="fig"}B). These proteins are involved in vesicle formation, suggesting their strong role in lipid body formation, thereby confining the TAGs into storage structures and promoting their accumulation. Lipid body formation can be also influenced by GTPases as COP II protein is inactivated by GTPases \[[@CR79]\], which we also found to be decreased during N starvation. Correlation networks need to be further analyzed since they also reach fatty acid and starch biosynthetic enzymes, pointing to key processes and promising targets for strain engineering. The increased C18:2 and the consequent alteration of the different fatty acids suggest a remodeling of the membranes, and thus fatty acids in membranes seem to be recycled into TAGs. TAG lipases and phospholipases were downregulated. N repletion led to an increased beta-oxidation for rapid lipid degradation and also increased the abundance of five acyl-CoA-oxidases. These were not correlated to enzymes of the BCA metabolism, demonstrating differential regulatory pathways. The lipase Cre01.g002400.t1.3 that was recovered after N replenishment was negatively correlated (\>\|0.9\|) to the translation initiation inhibitor (Cre12.g551350.t1.2) and a protein showing a DNA methyltransferase domain (Cre12.g508050.t1.2). This might indicate that the transcriptional response to N starvation is regulated not only by transcription factors, but also by epigenetic mechanisms, as suggested by \[[@CR50]\]. Can N recovery help us to understand lipid accumulation or metabolism and provide new targets for biotechnological improvement of oil production? {#Sec15} ------------------------------------------------------------------------------------------------------------------------------------------------- The study of specific changes occurring 5 h and 24 h after N resupply provides specific targets that could have potential use for bioengineering applications. Some of them have been depicted above (PEPC, CAH); however, there are specific targets that can be highlighted. One of the possible ways of intervention involves the enzymes and proteins related to lipid body stabilization and degradation with MLDP a key protein in the lipid biogenesis network. MLDP has been studied extensively and, despite its silencing by RNAi, does not lead to an increased lipid accumulation; however, the droplet size is significantly increased \[[@CR62]\]. This is a surprising result since our data, and also that of \[[@CR54]\], points out that the abundance of this protein is positively correlated to lipid accumulation. The combined manipulation of this protein together with beta oxidation, silencing acyl-CoA oxidases (especially Cre16.g687350.t1.2, Cre11.g467350.t1.2) and lipase (Cre01.g002400.t1.3), could result in an increased accumulation of lipids. Lipid transporters also play a key role in lipid metabolism, with Cre15.g641200.t1.2 a mitochondrial fatty acid transporter as a responsive candidate. g13764.t1, Cre13.g573150.t1.3 (hydroxylases) and Cre11.g467350.t1.2 (acyl-CoA oxidase) were accumulated (hydroxylases) or repressed (acyl-CoA oxidase) under N starvation, following the same trend as its transcripts. Interestingly, these genes were not reported to be significantly expressed under other stresses such as C, S, or Fe deprivation or oxidative stress according to AlgaePath \[[@CR80]\]. However, finding central metabolism enzymes related to lipid biogenesis and only responsive to N starvation (when comparing with the available NGS/proteomics datasets) is particularly difficult. Furthermore, modifying the gene expression of a number of genes is tricky, and a perfect flux model should be available in order to avoid unintended effects. Previous attempts to increase lipid accumulation by targeting specific genes have not been completely successful \[[@CR10]\]. This indicates that lipid accumulation is the result of a complex regulatory network linking cellular processes such as vesicle formation as well as processes of central metabolism such as G6P, branched chain amino acids, and energy metabolism. This is clearly demonstrated by the system-level analysis of our study. In this sense, our results and other proteomic and transcriptomic datasets \[[@CR23],[@CR33]-[@CR35],[@CR50],[@CR59]\] show many complementary genes/protein clusters responsive to stress \[[@CR80]\]. This suggests that there are major points of metabolic regulation based on common signaling elements that can be considered as potential targets for increasing lipid accumulation. Numerous DNA and RNA binding proteins and helicases showed differential accumulations in our assay (described above), but these candidates could not be easily proposed as biotechnological targets since their interaction network is unknown. On the other hand, the BRI1 suppressor (Cre01.g050850.t1.2), an inhibitor of BRI1, a receptor-like kinase which is responsive to brassinosteroids located both in plasma and in nuclear membranes \[[@CR81]\], was correlated to glycerol and C18:2, and it is known to initiate a signaling cascade leading to regulation of gene expression in the nucleus through BZR/BES proteins \[[@CR82]\]. BZR is known to block the metabolic switch in response to P deprivation \[[@CR83]\] and other abiotic stresses \[[@CR82]\] in *Arabidopsis* and also inhibits chloroplast development \[[@CR84]\], so a similar effect could be expected in *Chlamydomonas*. Thus, blocking not only the receptor, but also the transcription factor BZR by 14-3-3 proteins may be needed to respond to N starvation. In our dataset a 14-3-3 protein responding to N deprivation and recovery was quantified (Cre06.g257500.t1.2). The BRI1 suppressor was only detected when N was not present in the medium, showing a quick adaptive response to environment. FtsH chloroplast metalloproteases are closely related to development \[[@CR85]\], stress responses \[[@CR86]\], and chloroplast function \[[@CR87]\], also regulating lipid degradation in bacteria \[[@CR88]\]. We found three metalloproteases (Cre12.g485800.t1.2, Cre17.g720050.t1.2, g14586.t1) that were downregulated when N was absent from the media. Interestingly, and according to \[[@CR74]\] and considering all available datasets, Cre17.g720050.t1.2 responds only to N starvation, making this enzyme a potential candidate for further study. Conclusions {#Sec16} =========== The comprehensive analysis of systemic responses to N starvation and recovery in *C. reinhardtii* demonstrated that metabolism and growth are significantly affected at a system level. A complex network of stress-responsive proteins, metabolites, and physiological parameters was established, expanding our current understanding of physiological processes driven by a small set of proteins. Many uncharacterized proteins were identified by multivariate correlation network analysis to be involved in the response to N starvation. By N readdition it was possible to extract a list of proteins that showed a fast recovery effect, suggesting that they are highly involved in the re-establishment of vegetative cell growth. This study provides new insights and alterations to previous models and offers a complex dataset, which will be further analyzed towards increasing our biochemical understanding of the adaptive mechanisms to N starvation and recovery in *Chlamydomonas*. Methods {#Sec17} ======= Strains and cultures {#Sec18} -------------------- *Chlamydomonas reinhardtii* CC-503 *cw92,* mt+, agg1+, *nit1*, *nit2* (available at the Chlamydomonas Culture Collection, Duke University) cultures were grown in HEPES-Acetate-Phosphate medium supplemented with 7 mM NH~4~Cl (HAP + N; TAP medium in which Tris was replaced by 5 mM HEPES) at 25°C with shaking (120 rpm) in a 14:8 light:dark photoperiod (85 μmol m^-2^ s^-1^; Sylvania GroLux lamps). Cultures were pelleted down by centrifugation and resuspended in HAP -N (NH~4~Cl was replaced by 7 mM KCl) media to a final density of 1-3 × 10^5^ cells mL^-1^. Cells were sampled at times 0, 5, 24, and 72 h. After this sampling, NH~4~Cl was added to the HAP-N cultures to a final concentration of 7 mM, and then the cultures were sampled after 5 h (77 h) and 24 h (96 h). Physiological measurements {#Sec19} -------------------------- At each harvesting time the cell density was measured by employing a Thoma counting chamber and the fresh weight was determined gravimetrically. The photosynthetic rate was measured with an imaging/pulse-amplitude modulation fluorimeter (OS1-FL, Opti-Sciences). Total lipids were extracted from frozen pellets with 200 μL of a mixture of chloroform:isopropanol (1:1) and vigorous vortexing for 3 min. The samples were centrifuged (14.000 × g, 5 min, room temperature) and the supernatants were transferred to a new tube. The pellet was re-extracted with 500 μL of hexane and vigorous vortexing for 3 min. The samples were centrifuged, and the combined supernatants were dried in a speed vac. The amount of lipids was determined gravimetrically. Microscopy {#Sec20} ---------- The *Chlamydomonas* cultures were fixed in 3% (v/v) formaldehyde and kept at 4°C until staining. The cells were stained in a mixture of HEPES-Acetate-Phosphate medium with 3% (v/v) formaldehyde and 5 μg mL^-1^ of Nile red and incubated 15 min in the dark before imaging. Nile red was freshly added to the staining solution from a concentrated stock (0.1 g mL^-1^ in acetone). The stained cultures were directly observed by adding 15 μL to each slide under an LSM 780 confocal microscope (Zeiss, Germany) and Z series covering all of the cell height were captured. DIC, Nile red, and autofluorescence were recorded in different channels. From these stacks the four images closer to the middle section were selected and used for obtaining a maximum projection using the software Fiji (Figure [2](#Fig2){ref-type="fig"}). Nuclei isolation {#Sec21} ---------------- Nuclei were isolated following the protocol described in \[[@CR89]\] starting from 250 mg of fresh weight. Isolated nuclei were separated from cell debris by the use of a Percoll cushion. Quantitative proteome analysis (GeLC-LTQ-Orbitrap-MS) {#Sec22} ----------------------------------------------------- Proteins were extracted from frozen pellets (40 to 50 mg fresh weight), gel fractionated, and trypsin digested following a previously described protocol \[[@CR18]\]. Ten micrograms of digested peptides were loaded per injection into a one-dimensional nano-flow LC-Orbitrap/MS and resolved in a 90-min gradient from 5 to 40% (v/v) acetonitrile/0.1% (v/v) formic acid using a monolithic C18 column (Chromolith RP-18r, Merck, Darmstadt, Germany). MS analysis was performed on an Orbitrap LTQ XL mass spectrometer \[[@CR22]\]. The raw data were searched with the SEQUEST algorithm present in Proteome Discoverer version 1.3 (Thermo, Germany) as described by Valledor *et al.* \[[@CR17]\] using the *Chlamydomonas* genome v.5.3 (17,737 accessions), *Chlamydomonas* mitochondria (8 accessions), and chloroplast (76 accessions) databases. Only highly confident proteins, defined by at least two peptides with XCorr value greater than charge state +0.25 and 5% FDR, were considered for this work. Protein functions were identified using the BioMart tool available at Phytozome (<http://www.phytozome.org>), Mercator ([http://MapMan.gabipd.org/web/guest/app/mercator](http://mapman.gabipd.org/web/guest/app/mercator)), and the Algal Functional Annotation Tool \[[@CR28]\]. The identified proteins were quantitated by a label-free approach based on total ion count followed by an NSAF normalization strategy:$$\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {(NSAF)}_k=\raisebox{1ex}{${\left(PSM/L\right)}_k$}\!\left/ \!\raisebox{-1ex}{${\displaystyle {\sum}_{i=1}^N}{\left(PSM/L\right)}_i$}\right. $$\end{document}$$ in which the total number of spectra counts for the matching peptides from protein k (PSM) was divided by the protein's length (L), then divided by the sum of PSM/L for all N proteins within each sample. Before quantitation a second filtering step was performed to retain only those proteins with enough abundance (minimum abundance of 0.001) for a low-biased quantitation \[[@CR90],[@CR91]\]. Statistical analyses were conducted according to \[[@CR92],[@CR93]\]. Proteins were accounted for quantification only if they were present in all of the biological replicates (n = 4) at least one sampling time, or in five samples corresponding to different times. Missing spot volumes were determined from the dataset employing a sequential K-nearest neighbor algorithm. Univariate comparison between treatments was performed by the nonparametric Kruskal-Wallis test over log-transformed data for obtaining *P*-values (Additional file [2](#MOESM2){ref-type="media"}: Tables S1, S2, S8). Protein abundance values were scaled and subjected to principal component analysis (PCA), discriminant analysis (PLS-DA), sparse partial least square (sPLS), and heatmap-clustering analyses (Additional file [2](#MOESM2){ref-type="media"}: Tables S4-S6). GC-MS polar metabolite and fatty acid methyl ester (FAME) analyses {#Sec23} ------------------------------------------------------------------ Polar metabolites were extracted from frozen pellets (15 to 20 mg fresh weight) as described in \[[@CR94]\]. Fatty acids were extracted from frozen cell pellets (15 to 20 mg) previously disrupted with liquid nitrogen and glass beads in a homogenizer (15 seconds, maximum speed, Retsch Mixer Mill MM400). One mL of cyclo-hexane:water (1:0.4) was quickly added to the homogenized tissue, and mixing was repeated with the same conditions. The organic phase was recovered after centrifugation (24,000 × g, 10 min) and transferred to a new tube. Pellets were re-extracted following the same procedures, and both organic fractions were combined in a 2-mL tube and dried by speed vac. Sample derivatization (polar metabolites), methylesterification (lipids), and GC-MS measurements were carried out following the procedure previously developed in our group \[[@CR95]\] on a triple quad (TSQ Quantum GC; Thermo) instrument. Metabolites were identified based on their mass spectral characteristics and GC retention times, by comparison with retention times of reference compounds in an in-house reference library. Only metabolite peaks that were detected in all of the biological replicates for at least one sampling point or in more than 18 samples were further considered. Normalized abundance values for each metabolite were obtained by dividing peak areas by the total peak areas of each sample (sample-centric normalization). Significant differences between sampling times were assessed by One-way ANOVA (Additional file [2](#MOESM2){ref-type="media"}: Table S3), and multivariate analyses were conducted together with proteins and physiological data as described above. Bioinformatic tools {#Sec24} ------------------- All statistical procedures described above were performed using the software R v2.15.2 \[[@CR96]\] core functions plus the packages SeqKNN, mixOmics, and gplots. Confocal microscopy images were processed and analyzed using the last version of the software Fiji \[[@CR97]\] employing core packages. MapMan 3.5.1R2 \[[@CR42]\] was used to make a functional classification of the proteins. A map for the proteome of the JGI_236 version was specifically created using Mercator \[[@CR23]\] and the Algal Functional Annotation Tool \[[@CR28]\]. The accuracy of the classification was supervised by comparison to local databases, also adding some manual annotations. Experimental data were classified according to this newly developed map. We have used the dataset of transcriptomic data provided in \[[@CR33]\]. This dataset was generated using version 4 of the *Chlamydomonas* genome. To compare this dataset to ours, we have BLASTed Chlamy 4 against Chlamy 5.3 considering significant all hits with e-values lower than 10^-80^. The equivalences between Chlamy 4 and Chlamy 5, as the Illumina-based normalized abundance employed for this work, are given as supplementary material (Additional file [2](#MOESM2){ref-type="media"}: Table S9). Granger causality was tested with the statistical toolbox COVAIN \[[@CR38]\]. Parameter settings for COVAIN were Granger lag time 1 and Granger significance *P*-value 0.05 including a Benjamini-Hochberg correction. A list of all identified Granger causalities can be found in Additional file [2](#MOESM2){ref-type="media"}: Table S10. Additional files {#Sec25} ================ Additional file 1: Figure S1.Comparison of cell growth (fresh weight) of N repleted and depleted cultures of *C. reinhardtii* CC503 grown under the same conditions employed in this study. Cultures were grown in triplicate.Additional file 2: Table S1.List of the 1,534 quantified proteins in whole cell fractions. Protein abundance was determined following an NSAF approach. The mean abundance ± SD for each sampling time is indicated, as well as *P*- and q-values. Percentage of coverage and number of unique peptides used for identification and score are indicated. Deflines, E.C. number, and MapMan bins were manually curated. **Table S2.** List of proteins extracted from nuclei-enriched extracts. Spectra were processed as previously described. Deflines, E.C. number, and MapMan bins were manually curated. **Table S3.** List of the 52 quantified metabolites. Metabolite abundance was estimated from the peak areas of the indicated characteristic ions. The mean normalized abundances with respect to control ± SD for each sampling time are indicated as well as *P*-values. **Table S4.** PCA summary, sample scores, and variable loadings of the integrated (proteins, metabolites, physiological measurements) dataset. **Table S5.** PLS-discriminant analysis correlations of the integrated dataset. **Table S6.** sPLS analysis correlations of the integrated dataset employing proteins-metabolites as predictive variables and physiological measurements as responsive variables. **Table S7.** Values of individual replicates of the datasets employed for multivariate analyses. (a) Normalized abundances for proteins; (b, c) normalized abundances for polar metabolites and FAME; (d) physiological measurements. **Table S8.** List of nuclear proteins obtained after the joint analysis of total protein and nuclear extracts. Spectra were processed as previously described, but one step of *in silico* fractionation was added to remove non-nuclear contaminants. Deflines, E.C. number, and MapMan bins were manually curated. **Table S9.** Correspondence between 454 dataset (Miller *et al.* \[[@CR33]\]), Chlamydomonas 5.3, and MapMan analyses. Abundace of 454 detected transcripts is also provided. **Table S10.** Granger causality analyzed by COVAIN (Sun and Weckwerth, 2012 \[[@CR38]\]) between metabolites, proteins, and physiological data.Additional file 3: Figure S2.Representation of nitrogen starvation- and recovery-induced changes in photosynthetic pathways using MapMan. Individual plots show the variations in protein abundance at the six studied time points. Only differential proteins (*P* \< 0.05) were plotted. Protein abundances were normalized as a percentage of the maximal value in the time series.Additional file 4: Figure S3.Representation of simplified N transport, fixation, and assimilation pathways showing N starvation- and recovery-induced changes using MapMan. Individual plots show the variations in protein abundance at the six studied time points. Only differential proteins (*P* \< 0.05) were plotted. Protein abundances were normalized as a percentage of the maximal value in the time series.Additional file 5: Figure S4.Representation of ribosomal proteins using MapMan. Individual plots show the variations in protein abundance at the six studied time points. Only differential proteins (*P* \< 0.05) were plotted. Protein abundances were normalized as a percentage of the maximal value in the time series.Additional file 6: Figure S5.Metabolic changes in lipid biosynthetic pathways induced by N starvation. (A) MapMan-based representation of the evolution of the enzymes related to fatty acid biosynthesis and elongation. (B) Variations in the protein abundance of the functional groups involved in lipid metabolism. (C) Evolution of the cellular content of the indicated fatty acids during the time course experiment. (D) sPLS-based network showing the most correlated proteins and metabolites to glycerol and C18:2 nodes. C18:2 was connected to ammonia transport and assimilation proteins, whereas glycerol was mainly connected to glycerol biosynthesis proteins and NADH:ubiquinone oxidoreductases. Interestingly, Cre09.g405500.t1.3 (major lipid droplet protein, with a high importance in the model) links C18:2 and glycerol. This protein was manually annotated to major lipid droplet protein. Other proteins that link both metabolites are a ribosomal protein and g13518.t1 (annotated as amine oxidase). Network interpretation: edge color indicates correlation between nodes, circle diameter indicates the importance of each node within the network, and node color its radiality. Figure was plotted employing Cytoscape 2.8.3.Additional file 7: Figure S6.Representation of nitrogen starvation- and recovery-induced changes in nucleotide metabolism pathways. Individual plots show the variations in protein abundance at the six studied time points. Only differential proteins (*P* \< 0.05) were plotted. Protein abundances were normalized as a percentage of the maximal value in the time series.Additional file 8: Figure S7.Classification of the integrated dataset employing different multivariate methods such as principal component analysis, PCA (a,d), partial least squares discriminant analysis, PLS-DA (b,e), and sparse partial least squares, s-PLS (c,f), for which proteins and metebolites were used as predictive variables and physiological measurements as explanatory. For all analyses the first two components allowed an effective classification of the samples based on the sample time.Additional file 9: Figure S8.Simplified sPLS-based network constructed considering only highly correlated nodes (correlation \> \|0.85\|). The three main nodes established different correlations among them, with the glycerol-C18:2 cluster negatively correlated to the others. Interestingly, the fresh weight (FW) cluster is positively linked to the amino acids-myo-inositol cluster by a bridge constituted by the branched amino acids. Val, Leu, and Ile are positively correlated to a monofunctional catalase grouping with amino acids, and on the other side to an acyl-CoA oxidase linked to the FW. Surprisingly one enoyl-ACP-reductase, which is correlated to the amino acids cluster, is negatively correlated to glycerol-C18:2. Glycerol is also negatively correlated to Cre12.g536800.t1.2, a protein showing an uncharacterized domain which is related to FW. Network interpretation: edge color indicates correlation between nodes, circle diameter indicates the importance of each node within the network, and node color its radiality. Figure was plotted employing Cytoscape 2.8.3.Additional file 10: Figure S9.Detailed representation of the fresh weight (FW). FW is correlated to different kinds of enzymes: signaling enzymes, mainly GTPases and Ca2+ dependent, endopeptidases, oxidoreductases, and acyl-CoA synthases, oxidases, and dehydrogenases. Node size is dependent on stress value and node color on radiality. Edge color indicates correlation (red, positive, green negative). Figures were plotted employing Cytoscape 2.8.3. Only highly correlated nodes (correlation \> \|0.85\|) were plotted.Additional file 11: Figure S10.Detailed representation of the amino acid cluster. This cluster showed complex interactions, in which nitrogen assimilation and amino acid metabolism enzymes interact with lipid biosynthesis proteins. Based on this network, some relation between N metabolism and lipid biosynthesis can be established, for example, at the level of one enoyl-ACP-reductase, which correlated to Asp, Arg, Glu, and myo-inositol, or a 3-oxoacyl-ACP reductase linked to Thr and myo-inositol. Thus, myo-inositol is a metabolite that needs to be further investigated since it is in the center of this cluster, linking amino acids, lipids, and sugars. Node size is dependent on stress value and node color on radiality. Edge color indicates correlation (red, positive, green negative). Figures were plotted employing Cytoscape 2.8.3. Only highly correlated nodes (correlation \> \|0.85\|) were plotted. **Competing interests** The authors declare that they have no competing interests. **Authors' contributions** LV designed and performed the research, analyzed the data, and wrote the paper. TF contributed new GC-MS and FAME analysis procedures and performed research. LR-M performed research. SW contributed methods. WW designed the research, analyzed the data, and wrote the paper. All authors read, corrected, and approved the final manuscript. The authors wish to thank Dr. Monica Meijón for her invaluable help with confocal microscopy and Fiji. This research and LV work were supported by the European Union through a Marie Curie IEF grant (FP7-PEOPLE-2009-IEF-255109).

\section{Introduction} \input{Introduction} \section{Related Work} \input{RelatedWork} \section{Model and Notation} \input{Model} \section{Computation Complexity} In this section, we explore the \emph{hardness of approximation} for ARAs, FAMS and TSGs. In prior work on computation complexity of SSGs, researchers~\cite{Xu16a} have focused on hardness of exact computation providing general results relating the hardness of defender best response (DBR) problem (defined below) to the hardness of exact SSE computation. In contrast, we relate the hardness of approximation of the DBR problem to hardness of approximation of ARAs. We also prove that special cases of ARA such as FAMS and TSGs are also hard to approximate. First, we formally state the equilibrium computation problem in adversarial randomized allocation: given the assets, objects and assignment constraints of an adversarial randomized allocation problem as input, output the SSE utility and a set of pure strategies $P_1, \ldots, P_m$ and probabilities $p_1, \ldots, p_m$ that represents the SSE mixed strategy. We restrict $m$ to be polynomial in the input size. This is natural, since a polynomial time algorithm cannot produce an exponential size output. Also, as discussed in prior literature~\cite{Xu16a} the size of the support set of any mixed strategy is one more than the dimension of any pure strategy, which is the poly $kn+1$ in our case. Next, as has been defined in prior literature~\cite{Xu16a}, we state the defender best response (DBR) problem which will help in understanding the results. The DBR problem can be interpreted as the defender's best response to a given mixed strategy of the adversary. The DBR problem also shows up naturally as the slave problem in column generation based approaches to SSGs. While it is easy to show the hardness of approximation of given DBR problems, the question of how its relates to the hardness of approximation of SSE computation is open. \begin{definition} The DBR problem is $\max_{\mathbf{x} \in P} \mathbf{d} \cdot \mathbf{x}$ where $\mathbf{d}$ is a vector of positive constants. DBR is a combinatorial problem that takes the assignment constraints as inputs, and not $P$. \end{definition} Next, we state the standard definition of approximation \begin{definition} An algorithm for a maximization problem is $r$-approximate if it provides a feasible solution with value at least $OPT/r$, where $OPT$ is the exact maximum value. \end{definition} Note that lower $r$ means better approximation. Depending on the best $r$ possible, optimization problems are classified into various approximation complexity classes with increasing hardness of approximation in the following order PTAS, APX, log-APX, and poly-APX. We extensively use the well-known approximation preserving AP reduction between optimization problems for our results. We do not delve into the formal definition of complexity classes or AP reduction here due to lack of space and these concepts being standard~\cite{Ausiello}. Our first result shows that the ARA's approximation complexity is same as that of the DBR problem and in the worst case cannot be approximated. \begin{theorem} \label{ARAHard} The following hardness of approximation results hold for ARA problems \begin{itemize} \item ARA problems cannot be approximated by any bounded factor in polynomial time, unless $P=NP$. \item If the DBR problem for given ARA problem lies in some given approximation class (PTAS, APX, log-APX, poly-APX), then so does the ARA problem. \end{itemize} \end{theorem} \begin{proof}[Proof Sketch] The first result works by constructing a ARA from a NP hard unweighted ($\mathbf{d}=1$) DBR problem such that the feasibility of the constructed ARA solves the DBR problem, thereby ruling out any approximation. Such unweighted DBR problems exist (for example for FAMS). The second part of the proof works by constructing an ARA problem with one target and showing that the solution yields an approximate value for a relaxed DBR with $\mathbf{x} \in conv(P)$. Moreover, this solution is an expectation over integral points (pure strategies), thus, at least one integral point in the support set output by ARA also provides an approximation for the corresponding combinatorial DBR. \end{proof} As the above complexity result is a worst case analysis, one may wonder whether the above result holds for sub-classes of ARA problems. We show that strong versions of inapproximatibility also holds for FAMS and TSGs. \begin{theorem} \label{TSGInapp} TSGs cannot be approximated to any logarithmic factor in poly time, unless P=NP. \end{theorem} \begin{theorem} \label{FAMSHard} FAMS problems cannot be approximated to any logarithmic factor in poly time, unless P=NP. \end{theorem} Both the proofs above use AP reduction from max independent set. The proof for TSG follows from an observation that a special case of the TSG problem is the independent set problem itself, thus, any approximation for TSGs can provide an equivalent approximation for the max independent set problem. The proof for FAMS is much more involved. It involves constructing a poly number of FAMS instance (with varying number of resources) given any independent set problem. The FAMS instances are such that pure strategies corresponds to independent sets and the multiple instances are constructed such that the optimal exact solution for one of these instances provides a solution for the max independent set problem. Then, it is shown that any approximation $r$ for the FAMS problem yields at least one pure strategy (over all the FAMS instances) that corresponds to a better than $r$ approximation for the max independent set problem, thereby completing the AP reduction. See the supplementary material for details. \section{Approximation approach} Our approach to approximation first solves the $marginalLP$, which is quite fast in practice (see experiments) and provides an upper bound to the true value of the game. Then, we sample from the marginal solution, but unlike previous work~\cite{tsai2010urban}, we alter the sampled value to ensure that the final pure strategy output is valid. We describe an abstract sampling and alteration approach for ARA in this part, which we instantiate for FAMS and TSGs in the subsequent sub-sections. Recall that a constraint is given by an index set $S$ and the constraint is an equality if $n_S = N_S$. For our abstract approach we restrict our attention to ARAs with \emph{partitioned equality assignment constraints}, which means the index set $S$ for all equality constraints partitions the index set $\{1, \ldots, k\}\times \{1, \ldots, n\}$ of the allocation matrix. Further, for inequality constraints we assume $n_S = 0$. Call these problems as PE0-ARA; this class still includes FAMS and TSGs. For FAMS, which does not have equality constraints, we use dummy schedules $s_i$ to get partitioned equalities $\sum_j x_{i,j} + s_i = 1$; $s_i = 1$ denotes that resource $i$ is unallocated. Our abstract approximation approach for PE0-ARA is presented in Algorithm~\ref{abstractalgorithm}. \begin{algorithm}[t] \caption{Abstract Approximation}\label{abstractalgorithm} \DontPrintSemicolon \ForAll{$S \in $ EqualityConstraints}{ $\mathbf{x} \leftarrow CombSample(\mathbf{x}^m,S)$ } $\mathbf{x} \leftarrow FixViolatedInequalityConstraints(\mathbf{x})$ \; $\mathbf{x} \leftarrow FixEqualityConstraints(\mathbf{x})$ \; \end{algorithm} The Algorithm takes as input the marginal solution $\mathbf{x}^m$ from $marginalLP$ and produces a pure strategy. The first for loop (line 1-2) performs comb sampling for each equality constraint $S$ to produce integral values for the variables involved in $S$. Comb sampling was introduced in an earlier paper~\cite{tsai2010urban}; it provides the guarantee that $x_{i,j}^m$ is rounded up or down for all $(i,j) \in S$, the sample $x_{i,j}$ has expected value $E(x_{i,j}) = x^m_{i,j}$ for all $(i,j) \in S$ and equality S is still satisfied after the sampling. See Figure~\ref{IllusAlgo} for an example. Briefly, comb sampling works by creating $Z$ buckets of length one each, where $Z = \sum_{(i,j) \in S} \{x^m_{i,j}\}$, where $\{.\}$ denotes fractional part. Each of the $\{x^m_{i,j}\}$ length fraction is packed into the bucket (in any order and some of the $\{x^m_{i,j}\}$ fraction may have to be split into two buckets), then a number between $[0,1]$ is sampled randomly, say $z$, and for each bucket a mark is put at length $z$. Finally, the $(i,j)$ whose $\{x^m_{i,j}\}$ fraction lies on the marker $z$ for each bucket is chosen to be rounded up, and all other $x^m_{i,j}$ are rounded down. Observe that in expectation the output of comb sampling matches the marginal solution, thus, providing the same expected utility as the marginal solution. Recall that this expected utility is an upper bound on the optimal utility. However, the samples from comb sampling may not be valid pure strategies. Thus, in case the output of comb sampling is not already valid, the two abstract methods in line 3 and 4 modify the sample strategy by first decreasing the integral values to satisfy the violated inequalities and then increasing the integral values to satisfy the equalities. Such modification of the sampled strategy to obtain a valid strategy is guided by the principle that the change in defender utility between the sampled and the resultant valid strategy should be small, which ensures that change in expected utility from the marginal solution due to the modification is small. As the expected utility of the marginal solution is an upper bound on the optimal expected utility this marginal expected utility guided modification leads the output expected utility to be close to the optimal utility. These two methods on line 3 and 4 are instantiated with domain specific heuristics that implement the principle of marginal expected utility guided modification. Below, we show the instantiation for the TSG and FAMS domains. A sample execution of the above approach for TSGs is shown in Figure~\ref{IllusAlgo}. \subsection{TSG} \begin{figure}[t] \centering \includegraphics[width=\linewidth]{IllusAlgo.pdf} \caption{\small (Left to right) Sample execution for TSG: This uses the same TSG example as in Figure~\ref{Illus}. The marginal solution is the leftmost matrix which after CombSampling on each column becomes integral, e.g., 0.5 in the left column is rounded down to 0 and 1.5 rounded up to 2. Note that the CombSampling output satisfies all equalities, but exceeds the resource capacity 7 for X-ray. Next, allocation values are lowered (shown as red circle) to satisfy the X-Ray capacity but the equality constraint on third column gets violated. Next, allocation values are increased (again red circle) to fix the equality and that produces a valid pure strategy.}\label{IllusAlgo} \end{figure} The heuristics for TSG are guided by three \emph{observations}: (1) more effective resources are more constrained in their usage, (2) changing allocation for passenger categories with higher number of passengers changes the probability of detection of adversary by a smaller amount than changing allocation for category with fewer passengers and (3) higher risk passenger categories typically have lower number of passengers. Recall that for TSGs the inequalities are resource capacity constraints. Thus, for fixing violated inequalities we need to decrease allocation which decreases utility; we wish to keep the utility decrease small as that ensures that the expected utility does not move much further away from the upper bound marginal expected utility. Our approach for such decrease in allocation has the following steps: (a) prioritize fixing inequality of most violated resources first and (b) for each such inequality we attempt to lower allocation for passenger category with higher number of passengers. In light of the observations for TSG above this approach aims to keep the change in expected utility small. Specifically, observation 1 makes it likely that constraints for more effective resources are fixed in step a above. Observation 3 suggests that the changes in step b happens for lower risk passengers. Thus, step a aims to keep the allocation of effective resources for high risk passengers unchanged. This keeps the utility change small as changing allocation for high risk passengers can change utility by a large amount. Next, by observation 2, step b aims to minimize the change in probability of detecting the adversary by a low amount so that expected utility change in small. For example in Figure~\ref{IllusAlgo}, the inequality fix reduces the allocation for the third passenger category (column) which also has the highest number of passengers (15). Next, the equalities in TSGs are the constraints for every passenger category. For fixing equalities we need to increase allocation which increases utility; we wish to keep this utility increase high as it brings the expected utility closer to the upper bound marginal expected utility. Here we aim to do so by (a) prioritizing increase of allocation for categories with fewer people and (b) increasing allocation of those resources that have least slack in their resource capacity constraint (low slack could mean higher effectiveness). By Observation 1 low slack means that resource could be more effective and by Observation 2 fewer people means higher risk passengers. This ensures that higher risk passengers are screened more using more effective resources thereby raising the utility maximally. For example in Figure~\ref{IllusAlgo}, the equality for the third column is fixed by using the only available resource MD. Recall that TSGs differs from FAMS in that the allocation for TSGs can be non-binary. This offers an advantage for TSGs with respect to approximation, as small fractional changes in allocation do not change the overall allocation by much (0.5 to 1 is a 50\% change in FAMS, but 4.5 to 5 is less than 10\% change for TSGs). Thus, we assume here that the changes due to Algorithm~\ref{abstractalgorithm} do not reduce the probability of detecting an adversary in any passenger category (from the marginal solution) by more than $1/c$ factor, where $c > 1$ is a constant. This restriction is realistic as it is very unlikely that any passenger category will have few passengers and we only aim to change the allocation for passenger categories with a higher number of passengers. Hence we prove the following \begin{theorem} \label{TSGApp} Assume that Algorithm~\ref{abstractalgorithm} always successfully outputs a pure strategy and the change in allocation from the marginal strategy does not change the probability of detecting an adversary by more than $1/c$ factor. Then, the approximation approach above with the heuristic provides a $c$-approximation for TSGs. \end{theorem} As a remark, the above result does not violate the inapproximatability of TSGs since the above result holds for a restricted set of TSG problems. Also, the approximation for TSGs may sometimes fail to yield a valid pure strategy as satisfying the equalities may become impossible after using certain sequences of decreasing allocation. In our experiments we observe that the failure of obtaining a pure strategy after Algorithm~\ref{abstractalgorithm} is rare and also easily handled by repeating the approximation approach with a new sample (approximation runs in milli-secs). \subsection{FAMS} Recall that for FAMS the inequalities are the target allocation constraints: $\mathbf{x}[T] \leq 1$ and fixing violations for these involves decreasing allocation. Our heuristic is simple: the variables $x_{i,j}$ are set to zero (i.e., decreased) starting from those schedules $j$ that contain the most number of targets for which target allocation constraint is violated and do not contain any target for which the target allocation constraint is satisfied. We can guarantee to find a decrease in allocation that satisfies the constraint for $T$ without changing the allocation for targets that already satisfy constraints in the cases when the target $T$ with violated constraint (1) belongs to a schedule $j$ that exclusively contains that target $T$ ($x_{i,j}$ can be decreased without affecting any other constraint) or (2) $T$ belongs to only one schedule (all other targets in this schedule will violate their constraints). This approach ensures that we only work to fix the violated constraints and cause a minimal change in utility by leaving the satisfied constraints undisturbed. However, if in fixing a violated target allocation constraint for $T$ it becomes necessary to reduce allocation for another already satisfied target constraint, then sample uniformly from the $\geq 2$ schedules that $T$ belongs to in order to choose the $x_{i,j}$ allocation to reduce. Do this till all inequality constraints are satisfied. Then, we do nothing to fix equality constraints since we have only decreased $x_{i,j}$ and if any equality $\sum_j x_{i,j} + s_i = 1$ is not satisfied we can always set the dummy $s_i$ to be one. Also, observe that since we only always decrease allocations, we always find a pure strategy for any sample from Algorithm 1 (unlike TSGs). We prove the following: \begin{theorem} \label{FAMSApp} Let $C_t$ be the number of targets that share a schedule with any target $t$, and $C = \max_t C_t$. The approximation approach above with the heuristic provides a $2^Ck$-approximation for FAMS. \end{theorem} Next, we show experimental results that reveal the average case loss of our approximation approach (as opposed to the worst case approximation guarantees in this section). \section{Experimental Results} \begin{figure}[t] \centering \subfigure[Runtime (log-scale time)]{ \includegraphics[width=0.47\columnwidth]{FAMSRunTime.pdf} \label{fig:FAMSRunTime} } \subfigure[Solution quality]{ \includegraphics[width=0.47\columnwidth]{FAMSol.pdf} \label{fig:FAMSSol} } \caption{ASPEN and RAND Comparison} \label{AMS} \end{figure} \begin{figure}[t] \centering \subfigure[Runtime (log-scale time)]{ \includegraphics[width=0.47\columnwidth]{MGARunTime.pdf} \label{fig:MGARunTime} } \subfigure[Solution quality]{ \includegraphics[width=0.47\columnwidth]{MGASol.pdf} \label{fig:MGASol} } \caption{MGA and RAND Comparison} \label{MGA} \end{figure} Our set of experiments provide a comprehensive analysis of our randomized approach, which we name RAND. We compare RAND to the best know solver for TSGs called MGA; MGA~\cite{TSG2016} has been previously shown to outperform column generation based approaches by a large margin. For the FAMS problem the best known solver in literature for the general sum case is called the ASPEN~\cite{jain2010security}, which is a column generation based branch and price approach. Through private communication with the company (Avata Intelligence) managing the FAMS software, we know the FAMS problem is solved as a zero-sum problem for scalability and takes hours to complete. For our zero-sum case we implemented a plain column generation (CG) solver for FAMS, since branch and price is an overkill for the zero sum case. All results in this section are averaged over 30 randomly generated game instances. All game instances fix $U^t_s$ to $-1$ and randomly select $U^t_u$ from integers between $-2$ and $-10$. The utility for RAND is computed by generating 1000 pure strategies and taking their average as an estimate of the defender mixed strategy. All experiments were run using a system with Xeon 2.6 GHz processor and 4GB RAM. For FAMS, we vary the number of flights, keeping the number of resources fixed at 10 and number of schedules fixed at 1000 and 5 targets/schedule. The runtime \emph{in log scale} is shown in Figure~\ref{AMS}. CG hits the 60 min cut-off for 700 flights and the run time for RAND is much lower at only a few seconds. Next, we report the solution quality for RAND by comparing with the solution using CG. It can be seen that the solution gets better with increasing flights starting from 19\% loss at 300 flights to 5\% loss at 900 flights. Thus, the numbers show that we obtain large speed-ups up to factor of 1000x and are still able to extract 95\% utility for higher number of flights when the CG approach starts taking huge amount of time. For TSGs, we used six passenger risk levels, eight screening resource types and 20 screening team types. We vary the number of fights and we also randomly sample the team structure (how teams are formed from resources) for each of the 30 runs. The results in Figure~\ref{MGA} show runtime (in log scale) and defender utility values varying with number of flights (on x-axis). As can be seen, MGA only scales up to 110 flights before hitting the cut-off of 60 mins, while RAND takes only 10 sec for 110 flights. Also, the solution quality loss for RAND has a maximum averaged loss of 1.49\%. Thus, we obtain at-least 360X speed-ups with very minor loss. Finally, we test the scalability of RAND for FAMS and TSG, shown in Figure~\ref{ScaleRAND}. As can be seen, the runtime for RAND is low even with the highest number of flights we tested: 280 for TSG and 3500 for FAMS. The maximum runtime for FAMS was under 5 seconds. With TSGs the maximum runtime was under 25 secs. Overall, these results demonstrate the scalability of our approach with minor solution quality loss on average. Importantly, the scalability results shows that RAND can solve the FAMS and TSG problem for sizes that are required for a US wide deployment, and the results also provide evidence that further scalability with increasing number of flights is also possible. This also provides evidence that approximations is a viable approach to solving zero-sum SSGs at a large scale. \begin{figure}[t] \centering \subfigure[Runtime FAMS]{ \includegraphics[width=0.47\columnwidth]{FAMS_RR_Scale.pdf} \label{fig:FAMSRRRuntime} } \subfigure[Runtime TSG]{ \includegraphics[width=0.47\columnwidth]{TSG_RR_Scale.pdf} \label{fig:TSGRRRunTime} } \caption{Scalability of RAND} \label{ScaleRAND} \end{figure} \section{Conclusion} We studied approximations in SSGs both theoretically and practically. In fact, while our paper title talks about half a loaf (optimal solution), we practically obtain up to 97\% of the loaf (optimal solution). We obtain 1000x speedups with losses within 5\%. Thus, approximations open the door to solving very large scale zero-sum SSGs and we believe that we have laid a fertile ground for future research in approximating large scale zero-sum security games. Our approach also provided an avenue to solve the real world FAMS and airport screening problem in order to increase the scope of applicability of these already deployed application (in case of FAMS) or applications under test (airport screening). \clearpage \bibliographystyle{aaai} \section{Appendix} \subsection{ARA complex model} It is easy to change the probability to be $c_t = f(\{x_{i,j} ~|~ (i,j) \in T\})$ so that the probability is non-linear (as given by f) in the entries. Also, clearly increasing number of targets to be large is allowed as targets can be any subset of the indexes. A general sum variant can also be stated by specifying the adversary utilities in terms of $c_t$. Of course, as in SSGs, the general sum case cannot be written as a LP but can be written as multiple optimizations or a MILP. \subsection{Implementability} Viewing SSGs as ARAs provides an easy way of determining implementability using results from randomized allocation~\cite{budish2013designing}. First, we define \emph{bi-hierarchical assignment constraints} as those that can be partitioned into two sets $H_1, H_2$ such that two constraints $S, S'$ in the same partition ($H_1$ or $H_2$) it is the case that either $S \subseteq S'$ or $S' \subseteq S$ or $S \cap S' = \phi$. Then, we obtain the following sufficiency result \begin{proposition} All marginal strategies are implementable, or more formally $conv(P) = MgS$, if the assignment constraints are bi-hierarchical. \end{proposition} \begin{proof} The proof is just an application of Thm 1 in~\cite{budish2013designing}. \end{proof} In contrast to prior work that have identified cases of non-implementability~\cite{KorzhykCP10,letchford2013solving,TSG2016} for specific cases, this provides an easy way to characterize non-implementability across a wide range of SSGs. As Figure~\ref{Illus} reveals, both FAMS and TSG have non-implementable marginals due to overlapping constraints. Next, as defined in~\cite{budish2013designing}, \emph{canonical assignment constraints} are those that impose constraints on all rows and columns of the matrix (with possibly additional constraints), using which we obtain the following necessity result \begin{proposition} Given canonical assignment constraints, if all marginal strategies are implementable then the assignment constraints are bi-hierarchical. \end{proposition} \begin{proof} The proof is just an application of 2 in~\cite{budish2013designing}. \end{proof} \subsection{Proofs} \textbf{Proof of Theorem~\ref{ARAHard}}: \begin{proof} For the first part, given any ARA problem with a NP hard DBR-U instance (for the decision version of DBR-U), we construct another ARA instance such that the feasibility problem for that ARA instance solves the DBR-U decision problem. Thus, as the feasibility is NP Hard, there exists no approximation. First, since ARA problem is so general there exists DBR-U problems that are NP Hard. For example the DBR-U problem for FAMS has been shown to be NP Hard~\cite{Xu16a}. Given the hard DBR-U problem, form an ARA problem with by adding the constraint $\mathbf{1} \cdot \mathbf{x} = k$. Also, let there be only one target $t$ in the problem, so that the objective becomes $U(\mathbf{x}, t)$ insted of $z$ and all constraints in the optimization are just the marginal space constraints and $\mathbf{1} \cdot \mathbf{x} = k$. Now, the existence of any solution of the optimization gives a feasible point as $\mathbf{x} = \sum_m a_m \mathbf{x_m}$, where the integral points $x_m$ is solution to the question of does there exist a solution of the DBR-U optimization problem with value $k$. Next, a binary search on $k$ can answer the decision DBR-U problem of whether there exists a solution for decision DBR-U with value $\geq k$. Thus, existence of any solution for ARA is NP Hard, thus, no approximation is easy. For the second part, we present a AP approximation preserving reduction (with problem mapping that doe not depend on approximation ratio); such a reduction preserves membership in PTAS, APX, log-APX, Poly-APX (see ). Given any DBR problem, we construct the ARA problem with one target such that $T = \{1, \ldots, k\}\times \{1, \ldots, n\}$. Choose the weights $w_{i,j}$'s such that $w_{i,j} \propto d_{i,j}$ and $w_{i,j} \leq 1/\max_{\mathbf{x} \in MgS} \sum_{i,j} x_{i,j}$. Observe that $\max_{\mathbf{x} \in MgS} \sum_{i,j} x_{i,j}$ is computable efficiently and $\max_{\mathbf{x} \in MgS} \sum_{i,j} x_{i,j} \geq \max_{\mathbf{x} \in conv(P)} \sum_{i,j} x_{i,j}$, thus, the ARA is well-defined. Thus, due to just one target, the ARA optimization is same as $\max_{\mathbf{x} \in conv(P)} \mathbf{w} \cdot \mathbf{x}$. Suppose we can solve this problem with $r$ approximation with the solution mixed strategy being $ \mathbf{x}^\epsilon = \sum_{i=1}^m a_i \mathbf{x_i}$ for some pure strategies $\mathbf{x_i}$. Now, since $w_{i,j} \propto d_{i,j}$ we also know that this solution also provides $r$ approximation for DBR-C. Let the optimal solution for DBR-C be $OPT$; note that $OPT$ is also the optimal solution for DBR. $ \mathbf{x}^\epsilon$ provides a solution value $\mathbf{w} \cdot \mathbf{x}^\epsilon \geq OPT/r$. Further, as the objective is linear in $\mathbf{x}$ and $ \mathbf{x}^\epsilon = \sum_{i=1}^m a_i \mathbf{P_i}$, it must be the case that there exists a $j \in \{1, \ldots, m\}$ such that $\mathbf{w} \cdot \mathbf{P_j} \geq \mathbf{w} \cdot \mathbf{x}^\epsilon \geq OPT/r$. Thus, since $\mathbf{P_j} \in P$, $\mathbf{P_j}$ provides $r$ approximation for DBR. Since, $m$ the number of the pure strategies in support of $\mathbf{x}^\epsilon$ is polynomial, $\mathbf{P_j}$ can be found in polynomial time by a linear search. \end{proof} \textbf{Proof of Theorem~\ref{TSGInapp}}: \begin{proof} Given an independent set problem with $V$ vertices, we construct a TSG with $\{1, \ldots, V + 1\}$ team types, where each team type in $1, \ldots, V$ corresponds to a vertex. The $V+1$ team is special in the sense that it does not correspond to any vertex and it is made up of just one resource with a very large resource capacity $2V$. Construct just one passenger category with passengers $N = V+1$. Since, there is just one passenger category (and target) we will use $x_i$ as the matrix entries instead of $x_{i,j}$. Choose $U^t_s = V+1$ and $U^t_u = 0$ and efficiencies $E_i = 1$ for all teams, except $E_{V+1} = 0$. Then, the objective of the integer LP is $\sum_{i=1}^V x_i = \mathbf{1}_V \cdot \mathbf{x}$ where $ \mathbf{1}_V $ is a vector with first $V$ components as 1 and last component as 0. Next, have resources for every edge $(i,k) \in E$ with resource capacity $1$. This provides the inequality $\sum_{(i,k) \in E} x_i + x_j \leq 1$. Also, we have $x_{V+1} \leq 2V$. Inspection of every passengers provides the constraints $\sum_{i=1}^{V+1} x_i = V+1$. Treating $x_{V+1}$ as a slack, we can see that the constraint $x_{V+1} \leq 2V$ and $\sum_{i=1}^{V+1} x_i = V+1$ are redundant. For the left over constraints $\sum_{(i,k) \in E} x_i + x_j \leq 1$, we can easily check that any valid integral assignment (pure strategy) is an independent set. Moreover, the objective $\sum_{i=1}^V x_i $ tries to maximize the independent set. The optimal value of this optimization over $conv(P)$ is an extreme point which is integral and equal to the maximum independent set OPT. Thus, suppose a solution $\mathbf{x}^\epsilon$ to the SSE problem with value $\geq OPT/r$. Further, as the objective is linear in $\mathbf{x}$ and $ \mathbf{x}^\epsilon = \sum_{i=1}^m a_i \mathbf{P_i}$, it must be the case that there exists a $j \in \{1, \ldots, m\}$ such that $ \mathbf{1}_V \cdot\mathbf{P_j} \geq \mathbf{1}_V \cdot \mathbf{x}^\epsilon \geq OPT/r$. Thus, since $\mathbf{P_j} \in P$, $\mathbf{P_j}$ provides $r$ approximation for maximum independent set. Since, $m$ the number of the pure strategies in support of $\mathbf{x}^\epsilon$ is polynomial, $\mathbf{P_j}$ can be found in polynomial time by a linear search. \end{proof} \textbf{Proof of Theorem~\ref{FAMSHard}}: \begin{proof} We provide an AP reduction from independent set. As max independent set is poly APX complete, this rules out any log factor approximation. Given an independent set maximization problem with vertices $V$ and edges $E$ construct the following FAMS problems, one for each $k$. Use $2n - k$ resources. All resources can be assigned to any schedule. Construct schedules $s_1, \ldots, s_n$ corresponding to the vertices $v_1, \ldots, v_n$. Construct target $t_{e}$ corresponding to every edge $e = (u,v)$ such that $t_e \in s_u$ and $t_e \in s_v$. All $t_e$'s have the same value for being defended or undefended and that value is $n+2$; thus, these targets do not need to covered but impose the constraint that $s_u$ and $s_v$ cannot be simultaneously defended. Thus, it is clear that any allocation of resources to $s_1, \ldots, s_n$ corresponds to an independent set. Next, consider additional $2n$ targets and expand the schedules such that $t_{i}, t_{i+1} \in s_{i}$. Further, add more singleton schedules $s_{n+1}, \ldots, s_{3n}$ with $t_{i} \in s_{n+i}$. All additional targets $t_1, \ldots, t_{2n}$ provide value $k$ when defended and $k - 2n$ otherwise. Thus, the expected utility of defending an additional target $t$ given coverage $c_t$ is $c_t(k) + (1-c_t)(k-2n) = 2n*c_t + k - 2n$. First, assume we have a poly time algorithm to approximately compute the SSE with approx factor $r$ ($r > 1$). We will run this poly time algorithm with resources $2$ to $2n-1$ which is again a poly time overall, and also the overall output size is poly. For the given max independent set problem, let the solution be $k^*$. Observe that for problems with resources $2n - k$ where $k \leq k^*$, all valuable (additional) targets can be covered by covering $k^*$ schedules with $2k^*$ targets in $s_1, \ldots, s_n$ and using the remaining $\geq 2n - 2k^*$ resources to cover the remaining $2n - 2k^*$ targets. This provides utility of $k$ for the SSE. In particular, the utility with $2n -k^*$ resources is $k^*$. Also note that for every problem, there is always a trivial allocation of $2n-k$ resources to the $2n$ singleton schedules such that coverage of each target is $1 - k/2n$. (this is deducible as the allocation to singleton schedules is unconstrained and can be implemented in poly time by Birkhoff-von Neumann result as provided in~\cite{KorzhykCP10}). This trivial allocation provides an utility 0. Also the following result will be useful: given approximation factor $r$ for the case with $2n - k^*$ resources then one of the pure strategy output for this case will have at least $k^* - l_{\min}$ schedules among $s_1, \ldots, s_n$ covered where $l_{\min} = \lfloor \arg\!\min_l \frac{k^*}{k*-l} > r \rfloor$. Before proving the above note that by definition $\frac{k^*}{k*-l_{\min} - 1} > r$ and $\frac{k^*}{k*-l_{\min}} < r$. To prove the result in last paragraph consider the contra-positive: suppose all pure strategies output cover at most $k^* - l_{\min}-1 $ schedules among $s_1, \ldots, s_n$, then in each pure strategy at least $l_{\min}+1$ targets are not covered (since 2 targets are covered for the $k^* - l_{\min}-1 $ schedules and rest of resources can cover only 1 target). Then the coverage of the least covered target in the mixed strategy formed using such pure strategies is $\leq 1 - (l_{\min}+1)/2n$. The utility for this least covered target is $\leq k^* - l_{\min} - 1$. The overall utility has to be lower than utility for any target, hence the utility is $\leq k^* - l_{\min} - 1$. The optimal utility is $k^*$. Thus, by definition of approximation ratio $r$ we must have $k^* - l_{\min} - 1 \geq k^*/r$ or re-arranging $\frac{k^*}{k^*-l_{\min} - 1} \leq r $ but by definition of $l_{\min}$ we must have $\frac{k^*}{k*-l_{\min} - 1} > r$ hence a contradiction. The above result also shows that at least one pure strategy yields an independent set of size $k^* - l_{\min}$. As $k^*$ is the max size of independent sets we obtain and approximation ratio $r'$ for the max independent set problem such that $r' < \frac{k^*}{k^* - l_{\min}} < r$. Thus, we obtain a better approximation $r'$ given $r$ approximation for the SSE. Thus, we have an AP reduction. \end{proof} \textbf{Proof of Theorem~\ref{FAMSApp}}: \begin{proof} Consider the event of a target $t$ having an infeasible assignment after the comb sampling. Call this event $E_t$. Let $C_{t,i}$ be the event that resource $i$ covers this target $t$. Then, $P(E_t) = \sum_{i} P(E_t|C_{t,i})P(C_{t,i})$. From the guarantees of comb sampling we know that $P(C_{t,i}) = \sum_{j: (i,j) \in T} x^m_{i,j} \leq 1$ and $P(x_{i,j} = 1) = x^m_{i,j}$. Also, by comb sampling if $x_{i,j} = 1$ then $x_{i,j'} = 0$ for any $j'\neq j$. Next, we know that $P(E_t|C_{t,i})$ is the probability that the any of the other $x_{i',j}$ is assigned a one, which is $1 - $ the probability that all other $x_{i',j}$ are assigned 0. Thus, $$P(E_t|C_{t,i}) = 1 - \prod_{i' \neq i} (1- P(C_{t,i})) $$ Let $p_{t,i} = P(C_{t,i})$. Considering the fact that $\prod_i (1 - p_{t,i}) > 1 - \sum_i p_{t,i}$, we get $$1 - \prod_{i' \neq i} (1- P(C_{t,i})) \leq \sum_{(i',j): i'\neq i \land (i',j) \in T} x^m_{i',j} \leq 1 - \sum_j x^m_{i,j}$$ where the last inequality is due to the fact that $\sum_{(i,j) \in T} x^m_{i,j} \leq 1$. Thus, $P(E_t) \leq \sum_{i} (1 - p_{t,i})p_{t,i} \leq \sum_{i} p_{t,i} - \sum_{i} (p_{t,i})^2$. Next, we know from standard sum of squares inequality that $\sum_{i} (p_i)^2 \geq (\sum_{i} p_i)^2/k$. Thus, we get $P(E_t) \leq (\sum_{i} p_i) (1 - \sum_{i} p_i/k)$ The RHS is maximized when $\sum_{i} p_i = 1$, thus, $P(E_t) \leq 1 - 1/k$. Also, then $P(\lnot E_t) \geq 1/k$ Now consider the coverage of target $t$: $x^m_t = \sum_{(i,j) \in T} x^m_{i,j}$. According to our algorithm the allocation for target $t$ continues to remain $1$ with probability $(1/2)^C$ if its allocation is already feasible after comb sampling (and we always obtain a pure strategy). This is because this target shares schedules with $C$ other targets and thus in the worst case may be reduced with $1/2$ probability for each of the $C$ targets. We do a worst case analysis and assume that no resource is allocated to a target when the sampled allocation is infeasible for that target. Thus, let $y_t$ denote the random variable denoting that target $t$ is covered. Thus, $E(y_t) = P(y_t = 1) = P(y_t = 1|E_t) P(E_t) + P(y_t = 1|\lnot E_t)P(\lnot E_t)$. Now, $P(y_t = 1|\lnot E_t)$ is same as $x^m_t/2^C$ and we assumed the worst case of $P(y_t = 1|E_t) = 0$. Thus, we have $E(y_t) \geq x^m_t/2^Ck$. As the utilities are linear in $y_t$, we have the utility for $t$ as $U_t \geq U_t^m/2^Ck$, where $U_t^m$ is the utility under the marginal $\mathbf{x}^m$. Thus, if $t^*$ is the choice of adversary under the marginal $\mathbf{x}^m$ we know that $U_{t^*}^m$ is the lowest utility for the defender over all targets $t$. Hence, we can conclude that the utility with the approximation is at least $U_{t^*}^m/2^Ck$ \end{proof} \textbf{Proof of Theorem~\ref{TSGApp}}: \begin{proof} The main assumption in the proof is that the steps after after comb sampling changes the probability of detecting an adversary in passenger category $j$ by at most $1/c$. Also, by assumption of the theorem since Algorithm~\ref{abstractalgorithm} does not fail ever, the change in utility for any passenger category $j$ is at most a factor of $1/c$. By similar reasoning as for FAMS, we conclude that this provides a $c$-approximation. \end{proof} \subsection{FAMS} We model zero-sum FAMS in the ARA model. The FAMS problem is to allocate federal air marshal (FAMS) to flights to and from US in order to prevent hijacking attacks. The allocation is constrained by the number of FAMS available and the fact that each FAMS must be scheduled on round trips that take them back to their home airport. Thus, the main technical complication arises from the presence of schedules. A schedule is a subset of flights that has to be defended together, e.g., flight f1 and f2 should be defended together as they form a round trip for the air marshal. Air marshals are allocated to schedules, no flight can have more than one air marshal and some schedules cannot be defended by some air marshals. The adversary attacks a flight. Then, we capture the FAMS domain in the above model by mapping schedules in FAMS to objects (on columns) and air marshal in FAMS to assets (on rows). See Figure~\ref{Illus} for an illustrative example. The assignment constraints include the constraint for each resource $i$: $\sum_j x_{i,j} \leq 1$, which states that every resource can be assigned at most once. If an air marshal $i$ cannot be assigned to schedule $j$ then add the constraint $ x_{i,j} = 0$. A target $t$ in the abstract model maps to a flight $f$ in FAMS, and the set $T$ are all the indexes for all schedules that include this flight: $\{(i,j) ~|~$ flight f is in schedule $j\}$. The constraint that a flight cannot have more than one air marshal is captured by adding the \emph{target allocation constraint} $\mathbf{x}[T] \leq 1$. The probability of defending a target (flight) is $c_t = \mathbf{x}[T]$, hence the weights $w_{i,j}$'s in ARA are all ones. \subsection{TSG} We model TSGs using the Bayesian formulation of ARA. The TSG problem is how to allocate screening resources to screenees in order to screen optimally, which we elaborate in the context of airline passenger screening. In TSGs, different TSG resources such as X-Rays, Metal Detector act in teams to inspect an airline passenger. The possible teams are given. Passengers are further grouped into passenger categories with a given $N_c$ number of passengers in each category $c$. The allocation is of resource teams to passenger categories. There are \emph{resource capacity constraints} for each resource usage (not on teams but on each resource). Further, all passengers need to be screened. Each resource team $i$ has an effectiveness $E_i < 1$ of catching the adversary. Observe that, unlike SSGs, the allocation in TSGs is not just binary $\{0,1\}$ but any positive integer within the constraints. The passenger category $c$ is a tuple of risk level and flight $(r, f)$; the adversary's action is to choose the flight $f$ but he is probabilistically assigned his risk level. Then, we capture the TSG domain in the above abstract model by mapping passenger categories in TSGs to objects (on columns) and resource teams in TSGs to assets (on rows). See Figure~\ref{Illus} for an illustrative example. The capacity constraint for each resource $r$ is captured by specifying the constraint $\mathbf{x}[S] \leq N_S$ which contains all indexes of teams that are formed using the given resource $r$: $S = \{(i,j) ~|~$ team $i$ is formed using resource $r\}$ with $N_S$ equal to the resource capacity bound for resource $r$. For every passenger category $j$, the constraint $\sum_i x_{i,j} = N_j$ enforces that all passengers are screened. A target $t$ in TSG is simply a passenger category $j$, thus, the set $T$ is $\{(i,j)~|~ j$ is given passenger category$\}$. The probability of detecting an adversary in category $j$ is given by $\sum_{(i,j) \in T} E_i x_{i,j}/N_j$, hence the weights $w_{i,j}$ are $E_i/N_j$; since $E_i < 1$ it is easy to check that $\sum_{(i,j) \in T} w_{i,j} x_{i,j} \leq 1$ for any $T$. The adversary type is the risk level $r$, and each type $r$ of adversary can choose a flight $f$, thus, choosing a target which is the passenger category $(r,f)$. The probability of the adversary having a particular risk level is given.

**Responsible Editor** Stig Wall, Umeå University, Sweden Background {#S0001} ========== Hearing loss is one of the most common developmental disorders identifiable at birth which, if left undetected, has a negative impact on a child's speech, language, cognitive, educational and socio-emotional development \[[1](#CIT0001),[2](#CIT0002)\]. Approximately 0.5 to 5 in every 1000 neonates and infants have congenital, early childhood onset sensorineural or severe-to-profound hearing loss \[[3](#CIT0003)\]. Hearing loss may lead to developmental delay and difficulty progressing in school if timely and optimal interventions are not provided \[[3](#CIT0003)\]. These children are therefore at a greater risk for failure and drop-out from school thus placing the children at an economic disadvantage \[[4](#CIT0004)\]. Even a unilateral hearing loss in children poses significant risk factors such as increased rates of grade failure, the need for additional educational assistance, and perceived behavioral issues in the classroom \[[5](#CIT0005)--[7](#CIT0007)\]. Unfortunately, there are limited prospects of identifying hearing loss in children, particularly within developing regions such as sub-Saharan Africa where an estimated 6.8 million children suffer from permanent disabling hearing loss \[[5](#CIT0005),[8](#CIT0008),[9](#CIT0009)\]. This may be attributed to the absence of early hearing detection and identification (EHDI) programs due to reasons including limited human resources for ear and hearing care, a lack of appropriate equipment, costs and other health care priorities \[[10](#CIT0010),[11](#CIT0011)\]. The World Health Organization (WHO) estimates that there is only one audiologist per 0.5 million to 6.25 million people in the developing world \[[11](#CIT0011)\], with countries in sub-Saharan Africa presenting with a ratio of one audiologist per 0.8 million people \[[12](#CIT0012)\]. Community-based hearing programs have been proposed as a way to improve access to ear and hearing care \[[13](#CIT0013)\]. WHO primary ear and hearing care training manuals recommend that primary health care workers and community healthcare workers (CHWs) in LMICs are trained to stimulate and encourage greater prioritization of prevention, identification and treatment of hearing loss \[[14](#CIT0014)\]. Prevention or early identification can reduce the negative consequences of a hearing loss, is usually less expensive and can often be implemented at a community level \[[13](#CIT0013),[15](#CIT0015)\]. Costs may also be reduced by using innovative technologies using mobile health or mHealth applications \[[13](#CIT0013),[16](#CIT0016)\]. With the widespread penetration of 4.92 billion mobile phones worldwide, of which more than 3.74 billion are smartphones, mHealth hearing applications are demonstrating promise to improve access to hearing services in low- and middle-income countries (LMICs) \[[16](#CIT0016)--[20](#CIT0020)\]. One such mHealth solution, validated in various contexts, is the hearScreen^TM^ solution that allows a low-cost alternative to conventional hearing screening equipment whilst adhering to required acoustic calibration standards \[[21](#CIT0021),[22](#CIT0022)\]. The hearScreen^TM^ mHealth solution allows for pre-specified screening protocols with automated sequences to be employed by non-specialist personnel \[[21](#CIT0021),[23](#CIT0023)\]. This means generalist healthcare workers or schoolteachers can be trained to operate the device, after which patient-specific data and results collected on the smartphone application can be uploaded to a centralized cloud-based server through cellular networks. This allows for asynchronous point-of-care diagnostics in difficult to reach populations, with cloud-based data management, surveillance and referrals that in turn may reduce the demand placed on already limited professional ear and hearing health human resources in developing countries. A smartphone-based application may also offer other benefits essential when testing in informal settings, including environmental noise monitoring, quality control indices of test operators and data management \[[21](#CIT0021),[22](#CIT0022)\]. The hearScreen^TM^ software employs noise monitoring algorithms which provide operators with real-time feedback on ambient noise levels, thereby providing a guide to minimize the effect of noise levels when testing in varying noise conditions \[[21](#CIT0021),[22](#CIT0022)\]. The hearScreen^TM^ automatically retests frequencies where maximum noise levels were exceeded \[[21](#CIT0021),[22](#CIT0022)\]. Furthermore, smartphone applications can employ a geotag feature to immediately link patients to their closest hearing health providers or primary healthcare facility with text message notifications. Pure tone audiometry screening in schools using the hearScreen^TM^ application has demonstrated a low-cost, accurate and efficient asynchronous screening solution that could be facilitated by non-specialist personnel with limited training \[[22](#CIT0022)\]. However, no systematic clinical validation has been conducted for using this solution to identify younger, more difficult to test preschool children. Early childhood development (ECD) centers are aimed at providing emotional, cognitive and physical development of children from birth to school-going age in addition to a focus on a child's nutrition, health, psychological and other needs \[[24](#CIT0024)\]. With the integration of asynchronous, low-cost mHealth technologies, early childhood development centers in LMICs could therefore have the potential to serve as the first point of access to preventative hearing healthcare services for children from underserved populations prior to school entry. Therefore this study set out to determine the feasibility of a low-cost, ECD hearing screening program for preschool children operated by CHWs using an mHealth point-of-care diagnostic and cloud-based data management, surveillance and referral system. Methods {#S0002} ======= Context {#S0002-S2001} ------- The study was conducted in the community of Mamelodi, City of Tshwane, Gauteng, South Africa. Mamelodi is situated approximately 20 km east of the city. The unofficial population of Mamelodi is currently close to one million. Census indicated 110,703 households within the community of which only 61% are formal dwellings \[[25](#CIT0025)\]. Participants {#S0002-S2002} ------------ Non-probability purposive sampling was used to select participants. Initially, three CHWs were trained to map ECD centers and conduct hearing screenings. An additional two CHWs joined the project during the last three months to assist with the workload. CHWs were first trained to map ECD centers using the facility-mapping feature of the hearScreen^TM^ application. This feature allowed CHWs to record the name of the ECD facility, geolocation, contact person and number of children enrolled. CHWs were thereafter trained to conduct hearing screenings using the hearScreen^TM^ application. These participants had no formal training on hearing healthcare. Prior to implementation of the project, CHWs received a training session during which they were provided with information regarding ear and hearing healthcare, and its importance, as well as training and hands-on practice with the hearing screening smartphone application. Two hundred and fifty ECD centers were mapped in the community of Mamelodi East and West. ECD centers (crèches) included both public and private facilities that provided learning and support to children between the ages of three to six years. These ECD centers were often informal in nature and based in the homes of local community members. Once consent was obtained from the principal of the ECD centers to conduct hearing screenings, consent letters were sent to the parents/caregivers. Data was collected over a 12-month period with the exception of 3 vacation periods. Equipment {#S0002-S2003} --------- The hearScreen^TM^ smartphone application was initially operated on Samsung Trend Plus (S5301) smartphones (Android OS, 4.0). In July 2016, the Samsung Trend plus smartphones were replaced with Samsung J2 Galaxy smartphones (Andriod OS, 5.1), which operated an upgraded version of the hearScreen^TM^ software. Smartphones were connected to supra-aural Sennheiser HD280 Pro headphones (Sennheiser, Wedemark, Germany). The hearScreen^TM^ calibration function was used to calibrate the headphones according to prescribed standards (ISO 389--1:1998) adhering to equivalent threshold sound pressure levels determined for this headphone \[[26](#CIT0026)\]. The hearScreen^TM^ application records a quality index of test operators, which gives an objective measure of their screen performance. During each test, a randomized false presentation of a sound is presented. The purpose of this presentation is to determine if the screener correctly records the response by the person tested as 'no response.' The hearScreen^TM^ then calculates a quality index on the tests completed by each screener to provide an indication of the reliability of the tests conducted by the test operator. This quality index was monitored throughout the project in order to guide and retrain testers when needed. Noise levels are also recorded by the smartphone hearing screening application for each child during testing. In order to minimize false-positive results caused by exceeded noise levels, testing was not conducted at 0.5kHz. The hearScreen^TM^ solution has been validated to monitor noise accurately \[[21](#CIT0021)\]. Data collected by the smartphone was automatically uploaded to a secure cloud-based server through a 3G cellular network. Procedures {#S0002-S2004} ---------- Once mapping was completed, hearing screenings were only conducted on a set test date if consent was granted by both the ECD center and the parent/caregiver. ECD center staff allocated a room with the least noise possible for testing. Children were provided with a simple explanation and demonstration of what was required of them. The hearing screening application employed automated test protocols. In order to ensure that the child understood what was expected, testing began in the left ear with an initial conditioning tone at 1 kHz at an intensity level of 35 dB HL. Thereafter, a sweep was performed at the test frequencies of 1, 2 and 4 kHz bilaterally at a screening intensity of 25 dB HL \[[21](#CIT0021)\]. The smartphone microphone measured noise levels in the environment and employed a smart noise-monitoring algorithm that only initiated a rescreen if noise levels exceeded maximum permissible ambient noise levels (MPANLs) when there was no response from a patient. The stimulus was repeated once if the child did not respond at any frequency. Once data was collected for the left ear, the same procedure was repeated in the right ear. Failure to respond at 25 dB HL at any frequency in any ear constituted an initial fail. In such cases, children were reconditioned and an immediate rescreen was initiated which followed the same procedure \[[2](#CIT0002)\]. If a child referred the immediate rescreen at the ECD center, he/she was referred to his/her local clinic for a second hearing screening followed by a diagnostic hearing assessment if necessary. This was done by automatically sending a text message notification to parents via the cloud-based server. Additionally, results and test quality were remotely monitored from the cloud-based data management portal. ECD facilities were provided with a summary of results for educational interventions. Data analysis {#S0002-S2005} ------------- Data were extracted from the cloud-based server to an MS Excel (2011) sheet and analyzed using SPSS v24 (Chicago, Illinois). Referral rates, test times, noise levels and quality indices of testers were analyzed using descriptive statistical measures. Binomial logistic regression analysis was used to determine the effects of age, gender and exceeded MPANLs on referral rates in children, with p \< 0.05 used to indicate a significant effect. Frequency distributions were also used to analyze the quality indices of tester. Results {#S0003} ======= A total of 6424 children (3446 females, 2978 males) between the ages of 3 and 6 years were screened at ECD facilities. Initial screen referral rates were 34.8% ([Table 1](#T0001)), with no significant difference between left and right ears (p \> 0.05, chi-square). A total of 2227 children were rescreened automatically after the initial failed screen, resulting in an overall referral rate of 24.9% varying from 19.6 to 45.8% for children 6 and 3 years of age, respectively. A rescreen was not completed for nine participants due to a tester inadvertently selecting to skip the procedure. Mean test duration, including both initial and rescreen test times for both ears, was 68 seconds (SD 2.8) for participants who passed and 258.5 seconds (SD 251.2) for those who failed.10.1080/16549716.2018.1467077-T0001Table 1.Referral rate for smartphone hearing screenings in ECD centers. Screening at ECD centersSecond screening at clinicsParticipants (n)Referral Rate (%)Participants (n)Referral Rate (%)Initial screen642434.861750.2Left overall642425.261736Left 1kHz642418.561727.9Left 2kHz642414.161719.6Left 4kHz64249.861718.3Right overall642426.461739.9Right 1kHz64242161730.6Right 2kHz642413.761722.2Right 4kHz64241161721.1Immediate rescreen222770.330980.6Left overall222752.930959.9Left 1 kHz222749.230946.9Left 2 kHz222740.230929.4Left 4 kHz222731.630932.4Right overall222757.530968.9Right 1 kHz22275530956.3Right 2 kHz222741.530940.5Right 4 kHz222735.430935.3Overall referral result642424.961740.5 Average noise levels recorded during the initial screen at ECD centers were 44, 41 and 40 dB for 1, 2 and 4kHz respectively. MPANLs were exceeded occasionally during the initial and immediate rescreens conducted at 1, 2 and 4 kHz at both ECD centers and clinics ([Table 3](#T0003)).10.1080/16549716.2018.1467077-T0002Table 2.Referral rate in children according to gender and age groups. Screening at ECD centersRescreening at clinicsParticipants (n)Referral Rate (%)Participants (n)Referral Rate (%)Gender    Female344626.924341.2Male297822.737440.1Age groups    3 years50445.871624 years151930141395 years22592219536.96 years214219.621037.610.1080/16549716.2018.1467077-T0003Table 3.Instances where noise levels exceeded MPANLs during smartphone screening. Frequencies1kHz2kHz4kHzLeftRightLeftRightLeftRightScreening at ECD centersInitial Screen (n = 6424)8.3%7.2%0.9%0.9%0.9%0.9%Immediate rescreen (n = 2227)6%5.7%0.3%0.2%0.2%0.1%Second screening at clinicsInitial Screen (n = 617)6.2%6.5%1.5%1.6%1.5%1.5%Immediate rescreen (n = 309)5.5%7.1%1.0%0.6%0.6%0.3% A binomial logistic regression was performed to ascertain the effect of gender, and noise levels at each frequency, as well as age on overall results obtained at the ECD centers. The logistic regression model was statistically significant (χ^2^(8) = 185.412, p \< 0.001) and correctly classified 75.1% of the cases. Referral rates were significantly affected by age (p \< 0.01; *B*: −0.004; 95% CI lower: 0.996, 95% CI upper: 0.997), gender (p \< 0.01; *B*: 0.231; 95% CI lower: 1.122, 95% CI upper: 1.415), and noise levels at 1kHz in the left ear (p \< 0.01; *B*: 0.356; 95% CI lower: 1.103, 95% CI upper: 1.847). Females were 1.26 times more likely to fail compared to males, and increasing age was associated with a decreased likelihood of failure (Table 2). Failures were monitored and referred to their local clinic for a follow-up via the cloud-based data management system. The follow-up constituted a rescreen and diagnostic test if indicated. A total of 617 children attended their follow-up appointments ([Table 1](#T0001)), indicating a follow-up return rate of 39.4%. The overall follow-up screen referral rate was 40.5%. The mean test duration recorded was 170.7 seconds (SD 199.3) and 141.5 seconds (SD 188.2) for pass and failure rates respectively. Quality indices of test operators were monitored throughout the test period. [Table 4](#T0004) displays the increase in quality indices of the first three test operators over a five-month period.10.1080/16549716.2018.1467077-T0004Table 4.Quality index of test operators. Month 1Month 2Month 3Month 4Month 5Tester 1No. of children screened92270282142179Quality index (%)95969999100Tester 2No. of children screened71261245178189Quality index (%)9281969999Tester 3No. of children screened58202166100129Quality index (%)69909697100 Discussion {#S0004} ========== Performing hearing screenings within preschool aged populations is important to identify hearing health concerns that may interfere with language development and future school success \[[2](#CIT0002)\]. Within LMICs, ECD centers have the potential to serve as the first point of access to identify these children. This study provides a baseline for the implementation of a low-cost, ECD hearing screening program operated by CHWs using an mHealth point-of-care diagnostic and cloud-based data management and referral system. The referral rate in the preschool-aged population (3--6 years) using the hearScreen^TM^ application was 24.9%. Studies using conventional pure tone audiometry reported similar referral rates of 21.5% (2--6 years) and 21.3% (3.5--6 years) \[[27](#CIT0027),[28](#CIT0028)\]. In contrast, a recent study conducted using the hearScreen^TM^ smartphone application indicated a significantly lower referral rate of 4.3% for older children aged 5 to 7 \[[21](#CIT0021)\]. Higher referral rates in the current study are likely due to the fact that testing was conducted in a poor community where risk factors such as otitis media are higher, and due to the inclusion of younger children aged three and four years who presented with high referral rates. Results indicated a lower risk of failure in older compared to younger children. A previous study reported similar findings with a decrease in referral rate as the age of children increase \[[29](#CIT0029)\]. Referral rates of the current study were greatest in children aged 3 years (45.8%) as opposed to older children, which prompted the researchers to discontinue testing this age group. A study by Sideris and Glattke \[[28](#CIT0028)\] found that children younger than four years were often unable to perform pure tone screening, suggesting that pure tone audiometry requires a higher level of cognitive maturity. Additionally, the incidence of acute otitis media and otitis media with effusion is high in LMICs, with a higher incidence in children between the ages of two and five years, thus adding to a higher referral rate \[[30](#CIT0030)--[32](#CIT0032)\]. Environmental noise can also have an effect on referral rates, particularly when testing within an ECD setting where noise levels often fluctuate due to the absence of a sound-treated room, children leaving or entering the test environment, testers providing instructions, or groups of children walking past the test room \[[2](#CIT0002),[21](#CIT0021),[29](#CIT0029)\]. A smartphone-based mHealth solution like hearScreen^TM^ utilizes integrated noise monitoring, providing operators with real-time feedback on noise levels to allow testers to minimize noise levels before continuing with tests. Results and corresponding noise levels analyzed on the centralized cloud-based server indicated that only noise levels at 1 kHz had a significant effect on referral rates obtained in comparison to 2 and 4 kHz test frequencies. Previous studies using the hearScreen^TM^ application also reported similar effects when testing at the lower frequency of 1 kHz \[[21](#CIT0021)--[23](#CIT0023)\]. Increasing screening intensities to 30 dB HL at 1 kHz to compensate for high noise levels in future community-based studies could reduce the incidence of exceeded MPANLs and false-positive results, but may decrease the validity of the screening process as mild losses may be missed \[[23](#CIT0023),[29](#CIT0029)\]. The European Consensus Statement on Hearing, Vision, and Speech Screening in Pre-School and School-age Children indicated that although hearing screenings will produce over-referrals, false positives are preferred over false negatives \[[33](#CIT0033)\]. Gender effects were evident in smartphone hearing screening outcomes, with females more likely to refer than males. Mahomed-Asmail et al. \[[22](#CIT0022)\] also reported a significantly higher referral rate in school-aged females using conventional screening; however, these gender effects were not evident when using smartphone hearing screening. One possible reason was attributed to hair length or styles in girls that could have affected headphone placement. Further investigations on gender-specific results are needed. During follow-up appointments at clinics, ECD screening results were initially pulled from the cloud-based data system and analyzed. Thereafter, children received a second rescreen at the clinic. This was done in order to avoid unnecessary diagnostic assessments and to reduce the workload on already strained audiologists. Less than half (40.5%) of the children who failed their ECD screening failed the second screen at the clinic. The referral rate dropped by a further 5% (35.7%) when excluding more difficult to test children aged three to four years. Some influences which may have contributed to the difference in referral rates include ambient noise levels, headphone placement, visual distractions, and examiner instructions and expertise \[[29](#CIT0029)\]. Mean test durations, including the initial and immediate rescreen of both ears, were 177.8 and 174.3 seconds when testing at ECD centers and clinics respectively. Wu et al. \[[34](#CIT0034)\], reported a slightly shorter test duration of 149.4 seconds when using a smartphone hearing screening application. Higher test times at ECD centers could be attributed to the longer test times recorded with failure rates (258.5 seconds) in comparison to pass rates (68 seconds) due to the additional time required to recondition the child being rescreened. Furthermore, mean screen times were significantly higher for 3-year-olds when testing at ECD centers (193.4 seconds) and clinics (239.8 seconds). Only initiating a rescreen for failed frequencies during rescreens, instead of repeating the entire screen sequence, may reduce test times. Although parents were notified of their child's result via an SMS sent automatically from the cloud-based system, a low follow-up return rate of 39.4% was found. We suspect that this rate was affected by a long waiting period before follow-up appointments, parents changing their mobile phone number and not notifying the ECD center, and difficulties with taking leave from work, which may result in loss of income for informal workers. Other reasons that may account for non-attendance include lack of transportation, fear and uncertainty about the referral clinic, lack of education regarding hearing loss, and a lack of visibility of services \[[35](#CIT0035)\]. More precise reasons for non-attendance should be investigated in future studies. Incorporating a system to send a second text message reminder three days prior to a child's appointment may assist in improving follow-up rates \[[36](#CIT0036),[37](#CIT0037)\]. Alternatively, immediate onsite hearing assessments could be incorporated into the screening program. Immediate onsite automated audiometry could motivate parents to attend follow-up appointments by providing an immediate indication of the severity of a hearing problem and thereby also reduce the number of appointments that parents need to attend at clinics \[[38](#CIT0038)\]. Using onsite automated diagnostic audiometry, facilitated by the same smartphone, could ensure direct referrals for audiological or medical intervention, and may also reduce false positive results \[[39](#CIT0039)--[41](#CIT0041)\]. In turn, this will improve the cost-effectiveness, feasibility and credibility of the screening program with parents and physicians \[[29](#CIT0029)\]. Asynchronous cloud-based monitoring and surveillance allowed for quality indices of test operators to be monitored throughout the test period to ensure quality control ([Table 4](#T0004)). This guided project managers to provide feedback, additional information and more training to CHWs to ensure reliable test results. High-quality indices of test operators show that CHWs can successfully screen for hearing loss in children. The integrated cloud-based data management system also allowed for advanced features like location-based referrals via text message, reporting and determining follow-up return rates at clinics. Conclusion {#S0005} ========== ECD hearing screening programs using an mHealth point-of-care diagnostics and cloud-based data management and referral systems can be successfully implemented by CHWs within LMICs to identify children prior to school entry. This mHealth model provides a means to improve the cost-effectiveness, quality and efficiency of, and access to, hearing health services in poorer communities, particularly where hearing healthcare providers are unavailable. Quality control features including integrated noise monitoring, quality control indices of test operators and data management allow for asynchronous remote management to ensure reliable testing and to intervene when necessary. Age contributed significantly to high referral rates, suggesting an optimal screening age of five to six years of age. Environmental noise also posed a challenge when testing at the frequency of 1 kHz. Methods to improve the parental follow-up rate should be explored in future studies. Acknowledgments =============== We would like to thank all the participants involved with this study, including the team leaders and community health workers of the community of Mamelodi. Author contributions {#S0006} ==================== All authors listed have met all the International Committee of Medical Journal Editors (ICMJE) criteria for authorsip and have contributed equally to this work. Disclosure statement {#S0007} ==================== The hearScreen™ application is intellectual property owned, patented, and trademarked by the University of Pretoria. The product has been developed for commercialization. More information is available at www.hearscreen.co.za. Ethics and consent {#S0008} ================== Ethical clearance for this study was obtained from the Ethics Committee, University of Pretoria, South Africa. Assent of participants and consent from their parent/caregiver was obtained prior to any data collection. Paper context {#S0009} ============= Childhood hearing screening programs within low- and middle-income countries are few, due to limited financial and human resources. This study investigated the feasibility of a low-cost smartphone-based hearing-screening program for preschool children operated by non-specialist healthcare workers. Such programs can successfully aid in prevention, early identification and treatment of disabling hearing loss prior to school entry. Additionally, features such as automated test sequences, integrated noise monitoring and cloud-based data management allow for asynchronous point-of-care diagnostics in difficult to reach populations.

Fistulising disease complicates Crohn's disease (CD) in up to 40% of patients.[@b1]^--^[@b3] Fistulas rarely heal spontaneously and usually require medical therapy or surgery.[@b1] Antibiotics and immunosuppressive agents have been widely used for treatment, although their efficacy for the sustained closure of fistulas has not been proved.[@b1] Tumour necrosis factor (TNF) antagonists specifically target the elevated concentrations of TNF that contribute to the pathological inflammation in CD and represent a significant therapeutic advance in the treatment of patients with CD. Infliximab, a chimeric monoclonal antibody to TNF, has been shown to be effective for the treatment of fistulas.[@b4] In patients with an initial response to infliximab induction therapy, there is an increased likelihood of sustained response if infusions are continued every 8 weeks.[@b5] Clinical trials have demonstrated that adalimumab, a self-injected fully human monoclonal antibody to TNF, is effective for the induction and maintenance of remission in patients with moderate to severe CD.[@b6]^--^[@b10] In addition, in patients with CD who have lost response to or have become intolerant of infliximab, adalimumab has been demonstrated to be safe and efficacious in regaining a medical response.[@b11] The Crohn's Trial of the Fully Human Antibody Adalimumab for Remission Maintenance (CHARM)[@b6] was a large, phase III, randomised, double-blind, placebo controlled, 56-week study of patients with moderate to severe CD who may or may not have received TNF antagonist therapy previously. The primary objective was to assess the benefit of two adalimumab dosing regimens in maintaining clinical remission at 26 and 56 weeks. Among patients who responded to adalimumab, both dosing regimens (40 mg of adalimumab every other week (eow) and weekly) were statistically significantly more effective than placebo in maintaining remission to 56 weeks. Overall efficacy in fistula closure was also assessed, with significant effects of adalimumab therapy on fistula closure observed at both weeks 26 and 56.[@b6] The objectives of the present analysis of patients with fistulas in CHARM were as follows: (1) to describe fully the demographics, disease characteristics and safety outcome of these patients with fistulas; (2) to describe a new statistical approach that was developed to provide a more accurate and longitudinal approximation of fistula burden than was possible with previous approaches by calculating the number of draining fistulas per day for each individual patient and (3) to evaluate the 2-year maintenance of fistula healing during treatment with adalimumab in an open-label extension study (called the ADHERE trial---Additional Long-Term Dosing with HUMIRA to Evaluate Sustained Remission and Efficacy in Crohn's disease). METHODS {#s2} ======= Study design {#s2a} ------------ Detailed CHARM study methodology was reported previously.[@b6] CHARM was a 56-week, multicentre, randomised, double-blind, placebo controlled trial with a 4-week open-label induction period. Patients successfully completing CHARM could enroll in an open-label extension study (ADHERE) ([fig 1](#GUT-58-07-0940-f01){ref-type="fig"}). At baseline, patients received open-label adalimumab 80 mg subcutaneously followed by 40 mg at week 2. At week 4, all patients still enrolled were stratified by whether or not they achieved a clinical response (defined as achieving a decrease in the Crohn's disease activity index (CDAI) of ⩾70 points compared with baseline). Patients were then randomly assigned within each strata in a 1 : 1 : 1 ratio to one of three treatment groups: adalimumab 40 mg eow, adalimumab 40 mg weekly, or placebo.[@b6] {#GUT-58-07-0940-f01} At or after week 12, a patient with a flare or non-response could be switched to open-label treatment with 40 mg of adalimumab eow, which could be increased to 40 mg weekly if needed. Continued non-response on open-label adalimumab 40 mg weekly resulted in withdrawal from the study. A flare was defined as a recurrence of very active disease (CDAI increase ⩾70 points from week 4 and a CDAI \>220). Non-response was defined as failure to achieve 70-point response at any visit at or after week 12. All patients who successfully completed CHARM could enter the extension study (ADHERE). Patients who completed CHARM on placebo therapy received open-label adalimumab eow. Patients who had switched to open-label adalimumab eow or weekly continued on those regimens, and patients could switch from eow to weekly dosing in the event of flares or non-response. Of note, patients could not decrease their dosing intervals back to eow once they began receiving weekly dosing in either study. In addition, patient visits in ADHERE (weeks 0, 2, 4, 8 and 12 and then every 12 weeks) were much less frequent than in CHARM (every 2 weeks to week 8, every 4 weeks to week 20, every 6 weeks to week 32 and then every 8 weeks to study end in the original CHARM controlled trial). Patients {#s2b} -------- Patients all had moderately to severely active CD, defined as a CDAI score between 220 and 450.[@b12] Enrollment with a history of infliximab treatment was permissible only if infliximab had been discontinued at least 12 weeks before the screening visit and the patient had experienced an initial response to the agent (as judged by the investigator). Demographic and baseline disease severity data, concomitant medication use at baseline, previous history of TNF antagonist use and smoking history were collected for the entire population of patients in CHARM, but only patients with draining fistulas at screening and baseline are included in the subgroup analyses reported here. Baseline fistula data, that is the number of fistulas and their location, were noted and recorded. Efficacy assessments {#s2c} -------------------- The primary endpoint in CHARM was clinical remission and the study was statistically powered based on predicted remission rates. Fistula healing and the mean number of draining fistulas per day were predetermined secondary endpoints in CHARM; however, the study was not powered for secondary endpoints. Patients classified as having fistulas were only those with draining fistulas at both screening and at baseline of CHARM. Fistulas were assessed for spontaneous drainage or drainage with gentle compression at each study visit. Complete fistula healing was defined as the absence of drainage under both these circumstances. In addition to these standard assessments of fistula closure at set time points, a new, prespecified statistical method was used to determine the number of draining fistulas per day. Fistula closure rates for patients with fistulas at baseline were also stratified by baseline immunosuppressant use (yes or no), baseline CD-related antibiotic use (yes or no) and previous history of use of any TNF antagonist (yes or no), as well as the reason for the discontinuation of the previous TNF antagonist (loss of response, adverse reaction, both). Two post-hoc analyses were conducted to assess the sustained effect of adalimumab on fistula efficacy to 2 years of treatment (week 60 of ADHERE). The first analysis evaluated long-term fistula healing to 2 years of adalimumab therapy. The second analysis evaluated the maintenance of fistula healing in all patients with healed fistulas at the end of CHARM. In the primary fistula analysis specified in the statistical analysis plan, complete fistula healing (also referred to as complete fistula closure) was defined as the absence of draining fistulas for at least the last two post-baseline evaluations on or before the study visit in CHARM. Because of the reduced frequency of visits in ADHERE, fistula healing was defined as occurring at the time point of the visit, rather than for at least two consecutive visits, as was the case for the primary CHARM fistula analysis. For consistency, this definition of healing at a single time point for the 2-year analysis was used starting from the baseline of CHARM. Safety assessments {#s2d} ------------------ For evaluation of safety in the subset of patients with fistulas at baseline, serious adverse events and other adverse events of interest were collected throughout the study and its extension. Statistical analysis {#s2e} -------------------- Fistula results were evaluated for all randomly assigned patients with draining fistulas at both baseline and screening visits. The prespecified statistical analysis plan stipulated that fistula data from both adalimumab treatment groups (weekly and eow) be combined (owing to the anticipated small sample sizes of patients with fistulas) and compared with the placebo group. All analyses were based on two-sided tests with α  =  0.05. Continuous variables were compared using analyses of covariance adjusted for baseline values and discrete variables were compared using χ^2^ tests. To calculate the draining fistulas per day, a new statistical approach was developed a priori to express the total fistula burden of each individual patient over the course of their blinded participation in CHARM. Starting at week 0 and ending at week 56 (or study discontinuation), the number of draining fistulas at each pair of consecutive evaluations was averaged and multiplied by the elapsed days between the evaluations to obtain the number of draining fistula days. The total number of draining fistula days was summed from week 0 to week 56 (or study discontinuation) and divided by the total days in the study to obtain the average number of draining fistulas per day. The difference between each blinded adalimumab group and the placebo group was assessed with least-squares means from the analysis of variance with treatment as the only factor. Of note, the number of study days excluded days after discontinuation of double-blind study drug. Similarly, draining fistulas observed after study discontinuation or after the discontinuation of double-blind treatment were excluded from the analysis. If a missing evaluation had non-missing double-blind evaluations before and after it, the average of the two non-missing evaluations was used to estimate the missing evaluation. This analysis was conducted for all randomly assigned patients and randomly assigned responders (defined as patients who achieved ⩾70 point decreases in CDAI scores at week 4 in CHARM). The first analysis of long-term fistula healing to 2 years of adalimumab therapy included only those patients who were randomly assigned to adalimumab in CHARM and who entered ADHERE (analysis 1). Data were summarised at each visit using an observed-case analysis and non-responder imputation (NRI). In the second analysis of maintenance of fistula healing and response, all patients with healed fistulas at the end of CHARM (ie, both adalimumab and placebo-treated patients) were followed into ADHERE (analysis 2). In this analysis, data were summarised at weeks 24 and 60 of ADHERE (representing approximately 18 months and 2 years of therapy, respectively) using both observed and NRI analyses. With NRI analyses, the assumption is that patients with missing data or those who had left the trial were failures for the endpoint in question (ie, fistula healing). RESULTS {#s3} ======= Patients {#s3a} -------- The demographics and patient disposition for the overall CHARM patient population were previously described.[@b6] Of the 778 patients randomly assigned at week 4 to receive placebo (n  =  261), adalimumab 40 mg eow (n  =  260) or adalimumab 40 mg weekly (n  =  257), there was a total of 117 patients with draining fistulas at screening and baseline (placebo, n  =  47; 40 mg eow, n  =  30; and 40 mg weekly, n  =  40; [fig 2A](#GUT-58-07-0940-f02){ref-type="fig"}). [Figs 2B and C](#GUT-58-07-0940-f02){ref-type="fig"} display the numbers of patients included in the long-term fistula efficacy analyses (to 2 years from CHARM baseline) and the reasons for exclusion from analysis. Baseline characteristics are presented in [table 1](#GUT-58-07-0940-t01){ref-type="table"}. ###### Patient disposition. (A) Patient disposition to week 56 of the Crohn's Trial of the Fully Human Antibody Adalimumab for Remission Maintenance (CHARM). Randomly assigned responders achieved a 70-point or greater decrease in the Crohn's disease activity index score at week 4 in CHARM. (B) Patient disposition for sustained fistula response over 2 years of adalimumab therapy in patients randomly assigned to adalimumab in CHARM and who enrolled in the Additional Long-Term Dosing with HUMIRA to Evaluate Sustained Remission and Efficacy in Crohn's disease (ADHERE) study (analysis 1). (C) Patient disposition for sustained fistula response in all patients with healed fistulas at the end of CHARM (includes adalimumab and placebo-treated patients who entered ADHERE) (analysis 2). \*Denominator for non-responder imputation (NRI) analyses. †Denominator for observed-case analysis at the end of CHARM. ‡Denominator for observed-case analysis to 2 years of therapy. eow, every other week.    ###### Baseline demographics and clinical characteristics Characteristics All patients (n  =  854) Patients with fistulas at baseline ----------------------------------------------- -------------------------- ------------------------------------ ------------------ ------------------ Male, no (%) 326 (38.2) 15 (31.9) 34 (48.6) 49 (41.9) Age in years, mean (SD) 37.1 (11.9) 36.5 (10.1) 35.9 (12.2) 36.1 (11.4) Baseline CDAI score, mean (SD) 313.1 (62.0) 308.0 (59.3) 318.4 (55.8) 314.3 (57.2) CRP (mg/dl)     Mean (SD) 2.3 (3.4) 3.6 (4.7) 2.8 (3.0) 3.1 (3.8)     Median (range) 0.9 (0.02--35.0) 2.3 (0.12--28.7) 1.8 (0.06--12.3) 1.9 (0.06--28.7) CRP concentration ⩾1.0 mg/dl (10 mg/l), n (%) 407 (47.7) 32 (68.1) 42 (60.0) 74 (63.2) Previous TNF antagonist exposure, n (%) 424 (49.6) 31 (66.0) 41 (58.6) 72 (61.5) Concomitant medication, n (%)     Any glucocorticoid\* 376 (44.0) 20 (42.6) 29 (41.4) 49 (41.9)         Prednisone 244 (28.6) 14 (29.8) 18 (25.7) 32 (27.4)         Budesonide 100 (11.7) 9 (19.1) 2 (2.9) 11 (9.4)     Any immunosuppressive agent 399 (46.7) 26 (55.3) 31 (44.3) 57 (48.7)         Azathioprine 275 (32.2) 21 (44.7) 24 (34.3) 45 (38.5)         6-Mercaptopurine 81 (9.5) 4 (8.5) 3 (4.3) 7 (6.0)         Methotrexate 90 (10.5) 3 (6.4) 7 (10.0) 10 (8.5)     5-Aminosalicylates† 334 (39.1) 13 (27.7) 24 (34.3) 37 (31.6) Current smoker, no (%) 303 (35.5) 18 (38.3) 20 (28.6) 38 (32.5) \*Includes betamethasone, budesonide, dexamethasone, deflazacort, cortisone, cloprednol, fluocortolone, glucocorticoids, hydrocortisone, methylprednisolone, prednisolone, prednisone, paramethasone and prednylidene. †Aminosalicylic acid, balsalazide, mesalazine, olsalazine, sulfasalazine. CDAI, Crohn's disease activity index; CRP, C-reactive protein; TNF, tumour necrosis factor. Baseline characteristics of the patients with draining fistulas at screening and baseline were similar compared with the remainder of the patients, with the exception of numerically greater C-reactive protein concentrations and a previous history of TNF antagonist exposure in the patients with fistulising disease. Baseline fistula data are presented in [table 2](#GUT-58-07-0940-t02){ref-type="table"}. ###### Baseline fistula data Characteristic Placebo (n  =  47) Adalimumab 40 mg eow (n  =  30) Adalimumab 40 mg weekly (n  =  40) --------------------------------------- -------------------- --------------------------------- ------------------------------------ Draining cutaneous fistulas,\* no (%)     One 30 (63.8) 19 (63.3) 23 (57.5)     Two 7 (14.9) 6 (20.0) 7 (17.5)     Three 6 (12.8) 1 (3.3) 5 (12.5)     Four 4 (8.5) 4 (13.3) 5 (12.5) Perianal fistulas,† no (%)     One 29 (64.4) 19 (63.3) 21 (55.3)     Two 7 (15.6) 6 (20.0) 7 (18.4)     Three 6 (13.3) 1 (3.3) 5 (13.2)     Four 3 (6.7) 4 (13.3) 5 (13.2) \*Draining cutaneous fistulas includes perianal fistulas (n  =  113) and abdominal fistulas (n  =  4). †n  =  45 for placebo; n  =  30 for 40 mg adalimumab eow; n  =  38 for 40 mg adalimumab weekly. eow, every other week. Efficacy {#s3b} -------- Adalimumab therapy was associated with progressive increases in fistula closure over time, with separation in the rates of complete closure between placebo and adalimumab groups evident as early as 2 weeks after randomisation. Statistically significant differences in fistula closure between placebo and adalimumab were first observed at 16 weeks ([fig 3](#GUT-58-07-0940-f03){ref-type="fig"}). Baseline immunosuppressant or CD-related antibiotic use had no apparent effect on the rates of fistula closure in patients receiving adalimumab or placebo at weeks 26 or 56, although the numbers of patients in each group were relatively small ([table 3](#GUT-58-07-0940-t03){ref-type="table"}). In addition, whether patients were naive to or experienced with TNF antagonists before receiving adalimumab did not appear to effect fistula healing (data not shown). ###### Percentage of patients with no draining fistulas at weeks 26 and 56 for all patients with fistulas at baseline (n  =  117) stratified by baseline concomitant therapies Subgroup No draining fistulas\* --------------------------------------- ------------------------ ---------------- No baseline immunosuppressant use     Placebo (n  =  21) 19 19     Both adalimumab groups (n  =  39) 33 (p = 0.369) 36 (p = 0.241) Baseline immunosuppressant use     Placebo (n  =  26) 8 8     Both adalimumab groups (n  =  31) 26 (p = 0.092) 29 (p = 0.051) No baseline CD-related antibiotic use     Placebo (n  =  28) 14 14     Both adalimumab groups (n  =  44) 32 (p = 0.162) 36 (p = 0.059) Baseline CD-related antibiotic use     Placebo (n  =  19) 11 11     Both adalimumab groups (n  =  26) 27 (p = 0.264) 27 (p = 0.264) \*Patients who had no draining fistulas at the last two post-baseline evaluations in the double-blind period on or before the week 26 or week 56 visits were classified as no; otherwise, patients were classified as yes. p Values from χ^2^ tests comparing both adalimumab groups combined versus placebo. CD, Crohn's disease. {#GUT-58-07-0940-f05} The new statistical methodology for determining fistula burden was used to calculate the number of draining fistulas per day during the double-blind period. For all randomly assigned patients and also all randomly assigned responder patients (those with week 4 CDAI decreases ⩾70 points vs baseline), there were significant decreases in the mean number of draining fistulas per day among adalimumab-treated patients compared with placebo-treated patients during the double-blind treatment period ([fig 4](#GUT-58-07-0940-f04){ref-type="fig"}). For all randomly assigned patients, the mean number of draining fistulas per day was 1.34 for placebo compared with a mean of 0.88 for the combined adalimumab groups (p = 0.002). Results were similar for the randomly assigned responders (mean 1.15 for placebo, mean 0.76 for adalimumab groups combined, p = 0.043). The effect of adalimumab on the number of draining fistulas in each subgroup (based on immunosuppressant, antibiotic, or previous TNF antagonist exposure) was similar to that observed for the overall placebo and adalimumab groups (data not shown). Fistula healing over time for all patients with fistulas who were randomly assigned to adalimumab is provided in [table 4](#GUT-58-07-0940-t04){ref-type="table"} (observed analysis). Of the adalimumab-treated patients with fistulas at baseline (n  =  70) approximately 60% (22/37) had healed fistulas after 2 years of therapy ([table 4](#GUT-58-07-0940-t04){ref-type="table"}). Long-term fistula healing over time using NRI analysis is provided in [table 5](#GUT-58-07-0940-t05){ref-type="table"}. Because patients in CHARM and ADHERE were managed according to inflammatory activity and therapy dosage changes were not predicated on fistula response (fistula is part of CDAI, which is an inflammatory scale assessing disease activity), NRI is a very conservative approach to the data analysis. ###### Long-term efficacy of fistula healing with adalimumab to 2 years from CHARM baseline: observed analysis\* Time point Observed analysis n/No (%) --------------------- ---------------------------- 6 Months in CHARM 32/58 (55) 1 Year in CHARM 29/50 (58) 24 Weeks in ADHERE 25/42 (60) 36 Weeks in ADHERE 26/40 (65) 48 Weeks in ADHERE 23/37 (62) 60 Weeks in ADHERE† 22/37 (59) \*Includes all adalimumab-treated patients with fistulas at the baseline of the Crohn's Trial of the Fully Human Antibody Adalimumab for Remission Maintenance (CHARM) remaining in CHARM or the Additional Long-Term Dosing with HUMIRA to Evaluate Sustained Remission and Efficacy in Crohn's disease study (ADHERE; open-label extension) at the time points specified. Fistula efficacy defined as complete healing/closure of draining fistulas at time point since baseline of CHARM. †60 weeks in ADHERE represents approximately 2 years of adalimumab therapy. ###### Long-term efficacy of fistula healing with adalimumab to 2 years from CHARM baseline: NRI analysis\* Time point NRI (n  =  70) --------------------- ---------------- 6 Months in CHARM 32 (46) 1 Year in CHARM 29 (41) 24 Weeks in ADHERE 25 (36) 36 Weeks in ADHERE 26 (37) 48 Weeks in ADHERE 23 (33) 60 Weeks in ADHERE† 22 (31) \*Includes all adalimumab-treated patients (every other week and weekly groups combined, N  =  70) with fistulas at baseline of the Crohn's Trial of the Fully Human Antibody Adalimumab for Remission Maintenance (CHARM). Fistula efficacy defined as complete healing/closure of draining fistulas at time point since baseline of CHARM. †60 weeks in the Additional Long-Term Dosing with HUMIRA to Evaluate Sustained Remission and Efficacy in Crohn's disease (ADHERE) study represents approximately 2 years of adalimumab therapy. NRI, non-responder imputation. Long-term maintenance of fistula healing rates for patients with healed fistulas at the end of CHARM (n  =  40) are provided in [fig 5](#GUT-58-07-0940-f05){ref-type="fig"}. Of these patients, 90% (28/31) maintained healed fistulas after one additional year of treatment in ADHERE (2 years from CHARM baseline, observed analysis). In the NRI analysis, 75% of patients maintained healing to 2 years of therapy. {#GUT-58-07-0940-f03} Safety {#s3c} ------ Adalimumab was well tolerated for up to 56 weeks of treatment in patients with fistulas at baseline ([table 6](#GUT-58-07-0940-t06){ref-type="table"}), similar to the findings for all randomly assigned patients reported by Colombel *et al*.[@b6] Safety data in the fistula cohort to 2 years from CHARM baseline are presented in [table 7](#GUT-58-07-0940-t07){ref-type="table"}. With respect to the serious events of abscess, eight events were reported in the 70 adalimumab-treated patients in the fistula cohort who received 2 years of therapy; one additional event was reported in a patient originally randomly assigned to placebo. ###### Summary of safety in adalimumab and placebo-treated patients with fistulas to 56 weeks in CHARM Event Placebo (n  =  47) All adalimumab (n  =  70) -------------------------------------------------------------- -------------------- --------------------------- Adverse event 38 (80.9) 59 (84.3) Serious adverse event 5 (10.6) 9 (12.9) Adverse event leading to discontinuation of study medication 3 (6.4) 4 (5.7) Infectious adverse event 16 (34.0) 31 (44.3) Serious infectious adverse event\* 2 (4.3) 5 (7.1) Abscess (all) 5 (10.6) 8 (11.4) Malignant neoplasm 0 0 Injection-site reaction (all) 2 (4.3) 3 (4.3) Opportunistic infection† 1 (2.1) 0 Congestive heart failure 0 0 Demyelinating disorder 0 0 Death 0 0 \*Both of the placebo-treated patients had an abdominal abscess. The serious infectious adverse events for the five adalimumab-treated patients were a pulmonary embolus with pneumonia (n  =  1); intra-abdominal abscess (n  =  1); perianal abscess (n  =  2) and scrotal abscess (n  =  1). †Oral candidiasis. CHARM, Crohn's Trial of the fully Human Antibody Adalimumab for Remission Maintenance. ###### Summary of safety in cohort of patients with fistulas to 2 years from CHARM baseline Event All patients with fistula (adalimumab and placebo) (n  =  117) All adalimumab patients with fistula (eow and weekly combined) (n  =  70) -------------------------------------------------------------- ---------------------------------------------------------------- --------------------------------------------------------------------------- Adverse event 112 (95.7) 70 (100) Serious adverse event 33 (28.2) 22 (31.4) Adverse event leading to discontinuation of study medication 19 (16.2) 13 (18.6) Infectious adverse event 76 (65.0) 53 (75.7) Serious infectious adverse event 13 (11.1)\* 10 (14.3)† Serious gastrointestinal adverse event 15 (12.8)‡ 13 (18.6)§ Malignant neoplasm 1 (0.9) 0 Injection-site pain 7 (6.0) 4 (5.7) Opportunistic infection 0 0 Congestive heart failure 0 0 Demyelinating disorder 0 0 Death 0 0 \*Fifteen events were reported in 13 patients: nine events of abscess (four perianal, two abdominal, one rectal, one scrotal, one not specified), one event of clostridial infection, one event of sepsis, one event of pneumonia, one event of tuberculosis, one event of otitis media and one event of post-procedural infection. †Twelve events were reported in 10 patients: eight events of abscess (three perianal, two abdominal, one rectal, one scrotal, one not specified), one event of clostridial infection, one event of pneumonia, one event of tuberculosis and one event of otitis media. ‡Seventeen events were reported in 15 patients: 10 events of Crohn's disease (CD), four events of obstruction (two intestinal, two small intestinal), two events of fistula (one anal, one intestinal) and one event of abdominal mass. §Fifteen events were reported in 13 patients: eight events of CD, four events of obstruction (two intestinal, two small intestinal), two events of fistula (one anal, one intestinal), and one event of abdominal mass. CHARM, Crohn's Trial of the fully Human Antibody Adalimumab for Remission Maintenance. DISCUSSION {#s4} ========== These data extend and define the beneficial effect of adalimumab on fistula healing in patients with CD and provide the first long-term fistula healing results with an anti-TNF therapy. This study also introduces and utilises a new methodology for evaluating the effect of treatment in patients with fistulising disease. The demographic characteristics of patients with fistulising disease did not differ substantially from the overall study population in CHARM. In this subgroup of patients whose CD was often difficult to manage because of fistulising disease, adalimumab still demonstrated a consistent and significant reduction of draining fistulas when compared with placebo in patients with moderate to severe CD. The effect of adalimumab on fistula closure was durable, as demonstrated by three-quarters of the adalimumab-treated patients (NRI; 90% with observed analyses) who had healed fistulas at the end of CHARM maintaining closure following an additional year of open-label therapy. The efficacy of adalimumab was consistent in subgroups of patients with fistulas who were or were not receiving concomitant therapy with immunosuppressive agents or antibiotics, as well as in patients naive to and experienced with TNF antagonist therapy. However, caution should be exercised in interpreting these analyses as a result of the small numbers of patients in each subgroup. ###### Sample calculation of the mean number of draining fistulas per day. (A) Hypothetical patient 1---minimal disease activity yet not achieving complete healing\* at either week 26 or week 56. (B) Hypothetical patient 2---significant disease activity yet achievement of complete healing\* at week 56. \*Complete healing, no draining fistulas for at least the last two post-baseline evaluations in the double-blind period.   In addition to standard assessments of fistula closure at set time points, a new, prespecified statistical method was used to determine the number of draining fistulas per day. Specific examples of the statistical analysis are depicted in [fig 6](#GUT-58-07-0940-f06){ref-type="fig"}. With the definition of complete fistula healing used in this analysis, hypothetical patient 1 ([fig 6A](#GUT-58-07-0940-f06){ref-type="fig"}) would have been assessed as not having successful closure of fistulas at either individual time point (week 26 or week 56), whereas hypothetical patient 2 ([fig 6B](#GUT-58-07-0940-f06){ref-type="fig"}) would have been assessed as having successful closure at week 56. With the new method of measuring fistulas per day reported here, it becomes evident that hypothetical patient 1 actually had a lower fistula burden over the study period than did hypothetical patient 2. Evaluation of "fistulas per day" more accurately reflects disease activity over the entire study period, instead of focusing on fistula counts at single points in time as does the definition of fistula healing. By providing a longitudinal perspective, this new methodology may yield a more accurate picture of the total fistula experience of each individual patient over the full course of participation in the study. This analysis, which facilitates the comparison of subgroups of patients within an individual study, was prespecified in the statistical analysis plan for the CHARM study. One limitation of this method is that between the visits, the numbers of draining fistulas on each day are assumed to remain the same and approximate the average number of draining fistulas observed at each visit. This assumption requires validation in appropriately constructed clinical trials. {#GUT-58-07-0940-f04} Infliximab is the only other TNF antagonist shown to be effective in inducing and maintaining fistula closure.[@b4] [@b5] In the ACCENT II study (Crohn's Disease Clinical Study Evaluating Infliximab in a New Long-Term Treatment Regimen II), 36% of patients who had responded to induction therapy and received maintenance infliximab therapy maintained complete fistula closure compared with 19% of patients receiving placebo (p = 0.009).[@b5] In ACCENT II, patients were included based on a minimal duration of fistula activity; only patients who responded to initial induction therapy were included in the analysis of maintenance efficacy.[@b5] In CHARM, patients were included based on the overall activity of their CD and not on fistula activity, and all patients with fistulas were included in the analysis irrespective of initial response to adalimumab. Despite these differences, the fistula closure rates from ACCENT II at week 54 are generally comparable to the fistula closure rates observed with adalimumab treatment at week 56. In two studies of certolizumab pegol (PRECiSE 1 and 2), a pegylated Fab′ fragment of an anti-TNF monoclonal antibody, the effect of certolizumab on fistula closure was not statistically significantly different compared with placebo.[@b13] [@b14] Patients with fistulas at baseline tolerated sustained adalimumab treatment well. The safety profile of adalimumab in CHARM was consistent with studies of adalimumab in rheumatological disorders and with previous studies of patients with CD. The safety of TNF antagonists in patients with fistulising disease is of concern, because of the greater risk of infection at baseline (especially abscesses) for these patients.[@b15] In the CHARM fistula cohort, there were no differences in the rates of adverse events, including infectious adverse events and, specifically, abscesses, in patients who received placebo compared with those who received adalimumab at either dosage. This observation is consistent with previous reports.[@b16] CONCLUSIONS {#s5} =========== Adalimumab therapy resulted in a significant decrease in the number of draining fistulas per day compared with placebo. Significant and complete healing of draining fistulas was observed in adalimumab-treated patients, and long-term healing of draining fistulas was maintained over 2 years from the baseline of the CHARM study. Adalimumab was well tolerated, with a safety profile consistent with previous studies in patients with CD. The authors would like to thank Dana L Randall and Cathryn Carter of Arbor Communications Inc, Ann Arbor, Michigan, USA, as well as Michael A Nissen of Abbott Laboratories, for writing and editing assistance. **Funding:** This study was supported by a research grant from Abbott Laboratories, Abbott Park, Illinois, USA. The writing and editing assistance for this paper was also funded by Abbott. [^1]: ▸ Competing interests: Declared (the declaration can be viewed on the *Gut* website at <http://gut.bmj.com/content/vol58/issue7>)

Introduction {#Sec1} ============ Bovine coronavirus (BCoV) is amember of the order *Nidovirales*, *Coronaviridae* family, which was recently classified as member of the specie *Betacoronavirus 1* in the genus *Betacoronavirus* \[[@CR1]\]. BCoV are frequently circulating in cattle farms worldwide, causing both enteric and respiratory disease in calves and adult cattle \[[@CR2]\]. Because the respiratory and enteric disorders are similar to the other infectious diseases, a specific test is needed for a conclusive diagnosis of BCoV infection. Besides sensitive tests are required to detect BCoV especially at early or late stages of disease when they have low levels of viral shedding. The current methods used for the diagnosis of BCoV consist mostly of the detection of viral RNA by conventional ( RT-PCR) and real-time reverse transcription polymerase chain reaction (qRT- PCR) \[[@CR3]--[@CR6]\]. Recently, isothermal nucleic acid amplification techniques, such as recombinase polymerase amplification (RPA) assay and reverse transcription loop-mediated isothermal amplification (RT-LAMP), have been developed for rapid detection of BCoV \[[@CR7], [@CR8]\]. However, there are still few reports evaluating these new assays in large-scale epidemiological studies of BCoV infections. In this chapter, we describe a sensitive and specific conventional RT-PCR assay that has been successfully applied for diagnosis of both enteric and respiratory bovine coronaviruses \[[@CR9]--[@CR12]\]. Materials {#Sec2} ========= RNA Extraction {#Sec3} -------------- Sodium dodecyl sulfate (SDS) 10 %.Ultrapure phenol--chloroform--isoamyl alcohol (25:24:1, v/v) (Invitrogen).Silicon Dioxide (SiO~2~) (Sigma).Guanidine isothiocyanate (GuSCn) (Invitrogen).Acetone PA.Ethanol solution 70 % (in water).Diethylpyrocarbonate (DEPC)-treated water (Invitrogen).EDTA.Hydrochloric acid 32 %---HCl 32 % (Sigma).Triton X-100 (Invitrogen).Tris (hydroxymethyl)aminomethane (TRIS).Lysis buffer L6.Washing buffer L2. Clinical Samples (*See***Note**[**1**](#Sec23){ref-type="sec"}) {#Sec4} --------------------------------------------------------------- Fecal samples: The samples were prepared either as 10 % (w/v) suspensions of solid or semisolid feces in 0.01 M phosphate-buffered saline (PBS) pH 7.2 (137 mM NaCl; 3 mM KCl; 8 mM Na~2~HPO~4~; 15 mM KH~2~PO~4~) or as 50 % (v/v) suspensions of liquid feces in 0.01 M PBS and centrifuged at 3000 × *g* for 15 min at 4 °C. The supernatant is transferred to a sterile tube. Separate an aliquot of 400 μl for RNA extraction. Nasopharyngeal swab samples: the tip of swab containing nasopharyngeal secretions is soaked in 1 ml of sterile saline solution or 0.01 M PBS. The swab is vortex-mixed for 15 s and then discarded. Centrifuge at 3000 × *g* for 15 min at 4 °C. The supernatant is transferred to a sterile tube. Separate an aliquot of 450 μl for RNA extraction. Preparing Reagent and Solutions Used for RNA Extraction {#Sec5} ------------------------------------------------------- ### Silica Hydration Process {#Sec6} Silica suspension is prepared as described by Boom et al. \[[@CR13]\] with minor modifications.Suspend six grams of silicon dioxide (SiO~2~) in 50 ml of sterile distilled water in a glass graduated cylinder.Slowly stir and keep at rest at room temperature for 24 h for the silica coarse particles to settle.Remove and discard 43 ml of the supernatant from the top with vacuum suction or pipette. Then, resuspend the silica by adding 50 ml of sterile distilled water.Slowly stir and leave for 5 h.Remove and discard 44 ml of the supernatant from the top with vacuum suction or pipette.Add 60 μl of concentrated HCl (32 % w/v) to adjust pH (pH = 2.0).Mix the final content (resulting suspension of silica coarse) and divide into small aliquots (4 ml) into glass bottles with autoclavable cap, tightly closed.Sterilize in an autoclave at 121 °C for 20 min. Store at 4 °C. Prepare of Lysis Buffer L6 and Washing Buffer L2 {#Sec7} ------------------------------------------------ Prepare the following fresh solutions before preparing lysis buffer: ### 0.1 M Tris--HCl (pH 6.4) {#Sec8} Dissolve 2.67 g of TRIS in 180 ml of distilled water. Add HCl 32 % to adjust the pH to 6.4. Stir vigorously on a magnetic stirrer. Fill up to volume 220 ml with distilled water and set aside at room temperature until use. ### 0.2 M EDTA (pH 8.0) {#Sec9} Add 1.86 g of EDTA in 15 ml of distilled water. Adjust the pH to 8.0 with NaOH. Fill up to volume 25 ml with distilled water and set aside at room temperature until use (*see* **Note** [**2**](#Sec23){ref-type="sec"}). ### Lysis Buffer L6 {#Sec10} Dissolve 120 g of GuSCN in 100 ml of 0.1 M Tris--HCl (pH 6.4) in a beaker.Add 22 ml of a 0.2 M EDTA solution (pH 8.0).Add 2.6 g of Triton X-100.Homogenize vigorously the solution (*see* **Note** [**3**](#Sec23){ref-type="sec"}).Transfer the solution into a glass bottle with autoclavable cap.Sterilize in an autoclave at 121 °C for 20 min. Store at 4 °C (*see* **Note** [**4**](#Sec23){ref-type="sec"}). ### Washing Buffer L2 {#Sec11} Dissolve 120 g of GuSCN in 100 ml of 0.1 M Tris--HCl (pH 6.4) in a beaker.Homogenize vigorously the solution (*see* **Note** [**3**](#Sec23){ref-type="sec"}).Transfer the solution into a glass bottle with autoclavable cap.Sterilize in an autoclave at 121 °C for 20 min. Store at 4 °C. Reverse Transcription {#Sec13} --------------------- Moloney Murine Leukemia Virus Reverse Transcriptase (M-MLV RT), 200 U/μl (Invitrogen).5× First-Strand Buffer: 250 mM Tris--HCl (pH 8.3), 375 mM KCl, 15 mM MgCl~2~ (Invitrogen).0.1 M dithiothreitol (DTT) (Invitrogen).20 μM reverse primer BCoV2: 5′-TGTGGGTGCGAGTTCTGC-3′ (nt 940--957).Deoxynucleotide triphosphates (dNTP) mix, 10 mM each (Invitrogen).DEPC-treated water (Invitrogen).Thermocycler (Applied Biosystems). Polymerase Chain Reaction {#Sec14} ------------------------- Platinum *Taq*DNA polymerase, 5 U/μl (Invitrogen).10× PCR buffer (Invitrogen).50 mM MgCl~2~ solution (Invitrogen).Deoxynucleotide triphosphates (dNTP) mix, 10 mM each.20 μM PCR forward primer, BCoV1: 5′-CGATGAGGCTATTCCGAC-3′ (nt 504--521).20 μM PCR reverse primer, BCoV2: 5′-TGTGGGTGCGAGTTCTGC-3′ (nt 940-957).Thermocycler (Applied Biosystems). Semi-nested Polymerase Chain Reaction {#Sec15} ------------------------------------- Platinum *Taq*DNA polymerase, 5 U/μl (Invitrogen).10× PCR buffer (Invitrogen).50 mM MgCl~2~ solution (Invitrogen).Deoxynucleotide triphosphates (dNTP) mix, 10 mM each.20 μM PCR forward primer, BCoV3: 5′-TTGCTAGTCTTGTTCTGGC-3′ (nt 707--725).20 μM PCR reverse primer, BCoV2: 5′-TGTGGGTGCGAGTTCTGC-3′ (nt 940-957).Thermocycler (Applied Biosystems). Gel electrophoresis {#Sec16} ------------------- 1× TRIS--borate--EDTA buffer pH 8.0 (0.89 M TRIS; 0.89 M boric acid; 0.02 M EDTA).UltraPure agarose (Invitrogen).123 bp DNA ladder markers (Invitrogen).Ethidium bromide, 10 mg/ml.Agarose gel electrophoresis system (Apelex).Power supply (Apelex).Gel Documentation and Analysis System (Kodak).Gel loading buffer (6×): 10 mM Tris--HCl (pH 7.6), 0.03 % bromophenol blue, 0.03 % xylene cyanol, 60 % glycerol, and 60 mM EDTA. Methods {#Sec17} ======= The protocols described below are routinely used for clinical diagnosis of BCoV in cases of neonatal diarrhea, winter dysentery in adult cattle, and respiratory syndrome in cattle of all ages \[[@CR9], [@CR10]\]. Firstly, we describe how to obtain high quality nucleic acid total for routine molecular biology applications such as PCR and sequencing. We present an efficient and reproducible protocol for extracting RNA from feces and nasopharyngeal swab samples for BCoV diagnosis. This protocol, which we call phenol chloroform silica method, is based on cellular lysis and protein denaturation by SDS and phenol--chloroform--isoamyl alcohol treatment followed by nucleic acid purification by the guanidinium thiocyanate and silica \[[@CR14]\]. Guanidinium thiocyanate is an excellent protein denaturant and hence very effective in inactivating nucleases. Using this protocol extraction it is possible to obtain high quality of RNA in the laboratory on a routine basis without the need for expensive commercial kits extractions. For RT-PCR assay we chose the N gene because it is highly conserved among BCoV strains (*see* **Note** [**5**](#Sec23){ref-type="sec"}). Also it is known that the N protein is the most abundant antigen in coronavirus-infected cells because its RNA template is the smallest and it has the most abundant sgRNA (subgenomic RNA) during transcription \[[@CR15]\]. This indicates that there is more available RNA for the N gene than for the other BCoV protein genes. Consequently, detection of the N gene RNA might be advantageous due to its high abundance in cells, facilitating a high sensitivity of the diagnostic technique. RNA Extraction {#Sec18} -------------- Transfer 450 μl of the fecal or nasopharyngeal suspensions into a 1.5 ml polypropylene microtube. Include at least one positive control (HRT-18 cell culture-adapted BCoV Mebus strain or BCoV wild type strains) and one negative control (DEPC-treated water) in all the RNA extraction procedures.Add 50 μl of SDS 10 % to a final concentration of 1 % (v/v).Vortex for 5 s and incubate in a water bath at 56 °C for 30 min.Add an equal volume (500 μl) of ultrapure phenol--chloroform--isoamyl alcohol to the microtube (*see* **Note** [**6**](#Sec23){ref-type="sec"}). For nasopharyngeal swab samples the treatment with phenol--chloroform--isoamyl alcohol (steps 2--5) are not performed.Vortex vigorously for 20 s to mix the phases and incubate in a water bath at 56 °C for 15 min.Centrifuge at 10,000 × *g* for 10 min to separate the phases (upper aqueous phase, interface, and lower organic phase).Use a micropipette to remove the aqueous phase (upper aqueous phase) into a new 1.5 ml polypropylene microtube. Discard the microtube containing the interface (between the organic and aqueous phases) and organic phase (lower) (*see* **Notes** [**7**](#Sec23){ref-type="sec"} and [**8**](#Sec23){ref-type="sec"}).Add 30 μl of hydrated silica (*see* **Note** [**9**](#Sec23){ref-type="sec"}).Add 900 μl of lysis buffer L6.Vortex vigorously for 20 s.Place the microtubes in a rack and then incubate on an orbital shaker for 30 min at room temperature.Briefly centrifuge the microtubes at 10,000 × *g* for 30 s to pellet the silica resin.Gently discard the supernatant by inversion (*see* **Note** [**10**](#Sec23){ref-type="sec"}).Add 500 μl of ice-cold washing buffer L2 to the pellet. Vortex for 5 s to resuspend the silica pellet.Briefly centrifuge the microtubes at 10,000 × *g* for 30 s to pellet the silica resin.Gently discard the supernatant by inversion (*see* **Note** [**10**](#Sec23){ref-type="sec"}).Once again, add 500 μl of ice cold washing buffer L2 to the pellet. Vortex for 5 s to resuspend the silica resin. Repeat steps 13 and 14.Add 1 ml of ice-cold ethanol 70 % to the silica pellet. Vortex for 5 s to resuspend the silica pellet.Briefly centrifuge the microtubes at 10,000 × *g* for 30 s to pellet the silica resin.Gently discard the supernatant by inversion (*see* **Note** [**10**](#Sec23){ref-type="sec"}).Once again add 1 ml of ice cold ethanol 70 % to the pellet. Vortex for 5 s to resuspend the silica resin. Repeat steps 17 and 18.Add 1 ml of ice-cold acetone PA to the silica pellet. Vortex for 5 s to resuspend the silica pellet.Centrifuge the microtubes at 10,000 × *g* for 60 s to pellet the silica resin.Gently discard the supernatant by inversion (*see* **Note** [**10**](#Sec23){ref-type="sec"}).Place the microtubes with open lids in an inverted position over a clean filter paper to drain off the acetone excess.Return the tubes to their normal position and keep them with lids open in a thermoblock (dry block heater) at 56 °C for 15 min to evaporate the acetone and dry the silica pellet.Add 50 μl of DEPC water. Vortex for 5 s to resuspend the silica pellet.Incubate the microtubes with the lids closed in a water bath at 56 °C for 15 min to elute nucleic acid from the silica resin.Centrifuge the microtubes at 10,000 × *g* for 2 min to pellet the silica resin.Remove the supernatant (nucleic acid eluted) with a micropipette into a new 0.5 ml polypropylene microtube. Discard the microtube containing the silica pellet.Store the eluted nucleic acid at −20 or −80 °C until use (*see* **Note** [**11**](#Sec23){ref-type="sec"}). Reverse Transcription (*See***Note**[**12**](#Sec23){ref-type="sec"}) {#Sec19} --------------------------------------------------------------------- Prepare a reverse transcription (RT) master mix in a volume sufficient for the number of reactions plus 1 in a sterile 1.5 ml polypropylene microtube (label it "RT mix"). The volume of each RT reagent per reaction and the initial and final concentrations are shown in Table [1](#Tab1){ref-type="table"}.Table 1Components of reverse transcription reactionReagentVolume per reaction (μl)Volume mix for *N* reactions (μl)Final concentration5× First strand buffer44 × *N*1×0.1 mM DTT22 × *N*0.01 mM10 mM dNTP1*N*0.5 mMMMLV reverse transcriptase (200 U/μl)0.50.5 × *N*100 U/reactionWater2.52.5 × *N*--Total volume of RT master mix1010 × *N*--Vortex and centrifuge the tube briefly. Keep the microtube "RT mix" on an ice bath until use.Label appropriately the 0.5 μl polypropylene microtubes with the sample identification.Add to each microtube 5 μl of the corresponding sample (eluted RNA), 1 μl of reverse primer BCoV2 and 4 μl of DEPC-treated water.Incubate the microtubes at 97 °C in a heat block for 4 min.Immediately after denaturing place on ice for at least 5 min.Add 10 μl of "RT mix" solution into each microtube.Vortex and centrifuge the tubes briefly.In a thermocycler, incubate the microtubes at 42 °C for 30 min, inactivate the transcription reaction at 95 °C for 5 min and then chill the samples on ice bath.Store the cDNA samples at −20 °C until use. PCR Assay {#Sec20} --------- Prepare a PCR master mix in a volume sufficient for the number of reactions plus 1 in a sterile 1.5 ml polypropylene microtube and label it "PCR mix." The volume of each PCR reagent per reaction and the initial and final concentrations are shown in Table [2](#Tab2){ref-type="table"}.Table 2Components of the PCR assayReagentVolume per reaction (μl)Volume mix for N reactions (μl)Final concentration10× First strand buffer7.57.5 × *N*1.5×MgCl~2~, 50 mM22 × *N*2 mMdNTP, 10 mM44 × *N*0.8 mMReverse primer BCoV2 (20 μM)1N0.4 μMForward primer BCoV1 (20 μM)1N0.4 μMDNA polymerase (5 U/μl)0.50.5 × *N*2.5 U/reactionWater2626 × *N*--Total4242 × *N*--*N* = number of 0.2 ml tubesVortex and centrifuge the tube briefly. Keep the " PCR mix" on ice bath.Dispense 42 μl of the " PCR mix" into separate 0.5 polypropylene microtube and label the tubes accordingly.Add 8 μl of cDNA generated from the reverse transcription reactions into the correspondent tubes.Vortex and centrifuge the tubes briefly.Run the PCR under the conditions shown in Table [3](#Tab3){ref-type="table"}.Table 3Conditions for PCR assayStepTemperature (°C)Time1. Heat activation944 min2. Thermal cycling (40 cycles) Denaturing step941 min Annealing step551 min Extension721 min3. Final extension727 min4. Soak 4∞ Semi-nested PCR Assay {#Sec21} --------------------- Prepare a SN- PCR master mix in a volume sufficient for the number of reactions plus 1 in a sterile 1.5 ml polypropylene microtube and label it "SN-PCR mix". The volume of each PCR reagent per reaction and the initial and final concentration are shown in Table [4](#Tab4){ref-type="table"}. Vortex and centrifuge the tube briefly. Keep the "SN-PCR mix" on ice bath.Table 4Components of the semi-nested PCR assayReagentVolume per reaction (μl)Volume mix for *N* reactions (μl)Final concentration10× First strand buffer55 × *N*1×MgCl~2~, 50 mM22 × *N*2 mMdNTP, 10 mM44 × *N*0.8 mMReverse primer BCoV2 (20 μM)1N0.4 μMForward primer BCoV3 (20 μM)1N0.4 μMDNA polymerase (5 U/μl)0.50.5 × *N*2.5 U/reactionWater33.533.5 × *N*--Total4747 × *N*--*N* = number of 0.2 ml tubesDispense 47 μl of the "SN- PCR mix" into separate 0.5 polypropylene microtube and label the tubes accordingly.Add 3 μl of the first amplification product ( PCR assay) into the correspondent tubes.Vortex and centrifuge the tubes briefly.Run the SN- PCR under the conditions shown in Table [5](#Tab5){ref-type="table"}.Table 5Conditions for the semi-nested PCR assayStepTemperature (°C)Time1. Heat activation944 min2. Thermal cycling (30 cycles) Denaturing step941 min Annealing step551 min Extension721 min3. Final extension727 min4. Soak 4∞After the run, analyze the SN- PCR products by gel electrophoresis. Alternatively, the products can be kept at −20 °C for short-term storage. Agarose Gel Electrophoresis {#Sec22} --------------------------- Prepare 2 % agarose gel by weighing out 2 g de agarose powder (for gel-tray dimension 14 × 10 × 0.7 cm).Add it into a 250 ml Erlenmeyer flask containing 100 ml 1× TBE buffer.Heat the agarose powder and electrophoresis buffer in a microwave oven until the agarose is completely melted.Let agarose solution cool down to approximately 55 °C.Add 5 μl of ethidium bromide (10 mg/ml) to a final concentration of approximately 0.5 μg/ml (*see* **Notes** [**13**](#Sec23){ref-type="sec"}--[**15**](#Sec23){ref-type="sec"}).Pour the solution into a sealed gel casting tray containing a gel comb.Let sit for at least 30 min to solidify at room temperature (the solidified gel is opaque in appearance).Remove the seal and comb. Place the gel and the plastic tray horizontally into the electrophoresis chamber with the wells at the cathode side.Cover the gel with 1× TBE buffer.Pipet 0.5 μl of the DNA markers, 2 μl of 6× gel loading dye and 8 μl de water on a Parafilm sheet and mix well (pipetting up and down).Pipet 10 μl of the SN- PCR products and 2 μl of 6× gel loading dye on a Parafilm sheet and mix well (pipetting up and down).Pipet the samples into the wells.Close the lid and connect the power leads on the electrophoresis apparatus.Turn on power supply and apply a voltage of 80--100 V. Run the gel for approximately 40 min.Turn off the power supply when the tracking dye has migrated a sufficient distance.Remove the cover and retrieve the gel (*see* **Note** [**16**](#Sec23){ref-type="sec"}).Place the gel on an ultraviolet transilluminator. Switch on the power of the gel photo-documentation machine (*see* **Note** [**17**](#Sec23){ref-type="sec"}).Adjust the position of the gel and record the results. The size of the expected product for BCoV is 454 bp and 251 bp for the first and second round of amplification, respectively. Notes {#Sec23} ===== Clinical samples must be store at 4 °C for up to 48 h or frozen or −80 °C for longer periods of time. The sensitivity of the PCR could be reduced if you repeat freezing and thawing of fecal samples. Store samples in multiple small aliquots to avoid repeated freeze--thaw cycles which accelerate degradation of nucleic acids, especially RNA.The disodium salt of EDTA will not dissolve until the pH of the solution is adjusted to 8.0 by the addition of NaOH.Dissolution of GuSCN is facilitated by heating in a 60--65 °C water bath under continuous shaking.Buffers L6 and L2 are stable for at least 3 weeks at room temperature in the dark \[[@CR13]\].The three primers used in SN- PCR were designed from the highly conserved region of the N gene of the Mebus strain (GenBank accession number U00735).Phenol is highly corrosive and can cause severe burns to skin. Personal protective equipment such as gloves, safety glasses, and a lab coat are essentials whenever working with phenol.Be careful not to transfer any of the protein at the interphase.Phenol and chloroform are considered hazardous chemicals by waste environment protection agencies and similar bodies around the world. Discard phenol--chloroform--isoamyl alcohol waste in a properly labeled container. For more information regarding hazardous waste management, contact your state hazardous waste agency.Be careful to stir the tube containing hydrated silica before use to resuspend settled silica.Be careful not to disturb the silica pellet when discarding the supernatant. Attention: GuSCN-containing waste should be collected in a strong alkaline solution (10 N NaOH).If you perform the RT-PCR on the same day, RNA extraction can be conserved at 4 °C. For a longer conservation period, RNA must be stored at −80 °C.It is recommended to carry out the cDNA synthesis immediately after the RNA preparation. Being more stable than RNA, the storage of a first-strand cDNA is less critical.Alternatively, the gel can be stained after DNA separation by electrophoresis. Place the gel into a container filled with 100 ml of TBE running buffer and 5 μl of ethidium bromide. Let the gel soak in this solution for 20--30 min. Wash the gel briefly with water before visualization under UV light.Ethidium bromide is mutagenic and moderately toxic and must be handled with care. Wear a lab coat, eye protection, and gloves when working with this chemical.Liquid (running buffer) and solid (agarose gel, gloves, tips) waste contaminated with ethidium bromide must be managed as a hazardous waste. No liquids should be placed in the containers with the solid wastes. For more information regarding hazardous waste management, contact your state hazardous waste agency.DNA will diffuse within the gel over time. Therefore, examination or photography should take place shortly after cessation of electrophoresis.Exposure to UV light can cause severe skin and eye damage. Wear safety glasses and close the photography hood before turning on the UV transilluminator. Be certain to turn off UV light immediately after gel image is captured. Alfieri, A.A. and Alfieri A.F. are recipients of the CNPq fellowships.

Various types of immunocompetent cells, such as macrophages \[[@r2], [@r35], [@r41]\], dendritic cells \[[@r6], [@r35], [@r42]\], various types of T cells \[[@r5], [@r9], [@r15]\], and B cells/plasma cells \[[@r11]\], exist in large numbers in the lamina propria of the intestinal mucosa. Several of these cell types are localized in specific portions in the mucosa of the small intestine. For example, in the mouse small intestine, plasma cells are mainly located in the lower half of the intestinal villus \[[@r29]\], whereas TCRαβ^+^ CD8^+^ T cells are mainly localized in the villous epithelium, probably in a CD103/integrin alpha E-dependent manner \[[@r33]\]. CX3CR1^+^ macrophages reside in whole portions of the mouse colonic lamina propria, while CD169^+^ macrophages preferentially localize near the muscularis mucosa \[[@r2]\]. However, the mechanisms of localization and migration of immunocompetent cells mostly remain unclear in the small intestinal mucosa, although they have been eagerly investigated in mucosal lymphatic tissues of the intestines, such as the Peyer's patch and isolated lymphoid follicles \[[@r16], [@r22], [@r28], [@r32], [@r43]\]. A number of studies have reported that bacterial stimulation affects the localization and differentiation of various lymphocytes. In these studies, bacteria or bacterial components were shown to induce various effects on lymphocyte lineages *in vitro*, such as induction of proliferation and differentiation of B cells \[[@r4], [@r19], [@r30]\] and T cells \[[@r17], [@r19]\]. Moreover, the proliferation, differentiation and localization of several lymphocyte lineages, especially IgA-producing cells \[[@r7], [@r10], [@r18], [@r26]\] and TCRαβ^+^ CD8^+^ intraepithelial lymphocytes \[[@r3], [@r12], [@r13], [@r39]\], have been reported to be affected by the settlement of indigenous bacteria in germ-free animals *in vivo*. These reports suggest that the localization and migration of both TCRαβ^+^ CD8^+^ T cells and IgA^+^ cells are sensitive to bacteria-derived stimulation. However, the effects of indigenous bacteria on TCRαβ^+^ CD8^+^ T cells and IgA^+^ cells under physiological conditions have not been investigated *in vivo*. In the rat small intestine, transient hyperproliferation of indigenous bacteria, which is observed as the expansion of bacterial colonies into the intervillous spaces, often occurs on the intestinal villus \[[@r45]\] and accelerates several host defenses, such as the secretion of bactericidal substances from Paneth cells \[[@r46]\] and acceleration of the epithelial migration \[[@r31]\]. However, whether such bacterial hyperproliferation affects the localization of the TCRαβ^+^ CD8^+^ T and IgA^+^ B cell lineages has never been clarified. Therefore, the primary aim in the present study was to clarify the effect of the expansion of bacterial colonies into the intervillous spaces against the localization of TCRαβ^+^ CD8^+^ T cell and IgA^+^ B cell lineage in the mucosa of the rat small intestine. In addition, we found that many lymphocytes, frequently with immunopositivity for L-selectin, were present in the central lymph vessel of intestinal villi in the mucosa attached to the lymphoid follicles of the Peyer's patches with bacterial colonies in the intervillous spaces in the preliminary examination, suggesting that the expansion of bacterial colonies into the intervillous spaces might induce irregular migration of L-selectin^+^ lymphocytes in the mucosa around the Peyer's patches. Therefore, the secondary aim of the present study was to investigate the specificity of lymphocyte migration in response to the expansion of bacterial colonies into the intervillous spaces on the mucosa around the Peyer's patches with a focus on several lymphocyte lineages, including TCRαβ^+^ CD8^+^ T cells, IgA^+^ B cell lineage and L-selectin^+^ lymphocytes. MATERIALS AND METHODS {#s1} ===================== Animals ------- Ten male Wistar rats aged 7 weeks (Japan SLC Inc., Hamamatsu, Japan) were maintained under specific pathogen-free conditions in individual ventilated cages (Sealsafe Plus; Tecniplast S.p.A, Buguggiate, Italy). They were permitted free access to water and food (Lab MR Stock; Nosan Corp., Yokohama, Japan). The animal facility was maintained under conditions of a 12 hr light/dark cycle (light on: 8:00 AM; light off: 20:00 PM) at 23 ± 2°C and 50 ± 10% humidity. Clinical and pathological examinations in all animals confirmed that there were no signs of disorder. This experiment was approved by the Institutional Animal Care and Use Committee (permission number: 25-06-01) and carried out according to the Kobe University Animal Experimentation Regulations. Tissue preparation ------------------ Collection of tissues from rats was conducted from 9:00 AM to 12:00 PM. After euthanasia by overdose peritoneal injection of pentobarbital sodium (Kyoritsu Seiyaku Corp., Tokyo, Japan), tissue blocks were removed from the ileum including the Peyer's patches. Then, the tissue blocks were immersion-fixed in 4.0% paraformaldehyde fixative in 0.1 M phosphate buffer for 6 hr at 4°C, and snap-frozen in liquid nitrogen as described in a previous study \[[@r47]\]. Four micrometer-thick sections were cut using a Coldtome Leica CM1950 (Leica Biosystems, Nussloch, Germany), placed on slide glasses precoated with 0.2% 3-aminopropyltriethoxysilane (Shin-Etsu Chemical Co., Tokyo, Japan), and stored at −30°C until use. Enzyme immunohistochemistry and double immunofluorescence method ---------------------------------------------------------------- Detection of antigens was conducted using the indirect method of enzyme immunohistochemical and immunofluorescence analysis with the antibodies shown in [Table 1](#tbl_001){ref-type="table"}Table 1.List of antibodyAntibodyManufacturersDiluted concentrationManufacturer's specification formPrimary antibodyanti-CD8 mouse IgGThermoFisher Scientific, Waltham, MA, U.S.A.1:200MA1-70003anti-CD19 mouse IgGThermoFisher Scientific1:100MA5-24372anti-IgA goat IgGSeikagaku Corp., Tokyo, Japan1:20,000270352anti-IgG goat IgGJackson ImmunoResearch Inc., West Grove, PA, U.S.A.1:20,000112-005-143anti-IgM goat IgGAbcam, Cambridge, MA, U.S.A.1:800ab98365anti-TCRαβ mouse IgGThermoFisher Scientific1:200MA5-17541anti-L-selectin goat IgGR&D Systems Inc., Minneapolis, MN, U.S.A.1:400AF1534anti-Ia mouse IgGAbD Serotec, Oxford, U.K.1:1,600MCA46RSecondary antibodyhorseradish peroxidase-conjugated anti-mouse IgG rat IgGJackson ImmunoResearch Inc.1:100415-035-166horseradish peroxidase-conjugated anti-goat IgG donkey IgGJackson ImmunoResearch Inc.1:400705-035-147Rhodamine Red^TM^-X (PRX) conjugated anti-mouse IgG rat IgG (H+L)Jackson ImmunoResearch Inc.1:400415-295-166Alexe Fluor 488 conjugated anti-goat IgG donkey IgGAbcam1:400ab150133. Briefly, after rinsing with 0.05% Tween-added 0.01 M phosphate buffered saline (TPBS; pH 7.4), the sections used for the detection of all antigens except TCRαβ were heated for 20 min at 80°C for antigen retrieval. Then the sections were immersed in absolute methanol for 30 min, followed by 0.5% H~2~O~2~ for 30 min. After each preparation step, the sections were rinsed three times in TPBS to remove any reagent residues. After blocking with Blocking One Histo (Nacalai Tesque Inc., Kyoto, Japan) for 1 hr at room temperature (r.t.), the sections for enzyme immunohistochemical analysis were reacted with primary antibody for 18 hr at 6°C. Primary antibodies for the detection of IgA, IgG and IgM were used at considerably low concentrations in order to detect these immunoglobulins exclusively in B cells and plasma cells, and not in the interstitial fluid. The sections for the immunofluorescence analysis were reacted with one of two sets of paired antibodies---anti-L-selectin antibody/anti-TCRαβ antibody or anti-L-selectin antibody/anti-CD19 antibody---for 18 hr at 6°C. The antibody specificity for each antigen in rats is given in the respective manufacturer's manual. Then, the sections to be applied to the enzyme immunohistochemical analysis were incubated with horseradish peroxidase-conjugated secondary antibody for 1 hr at r.t. Finally, the sections for the enzyme immunohistochemical analysis were incubated with 3,3′-diaminobenzidine (Dojindo Laboratories, Mashiki, Japan) containing 0.03% H~2~O~2~ and counterstained with hematoxylin. The sections for immunofluorescence analysis were reacted with fluorochrome-conjugated secondary antibody and DAPI (diluted at 1:1,000; Dojindo Laboratories) for 1 hr at r.t. Control sections were incubated with TPBS or non-immunized goat IgG (Peprotech, Rocky Hill, NJ, U.S.A.) and mouse IgG (Santa Cruz Biotechnology, Santa Cruz, CA, U.S.A.) instead of the primary antibody. The sections of immunohistochemistry and double immunofluorescence were observed by Nikon Eclipse E600 Fluorescence Microscope and photographed by DS-Fi1c digital camera (Nikon, Tokyo, Japan). Fluorescence filter set B-2A, G-2A and DAPI (Nikon) were used for the observation of double immunofluorescence. Definition of tissue elements ----------------------------- Sections from each tissue block were stained with hematoxylin-eosin for observation of the general structure and the expansion level of bacterial colonies. The following two types of mucosa were investigated: 1) ordinary mucosa, which were defined as mucosa without any mucosal lymphatic tissue such as Peyer's patches, isolated lymphatic follicles or cryptopatches; and 2) follicle-associated mucosa (FAM), which accompanied the parafollicular area under the muscularis mucosae in Peyer's patches in the rat ileum. Then, the intestinal villus was classified into two groups: intestinal villus in the ordinary mucosa (IV) and intestinal villus in FAM (IV-FAM), including the follicle-associated intestinal villus (FAIV), which we defined as the intestinal villus just adjacent to the lymphatic follicle in our previous study \[[@r48]\]. Moreover, the intestinal crypts were classified into two groups: intestinal crypts in the ordinary mucosa (IC) and intestinal crypts in FAM (IC-FAM), including follicle-associated intestinal crypts (FAIC), which we defined as intestinal crypts just adjacent to the lymphatic follicle in our previous study \[[@r48]\] ([Fig. 1a](#fig_001){ref-type="fig"}Fig. 1.a) Diagram of intestinal villus and intestinal crypt in ordinary mucosa (IV and IC) (green area) and intestinal villus and intestinal crypt in follicle-associated mucosa (IV-FAM and IC-FAM) (blue area) of the rat Peyer's patch. The dotted line represents the germinal center. b) Diagram of IV and IV-FAM without the bacterial colonies in the intervillous spaces (nb-IV and nb-IV-FAM) or with the bacterial colonies in the intervillous space (b-IV and b-IV-FAM). Green ovals on the epithelium represent indigenous bacteria. b-IC, IC adjacent to b-IV. b-IC-FAM, IC-FAM adjacent to b-IV-FAM. FAIC, follicle-associated intestinal crypt. FAIV, follicle-associated intestinal villus. nb-IC, IC adjacent to nb-IV. nb-IC-FAM, IC-FAM adjacent to nb-IV-FAM. PA, parafollicular area.). The intestinal villus was divided into three portions from the crypt orifice to the apex of the intestinal villus: the basal, middle and apical portions. Furthermore, IV and IV-FAM were classified into two groups based on the degree of expansion of bacterial colonies on their epithelia: nb-IV and nb-IV-FAM, in which indigenous bacteria existed only on the epithelium of the villous apex or did not exist on the epithelia at all; and b-IV and b-IV-FAM, in which the indigenous bacterial colonies were expanded from the villous apex to the intervillous space or the space between the epithelium of IV-FAM and follicle-associated epithelium. Intestinal crypts adjacent to nb-IV, b-IV, nb-IV-FAM and b-IV-FAM were defined as nb-IC, b-IC, nb-IC-FAM and b-IC-FAM, respectively ([Fig. 1b](#fig_001){ref-type="fig"}). Abbreviations as defined above were listed in [Table 2](#tbl_002){ref-type="table"}Table 2.List of abbreviations in the present studyAbbreviations related to intestinal villusAbbreviations related to intestinal cryptIVintestinal villus in ordinary mucosaICintestinal crypt in ordinary mucosanb-IVIV without bacterial colonies in intervillous spacenb-ICIC adjacent to nb-IVb-IVIV with bacterial colonies in intervillous spaceb-ICIC adjacent to b-IVIV-FAMintestinal villus in follicle-associated mucosaIC-FAMintestinal crypt in follicle-associated mucosanb-IV-FAMIV-FAM without bacterial colonies in intervillous spacenb-IC-FAMIC-FAM adjacent to nb-IV-FAMb-IV-FAMIV-FAM with bacterial colonies in intervillous spaceb-IC-FAMIC-FAM adjacent to b-IV-FAMFAIVfollicle-associated intestinal villusFAICfollicle-associated intestinal crypt. Histoplanimetry --------------- For counting the number of each immunopositive cell in the epithelium and lamina propria and migrating cells in the central lymph vessel, 5 nb-IVs, b-IVs, nb-IV-FAMs and b-IV-FAMs, whose central axes were longitudinally cut, were randomly chosen from 5 rats, respectively. Then, the number of TCRαβ^+^ and CD8^+^ cells in the epithelium was counted in nb-IVs, b-IVs, nb-IV-FAMs and b-IV-FAMs. The number of IgA^+^ cells per 100-*µ*m^2^ in the lamina propria was calculated in the basal, middle and apical portions of each intestinal villus and around nb-ICs, b-ICs, nb-IC-FAMs and b-IC-FAMs. The number of migrating cells per 1,000-*µ*m^2^ in the central lymph vessel, which lumen was confirmed to be located on the central axes of intestinal villus and not contained erythrocytes unlike the blood vessels, of the intestinal villus was also calculated in nb-IV, b-IV, nb-IV-FAM and b-IV-FAM using hematoxylin-eosin-stained sections. The area of the lamina propria or the central lymph vessel were determined using Image-J software (National institutes of health, Bethesda, MD, U.S.A.) to calculate the number of each cell per each unit area. Furthermore, one hundred migrating cells in the central lymph vessel of b-IV-FAM were randomly chosen from 5 rats, respectively. Then, the ratios of L-selectin^+^, TCRαβ^+^, CD8^+^, CD19^+^, IgA^+^, IgM^+^, IgG^+^ and Ia^+^ migrating cell were calculated, respectively. Statistical analysis -------------------- The normality of distribution was first assessed by the Kolmogorov--Smirnov test. For parametric variables in multiple comparisons, two-way ANOVA was performed, followed by the Tukey--Kramer test for *post hoc* comparison. For non-parametric variables in multiple comparisons, the Kruskal--Wallis test was performed, followed by the Steel--Dwass test was performed for *post hoc* comparison. RESULTS {#s2} ======= The effect of bacterial colonies in the intervillous spaces against localization of CD8^+^ T cell or IgA^+^ B cell lineages in the rat ileum -------------------------------------------------------------------------------------------------------------------------------------------- Immunopositivity for TCRαβ was observed in the cellular membrane of the small and oval cells in ordinary mucosa and FAM ([Fig. 2a](#fig_002){ref-type="fig"}Fig. 2.Localization of TCRαβ^+^ cells in the intestinal villus in ordinary mucosa (IV) and that in follicle-associated mucosa (FAM) (IV-FAM) and around the intestinal crypt in ordinary mucosa (IC) and that in FAM (IC-FAM). a) High magnification image of TCRαβ^+^ cells in the lamina propria of IV. Immunopositivity for TCRαβ is observed in the cellular membrane of small and oval cells (arrowhead). b--i) TCRαβ^+^ cells are visible in the lamina propria (black arrowheads) and epithelium (arrows) of IV or IV-FAM without bacterial colonies in the intervillous space (nb-IV (b) or nb-IV-FAM (f)) and IV or IV-FAM with bacterial colonies in the intervillous space (b-IV (d) or b-IV-FAM (h)). TCRαβ^+^ cells in the epithelium (arrows) and lamina propria (black arrowheads) are more abundant in b-IV-FAM (h) than in nb-IV-FAM (f). TCRαβ^+^ cells in the lamina propria (black arrowheads) are scarce around IC adjacent to nb-IV (nb-IC) (c), IC adjacent to nb-IV-FAM (nb-IC-FAM) (g) and IC adjacent to b-IV (b-IC) (e), and moderately present around IC adjacent to b-IV-FAM (b-IC-FAM) (i). Red arrowheads indicate bacterial colonies in the intervillous space. Bar=10 *µ*m. j) The number of TCRαβ^+^ cells in the epithelium of nb-IV, b-IV, nb-IV-FAM and b-IV-FAM per 100-*µ*m-long epithelium. Asterisks, *P*\<0.05. Double asterisks, *P*\<0.01. Each value represents the mean ± SE.). A few TCRαβ^+^ cells were detected in the lamina propria throughout nb-IV and around nb-IC ([Fig. 2b, 2c)](#fig_002){ref-type="fig"}. Moreover, TCRαβ^+^ cells were localized in the epithelium of nb-IV, but rarely detected in that of nb-IC. TCRαβ^+^ cells in the epithelium of nb-IV were moderately detected in the basal portion and decreased toward the apical portion. These distributions of TCRαβ^+^ cells in nb-IV and nb-IC did not differ substantially from those in b-IV and b-IC ([Fig. 2b--e](#fig_002){ref-type="fig"}). The distributions of TCRαβ^+^ cells in nb-IV-FAM and nb-IC-FAM were similar to those in IV and IC, respectively ([Fig. 2f, 2g](#fig_002){ref-type="fig"}). On the other hand, TCRαβ^+^ cells were more abundant in the epithelium throughout b-IV-FAM, especially in apical portion, and in the lamina propria throughout b-IV-FAM and around b-IC-FAM compared with nb-IV-FAM and nb-IC-FAM, respectively ([Fig. 2f--i](#fig_002){ref-type="fig"}). The number of TCRαβ^+^ cells in the epithelium was significantly greater in b-IV-FAM than in nb-IV-FAM, and significantly greater in IV-FAM than in IV, although no significant difference was observed between the number of TCRαβ^+^ cells in the epithelia of nb-IV and b-IV ([Fig. 2j](#fig_002){ref-type="fig"}). Immunopositivity for CD8 was observed in the cellular membrane of the small and oval cells in ordinary mucosa and FAM ([Fig. 3a](#fig_003){ref-type="fig"}Fig. 3.Localization of CD8^+^ cells in the intestinal villus in ordinary mucosa (IV) and that in follicle-associated mucosa (FAM) (IV-FAM) and around the intestinal crypt in ordinary mucosa (IC) and that in FAM (IC-FAM). a) High magnification image of CD8^+^ cells in the lamina propria and epithelium of IV. Immunopositivity for CD8 is observed in cellular membrane of small and oval cells (arrowhead). b--i) CD8^+^ cells are visible in the lamina propria (black arrowheads) and epithelium (arrows) of IV or IV-FAM without bacterial colonies in the intervillous space (nb-IV (b) or nb-IV-FAM (f)) and IV or IV-FAM with bacterial colonies in the intervillous space (b-IV (d) or b-IV-FAM (h)). CD8^+^ cells in the lamina propria (black arrowheads) are scarce in nb-IV (b), b-IV (d) and nb-IV-FAM (f), and moderately present in b-IV-FAM (h). CD8^+^ cells in the epithelium (arrows) are more abundant in b-IV (d) and b-IV-FAM (h) compared with nb-IV (b) and nb-IV-FAM (f), respectively. CD8^+^ cells in the lamina propria (black arrowheads) are scarce around IC adjacent to nb-IV (nb-IC) (c), IC adjacent to b-IV (b-IC) (e) and IC adjacent to nb-IV-FAM (nb-IC-FAM) (g), and moderately present around IC adjacent to b-IV-FAM (b-IC-FAM) (i). Red arrowheads indicate bacterial colonies in the intervillous space. Bar=10 *µ*m. j) The number of CD8^+^ cells in a 100-*µ*m segment of the epithelium of nb-IV, nb-IV-FAM, b-IV and b-IV-FAM. Asterisks, *P*\<0.05. Double asterisks, *P*\<0.01. Each value represents the mean ± SE.). CD8^+^ cells in the lamina propria were scarce in nb-IV, b-IV and nb-IV-FAM and around nb-IC, b-IC and nb-IC-FAM, but moderately detected in b-IV-FAM and around b-IC-FAM ([Fig. 3b--i](#fig_003){ref-type="fig"}). CD8^+^ cells in the epithelium were moderately detected in the basal portion and decreased toward the apical portion in nb-IV and nb-IV-FAM ([Fig. 3b, 3f](#fig_003){ref-type="fig"}). CD8^+^ cells were more abundant in the epithelium throughout b-IV and b-IV-FAM, especially in the apical portion, compared with nb-IV and nb-IV-FAM, respectively ([Fig. 3b--i](#fig_003){ref-type="fig"}), which was confirmed by histoplanimetry ([Fig. 3j](#fig_003){ref-type="fig"}). Immunopositivity for CD19, which was used as a B cell marker, was observed in the cellular membrane of the small and oval cells in FAM ([Fig. 4a](#fig_004){ref-type="fig"}Fig. 4.Localization of CD19^+^ cells in the intestinal villus in ordinary mucosa (IV) and that in follicle-associated mucosa (FAM) (IV-FAM) and around the intestinal crypt in ordinary mucosa (IC) and that in FAM (IC-FAM). a) High magnification image of CD19^+^ cells in the lamina propria of IV-FAM. Immunopositivity for CD19 is observed in the cellular membrane of small and oval cells (arrowhead). b--g) CD19^+^ cells (black arrowheads) are visible in the lamina propria of IV-FAM with bacterial colonies in the intervillous space (b-IV-FAM) (d) and that around IC-FAM adjacent to b-IV-FAM (b-IC-FAM) (g), but not in the lamina propria of IV or IV-FAM without bacterial colonies in the intervillous spaces (nb-IV (b) or nb-IV-FAM (c)) and around IC adjacent to nb-IV (nb-IC) (e) or nb-IV-FAM (nb-IC-FAM) (f). Red arrowheads indicate bacterial colonies in the intervillous space. Bar=10 *µ*m.). In IV and IC, CD19^+^ cells were not detected in the lamina propria regardless of the presence of bacterial colonies in the intervillous spaces ([Fig. 4b, 4e](#fig_004){ref-type="fig"}). CD19^+^ cells were also not detected in the lamina propria in nb-IV-FAM and that around nb-IC-FAM ([Fig. 4c, 4f](#fig_004){ref-type="fig"}), while they were present at moderate levels in the lamina propria from the basal to middle portions of b-IV-FAM and that around b-IC-FAM ([Fig. 4d, 4g](#fig_004){ref-type="fig"}). Immunopositivities for IgA, IgM and IgG were observed in the cytoplasm of the small and oval cells in ordinary mucosa and FAM ([Fig. 5a--c](#fig_005){ref-type="fig"}Fig. 5.Localization of IgA^+^, IgM^+^ and IgG^+^ cells in the intestinal villus in ordinary mucosa (IV) and that in follicle-associated mucosa (FAM) (IV-FAM) and around the intestinal crypt in ordinary mucosa (IC) and that in FAM (IC-FAM). a--c) High magnification image of IgA^+^ (a), IgM^+^ (b) and IgG^+^ cells (c) in the lamina propria of IV. Immunopositivity for IgA, IgM or IgG is observed in the cytoplasm of small and oval cells in IV (arrowhead). d--k) IgA^+^ cells (black arrowheads) are visible in the lamina propria of IV or IV-FAM without bacterial colonies in the intervillous space (nb-IV (d) or nb-IV-FAM (f)) and IV or IV-FAM with bacterial colonies in the intervillous space (b-IV (e) or b-IV-FAM (g)) and the lamina propria around IC adjacent to b-IV (b-IC) (i) and IC adjacent to b-IV-FAM (b-IC-FAM) (k), but scarce around IC adjacent to nb-IV (nb-IC) (h) and IC adjacent to nb-IV-FAM (nb-IC-FAM) (j). IgA^+^ cells are visible in the basal and middle portions of nb-IV (d) and nb-IV-FAM (f), but are visible in the entirety of b-IV (e) and b-IV-FAM (g). Red arrowheads indicate bacterial colonies in the intervillous space. l--o) The number of IgA^+^ cells per 100-*µ*m^2^ in the lamina propria of the apical (l), middle (m) and basal portions (n) of nb-IV, nb-IV-FAM, b-IV and b-IV-FAM and that around nb-IC, nb-IC-FAM, b-IC, and b-IC-FAM (o). Asterisks, *P*\<0.05. Double asterisks, *P*\<0.01. Each value represents the mean ± SE. p--s) IgM^+^ cells (arrowheads) are visible in the lamina propria of nb-IV (p) and nb-IV-FAM (r) and that around nb-IC (q) and nb-IC-FAM (s), with IgM^+^ cells being abundant in nb-IV-FAM and nb-IC-FAM. t) IgG^+^ cells (arrowhead) are scarce in the lamina propria of nb-IV and that around nb-IC. Bar=10 *µ*m.). IgA^+^ cells in the lamina propria were moderately localized in the basal and middle portions and were scarce in the apical portion of nb-IV ([Fig. 5d](#fig_005){ref-type="fig"}) and nb-IV-FAM ([Fig. 5f](#fig_005){ref-type="fig"}) and around nb-IC ([Fig. 5h](#fig_005){ref-type="fig"}) and nb-IC-FAM ([Fig. 5j](#fig_005){ref-type="fig"}). On the other hand, IgA^+^ cells in the lamina propria were more frequent in b-IV and b-IV-FAM ([Fig. 5e, 5g](#fig_005){ref-type="fig"}) and around b-IC and b-IC-FAM ([Fig. 5i, 5k](#fig_005){ref-type="fig"}), compared with nb-IV, nb-IV-FAM, nb-IC and nb-IC-FAM, respectively. The number of IgA^+^ cells in the lamina propria was significantly greater in the apical portions of b-IV and b-IV-FAM ([Fig. 5l](#fig_005){ref-type="fig"}) and in the middle portions of b-IV ([Fig. 5m](#fig_005){ref-type="fig"}), but not b-IV-FAM, compared with those of nb-IV and nb-IV-FAM, respectively. The numbers of IgA^+^ cells in the lamina propria in the basal portion were not different between nb-IV and b-IV, while these numbers were smaller in b-IV-FAM than in nb-IV-FAM ([Fig. 5n](#fig_005){ref-type="fig"}). IgA^+^ cells in the lamina propria were also significantly more abundant around b-IC and b-IC-FAM compared with nb-IC and nb-IC-FAM, respectively ([Fig. 5o](#fig_005){ref-type="fig"}). IgM^+^ cells were moderately localized in the lamina propria around the IC and the lamina propria of the basal portion of IV, and were decreased toward the lamina propria of the apical portion of IV ([Fig. 5p, 5q](#fig_005){ref-type="fig"}). IgM^+^ cells in the lamina propria were more frequently detected in IV-FAM and around IC-FAM than IV and IC, respectively ([Fig. 5r, 5s](#fig_005){ref-type="fig"}). IgG^+^ cells were scarcely detected in the lamina propria of IV, IC, IV-FAM and IC-FAM ([Fig. 5t](#fig_005){ref-type="fig"}). The above findings on IgM and IgG were similar regardless of the presence of bacterial colonies in the intervillous spaces of IV and IV-FAM. The effect of bacterial colonies in the intervillous spaces against localization of L-selectin^+^ lymphocytes in rat ileum -------------------------------------------------------------------------------------------------------------------------- Migrating cells were frequently accumulated in the central lymph vessel of b-IV-FAM, but were rarely accumulated in those of nb-IV, b-IV and nb-IV-FAM ([Fig. 6a, 6b](#fig_006){ref-type="fig"}Fig. 6.Migrating cells in the central lymph vessel of the intestinal villus in ordinary mucosa (IV) and that in follicle-associated mucosa (IV-FAM). a, b) Migrating cells are not visible in the central lymph vessel of IV-FAM without bacterial colonies in the intervillous space (nb-IV-FAM) (a), but are abundantly visible in the central lymph vessel of IV-FAM with bacterial colonies in the intervillous space (b-IV-FAM) (b). Red arrowheads indicate bacterial colonies in the intervillous space. c) The number of migrating cells per 1,000-*µ*m^2^ in the central lymph vessel of IV without bacterial colonies in the intervillous space (nb-IV), IV with bacterial colonies in the intervillous space (b-IV), nb-IV-FAM and b-IV-FAM. Asterisks, *P*\<0.05. Double asterisks, *P*\<0.01. Each value represents the mean ± SE. d--k) Immunopositivity for each cellular marker in the migrating cells in the central lymph vessel of b-IV-FAM. Migrating cells in the central lymph vessel of b-IV-FAM are immunopositive (arrowheads) for L-selectin (d), TCRαβ (e), CD8 (f), CD19 (g), IgA (h), IgM (i), IgG (j) and Ia (k). Dashed lines, central lymph vessel. Bar=10 *µ*m.). The migrating cells were significantly more abundant in the central lymph vessel of b-IV-FAM than in those of nb-IV, b-IV or nb-IV-FAM, while the levels of migrating cells in the central lymph vessel were not significantly different between nb-IV and b-IV ([Fig. 6c](#fig_006){ref-type="fig"}). Migrating cells in the central lymph vessel of b-IV-FAM were immunopositive for L-selectin (94.4 ± 1.60%), TCRαβ (58.6 ± 2.29%), CD8 (26.2 ± 1.11%), CD19 (28.2 ± 0.86%), IgA (2.8 ± 0.86%), IgM (30.2 ± 2.59%), IgG (3.2 ± 1.06%) and Ia (29.6 ± 2.04%) ([Fig. 6d--k](#fig_006){ref-type="fig"}). L-selectin^+^ cells were not detected in the lamina propria of IV or the lamina propria around IC regardless of the presence of bacterial colonies in the intervillous spaces of IV ([Fig. 7a, 7b](#fig_007){ref-type="fig"}Fig. 7.Localization of L-selectin^+^ cells in the intestinal villus in ordinary mucosa (IV) and that in follicle-associated mucosa (FAM) (IV-FAM) and around the intestinal crypt in ordinary mucosa (IC) and that in FAM (IC-FAM). a, b) Immunopositivity for L-selectin is not visible in the lamina propria of IV without bacterial colonies in the intervillous space (nb-IV) (a) and IC adjacent to nb-IV (nb-IC) (b). c--f) L-selectin^+^ cells are scarce in the lamina propria of IV-FAM without bacterial colonies in the intervillous space (nb-IV-FAM) (c) and that around IC adjacent to nb-IV-FAM (nb-IC-FAM) (d), but are abundant in that of IV-FAM with bacterial colonies in the intervillous space (b-IV-FAM) (e) and that around IC adjacent to b-IV-FAM (b-IC-FAM) (f). L-selectin^+^ cells are continuously distributed from the parafollicular area to the lamina propria around b-IC-FAM (f). Red arrowheads indicate bacterial colonies in the intervillous space. g, h) Double immunofluorescence against L-selectin (green)/TCRαβ (red) (g) and L-selectin (green)/CD19 (red) (h) in the lamina propria around b-IC-FAM. Co-localizations of L-selectin/TCRαβ (g) or CD19 (h) (yellow arrows) are visible in the lamina propria around b-IC-FAM. Dotted lines represent the basal membrane of the epithelium. Bar=10 *µ*m.), although in rare cases they were detected in the central lymph vessel of IV regardless of the presence of bacterial colonies in the intervillous spaces of IV. Small numbers of L-selectin^+^ cells were detected in the lamina propria around nb-IC-FAM, but L-selectin^+^ cells were scarce in nb-IV-FAM ([Fig. 7c, 7d](#fig_007){ref-type="fig"}). On the other hand, L-selectin^+^ cells were frequently found in the lamina propria of b-IV-FAM and that around b-IC-FAM. Moreover, L-selectin^+^ cells were often continuously distributed from the parafollicular area of Peyer's patch to the lamina propria around b-IC-FAM, and occasionally expanded into the middle portion of b-IV-FAM ([Fig. 7e, 7f](#fig_007){ref-type="fig"}). Double immunofluorescence for L-selectin and TCRαβ or CD19 showed that L-selectin^+^ cells from the parafollicular area to IV-FAM were frequently immunopositive for TCRαβ^+^ ([Fig. 7g](#fig_007){ref-type="fig"}) and occasionally immunopositive for CD19 ([Fig. 7h](#fig_007){ref-type="fig"}). Co-localizations of L-selectin/TCRαβ or CD19 were also found in the migrating cells in the central lymph vessel of b-IV-FAM. The changes of histological localization of several lymphocyte lineages in response to the expansion of bacterial colonies into the intervillous spaces are summarized in [Fig. 8.](#fig_008){ref-type="fig"}Fig. 8.Schematic diagram of localization of the lymphocyte lineage in the ordinary mucosa and follicle-associated mucosa (FAM). Gray ovals on the epithelium represent indigenous bacteria. DISCUSSION {#s3} ========== TCRαβ^+^ CD8^+^ intraepithelial lymphocytes reside in the intestinal epithelium of the mouse small intestine, likely in CD103/integrin alpha E-dependent manner \[[@r33]\]. The number and activity of TCRαβ^+^ intraepithelial lymphocytes have been reported to increase in mice with monoassociation of segmented filamentous bacteria compared with germ-free mice \[[@r37], [@r38]\]. In the present study, the numbers of TCRαβ^+^ or CD8^+^ cells were increased in the epithelia of the intestinal villi of the ordinary mucosa and FAM in response to the expansion of bacterial colonies into the intervillous spaces in the rat ileum. These findings suggest that the number of TCRαβ^+^ CD8^+^ T cells in the villous epithelium could be affected even by transient proliferation of indigenous bacteria. Intraepithelial lymphocytes have been suggested to contribute to the formation of an intestinal epithelial barrier against bacterial invasion in various ways. For example, intraepithelial lymphocytes are suggested to promote the expression of claudin-1 by epithelial cells and increase mucus thickness in response to bacterial stimulation \[[@r21]\]. In addition, intraepithelial lymphocytes are suggested to produce bactericidal lectin RegIIIβ in response to bacterial stimulation \[[@r14]\]. Therefore, an increase of intraepithelial lymphocytes in response to the expansion of bacterial colonies into the intervillous spaces might contribute to the enhancement of local intestinal epithelial intensity and secretory host defense mediated by the mucus and innate antimicrobial factors against proliferating bacteria under physiological conditions. Monoassociation of segmented filamentous bacteria in germ-free mice has been shown to increase the number of IgA-producing cells in the lamina propria \[[@r38]\]. Bacterial stimulation has been suggested to increase IgA^+^ cells in the lamina propria and fecal IgA in the mouse small intestine \[[@r34]\]. In the present study, IgA^+^ cells possessed IgA-immunopositivity in their cytoplasms and changed their localization in response to the expansion of bacterial colonies into the intervillous spaces of IV and IV-FAM, while the localizations of IgG^+^, IgM^+^ and CD19^+^ cells were not changed and thus did not correspond with the change of localization of IgA^+^ cells in response to the expansion of bacterial colonies. CD19 is expressed in murine and human pro-B cells, pre-B cells and B cells, but not plasma cells \[[@r20]\]. These findings suggest that the IgA^+^ cells that changed their localization in response to the expansion of bacterial colonies into the intervillous spaces in the present study were plasma cells. IgA are secreted from the intestinal crypt \[[@r27]\], coat the indigenous bacteria in the feces \[[@r40]\] and inhibit the bacterial adherence on the mucosal epithelium \[[@r44]\]. Based on these previous findings, our present result suggests that IgA^+^ cells, probably plasma cells, were increased around the intestinal crypt in response to the expansion of bacterial colonies into the intervillous spaces and contribute to the local host defenses against the indigenous bacteria on the epithelium of intestinal villi. In addition, IgA^+^ cells increased not only around the intestinal crypt, but also in the apical portions of IV and IV-FAM in response to the expansion of bacterial colonies into the intervillous spaces. Lipopolysaccharide, one of the Gram-negative bacterial constituents, is absorbed from the intestine into the blood stream in humans even under healthy conditions \[[@r1]\], and this process is accelerated during the postprandial period by a high-fat diet \[[@r8]\]. Intestinal IgA have been shown to bind lipopolysaccharide from indigenous bacteria \[[@r23]\]. Considering that Gram-negative bacteria are more abundant in deep portions of intervillous spaces of the rat small intestine than Gram-positive bacteria \[[@r24]\], an increase of IgA^+^ cells---probably plasma cells---in the apical portion of the intestinal villus in response to the expansion of bacterial colonies into the intervillous spaces might contribute to the protection against or neutralization of absorbed bacterial constituents such as lipopolysaccharide from Gram-negative bacteria. Most of the studies on Peyer's patches have concentrated on the characteristics of the lymphatic follicle. However, we previously reported that the FAIV, which is the intestinal villus just adjacent to the lymphatic follicle and is included in the IV-FAM, have special mechanism for recognition of bacteria \[[@r48]\], and that the follicle-associated intestinal crypts, which are included in the IC-FAM, have an unusual composition of epithelial cells \[[@r25]\] in the rat Peyer's patches. In the present study, L-selectin^+^ cells, which also express TCRαβ or CD19, were numerously detected in the central lymph vessel of b-IV-FAM, although migrating cells themselves were rarely detected in the central lymph vessel even in nb-IV, b-IV and nb-IV-FAM. In addition, L-selectin^+^ cells were continuously localized from the parafollicular area to the lamina propria of the middle portion of b-IV-FAM and were immunopositive for TCRαβ or CD19. These findings suggest that L-selectin^+^ B cells and T cells migrate from the parafollicular area of Peyer's patches to the lamina propria of IV-FAM and enter the central lymph vessel of IV-FAM in response to the expansion of bacterial colonies into the intervillous spaces of the IV-FAM, although the host defenses against indigenous bacteria by CD8^+^ T cell and IgA^+^ B cell lineages were probably common in both the ordinary mucosa and FAM. It has been clarified that lymph vessel of mesentery from area with Peyer's patches include more abundant migrating cells than that from area without Peyer's patches in rat small intestine \[[@r36]\]. From these findings, the expansion of bacterial colonies into the intervillous spaces might induce the active replacement of T and B cells in IV-FAM and abundant supplementations of the lymphocyte toward the mesenteric lymph nodes from IV-FAM in the rat ileum, and these actions could contribute to the effective induction of an immune response. This peculiar migration of lymphocytes in IV-FAM suggest the existence of specific signals for chemotaxis, which could regulate the migration of lymphocytes, in the FAM. In the future, the expression of chemokines and its receptor and the peculiar intracellular communication among various immunocompetent cells in IV-FAM is needed to be investigated. This work was financially supported in part by a Grant-in-Aid for Scientific Research (nos. 15K07766 and 16K18813) from the Japan Society for the Promotion of Science.

\section{Introduction} In the local Universe, three types of bulges have been defined \citep[see][for recent reviews]{Kormendy04,Athanassoula05}. {\it Classical bulges} are dynamically hot systems with red colors, high concentration, and formation processes perhaps similar to elliptical galaxies. {\it Pseudobulges} \citep{Kormendy93} have lower central densities and bluer colors than classical bulges, often with ongoing star formation. They are believed to form through secular processes that drive gas and stars from the disk inward. {\it Boxy-peanut shaped bulges}, sometimes considered as a subclass of pseudobulges, are formed via the natural evolution of barred galaxies. The ideas on the formation histories of bulges can be confronted with direct observations of high-redshift bulge precursors. Of critical importance is the search for of strong star formation episodes in the nuclei of galaxies which would qualify as late bulge formation. Recent works provide differing accounts of the abundance of blue bulges. \citet{Menanteau01} argue for a $30\%-50\%$ of young field spheroids in the two {\it Hubble Deep Fields} (HDF). In contrast, \citet{Koo05} found that $85\%$ of luminous field bulges at $z\sim0.8$ in the Groth strip are nearly as red as local bulges, arguing for very little growth of massive bulges at $z<1$. \citet{MacArthur07}, working with GOODS-N field disk galaxies, noted that bulge color is related to mass: the most massive bulges are as red as massive spheroids, perhaps from mass assembly at high redshift; smaller bulges show more diverse star formation histories, with significant star formation at $z<1$. In our own analysis of nuclear colors of disk galaxies at $0<z<1$ in the Groth strip \citep[DB08]{Dominguez08II}, we found that $60\%$ of $0<z<1$ bulges define a passive evolution red sequence (RS); the remaining $40\%$ have bluer colors that may trace ongoing bulge formation/growth. For such late-forming bulges two questions arise. First, what type of $z=0$ bulge (classical, pseudobulge) may be the descendant of an observed star-forming bulge at intermediate $z$. Second, who are their precursors. These may be high-$z$ RS classical bulges rejuvenated due to e.g., gas infall and ensuing star formation. Rejuvenation models have often been advocated \citep[e.g.][]{Ellis01,Koo05}. Alternatively, blue bulges may be the result of centrally concentrated star formation in the disk, perhaps through redistribution of gas through secular or environmental processes. Distinguishing rejuvenated bulges from bona-fide young bulges growing from their host disks can be done by measuring the relationship between central surface brightness and color. For the general galaxy population in the local Universe, color is strongly correlated with surface density, in the sense that redder galaxies are denser \citep{Kauffmann03dens,Bell00II}, suggesting that galaxies with high surface densities formed the bulk of their stars at earlier epochs than galaxies with lower surface densities. For galaxies with bulges, \citet{Drory07} conclude that galaxies with classical bulges have red global colors and high central surface brightness, whereas galaxies harbouring pseudobulges show bluer global colors and moderate central surface brightness. In this letter we present the redshift evolution of the relationship between bulge colors and central surface brightness, $\mu_0$. We work with a diameter-limited sample in the redshift range $0.1<z<1.3$. We quantify the $z$-evolution of the $\mu_0$ distribution for bulge and non-bulge galaxies, and investigate the ability of color-$\mu_0$ diagram to segregate different types of bulges at $z\sim1$. Our main result is that, beyond $z\sim0.5-0.8$, blue nuclei appear at increasingly high $\mu_0$, tracing episodes of nuclear star formation that lead to bulge growth, most probably by secular processes. Their densities make them candidate precursors of today's pseudobulges, we do not find evidences of rejuvenation of clasical bulges. Interestingly, the AGN fraction increases as well among bulge galaxies at $z>0.8$. A cosmology with $\Omega_M=0.3$, $\Omega_\Lambda=0.7$, $H_0=70$ km\,s$^{-1}$\,Mpc$^{-1}$\ is assumed througout. Magnitudes are expressed in the Vega system. \begin{table} \begin{center} \caption{Bulge sample characteristics} \label{tab:medians} \begin{tabular}{l c c c c} \hline \hline & quantity & $<\mu_0>$ & $\sigma_{\mu_0}$ & $\%$ RS \\ & (1) & (2) & (3) & (4) \\ \hline $0.0<z<0.5$ & 50 & 18.79 & 1.12 & 43.6 \\ $0.5<z<0.8$ & 41 & 18.08 & 1.07 & 45.9 \\ $0.8<z<1.3$ & 22 & 17.50 & 0.87 & 31.8 \\ \hline \hline \end{tabular} \end{center} \begin{footnotetext} TNote.- (1) Number of galaxies with bulge and without AGN at each redshift range; (2) and (3) median observed $F435W$-band $\mu_0$ and standard deviation; (4) percentage of bulges that belong to the RS, using the cut $(U-B)>0.26$ for bulge galaxies with $\mu_0>19.5$. \end{footnotetext} \end{table} \begin{table} \begin{center} \caption{Non-bulge sample characteristics} \label{tab:mediansNB} \begin{tabular}{l c c c} \hline \hline & quantity & $<\mu_0>$ & $\sigma_{\mu_0}$ \\ & (1) & (2) & (3) \\ \hline $0.0<z<0.5$ & 92 & 20.57 & 0.99 \\ $0.5<z<0.8$ & 61 & 20.09 & 0.75 \\ $0.8<z<1.3$ & 51 & 19.39 & 0.82 \\ \hline \hline \end{tabular} \end{center} \begin{footnotetext} TNote.- (1) Number of galaxies without bulge and without AGN at each redshift range; (2) and (3) median observed $F435W$-band $\mu_0$ and standard deviation. \end{footnotetext} \end{table} \section{Data} \label{sec:data} Sample selection followed identical procedures as in DB08, and we refer the reader to \citet{Dominguez08I} for a detailed description, and to \citet{Dominguez08III} for details specific of the present sample. In brief, samples were selected from the ACS/HST GOODS-N \citep{Giavalisco04}, using catalog and image versions v1.1 and v1.0, respectively. The images cover 160 arcmin$^2$ at 0.03 arcsec/pixel through multi-drizzle. We started from a diameter-limited parent sample with \textsc{SExtractor}\ radius above 1.4\arcsec\ (398 galaxies). All have spectroscopic redshifts \citep{Cowie04,Wirth04}, in the range $0.1<z<1.3$. We selected a subsample of galaxies with prominent bulges using the standard {\it mark the disk} method of assuming an exponential profile for the disk. We define a bulge prominence index $\eta$ that measures the excess central surface brightness over the inward extrapolation of the exponential profile of the disk; we take as galaxies with measurable bulges those with $\eta>1$ mag\,arcsec$^{-2}$. The subsample of galaxies with measurable bulges comprises 131 galaxies, after excluding 53 strongly-distorted merger candidates. The remainder 214 galaxies, classified as disk galaxies without bulge, constitute a comparison sample. Galaxies in the latter sample may contain small bulges, unresolved by our images. Indeed, our bulge sample comprises physically large bulges, though not restricted to classical bulges; e.g., simulations show that the prototype peanut-shaped bulge of NGC~5965 would be clearly distinguished out to $z\sim1$ in our survey. See DB08, \S3, for a detailed analysis of selection effects in our bulge sample. To investigate selection biases in our diameter-limited sample, we have additionally selected samples by physical radius (above 10.8 kpc); we will show that their behavior is very similar to the angular-limited ones. The central surface brightness, $\mu_0$ in rest-frame $F435W$, was measured on a central aperture with a physical diameter of 1~kpc, which corresponds to the size of the image PSF for the most distant galaxy in the sample. We tested deconvolution, and bulge-disk decomposition, but both provided dubious results. We calculated $\mu_0$ using the nearest observed band to the rest-frame $F435W$ one at each redshift; it was corrected of the term $(1+z)^4$ of the cosmological dimming, and of inclination and dust, the latter following \citet{Tuffs04} and \citet{Moellenhoff06}. Nuclear colors were extracted from the same 1~kpc central aperture where we measured $\mu_0$. Integrated galaxy colors were obtained using \textsc{MAG\_BEST} \textsc{SExtractor}\ magnitudes. For galaxies harbouring a broad-line or a narrow-line AGN, nuclear colors were measured just in the origin of the color profile, and global colors excluded an inner region of radius 2~kpc, to avoid AGN contamination. Rest-frame $(U-B)$ colors were measured using the nearest observed colors at each redshift. The corrections for filter differences were calculated using SEDs obtained from the best-fit solution, delivered by \textsc{HyperZ}\ \citep{hyperz}, to photometric points in the bands $F435W,F606W,F775W,F850LP$. \section{Results and discussion} \label{sec:results} In Fig.~\ref{fig:magsupcenaper}, first and second rows, we plot galaxy nuclear and global rest-frame $(U-B)$ colors vs $\mu_0$, for three redshift ranges: $0.0<z<0.5$, $0.5<z<0.8$ and $0.8<z<1.3$. The horizontal solid and dotted lines trace the mean color of the RS and its blue boundary, respectively, at each redshift bin, from \citet{Kriek08}. The dashed lines trace linear least-squares fits to the color-$\mu_0$ distribution of bulge galaxies in the $z < 0.5$ bin ($U-B=-0.27\mu_0+5.27$ for nuclear colors; $U-B=-0.24\mu_0+4.71$ for global colors); these lines are also plotted in the corresponding higher $z$ bins for reference. The dot-dashed lines trace fits to the bulge galaxy colors in the $0.5<z<0.8$ range ($U-B=-0.30\mu_0+5.81$ for nuclear colors; $U-B=-0.25\mu_0+4.71$ for global colors). In the range $0.0<z<0.5$, a strong correlation exists between the $(U-B)$ color and $\mu_0$; both nuclear and global colors are redder in galaxies with higher $\mu_0$. Galaxies without bulges are confined to the blue, faint region of the color-$\mu_0$ diagram, whereas those with bulges cover the entire width of the distribution, with a concentration at the red, high-$\mu_0$ end. The trend is very strong (Spearman Rank correlation coefficient $\ensuremath{{R_\mathrm{S}}}=-0.68$, $\ensuremath{{P_\mathrm{null}}}=2\times10^{-4}\%$ for nuclear colors of bulge galaxies; $\ensuremath{{R_\mathrm{S}}}=-0.74$, $\ensuremath{{P_\mathrm{null}}}=2\times10^{-5}\%$ for global colors) with one single deviant point, a galaxy harbouring a broad-line AGN. Trends are similarly strong ($\ensuremath{{R_\mathrm{S}}}=-0.69$, $\ensuremath{{P_\mathrm{null}}}=0.19\%$ for nuclear colors of bulge galaxies) when the sample is limited by physical diameter (\S~\ref{sec:data}). This shows that the color-$\mu_0$ correlation is not driven by large galaxies at the bright end, and small ones at the faint end. Because red stellar populations, either from old age or from dust, are dimmer than younger populations, the color-$\mu_0$ trend implies that redder bulges have higher central stellar surface mass densities. Hence, at $z<0.5$, central surface density correlates with color, and, hence, with the star formation history of the bulge: denser bulges stopped forming stars in earlier epochs than lower density bulges. Not only nuclear colors, but also global colors correlate with central surface density (Fig.~\ref{fig:magsupcenaper}d), as found in the local Universe \citep {Kauffmann03dens}. Interestingly, $\mu_0$ for galaxies \textsl{without} bulges also correlate with nuclear colors ($\ensuremath{{R_\mathrm{S}}}=-0.26$, $\ensuremath{{P_\mathrm{null}}}=1.6\%$) and with global colors ($\ensuremath{{R_\mathrm{S}}}=-0.37$, $\ensuremath{{P_\mathrm{null}}}=0.06\%$), i.e., the scaling of central surface brightness and color is not entirely due to the presence of the bulge. If the two extremes of the color-$\mu_0$ distribution trace classical and pseudobulges \citep{Drory07}, then Fig.~\ref{fig:magsupcenaper}a indicates that the two types of bulges are found up to $z\sim0.5$. Bluer bulges occupy the color-$\mu_0$ locus of star-forming galaxies, and must be forming stars. Notably, all of such galaxies, presenting lower $\mu_0$ than their redder counterparts, must have lower surface densities too. It is unlikely that such blue bulges are rejuvenated classical bulges. If they were, the high density of the underlying classical bulge, plus the high brightness of the burst, would lead to higher surface brightness than red, classical bulges, contrary to what is observed. We quantify these statements below. At the intermediate redshift bin $0.5<z<0.8$ (Fig.~\ref{fig:magsupcenaper}b,e), despite a brighter $\mu_0$ cutoff due to cosmological dimming, the color-$\mu_0$ relation remains in place ($\ensuremath{{R_\mathrm{S}}}=-0.42[-0.39]$, $\ensuremath{{P_\mathrm{null}}}=0.85\%[2.6\%]$ for nuclear colors of bulge galaxies [selected by physical radius], and $\ensuremath{{R_\mathrm{S}}}=-0.73$, $\ensuremath{{P_\mathrm{null}}}=3.6\times10^{-4}\%$ for global colors). Linear fits to the nuclear and to the global color-$\mu_0$ distributions yield very similar results to those at $z<0.5$. However, blue deviant nuclei are found around $\mu_0\sim18.5$ and $(U-B)\sim-0.2$ (Fig.~\ref{fig:magsupcenaper}b). The color profiles of those objects (not shown) shows mild blueing toward the center, which provides evidence for enhanced star formation in the nuclear region of the galaxy at redshifts $0.5<z<0.8$. Two of the four blue deviant bulges may be minor merger remnants, and one hosts a bar, which suggests that these bulges might be growing by secular processes caused by disk instabilities or by satellite accretion \citep{eliche06}. Like in the lower redshift bin, we do not find evidence for rejuvenation of red, classical bulges at $0.5<z<0.8$: the three blue, high-surface brightness nuclei in Figure~\ref{fig:magsupcenaper}b are AGN hosts. Finally, at $0.8<z<1.3$, while global colors still correlate with $\mu_0$ (Fig.~\ref{fig:magsupcenaper}f: $\ensuremath{{R_\mathrm{S}}}=-0.48$, $\ensuremath{{P_\mathrm{null}}}=2.7\%$), nuclear colors (Fig.~\ref{fig:magsupcenaper}c) show a null trend ($\ensuremath{{R_\mathrm{S}}}=0.10$, $\ensuremath{{P_\mathrm{null}}}=64.3\%$). Figure~\ref{fig:magsupcenaper}c shows a population of blue, high surface brightness nuclei with $\mu_0=16-18$, i.e., as high as the red, classical bulges. In the previous paragraph we already described blue, deviant nuclei at $0.5<z<0.8$, but of a lower $\mu_0=18.0-19.5$. The blue colors and high surface brightness must trace strong episodes of star formation concentrated in the galaxy central regions. The phenomenon may be related to the HDF-N field spheroidals with blue cores \citep{Menanteau01} or to the luminous compact blue galaxies in the HDF-FF \citep{Guzman97}. One of the blue nuclei harbours a narrow-line AGN; the others are undoubtedly stellar. Examples (non-AGN cases) are shown in Figure~\ref{fig:morphology}. One shows signs of interaction; the others are normal spiral galaxies. The color profiles reveal strongly inverted color profiles, bluer inward, in the inner $1-2$ kpc. Selection effects also contribute against a color-$\mu_0$ correlation in $0.8<z<1.3$. Missing small galaxies, due to our apparent diameter selection, is probably unimportant: the trends and correlation coefficients in Figures~\ref{fig:magsupcenaper}a-c are similar when the samples are selected by physical diameter $D>21.6$ kpc. But galaxies with small \textsl{bulges} may have been classified in the bulge-less class due to resolution and dimming effects (DB08, \S3). Hence, it is plausible that the color-$\mu_0$ correlation survives at $0.8<z<1.3$ if small bulges do exist at those redshifts. In any case, the $0.8<z<1.3$ color-$\mu_0$ distribution remains peculiar for the abundance of blue nuclei as bright as the classical bulges: $27\%$ of the $16<\mu_0<18$ nuclei have colors below $(U-B)=0$. Figure~\ref{fig:magsupcenaper} clarifies the redshift evolution of the color-$\mu_0$ distribution. The linear fits to the bulges, plotted in the diagrams, show that there is no significant change between the ranges $z<0.5$ and $0.5<z<0.8$. But for $0.8<z<1.3$, the distribution lies well below the fit to the first range, with brighter $\mu_0$ and bluer colors. The variation in surface brightness is shown in the $\mu_0$ histograms (Fig.~\ref{fig:magsupcenaper}ghi): the peaks of the distributions of both bulge and non-bulge galaxy samples shift toward brigther $\mu_0$, by $\sim1.2$ mag, from $z\sim0$ to $z\sim1.3$ (see Tables~\ref{tab:medians} and~\ref{tab:mediansNB}; median values do not change significantly if we impose a common selection limit at all $z$, $\mu_0=19.5[21]$ for the bulge [non-bulge] sample): Both galaxies with and without bulges were brighter and bluer at $z\sim1$ and faded and reddened, toward $z\sim0$. Most likely, the centrally-concentrated star forming events contribute to bulge formation inside disks. Who are the precursors, and who are the descendants of such blue, high-surface brightness nuclei? In particular, are they rejuvenated classical bulges that will evolve back into the RS once the burst has faded? We estimate the evolution of starbursting nuclei in color-$\mu_0$ space using composite stellar population (CSP) models from \citet[BC03]{Bruzual03}, with exponentially-decaying star formation rates (SFR), without dust, and build tracks intended to match the span of color-$\mu_0$ at $z\sim0$. For the reddest track, we look for models with $(U-B)\sim0.5$, and assign them a surface brightness of $\mu_0=17.5$. Color and surface brightness are tracked back in time using BC03 and the variation with $z$ of the angular size of the galaxy. For a canonical $\ifmmode{z_{\rm f}}\else{$z_{\rm f}$}\fi=2.0$, and a moderate $\tau=0.8$ Gyr, the model evolves along the orange lines in Fig.~\ref{fig:magsupcenaper}abc, becoming both bluer and brighter as we move back in time. At $0.8<z<1.3$, the evolution fails to reach the region of the blue, high-surface brightness colors of the starbursting nuclei seen at those redshifts. The bluer colors at $z\sim0.8$ require a younger age, but a younger age would yield too high a surface brightness. Indeed, evolution is not vertical but diagonal in the color-$\mu_0$ plane, as population ageing implies both reddening and dimming. Taking into account that starbursting nuclei are likely composite old plus young populations does not help, as the underlying, more stable old population would contribute to the surface brightness at all ages, making the problem worse. Hence, the track shows that the blue, high-surface brightness nuclei seen at $0.8<z<1.3$ are not descendants, nor precursors, of the red, high-surface brightness classical bulges. For the bluest track, we impose a $z=0$ color of $(U-B)=0.1$, and assign a $z=0$ surface brightness $\mu_0=20$. This track (green line on Fig.~\ref{fig:magsupcenaper}abc) was modeled using a $\ifmmode{z_{\rm f}}\else{$z_{\rm f}$}\fi=1.0$ and a $\tau=2$ Gyr. Blue deviant nuclei with intermediate surface brightness at $0.5<z<0.8$ fit the track evolution. An intermediate model with a slightly brighter $\mu_0$ at $z\sim0$, $\mu_0=19$, and a $\ifmmode{z_{\rm f}}\else{$z_{\rm f}$}\fi=1.5$ (violet line on Fig.~\ref{fig:magsupcenaper}abc), nicely fits the blue, high-surface brightness nuclei at $0.8<z<1.3$. This model links the blue, high-$\mu_0$ deviant bulges at $0.8<z<1.3$ with moderate color, intermediate-$\mu_0$ bulges at $z=0$; it supports the idea that these bulges are precursors of local Universe pseudobulges, rather than old, dense bulges. Classifying as AGN those galaxies with $L_X>10^{42}$~erg\,s$^{-1}$ in the 2--8 keV band of \textit{Chandra} \citep{Alexander03,Treister06}, $17\%(5\%)$ of the galaxies with (without) bulges host an AGN in the $0.5<z<1.3$ range. This highlights that the preference of AGN to live in galaxies with bulges extends up to $z\sim1.3$. In Fig.~\ref{fig:magsupcenaper} we denote with specific symbols three AGN types, namely broad-line (BL), narrow-line (NL), and low-luminosity (LL; no AGN spectral lines), from \citet{Treister06}. Except for the three BL AGN, the nuclear colors and $\mu_0$ of galaxies with AGN are in the general region occupied by the non-AGN galaxies -- both in the main color-$\mu_0$ sequence, and in the blue, high-$\mu_0$ region of the starbursting nuclei. Their nuclear colors are likely to be affected by the AGN, especially for the more X-luminous (nuclear colors for our AGN galaxies get bluer with increasing X-ray luminosity). However, the size of the blue nuclear region, and the comparison with synthetic color profiles built with combinations of a disk, a bulge, and a nuclear source with varying colors and luminosity ratios, do suggest that most blue regions are spatially resolved, hence that star formation coexists with the AGN. The trigger of the AGN activity in these galaxies is unclear. Morphologies for the three BL AGN hosts reveal a normal, bulge-dominated spiral structure, with a central bar in two of them. The NL AGN galaxies show a more mixed variety of types, including early-type galaxies, normal and barred spirals, and distorted galaxies. Hence, AGN activity is not associated with merging or interacting situations, as noted by \citet{Grogin05}. \begin{figure*} \begin{center} \includegraphics[angle=0,width=0.95\textwidth]{f1.eps}% \end{center} \caption{{\bf a,b,c)} and {\bf d,e,f)} galaxy nuclear and global rest-frame $(U-B)$, respectively, vs $F435W$-band $\mu_0$. {\it Intermediate red filled circles}: normal bulge galaxies. {\it Small green open circles}: bulge galaxies hosting a LL AGN. {\it Small dark blue filled circles}: bulge galaxies harbouring a NL AGN. {\it Large blue filled circles}: bulge galaxies hosting a BL AGN. {\it Blue crosses}: normal non-bulge galaxies. {\it Green asterisks}: non-bulge galaxies harbouring a LL AGN. {\it Blue asterisks}: non-bulge galaxies hosting a NL AGN. Colors and $\mu_0$ errors are derived from observed errors. {\it Horizontal solid and dotted lines}: mean color and blue boundary of the RS. {\it Dashed lines}: linear least-squares fits to the color-$\mu_0$ distribution of bulge galaxies in the $z<0.5$ bin. {\it Dot-dashed lines}: least-squares fits at $0.5<z<0.8$ bin. {\it Orange triangle tracks}: CSP model matching RS bulges at $z\sim0$ ($\mu_0=17.5$, $U-B\sim0.5$) with a SFR timescale $\tau=0.8$ Gyr and $\ifmmode{z_{\rm f}}\else{$z_{\rm f}$}\fi=2$ Gyr; each triangle represents the color-$\mu_0$ at redshifts, from right to left, $z=1.2,0.8,0.5,0.1$. {\it Violet square tracks}: CSP model matching $z=0$ bulges with $\mu_0=19$ and $(U-B)\sim0.2$; the model has a SFR $\tau=2$ Gyr and $\ifmmode{z_{\rm f}}\else{$z_{\rm f}$}\fi=1.5$ Gyr. Each square represents the color-$\mu_0$ at redshifts $z=1.2,0.8,0.5,0.1$. {\it Green diamond tracks}: CSP model matching $z=0$ bulges with $\mu_0=20$ and $(U-B)\sim0.1$; the model has a SFR $\tau=2$ Gyr and $\ifmmode{z_{\rm f}}\else{$z_{\rm f}$}\fi=1$ Gyr. Each diamond represents the color-$\mu_0$ at redshifts $z=0.8,0.5,0.3,0.1$. {\it Red vectors}: reddening for bulge galaxies with $B/T=0.4$, central $B$-band opacity $\tau=4$, and inclinations of 40\deg, for nuclear colors ({\bf a,b,c}) and global colors ({\bf d,e,f}). {\it Blue vectors}: reddening for pure disk galaxies with central $B$-band opacity $\tau=4$, and inclinations of of 40\deg, for nuclear and global colors. {\bf g,h,i)} $F435W$-band $\mu_0$ histograms for the three redshifts ranges. {\it Solid red lines}: normal bulge galaxies. {\it Dashed black lines}: normal non-bulge galaxies.} \label{fig:magsupcenaper} \end{figure*} \begin{figure*} \begin{center} \begin{minipage}[c]{0.2\textwidth}% \centering \includegraphics[angle=0,height=\textwidth]{f2.eps}% \end{minipage}% \begin{minipage}[c]{0.3\textwidth}% \includegraphics[angle=0,width=\textwidth]{f3.eps}% \end{minipage}% \begin{minipage}[c]{0.3\textwidth}% \includegraphics[angle=0,width=\textwidth]{f4.eps \end{minipage}% \vfill \begin{minipage}[c]{0.2\textwidth}% \includegraphics[angle=0,width=\textwidth]{f5.eps}% \end{minipage}% \begin{minipage}[c]{0.3\textwidth}% \includegraphics[angle=0,width=\textwidth]{f6.eps}% \end{minipage}% \begin{minipage}[c]{0.3\textwidth}% \includegraphics[angle=0,width=\textwidth]{f7.eps \end{minipage}% \vfill \begin{minipage}[c]{0.2\textwidth}% \includegraphics[angle=0,width=\textwidth]{f8.eps}% \end{minipage}% \begin{minipage}[c]{0.3\textwidth}% \includegraphics[angle=0,width=\textwidth]{f9.eps}% \end{minipage}% \begin{minipage}[c]{0.3\textwidth}% \includegraphics[angle=0,width=\textwidth]{f10.eps \end{minipage}% \end{center} \caption{$F850LP$ image, observed $F850LP$ surface brightness profile averaged over both semi-major axes, and observed $(F435W-F850LP)$ color profile on both semi-minor axes for the three non-AGN galaxies at $0.8<z<1.3$ with high $\mu_0$ and nuclear positive color gradient.} \label{fig:morphology} \end{figure*} \section{Conclusions} \label{sec:conclusions} In the GOODS-N field, $27\%$ of very blue, high-$\mu_0$ bulges in $0.8<z<1.3$ trace an epoch of important bulge formation inside disks. The fraction of blue cores is close to that for HDF field spheroidals \citep{Menanteau01}. Similar but weaker nuclear star formation is present at $0.5<z<0.8$, pointing at bulge-forming events that become weaker with cosmic time, and lead to a mass-age relation for bulges \citep{MacArthur07}. The data imply important bulge growth at $z<1$. Models predict that such starbursting nuclei evolve into moderate surface-brightness, intermediate color $z=0$ pseudobulges rather than classical bulges, and argue against rejuvenation processes for $z\sim1$ dense and old bulges. The latter are already abundant at $0.8<z<1.3$, as found by \citep{Koo05}. At lower redshifts, $0.0<z<0.8$, the correlation of nuclear and global colors with $\mu_0$, and with central density, suggests that classical and pseudobulges exist at least up to $z=0.8$. The high frecuency of AGNs in $z>0.5$ galaxies with bulges suggests a link between AGN activity and bulge growth in morphologically normal disk galaxies at $z\sim1$. \acknowledgments We thank the anonymous referee for suggestions that improved the paper and Ignacio Trujillo and Mercedes Prieto for useful discussions. This work is based on archival data from the HST, obtained from the data archive at STScI, which is operated by AURA under NASA contract NAS5-26555. Support has been provided by the Spanish Programa Nacional de Astronom\'\i a y Astrof\'\i sica through project number AYA2006-12955 and through the Consolider-Ingenio 2010 Program grant CSD2006-00070: First Science with the GTC (http://www.iac.es/consolider-ingenio-gtc/).

This month we have our first JCE articles on Covid-19 coming through an expedited review process. The new submissions are reviewed weekly. Although we are only able to consider papers that methodologic in focus, it is good to see the substantial interest in the method challenges from authors. This is attracting the formation of new multidisciplinary author groups working virtually that may be the prelude to a new normal. In this issue we have 2 series addressing important methods issues. The first series proposes a revolution of the ecosystems of systematic reviews. [Ravaud et al](10.1016/j.jclinepi.2020.01.024){#intref0010a} argue dramatic changes are needed in the ecosystem world of systematic reviews and meta-analyses of therapeutic interventions if these are to be truly stakeholder-needs-driven. There is a mismatch between and among the primary research enterprise and the evidence synthesis enterprise; the planning, coordination and conduct of systematic reviews; there is fragmentation of available evidence that hampers providing useful information for health decision making. Some concrete solutions are proposed through the 4 papers and then vigorously commented on in 2 accompanying commentaries. These issues and their resolution is currently being actively debated within the Cochrane Collaboration. The second series is a Controversy and Debate Series. This begins with a meta-epidemiology review to look at 'subjective' versus 'objective' when contrasting patient reported outcomes with other types of outcomes such as clinician reported outcomes, laboratory tests, imaging etc. [Berthelson et al](10.1016/j.jclinepi.2020.03.025){#intref0010} argue that their meta-epidemiology study demonstrated not only is the patient the most relevant source of data for patient care, but the clinimetric performance in terms of truth discrimination and feasibility of patient reported outcomes equals the performance of clinician reported outcomes, laboratory tests, imaging etc. This paper then becomes the focus of a debate over three subsequent papers on the methodological validity of such meta-epidemiological studies in avoiding a) various design related biases such as inadequate sequence generation and inadequate allocation concealment cause bias that inflate estimates of treatment effects; and b) misclassification of trial characteristics, non-reporting biases affecting the pool of trials available for analysis, and the generalisability of the meta-analyses sampled. Just asking a question without any additional intervention can induce behavior change so this needs to be considered in any study. As part of a research program \[MEasurement Reactions In Trials (MERIT) \] to develop MRC/NIHR guidance on minimising the risk of bias in trials of healthcare interventions as a result of measurement reactivity, [Miles et al](10.1016/j.jclinepi.2020.03.014){#intref0015} report on an updated evidence-base for new guidance. They found further evidence for an overall small but heterogeneous effect after adding 10 new trials to the 33 trials across trials of conditions including alcohol, blood donation, diet, flossing, health checks, physical activity, screening, sexual behavior and vaccination. This reinforces the importance of considering this in such studies, especially where the effect size is likely to be small. Publicly available protocols are increasingly felt to be essential for best practice in systematic reviews in most health disciplines--although this is still being debated amongst environment health journal editors. The most common approach is to register these on a publicly available database such as PROSPERO. However although this requires a standardised format this is not peer reviewed. Organisations such as the Campbell and Cochrane Collaborations require these to be peer reviewed but most other journals so not require this. In this issue Rombey et al report on a survey of 49 authors of systematic reviews registered on PROSPERO to assess the perceived importance of such protocols being peer reviewed. Many agreed that peer review would or did improve the protocol but less than half felt this should be required. This needs more empiricval studies. Searching for reports of studies presented at a conference is an acknowledged approach to study identification in systematic reviews, since this allows identification of newly emerging studies, or updated findings of on-going studies, potentially ahead of journal publication; thus identifying data that can help minimise the introduction of bias into systematic reviews, since a significant proportion of studies that are reported at conferences are never published. Handsearching has traditionally been the method used to search for reports of studies presented at conferences but it is resource intensive. Cooper et al report a case study of searching conference proceedings presented at the American Society of Hematology conferences between 2016 and 2018 to assess different approaches used to try and make this more efficient. Searching Embase, PubMed or Conference Proceedings Citation Index- Science were all inadequate; keyword searching performed better and did save some time. Efficiency of exporting would be improved if journals permitted bulk downloads. The degree of robust trustworthiness of the science is the cornerstone of clinically important conclusions made in systematic reviews. This is assessed using validated instruments to assess risk of bias such as the Cochrane Risk of Bias measure. This does requite expertise and judgment so it is recommended that this be assessed independently by two assessors. [Bertizzolo et al](10.1016/j.jclinepi.2020.03.008){#intref0020} have previously shown that suboptimal agreement in risk-of-bias assessment, ranging from 81% agreement for random sequence generation to 57% for incomplete outcome data.; such disagreement dilutes the confidence the conclusion. Sometimes this is due to one assessor missing the information due to it being placed in varying parts of reports but more than two-thirds of disagreements this was related to differences in interpretation of the same information. In this issue these authors analysed external factors that might explain this beyond training--they found that assessors were more likely to allocate a good rating \[low risk of bias\] if any of the review authors are authors of the included RCTs and if the RCT in question is relatively old. Experience as assessed by number of reviews by the first author was not found to be a factor. Systematic reviews to assess the efficacy and effectiveness of interventions require the assessment of the trade off between benefit and harm. However as [Li et al](10.1016/j.jclinepi.2020.03.017){#intref0025} point out journals and databases continue to favor assessment of benefits over assessment of harms despite increasing emphasis from reporting guidelines on harms such as the Cochrane Handbook and the PRISMA harms statement. Of 120 systematic reviews, including 60 Cochrane and 60 non Cochrane ones. only nine (7.5%) reviews clearly defined safety outcomes, and seven (5.8%) defined a primary safety outcome; none stated whether the primary safety outcome was pre-defined. Journal editors should require systematic review authors to report safety information. [Puljak et al](10.1016/j.jclinepi.2020.03.019){#intref0030} report that 85% of 508 non-Cochrane systematic reviews published in Pubmed in June 2018 indicated they used the Cochrane Risk of Bias instrument. However in 41% this instrument was not correctly applied--most often due to the omission of the required details to justify the judgment of high/low/unclear risk of bias. If the non-Cochrane systemtic reviews are to be used for policy and practice decisions these details need to be made available--if space in the written journal is limited this information should be made available in a web appendix. Assessmemt of Risk of Bias in observational study designs is also controversial, even for the most widely used STROBE *(* ***ST*** *rengthening the* ***R*** *eporting of* ***OB*** *servational studies in* ***E*** *pidemiology)* instrument--a reporting guideline published in 2007 that consists of a checklist of 22 items that are considered essential for good reporting of observational studies. It has been widely endorsed by biomedical journals but few of these journals actually insisit it is used and submitted with manuscripts As part of a quantitative study on its use reported elsewhere Sharp et al report on over 150 qualitative comments made by respondents. Some of these are important for the clinical epidemiology community to consider given the increasing pressure to use such reporting guidelines. On the positive side the systematic process is an important part of scientific reporting and this structure is extremely useful for training of authors and journal reviewers. On the negative side several comments were made re timing--ie these issues should be addressed at the time of the study being designed; also some experienced investigators felt this was an inflexible overreach \[forcing the researchers into a 'straightjacket'\] and not being taken seriously by the journals in their decision-making. The authors of this paper recommend a review of the STROBE instrument 's content, training for early career researchers, and the need for better incentive and enforcement mechanisms. Three articles address statistical issues. Bayesian analyses have enormous conceptual appeal over traditional frequentist analyses since Bayesian statistics better incorporates prior probabilities based on prior knowledge from previous experiments. This approach is slowly gaining acceptance in clinical epidemiology as reflected in many statistical texts, its acceptance by many leading journals, the recommendation by the US Food and Drug Administration for Phase III trials evaluating medical devices, and its acceptance by the Cochrane Collaboration for traditional and especially network meta-analyses. It is often chosen when trials do not achieve the sample size needed to meet type 1 and type 2 estimates for a frequentist analysis. In this issue [Ferreira et al](10.1016/j.jclinepi.2020.03.021){#intref0035} report on 49 trials between 2014 and 2016 that used a Bayes analysis as the main analysis. Although many were in major journals, few met best practice standards; especially concerning is the finding that a third failed to justify the priors and if those that did, few were based on robust prior studies. The authors emphasize the need for training in the use of Bayesian methods, especially in the context of validating medical devices. The use of Bayesian guidelines should be encouraged by the editorial boards. Forecasting cancer prevalence is important for health resource planning but is a complex task, depending on past and future cancer incidence rates, past and future survival rates among cancer patients, the past and future population size and age-sex-distribution. Novikov et al report on the improvement over traditional approaches of the PIAMOD model(Prevalence Incidence Analysis Model) that used data from 1983 to 2013, to predict the cases in 2014-2016. Selecting forecasting models based on such "validation" is recommended. Xu et al report that they are concerned about the practice of excluding studies with no events in both arms in meta-analyses. They identified 442 systematic reviews in the Cochrane Library where at least one RCT had zero effects in both arms and then excluded these results in the meta-analysis. There are methods such as the generalised linear mixed model for including these. They demonstrate that if these are included over 10% of meta analyses the statistical significance of *P*-value and width of confidence interval is changed, often substantively. These results suggest the practice of excluding this data should be stopped. Finally As Tunji-Ajayi et al report whilst there is now a significant body of research asking various methodological questions about recruitment and using an array of methods to answer these questions, the same cannot be said for retention. Missing data in trials is of concern as it has the potential to introduce bias and make the trial results unreliable or in some cases unusable. This is no doubt contributed to by the fact that most consent forms highlight the right to withdraw without describing the trial consequences. They argue that it seems much more sensible to mitigate problems of missing data by designing effective approaches and strategies to maximise data collection. These authors interviewed 21 individuals \[trial participants, trial surgeons and research nurses\]. Retention was rarely discussed during clinical trial consultations. If retention was discussed it only made up 3% (at best) of the consultation content. This needs to change.

The Gulf Coast tick, *Amblyomma maculatum* ([Figure](#F1){ref-type="fig"}), is a Nearctic and Neotropical hard tick found in coastal areas of the southern United States, with inland range extensions in Kansas, Oklahoma, and some other states. It is also found in regions of several Central and South American countries that border the Gulf of Mexico and Caribbean Sea, including Mexico, Guatemala, Belize, Nicaragua, Honduras, Costa Rica, Colombia, Venezuela, and some parts of Ecuador and Peru ([@R1]). {#F1} *Rickettsia parkeri,* a member of the spotted fever group rickettsiae, was initially identified in Gulf Coast ticks in 1937([@R2])*.* In 2004, the first confirmed human infection with *R. parkeri* was reported ([@R3]). Since that report, confirmed cases of *R. parkeri* rickettsiosis have been identified in other persons in Mississippi, Virginia, and possibly other US states ([@R4]*--*[@R6]). Only a few studies, each conducted \>50 years ago, document the occurrence of *R. parkeri* in *A. maculatum* ticks ([@R2]*,*[@R7]*,*[@R8]). No contemporary surveys have documented the range of *R. parkeri* in the United States or the frequency of *R. parkeri* infection in collections of individual Gulf Coast ticks. The Study ========= *A. macu*latum ticks collected during 1996--2005 were evaluated by molecular methods for evidence of infection with *R. parkeri*. Ticks were collected from various locations in Florida, Georgia, Kentucky, Mississippi, Oklahoma, and South Carolina. Most were questing adults collected from vegetation by using flannel cloth flags; a few crawling, nonattached, nonengorged adults were obtained (4 from a coyote and 3 from human hosts), and 1 engorged nymph was removed from a cotton rat. Ticks were preserved in 70% ethanol or frozen at --80ºC until evaluation. Most individual ticks were minced with a sterile scalpel blade, and DNA was extracted by using a QIAamp Mini Kit (QIAGEN, Inc., Valencia, CA, USA). Others were minced or crushed after freezing in liquid nitrogen, and DNA was extracted by using an IsoQuick nucleic acid extraction kit (ORCA Research, Bothell, WA, USA). DNA extracts were evaluated by using nested or heminested PCR assays designed to amplify a segment of the *rompA* gene. For the primary stage of each assay, 5 µL of extract and primers 190--70 and 190--701 ([@R9]) were used. For the nested reaction, 2 µL of completed primary reaction was used as template with primers 190-FN1 (5′-AAG CAA TAC AAC AAG GTC-3′) and 190-RN1 (5′-TGA CAG TTA TTA TAC CTC-3′); for the heminested reaction, primers 190--FN1 and 190--701 were used. All reactions were prepared by using a High Fidelity PCR Master Kit (Roche Diagnostics, Indianapolis, IN, USA) with final primer concentrations of 300 nmol in a total volume of 50 μL. Thermalcycler parameters for the primary stage consisted of an initial denaturation period of 2 min at 94^o^C, followed by 40 cycles of 15 s at 94^o^C, 30 s at 60^o^C, 45 s at 72^o^C, and a 5-min extension period at 72^o^C. For the nested and heminested stages, the annealing temperature was changed to 55^o^C and the number of cycles was reduced to 30. PCR products (10 μL) were separated by electrophoresis in 2% agarose gels containing ethidium bromide. For each positive reaction, the remaining 40 μL was subjected to gel electrophoresis, and products of the appropriate size were excised. DNA was purified from the gel fragments by using the QIAquick Gel Extraction Kit (QIAGEN). Purified PCR products were sequenced using the PCR primers and the GenomeLab DTCS Quick Start Kit (Beckman Coulter, Fullerton, CA, USA). For some products, additional sequencing primers 190-SF3 (5′-GGT ACT ACT CCC GTA GGT C-3′) and 190-SR2 (5′-CCG GCA GTA AKA GTA ACA G-3′) were used to obtain complete sequences for both strands. Sequences were detected by using a Beckman CEQ 8000 automated sequencer. Sequence similarities were determined by using the BLAST program (version 2.0, National Center for Biotechnology Information, [www.ncbi.nlm.nih.gov/blast](http://www.ncbi.nlm.nih.gov/blast)). Sequence-length reaction products (excluding primers) were 590 bp for primary, 559 bp for heminested, and 540 bp for nested. DNA of *R. parkeri* was amplified from 21 (11.5%) of 182 male and female adult *A. maculatum* ticks collected in Georgia (11/64), Florida (7/89), Kentucky (1/1), Mississippi (1/24), and South Carolina (1/4), and from 1 engorged nymph collected in Oklahoma ([Table](#T1){ref-type="table"}). A unique *rompA* sequence (GenBank accession no. EF372578) amplified from 5 adult female ticks collected in Florida, Georgia, and Mississippi showed closest homology (≈94%) to several other *rompA* sequences (GenBank accession nos. EF063690, AY093696, AF120021, and DQ365801). PCR amplification of a 208-bp segment of the rickettsial 17-kDa antigen gene ([@R10]) from these same 5 ticks (GenBank accession no. EF372579) showed 100% homology with the corresponding sequences of *Rickettsia* sp. Arahna (AY360215), *R. montanensis* (DQ402377), *Rickettsia* sp. Hf332 (AB114804), *Rickettsia* sp. Is-1 (DQ344620), and "R. gravesii" (DQ269436). ###### Gulf Coast ticks (*Amblyomma maculatum*) infected with spotted fever group rickettsiae, United States, 1996--2005 *Rickettsia* sp. identified State County Total ticks positive/total ticks tested\* Year collected Stage, sex† Source ------------------------------------ ---------------- ------------ ------------------------------------------- ---------------- -------------- --------------------- *Rickettsia parkeri* Florida Duval 1/9 1999 Adult, F Vegetation Franklin 3/25 2004 Adults, M, F Vegetation 3/38 2005 Adults, M, F Vegetation Georgia Bulloch 2/20 1999 Adults, M, F Vegetation 5/16 2003 Adults, M, F Vegetation 3/24 2005 Adults, M, F Vegetation McIntosh 1/4 2005 Adult, F Vegetation Kentucky Montgomery 1/1 2003 Adult, M Human Mississippi Copiah 1/9 2002 Adult, M Vegetation Oklahoma Pittsburgh 1/1 1997 Nymph‡ *Sigmodon hispidis* South Carolina Anderson 1/4 2005 Adult, M *Canis latrans* Noncharacterized *Rickettsia* sp.§ Florida Nassau 1/13 1996 Adult, F Vegetation Franklin 1/38 2005 Adult, F Vegetation Mississippi Copiah 1/9 2002 Adult, F Vegetation Georgia Bulloch 1/20 1999 Adult, F Vegetation 1/24 2005 Adult, F Vegetation ­­­­­­­­­­­­­­\*Some specimens obtained from a given county during a given year were not collected synchronously.
†Identified by PCR; nonengorged specimens unless otherwise specified.
‡Engorged specimen.
§Based on a unique *rompA* sequence. Precise estimates of infection prevalence could not be assessed from these data because most of the ticks evaluated in this study were collected as relatively small sample sizes or were collected in a discontinuous manner as multiple samples from the same sites over several weeks or months in a given year. However, some collections were flagged synchronously at a single location, including those in Copiah County, Mississippi, during July 2002 (n = 9) and in Franklin County, Florida, during July 2004 (n = 25) and July 2005 (n = 27). Infection prevalence for each of these 3 collections was 11%--12%, which suggests that *R. parkeri* may be a relatively common inhabitant of some populations of Gulf Coast ticks*.* By comparison, the estimated prevalence of infection of tick vectors with *R. rickettsii*, the etiologic agent of Rocky Mountain spotted fever (RMSF), is typically much lower than that found by surveys elsewhere, which identified *R. rickettsii* in only 0.05%--1.3% of the collected specimens: 3,705 *Dermacentor andersoni* ticks from canyons bordering the Bitterroot Valley of Montana; 2,123 and 310 *D. variabilis* ticks from RMSF-endemic areas of North Carolina and Ohio, respectively; and 669 *A. aureolatum* ticks from São Paulo, Brazil ([@R11]*--*[@R14])*.* Conclusions =========== *R. parkeri* has been isolated in culture from Gulf Coast ticks collected in Alabama, Florida, Georgia, Mississippi, and Texas (\[*2,7,8*\]*,* C. Paddock, unpub. data). These results, combined with data from the present study, suggest that in the United States *R. parkeri* can be found anywhere that *A.* *maculatum* ticks are found. In this context, persons exposed to habitats in any region infested by Gulf Coast ticks are potentially vulnerable to infection with *R. parkeri*. A previously undescribed *rompA* sequence, identified in a few ticks collected during this survey, may represent a novel species of spotted fever group rickettsiae associated with the Gulf Coast tick. Attempts to isolate and characterize this species are in progress. Collectively, these findings suggest that the role of *A. maculatum* in the ecology of various spotted fever group rickettsiae deserves further attention. These results and the recent discovery of *R. parkeri* in *A. triste* ticks in Uruguay ([@R15]) indicate that much remains to be learned about *R. parkeri* and other rickettsiae of human-biting ticks in the Western Hemisphere and their relative contributions to the epidemiology of New World spotted fevers. *Suggested citation for this article*: Sumner JW, Durden LA, Goddard J, Stromdahl EY, Clark KL, Reeves WK, et al. Gulf Coast ticks (*Amblyomma maculatum*) and *Rickettsia parkeri,* United States. Emerg Infect Dis \[serial on the Internet\]. 2007 May \[*date cited*\]. Available from <http://www.cdc.gov/eid/content/13/5/751.htm> Mr Sumner is a molecular biologist at the Centers for Disease Control and Prevention, where he has worked extensively on the molecular detection of various *Rickettsia*, *Ehrlichia*, and *Bartonella* spp. for \>15 years. His research interests now focus primarily on PCR-based evaluation formalin-fixed tissues for bacterial and viral pathogens.

\section{Introduction} \noindent Combining quantum mechanics and general relativity, and understanding the quantum properties of the spacetime geometry, is a central open problem in fundamental physics. Among the various approaches presently pursued to address this problem (see \cite{Puna} for an overview), is loop quantum gravity (see \cite{Loops} for a recent overview). The theoretical framework of loop quantum gravity has evolved through numerous twists and re-foundations \cite{basic,Area,Barbero,RovDP,nodes,Helvetica,altri}, and it is easy for the non-experts to loose track of the basics of the theory. Furthermore, the basic structures have been found via very roundabout paths, involving mathematical tools such as Penrose's spinor and binor calculus, Temperley-Lieb algebras, the Kauffman-axioms, C*-algebraic techniques, Gelfand representation theory, infinite dimensional measures, generalized connections, projective limits and so on. Each of these tools adds a valuable new perspective and possible new handles to the overall picture, but it is now clear that none of these tools is strictly necessary for the definition of the basics theory and the derivation of the main results. Therefore, it is now a good time for a simple presentation of the basics of the loop representation and a simple derivation of its main results. We provide here such a derivation, confining ourselves to the definition of the (unconstrained) state space of the theory, and the quantization of the area. The results we present were obtained in References \cite{basic,Area,Barbero,RovDP,nodes,Helvetica}. The derivation here is slightly original, but our aim is mainly to provide a simple introduction to the subject; we refer to the original works for some proofs and details. We organize the presentation in two parts. The first part (Section \ref{Math}) is purely mathematical: we construct an Hilbert space ${\cal H}$ and certain operators on ${\cal H}$, and we solve the spectral problem for these operators. In the second part (Section \ref{Phys}), we present the motivations for a physical interpretation of the structures defined in the first part. We add two appendices. In the first, we recall a few relevant facts from $SU(2)$ representation theory, which clarify and make more explicit the construction presented in the main text; in the second, we discuss a special case disregarded in the main text. \section{Mathematics} \label{Math} In this section, we define the basic tools of the kinematics of the loop representation. These will get a physical interpretation in Section \ref{Phys}. \subsection{The Hilbert space} Let $M$ be a fixed three-dimensional compact smooth manifold. For concreteness, we choose $M$ to be (topologically) a three-sphere. Let $A$ be an $SU(2)$ connection on $M$; that is, $A$ is a smooth 1-form with values in $su(2)$, the Lie algebra of $SU(2)$. We denote the space of the smooth $su(2)$ valued 1-forms $A$ on $M$ as $\cal A$. The space $\cal A$, equipped with the {\em sup\/} norm is a topological space. We denote the space of the continuous functions $\Psi(A)$ on $\cal A$ as $L$. Equipped with the pointwise topology, $L$ is a topological vector space. An important class of functions in $L$ is given by the cylindrical functions defined as follows. Let $\Gamma$ be a graph embedded in $M$. That is, $\Gamma$ is a collection of a finite number $n$ of (oriented) piecewise analytic curves embedded in $M$, which we call links and denote as $\gamma_{1},\ldots,\gamma_{n}$, which may overlap only at their end-points, which we call nodes, and denote as $p_{1},\ldots,p_{m}$. The number of lines meeting in a node $p$ is called the valence of the node. A node may have any valence equal or larger than 1. Given a curve $\gamma$ and a connection $A$ defined on a manifold, the holonomy (or parallel transport matrix) of $A$ along $\gamma$ is defined (See Appendix A); it is an element of $SU(2)$, and we denote it as $U(\gamma,A)$. Given a graph $\Gamma$ with links $\gamma_{1},\ldots,\gamma_{n}$, and a (Haar-integrable) complex function $f$ on $(SU(2))^{n}$, we define the function \begin{equation} \Psi_{\Gamma,f}(A)\equiv f(U(\gamma_{1},A),\ldots, U(\gamma_{n},A)). \label{cylindrical} \end{equation} We call functions of the form (\ref{cylindrical}) cylindrical. The cylindrical functions are dense in $L$. Any cylindrical function $\Psi_{\Gamma',f'}(A)$, based on a graph $\Gamma'$, can be rewritten as as a cylindrical function $\Psi_{\Gamma,f}(A)$ based on a larger graph $\Gamma$ containing $\Gamma'$, by simply choosing $f$ to be independent from the links in $\Gamma$ but not in $\Gamma'$. Furthermore, any link of a graph can be broken in two links, separated by a (bivalent) node. Therefore it is clear that any two given cylindrical functions $\Psi_{\Gamma',f'}$ and $\Psi_{\Gamma'{}',g'{}'}$ can always be rewritten as based on the same graph $\Gamma$ (choosing $\Gamma$ as the smallest graph having both $\Gamma'$ and $\Gamma'{}'$ as subgraphs). Using this, we define the following quadratic form, defined for any two cylindrical functions. \begin{equation} (\Psi_{\Gamma,f},\Psi_{\Gamma,g}) \equiv \int_{(SU(2))^{n}} dU_{1}\ldots dU_{n}\ \overline{f(U_{1},\ldots,U_{n})}\ g(U_{1},\ldots,U_{n}). \label{scalarproduct} \end{equation} We write this in the suggestive form \begin{equation} (\Psi_{\Gamma,f},\Psi_{\Gamma,g}) = \int_{\cal A} {\cal D}A\ \, \overline{\Psi_{\Gamma,f}(A)}\ \Psi_{\Gamma,g}(A), \end{equation} where the meaning of the infinite dimensional integration is given only by equation (\ref{scalarproduct}). The quadratic form (\ref{scalarproduct}) defines the norm $||\Psi_{\Gamma,f}||^{2}\equiv (\Psi_{\Gamma,f},\Psi_{\Gamma,f})$ and can be extended to $L$ by continuity. By factoring away the zero-norm subspace and closing $L$ in norm, we obtain a (non-separable) Hilbert space $\cal H$. This Hilbert space plays a key role in the following. \subsection{Gauge transformations and Diffeomorphisms} The connection $A$ transforms under local $SU(2)$ gauge transformations $A \mapsto A_{V} = V^{-1}AV + V^{-1}dV$, where $V$ is a smooth map from $M$ to $SU(2)$. The Hilbert space $\cal H$ carries a natural representation of the group of the these gauge transformations $\Psi(A) \mapsto \Psi(A_{V})$. This representation is unitary, because the scalar product (\ref{scalarproduct}) is invariant. In fact, the parallel trasport matrices transform as $U(\gamma,A) \mapsto U(\gamma,A_{V})= V(x_{i})U(\gamma,A) V(x_{f})$, where $x_{i}$ and $x_{f}$ are the initial and final points of $\gamma$; and the Haar integral in (\ref{scalarproduct}) is $SU(2)$ invariant. The gauge invariant functions in $\cal H$ (which satisfy $\Psi(A)=\Psi(A_{V})$) form a proper subspace of $\cal H$, which we denote ${\cal H}_{0}$. The connection $A$ transforms under a diffeomorphism $\phi: M\to M$ as a one form $A \mapsto \phi^*A$. The Hilbert space $\cal H$ carries a natural representation of the diffeomorphism group $\Psi(A) \mapsto \Psi(\phi^*A)$. This representation is unitary, because the scalar product (\ref{scalarproduct}) is invariant. In fact, the parallel trasport matrices transform as $U(\gamma,A) \mapsto U(\gamma,\phi^* A) = U(\phi^{-1}\cdot\gamma,A)$, where $[\phi^{-1} \cdot \gamma](x) = \gamma(\phi(x))$, and this transformation has no effect on the right hand side of (\ref{scalarproduct}). Thus, the Hilbert space $\cal H$ carries a natural unitary representation of the group of the diffeomorphisms of $M$.\footnote{There are no finite norm diffeomorphism invariant functions in $\cal H$ (satisfying $\Psi(A)=\Psi(\phi^*A)$), but a space of diffeomorphism invariant (infinite norm) ``generalized functions'' can be naturally defined, and can be equipped by a natural Hilbert structure, using general techniques.} \subsection{An orthonormal basis} Consider a graph $\Gamma$ in $M$. Associate an irreducible representation $j_{i}$ of $SU(2)$ to each link $\gamma_{i}$ of the graph. This is called coloring of the links. We denote the Hilbert space on which the representation $j_{i}$ is defined as $H_{i}$. Let $p$ be a $n$-valent node of $\Gamma$ and let $\gamma_{1}\ldots\gamma_{n}$ be the $n$ links that meet in $p$, and $j_{1}\ldots j_{n}$ their associated representations. Consider the tensor product of the Hilbert spaces of these representations $H_{(j_{1})}\otimes \ldots \otimes H_{(j_{n})}$. Let $H_{p}$ be the invariant subspace (the spin-zero component) of this tensor product. Assume that $H_{p}$ has dimension equal or larger than one -- this amounts to restricting the valence of the nodes to be strictly larger than one, and to a restriction on the coloring of the links. Let us fix an orthonormal basis in $H_{p}$. Associate an element $v$ of this basis to the node $p$. Do the same for all the nodes of the graph -- that is: associate a basis element $v_{r}$ of $H_{p_{r}}$ to each node $p_{r}$. This is called coloring of the nodes. A colored graph, namely a triplet $s=(\Gamma,j_{i},v_{r})$ is called an (embedded) spin network. A spin network $s$ defines a cylindrical function $\Psi_{s}(A)$, called a ``spin network state'', as follows. The graph of the cylindrical function is the graph of the spin network. For each link $\gamma_i$, the quantity $j_{i}[U(\gamma_{i},A)]$ is a linear operator in the Hilbert space $H_{j_{i}}$. It can be viewed as an operator from the $H_{j_{i}}$ associated to its initial point to the $H_{j_{i}}$ associated to its final point. These operators can be contracted with the invariant tensors $v_{r}$ at the nodes (see Appendix A), obtaining a number $\Psi_{s}(A)$ that depends only on the spin network and the connection. We now have the remarkable result that the spin network states $\Psi_{s}$ (for a fixed choice of a basis in $H_{p}$ at each node) form an orthogonal basis in ${\cal H}_{0}$. This can be proven by direct computation, using standard $SU(2)$ representation theory machinery. The normalization of the spin network states, to obtain an orthonormal basis, is simple; it is given explicitly in \cite{RovDP}. \subsection{The operator $E(\Sigma)$} We now introduce some operators on $\cal H$. It is easier to work with indices. Let $x^{a}$ be local coordinates on $M$, $a,b,\ldots=1,2,3$ be tangent indices and $i,j,\ldots=1,2,3$ be indices on the Lie algebra of $SU(2)$. Thus the components of the connection are given by $A(x) = A_{a}(x) dx^{a}= A^{i}_{a}(x) \tau^{i} dx^{a} $, where the $\tau^{i}$ are the three generators of $SU(2)$. The first operator we define is a functional derivative ``$-i \frac{\delta}{\delta A_{a}^{i}(x)}$'' operator. A functional derivative is a distribution and needs to be suitably smeared. Because of the peculiar structure of the functions in $\cal H$, which are (limits of sequences of) functions with support on one dimension, it is sufficient to smear the functional derivative in two dimensions only, in order to have a well defined operator. Let $\Sigma: \vec{\sigma}\mapsto x^{a}(\vec\sigma)$ be an oriented (2d) surface in $M$, and let $\vec\sigma=(\sigma^{1},\sigma^{2})$ be coordinates on $\Sigma$. We define the operator \begin{equation} E^{i}(\Sigma) \equiv -i \int_{\Sigma}d\sigma^{1}d\sigma^{2} \epsilon_{abc} \frac{\partial x^{a}(\vec\sigma)}{\partial \sigma^{1}} \frac{\partial x^{b}(\vec\sigma)}{\partial \sigma^{2}} \frac{\delta}{\delta A_{c}^{i}(x(\vec\sigma))}, \label{E} \end{equation} where $\epsilon_{abc}$ is the completely antisymmetric tensor-density with $\epsilon_{123}=1$. It is easy to verify that the operator $E^{i}(\Sigma)$ is geometrically well defined, that is, it is independent of the coordinates $\vec\sigma $ and $x^a$ chosen. The operator $E^{i}(\Sigma)$ is well defined on the cylindrical functions. To see how this may happen, consider its action on the spin network state $\Psi_{s}(A)$. We denote the intersection points between the spin network $s$ and the surface $\Sigma$ as ``punctures''. To begin with, let us consider the simplest case in which the surface $\Sigma$ and the spin network $s$ intersect on single puncture $p$, where $p$ lies on (the interior of) the link $\gamma$. Let $j$ be the color of $\gamma$. A standard result, which can be obtained, for instance, by taking the first variation of the differential equation (\ref{holonomy}) defining the holonomy (see for instance \cite{Ted}), is \begin{equation} \frac{\delta}{\delta A_{a}^{i}(x)}\ U(\gamma,A) = \int_{\gamma} ds\ \frac{dx^{a}(s)}{ds}\ \delta^{3}(\gamma(s),x) \ U(\gamma(0,s),A)\ \tau^{i}\ U(\gamma(s,1),A). \label{derivative} \end{equation} Here, $s\in[0,1]$ is a coordinate along the curve $\gamma: s \to x^{a}(s)$ and the curves $\gamma(0,s)$ and $\gamma(s,1)$ are the two segments in which the point with coordinate $s$ cuts $\gamma$. It follows that the derivative of a matrix in the representation $j$ is \begin{equation} \frac{\delta}{\delta A_{a}^{i}(x)}\ j[U(\gamma,A)] = \int_{\gamma} ds\ \frac{dx^{a}(s)}{ds}\ \delta^{3}(\gamma(s),x) \ \ j[U(\gamma(0,s),A)]\ \tau^{i}_{(j)}\ j[U(\gamma(s,1),A)], \label{jderivative} \end{equation} where $\tau^{i}_{(j)}$ are the generators of the spin $j$ representation of $SU(2)$. Let us isolate $j[U(\gamma,A)]$ ($\gamma$ being the link that crosses $\Sigma$) in the state and write the state $\Psi_{s}(A)$ as \begin{equation} \Psi_{s}(A) = \Psi_{(s-\gamma)}^{lm}(A)\ j[U(\gamma,A)]_{lm}, \end{equation} where $l$ and $m$ are indices in the Hilbert space of the representation associated to $\gamma$. From (\ref{E}), the action of the operator $E^{i}(\Sigma)$ on $\Psi_{s}$ is \begin{equation} E^{i}(\Sigma)\ \Psi_{s}(A) = -i\int_{\Sigma}d\sigma^{1}d\sigma^{2} \epsilon_{abc} \frac{\partial x^{a}}{\partial \sigma^{1}} \frac{\partial x^{b}}{\partial \sigma^{2}} \ \Psi_{(s-\gamma)}^{lm}(A)\ \frac{\delta}{\delta A_{c}^{i}(x)} j[U(\gamma,A)]_{lm}. \end{equation} Using (\ref{jderivative}), we obtain \begin{eqnarray} E^{i}(\Sigma) \Psi_{s}(A) &=& -i \int_{\Sigma}d\sigma^{1}d\sigma^{2} \int_{\gamma} ds \ \epsilon_{abc} \frac{\partial x^{a}(\vec\sigma)}{\partial \sigma^{1}} \frac{\partial x^{b}(\vec\sigma)}{\partial \sigma^{2}} \ \frac{dx^{c}(s)}{ds}\ \delta^{3}(\gamma(s),x(\vec\sigma)) \nonumber \\ && \times \ \ \Psi_{(s-\gamma)}^{lm}(A) \ \left(j[U(\gamma(0,s),A)]\ \tau^{i}_{(j)}\ j[U(\gamma(s,1),A)]\right)_{lm}. \label{full} \end{eqnarray} Remarkably, the three partial derivatives combine to produce the jacobian for the change of integration coordinates from $(\sigma^{1},\sigma^{2},s)$ to $x^{1},x^{2},x^{3}$. If this jacobian is non vanishing, we perform the change of integration coordinates and then we can integrate away the delta function, obtaining \begin{equation} E^{i}(\Sigma)\ \Psi_{s}(A) = -i \Psi_{(s-\gamma)}^{lm}(A) \ \left(j(U[\gamma(0,s),A)]\ \tau^{i}_{(j)}\ j[U(\gamma(s,1),A)]\right)_{lm}. \end{equation} Thus, the effect of the operator $E^{i}(\Sigma)$ on the state $\Psi_{s}(A)$ is simply the insertion of the matrix $-i\tau^{i}_{(j)}$ in the point corresponding to the puncture. If, on the other hand, the jacobian vanishes, then the entire integral vanishes. This happens if the tangent to the loop $\frac{dx^{a}(s)}{ds}$ is tangent to the surface. In particular, for instance, this happens if the loop $\gamma$ lies entirely on the surface, in which case the puncture is not an isolated point. Therefore only isolated punctures contribute to $E^{i}(\Sigma) \Psi_{s}(A)$. The key of the above computation is the analytical expression for the (integer) intersection number $I[\Sigma, \gamma]$ between a surface $\Sigma$ and a loop $\gamma$ \begin{equation} I[\Sigma, \gamma] = \int_{\Sigma}d\sigma^{1}d\sigma^{2} \int_{\gamma} ds \ \epsilon_{abc} \frac{\partial x^{a}(\vec\sigma)}{\partial \sigma^{1}} \frac{\partial x^{b}(\vec\sigma)}{\partial \sigma^{2}} \ \frac{dx^{c}(s)}{ds}\ \delta^{3}(\gamma(s),x(\vec\sigma)). \end{equation} This integral is independent of the coordinates $\vec\sigma, s$ and $x^{a}$, and yields an integer: the (oriented) number of punctures. (The sign is determined by the relative orientation of surface and loop.) Finally, it is easy to see that what happens if the surface $\Sigma$ and the spin network $s$ intersect in more than one puncture (along the same or different links). In this case $E^{i}(\Sigma) \Psi_{s}(A)$ is a sum of one term per puncture -- each term being given by the insertion of a $\tau$ matrix. Finally, a bit more care is required for the computation of $E^{i}(\Sigma) \Psi_{s}(A)$ when the punctures are also nodes of $s$. This case is discussed in Appendix B. \subsection{The operator $A(\Sigma)$} Clearly $E^{i}(\Sigma)$ is not an $SU(2)$ gauge invariant operator. In fact, we have seen in the previous section that acting on spin network states, which are in ${\cal H}_{0}$, it gives states which, in general, are not in ${\cal H}_{0}$. Consider the operator \begin{equation} E(\Sigma)=\sqrt{E^{i}(\Sigma)E^{i}(\Sigma)}. \label{Edef} \end{equation} Acting on a state $\Psi_{s}$ that intersects $\Sigma$ only once, it gives \begin{eqnarray} E^{2}(\Sigma)\Psi_{s}(A) &=& - \Psi_{(s-\gamma)}^{lm}(A) \ \left(j(U[\gamma(0,s),A))\ \tau^{i}_{(j)}\ \tau^{i}_{(j)} \ j[U(\gamma(s,1),A)]\right)_{lm} \nonumber \\ &=& \Psi_{(s-\gamma)}^{lm}(A) \left(j[U(\gamma(0,s),A))\ j(j+1)\ j[U(\gamma(s,1),A)]\right)_{lm} \nonumber \\ &=& j(j+1)\ \Psi_{s}(A). \end{eqnarray} (The relative orientation of the surface and the link becomes irrelevant because of the square.) Thus \begin{equation} E(\Sigma)\ \Psi_{s}(A) = \sqrt{j(j+1)}\ \Psi_{s}(A). \label{jjpu} \end{equation} This is a nice result: not only $E(\Sigma)\Psi_{s}$ is in ${\cal H}_{0}$, the space of the $SU(2)$ gauge invariant states, but we even have that the spin network state $\Psi_{s}$ is an eigenstate. Both properties, however, hold only because there is a {\em single\/} puncture. It is easy to see that if there are two (or more) punctures, the cross terms spoil the gauge invariance of $E^{2}(\Sigma)\Psi_{s}$. We are now going to define a $SU(2)$ invariant operator by eliminating the cross terms. To this purpose, let $\Sigma^{(n)}=(\Sigma^{(n)}_{k}, k=1\ldots n)$ be a sequence of increasingly fine partitions of $\Sigma$ in $n$ small surfaces. We define \begin{equation} A(\Sigma) = \lim_{n\to\infty} \sum_{k} E(\Sigma^{(n)}_{k}). \label{Adef} \end{equation} For every given spin network state $\Psi_{s}$, there is a partition $(n)$ sufficiently fine for which each puncture $i$ falls on a different small surface $\Sigma_{k}$. For these $k$, equation (\ref{jjpu}) holds, while the other $E^{i}(\Sigma_{k})$ vanish. From that $(n)$ on, the limit is trivial. And therefore \begin{equation} A(\Sigma)\ \Psi_{s}=\sum_{i\in(s\cap\Sigma)} \sqrt{j_{i}(j_{i}+1)} \ \Psi_{s}, \label{A} \end{equation} where the index $i$ labels the punctures. This formula holds for all cases except for the case in which a node of $p$ is on $\Sigma$, for which see Appendix B. The operator $A(\Sigma)$ is well defined on ${\cal H}_{0}$. It is diagonalized by spin network states. It is self-adjoint. Its spectrum includes the main sequence given by (\ref{A}), namely \begin{equation} A = \sum_{i} \sqrt{j_{i}(j_{i}+1)}. \label{spettrom} \end{equation} The full spectrum is computed taking also the cases with nodes into account (see Appendix B). It is given by \begin{equation} A = \sum_{j_{i}^{u},j_{i}^{d},j_{i}^{t}} \sqrt{j^{u}_{i}(j^{u}_{i}+1) + j^{d}_{i}(j^{d}_{i}+1) -1/2 j^{t}_{i}(j^{t}_{i}+1)}. \label{sptutto} \end{equation} In conclusion, for each surface $\Sigma$ in $M$ there is a self-adjoint operator $A(\Sigma)$, with spectrum (\ref{sptutto}), which is defined by equations (\ref{E}), (\ref{Edef}) and (\ref{Adef}), and which is diagonalized by a basis of spin network states. \section{Physics} \label{Phys} We now provide a physical interpretation to the construction of the previous section. General relativity can be seen as a diffeomorphism invariant theory for a $SU(2)$ connection $A_{a}^{i}(x)$ \cite{Barbero,Ashtekar}. In this formalism, the conjugate momentum observables $E^{a}_{i}(x)$ must be interpreted as the inverse densitized triad. That is, the three-dimensional metric $g_{ab}(x)$ is defined by \begin{equation} (\det g) g^{ab}(x) = E^{a}_{i}(x) E^{b}_{i}(x). \label{g} \end{equation} The space $\cal A$ defined in Section \ref{Math} has thus a natural interpretation as the configuration space of general relativity in this formalism. The two (kinematical) gauge groups of the theory, namely local $SU(2)$ and the group of the diffeomorphisms act naturally in $\cal A$. To quantize the theory canonically, we need a Hilbert space of states on which an algebra of observables is represented. The simplest possibility is to consider a space of functions over configuration space. The Hilbert space $\cal H$ is an example of such a space. It represents an interesting candidate particularly because it carries natural unitary representations of the two invariance groups of the theory. The analogous choice would not be viable in a conventional Yang-Mills theory, because $\cal H$ is a ``far too big'' state space (in fact, non-separable), obtained by assigning finite norm to far ``too singular'' functions, having domain of dependence on one-dimensional curves. But the presence of the diffeomorphism group changes the game drastically, because the physical states invariant under (a suitable quantum version of) the diffeomorphisms are {\em de facto\/} ``smeared all over the manifold'', and form a separable Hilbert space. On the other hand, the state spaces on which conventional quantum field theory is built (say, Fock space) do not carry a sensible unitary representation of the diffeomorphism group. Their structure relies heavily on the existence of a background metric, and the difficulty of making sense of background independent notions in field theory (or string theory) is well known (see for instance the discussion in \cite{Witten}). The physical hypothesis that loop quantum gravity explores is that the correct mathematics for describing in a non-perturbative and background independent fashion a diffeomorphism invariant theory --a theory {\em of\/} spacetime, and not a theory {\em in\/} spacetime-- must be searched in rather different realms than the usual splendor of {\em conventional\/} quantum field theory. Thus we interpret the states $\Psi(A)$ in $\cal H$ as quantum states of the gravitational field (prior to imposing the quantum constraints). The conjugate momentum operator $\frac{\delta}{\delta A_{a}^{i}(x)}$ is then to be interpreted as the quantum operator corresponding to the inverse densitized triad $E^{a}_{i}(x)$ (the two have correctly, of course, the same geometrical transformation properties). More precisely, the canonical commutation relations between $A$ and $E$ are \begin{equation} \{A_{a}^{i}(x), E^{b}_{j}(y) \} = G\ \delta_{a}^{b}\ \delta_{j}^{i} \ \delta^{3}(x,y). \end{equation} (The presence of the Newton constant $G$ is due to the fact that the momentum conjugate to $A$, which is defined as the derivative of the action with respect to $\frac{dA}{dt}$, is actually $\frac{1}{G}E$. The Newton constant enters the canonical formalism and the quantum theory precisely in the same manner the mass of a free particle does.) Therefore the operator corresponding to $E$ is $-i \hbar G \frac{\delta}{\delta A}$. This operator --suitably smeared-- together with the holonomy of $A$, forms a operator algebra which quantizes the corresponding classical Poisson algebra. What is then the physical interpretation of the operator $A(\Sigma)$ constructed above? The classical limit of $A(\Sigma)$ is obtained replacing $-i \hbar G \frac{\delta}{\delta A}$ with $E$ in its definition. (Let us put $\hbar G = 1$ for a moment.) We obtain from (\ref{Adef}), (\ref{Edef}) and (\ref{E}) \begin{eqnarray} A(\Sigma) &=& \lim_{n\to\infty} \sum_{k} \sqrt{E^{i}(\Sigma^{(n)}_{k})E^{i}(\Sigma^{(n)}_{k})}. \label{uno} \\ E^{i}(\Sigma^{(n)}_{k}) &=& \int_{\Sigma^{(n)}_{k}}d\sigma^{1}d\sigma^{2} \ \epsilon_{abc} \frac{\partial x^{a}}{\partial \sigma^{1}} \frac{\partial x^{b}}{\partial \sigma^{2}} \ E^{ci}(x(\vec\sigma)). \label{inte} \end{eqnarray} For a sufficiently fine triangulation $(n)$, $\Sigma^{(n)}_{k}$ is arbitrarily small. We can thus replace the integration in (\ref{inte}) with the value of the integrand in a point $x_{k}$ in $\Sigma^{(n)}_{k}$, times the coordinate area $a(\Sigma^{(n)}_{k})$ of the small surface \begin{equation} E^{i}(\Sigma^{(n)}_{k}) = a(\Sigma^{(n)}_{k}) \ \epsilon_{abc} \frac{\partial x^{a}}{\partial \sigma^{1}} \frac{\partial x^{b}}{\partial \sigma^{2}} E^{ci}(x_{k}). \end{equation} Inserting this in (\ref{uno}), we obtain precisely the definition of Riemann integral, yielding \begin{equation} A(\Sigma) = \int_{\Sigma} d\sigma^{1} d\sigma^{2} \sqrt{\epsilon_{abc} \frac{\partial x^{a}}{\partial \sigma^{1}} \frac{\partial x^{b}}{\partial \sigma^{2}} E^{ci}(x(\vec\sigma))\ \epsilon_{def} \frac{\partial x^{d}}{\partial \sigma^{1}} \frac{\partial x^{e}}{\partial \sigma^{2}} E^{fi}(x(\vec\sigma))}. \end{equation} This expression is independent from the choice of both $x$ and $\sigma$ coordinates. The easiest way to evaluate it is to choose $x^{3}=0$ on $\Sigma$ and $\sigma^{1}=x^{1}, \sigma^{2}=x^{2}$. We obtain \begin{equation} A(\Sigma) = \int_{\Sigma} d\sigma^{1} d\sigma^{2} \sqrt{E^{3i}(x) E^{3i}(x)} = \int_{\Sigma} d\sigma^{1} d\sigma^{2} \sqrt{(\det g) g^{33}}. \end{equation} where we have used (\ref{g}). Using the formula for the inverse of a matrix, we have \begin{equation} A(\Sigma) = \int_{\Sigma} d\sigma^{1} d\sigma^{2} \sqrt{g_{11}g_{22}-g_{12}g_{12}} = \int_{\Sigma} d\sigma^{1} d\sigma^{2} \sqrt{\det {}^{2}g}. \end{equation} where ${}^{2}g$ is the 2d metric induced by $g_{ab}$ on $\Sigma$. The last equation shows that the result is covariant and is immediately recognized as the physical area of $\Sigma$. In conclusion, and restoring physical units, we have \begin{equation} \hbar G\ A(\Sigma) = {\rm Area\ of}\ \Sigma . \end{equation} The area depends on the metric and the metric is the gravitational field. In quantum gravity, the gravitational field is given by a physical operator, and therefore the area is represented by an operator as well. According to standard quantum mechanics rules, the spectrum of the operator corresponds to the possible outcomes of individual measurements of the corresponding physical quantity. Therefore the hypothesis that the quantum theory of gravity can be based on the Hilbert space $\cal H$ --which is the basic hypothesis of loop quantum gravity-- yields the physical result that Planck scale measurements of the area can yield only quantized outcomes. These are labeled (see (\ref{fullspectrum})) by $n$-tuplets of triplets of half integers $(j_{i}^{u}, j_{i}^{d}, j_{i}^{t}), i=1,\ldots,n$ and given by \begin{equation} A = \frac{1}{2}\ \hbar G\ \sum_{i} \sqrt{2 j_{i}^{u}(j_{i}^{u}+1) + 2 j_{i}^{d}(j_{i}^{d}+1) - j_{i}^{t}(j_{i}^{t}+1)} . \label{spettro} \end{equation} For $j_{i}^{t}=0$ and $j_{i}^{u}=j_{i}^{d}=j_{i}$, this reduces to the spectrum \begin{equation} A = \hbar G\ \sum_{i}\ \sqrt{j_{i}(j_{i}+1)}. \label{spettrocorto} \end{equation} A few comments are needed, in order to qualify this result. \begin{itemize} \item The overall multiplicative factor in front of the right hand side of (\ref{spettro}) is uncertain. In fact, it has been noted that the loop quantization of a $SU(2)$ theory contains an arbitrary parameter, the Immirzi parameter, which rescales the spectrum. See \cite{Immirzi} for details. \item The operator $A(\Sigma)$ is invariant under $SU(2)$ gauge transformations, but {\em not\/} under three or four dimensional diffeomorphisms. Therefore, strictly speaking it is not an observable of the theory, and we cannot directly give its spectrum physical meaning. The failure of $A(\Sigma)$ to be diff-invariant is a consequence of the fact that the area of an abstract surface defined in terms of coordinates is not a diff invariant concept. In fact, physical measurable areas in general relativity correspond to surfaces defined by physical degrees of freedom, for instance matter (the area of the surface this table) or the gravitational field itself (the area of an event horizon). However, it is reasonable to expect that the fully gauge invariant operator corresponding to a physically defined area (say defined by matter) has precisely the same mathematical form as the non gauge invariant operator studied here. The reason is that one can use matter degrees of freedom to gauge-fix the diffeomorphisms -- so that a non-diff-invariant quantity in pure gravity corresponds to a diff-invariant quantity in a gravity+matter theory. This expectation has been confirmed explicitly in a number of cases \cite{Area,Leearea}. \item Objections have been raised about the last point. See in particular \cite{Amelino}. Some objections are based on the intuition that the position of the matter defining the surface could be subjected to quantum fluctuations, preventing the possibility of defining a sharp surface. This objection is incorrect. Neither the position of matter, nor the area of a surface, have physical independent reality. It is only the gravitational field in the location determined by the matter, or, the other way around, the location of the matter in the gravitational field, that have physical reality. The two do not form independent sets of degrees of freedom subjected to independent quantum fluctuations. See \cite{observable} for a detailed discussion of this point. \item A result analogous to the one for the area holds for the volume. \item There are many spin network bases, because a basis in each $H_{p}$ must be selected in order to have a fully determined choice. For fixed $\Sigma$, we can choose a basis (in fact, many bases) that diagonalizes $A(\Sigma)$. However, we cannot choose a basis that diagonalizes $A(\Sigma)$ for every $\Sigma$. To see how this may happen, consider a the 4-valent node described in Appendix A. The area of a surface $\Sigma_{1}$ that separates $\gamma_{1}$ and $\gamma_{2}$ from $\gamma_{3}$ and $\gamma_{4}$ is diagonalized by the basis (\ref{base1}), while the area of a surface $\Sigma_{2}$ that separates $\gamma_{1}$ and $\gamma_{3}$ from $\gamma_{2}$ and $\gamma_{4}$ is diagonalized by the basis (\ref{base2}). Thus, $A(\Sigma_{1})$ and $A(\Sigma_{2})$ do not commute, and there is no basis that diagonalizes all area operators. \item The spectrum of the area operator is composed by two sequences. A main sequence (\ref{spettrocorto}), and a secondary sequence, given by (\ref{spettro}) with $j^{(t)}\ne 0$, generated by the states with nodes on the surface. Based on the intuition that such states are --in a sense-- degenerate, doubts have been raised on whether the secondary sequence must be considered a prediction of the theory on the same ground as the main sequence (\ref{spettrocorto}). \item Obviously, direct experimental verification of the predicted spectrum (\ref{spettro}) is far outside the present possibilities. Possibilities of empirical corroboration, if any, are likely to be indirect. For instance, future observations of the cosmic background gravitational radiation might perhaps reveal an imprint of the spectrum. This possibility has not yet been investigated. On the other hand, the result (\ref{spettrocorto}) is the basis of all applications of loop quantum gravity to black hole thermodynamics \cite{Helvetica,bh}. \end{itemize} In summary, we have defined the unconstrained state space of quantum gravity in the loop representation, and the basic operators of the theory. We have then defined an operator $A(\Sigma)$, associated to each surface $\Sigma$, and computed its spectrum. We have finally shown that the classical limit of $A(\Sigma)$ is the geometrical area of the surface $\Sigma$. This leads to the physical hypothesis that the spectrum of $A(\Sigma)$ describes the quantizations of the physically measurable area. This constructions puts the old suggestion that at the Planck scale geometry is discrete (in the quantum sense) in a precise framework and gives it a precise and non trivial quantitative form. For a discussion of the rest of the loop quantum gravity and the problems still open in the theory, and for detailed references, see \cite{Loops}. \section*{Acknowledgments} We thank Don Marolf for a useful exchange. This work has been supported by NSF Grant PHY-95-15506. \section*{Appendix A: Details on spin network states} We add a few details in order to clarify the construction of the previous subsection. \subsection{$SU(2)$} We begin by recalling some basic facts about $SU(2)$ representation theory, in order to fix our notation. Let $U^{A}{}_{B}$, where $A,B=0,1$ be a $SU(2)$ matrix. These matrices act on $C^{2}$ vectors $v^{A}$, and leave the unit antisymmetric objects $\epsilon^{AB}$ and $\epsilon_{AB}$ (where $\epsilon^{01}=\epsilon_{01}=1$) invariant. Using $\epsilon_{AB}$ (from, say, the left), we can lower the first index of the matrix $U$, defining $U_{AB}=\epsilon_{AC} U^{C}{}_{B}$. The irreducible representations of $SU(2)$ are labeled by their spin $j$, which is a nonnegative half-integer. The spin-$j$ representation has dimension $2j+1$ and can be obtained as the symmetrized tensor product of $2j$ copies of the fundamental (2-dimensional) representation. Thus the Hilbert space $H_{(j)}$ of the $j$ representation is the symmetrized tensor product of $2j$ copies of $C^{2}$. We write vectors in $H_{(j)}$ (in ``spinor'' notation) as completely symmetric tensors with $2j$ indices $u^{A_{1}\ldots A_{2j}}$ (there are $2j+1$ unordered combinations of $2j$ zero's and one's). For every $U\in SU(2)$, $j[U]$ is then obtained by symmetrizing the indices of the (tensor) product of $2j$ times $U^{A}{}_{B}$. \begin{equation} j[U]^{A_{1}\ldots A_{2j}}{}_{B_{1}\ldots B_{2j}} =U^{(A_{1}}{}_{(B_{1}}\ldots U^{A_{2j})}{}_{B_{2j})}. \end{equation} Again, we can lower all upper indices of $j[U]$ with $\epsilon_{AB}$'s. \subsection{The invariant tensors $v$} The invariant tensors $v$ defined in Section \ref{Math} are $SU(2)$ invariant vectors in the tensor product $H_{(j_{1})}\otimes \ldots \otimes H_{(j_{n})}$ of the representations $j_{1},\ldots,j_{n}$ associated to the links $\gamma_{1},\ldots,\gamma_{n}$ that meet at a $n$-valent node $p$. Thus they have the form $v^{(A_{1}\ldots A_{2j_{1}}),\ldots,(D_{1}\ldots D_{2j_{n}})}$. Since the only invariant object is $\epsilon^{AB}$, they are given by suitably symmetrized sequences of $\epsilon^{AB}$'s. Recall we denote the space of such vectors (the spin zero subspace of the tensor product) as $H_{p}$. $H_{p}$ has dimension zero if the valence of $p$ is one. If $p$ is bi-valent, $H_{p}$ has dimension zero unless the two links are associated to the same representation. In this case, the only invariant tensor is \begin{equation} v^{A_{1}\ldots A_{2j},\ B_{1}\ldots B_{2j}} =\epsilon^{(A_{1}(B_{1}}\ldots\ \epsilon^{A_{2j})B_{2j})}. \label{due} \end{equation} The most interesting case is if $p$ is trivalent, then $H_{p}$ has dimension zero unless the three representations satisfy the Clebsh-Gordon condition $|j_{2}-j_{3}|\le j_{1}\le j_{2}+j_{3}$ (the angular momentum addition condition: each one of three representations is in the tensor product of the other two). If they do satisfy these conditions, then $v$ is the Wigner 3-$j$ symbol, namely the symmetric form of the Clebsh-Gordon coefficients, which can be written as \begin{equation} v^{A_{1}\ldots A_{2j_{1}},\ B_{1}\ldots B_{2j_{2}},\ C_{1}\ldots C_{2j_{3}}} = \epsilon^{A_{1}B_{1}} \ldots \epsilon^{A_{a}B_{a}} \ \ \epsilon^{B_{a+1}C_{1}} \ldots \epsilon^{B_{a+b}C_{b}}\ \ \epsilon^{A_{a+1}C_{b}} \ldots \epsilon^{A_{a+c}C_{b+c}} \label{tre} \end{equation} where complete symmetrization in the $A$ indices, in the $B$ indices and in the $C$ indices is understood, and \begin{equation} 2j_{1}=a+c, \ \ \ 2j_{2}=a+b, \ \ \ 2j_{3}=b+c. \label{cg} \end{equation} Notice that the existence of three integers $a$, $b$ and $c$ satisfying (\ref{cg}) is equivalent to the Clebsh-Gordon condition on the three spins $j_{1}, j_{2}, j_{3}$. Indeed, the Clebsh-Gordon condition is equivalent to the possibility of rooting lines across a fork: if we have a triple node with $2j_{1}, 2j_{2}, 2j_{3}$ lines coming along each of the three links $\gamma_{1}, \gamma_{2}, \gamma_{3}$, then $j_{1}, j_{2}, j_{3}$ satisfy the Clebsh-Gordon condition if and only if the lines can be consistently rooted. Then $a$ lines connect $\gamma_{1}$ and $\gamma_{2}$, $b$ lines connect $\gamma_{2}$ and $\gamma_{3}$ and $c$ lines connect $\gamma_{3}$ and $\gamma_{1}$. For higher valence $n$, the $n$-indices tensor $v$ can be constructed by contracting 3-$j$ symbols. It is easier at this point to shift to vector notation. That is, to write $u^{A_{1},\ldots A_{2j}}$ as $u_{(j)}^{m}$ where $m=1\ldots(2j+1)$. Then we write the representation matrices as $j[U]^{a}{}_{b}$ (and $j[U]_{ab}$) and we write (\ref{due}) and (\ref{tre}) as $q_{(j_{1})(j_{2})}^{lm}$ and $v^{lmn}_{(j_{1}) (j_{2}) (j_{3})}$. The $4$-index $v$ have the form $v^{lmnp}_{(j_{1}) (j_{2}) (j_{3})(j_{4})}$. A basis in the space of these objects is given as \begin{equation} v^{lmnp}_{(j_{1}) (j_{2}) (j_{3})(j_{4})}= c^{j}\ e^{lmnp}_{j}{}_{(j_{1}) (j_{2}) (j_{3})(j_{4})} \end{equation} where the basis elements are \begin{equation} e^{lmnp}_{j}{}_{(j_{1}) (j_{2}) (j_{3})(j_{4})}= v^{lmr}_{(j_{1}) (j_{2}) (j)}\ \ q^{(j)}_{rs} \ \ v^{snp}_{(j)(j_{3}) (j_{4})}. \label{four} \end{equation} Here $j$ is any representation such that the Clebsh-Gordon condition are satisfied (between $j_{1}, j_{2}, j$ as well as between $j, j_{3}, j_{4}$). Since there is a finite number of representations $j$ satisfying these Clebsh Gordon conditions, the $v$ span a finite dimensional space, which is $H_{p}$. A direct calculation shows that the basis vectors (with different $j$'s) (\ref{four}) are orthogonal to each other. We can view equation (\ref{four}) as representing the expansion of a 4-valent node into two 3-valent nodes, joined by a ``virtual'' link with color $j$. Notice that in (\ref{four}) we could have grouped the 4 representations in couples differently. Namely, there exists another basis \begin{equation} a^{lmnp}_{j}{}_{(j_{1}) (j_{2}) (j_{3})(j_{4})}= v^{lnr}_{(j_{1}) (j_{3}) (j)}\ \ q^{(j)}_{rs} \ \ v^{smp}_{(j)(j_{2}) (j_{4})}. \label{base1} \end{equation} According to the recoupling theorem, the matrix of the change of basis is given by the Wigner 6-$j$ symbols \begin{equation} a^{abcd}_{j}{}_{(j_{1}) (j_{3}) (j_{2})(j_{4})}= \sum_{k} \ \left( \begin{array}{ccc}j_{1}&j_{2}& j\\ j_{3}&j_{4}& k\end{array} \right)\ \ e^{abcd}_{k}{}_{(j_{1}) (j_{2}) (j_{3})(j_{4})}. \label{base2} \end{equation} The technique of expanding a node in 3-valent nodes joined by virtual links can be used for constructing the basis of $H_{p}$ in general. For every $n$-valent node, we write a ``virtual'' tree-like (no closed loops) graph with $n$ open ends. The dimension of $H_{p}$ is given by counting the number of colorings of the virtual links which are (Clebsh-Gordon-) compatible at all virtual 3-valent nodes. Each topologically distinct virtual decomposition defines a basis in $H_{p}$. \subsection{Holonomies} We recall that given a loop $\gamma:s\in[0,1]\to \gamma(s)\in M$ and a connection $A$ on $M$, the parallel propagator $U(\gamma,A)$ is the value for $s=1$ of the $SU(2)$ matrix $U(s)$ that solves the differential equation \begin{equation} \frac{dU(s)}{ds} + \frac{d\gamma^a}{ds} A_a(\gamma(s)) U(s) = 0 \label{holonomy} \end{equation} with the initial condition $U(0)=1$, the identity element of $SU(2)$. The solution of this equation is formally written as \begin{equation} U(\gamma,A) = {\cal P}\, e^{\int^1_{0}\,A_a(\gamma(s))\ \frac{d\gamma^a}{ds}\ ds}, \end{equation} and can be given explicitly as a power series in $A$ by expanding the exponential in powers and path-ordering the powers of $A$ along $\gamma$. (That is ${\cal P}\, A(\gamma(s'))A(\gamma(s)) \equiv A(\gamma(s)) A(\gamma(s'))$ if $s<s'$.) \subsection{Spin network states} Given a spin network $s=(\Gamma,j_{i},v_{r})$, the (representative in the representation $j_{i}$ of the) parallel propagator of the connection $A$ along each link $\gamma_{i}$ defines the matrix $j[U(\gamma_{i},A)]_{lm}$. By contracting these matrices with the invariant tensors $v_{r}$ at each node $p_{r}$, we obtain the spin network state $\Psi_{s}(A)$. As a simple example, consider a graph $\Gamma$ formed by two nodes $p$ and $q$ in $M$ joined by three links $\gamma_{1}, \gamma_{2}, \gamma_{3}$. Then \begin{equation} \Psi_{(\Gamma,j_{1}j_{2}j_{3})}= v^{lmn}_{(j_{1}) (j_{2}) (j_{3})}\ j[U(\gamma_{1},A)]_{lp}\ j[U(\gamma_{2},A)]_{mq}\ j[U(\gamma_{3},A)]_{nr}\ v^{pqr}_{(j_{1}) (j_{2}) (j_{3})}. \label{example} \end{equation} (In this example, there is no need of indicating explicitly the coloring of the nodes, because they are both trivalent and therefore admit a unique coloring.) Notice that in (\ref{example}) the pattern of the indices reflects the topology of the graph $\Gamma$. In fact, spin network originated from Penrose's graphical notation for tensors. For more details, see \cite{RovDP}. \section*{Appendix B: Nodes on $\Sigma$} In this Appendix we discuss the action of $E^{i}(\Sigma)$ and $A(\Sigma)$ on $\Psi_{s}$ in the case that we have neglected in the main text, namely the case in which $s$ includes nodes which lies on $\Sigma$. Let us assume that there is a single node $p$ of $s$ on $\Sigma$; the case in which there is more than one is then trivial. Assume that $p$ is an $n-$valent node. We must distinguish the links that meet in $p$ in three classes, which we denote ``up'', ``down'' and ``tangential''. The tangential links are the ones that overlap with $\Sigma$ for a finite interval. The others links are naturally separated into two classes according to the side of $\Sigma$ they lie (which we arbitrarily label ``up'' and ``down''). We chose a basis in $H_{p}$ as follows. We decompose the $n$ valent node $p$ into a tree-like virtual graph (see Appendix A) by combining all ``up'' links into a virtual link ``up'', all ``down'' links into a virtual link ``down'', and all ``tangential'' links into a virtual link ``tangential''. We call $j^{u}, j^{d}$ and $j^{t}$, respectively, the colors of these three virtual links. These three virtual links then meet together, at the node $v_{(j^{u})(j^{d})(j^{t})}^{lmn}$. By inspecting equation (\ref{full}), we see that the operator $E_{i}(\Sigma)$ fails to ``see'' the tangential links, because for these links the three derivatives in \begin{equation} \epsilon_{abc} \frac{\partial x^{a}(\vec\sigma)}{\partial \sigma^{1}} \frac{\partial x^{b}(\vec\sigma)}{\partial \sigma^{2}} \ \frac{dx^{a}(s)}{ds} \label{jac} \end{equation} are coplanar, and the above expression vanishes. Therefore the tangential links do not contribute. On the other hand, the other two set of links contribute (by symmetry) in equal fashion. Thus $E^{i}(\Sigma)$ inserts a $\tau_{(j^{u})}$ matrix in the ``up'' link plus a $\tau_{(j^{d})}$ matrix in the ``down'' virtual link. In other words, if we write $\Psi_{s}=\Psi_{s}{}_{lmn}\ v_{(j^{u})(j^{d})(j^{t})}^{lmn}$, we have \begin{equation} E^{i}(\Sigma)\ \Psi_{s} = \Psi_{s}{}_{lmn}\ \frac{1}{2}\left(\tau^{i}_{(j^{u})}{}_{p}^{l}\ \delta {}_{q}^{m} - \delta_{p}^{l}\ \tau^{i}_{(j^{d})}{}_{q}^{m} \right) \ v_{(j^{u})(j^{d})(j^{t})}^{pqn}. \label{action} \end{equation} Notice the minus sign coming from the inverse orientation of the jacobian (\ref{jac}) in the two cases. For a more complete derivation of (\ref{action}), see \cite{nodes}. Since $v$ is an invariant tensor, \begin{equation} \left(\tau^{i}_{(j^{u})}+\tau^{i}_{(j^{d})}+ \tau^{i}_{(j^{t})}\right)\ v_{(j^{u})(j^{d})(j^{t})}=0, \end{equation} from which \begin{equation} (\tau^{i}_{(j^{u})}+\tau^{i}_{(j^{d})})^{2} = (\tau^{i}_{(j^{t})})^{2} \end{equation} when acting on $v_{(j^{u})(j^{d})(j^{t})}$. Using this, simple algebra yields \begin{equation} (\tau^{i}_{(j^{u})}-\tau^{i}_{(j^{d})})^{2} = 2(\tau^{i}_{(j^{u})})^{2}+2(\tau^{i}_{(j^{u})})^{2}- (\tau^{i}_{(j^{u})})^{2}= 2 j^{u}(j^{u}+1) + 2 j^{d}(j^{d}+1) - j^{t}(j^{t}+1). \end{equation} In conclusion \begin{equation} A(\Sigma)\ \Psi_{s} = \frac{1}{2} \sqrt{2 j^{u}(j^{u}+1) + 2 j^{d}(j^{d}+1) - j^{t}(j^{t}+1)}\ \Psi_{s}. \end{equation} And in general \begin{equation} A(\Sigma)\ \Psi_{s} = \frac{1}{2} \sum_{i\in(s\cap\Sigma)} \sqrt{2 j_{i}^{u}(j_{i}^{u}+1) + 2 j_{i}^{d}(j_{i}^{d}+1) - j_{i}^{t}(j_{i}^{t}+1)}\ \Psi_{s}. \label{fullspectrum} \end{equation} which is valid for all spin network states. It includes also the case of a link $\gamma$ crossing $\Sigma$, because we can always rewrite $\gamma$ as two links joined by a 2-valent node on $\Sigma$, with $j^{t}=0$ and $j^{u}=j^{d}$. In this case, (\ref{fullspectrum}) reduces to (\ref{A}).