Datasets:

kjappelbaum
/

chemnlp-arxiv

Dataset card Viewer Files Files and versions Community

Split (1)

train · 1.56M rows

text stringlengths 0 643k ⌀	meta stringlengths 137 151 ⌀
# Introduction Let $G$ be a simple undirected graph with the vertex set $V(G)$ and the edge set $E(G)$. A vertex with degree one is called a pendant vertex. The distance between the vertices $u$ and $v$ in graph $G$ is denoted by $d_G(u,v)$. A cycle $C$ is called chordless if $C$ has no cycle chord (that is an edge not in the edge set of $C$ whose endpoints lie on the vertices of $C$). The Induced subgraph on vertex set $S$ is denoted by $\langle S\rangle$. A path that starts in $v$ and ends in $u$ is denoted by $\stackrel\frown{v u}$. A traceable graph is a graph that possesses a Hamiltonian path. In a graph $G$, we say that a cycle $C$ is formed by the path $Q$ if $\| E(C) \setminus E(Q) \| = 1$. So every vertex of $C$ belongs to $V(Q)$. In 2011 the following conjecture was proposed: Conjecture $\,$ also appears in Problem 516. There are a few partial results known for Conjecture . Kostochka noticed that the Petersen graph, the prisms over cycles, and many other graphs have a decomposition desired in Conjecture . Ozeki and Ye proved that the conjecture holds for 3-connected cubic plane graphs. Furthermore, it was proved by Bachstein that Conjecture $\,$ is true for every 3-connected cubic graph embedded in torus or Klein-bottle. Akbari, Jensen and Siggers showed that Conjecture $\,$ is true for Hamiltonian cubic graphs. In this paper, we show that Conjecture $\,$ holds for traceable cubic graphs. # Results Before proving the main result, we need the following lemma. Remark 1. [\[remark:1\]]{#remark:1 label="remark:1"} Let $C$ be a cycle formed by the path $Q$. Then clearly there exists a chordless cycle formed by $Q$. Now, we are in a position to prove the main result. Remark 2. [\[remark:2\]]{#remark:2 label="remark:2"} Indeed, in the proof of the previous theorem we showed a stronger result, that is, for every traceable cubic graph there is a decomposition with at most two cycles.	{'timestamp': '2016-07-19T02:04:55', 'yymm': '1607', 'arxiv_id': '1607.04768', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04768'}
# Principle of nano strain-amplifier Figure [\[fig:fig1\]](#fig:fig1){reference-type="ref" reference="fig:fig1"}(a) and 1(b) show the concept of the conventional structures of piezoresistive sensors. The piezoresistive elements are either released from, or kept on, the substrate. The sensitivity ($S$) of the sensors is defined based on the ratio of the relative resistance change ($\Delta R/R$) of the sensing element and the strain applied to the substrate ($\varepsilon_{sub}$): \[S = (\Delta R/R)/\varepsilon_{sub} \label{eq:sensitivity}\] In addition, the relative resistance change $\Delta R/R$ can be calculated from the gauge factor ($GF$) of the material used to make the piezoresistive elements: $\Delta R/R = GF \varepsilon_{ind}$, where $\varepsilon_{ind}$ is the strain induced into the piezoresistor. In most of the conventional strain gauges as shown in Fig. [\[fig:fig1\]](#fig:fig1){reference-type="ref" reference="fig:fig1"} (a,b), the thickness of the sensing layer is typically below a few hundred nanometers, which is much smaller than that of the substrate. Therefore, the strain induced into the piezoresistive elements is approximately the same as that of the substrate ($\varepsilon_{ind} \approx \varepsilon_{sub}$). Consequently, to improve the sensitivity of strain sensors (e.g. enlarging $\Delta R/R$), electrical approaches which can enlarge the gauge factor ($GF$) are required. Nevertheless, as aforementioned, the existence of the large gauge factor in nanowires due to quantum confinement or surface state, is still considered as controversial. It is also evident from Eq. [\[eq:sensitivity\]](#eq:sensitivity){reference-type="ref" reference="eq:sensitivity"} that the sensitivity of strain sensors can also be improved using a mechanical approach, which enlarges the strain induced into the piezoresistive element. Figure [\[fig:fig1\]](#fig:fig1){reference-type="ref" reference="fig:fig1"}(c) shows our proposed nano strain-amplifier structure, in which the piezoresistive nanowires are locally fabricated at the centre of a released bridge. The key idea of this structure is that, under a certain strain applied to the substrate, a large strain will be concentrated at the locally fabricated SiC nanowires. The working principle of the nano strain-amplifier is similar to that of the well-known dogbone structure, which is widely used to characterize the tensile strength of materials. That is, when a stress is applied to the dogbone-shape of a certain material, a crack, if generated, will occur at the middle part of the dogbone. The large strain concentrated at the narrow area located at the centre part with respect to the wider areas located at outer region, causes the crack. Qualitative and quantitative explanations of the nano strain-amplifier are presented as follows. For the sake of simplicity, the released micro frame and nanowire (single wire or array) of the nano strain-amplifier can be considered as solid springs, Fig. [\[fig:fig1\]](#fig:fig1){reference-type="ref" reference="fig:fig1"}(d). The stiffness of these springs are proportional to their width ($w$) and inversely proportional to their length (l): $K \propto w/l$. Consequently, the model of the released nanowire and micro frames can be simplified as a series of springs, where the springs with higher stiffness correspond to the micro frame, and the single spring with lower stiffness corresponds to the nanowire. It is well-known in classical physics that, for serially connected springs, a larger strain will be concentrated in the low--stiffness string, while a smaller strain will be induced in the high--stiffness string. The following analysis quantitatively explained the amplification of the strain. When a tensile mechanical strain ($\varepsilon_{sub}$) is applied to the substrate, the released structure will also be elongated. Since the stiffness of the released frame is much smaller than that of the substrate, it is safe to assume that the released structure will follows the elongation of the substrate. The displacement of the released structure $\Delta L$ is: \[\Delta L = \Delta L_m + \Delta L_n = L_m \varepsilon_m + L_n \varepsilon_n \label{eq:displacement}\] where $L_m$, $L_n$ are the length; $\Delta L_m$, $\Delta L_n$ are the displacement; and $\varepsilon_m$, $\varepsilon_n$ are the strains induced into the micro spring and nano spring, respectively. The subscripts m and n stand for the micro frames and nanowires, respectively. Furthermore, due to the equilibrium of the stressing force ($F$) along the series of springs, the following relationship is established: $F= K_m\Delta L_m = K_n \Delta L_n$, where $K_m$, $K_n$ are the stiffness of the released micro frames and nanowires, respectively. Consequently the relationship between the displacement of the micro frame (higher stiffness) and nanowires (lower stiffness) is: \[\frac{\Delta L_m}{\Delta L_n}=\frac{K_n}{K_m}=\frac{L_mw_n}{L_nw_m} \label{eq:euili}\] Substituting Eqn. [\[eq:euili\]](#eq:euili){reference-type="ref" reference="eq:euili"} into Eqn. [\[eq:displacement\]](#eq:displacement){reference-type="ref" reference="eq:displacement"}, the strain induced into the locally fabricated nanowires is: \[\varepsilon_n = \frac{\Delta L_n}{L_n} = \frac{1}{1-\frac{w_m-w_n}{w_m}\frac{L_m}{L}}\varepsilon_{sub} \label{eq:strainamp}\] Equation [\[eq:strainamp\]](#eq:strainamp){reference-type="ref" reference="eq:strainamp"} indicates that increasing the ratio of $w_m/w_n$ and $L_m/L_n$ significantly amplifies the strain induced into the nanowire from the strain applied to the substrate. This model is also applicable to the case of nanowire arrays, in which $w_n$ is the total width of all nanowires in the array. The theoretical model is then verified using the finite element analysis (FEA). In the FEA simulation, we compare the strain induced into (i) non released nanowires, (ii) the conventionally released nanowires, and (iii) our nano strain-amplifier structure, using COMSOL Multiphysics . In our nano strain amplifying structure, the width of the released frame was set to be 8 $\mu$m, while the width of each nanowire in the array (3 wires) was set to be 370 nm. The nanowires array structure was selected as it can enhance the electrical conductance of the SiC nanowires resistor which makes the subsequent experimental demonstration easier. The ratio between the length of nanowires and micro bridge was set to be 1: 20. With this geometrical dimensions, strain induced into nanowires array $\varepsilon_n$ was numerically calculated to be approximately 6 times larger than $\varepsilon_{sub}$, Eqn. [\[eq:strainamp\]](#eq:strainamp){reference-type="ref" reference="eq:strainamp"}. The simulation results show that for all structure, the elongation of non-released and released nanowires follow that of the substrate. In addition, strain was almost completely transferred into conventional released and non-released structures. Furthermore, the ratio of the strain induced in to the locally fabricated nanowires was estimated to be 5.9 times larger than that of the substrate, Fig. [\[fig:fig2\]](#fig:fig2){reference-type="ref" reference="fig:fig2"}. These results are in solid agreement with the theoretical analysis presented above. For a nanowire array with an average width of 470 nm, the amplified gain of strain was found to be 4.5. Based on the theoretical analysis, we conducted the following experiments to demonstrate the high sensitivity of SiC nanowire strain sensors using the nano strain-amplifier. A thin 3C-SiC film with its thickness of 300 nm was epitaxially grown on a 150 mm diameter Si wafer using low pressure chemical vapour deposition. The film was in situ doped using Al dopants. The carrier concentration of the p-type 3C-SiC was found to be $5 \times 10^{18}$ cm$^{-3}$, using a hot probe technique. The details of the characteristics of the grown film can be found elsewhere. Subsequently, I-shape p-type SiC resistors with aluminum electrodes deposited on the surface were patterned using inductive coupled plasma (ICP) etching. As the piezoresistance of p-type 3C-SiC depends on crystallographic orientation, all SiC resistors of the present work were aligned along \[110\] direction to maximize the piezoresistive effect. Next, the micro scale SiC resistors were then released from the Si substrate using dry etching (XeF$_2$). Finally, SiC nanowire arrays were formed at the centre of the released bridge using focused ion beam (FIB). Two types of nanowire array were fabricated with three nanowires for each array. The average width of each nanowire in each type were 380 nm and 470 nm, respectively. Figure [\[fig:fig3\]](#fig:fig3){reference-type="ref" reference="fig:fig3"} shows the SEM images of the fabricated samples, including the conventional released structure, non-released nanowires, and the nano strain-amplifier. The current voltage (I-V) curves of all fabricated samples were characterized using a HP 4145 parameter analyzer. The linear relationship between the applied voltage and measured current, indicated that Al made a good Ohmic contact with the highly doped SiC resistance, Fig. [\[fig:IV\]](#fig:IV){reference-type="ref" reference="fig:IV"}. Additionally, the electrical conductivity of both nanowires and micro frame estimated from the I-V curve and the dimensions of the resistors shows almost the same value. This indicated that the FIB process did not cause a significant surface damage to the fabricated nanowires. The bending experiment was used to characterize the piezoresistive effect in micro size SiC resistors and locally fabricated SiC nanowire array. In this experiment one end of the Si cantilever (with a thickness of 625 $\mu$m, and a width of 7 mm) was fixed while the other end was deflected by applying different forces. The distance from the fabricated nanowires to the free end of the Si cantilever was approximately 45 mm. The strain induced into the Si substrate is $\varepsilon_\text{sub} = Mt/2EI$, where $M$ is the applied bending moment; and $t$, $E$ and $I$ are the thickness, Young's modulus and the moment of inertia of the Si cantilever, respectively. The response of the SiC resistance to applied strain was then measured using a multimeter (Agilent A). The relative resistance change ($\Delta R/R$) of the micro and nano SiC resistors was plotted against the strain induced into the Si substrate $\varepsilon_{sub}$, Fig. [\[fig:DRR\]](#fig:DRR){reference-type="ref" reference="fig:DRR"}(a). For all fabricated samples, the relative resistance change shows a good linear relationship with the applied strain ($\varepsilon_{sub}$). In addition, with the same applied strain to the Si substrate, the resistance change of the SiC nanowires using the nano strain-amplifier was much larger than that of the the SiC micro resistor and the conventional non-released SiC nanowires. In addition, reducing the width of the SiC nanowires also resulted in the increase of the sensitivity. The magnitude of the piezoresistive effect in the nano strain-amplifier as well as conventional structures were then quantitatively evaluated based on the effective gauge factor ($GF_{eff}$), which is defined as the ratio of the relative resistance change to the applied strain to the substrate: $GF_{eff} = (\Delta R/R)/\varepsilon_{sub}$. Accordingly, the effective gauge factor of the released micro SiC was found to be 28, while that of the non-released SiC nanowires was 35. From the data shown in Fig. [\[fig:DRR\]](#fig:DRR){reference-type="ref" reference="fig:DRR"}, the effective gauge factor of the 380 nm and 470 nm SiC nanowires in the nano strain-amplifier were calculated as 150 and 124, respectively. Thus for nanowire arrays with average widths of 380 nm and 470 nm, the sensitivity of the nano strain-amplifier was 5.4 times and 4.6 times larger than the bulk SiC, respectively. These results were consistent with analytical and numerical models presented above. The relative resistance change of the nano strain-amplifier also showed excellent linearity with the applied strain, with a linear regression of above 99%. The resistance change of the nano strain-amplifier can also be converted into voltage signals using a Wheatstone bridge, Fig. [\[fig:DRR\]](#fig:DRR){reference-type="ref" reference="fig:DRR"}(b). The output voltage of the nano strain-amplifier increases with increasing tensile strains from 0 ppm to 180 ppm, and returned to the initial value when the strain was completely removed, confirming a good repeatability after several strain induced cycles. The linearity of the relative resistance change, and the repeatability indicate that the proposed structure is promising for strain sensing applications. In conclusion, this work presents a novel mechanical approach to obtain highly sensitive piezoresistance in nanowires based on a nano strain-amplifier. The key factor of the nano strain-amplifier lies on nanowires locally fabricated on a released micro structure. Experimental studies were conducted on SiC nanowires, confirming that by utilizing our nano strain-amplifier, the sensitivity of SiC nanowires was 5.4 times larger than that of conventional structures. This result indicated that the nano strain-amplifier is an excellent platform for ultra sensitive strain sensing applications.	{'timestamp': '2016-07-18T02:07:38', 'yymm': '1607', 'arxiv_id': '1607.04531', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04531'}
# Introduction {#intro} Gas has a fundamental role in shaping the evolution of galaxies, through its accretion on to massive haloes, cooling and subsequent fuelling of star formation, to the triggering of extreme luminous activity around super massive black holes. Determining how the physical state of gas in galaxies changes as a function of redshift is therefore crucial to understanding how these processes evolve over cosmological time. The standard model of the gaseous interstellar medium (ISM) in galaxies comprises a thermally bistable medium (@Field:1969) of dense ($n \sim 100$ cm$^{-3}$) cold neutral medium (CNM) structures, with kinetic temperatures of $T_{\rm k} \sim 100$ K, embedded within a lower-density ($n \sim 1$ cm$^{-3}$) warm neutral medium (WNM) with $T_{\rm k} \sim 10^{4}$ K. The WNM shields the cold gas and is in turn ionized by background cosmic rays and soft X-rays (e.g. @Wolfire:1995 [@Wolfire:2003]). A further hot ($T_{\rm k} \sim 10^{6}$ K) ionized component was introduced into the model by, to account for heating by supernova-driven shocks within the inter-cloud medium. In the local Universe, this paradigm has successfully withstood decades of observational scrutiny, although there is some evidence (e.g. @Heiles:2003b; @Roy:2013b; @Murray:2015) that a significant fraction of the WNM may exist at temperatures lower than expected for global conditions of stability, requiring additional dynamical processes to maintain local thermodynamic equilibrium. Since atomic hydrogen (H i) is one of the most abundant components of the neutral ISM and readily detectable through either the 21 cm or Lyman $\alpha$ lines, it is often used as a tracer of the large-scale distribution and physical state of neutral gas in galaxies. The 21 cm line has successfully been employed in surveying the neutral ISM in the Milky Way (e.g. @McClure-Griffiths:2009 [@Murray:2015]), the Local Group (e.g. @Kim:2003 [@Bruns:2005; @Braun:2009; @Gratier:2010]) and low-redshift Universe (see @Giovanelli:2016 for a review). However, beyond $z \sim 0.4$ (@Fernandez:2016) H i emission from individual galaxies becomes too faint to be detectable by current 21 cm surveys and so we must rely on absorption against suitably bright background radio (21 cm) or UV (Lyman-$\alpha$) continuum sources to probe the cosmological evolution of H i. The bulk of neutral gas is contained in high-column-density damped Lyman-$\alpha$ absorbers (DLAs, $N_{\rm HI} \geq 2 \times 10^{20}$ cm$^{-2}$; see @Wolfe:2005 for a review), which at $z \gtrsim 1.7$ are detectable in the optical spectra of quasars. Studies of DLAs provide evidence that the atomic gas in the distant Universe appears to be consistent with a multi-phase neutral ISM similar to that seen in the Local Group (e.g. @Lane:2000; @Kanekar:2001c; @Wolfe:2003b). However, there is some variation in the cold and warm fractions measured throughout the DLA population (e.g. @Howk:2005; @Srianand:2005 [@Lehner:2008]; @Jorgenson:2010; @Carswell:2011 [@Carswell:2012; @Kanekar:2014a]; @Cooke:2015; @Neeleman:2015). The 21-cm spin temperature affords us an important line-of-enquiry in unraveling the physical state of high-redshift atomic gas. This quantity is sensitive to the processes that excite the ground-state of H i in the ISM (@Purcell:1956 [@Field:1958; @Field:1959b; @Bahcall:1969]) and therefore dictates the detectability of the 21 cm line in absorption. In the CNM the spin temperature is governed by collisional excitation and so is driven to the kinetic temperature, while the lower densities in the WNM mean that the 21 cm transition is not thermalized by collisions between the hydrogen atoms, and so photo-excitation by the background Ly $\alpha$ radiation field becomes important. Consequently the spin temperature in the WNM is lower than the kinetic temperature, in the range $\sim$`<!-- -->`{=html}1000--5000 K depending on the column density and number of multi-phase components (@Liszt:2001). Importantly, the spin temperature measured from a single detection of extragalactic absorption is equal to the harmonic mean of the spin temperature in individual gas components, weighted by their column densities, thereby providing a method of inferring the CNM fraction in high-redshift systems. Surveys for 21 cm absorption in known redshifted DLAs have been used to simultaneously measure the column density and spin temperature of H i (see @Kanekar:2014a and references therein). There is some evidence for an increase (at $4\,\sigma$ significance) in the spin temperature of DLAs at redshifts above $z = 2.4$, and a difference (at $6\,\sigma$ significance) between the distribution of spin temperatures in DLAs and the Milky Way (@Kanekar:2014a). The implication that at least 90 per cent of high-redshift DLAs may have CNM fractions significantly less than that measured for the Milky Way has important consequences for the heating and cooling of neutral gas in the early Universe and star formation (e.g. @Wolfe:2003a). However, these targeted observations rely on the limited availability of simultaneous 21 cm and optical/UV data for the DLAs and assumes commonality between the column density probed by the optical and radio sight-lines. The first issue can be overcome by improving the sample statistics through larger 21 cm line surveys of high-redshift DLAs, but the latter requires improvements to our methodology and understanding of the gas distribution in these systems. There are also concerns about the accuracy to which the fraction of the source structure subtended by the absorber can be measured in each system, which can only be resolved through spectroscopic very long baseline interferometry (VLBI). It has been suggested that the observed evolution in spin temperature could be biased by assumptions about the radio-source covering factor (@Curran:2005) and its behaviour as a function of redshift (@Curran:2006b [@Curran:2012b]). In this paper we consider an approach using the statistical constraint on the average spin temperature achievable with future large 21 cm surveys using precursor telescopes to the Square Kilometre Array (SKA). This will enable independent verification of the evolution in spin temperature at high redshift and provide a method of studying the global properties of neutral gas below $z \approx 1.7$, where the Lyman $\alpha$ line is inaccessible using ground-based observatories. In an early attempt at a genuinely blind 21 cm absorption survey, used pilot data from the Arecibo Legacy Fast Arecibo L-band Feed Array (ALFALFA) survey to obtain upper limits on the column density frequency distribution from 21 cm absorption at low redshift ($z \lesssim 0.06$). However, they also noted that the number of detections could be used to make inferences about the ratio of the spin temperature to covering factor. Building upon this work, found that their upper limits on the frequency distribution function measured from the 40 per cent ALFALFA survey ($\alpha$.40; @Haynes:2011) could only be reconciled with measurements from other low-redshift 21 cm surveys if the typical spin temperature to covering factor ratio was greater than 500 K. At higher redshifts, found that the number density of 21 cm absorbers in known Mg ii absorbers appeared to decrease with redshift above $z \sim 1$, consistent with a reduction in the CNM fraction. We pursue this idea further by investigating whether future wide-field 21 cm surveys can be used to measure the average spin temperature in distant galaxies that are rich in atomic gas. # The expected number of intervening H i absorbers {#section:expected_number} We estimate the expected number of intervening H i systems towards a sample of background radio sources by evaluating the following integral over all sight-lines \[\label{equation:expected_number} \mu = \iint{f(N_{\rm HI},X)\,\mathrm{d}X\,\mathrm{d}N_{\rm HI}},\] where $f(N_{\rm HI}, X)$ is the frequency distribution as a function of column density ($N_{\rm HI}$) and comoving path length ($X$). We use the results of recent surveys for 21 cm emission in nearby galaxies (e.g. @Zwaan:2005) and high-redshift Lyman-$\alpha$ absorption in the Sloan Digitial Sky Survey (SDSS; e.g. @Prochaska:2005; @Noterdaeme:2009), which show that $f(N_{\rm HI}, X)$ can be parametrized by a gamma function of the form \[f(N_{\rm HI}, X) = \left({f_{\ast} \over N_{\ast}}\right)\left({N_{\rm HI} \over N_{\ast}}\right)^{-\beta}\exp{\left(-{N_{\rm HI} \over N_{\ast}}\right)}\,\mathrm{cm}^{2},\] where $f_{\ast} = 0.0193$, $\log_{10}(N_{\ast}) = 21.2$ and $\beta = 1.24$ at $z = 0$ (@Zwaan:2005), and $f_{\ast} = 0.0324$, $\log_{10}(N_{\ast}) = 21.26$ and $\beta = 1.27$ at $z \approx 3$ (@Noterdaeme:2009). While the observational data do not yet constrain models for evolution of the H i distribution at intermediate redshifts between $z \sim 0.1$ and $3$[^1], it is known to be much weaker than the significant decline seen in the global star-formation rate and molecular gas over the same epoch (e.g. @Lagos:2014). We therefore carry out a simple linear interpolation between the low and high redshift epochs to estimate $f(N_{\rm HI},X)$ as a function of redshift. The probability of detecting an absorbing system of given column density depends on the sensitivity of the survey, the flux density and structure of the background source and the fraction of H i in the lower spin state, given by the spin temperature. We express the column density ($N_{\rm HI}$; in atoms cm$^{-2}$) in terms of the optical depth ($\tau$) and spin temperature ($T_{\rm spin}$; in K) by \[\label{equation:column_density} N_{\rm HI} = 1.823\times10^{18}\,T_{\rm spin} \int{\tau(v)\mathrm{d}v},\] where the integral is performed across the spectral line in the system rest-frame velocity $v$ (in km s$^{-1}$). We then express the optical depth in terms of the observables as \[\tau =-\ln\left[1 + {\Delta{S}\over c_{\rm f}S_{\rm cont}}\right],\] where $\Delta{S}$ is the observed change in flux density due to absorption, $S_{\rm cont}$ is the background continuum flux density and $c_{\rm f}$ is the (often unknown) fraction of background flux density subtended by the intervening gas. We assume that a single intervening system can be described by a Gaussian velocity distribution of full width at half maximum (FWHM) dispersion ($\Delta{v_{\rm 50}}$) and peak optical depth ($\tau_{\rm peak}$), so that can be re-written as \[\label{equation:column_density_gaussian} N_{\rm HI} = 1.941\times10^{18}\,T_{\rm spin}\,\tau_{\rm peak}\,\Delta{v_{\rm 50}}.\] If we further assume that the rms spectral noise is Gaussian, with a standard deviation $\sigma_{\rm chan}$ per independent channel $\Delta{v_{\rm chan}}$, then the 5$\sigma$ column density detection limit is given by \[N_{5\sigma} \approx 1.941\times10^{18}\,T_{\rm spin}\,\tau_{\rm 5\sigma}\,\Delta{v_{\rm conv}},\] where \[\label{equation:optical_depth_limit} \tau_{5\sigma} \approx-\ln\left[1-{5\,\sigma_{\rm chan}\over c_{\rm f}\,S_{\rm cont}}\sqrt{\Delta{v}_{\rm chan}\over \Delta{v_{\rm conv}}}\right],\] and $\Delta{v_{\rm conv}} \approx \sqrt{\Delta{v}_{\rm chan}^{2} + \Delta{v}_{50}^{2}}$, which is the observed width of the line, given by the convolution of the physical velocity distribution and the spectral resolution of the telescope. We now redefine $\mu$ as the expected number of intervening H i detections in our survey as a function of the column density sensitivity along each sight-line where each comoving path element $\delta{X}(z)$[^2] in the integral defined by is given by \[\delta{X}(z)= \begin{cases} {\delta{z}\,(1+z)^{2}\over \sqrt{(1+z)^{2}(1+z\Omega_{\rm M})-z(z+2)\Omega_{\rm \Lambda}}}, & \text{if}\ N_{\rm HI} \geq N_{5\sigma}, \\ 0, & \text{otherwise}. \end{cases}\] To calculate the column density sensitivity for each comoving element we draw random samples for $\Delta{v}_{50}$ and $c_{\rm f}$ from continuous prior distributions based on existing evidence. In the case of $\Delta{v}_{50}$ we use a log-normal distribution obtained from a simple least-squares fit to the sample distribution from previous 21-cm absorption surveys reported in the literature (see )[^3], assuming that this correctly describes the true distribution for the population of DLAs. However, direct measurement of the H i covering factor is significantly more difficult and so for the purposes of this work we draw random samples assuming a uniform distribution between 0 and 1. In , we show a comparison between this assumption and the sample distribution estimated by from their main sample of 37 quasars. Kanekar et al. used VLBI synthesis imaging to measure the fraction of total quasar flux density contained within the core, which was then used as a proxy for the covering factor. By carrying out a two-tailed Kolmogorov-Smirnov (KS) test of the hypothesis that the Kanekar et al. data are consistent with our assumed uniform distribution, we find that this hypothesis is rejected at the 0.05 level, but not at the 0.01 level (this outcome is dominated by the paucity of quasars in the sample with $c_{\rm f} \lesssim 0.2$). It is therefore possible that the population distribution of H i covering factors may deviate somewhat from the uniform distribution assumed in this work. We discuss the implications of this further in . # A 21 cm absorption survey with ASKAP {#section:all_sky_survey} We use the Australian Square Kilometre Array Pathfinder (ASKAP; @Johnston:2007) as a case study to demonstrate the expected results from planned wide-field surveys for 21 cm absorption (e.g. the ASKAP First Large Absorption Survey in H i--Sadler et al., the MeerKAT Absorption Line Survey--Gupta et al., and the Search for HI absorption with AperTIF--Morganti et al.). ASKAP is currently undergoing commissioning. Proof-of-concept observations with the Boolardy Engineering Test Array (@Hotan:2014) have already been used to successfully detect a new H i absorber associated with a probable young radio galaxy at $z = 0.44$ (@Allison:2015a). Here we predict the outcome of a future 2 h-per-pointing survey of the entire southern sky ($\delta \leq +10\degr$) using the full 36-antenna ASKAP in a single 304 MHz band between 711.5 and 1015.5 MHz, equivalent to H i redshifts between $z = 0.4$ and 1.0. Our expectations of the ASKAP performance are based on preliminary measurements by using the prototype Mark II phase array feed. We estimate the noise per spectral channel using the radiometer equation \[\sigma_{\rm chan} = {S_{\rm system} \over \sqrt{n_{\rm pol}\,n_{\rm ant}\,(n_{\rm ant}-1)\,\Delta{t}_{\rm in}\,\Delta{\nu}_{\rm chan}}},\] where $S_{\rm system}$ is the system equivalent flux density, $n_{\rm pol}$ is the number of polarizations, $n_{\rm ant}$ is the number of antennas, $\Delta{t}_{\rm in}$ is the on-source integration time and $\Delta{\nu}_{\rm chan}$ is the spectral resolution in frequency. The sensitivity of the telescope in the 711.5-1015.5 MHz band is expected to vary between $S_{\rm system} \approx 3200$ and $2000$ Jy, with the largest change in sensitivity between 700 and 800 MHz. ASKAP has dual linear polarization feeds, 36 antennas and a fine filter bank that produces 16 416 independent channels across the full 304 MHz bandwidth, so the expected noise per 18.5 kHz channel in a 2 h observation is approximately 5.5-3.5 mJy beam$^{-1}$ across the band. In the case of an actual survey, the true sensitivity will of course be recorded in the spectral data as a function of redshift (see e.g. @Allison:2015a), but for the purposes of the simulated survey presented in this work we split the band into several frequency bins to capture the variation in sensitivity and velocity resolution (which is in the range 7.8 km s$^{-1}$ at 711.5 MHz to 5.5 km s$^{-1}$ at 1015.5 MHz). In order to simulate a realistic survey of the southern sky we select all radio sources south of $\delta = +10\degr$ from catalogues of the National Radio Astronomy Observatory Very Large Array Sky Survey (NVSS, $\nu = 1.4$ GHz, $S_{\rm src} \gtrsim 2.5$ mJy; @Condon:1998), the Sydney University Molonglo SkySurvey ($\nu = 843$ MHz, $S_{\rm src} \gtrsim 10$ mJy; @Mauch:2003) and the second epoch Molonglo Galactic Plane Survey ($\nu = 843$ MHz, $S_{\rm src} \gtrsim 10$ mJy; @Murphy:2007). The source flux densities, used to calculate the optical depth limit in , are estimated at the centre of each frequency bin by extrapolating from the catalogue values and assuming a canonical spectral index of $\alpha =-0.7$. In , we show the resulting cumulative distribution of radio sources in our sample as a function of flux density across the band. For any given sight-line, the redshift interval over which absorption may be detected is dependent upon the distance to the continuum source. The lack of accurate spectroscopic redshift measurements for most radio sources over the sky necessitates the use of a statistical approach based on a model for the source redshift distribution. We therefore apply a statistical weighting to each comoving path element $\delta{X}(z)$ such that the expected number of absorber detections is now given by \[\label{equation:weighted_sum_number} \mu = \iint{f(N_{\rm HI},X)\,\mathcal{F}_{\rm src}(z^{\prime} \geq z)\,\mathrm{d}X\,\mathrm{d}N_{\rm HI}},\] where \[\mathcal{F}_{\rm src}(z^{\prime} \geq z) = {\int_{z}^{\infty} \mathcal{N}_{\rm src}(z^{\prime})\mathrm{d}z^{\prime}\over\int_{0}^{\infty}\mathcal{N}_{\rm src} (z^{\prime})\mathrm{d}z^{\prime}},\] and $\mathcal{N}_{\rm src}(z)$ is the number of radio sources as a function of redshift. To estimate $\mathcal{N}_{\rm src}(z)$ we use the Combined EIS-NVSS Survey Of Radio Sources (CENSORS; @Brookes:2008), which forms a complete sample of radio sources brighter than 7.2 mJy at 1.4 GHz with spectroscopic redshifts out to cosmological distances. In we show the distribution of CENSORS sources brighter than 10 mJy beyond a given redshift $z$, and the corresponding analytical function derived from the model fit of, given by \[\mathcal{N}_{\rm src}(z) \approx 1.29 + 32.37z-32.89z^{2} + 11.13z^{3}-1.25z^{4},\] which we use in our analysis. For the redshifts spanned by our simulated ASKAP survey, the fraction of background sources evolves from 87 per cent at $z = 0.4$ to 53 per cent at $z = 1.0$. We assume that this redshift distribution applies to any sight-line irrespective of the continuum flux density. However, this assumption is only true if the source population in the target sample evolves such that the effect of distance is nullified by an increase in luminosity. Given this criterion, and the sensitivity of our simulated survey, we limit our sample to sources with flux densities between 10 and 1000 mJy, which are dominated by the rapidly evolving population of high-excitation radio galaxies and quasars (e.g. @Jackson:1999 [@Best:2012; @Best:2014; @Pracy:2016]) and for which the redshift distribution is known to be almost independent of flux density (e.g. @Condon:1984 [@Condon:1998]). In , we show the number of sources from this sub-sample as a function of opacity sensitivity \[as defined by \], drawing random samples of the line FWHM and covering factor from the distributions shown in and . There are approximately 190 000 sightlines with sufficient sensitivity to detect absorption of optical depth greater than $\tau_{5\sigma} \approx 1.0$ and 25 000 sensitive to optical depths greater than $\tau_{5\sigma} \approx 0.1$. Since this distribution converges at optical depth sensitivities greater than $\tau_{5\sigma} \approx 5$, the population of sources fainter than 10 mJy, which are excluded from our simulated ASKAP survey, would not significantly contribute to further detections of absorption. Similarly, while sources brighter than 1 Jy are good probes of low-column-density H i gas, they do not constitute a sufficiently large enough population to significantly affect the total number of absorber detections expected in the survey and can also be safely excluded. Based on these assumptions, we can estimate the number of absorbers we would expect to detect in our survey with ASKAP as a function of spin temperature. In , we show the expected detection yield as a cumulative function of column density. We show results for two scenarios where the spin temperature is fixed at a single value of either 100 or 1000 K, and the line FWHM and covering factors are drawn from the random distributions shown in and . We find that for both these cases the expected number of detections is not sensitive to column densities below the DLA definition of $N_{\rm HI} = 2\times 10^{20}$ cm$^{-2}$. We also show in the expected total detection yield (integrated over all H i column densities) as a function of a single spin temperature $T_\mathrm{spin}$ and line FWHM $\Delta{v}_\mathrm{50}$. We find that for typical spin temperatures of a few hundred kelvin (consistent with the typical fraction of CNM observed in the local Universe) and a line FWHM of approximately $20$ km s$^{-1}$, a wide-field 21 cm survey with ASKAP is expected to yield $\sim 1000$ detections. However, even moderate evolution to a higher spin temperature in the DLA population should see significant reduction in the detection yield from this survey. # Inferring the average spin temperature {#section:spin_temp} We cannot directly measure the spin temperatures of individual systems without additional data from either 21 cm emission or Lyman-$\alpha$ absorption. However, from it is evident that the total number of absorbing systems expected to be detected with a reasonably large 21 cm survey is strongly dependent on the assumed value for the spin temperature. Therefore, by comparing the actual survey yield with that expected from the known H i distribution, we can infer the average spin temperature of the atomic gas within the DLA population for a given redshift interval. Assuming that the total number of detections follows a Poisson distribution, the probability of detecting $\mathcal{N}$ intervening absorbing systems is given by \[p(\mathcal{N}\|\overline{\mu}) = {\overline{\mu}^{\mathcal{N}} \over \mathcal{N}!} \mathrm{e}^{-\overline{\mu}},\] where $\overline{\mu}$ is the expected total number of detections given by the integral \[\overline{\mu} = \iiint{\mu(T_{\rm spin},\Delta{v}_{50},c_{\rm f})\rho(T_{\rm spin},\Delta{v}_{50},c_{\rm f})\mathrm{d}T_{\rm spin}\mathrm{d}\Delta{v}_{50}\mathrm{d}c_{\rm f}},\] and $\rho$ is the distribution of systems as a function of spin temperature, line FWHM and covering factor. We assume that all three of these variables are independent[^4] so that $\rho$ factorizes into functions of each. We then marginalise over the covering factor and line width distributions shown in and so that the expression for $\overline{\mu}$ reduces to \[\label{equation:harmonic_spin_temp} \overline{\mu} = \int{\mu(T_{\rm spin})\rho(T_{\rm spin})\mathrm{d}T_{\rm spin}} = \mu(\overline{T}_{\rm spin}),\] where $\overline{T}_{\rm spin}$ is the harmonic mean of the unknown spin temperature distribution, weighted by column density. This is analogous to the spin temperature inferred from the detection of absorption in a single intervening galaxy averaged over several gaseous components at different temperatures (e.g. @Carilli:1996). In the event of the survey yielding $\mathcal{N}$ detections, we can calculate the posterior probability density of $\overline{T}_{\rm spin}$ using the following relationship between conditional probabilities \[p(\overline{T}_{\rm spin}\|\mathcal{N}) = {p(\mathcal{N}\|\overline{T}_{\rm spin})p(\overline{T}_{\rm spin})\over p(\mathcal{N})},\] where $p(\overline{T}_{\rm spin})$ is our prior probability density for $\overline{T}_{\rm spin}$ and $p(\mathcal{N})$ is the marginal probability of the number of detections, which can be treated as a normalizing constant. The minimally informative Jeffreys prior for the mean value $\mu$ of a Poisson distribution is $1/\sqrt{\mu}$ (@Jeffreys:1946)[^5]. From it therefore follows that a suitable form for the non-informative spin temperature prior is $p(\overline{T}_{\rm spin}) = 1/\sqrt{\overline{\mu}}$, so that \[\label{equation:tspin_prob} p(\overline{T}_{\rm spin}\|\mathcal{N}) = C^{-1}\,{\overline{\mu}^{(\mathcal{N}-1/2)} \over \mathcal{N}!} \mathrm{e}^{-\overline{\mu}},\] where the distribution is normalised to unit total probability by evaluating the integral \[C = \int{{\overline{\mu}^{(\mathcal{N}-1/2)} \over \mathcal{N}!} \mathrm{e}^{-\overline{\mu}}}\,\mathrm{d}\overline{T}_{\rm spin}.\] The probabilistic relationship given by and the expected detection yield derived in can be used as a frame-work for inferring the harmonic-mean spin temperature using the results of any homogeneous 21-cm survey. We have assumed that we can accurately distinguish intervening absorbing systems from those associated with the host galaxy of the radio source. However, any 21-cm survey will be accompanied by follow-up observations, at optical and sub-mm wavelengths, which will aid identification. Furthermore, future implementation of probabilistic techniques to either use photometric redshift information or distinguish between line profiles should provide further disambiguation. Of course we have not yet accounted for any error in our estimate of $\overline{\mu}$, which will increase our uncertainty in $\overline{T}_\mathrm{spin}$. In the following section we discuss these possible sources of error and their effect on the result. # Sources of error {#section:errors} Our estimate of the expected number of 21 cm absorbers is dependent upon several distributions describing the properties of the foreground absorbing gas and the background source distribution. For future large-scale 21 cm surveys, the accuracy to which we can infer the harmonic mean of the spin temperature distribution will eventually be limited by the accuracy to which we can measure these other distributions. In this section, we describe these errors and their propagation through to the estimate of $\overline{T}_\mathrm{spin}$, summarizing our results in . ## The covering factor {#section:covering_factor} ### Deviation from a uniform distribution between 0 and 1 The fraction $c_{\rm f}$ by which the foreground gas subtends the background radiation source is difficult to measure directly and is thereby a significant source of error for 21 cm absorption surveys. In this work, we have assumed a uniform distribution for $c_{\rm f}$, taking random values between 0 and 1. In , we tested this assumption by comparing it with the distribution of flux density core fractions in a sample of 37 quasars, used by as a proxy for the covering factor. By carrying out a two-tailed KS test, we found some evidence (at the 0.05 level) that this quasar sample was inconsistent with our assumption of a uniform distribution between 0 and 1. Noticeably there seems to be an under-representation of quasars in the Kanekar et al. sample with estimated $c_{\rm f} \lesssim 0.2$. In the low optical depth limit, the detection rate is dependent on the ratio of spin temperature to covering factor, in which case a fractional deviation in $c_{\rm f}$ will propagate as an equal fractional deviation in $\overline{T}_\mathrm{spin}$. Based on the difference seen in the covering factor distribution of the Kanekar et al. sample and the uniform distribution, we assume that the spin temperature can deviate by as much as $\pm$`<!-- -->`{=html}10 per cent. ### Evolution with redshift We also consider that the covering factor distribution may evolve with redshift, which would mimic a perceived evolution in the average spin temperature. Such an effect was proposed by and, who claimed that the relative change in angular-scale behaviour of absorbers and radio sources between low-and high-redshift samples could explain the apparent evolution of the spin temperature found by. To test for this effect in their larger DLA sample, considered a sub-sample at redshifts greater than $z = 1$, for which the relative evolution of the absorber and source angular sizes should be minimal. While the significance of their result was reduced by removing the lower redshift absorbers from their sample, they still found a difference at $3.5\,\sigma$ significance between spin temperature distributions in the two DLA sub-samples separated by a median redshift of $z = 2.683$. Future surveys with ASKAP and the other SKA pathfinders will search for H i absorption at intermediate redshifts ($z \sim 1$), where the relative evolution of the absorber and source angular sizes is expected to be more significant than for the higher redshift DLA sample considered by. We therefore consider the potential effect of this cosmological evolution on the inferred value of $\overline{T}_\mathrm{spin}$. We approximate the covering factor using the following model of \[\label{equation:covering_factor} c_{f} \approx \begin{cases} \left({\theta_{\rm abs}\over \theta_{\rm src}}\right)^{2}, & \text{if}\ \theta_{\rm abs} < \theta_{\rm src}, \\ 1, & \text{otherwise}, \end{cases}\] where $\theta_{\rm abs}$ and $\theta_{\rm src}$ are the angular sizes of the absorber and background source, respectively. Under the small-angle approximation $\theta_{\rm abs} \approx {d_{\rm abs}/D_{\rm abs}}$ and $\theta_{\rm abs} \approx {d_{\rm src}/D_{\rm src}}$, where $d_{\rm abs}$ and $D_{\rm abs}$ are the linear size and angular diameter distance of the absorber, and likewise $d_{\rm src}$ and $D_{\rm src}$ are the linear size and angular diameter distance of the background source. Assuming that the ratio $d_{\rm abs}/d_{\rm src}$ is randomly distributed and independent of redshift, any evolution in the covering factor is therefore dominated by relative changes in the angular diameter distances. We calculate the expected angular diameter distance ratio at a redshift $z$ by \[\left\langle{D_{\rm abs}\over D_{\rm src}}\right\rangle_{z} = D_{\rm abs}(z){\int_{z}^{\infty} \mathcal{N}_{\rm src}(z^{\prime})D_{\rm src}(z^{\prime})^{-1}\mathrm{d}z^{\prime}\over\int_{z}^{\infty}\mathcal{N}_{\rm src} (z^{\prime})\mathrm{d}z^{\prime}},\] which, for the source redshift distribution model given by, evolves from 0.7 at $z = 0.4$ to 1.0 at $z = 1.0$ (see ). We note that this is consistent with the behaviour measured by for the total sample of DLAs observed at 21 cm wavelengths. By applying this as a correction to the otherwise uniformly distributed covering factor (using ), we find that the inferred value of $\overline{T}_{\rm spin}$ systematically increases by approximately 30 per cent. ## The $\bmath{N_{\rm HI}}$ frequency distribution ### Uncertainty in the measurement of $f(N_{\rm HI},X)$ We assume that $f(N_{\rm HI}, X)$ is relatively well understood as a function of redshift by interpolating between model gamma functions fitted to the distributions at $z = 0$ and $3$. However, these distributions were measured from finite samples of galaxies, which of course have associated uncertainties that need to be considered. In the case of the data presented by and, both have typical measurement uncertainties in $f(N_{\rm HI}, X)$ of approximately 10 per cent over the range of column densities for which our simulated ASKAP survey is sensitive (see ). This will propagate as a 10 per cent fractional error in the expected number of absorber detections, and contribute a similar percentage uncertainty in the inferred average spin temperature. ### Correcting for 21 cm self-absorption In the local Universe, showed that self-absorption from opaque H i clouds identified in high-resolution images of the Local Group galaxies M31, M33 and the Large Magellanic Cloud may necessitate a correction to the local atomic mass density of up to 30 per cent. Although it is not yet clear whether this small sample of Local Group galaxies is representative of the low-redshift population, it is useful to understand how this effect might propagate through to our average spin temperature measurement. We therefore replace the gamma-function parametrization of the local $f(N_{\rm HI})$ given by with the non-parametric values given in table 2 of, and recalculate $\overline{T}_{\rm spin}$. For an all-sky survey with the full 36-antenna ASKAP we find that $\overline{T}_{\rm spin}$ increases by $\sim$`<!-- -->`{=html}30 for 100 detections and $\sim$`<!-- -->`{=html}10 per cent for 1000 detections. Note that the correction increases for low numbers of detections, which are dominated by the highest column density systems. ### Dust obscuration bias in optically-selected DLAs At higher redshifts, it is possible that the number density of optically-selected DLAs could be significantly underestimated as a result of dust obscuration of the background quasar (@Ostriker:1984). This would cause a reduction in the $f(\mbox{H\,{\sc i}}, X)$ measured from optical surveys, thereby significantly underestimating the expected number of intervening 21 cm absorbers at high redshifts. The issue is further compounded by the expectation that the highest column density DLAs ($N_{\rm HI} \gtrsim 10^{21}$ cm$^{-2}$), for which future wide-field 21 cm surveys are most sensitive (see ), may contain more dust than their less-dense counterparts. This conclusion was supported by early analyses of the existing quasar surveys at that time (e.g. @Fall:1993), which indicated that up to 70 per cent of quasars could be missing from optical surveys through the effect of dust obscuration, albeit with large uncertainties. However, subsequent optical and infrared observations of radio-selected quasars (e.g. @Ellison:2001; @Ellison:2005; @Jorgenson:2006), which are free of the potential selection biases associated with these optical surveys, found that the severity of this issue was substantially over-estimated and that there was minimal evidence in support of a correlation between the presence of DLAs and dust reddening. Furthermore, the H i column density frequency distribution measured by was found to be consistent with the optically-determined gamma-function parametrization of, with no evidence of DLA systems missing from the SDSS sample at a sensitivity of $N_{\rm HI} \lesssim 5 \times 10^{21}$ cm$^{-2}$. Although radio-selected surveys of quasars are free of the selection biases associated with optical surveys, they do typically suffer from smaller sample sizes and are therefore less sensitive to the rarer DLAs with the highest column densities. Another approach is to directly test whether optically-selected quasars with intervening DLAs, selected from the SDSS sample, are systematically more dust reddened than a control sample of non-DLA quasars. Comparisons in the literature are based on several different colour indicators, which include the spectral index (e.g. @Murphy:2004 [@Murphy:2016]), spectral stacking (e.g. @Frank:2010 [@Khare:2012]) and direct photometry (e.g. @Vladilo:2008; @Fukugita:2015). The current status of these efforts is summarized by, showing broad support for a missing DLA population at the level of $\sim$`<!-- -->`{=html}5 per cent but highlighting that tension still exists between different dust measurements. No substantial evidence has yet been found to support a correlation between the dust reddening and H i column density in these optically selected DLA surveys (e.g. @Vladilo:2008 [@Khare:2012; @Murphy:2016]). In an attempt to reconcile the differences and myriad biases associated with these techniques, carried out a statistically-robust meta-analysis of the available optical and radio data, using a Bayesian parameter estimation approach to model the dust as a function of column density and metallicity. They found that the expected fraction of DLAs missing from optical surveys is 7 per cent, with fewer than 28 per cent missing at 3 $\sigma$ confidence. Based on this body of work we therefore assume that approximately 10 per cent of DLAs are missing from the SDSS sample of and consider the affect on our estimate of $\overline{T}_{\rm spin}$. We further assume that there is no dependance on column density, an assumption which is supported by the aforementioned observational data for the range of column densities to which our 21 cm survey is sensitive. We find that increasing the high-redshift column density frequency distribution by 10 per cent introduces a systematic increase of approximately 3 per cent in the expected number of detections for the redshifts covered by our ASKAP surveys. We note that this error will increase significantly for 21 cm surveys at higher redshifts where the optically derived $f(\mbox{H\,{\sc i}}, X)$ dominates the calculation of the expected detection rate. ## The radio source background As described in , we weight the comoving path-length for each sight-line by a statistical redshift distribution in order to account for evolution in the radio source background. We use the parametric model of, which is derived from fitting the measured redshifts of for CENSORS sources brighter than 10 mJy, and assume that this applies to all sources in the range 10-1000 mJy. In , we show the cumulative distribution of sources located behind a given redshift and the associated measurement uncertainty given by the errorbars. For the intermediate redshifts covered by the ASKAP survey, the fractional uncertainty in this distribution increases from $\sigma_{\mathcal{F}_{\rm src}}/\mathcal{F}_{\rm src} \approx 3.5$ to 8 per cent between $z = 0.4$ and 1.0, which propagates through to a similar fractional uncertainty in $\overline{T}_{\rm spin}$. However, for higher redshifts this fractional uncertainty increases rapidly at $z > 2$, to more than 50 per cent at $z = 3$, reflecting the paucity of optical spectroscopic data for the high-redshift radio source population. Understanding how the radio source population is distributed at lower flux densities and at higher redshifts is therefore a concern for the future 21 cm absorption surveys undertaken with the SKA mid-and low-frequency telescopes (see @Kanekar:2004 and @Morganti:2015 for reviews). ::: # Expected results for future 21-cm absorption surveys {#section:tspin_results} In the top panel of we show the results of applying our method for inferring $\overline{T}_{\rm spin}$ to the simulated all-southern-sky H i absorption survey with ASKAP described in . We account for the uncertainties in the expected detection rate $\overline{\mu}$, discussed in , by using a Monte Carlo approach and marginalizing over many realizations. A yield of 1000 absorbers from such a survey would imply an average spin temperature of $\overline{T}_\mathrm{spin} = 127^{+14}_{-14}\,(193^{+23}_{-23})$ K[^6], where values in parentheses denote the alternative posterior probability resulting from the systematic errors discussed in . This scenario would indicate that a large fraction of the atomic gas in DLAs at these intermediate redshifts is in the classical stable CNM phase. Conversely, a yield of only 100 detections would imply that $\overline{T}_\mathrm{spin} = 679^{+64}_{-65}\,(1184^{+116}_{-120})$ K, indicating that less than 10 per cent of the atomic gas is in the CNM and that the bulk of the neutral gas in galaxies is significantly different at intermediate redshifts compared with the local Universe. We also consider the effect of reducing the sky area and array size, which is relevant for planned early science surveys with ASKAP and other SKA pathfinder telescopes. In the bottom panel of , we show the spin temperatures inferred when observing a random 1000 deg$^{2}$ field for 12 h per pointing, between $z_{\rm HI} = 0.4$ and $1.0$, using a 12-antenna version of ASKAP. We find that detection yields of 30 and 3 from such a survey would give inferred spin temperatures of $\overline{T}_{\rm spin} =134^{+23}_{-27}\,(209^{+40}_{-47})$ and $848^{+270}_{-430}\,(1535^{+513}_{-837})$ K, respectively. The significant reduction in telescope sensitivity and sky-area, compensated by the increase in integration time per pointing planned for early-science, results in a factor of 30 decrease in the expected number of detections and therefore an increase in the sample variance and uncertainty in $\overline{T}_{\rm spin}$. However, this result demonstrates that we expect to be able to distinguish between the limiting cases of CNM-rich or deficient DLA populations even during the early-science phases of the SKA pathfinders. For example 30 detections with the early ASKAP survey rules out an average spin temperature of 1000 K at high probability. # Conclusions We have demonstrated a statistical method for measuring the average spin temperature of the neutral ISM in distant galaxies, using the expected detection yields from future wide-field 21 cm absorption surveys. The spin temperature is a crucial property of the ISM that can be used to determine the fraction of the cold ($T_{\rm k} \sim 100$ K) and dense ($n \sim 100$ cm$^{-2}$) atomic gas that provides sites for the future formation of cold molecular gas clouds and star formation. Recent 21 cm surveys for H i absorption in Mg ii absorbers and DLAs towards distant quasars have yielded some evidence of an evolution in the average spin temperature that might reveal a decrease in the fraction of cold dense atomic gas at high redshift (e.g. @Gupta:2009 [@Kanekar:2014a]). By combining recent specifications for ASKAP, with available information for the population of background radio sources, we show that strong statistical constraints (approximately $\pm10$ per cent) in the average spin temperature can be achieved by carrying out a shallow 2-h per pointing survey of the southern sky between redshifts of $z = 0.4$ and $1.0$. However, we find that the accuracy to which we can measure the average spin temperature is ultimately limited by the accuracy to which we can measure the distribution of the covering factor, the $N_{\rm HI}$ frequency distribution function and the evolution of the radio source population as a function of redshift. By improving our understanding of these distributions we will be able to leverage the order-of-magnitude increases in sensitivity and redshift coverage of the future SKA telescope, allowing us to measure the evolution of the average spin temperature to much higher redshifts.	{'timestamp': '2016-08-09T02:04:29', 'yymm': '1607', 'arxiv_id': '1607.04828', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04828'}
# Introduction Given $\rho>0$, we consider the problem \[\label{eq:main_prob_U} \begin{cases}-\Delta U + \lambda U = \|U\|^{p-1}U & \text{in }\Omega,\smallskip\\ \int_\Omega U^2\,dx = \rho, \quad U=0 & \text{on }\partial\Omega, \end{cases}\] where $\Omega\subset{\mathbb{R}}^N$ is a Lipschitz, bounded domain, $1<p<2^-1$, $\rho>0$ is a fixed parameter, and both $U\in H^1_0(\Omega)$ and $\lambda\in{\mathbb{R}}$ are unknown. More precisely, we investigate conditions on $p$ and $\rho$ (and also $\Omega$) for the solvability of the problem. The main interest in [\[eq:main_prob_U\]](#eq:main_prob_U){reference-type="eqref" reference="eq:main_prob_U"} relies on the investigation of standing wave solutions for the nonlinear Schrödinger equation \[i\frac{\partial \Phi}{\partial t}+\Delta \Phi+ \|\Phi\|^{p-1}\Phi=0,\qquad (t,x)\in {\mathbb{R}}\times \Omega\] with Dirichlet boundary conditions on $\partial\Omega$. This equation appears in several different physical models, both in the case $\Omega={\mathbb{R}}^N$, and on bounded domains. In particular, the latter case appears in nonlinear optics and in the theory of Bose-Einstein condensation, also as a limiting case of the equation on ${\mathbb{R}}^N$ with confining potential. When searching for solutions having the wave function $\Phi$ factorized as $\Phi(x,t)=e^{i\lambda t} U(x)$, one obtains that the real valued function $U$ must solve \[\label{eq:NLS}-\Delta U + \lambda U = \|U\|^{p-1}U ,\qquad U\in H^1_0(\Omega),\] and two points of view are available. The first possibility is to assign the chemical potential $\lambda\in{\mathbb{R}}$, and search for solutions of [\[eq:NLS\]](#eq:NLS){reference-type="eqref" reference="eq:NLS"} as critical points of the related action functional. The literature concerning this approach is huge and we do not even make an attempt to summarize it here. On the contrary, we focus on the second possibility, which consists in considering $\lambda$ as part of the unknown and prescribing the mass (or charge) $\\|U\\|_{L^2(\Omega)}^2$ as a natural additional condition. Up to our knowledge, the only previous paper dealing with this case, in bounded domains, is, which we describe below. The problem of searching for normalized solutions in ${\mathbb{R}}^N$, with non-homogeneous nonlinearities, is more investigated, even though the methods used there can not be easily extended to bounded domains, where dilations are not allowed. Very recently, also the case of partial confinement has been considered. Solutions of [\[eq:main_prob_U\]](#eq:main_prob_U){reference-type="eqref" reference="eq:main_prob_U"} can be identified with critical points of the associated energy functional \[\mathcal{E}(U) = \frac12\int_\Omega\|\nabla U\|^2\,dx-\frac{1}{p+1} \int_\Omega\|U\|^{p+1}\,dx\] restricted to the mass constraint \[{\mathcal{M}}_\rho=\{U\in H_0^1(\Omega): \\|U\\|_{L^2(\Omega)}=\rho\},\] with $\lambda$ playing the role of a Lagrange multiplier. A cricial role in the discussion of the above problem is played by the Gagliardo-Nirenberg inequality: for any $\Omega$ and for any $v\in H^1_0(\Omega)$, \[\label{sobest} \\|v\\|^{p+1}_{L^{p+1}(\Omega)} \leq C_{N,p} \\| \nabla v \\|_{L^2(\Omega)}^{N(p-1)/2} \\| v \\|_{L^2(\Omega)} ^{(p+1)-N(p-1)/2},\] the equality holding only when $\Omega={\mathbb{R}}^N$ and $v=Z_{N,p}$, the positive solution of $-\Delta Z + Z = Z^{p}$ (which is unique up to translations ). Accordingly, the exponent $p$ can be classified in relation with the so called $L^2$-critical exponent* $1+4/N$ (throughout all the paper, $p$ will be always Sobolev-subcritical and its criticality will be understood in the $L^2$ sense). Indeed we have that ${\mathcal{E}}$ is bounded below and coercive on ${\mathcal{M}}_\rho$ if and only if either $p$ is subcritical, or it is critical and $\rho$ is sufficiently small. The recent paper deals with problem [\[eq:main_prob_U\]](#eq:main_prob_U){reference-type="eqref" reference="eq:main_prob_U"} in the case of the spherical domain $\Omega = B_1$, when searching for positive solutions $U$. In particular, it is shown that the solvability of [\[eq:main_prob_U\]](#eq:main_prob_U){reference-type="eqref" reference="eq:main_prob_U"} is strongly influenced by the exponent $p$, indeed: - in the subcritical case $1<p<1+4/N$, [\[eq:main_prob_U\]](#eq:main_prob_U){reference-type="eqref" reference="eq:main_prob_U"} admits a unique positive solution for every $\rho>0$; - if $p=1+4/N$ then [\[eq:main_prob_U\]](#eq:main_prob_U){reference-type="eqref" reference="eq:main_prob_U"} admits a unique positive solution for \[0<\rho<\rho^=\left(\frac{p+1}{2C_{N,p}}\right)^{N/2}=\\|Z_{N,p}\\|^2_{L^2({\mathbb{R}}^N)},\] and no positive solutions for $\rho\geq\rho^$; - finally, in the supercritical regime $1+4/N<p<2^-1$, [\[eq:main_prob_U\]](#eq:main_prob_U){reference-type="eqref" reference="eq:main_prob_U"} admits positive solutions if and only if $0<\rho\leq\rho^$ (the threshold $\rho^$ depending on $p$), and such solutions are at least two for $\rho<\rho^$. In this paper we carry on such analysis, dealing with a general domain $\Omega$ and with solutions which are not necessarily positive. More precisely, let us recall that for any $U$ solving [\[eq:main_prob_U\]](#eq:main_prob_U){reference-type="eqref" reference="eq:main_prob_U"} for some $\lambda$, it is well-defined the Morse index \[m(U) = \max\left\{k: \begin{array}{l} \exists V\subset H^1_0(\Omega),\,\dim(V)= k:\forall v\in V\setminus\{0\}\smallskip\\ \displaystyle\int_\Omega \|\nabla v\|^2 + \lambda v^2-p\|U\|^{p-1}v^2\,dx<0 \end{array} \right\}\in{\mathbb{N}}.\] Then, if $\Omega=B_1$, it is well known that a solution $U$ of [\[eq:main_prob_U\]](#eq:main_prob_U){reference-type="eqref" reference="eq:main_prob_U"} is positive if and only if $m(U)=1$. Under this perspective, the results in can be read in terms of Morse index one--solutions, rather than positive ones: introducing the sets of admissible masses \[{\mathfrak{A}}_k ={\mathfrak{A}}_k(p,\Omega) := \left\{\rho>0: \begin{array}{l} \eqref{eq:main_prob_U} \text{ admits a solution $U$ (for some $\lambda$)}\\ \text{having Morse index }m(U)\leq k \end{array} \right\},\] then implies that ${\mathfrak{A}}_1(p,B_1)$ is a bounded interval if and only if $p$ is critical or supercritical, while ${\mathfrak{A}}_1(p,B_1)={\mathbb{R}}^+$ in the subcritical case. On the contrary, when considering general domains and higher Morse index, the situation may become much more complicated. We collect some examples in the following remark. Motivated by the previous remark, the first question we address in this paper concerns the boundedness of ${\mathfrak{A}}_k$. We provide the following complete classification. The proof of such result, which is outlined in Section [2](#sec:blow-up){reference-type="ref" reference="sec:blow-up"}, is obtained by a detailed blow-up analysis of sequences of solutions with bounded Morse index, via suitable a priori pointwise estimates (see ). In this respect, the regularity assumption on $\partial\Omega$ simplifies the treatment of possible concentration phenomena towards the boundary. The argument, which holds for solutions which possibly change sign, is inspired by, where the case of positive solutions is treated. Once Theorem [\[thm:bbd_index\]](#thm:bbd_index){reference-type="ref" reference="thm:bbd_index"} is established, in case $p\geq 1 + 4/N$ two questions arise, namely: 1. is it possible to provide lower bounds for $\sup{\mathfrak{A}}_k$? Is it true that $\sup{\mathfrak{A}}_k$ is strictly increasing in $k$, or, at least, that $\sup{\mathfrak{A}}_k > \sup{\mathfrak{A}}_1$ for some $k$? 2. is [\[eq:main_prob_U\]](#eq:main_prob_U){reference-type="eqref" reference="eq:main_prob_U"} solvable for every $\rho\in(0,\sup{\mathfrak{A}}_k)$, or at least can we characterize some subinterval of solvability? It is clear that both issues can be addressed by characterizing values of $\rho$ for which existence (and multiplicity) of solutions with bounded Morse index can be guaranteed. To this aim, it can be useful to restate problem [\[eq:main_prob_U\]](#eq:main_prob_U){reference-type="eqref" reference="eq:main_prob_U"} as \[\label{eq:main_prob_u} \begin{cases}-\Delta u + \lambda u = \mu\|u\|^{p-1}u & \text{in }\Omega,\\ \int_\Omega u^2\,dx = 1, \quad u=0 & \text{on }\partial\Omega, \end{cases} \qquad\text{where}\quad \begin{cases} U=\sqrt{\rho} u\\ \mu = \rho^{(p-1)/2}, \end{cases}\] where now $\mu>0$ is prescribed. Since \[\label{Emu} \text{both } \mathcal{E}_{\mu}(u) := \frac{1}{2}\int_{\Omega}\|\nabla u\|^2-\frac{\mu}{p+1}\int_{\Omega}\| u\|^{p+1} \qquad \text{and }{\mathcal{M}}={\mathcal{M}}_1=\{u: \\|u\\|_{L^2(\Omega)}=1\}\] are invariant under the ${\mathbb{Z}}_2$-action of the involution $u\mapsto-u$, solutions of [\[eq:main_prob_u\]](#eq:main_prob_u){reference-type="eqref" reference="eq:main_prob_u"} can be found via min-max principles in the framework of index theories (see e.g. ). Notice that in the supercritical case ${\mathcal{E}}_\mu$ is not bounded from below on ${\mathcal{M}}$. Following, it can be convenient to parameterize solutions to [\[eq:main_prob_u\]](#eq:main_prob_u){reference-type="eqref" reference="eq:main_prob_u"} with respect to the $H^1_0$-norm, therefore we introduce the sets \[\label{eq:defBU} \mathcal{B}_\alpha:=\left\{u\in {\mathcal{M}}:\,\int_\Omega \|\nabla u\|^2\,dx<\alpha\right\},\quad\quad \mathcal{U}_\alpha:=\left\{u\in {\mathcal{M}}:\,\int_\Omega \|\nabla u\|^2\,dx=\alpha\right\}.\] Introducing the first Dirichlet eigenvalue of $-\Delta$ in $H^1_0(\Omega)$, $\lambda_1(\Omega)$, we have that the sets above are non-empty whenever $\alpha> \lambda_1(\Omega)$. Since we are interested in critical points having Morse index bounded from above, following we introduce the following notion of genus. We remark that this notion of genus is different from the classical one of Krasnoselskii genus, which is well suited for estimates of the Morse index from below, rather than above. Nonetheless, $\gamma$ shares with the Krasnoselskii genus most of the main properties of an index. In particular, by the Borsuk-Ulam Theorem, any set $A$ homeomorphic to the sphere ${\mathbb{S}}^{m-1} := \partial B_1 \subset {\mathbb{R}}^m$ has genus $\gamma(A) = m$. Furthermore, we show in Section [3](#sec:2const){reference-type="ref" reference="sec:2const"} that $\Sigma^{(k)}_{\alpha}$ is not empty, provided $\alpha>\lambda_k(\Omega)$ (the $k$-th Dirichlet eigenvalue of $-\Delta$ in $H^1_0(\Omega)$). Equipped with this notion of genus we provide two different variational principles for solutions of [\[eq:main_prob_u\]](#eq:main_prob_u){reference-type="eqref" reference="eq:main_prob_u"} (and thus of [\[eq:main_prob_U\]](#eq:main_prob_U){reference-type="eqref" reference="eq:main_prob_U"}). The first one is based on a variational problem with two constraints, which was exploited as the main tool in proving the results in. As a matter of fact, the results in were obtained by a detailed analysis of the map $\alpha \mapsto \mu_\alpha$ in the case $k=1$, i.e. when dealing with \[M_{\alpha,1} = \max\left\{\\|u\\|_{L^{p+1}}^{p+1}: \\|u\\|_{L^2}^2=1,\,\\|\nabla u\\|_{L^2}^2=\alpha \right\}.\] In the present paper we do not investigate the properties of the map $\alpha \mapsto \mu_\alpha$ for general $k$, but we rather prefer to exploit the characterization of $M_{\alpha,k}$ in connection with a second variational principle, which deals with only one constraint. The link between Theorem [\[thm:genus_2constr\]](#thm:genus_2constr){reference-type="ref" reference="thm:genus_2constr"} and Theorem [\[thm:genus_1constr\]](#thm:genus_1constr){reference-type="ref" reference="thm:genus_1constr"} is that we can provide explicit estimates of $\hat \mu_k$ (and hence of $\hat\rho_k$) in terms of the map $\alpha\mapsto M_{\alpha,k}$ (see Section [4](#sec:1const){reference-type="ref" reference="sec:1const"}). We stress that the above results hold for any Lipschitz $\Omega$. As a first consequence, this allows to extend the existence result in to non-radial domains. We observe that the exponent of $\lambda_1(\Omega)$ in the supercritical threshold is negative, therefore such threshold decreases with the size of $\Omega$. Once the first thresholds have been estimated, we turn to the higher ones: by exploiting the relations between $M_{\alpha,k}$ and $c_k$, we can show that the thresholds obtained for Morse index one--solutions in Theorem [\[thm:intro_GS\]](#thm:intro_GS){reference-type="ref" reference="thm:intro_GS"} can be increased, by considering higher Morse index--solutions, at least for some exponent. Beyond existence results for [\[eq:main_prob_U\]](#eq:main_prob_U){reference-type="eqref" reference="eq:main_prob_U"}, also multiplicity results can be achieved. A first general consideration, with this respect, is that Theorem [\[thm:genus_1constr\]](#thm:genus_1constr){reference-type="ref" reference="thm:genus_1constr"} holds true also when using the standard Krasnoselskii genus instead of $\gamma$; this allows to obtain critical points having Morse index bounded from below (see ), and therefore to obtain infinitely many solutions, at least when $\rho$ is less than some threshold. More specifically, we can also prove the existence of a second solution in the supercritical case, thus extending to any $\Omega$ the multiplicity result obtained in for the ball. Indeed, on the one hand, in the supercritical case ${\mathcal{E}}_\mu$ is unbounded from below; on the other hand the solution obtained in Theorem [\[thm:genus_1constr\]](#thm:genus_1constr){reference-type="ref" reference="thm:genus_1constr"}, for $k=1$, is a local minimum. Thus the Mountain Pass Theorem applies on $\mathcal{M}$, and a second solution can be found for $\mu<\hat\mu_1$, see Proposition [\[mpcritlev\]](#mpcritlev){reference-type="ref" reference="mpcritlev"} for further details (and also Remark [\[rem:further_crit_lev\]](#rem:further_crit_lev){reference-type="ref" reference="rem:further_crit_lev"} for an analogous construction for $k\ge2$). To conclude this introduction, let us mention that the explicit lower bounds obtained in Theorem [\[thm:intro_GS\]](#thm:intro_GS){reference-type="ref" reference="thm:intro_GS"} can be easily applied in order to gain much more information also in the case of special domains, as those considered in Remark [\[rem:specialdomains\]](#rem:specialdomains){reference-type="ref" reference="rem:specialdomains"}. For instance, we can prove then following. Therefore our starting problem in $\Omega=B$ can be solved for any mass value also in the critical and supercritical regime, at least for $p$ smaller than this further critical exponent $1+4/(N-1) > 1+ 4/N$. Of course, higher masses require higher Morse index--solutions. In particular, since by we know that ${\mathfrak{A}}_1(B,1+4/N) = (0,\\|Z_{N,p}\\|_{L^2})$, we have that for larger masses, even though no positive solution exists, nodal solutions with higher Morse index can be obtained: in such cases [\[eq:main_prob_U\]](#eq:main_prob_U){reference-type="eqref" reference="eq:main_prob_U"} admits nodal ground states with higher Morse index. The paper is structured as follows: in Section [2](#sec:blow-up){reference-type="ref" reference="sec:blow-up"} we perform a blow-up analysis of solutions with bounded Morse index, in order to prove Theorem [\[thm:bbd_index\]](#thm:bbd_index){reference-type="ref" reference="thm:bbd_index"}; Section [3](#sec:2const){reference-type="ref" reference="sec:2const"} is devoted to the analysis of the variational problem with two constraints [\[maxmin\]](#maxmin){reference-type="eqref" reference="maxmin"} and to the proof of Theorem [\[thm:genus_2constr\]](#thm:genus_2constr){reference-type="ref" reference="thm:genus_2constr"}; that of Theorems [\[thm:genus_1constr\]](#thm:genus_1constr){reference-type="ref" reference="thm:genus_1constr"}, [\[thm:intro_GS\]](#thm:intro_GS){reference-type="ref" reference="thm:intro_GS"} and Proposition [\[thm:intro_3\>1\]](#thm:intro_3>1){reference-type="ref" reference="thm:intro_3>1"} is developed in Section [4](#sec:1const){reference-type="ref" reference="sec:1const"}, by means of the variational problem with one constraint [\[infsuplev\]](#infsuplev){reference-type="eqref" reference="infsuplev"}; finally, Section [5](#sec:symm){reference-type="ref" reference="sec:symm"} contains the proof of Theorem [\[pro:symm\]](#pro:symm){reference-type="ref" reference="pro:symm"}. Notation. We use the standard notation $\{\varphi_k\}_{k\geq1}$ for a basis of eigenfunctions of the Dirichlet laplacian in $\Omega$, orthogonal in $H^1_0(\Omega)$ and orthonormal in $L^2(\Omega)$. Such functions are ordered in such a way that the corresponding eigenvalues $\lambda_k(\Omega)$ satisfy \[0<\lambda_1(\Omega)<\lambda_2(\Omega)\leq\lambda_3(\Omega)\leq\dots,\] and $\varphi_1$ is chosen to be positive on $\Omega$. $C_{N,p}$ denotes the universal constant in the Gagliardo-Nirenberg inequality [\[sobest\]](#sobest){reference-type="eqref" reference="sobest"}, which is achieved (uniquely, up to translations and dilations) by the positive, radially symmetric function $Z_{N,p}\in H^1({\mathbb{R}}^N)$, with \[\\|Z_{N,p}\\|^2_{L^2({\mathbb{R}}^N)}=\left(\frac{p+1}{2C_{N,p}}\right)^{N/2}.\] Finally, $C$ denotes every (positive) constant we need not to specify, whose value may change also within the same formula. # Blow-up analysis of solutions with bounded Morse index {#sec:blow-up} Throughout this section we will deal with a sequence $\{(u_n,\mu_n,\lambda_n)\}_n \subset H^1_0(\Omega)\times{\mathbb{R}}^+\times{\mathbb{R}}$ satisfying \[\label{eq:auxiliary_n}-\Delta u_n+\lambda_n u_n=\mu_n \|u_n\|^{p-1}u_n,\qquad\int_\Omega u_n^2\, dx=1,\qquad \int_\Omega \|\nabla u_n\|^2\, dx=:\alpha_n.\] To start with, we recall the following result (actually, in, the result is stated for positive solution, but the proof does not require such assumption). Next we turn to the study of sequences having arbitrarily large $H^1_0$-norm. In particular, we will focus on sequences of solutions having a common upper bound on the Morse index \[m(u_n) = \max\left\{k: \begin{array}{l} \exists V\subset H^1_0(\Omega),\,\dim(V)= k:\forall v\in V\setminus\{0\}\smallskip\\ \displaystyle\int_\Omega \|\nabla v\|^2 + \lambda_n v^2-p\mu_n\|u_n\|^{p-1}v^2\,dx<0 \end{array} \right\}.\] Throughout this section we will assume that \[\label{eq:mainass_secMorse} \text{the sequence }\{(u_n,\mu_n,\lambda_n)\}_n\text{ satisfies \eqref{eq:auxiliary_n}, with }\alpha_n\to+\infty\text{ and }m(u_n)\leq \bar k,\] for some $\bar k\in{\mathbb{N}}$ not depending on $n$. The local description of the asymptotic behaviour of the solutions $U_n$ to [\[equn\]](#equn){reference-type="eqref" reference="equn"} with bounded Morse index can be carried out more conveniently by defining the sequence (see ) \[\label{defVn} V_n(y)=\varepsilon_n^{\frac{2}{p-1}}\, U_n(\varepsilon_n\,y+P_n),\quad y\in \Omega_n :=\frac{\Omega-P_n}{\varepsilon_n},\] where $P_n$ is defined before [\[tildepsn\]](#tildepsn){reference-type="eqref" reference="tildepsn"}, and $\varepsilon_n=\frac{1}{\sqrt{\lambda_n}}\to 0$. Then, $V_n$ satisfies \[\nonumber \left\{ \begin{array}{ll} -\Delta V_n+ V_n=\| V_n\|^{p-1} V_n, & \hbox{in}\,\,\Omega_n;\\ \|V_n\|\le \| V_n(0)\|=\big ({\varepsilon_n/\tilde\varepsilon_n}\big )^{\frac{2}{p-1}}\rightarrow \tilde\lambda^{-\frac{1}{p-1}}, & \hbox{in}\,\, \Omega_n;\\ V_n=0, & \hbox{on}\,\, \partial\Omega_n. \end{array} \right.\] As before, we have (up to a subsequence) $V_n\rightarrow V$ in $\mathcal{C}^1_{\mathrm{loc}}(\overline H)$ where $H$ is either ${\mathbb{R}}^N$ or a half space and $V$ solves \[\label{limprob2} \left\{ \begin{array}{ll} -\Delta V+ V=\| V\|^{p-1} V, & \hbox{in}\,\, H;\\ \|V\|\le \| V(0)\|=\tilde\lambda^{-\frac{1}{p-1}}, & \hbox{in}\,\, H;\\ V=0, & \hbox{on}\,\, \partial H. \end{array} \right.\] By recalling the discussion following [\[limprob1\]](#limprob1){reference-type="eqref" reference="limprob1"} we also have $m(V)<+\infty$. We collect some well known property of such a $V$ in the following result. Following the same pattern as in, we now analyze the global behaviour of a sequence $\{U_n\}$ of solutions to [\[equn\]](#equn){reference-type="eqref" reference="equn"} for $\lambda_n\to +\infty$, assuming that $\lim_{n\to +\infty} m(U_n)\leq\bar k<\infty.$ By the previous discussion, if $P^1_n$ is a sequence of points such that $\|U_n(P^1_n)\|=\\|U_n\\|_{L^{\infty}(\Omega)}$, we have $\|U_n(P^1_n)\|\rightarrow +\infty$ and ${\lambda_n}\,d(P^1_n,\partial\Omega)^2\rightarrow +\infty$. We now look for other possible sequences of (local) extremum points $P^i_n$, $i=2,3..$, along which $\|U_n\|$ goes to infinity. For any $R>0$, consider the quantity \[\nonumber h_1(R)=\limsup_{n\to +\infty} \Bigl (\lambda_n^{-\frac{1}{p-1}}\max_{\|x-P^1_n\|\ge R\,\lambda_n^{-1/2}} \|U_n(x)\| \Bigr ).\] We will prove that if $h_1(R)$ is not vanishing for large $R$, then there exists a 'blow-up' sequence $P^2_n$ for $u_n$, 'disjoint' from $P^1_n$. Indeed, let us suppose that \[\nonumber \limsup_{R\to +\infty} h_1(R)=4\delta>0.\] Hence, up to a subsequence and for arbitrarily large $R$, we have \[\label{ass1} \lambda_n^{-\frac{1}{p-1}}\max_{\|x-P^1_n\|\ge R\,\lambda_n^{-1/2}} \|U_n(x)\| \ge 2\delta.\] Since $U_n$ vanishes on $\partial\Omega$, there exists $P^2_n\in\Omega\backslash B_{R\,\lambda_n^{-1/2}}(P_n^1)$ such that \[\label{pn2} \|U_n(P_n^2)\|=\max_{\|x-P^1_n\|\ge R\,\lambda_n^{-1/2}}\|U_n(x)\|.\] Clearly, assumption [\[ass1\]](#ass1){reference-type="eqref" reference="ass1"} implies that $\|U_n(P_n^2)\|\rightarrow +\infty$. We first prove that the sequences $P_n^1$ and $P_n^2$ are far away each other. Furthermore, we also have that the blow-up points stay far away from the boundary. We can now iterate the previous arguments: let us define, for $k\ge 1$, \[\label{defhn} h_k(R)=\limsup_{n\to +\infty} \Bigl (\lambda_n^{-\frac{1}{p-1}}\max_{d_{n,k}(x)\ge R\,\lambda_n^{-1/2}} \|U_n(x)\| \Bigr ),\] where \[d_{n,k}(x): =\min\{\|x-P^i_n\|\,:\, i=1...,k\}\] and the sequences $P^i_n$ are such that \[\nonumber \sqrt{\lambda_n}\,d(P^i_n,\partial\Omega)\rightarrow +\infty;\quad \lambda_n^{1/2}\|P_n^i-P^j_n\|\rightarrow +\infty,\quad\quad i,j=1...,k,\quad i\neq j\] as $n\to +\infty$. Assume that \[\limsup_{n\to +\infty} h_k(R)=4\delta>0.\] As before, up to a subsequence and for arbitrarily large $R$, we have \[\label{assk} \lambda_n^{-\frac{1}{p-1}}\max_{d_{n,k}(x)\ge R\,\lambda_n^{-1/2}} \|U_n(x)\| \ge 2\delta\] and there exist $P^{k+1}_n$ so that \[\nonumber \|U_n(P_n^{k+1})\|=\max_{d_{n,k}(x)\ge R\,\lambda_n^{-1/2}}\|U_n(x)\|\] with $\lim_{n\to +\infty}\|U_n(P_n^{k+1})\|=+\infty$. Moreover, as in Lemma [\[disj\]](#disj){reference-type="ref" reference="disj"} we deduce that, for every $i=1...,k$ \[\label{limblowseqi} \lambda_n^{-\frac{1}{p-1}}\,U_n(\lambda_n^{-1/2}\,y+P^i_n): = V^i_n(y)\rightarrow V^i(y)\quad \textrm{in} \,\, \mathcal{C}^1_{{\mathrm{loc}}}({\mathbb{R}}^N)\] as $n\to +\infty$; hence, by [\[assk\]](#assk){reference-type="eqref" reference="assk"} and again from the vanishing of $V$ at infinity, we conclude that \[\lambda_n^{1/2}\|P_n^{k+1}-P^i_n\|\rightarrow +\infty\] as $n\to \infty$, for every $i=1...,k$. Setting now \[\nonumber \tilde\varepsilon^{k+1}_n: =\|U_n(P^{k+1}_n)\|^{-\frac{p-1}{2}}\quad \mathrm{and} \quad R_n^{(k+1)}:=\frac{1}{2}\,\frac{d_{n,k}(P^{k+1}_n)}{\tilde\varepsilon^{k+1}_n}\] we still have $\tilde\varepsilon^{k+1}_n\to 0$ and, by [\[assk\]](#assk){reference-type="eqref" reference="assk"}, $R_n^{(k+1)} \to +\infty$ as $n\to \infty$ (see Lemma [\[distbd\]](#distbd){reference-type="ref" reference="distbd"}). Then, by the same arguments as in Lemma [\[distbd\]](#distbd){reference-type="ref" reference="distbd"}, we get \[\label{maxpkn} \|U_n(P_n^{k+1})\|=\max_{\Omega\cap B_{R^{(k+1)}_n\tilde\varepsilon^{k+1}_n}(P^{k+1}_n)} \|u_n\|\,,\] and furthermore \[\lim_{n\to +\infty}\tilde\varepsilon_n^{k+1}\,\sqrt{\lambda_n}>0\,,\] so that by defining $R_n=:R^{(k+1)}_n\tilde\varepsilon_n^{k+1}\,\sqrt{\lambda_n}\rightarrow +\infty$ we have \[\label{maxpknball} \|U_n(P_n^{k+1})\|=\max_{\Omega\cap B_{R_n\lambda^{-1/2}_n}(P^{k+1}_n)}\| U_n\|.\] Now, by the same arguments as in, it turns out that the iterative procedure must stop after at most $\bar k-1$ steps, where $\bar k =\lim_{n\to +\infty} m(u_n)$. Thus, we have proved: We now show that the sequence $U_n$ decays exponentially away from the blow-up points. We now exploit the previous results to show that suitable rescalings of the solutions to [\[eq:auxiliary_n\]](#eq:auxiliary_n){reference-type="eqref" reference="eq:auxiliary_n"} converge (locally) to some bounded solution $V$ of \[\label{eqV}-\Delta V+ V=\| V\|^{p-1} V\] in ${\mathbb{R}}^N$. The previous lemma allows us to gain some information on the asymptotic behavior of the sequences $\lambda_n$, $\mu_n$ and $\\|u_n\\|_{L^{p+1}(\Omega)}$. We first provide some bounds for the solutions of the limit problem [\[eqV\]](#eqV){reference-type="eqref" reference="eqV"} which will be useful in the sequel. # Max-min principles with two constraints {#sec:2const} In this section we deal with the maximization problem with two constraints introduced in, aiming at considering more general max-min classes of critical points. Let ${\mathcal{M}}$ be defined in [\[Emu\]](#Emu){reference-type="eqref" reference="Emu"} and, for any fixed $\alpha>\lambda_1(\Omega)$, let $\mathcal{B}_\alpha$, $\mathcal{U}_\alpha$ be defined as in [\[eq:defBU\]](#eq:defBU){reference-type="eqref" reference="eq:defBU"}. We will look for critical points of the $\mathcal{C}^2$ functional \[f(u)=\int_{\Omega}\|u\|^{p+1},\quad\quad\quad u\in {\mathcal{M}},\] constrained to $\mathcal{U}_\alpha$. To start with, we notice that the topological properties of such set depend on $\alpha$. Of course ${\mathcal{U}}_\alpha$ is closed and odd, for any $\alpha$. Recalling Definition [\[def:genus\]](#def:genus){reference-type="ref" reference="def:genus"} we deduce that its genus $\gamma({\mathcal{U}}_\alpha)$ is well defined. Now we turn to the properties of the functional $f$. To start with, it satisfies the Palais-Smale (P.S. for short) condition on $\overline{\mathcal{B}}_{\alpha}$; more precisely, the following holds. We can combine the previous lemmas to prove one of the main results stated in the introduction. We conclude this section with the following estimate. # Min-max principles on the unit sphere in $L^2$ {#sec:1const} According to equation [\[Emu\]](#Emu){reference-type="eqref" reference="Emu"}, let ${\mathcal{M}}\subset H^1_0(\Omega)$ denote the unit sphere with respect to the $L^2$ norm and $\mathcal{E}_{\mu}$ the energy functional associated to [\[eq:main_prob_u\]](#eq:main_prob_u){reference-type="eqref" reference="eq:main_prob_u"}. In this section we are concerned with critical points of $\mathcal{E}_{\mu}$ on ${\mathcal{M}}$ (which, in turn, correspond to solutions of our starting problem [\[eq:main_prob_U\]](#eq:main_prob_U){reference-type="eqref" reference="eq:main_prob_U"}). By the Gagliardo-Nirenberg inequality [\[sobest\]](#sobest){reference-type="eqref" reference="sobest"}, setting $\\|\nabla u\\|^2_{L^2}=\alpha$, one obtains \[\label{eq:boundonboundEmu} \frac12\,\alpha-\mu\frac{C_{N,p}}{p+1}\,\alpha^{N(p-1)/4} \leq \mathcal{E}_{\mu}(u)\le \frac12\alpha.\] In particular, $\mathcal{E}_{\mu}$ is bounded on any bounded subset of ${\mathcal{M}}$, and it is bounded from below (and coercive) on the entire ${\mathcal{M}}$ for subcritical $p<1+4/N$ and for critical $p=1+4/N$ whenever $\mu< \frac{p+1}{2}C_{N,p}^{-1}$. In these cases, one can easily show that $\mathcal{E}_{\mu}$ satifies the P.S. condition and apply the classical minimax principle for even functionals on a closed symmetric submanifold (see e.g. ). In the complementary case, when $p$ is either supercritical, i.e. $p>1+4/N$, or critical and $\mu$ is large, then $\mathcal{E}_{\mu}$ is not bounded from below (see e.g. [\[minusinfty\]](#minusinfty){reference-type="eqref" reference="minusinfty"} below). In order to provide a minimax principle suitable for this case, we recall the Definition [\[def:genus\]](#def:genus){reference-type="ref" reference="def:genus"} of genus and that of $\mathcal{B}_\alpha$ (see equation [\[eq:defBU\]](#eq:defBU){reference-type="eqref" reference="eq:defBU"}). Furthermore, we denote with $K_{c}$ the (closed and symmetric) set of critical points of ${\mathcal{E}}_\mu$ at level $c$ contained in $\mathcal{B}_\alpha$. The following theorem is an adaptation of well known arguments of previous critical point theorems relying on index theory. We now provide a sufficient condition to guarantee the validity of assumption [\[ass2\]](#ass2){reference-type="eqref" reference="ass2"}. Exploiting the results above, we are ready to prove our main existence results. To conclude this section we prove that in the supercritical case, if $\mu$ is not too large, in addition to $(c_k)_k$ there is a further sequence of critical levels $(\bar c_k)$ of $\mathcal{E}_{\mu}$ constrained to $\mathcal{M}$. For concreteness, let us first consider the case $k=1$: since in such case $c_1$ is a local minimum of ${\mathcal{E}}_\mu$ in $\mathcal{M}$, and ${\mathcal{E}}_\mu$ is unbounded from below in $\mathcal{M}$, the critical level $\bar c_1$ is of mountain pass type. # Results in symmetric domains {#sec:symm} This section is devoted to the proof of Theorem [\[pro:symm\]](#pro:symm){reference-type="ref" reference="pro:symm"}, therefore we assume $1+4/N \leq p < 2^-1$. We perform the proof in the case of $\Omega=B$, but it will be clear that the main assumption on $\Omega$ is the following: - there is a tiling of $\Omega$, made by $h$ copies of a subdomain $D$, in such a way that from any solution $U_D$ of [\[eq:main_prob_U\]](#eq:main_prob_U){reference-type="eqref" reference="eq:main_prob_U"} on $D$ one can construct, using reflections, a solution $U_\Omega$ of [\[eq:main_prob_U\]](#eq:main_prob_U){reference-type="eqref" reference="eq:main_prob_U"} on $\Omega$. Then $U_\Omega$ has $h$ times the mass of $U_D$, and recalling Theorem [\[thm:intro_GS\]](#thm:intro_GS){reference-type="ref" reference="thm:intro_GS"} we deduce that [\[eq:main_prob_U\]](#eq:main_prob_U){reference-type="eqref" reference="eq:main_prob_U"} on $\Omega$ is solvable for any $\rho< h \cdot D_{N,p} \lambda_1(D)^{\frac{2}{p-1}-\frac{N}{2}}$. At this point, for a sequence $(D_k,h_k)_k$ of tilings satisfying (T), we obtain the solvability of [\[eq:main_prob_U\]](#eq:main_prob_U){reference-type="eqref" reference="eq:main_prob_U"} on $\Omega$ whenever \[\rho< h_k \cdot D_{N,p} \lambda_1(D_k)^{\frac{2}{p-1}-\frac{N}{2}},\] and if we can show that \[\label{eq:finaltarget} \frac{ h_k }{ \lambda_1(D_k)^{\frac{N}{2}-\frac{2}{p-1}}} \to +\infty\qquad\text{as }k\to+\infty,\] we deduce the solvability of [\[eq:main_prob_U\]](#eq:main_prob_U){reference-type="eqref" reference="eq:main_prob_U"} on $\Omega$ for every mass. Having this scheme in mind, it is easy to prove analogous results on rectangles and also in other kind of domains. Then let $B\subset{\mathbb{R}}^N$ be the ball (w.l.o.g. of radius one), and let \[D_k:=\left\{(r\cos\theta, r\sin\theta,x_3,\dots,x_N)\in B:-\frac{\pi}{k} < \theta < \frac{\pi}{k}\right\}\] Then $D_k$ satisfies (T)*, with $h_k=k$. In order to estimate $\lambda_1(D_k)$ we observe that, by elementary trigonometry, \[B'_k = B_{\frac{\sin(\pi/k)}{\sin(\pi/k)+1}}\left(\frac{1}{\sin(\pi/k)+1},0,0,\dots,0\right) \subset D_k,\] and therefore \[\lambda_1(D_k) \le \lambda_1(B'_k) \le C k^2,\] for some dimensional constant $C=C(N)$ and $k$ large. Then \[\frac{ h_k }{ \lambda_1(D_k)^{\frac{N}{2}-\frac{2}{p-1}}}\ge C \frac{k}{k^{{N}-\frac{4}{p-1}}} = C k^{1-{N}+\frac{4}{p-1}} = C k^{\frac{N-1}{p-1}\left[1+\frac{4}{N-1}-p\right]},\] and finally [\[eq:finaltarget\]](#eq:finaltarget){reference-type="eqref" reference="eq:finaltarget"} holds true whenever $p< 1+\frac{4}{N-1}$, thus completing the proof of Theorem [\[pro:symm\]](#pro:symm){reference-type="ref" reference="pro:symm"}.	{'timestamp': '2016-07-18T02:07:27', 'yymm': '1607', 'arxiv_id': '1607.04520', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04520'}
# Introduction {#sec:intro} Despite the immense popularity and availability of online video content via outlets such as Youtube and Facebook, most work on object detection focuses on static images. Given the breakthroughs of deep convolutional neural networks for detecting objects in static images, the application of these methods to video might seem straightforward. However, motion blur and compression artifacts cause substantial frame-to-frame variability, even in videos that appear smooth to the eye. These attributes complicate prediction tasks like classification and localization. Object-detection models trained on images tend not to perform competitively on videos owing to domain shift factors. Moreover, object-level annotations in popular video data-sets can be extremely sparse, impeding the development of better video-based object detection models. Girshik et aldemonstrate that even given scarce labeled training data, high-capacity convolutional neural networks can achieve state of the art detection performance if first pre-trained on a related task with abundant training data, such as 1000-way ImageNet classification. Followed the pretraining, the networks can be fine-tuned to a related but distinct domain. Also relevant to our work, the recently introduced models Faster R-CNN and You Look Only Once (YOLO) unify the tasks of classification and localization. These methods, which are accurate and efficient, propose to solve both tasks through a single model, bypassing the separate object proposal methods used by R-CNN. In this paper, we introduce a method to extend unified object recognition and localization to the video domain. Our approach applies transfer learning from the image domain to video frames. Additionally, we present a novel recurrent neural network (RNN) method that refines predictions by exploiting contextual information in neighboring frames. In summary, we contribute the following: - A new method for refining a video-based object detection consisting of two parts: (i) a pseudo-labeler, which assigns provisional labels to all available video frames. (ii) A recurrent neural network, which reads in a sequence of provisionally labeled frames, using the contextual information to output refined predictions. - An effective training strategy utilizing (i) category-level weak-supervision at every time-step, (ii) localization-level strong supervision at final time-step (iii) a penalty encouraging prediction smoothness at consecutive time-steps, and (iv) similarity constraints between pseudo-labels and prediction output at every time-step. - An extensive empirical investigation demonstrating that on the YouTube Objects dataset, our framework achieves mean average precision (mAP) of $68.73$ on test data, compared to a best published result of $37.41$ and $61.66$ for a domain adapted YOLO network. # Methods {#sec:method} In this work, we aim to refine object detection in video by utilizing contextual information from neighboring video frames. We accomplish this through a two-stage process. First, we train a pseudo-labeler, that is, a domain-adapted convolutional neural network for object detection, trained individually on the labeled video frames. Specifically, we fine-tune the YOLO object detection network, which was originally trained for the 20-class PASCAL VOC dataset, to the Youtube-Video dataset. When fine-tuning to the 10 sub-categories present in the video dataset, our objective is to minimize the weighted squared detection loss (equation [\[eqn:obj_det_loss\]](#eqn:obj_det_loss){reference-type="ref" reference="eqn:obj_det_loss"}) as specified in YOLO. While fine-tuning, we learn only the parameters of the top-most fully-connected layers, keeping the $24$ convolutional layers and $4$ max-pooling layers unchanged. The training takes roughly 50 epochs to converge, using the RMSProp optimizer with momentum of $0.9$ and a mini-batch size of $128$. As with YOLO, our fine-tuned $pseudo-labeler$ takes $448 \times 448$ frames as input and regresses on category types and locations of possible objects at each one of $S \times S$ non-overlapping grid cells. For each grid cell, the model outputs class conditional probabilities as well as $B$ bounding boxes and their associated confidence scores. As in YOLO, we consider a responsible bounding box for a grid cell to be the one among the $B$ boxes for which the predicted area and the ground truth area shares the maximum Intersection Over Union. During training, we simultaneously optimize classification and localization error (equation [\[eqn:obj_det_loss\]](#eqn:obj_det_loss){reference-type="ref" reference="eqn:obj_det_loss"}). For each grid cell, we minimize the localization error for the responsible bounding box with respect to the ground truth only when an object appears in that cell. Next, we train a Recurrent Neural Network (RNN), with Gated Recurrent Units (GRUs). This net takes as input sequences of pseudo-labels, optimizing an objective that encourages both accuracy on the target frame and consistency across consecutive frames. Given a series of pseudo-labels $\mathbf{x}^{(1)},..., \mathbf{x}^{(T)}$, we train the RNN to generate improved predictions $\hat{\mathbf{y}}^{(1)},..., \hat{\mathbf{y}}^{(T)}$ with respect to the ground truth $\mathbf{y}^{(T)}$ available only at the final step in each sequence. Here, $t$ indexes sequence steps and $T$ denotes the length of the sequence. As output, we use a fully-connected layer with a linear activation function, as our problem is regression. In our final experiments, we use a $2$-layer GRU with $150$ nodes per layer, hyper-parameters determined on validation data. The following equations define the forward pass through a GRU layer, where $\mathbf{h}^{(t)}_l$ denotes the layer's output at the current time step, and $\mathbf{h}^{(t)}_{l-1}$ denotes the previous layer's output at the same sequence step: \[\label{eqn:GRU} \begin{aligned} \mathbf{r}^{(t)}_l &= \sigma(\mathbf{h}^{(t)}_{l-1}W^{xr}_l + \mathbf{h}^{(t-1)}_lW^{hr}_l + \mathbf{b}^r_l)\\ \mathbf{u}^{(t)}_l &= \sigma(\mathbf{h}^{(t)}_{l-1}W^{xu}_l + \mathbf{h}^{(t-1)}_lW^{hu}_l + \mathbf{b}^u_l)\\ \mathbf{c}^{(t)}_l &= \sigma(\mathbf{h}^{(t)}_{l-1}W^{xc}_l + r_t \odot(\mathbf{h}^{(t-1)}_lW^{hc}_l) + \mathbf{b}^c_l)\\ \mathbf{h}^{(t)}_l &= (1-\mathbf{u}^{(t)}_l)\odot \mathbf{h}^{(t-1)}_l + \mathbf{u}^{(t)}_l\odot \mathbf{c}^{(t)}_l \end{aligned}\] Here, $\sigma$ denotes an element-wise logistic function and $\odot$ is the (element-wise) Hadamard product. The reset gate, update gate, and candidate hidden state are denoted by $\textbf{r}$, $\textbf{u}$, and $\textbf{c}$ respectively. For $S = 7$ and $B=2$, the pseudo-labels $\mathbf{x}^{(t)}$ and prediction $\hat{\mathbf{y}}^{(t)}$ both lie in $\mathbb{R}^{1470}$. ## Training We design an objective function (Equation [\[eqn:objective\]](#eqn:objective){reference-type="ref" reference="eqn:objective"}) that accounts for both accuracy at the target frame and consistency of predictions across adjacent time steps in the following ways: \[\label{eqn:objective} \mbox{loss} = \mbox{d\_loss} + \alpha \cdot \mbox{s\_loss} + \beta \cdot \mbox{c\_loss} + \gamma \cdot \mbox{pc\_loss}\] Here, d_loss, s_loss, c_loss and pc_loss stand for detection_loss, similarity_loss, category_loss and prediction_consistency_loss described in the following sections. The values of the hyper-parameters $\alpha=0.2$, $\beta=0.2$ and $\gamma=0.1$ are chosen based on the detection performance on the validation set. The training converges in 80 epochs for parameter updates using RMSProp and momentum $0.9$. During training we use a mini-batch size of $128$ and sequences of length $30$. ### Strong Supervision at Target Frame On the final output, for which the ground truth classification and localization is available, we apply a multi-part object detection loss as described in YOLO. \[\label{eqn:obj_det_loss} \begin{aligned} \mbox{detection\_loss} &= \lambda_{coord}\sum^{S^2}_{i=0}\sum^{B}_{j=0}\mathbbm{1}^{obj}_{ij}\big(\mathit{x}^{(T)}_i-\hat{\mathit{x}}^{(T)}_i\big)^2 + \big(\mathit{y}^{(T)}_i-\hat{\mathit{y}}^{(T)}_i\big)^2 \\ & + \lambda_{coord}\sum^{S^2}_{i=0}\sum^{B}_{j=0}\mathbbm{1}^{obj}_{ij}\big(\sqrt{w_i}^{(T)}-\sqrt{\hat{w}^{(T)}_i}\big)^2 + \big (\sqrt{h_i}^{(T)}-\sqrt{\hat{h}^{(T)}_i} \big)^2 \\ & + \sum^{S^2}_{i=0}\sum^{B}_{j=0}\mathbbm{1}^{obj}_{ij}(\mathit{C}_i-\hat{\mathit{C}_i})^2 \\ & + \lambda_{noobj}\sum^{S^2}_{i=0}\sum^{B}_{j=0}\mathbbm{1}^{noobj}_{ij}\big(\mathit{C}^{(T)}_i-\hat{\mathit{C}}^{(T)}_i\big)^2 \\ & + \sum^{S^2}_{i=0}\mathbbm{1}^{obj}_{i}\sum_{c \in classes}\big(p_i^{(T)}(c)-\hat{p_i}^{(T)}(c)\big)^2 \end{aligned}\] where $\mathbbm{1}^{obj}_{i}$ denotes if the object appears in cell $i$ and $\mathbbm{1}^{obj}_{ij}$ denotes that $j$th bounding box predictor in cell $i$ is responsible for that prediction. The loss function penalizes classification and localization error differently based on presence or absence of an object in that grid cell. $x_i, y_i, w_i, h_i$ corresponds to the ground truth bounding box center coordinates, width and height for objects in grid cell (if it exists) and $\hat{x_i}, \hat{y_i}, \hat{w_i}, \hat{h_i}$ stand for the corresponding predictions. $C_i$ and $\hat{C_i}$ denote confidence score of objectness at grid cell $i$ for ground truth and prediction. $p_i(c)$ and $\hat{p_i}(c)$ stand for conditional probability for object class $c$ at cell index $i$ for ground truth and prediction respectively. We use similar settings for YOLO's object detection loss minimization and use values of $\lambda_{coord} = 5$ and $\lambda_{noobj} = 0.5$. ### Similarity between Pseudo-labels and Predictions Our objective function also includes a regularizer that penalizes the dissimilarity between pseudo-labels and the prediction at each time frame $t$. \[\label{auto_enc_loss} \mbox{similarity\_loss} = \sum^T_{t=0}\sum^{S^2}_{i=0}\hat{C}^{(t)}_i\Big(\mathbf{x}^{(t)}_i-\hat{\mathbf{y}_i}^{(t)} \Big)^2\] Here, $\mathbf{x}^{(t)}_i$ and $\hat{\mathbf{y}_i}^{(t)}$ denote the pseudo-labels and predictions corresponding to the $i$-th grid cell at $t$-th time step respectively. We perform minimization of the square loss weighted by the predicted confidence score at the corresponding cell. ### Object Category-level Weak-Supervision Replication of the static target at each sequential step has been shown to be effective in. Of course, with video data, different objects may move in different directions and speeds. Yet, within a short time duration, we could expect all objects to be present. Thus we employ target replication for classification but not localization objectives. We minimize the square loss between the categories aggregated over all grid cells in the ground truth $\mathbf{y}^{(T)}$ at final time step $T$ and predictions $\hat{\mathbf{y}}^{(t)}$ at all time steps $t$. Aggregated category from the ground truth considers only the cell indices where an object is present. For predictions, contribution of cell $i$ is weighted by its predicted confidence score $\hat{C}^{(t)}_i$. Note that cell indices with positive detection are sparse. Thus, we consider the confidence score of each cell while minimizing the aggregated category loss. \[\label{category_supervision} \mbox{category\_loss} = \sum^T_{t=0}\bigg(\sum_{c \in classes} \Big(\sum^{S^2}_{i=0} \hat{C}^{(t)}_i\big(\hat{p}^{(t)}_i(c)\big)-\sum^{S^2}_{i=0}\mathbbm{1}^{obj^{(T)}}_i \big(p_i^{(T)}(c)\big)\Big) \bigg)^2\] ### Consecutive Prediction Smoothness Additionally, we regularize the model by encouraging smoothness of predictions across consecutive time-steps. This makes sense intuitively because we assume that objects rarely move rapidly from one frame to another. \[\label{prediction_smoothness} \mbox{prediction\_consistency\_loss} = \sum^{T-1}_{t=0}\Big(\hat{\mathbf{y}_i}^{(t)}-\hat{\mathbf{y}_i}^{(t+1)} \Big)^2\] ## Inference The recurrent neural network predicts output at every time-step. The network predicts $98$ bounding boxes per video frame and class probabilities for each of the $49$ grid cells. We note that for every cell, the net predicts class conditional probabilities for each one of the $C$ categories and $B$ bounding boxes. Each one of the $B$ predicted bounding boxes per cell has an associated objectness confidence score. The predicted confidence score at that grid is the maximum among the boxes. The bounding box with the highest score becomes the responsible prediction for that grid cell $i$. The product of class conditional probability $\hat{p}^{(t)}_i(c)$ for category type $c$ and objectness confidence score $\hat{C}^{(t)}_i$ at grid cell $i$, if above a threshold, infers a detection. In order for an object of category type $c$ to be detected for $i$-th cell at time-step $t$, both the class conditional probability $\hat{p}^{(t)}_i(c)$ and objectness score $\hat{C}^{(t)}_i$ must be reasonably high. Additionally, we employ Non-Maximum Suppression (NMS) to winnow multiple high scoring bounding boxes around an object instance and produce a single detection for an instance. By virtue of YOLO-style prediction, NMS is not critical. # Experimental Results {#sec:results} In this section, we empirically evaluate our model on the popular Youtube-Objects dataset, providing both quantitative results (as measured by mean Average Precision) and subjective evaluations of the model's performance, considering both successful predictions and failure cases. The Youtube-Objects dataset is composed of videos collected from Youtube by querying for the names of 10 object classes of the PASCAL VOC Challenge. It contains 155 videos in total and between 9 and 24 videos for each class. The duration of each video varies between 30 seconds and 3 minutes. However, only $6087$ frames are annotated with $6975$ bounding-box instances. The training and test split is provided. ## Experimental Setup We implement the domain-adaption of YOLO and the proposed RNN model using Theano. Our best performing RNN model uses two GRU layers of $150$ hidden units each and dropout of probability $0.5$ between layers, significantly outperforming domain-adapted YOLO alone. While we can only objectively evaluate prediction quality on the labeled frames, we present subjective evaluations on sequences. ## Objective Evaluation We compare our approach with other methods evaluated on the Youtube-Objects dataset. As shown in Table [1](#table:per_category_results){reference-type="ref" reference="table:per_category_results"} and Table [2](#table:final_mAP){reference-type="ref" reference="table:final_mAP"}, Deformable Parts Model (DPM) )-based detector reports mean average precision below $30$, with especially poor performance in some categories such as cat. The method of Tripathi et al(VPO) uses consistent video object proposals followed by a domain-adapted AlexNet classifier (5 convolutional layer, 3 fully connected) in an R-CNN-like framework, achieving mAP of $37.41$. We also compare against YOLO ($24$ convolutional layers, 2 fully connected layers), which unifies the classification and localization tasks, and achieves mean Average Precision over $55$. In our method, we adapt YOLO to generate pseudo-labels for all video frames, feeding them as inputs to the refinement RNN. We choose YOLO as the pseudo-labeler because it is the most accurate among feasibly fast image-level detectors. The domain-adaptation improves YOLO's performance, achieving mAP of $61.66$. Our model with RNN-based prediction refinement, achieves superior aggregate mAP to all baselines. The RNN refinement model using both input-output similarity, category-level weak-supervision, and prediction smoothness performs best, achieving $\mbox{68.73}$ mAP. This amounts to a relative improvement of $\mbox{11.5\%}$ over the best baselines. Additionally, the RNN improves detection accuracy on most individual categories (Table [1](#table:per_category_results){reference-type="ref" reference="table:per_category_results"}). [\[table:per_category_results\]]{#table:per_category_results label="table:per_category_results"} [\[table:final_mAP\]]{#table:final_mAP label="table:final_mAP"} ## Subjective Evaluation We provide a subjective evaluation of the proposed RNN model in Figure [\[fig:subjective1\]](#fig:subjective1){reference-type="ref" reference="fig:subjective1"}. Top and bottom rows in every pair of sequences correspond to pseudo-labels and results from our approach respectively. While only the last frame in each sequence has associated ground truth, we can observe that the RNN produces more accurate and more consistent predictions across time frames. The predictions are consistent with respect to classification, localization and confidence scores. In the first example, the RNN consistently detects the dog throughout the sequence, even though the pseudo-labels for the first two frames were wrong (bird). In the second example, pseudo-labels were motorbike, person, bicycle and even none at different time-steps. However, our approach consistently predicted motorbike. The third example shows that the RNN consistently predicts both of the cars while the pseudo-labeler detects only the smaller car in two frames within the sequence. The last two examples show how the RNN increases its confidence scores, bringing out the positive detection for cat and car respectively both of which fell below the detection threshold of the pseudo-labeler. ## Areas For Improvement The YOLO scheme for unifying classification and localization imposes strong spatial constraints on bounding box predictions since each grid cell can have only one class. This restricts the set of possible predictions, which may be undesirable in the case where many objects are in close proximity. Additionally, the rigidity of the YOLO model may present problems for the refinement RNN, which encourages smoothness of predictions across the sequence of frames. Consider, for example, an object which moves slightly but transits from one grid cell to another. Here smoothness of predictions seems undesirable. Figure [\[fig:failure_cases\]](#fig:failure_cases){reference-type="ref" reference="fig:failure_cases"} shows some failure cases. In the first case, the pseudo-labeler classifies the instances as dogs and even as birds in two frames whereas the ground truth instances are horses. The RNN cannot recover from the incorrect pseudo-labels. Strangely, the model increases the confidence score marginally for a different wrong category cow. In the second case, possibly owing to motion and close proximity of multiple instances of the same object category, the RNN predicts the correct category but fails on localization. These point to future work to make the framework robust to motion. The category-level weak supervision in the current scheme assumes the presence of all objects in nearby frames. While for short snippets of video this assumption generally holds, it may be violated in case of occlusions, or sudden arrival or departure of objects. In addition, our assumptions regarding the desirability of prediction smoothness can be violated in the case of rapidly moving objects. # Related Work {#sec:prior-art} Our work builds upon a rich literature in both image-level object detection,video analysis, and recurrent neural networks. Several papers propose ways of using deep convolutional networks for detecting objects. Some approaches classify the proposal regions into object categories and some other recent methods unify the localization and classification stages. Kalogeiton et alidentifies domain shift factors between still images and videos, necessitating video-specific object detectors. To deal with shift factors and sparse object-level annotations in video, researchers have proposed several strategies. Recently, proposed both transfer learning from the image domain to video frames and optimizing for temporally consistent object proposals. Their approach is capable of detecting both moving and static objects. However, the object proposal generation step that precedes classification is slow. Prest et al, utilize weak supervision for object detection in videos via category-level annotations of frames, absent localization ground truth. This method assumes that the target object is moving, outputting a spatio-temporal tube that captures this most salient moving object. This paper, however, does not consider context within video for detecting multiple objects. A few recent papers identify the important role of context in visual recognition. For object detection in images, Bell et aluse spatial RNNs to harness contextual information, showing large improvements on PASCAL VOC and Microsoft COCO object detection datasets. Their approach adopts proposal generation followed by classification framework. This paper exploits spatial, but not temporal context. Recently, Kang et alintroduced tubelets with convolutional neural networks (T-CNN) for detecting objects in video. T-CNN uses spatio-temporal tubelet proposal generation followed by the classification and re-scoring, incorporating temporal and contextual information from tubelets obtained in videos. T-CNN won the recently introduced ImageNet object-detection-from-video (VID) task with provided densely annotated video clips. Although the method is effective for densely annotated training data, it's behavior for sparsely labeled data is not evaluated. By modeling video as a time series, especially via GRU or LSTM RNNs, several papers demonstrate improvement on visual tasks including video classification, activity recognition, and human dynamics. These models generally aggregate CNN features over tens of seconds, which forms the input to an RNN. They perform well for global description tasks such as classification but require large annotated datasets. Yet, detecting multiple generic objects by explicitly modeling video as an ordered sequence remains less explored. Our work differs from the prior art in a few distinct ways. First, this work is the first, to our knowledge, to demonstrate the capacity of RNNs to improve localized object detection in videos. The approach may also be the first to refine the object predictions of frame-level models. Notably, our model produces significant improvements even on a small dataset with sparse annotations. # Conclusion We introduce a framework for refining object detection in video. Our approach extracts contextual information from neighboring frames, generating predictions with state of the art accuracy that are also temporally consistent. Importantly, our model benefits from context frames even when they lack ground truth annotations. For the recurrent model, we demonstrate an efficient and effective training strategy that simultaneously employs localization-level strong supervision, category-level weak-supervision, and a penalty encouraging smoothness of predictions across adjacent frames. On a video dataset with sparse object-level annotation, our framework proves effective, as validated by extensive experiments. A subjective analysis of failure cases suggests that the current approach may struggle most on cases when multiple rapidly moving objects are in close proximity. Likely, the sequential smoothness penalty is not optimal for such complex dynamics. Our results point to several promising directions for future work. First, recent state of the art results for video classification show that longer sequences help in global inference. However, the use of longer sequences for localization remains unexplored. We also plan to explore methods to better model local motion information with the goal of improving localization of multiple objects in close proximity. Another promising direction, we would like to experiment with loss functions to incorporate specialized handling of classification and localization objectives.	{'timestamp': '2016-07-20T02:04:30', 'yymm': '1607', 'arxiv_id': '1607.04648', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04648'}
# Introduction The wave-particle duality is an alternative statement of the complementarity principle, and it establishes the relation between nature of quantum entities. It can be illustrated in a two-way interferometer, where the apparatus can be set to observe behavior when a single path is or wave-like behavior, when the impossibility to define a path is shown by the interference. A modern approach to the wave-particle duality includes quantitative relations between quantities that represent the possible a priori knowledge of the which-way information () and the "quality" of the interference fringes (). Several publications in the literature contributed to the formulation of the quantitative analysis of the wave-particle duality. For a bipartite system give an extra which-way (path) information about the interferometric possibilities. The quantitative relations for systems composed by two particles were studied in. Therefore, the behavior of such quantities, in various regimes and situations, is essential to answer fundamental and/or technological questions of the quantum theory. an example is the so-called quantum eraser, Since its proposal has both theoretically and experimentally (see for example Refs. ). In , the authors explore the quantum eraser problem in model. In this work, an increase of visibility is achieved by performing appropriate projective measurements. An intrinsic relation between the complementarity quantities and the performed measurements is outlined: since were made in order to obtain an of the , the remaining quantities (Entanglement as measured by the concurrence, and the predictability) must obey a "complementary" behavior. In that case, and predictability increases, and entanglement decreases, since the measurements are made in order to the quantum eraser. In Reference only the maximization of the visibility was considered, in the present work we extend the analysis and consider maximization of predictability, visibility and concurrence. Also, in the previous work, only one value of the coupling constant was considered. In this contribution we consider a second coupling regime that allows for the comparison between stronger and weaker interactions. Some questions may arise from the analysis presented in: how is the behavior of the , predictability and Entanglement, for different strengths of the coupling between the cavities and the $N$ atoms? if one measure the qubits in order to maximize another complementarity quantity? For finite $N$, could resemble the reservoir (dissipative) limit? Moreover, one can think about a three-part control scheme: initially parts $A$ and $B$ possesses a maximally entanglement state, constituted by two qubits $q_A$ and $q_B$, respectively. A third part $R$ may have, in principle, full control of a group of $N$-qubits (each one we call as $q_i$); i.e. $R$ may control: (i) the initial state of each qubit $q_i$, (ii) the interaction strength between $q_i$ and $q_B$ and (iii) the measurement basis where each $q_i$ could be projected by $R$. Here we will focus in the control of item (iii), therefore the initial state of all $q_i$ and the coupling strenght will be fixed for each realization of the scheme. Thus, part $R$ is able to control which complementarity quantity of part $A$ they ($A$ and $R$) would like to maximize. For instance, if $A$ and $R$ desire that $q_A$ is in a superposition state, $R$ can choose which basis he/she will project each qubit in order to accomplish the task (quantum eraser task ). However, now $R$ and $A$ are able to choose another complementarity quantity: if they would like to obtain and/or maintain an Entangled state between $A$ and $B$, $R$ may project each $q_i$ in a basis chosen in order obtain an state nearly maximally Entangled (the same idea follows for the predictability). More than that, since part $R$ can adjust the strenght of the interaction between $q_B$ and $q_i$, he/she can study what is the best option of coupling to do each task (together with the freedom to choose the basis of projection). In that way, parts $A$, $B$ and $R$ are able to study in details the behavior of the complementarity quantities, for a variety of conditions. In the present work we answer the questions and provide a useful tool to implement the control scheme mentioned above, considering a similar model two entangled qubits, $q_{A}$ and $q_{B}$, and a third system ($R$) which is composed by $N$ qubits. It is well known that $R$ will play the role of a reservoir. As it is possible to measure each qubit of system $R$ after the interaction, we can control the evolution of $q_{A}$ and $q_{B}$, induced by the interaction with $R$, by selecting an adequate sequence of results of measurements performed in the qubits of $R$. Such control would allow us to make $q_{A}$ and $q_{B}$ approach a chosen asymptotic state. This scheme can be implemented in cavity-QED system, where $q_{A}$ and $q_{B}$ would be cavity modes, prepared in an entangled state with one excitation, and $N$ two level atoms, interacting with the cavities one at the time, would play the role of the qubits that compose the system $R$. We consider the complementarity quantities concurrence, predictability and to guide the manipulation over $R$ and to quantify the information present in each subsystem. We also consider two regimes for the strength of the coupling constant $g$ between $q_B$ and each qubit of $R$, $g= \frac{1}{4}$ and $g= 4$. The paper is organized as : in section [2](#model){reference-type="ref" reference="model"} we the principal quantities studied: , concurrence and predictability. the distinguishability between different parts of the global system. In subsection [2.1](#digression){reference-type="ref" reference="digression"} we briefly review the case where $q_A+q_B$ interacts in a dissipative reservoir, and how the complementarity quantities behave in this case. Section [3](#results){reference-type="ref" reference="results"} shows how we implement the projective measurements in $R$, and present our results and discussions for the behavior of the complementarity quantities and for the variation of distinguishability (after and before the measurements). In section [4](#conclusion){reference-type="ref" reference="conclusion"} we conclude our work. # Model and definitions {#model} Let us consider that initially qubits $q_A$ and $q_B$ were prepared in the entangled state $\ket{\psi(0)} = \frac{1}{\sqrt 2} (\ket{0_A 1_B} + \ket{1_A 0_B})$ and a third system $R$ composed by , prepared all in the ground state $\ket 0$. Each qubit of $R$ interacts one at a time with $q_B$ and after a sequence of these interactions, the information, initially stored in $q_A$ and $q_B$, will be distributed over the and qubits $q_A$ and $q_B$. As an example of our interaction model, consider the following dynamics governing the interaction of an atom (between a total of $N$ atoms) and a cavity ($q_B$) (see Figure [\[scheme\]](#scheme){reference-type="ref" reference="scheme"}). The Hamiltonian that gives the interaction between the $k$-th atom and $q_B$ is $\hat{H}^{(k)}= \omega \hat{b}^{\dagger }\hat{b}+ \frac{\omega}{2}\hat{\sigma}_{z}^{(k)}+ g (\hat{b}^{\dagger}\hat{\sigma}_{-}^{(k)}+\hat{b}\hat{\sigma}_{+}^{(k)}),$ where $\hat{b}^{\dagger }$ ($\hat{b}$) corresponds to the creation (annihilation) operator for $q_B$, $\omega$ their transition frequency, $\hat{\sigma}_{z}^{(k)}=\|1^{(k)}\rangle\langle 1^{(k)}\|-\|0^{(k)}\rangle\langle 0^{(k)}\|$, $\hat{\sigma}_{-}^{(k)}=\|0^{(k)}\rangle\langle 1^{(k)}\|$, $\hat{\sigma}_{+}^{(k)}=\|1^{(k)}\rangle\langle 0^{(k)}\|$, $g$ the coupling constant for the interaction between the $k$-th qubit of $R$ and $q_B$. As the initial state has one excitation and the Hamiltonian preserves the excitation number, the states of mode $B$ can be written in the basis $\lbrace\vert 0 \rangle, \vert 1 \rangle\rbrace$. Although constant in each preparation, we let the parameter $g$ free in order to quantify the strength of the interaction, since we will analyze different coupling regimes and the correspondent behavior of the Complementarity quantities. In this model, $\left\|0^{(k)} \right\rangle$ and $\left\|1^{(k)}\right\rangle$ stand for the levels $0$ and $1$ of the $k$-th interacting atom, respectively. After $n$ qubits of $R$ have interacted with $q_B$ (note the difference between $N$, the total number of qubits that are able to interact, and $n$, the number of qubits that will interact at a given time) the global system is left in the state \[\begin{aligned} \ket{\psi^{(n)}} &=& \frac{1}{\sqrt 2} \Big( a^n \ket{0_{A}} \ket{0_{res}} \ket{1_{B}} + \Gamma \ket{0_{A}} \ket{1_{res}} \ket{0_{B}} + \nonumber \\ && + \ket{1_{A}} \ket{0_{res}} \ket{0_{B}} \Big), \label{psin}\end{aligned}\] where $a = \cos\left( \frac{g T}{N} \right)$ and $b =-i \sin \left( \frac{g T}{N} \right)$, assuming the same interaction time between each qubit of $R$ and $q_{B}$, given by $\Delta t=\frac{T}{N}$. To simplify the notation we define a normalized state with one excitation in subsystem $R$: $\ket{1_{res}}= \frac{1}{\Gamma} \left( a^{n-1} b \ket{n} + \ldots + a b \ket{2} + b\ket{1} \right)$ $\Gamma^{2} = 1-a^{2n}$ $\ket{i}= \ket{0_1 0_2 \ldots 0_{i-1} ~1_i ~ 0_{i+1} \ldots 0_n}$ represents a state with an excitation in the $i$-th qubit and $\ket{0_{res}} = \ket{0_1 0_2 \ldots 0_n}$. For a general pure two qubit state: $\ket{\Psi} = \gamma_1 \ket{00} + \gamma_2 \ket{01} + \gamma_3 \ket{10} + \gamma_4 \ket{11},$ the complementary quantities are defined in the following way. The concurrence, which is related to the quantum correlation between the parts, is given by $C(\ket{\Psi}) = 2 \abs{\gamma_1 \gamma_4-\gamma_2 \gamma_3}$. The coherence between two orthogonal states gives the visibility, defined by $V = 2 \abs{\bra{\Psi}\sigma_{+}\ket{\Psi}} = 2 \abs{\gamma_1 \gamma_3^* + \gamma_2 \gamma_4^}$. Besides, the predictability measures the knowledge if one of the parts is in state $\ket{0}$ or $\ket{1}$, $P= \abs{\bra{\Psi}\sigma_{z}\ket{\Psi}} = \abs{\abs{\gamma_3}^2+\abs{\gamma_4}^2-\left( \abs{\gamma_1}^2 + \abs{\gamma_2}^2 \right)}$. The distinguishability, or the "measure of the possible which-path information that one can obtain" in an interferometer setup, is given by $D = \sqrt{C^2 + P^2}$. ## Continuous Limit-A digression {#digression} Defining $k = g^2 \frac{T}{N},$ one can write $a = \cos \left(\sqrt{\frac{k T}{N}} \right),$ where $T$ is the total time of interaction between $q_B$ and $R$ . We consider that the interaction time between each qubit of $R$ and $q_B$ is equal, given by $\Delta t=\frac{T}{N}$. We also assume that $0<\Delta t<\pi/2g.$ After $N$ interactions the reduced state in the subsystem $q_A$ is a statistical mixture $\rho_{A}=\frac{1}{2}\left(\vert 0_{A} \rangle \langle 0_{A} \vert + \vert 1_{A} \rangle \langle 1_{A} \vert\right)$, therefore $V_{q_A} = 0$ and $P_{q_A} = 0$. The concurrence can be calculated from the reduced state of the subsystem $q_A, q_B$ and is given by $C_{q_A, q_B} = a^N$. The limit $N \rightarrow \infty$ (and consequently $g \rightarrow \sqrt{\frac{N}{T}} \rightarrow \infty$; $\Delta t \rightarrow dt$) is well known in the Literature and it gives the reservoir limit (at a given temperature implicitly defined in $k$) of a qubit interacting with a Markovian pure dissipative reservoir. The term $a^n$ in [\[psin\]](#psin){reference-type="eqref" reference="psin"} is in this limit $\lim_{N \rightarrow \infty} a^N = \lim_{N \rightarrow \infty} \cos^N \sqrt{\frac{k T}{N}} = \mathrm{e}^{-kT/2}$, and consequently the concurrence $C_{q_A, q_B} = \mathrm{e}^{\frac{-k T}{2}}$ decays exponentially with $T$. # Results ## Complementarity quantities versus coupling intensity Similar to what was done in Ref. , let us now consider that, after $n$ interactions, the $i$-th qubit of $R$ is projected in the state: $\ket{M_i} = \alpha_i \ket{0_i} + \beta_i \ket{1_i},$ where $\alpha_i = \cos \theta_i$ e $\beta_i = \mathrm{e}^{i \phi_i} \sin \theta_i$ (this measure can be done experimentally, see for example ). The vector state $\ket{M_i}$ is an eigenstate of the operator with $\vec{n}=\left(\sin2\theta_{i} ~ \cos2\phi_{i},~ \sin2\theta_{i}~\sin2\phi_{i},~\cos2\theta_{i}\right)$ and , the Pauli matrices. One can, in principle, choose in which base ($\theta_i$ and $\phi_i$) the global state will be measured. projective measurements performed on the state [\[psin\]](#psin){reference-type="eqref" reference="psin"}, the projector is given by $\Pi = \Pi_1 \otimes \ldots \otimes \Pi_n,$ with $\Pi_i = \mathbb{I}_1 \otimes \ldots \otimes \mathbb{I}_{i-1} \otimes \ket{M_i}\bra{M_i} \otimes \ldots \mathbb{I}_n.$ Notice that the projective measure $\Pi$ acts only on the subsystem $R$. After $n$ projective measurements the normalized global state vector is given by: \[\begin{aligned} \ket \psi^{(n,M)} &=& \frac{1}{N}\Big( \gamma_1 \ket{0_A} \ket M \ket{0_B} + \gamma_2 \ket{0_A} \ket{M} \ket{1_B} \nonumber \\ &&+ \gamma_3 \ket{1_A} \ket M \ket{0_B} \Big), \label{rhoreducedM}\end{aligned}\] where $\ket M = \ket{M_1} \ldots \ket{M_n}$, $N = \sqrt{\abs{\gamma_1}^2 + \abs{\gamma_2}^2 + \abs{\gamma_3}^2},$ and \[\begin{aligned} \gamma_1 &=& \frac{1}{\sqrt 2} \left(b \sum_{i = 1}^{n} \left[ a^{i-1} \frac{\beta_i}{\alpha_i} \left( \prod_{j=1}^{n} \alpha_j \right) \right] \right),\nonumber \\ \gamma_2 &=& \frac{1}{\sqrt 2} \left(a^n \prod_{i=1}^{n} \alpha_i \right), \nonumber \\ \gamma_3 &=& \frac{1}{\sqrt 2} \left( \prod_{i=1}^{n} \alpha_i \right).\end{aligned}\] The information carried by the qubits of $R$ are now embodied in the measurement outcomes $\theta_i$ and $\phi_i$. The complementarity quantities after the measurements are given by: \[\begin{aligned} V^{(n,M)}_{q_{A}} &=& \frac{2 \abs{\gamma_1 \gamma_3^}}{N^2} \nonumber \\ P^{(n,M)}_{q_{A}} &=& \frac{\abs{\abs{\gamma_3}^2-\abs{\gamma_1}^2-\abs{\gamma_2}^2}}{N^2} \nonumber \\ C^{(n,M)}_{q_{A},q_{B}} &=& \frac{2 \abs{\gamma_2 \gamma_3}}{N^2}. \nonumber \label{complementarity} \end{aligned}\] Since the reduced state is pure [\[rhoreducedM\]](#rhoreducedM){reference-type="eqref" reference="rhoreducedM"}, the closure relation for complementarities : \[\left(C^{(n,M)}_{q_{A},q_{B}}\right)^2 +\left(P^{(n,M)}_{q_{A}}\right)^2 +\left(V^{(n,M)}_{q_{A}}\right)^2 = 1. \label{complem1}\] quantities depend explicit on the coefficients $\theta_i$ and $\phi_i$ of $\ket M_i$. Concerning the Complementarity quantities, one can project the global state so that $V^{(n,M)}_{q_{A}}, P^{(n,M)}_{q_{A}}$ or $C^{(n,M)}_{q_{A}, q_B}$ acquire the maximum allowed values, after $n$ measurements on the qubits of $R$ have interacted with $q_B$. In , the authors studied a similar maximization procedure, only for the $V^{(n,M)}_{q_{A}}$ to produce a multipartite quantum eraser the coefficients $\alpha_i$ and $\beta_i$ were chosen to obtain an in the , maintaining a standard value for the coupling parameter ($g T = 2 \pi$). Here we are interested in how each of the Complementarity quantities behaves, between $q_B$ and the $i$-th of $R$, while making projective measurements in each qubit of $R$. if the function to be maximized is the concurrence $C^{(n,M)}_{q_{A}, q_B}$, for example, the procedure gives the values of $\theta_i$ and $\phi_i$ that provides the maximum value of $C^{(n,M)}_{q_{A}, q_B}$, after $n$ qubits have interacted with $q_B$; then, in the possession of $\theta_i$ and $\phi_i$, one can evaluate $V^{(n,M)}_{q_{A}}$ and $P^{(n,M)}_{q_{A}}$. The same procedure is carried out in order to maximize the or the predictability. Therefore, we have all the Complementarity quantities for each function to be maximized: $V^{(n,M)}_{q_{A}}, P^{(n,M)}_{q_{A}}$ or $C^{(n,M)}_{q_{A}, q_B}$. In Figures [\[maxvis\]](#maxvis){reference-type="ref" reference="maxvis"}, [\[maxpre\]](#maxpre){reference-type="ref" reference="maxpre"} and [\[maxcon\]](#maxcon){reference-type="ref" reference="maxcon"} we show the Complementarity quantities, for the maximization of $V^{(n,M)}_{q_{A}}, P^{(n,M)}_{q_{A}}$ and $C^{(n,M)}_{q_{A}, q_B}$, respectively. We also present the behaviour for different couplings (each Figure (a) depicts $g = 4$ and each Figure (b) $g = \frac{1}{4}$). The solid curve is related to the $\ankb{Visibility}{visibility}$; the same follows for the $concurrence$ ( dashed curve) and the $predictability$ (black dotted curve). Note in Figures [\[maxvis\]](#maxvis){reference-type="ref" reference="maxvis"}a and [\[maxvis\]](#maxvis){reference-type="ref" reference="maxvis"}b, the is increasing function of $n$, when the measurements are made to maximize $V^{(n,M)}_{q_{A}}$ if one measurements in order to maximize another Complementarity quantity, the does not increase, and remains zero, as one can see in Figures [\[maxpre\]](#maxpre){reference-type="ref" reference="maxpre"} and [\[maxcon\]](#maxcon){reference-type="ref" reference="maxcon"}. [Differently from the visibility, always when the predictability P is maximized it achieves the maximum value for some finite value of $n$ (Figures [\[maxpre\]](#maxpre){reference-type="ref" reference="maxpre"}a and [\[maxpre\]](#maxpre){reference-type="ref" reference="maxpre"}b). Moreover, it achieves a maximum valeu also when other quantities are maximized (Figure [\[maxcon\]](#maxcon){reference-type="ref" reference="maxcon"}a). Using an interferometric analogy, predictability is the which-way information that is available in the interferometric system, while gives the quality of the interference pattern and concurrence is a measure of entanglement between $q_a$ (main system) and $q_B$ (which-way detector). Therefore, in Figures [\[maxpre\]](#maxpre){reference-type="ref" reference="maxpre"}a, [\[maxpre\]](#maxpre){reference-type="ref" reference="maxpre"}b and [\[maxcon\]](#maxcon){reference-type="ref" reference="maxcon"}a, since the system loses entanglement, with no acquisition of visibility, the predictability must increase.]{style="color: 0.00,0.00,0.00"} This feature can be understood in our approach, since the projective measurements will inherently modify the global system. In Table [\[table\]](#table){reference-type="ref" reference="table"} we show the states after $n$ interactions ($n = 1, 2$ and $10$), and after performing the maximization procedures. For instance, performing a maximization of $P_{q_A}^{(n,M)}$, the state $\ket{\psi}^{(n,M)}$ for $g T = 2 \pi \times 4$ tends to the state $\ket{1_A 0_B}$. [The dashed curves in Figures [\[maxvis\]](#maxvis){reference-type="ref" reference="maxvis"}, [\[maxpre\]](#maxpre){reference-type="ref" reference="maxpre"} and [\[maxcon\]](#maxcon){reference-type="ref" reference="maxcon"} show the concurrence as a function of $n$. For $g T = \frac{2 \pi}{4}$, if the function to be maximized is the concurrence itself $C^{(n,M)}_{q_{A}, q_B}$ (Figure [\[maxcon\]](#maxcon){reference-type="ref" reference="maxcon"}b) , it is possible to maintain the state almost maximally Entangled--$C^{(n,M)}_{q_{A}, q_B} \sim 1$--by choosing the proper values of $\theta_i$ and $\phi_i$. However, if the coupling is increased--Figures [\[maxvis\]](#maxvis){reference-type="ref" reference="maxvis"}a, [\[maxpre\]](#maxpre){reference-type="ref" reference="maxpre"}a, or [\[maxcon\]](#maxcon){reference-type="ref" reference="maxcon"}a--even if the projective measures were made to maximize the concurrence (Figure [\[maxcon\]](#maxcon){reference-type="ref" reference="maxcon"}a), Entanglement decreases to zero. This behavior is similar to two entangled qubits, where $q_B$, is coupled to a thermal reservoir (red solid curves), but in our case we have a finite number of interacting qubits with $q_B$ (where the maximum number of interacting qubits is $N = 20$). how the initial information (given by the distinguishability between $q_A$ and the $i$-th qubit) is distributed over the $N$ qubits . An interesting aspect concerning Figure [\[maxcon\]](#maxcon){reference-type="ref" reference="maxcon"}b, for $g T = \frac{2 \pi}{4}$, is that one can see an approximately steady behavior of the concurrence $C^{(n,M)}_{q_{A}, q_B}$, near the initial value $C^{(0)}_{q_{A}, q_B} = 1$. It is possible, therefore, to maintain the system $q_A + q_B$ in an approximately maximal Entangled state, notwithstanding the qubits of $R$ became dynamically correlated with $q_A$ and $q_B$ (Equation [\[rhon\]](#rhon){reference-type="eqref" reference="rhon"}). One can see from Table [\[table\]](#table){reference-type="ref" reference="table"}, for $g T = \frac{2 \pi}{4}$ and $n=10$, the state resemble the initial maximally entangled state corroborating Figure [\[maxcon\]](#maxcon){reference-type="ref" reference="maxcon"}b.]{style="color: 0.00,0.00,0.00"}	{'timestamp': '2017-01-26T02:07:48', 'yymm': '1607', 'arxiv_id': '1607.04617', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04617'}
null	null
null	null
null	null
# Introduction {#sec:intro} Compact stars have a large number of pulsation modes that have been extensively studied since the seminal work of Chandrasekhar on radial oscillations. In general, these modes are very difficult to observe in the electromagnetic spectrum; therefore most efforts have concentrated on gravitational wave asteroseismology in order to characterise the frequency and damping times of the modes that emit gravitational radiation. In particular, various works focused on the oscillatory properties of pure hadronic stars, hybrid stars and strange quark stars trying to find signatures of the equation of state of high density neutron star matter (see and references therein). More recently, compact star oscillations have attracted the attention in the context of Soft Gamma ray Repeaters (SGRs), which are persistent X-ray emitters that sporadically emit short bursts of soft $\gamma$-rays. In the quiescent state, SGRs have an X-ray luminosity of $\sim 10^{35}$ erg/s, while during the short $\gamma$-bursts they release up to $10^{42}$ erg/s in episodes of about 0.1 s. Exceptionally, some of them have emitted very energetic giant flares which commenced with brief $\gamma$-ray spikes of $\sim 0.2$ s, followed by tails lasting hundreds of seconds. Hard spectra (up to 1 MeV) were observed during the spike and the hard X-ray emission of the tail gradually faded modulated at the neutron star (NS) rotation period. The analysis of X-ray data of the tails of the giant flares of SGR 0526-66, SGR 1900+14 and SGR 1806-20 revealed the presence of quasi-periodic oscillations (QPOs) with frequencies ranging from $\sim$ 18 to 1840 Hz. There are also candidate QPOs at higher frequencies up to $\sim 4$ kHz in other bursts but with lower statistical significance; in fact, according to a more recent analysis only one burst shows a marginally significant signal at a frequency of around 3706 Hz. Several characteristics of SGRs are usually explained in terms of the magnetar model, assuming that the object is a neutron star with an unusually strong magnetic field ($B \sim 10^{15}$ G). In particular, giant flares are associated to catastrophic rearrangements of the magnetic field. Such violent phenomena are expected to excite a variety of oscillation modes in the stellar crust and core. In fact, recent studies have accounted for magnetic coupling between the crust and the core, and associate QPOs to global magneto-elastic oscillations of highly magnetized neutron stars. There has also been interest in the possible excitation of low order $f$-modes because of their strong coupling to potentially detectable gravitational radiation. In the present paper we focus on radial oscillations of neutron stars permeated by ultra-strong magnetic fields. These modes might be relevant within the magnetar model because they could be excited during the violent events associated with gamma flares. Since they have higher frequencies than the already known QPOs, they cannot be directly linked to them at present. However, it is relevant to know all the variety of pulsation modes of strongly magnetized neutron stars because the number of observations is still small and new features could emerge in future flares' data. On the other hand, in the case of rotating objects we can expect some amount of gravitational radiation from even the lowest ($l = 0$) quasi-radial mode making them potentially relevant for gravitational wave astronomy. # Equations of state {#sec2} ## Hadronic phase under a magnetic field {#sec:A} In this section we present an overview of the hadronic equations of state (EOS) used in this work. We describe hadronic matter within the framework of the relativistic non-linear Walecka (NLW) model. In this model we employ a field-theoretical approach in which the baryons interact via the exchange of $\sigma-\omega-\rho$ mesons in the presence of a magnetic field $B$ along the $z-$axis. The total lagrangian density reads: \[\label{lt} \mathcal{L}_{H}=\sum_{b}\mathcal{L}_{b}+\mathcal{L}_{m}+\sum_{l}\mathcal{L}_{l}+\mathcal{L}_B\.\] where $\mathcal{L}_{b}$, $\mathcal{L}_{m}$, $\mathcal{L}_{l}$ and $\mathcal{L}_{B}$ are the baryons, mesons, leptons and electromagnetic field Lagrangians, respectively, and are given by \[\begin{aligned} \mathcal{L}_{b} &=& \overline{\psi}_{b}\left(i\gamma_{\mu}\partial^{\mu}-q_{b}\gamma_{\mu}A^{\mu}-m_{b}+g_{\sigma b}\sigma \right. \nonumber \\ && \left.-g_{\omega b}\gamma_{\mu}\omega^{\mu}-g_{\rho b}\tau_{3b}\gamma_{\mu}\rho^{\mu}\right)\psi_{b}\,, \\ \mathcal{L}_{m} &=& \tfrac{1}{2}(\partial_{\mu}\sigma\partial^{\mu}\sigma-m_{\sigma}^{2}\sigma^{2})-U(\sigma)+ \tfrac{1}{2}m_{\omega}^{2}\omega_{\mu}\omega^{\mu} \nonumber \\ &&-\tfrac{1}{4}\Omega_{\mu\nu}\Omega^{\mu\nu}+ \tfrac{1}{2}m_{\rho}^{2}\vec{\rho}_{\mu}\cdot\vec{\rho}_{\mu}-\tfrac{1}{4}P^{\mu\nu}P_{\mu\nu} \,, \\ \mathcal{L}_{l} &=& \overline{\psi}_{l}\left(i\gamma_{\mu}\partial^{\mu}-q_{l}\gamma_{\mu}A^{\mu}-m_{l}\right)\psi_{l} \,, \\ \mathcal{L}_{B} &=&-\tfrac{1}{4}F^{\mu\nu}F_{\mu\nu} \. \end{aligned}\] where he $b$-sum runs over the baryonic octet $b\equiv N~(p,~n),~\Lambda,~\Sigma^{\pm,0},~\Xi^{-,0}$, $\psi_{b}$ is the corresponding baryon Dirac field, whose interactions are mediated by the $\sigma$ scalar, $\omega_{\mu}$ isoscalar-vector and $\rho_{\mu}$ isovector-vector meson fields. The baryon charge, baryon mass and isospin projection are denoted by $q_{b}$, $m_{b}$ and $\tau_{3b}$, respectively, and the masses of the mesons are $m_{\sigma}= 512~$ MeV, $m_{\omega}=783~$MeV and $m_{\rho}=770~$MeV. The strong interaction couplings of the nucleons with the meson fields are denoted by $g_{\sigma N}=8.910$, $g_{\omega N}=10.610$ and $g_{\rho N}=8.196$. We consider that the couplings of the hyperons with the meson fields are fractions of those of the nucleons, defining $g_{iH}=X_{iH}g_{iN}$, where the values of $X_{iH}$ are chosen as $X_{\sigma H}=0.700$ and $X_{\omega H}=X_{\rho H}=0.783$. The term $U(\sigma)=\frac{1}{3}\,bm_{n}(g_{\sigma N}\sigma)^{3}-\frac{1}{4}\,c(g_{\sigma N}\sigma)^{4}$ denotes the scalar self-interactions, with $c=-0.001070$ and $b=0.002947$. The mesonic and electromagnetic field tensors are given by their usual expressions $\Omega_{\mu\nu}=\partial_{\mu}\omega_{\nu}-\partial_{\nu}\omega_{\mu}$, ${\bf P}_{\mu\nu}=\partial_{\mu}\vec{\rho}_{\nu}-\partial_{\nu}\vec{\rho}_{\mu}-g_{\rho b}(\vec{\rho}_{\mu}\times\vec{\rho}_{\nu})$ and $F_{\mu\nu}=\partial_{\mu}A_{\nu}-\partial_{\nu}A_{\mu}$. The $l$-sum runs over the two lightest leptons $l\equiv e,\mu$ and $\psi_{l}$ is the lepton Dirac field. The symmetric nuclear matter properties at saturation density adopted in this work are given by the GM1 parametrization, with compressibility $K=300$ MeV, binding energy $B/A=-16.3$ MeV, symmetry energy $a_{sym}=32.5$ MeV, slope $L=94$ MeV, saturation density $\rho_{0}= 0.153$ fm$^{-3}$ and nucleon mass $m=938$ MeV. The following equations present the scalar and vector densities for the charged and uncharged baryons, respectively: \[\begin{aligned} &\label{densities1}\rho_{b}^{s}=\frac{\|q_{b}\|B \bar{m}_{b} }{2\pi^{2}}\sum_{\nu}^{\nu_{\mathrm{max}}}\sum_{s}\frac{\bar{m}_{b}^{c}}{\sqrt{ \bar{m}_{b}^2 + 2\nu \|q_{b}\|B}} \ln\bigg\|\frac{k_{F,\nu,s}^{\,b}+E_{F}^{\,b}}{\bar{m}_{b}^{c}} \bigg\|,\\ &\label{densities2}\rho_{b}^{v}=\frac{\|q_{b}\|B}{2\pi^{2}}\sum_{\nu}^{\nu_{\mathrm{max}}}\sum_{s}k_{F,\nu,s}^{\,b},\\ &\label{densities3}\rho_{b}^{s}=\frac{\bar{m}_{b}}{4\pi^{2}}\sum_{s}\bigg[E_{F}^{\,b}k_{F,s}^{\,b}-\bar{m}_{b}^{2}\ln\bigg\|\frac{k_{F,s}^{\,b}+E_{F}^{\,b}}{\bar{m}_{b}}\bigg\|\bigg] ,\\ &\label{densities4}\rho_{b}^{v}=\frac{1}{2\pi^{2}}\sum_{s}\bigg[\frac{1}{3}(k_{F,s}^{\,b})^{3}\bigg], \end{aligned}\]	{'timestamp': '2016-07-19T02:02:47', 'yymm': '1607', 'arxiv_id': '1607.04707', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04707'}
# Introduction Ultra-cold molecules can play important roles in studies of many-body quantum physics , quantum logic operations , and ultra-cold chemistry . In our recent studies of LiRb, motivated largely by its large permanent dipole moment , we have explored the generation of these molecules in a dual species MOT . In particular, we have found that the rate of generation of stable singlet ground state molecules and first excited triple state molecules through photoassociation, followed by spontaneous emission decay, can be very large . There have been very few experimental studies of triplet states in LiRb , in part because they are difficult to access in thermally-distributed systems. Triplet states of bi-alkali molecules are important to study for two reasons: first, Feshbach molecules, which are triplet in nature, provide an important association gateway for the formation of stable molecules ; also, photoassociation (PA) of trapped colliding atoms is often strongest for triplet scattering states. Mixed singlet-triplet states are usually required to transfer these molecules to deeply bound singlet states. We show an abbreviated set of potential energy curves (PEC), as calculated in Ref. , in Fig. [\[fig:PEC\]](#fig:PEC){reference-type="ref" reference="fig:PEC"}. The d$^3 \Pi$-D$^1 \Pi$ complex in LiRb, asymptotic to the Li 2p $^2P_{3/2, 1/2}$ + Rb 5s $^2S_{1/2}$ free atom state, has several features that can promote its utility in stimulated-Raman-adiabatic-passage (STIRAP) and photoassociation. First, the ab inito calculations of Ref. predict mixing between low vibrational levels of the $d \ ^3\Pi_1$ and the D$^1 \Pi$ states. Second, both legs of a STIRAP transfer process from loosely bound triplet-character Feshbach molecules to the rovibronic ground state can be driven with commercially-available diode lasers. And third, similar deeply bound $^3 \Pi$ resonances have been successfully used for short-range PA in RbCs . While an interesting discovery on its own, spontaneous decay of these states after PA can populate the $a \ ^3\Sigma^+ \ v^{\prime \prime}=0$ state; one RbCs team found spontaneous decay of these states even populated the $X \ ^1\Sigma^+ \ v^{\prime \prime}=0$ state. In the present work, we study the $d \ ^3\Pi_{\Omega}$ states of LiRb, from the asymptote to the most bound vibrational level. We have found signatures of state mixing between low-lying vibrational levels of the d$^3 \Pi_1$ and D$^1 \Pi$ levels. We have determined the term energies of the different spin-orbit components of the $d \ ^3\Pi$ state, as well as their vibrational energies. We have also observed alternation of the intensities of the rotational lines, a possible indication of a $p$-wave resonance in the scattering state, and determined the rotational constants of the lowest vibrational levels. # Experiment We have previously described the details of our experimental apparatus , and provide here only a brief summary. We trap $\sim 5 \times 10^7$ Li atoms and $\sim 2 \times 10^8$ Rb atoms in a dual species magneto-optical trap (MOT), $\lesssim$`<!-- -->`{=html}1 mK in temperature and 1 mm in diameter . Our Rb MOT is a spatial dark spot MOT . We photoassociate Li and Rb atoms to form LiRb molecules using either a 300 mW cw Ti:Sapphire laser or a 150 mW cw external cavity diode laser. After spontaneous decay to a distribution of vibrational levels of the $a \ ^3\Sigma^+$ state, we use two-color resonantly-enhanced-multi-photon-ionization (RE2PI) to ionize the molecules. The lasers that we use to drive the RE2PI process are a Nd:YAG-pumped, pulsed dye laser (PDL) for the first photon, tunable in the wavelength range between 667 nm-750 nm (14950-13300 cm$^{-1}$ frequency range), and part of the 532 nm pump laser for the second photon. The repetition rate of this system is 10 Hz, and it delivers $\sim$`<!-- -->`{=html}1.5 mJ/pulse of dye energy in a 4 mm diameter beam and a larger $\sim$`<!-- -->`{=html}2 mJ/pulse 532 nm beam to the MOT region. When the frequency of the dye laser is resonant with a transition from an initial state (populated through spontaneous decay from the PA state) to an intermediate bound state (in this experiment the $d \ ^3\Pi$ state), absorption of a dye photon and a 532 nm photon ionizes the molecule. We detect the molecular LiRb$^+$ ions using a time-of-flight spectrometer and a microchannel plate detector. In this paper we will adopt the following notation: $v^{\prime \prime}$ and $J^{\prime \prime}$ denote the vibrational and rotational levels of the $a \: ^3 \Sigma ^+$ and $X \ ^1\Sigma^+$states, $v$ and $J$ (without a prime) denote the vibrational and rotational levels of the PA resonances (and for these vibrational numbers, we count down from the asymptote using negative integers), and $v^{\prime}$ and $J^{\prime}$ denote vibrational and rotational labeling of other excited electronic states. We used two techniques, RE2PI and depletion spectroscopy, to measure the $d \ ^3\Pi$ bound states. In RE2PI spectroscopy, we tune the PA laser to either the $v=-11 \ J=1$ or $v=-8 \ J=1$ lines of the $2(0^-)$ long range state, from which spontaneous decay leads primarily to the vibrational levels of the $a \ ^3\Sigma^+$ state . We count the number of ions detected over the course of 100 laser pulses, and tune the laser frequency $\nu_c$ of the PDL in 0.35 cm$^{-1}$ increments. We record the number of ions detected, normalized by the number of laser pulses, as a function of the PDL frequency $\nu_c$. In order to reach the full range of vibrational levels of the $d \ ^3\Pi$ states, we used two different laser dyes in the PDL. An LDS 698 dye covered the 13950-14950 cm$^{-1}$ range, and an LDS 750 dye covered from 13300 to 13950 cm$^{-1}$. These dyes are difficult to work with because of short lifetimes and low power output. The LDS 750 dye in particular was very troublesome: it has a lifetime $\le$`<!-- -->`{=html}8 hours, produces low power ($\le$`<!-- -->`{=html}0.5 mJ/pulse for much of its range) and because it has a very broad pulse width (i.e. lots of spontaneous emission) the baseline noise of our RE2PI spectra is enhanced over what we have observed with other dyes. The $2(0^-) \ v=-11 \ J=1$ PA line at $\nu_a = 12516.89$ cm$^{-1}$ is relatively weak, but it decays almost exclusively to a single vibrational level ($v^{\prime \prime}=11$) of the $a \ ^3\Sigma^+$ state. This facilitates straight-forward identification of the vibrational levels of the intermediate state. Unfortunately, several vibrational levels of the $d \ ^3\Pi$ state do not appear in this spectrum, presumably due to poor Franck-Condon overlap with the $v^{\prime \prime}=11$ state. This problem was even more evident when using the LDS 750 dye in the PDL. For this reason, we collected several RE2PI spectra using the stronger $2(0^-) \ v=-8 \ J=1$ PA resonance $\nu_a = 12557.60$ cm$^{-1}$. This line decays to a wider spread of vibrational levels, giving more complete coverage of the vibrational lines of the $d \ ^3\Pi$ states, but also making our analysis more difficult, due to the increased congestion of the spectra and frequent overlap between individual lines. To explore the deeply-bound levels of the $d \ ^3\Pi$ states, we used a depletion spectroscopy technique. In these measurements, we used the 150 mW ECDL tuned to the $2(0^-) \ v=-5 \ J=1$ PA resonance at $\nu_a = 12575.05$ cm$^{-1}$ . Spontaneous decay of this state populates the $a \ ^3\Sigma^+ \ v^{\prime \prime}=13$ state. We tune the PDL laser frequency to the $(3) \ ^3\Pi_{0} \ v^{\prime}=6 \leftarrow a \ ^3\Sigma^+ \ v^{\prime \prime}=13$ one-color resonantly-enhanced two-photon ionization (REMPI) transition at 17736.6 cm$^{-1}$ . We then tune the frequency of the Ti:Sapphire laser into resonance with bound-to-bound transitions from the $a \ ^3\Sigma^+ \ v^{\prime \prime}=13$ state to ro-vibrational levels in the $d \ ^3\Pi$ state. Exciting these transitions depletes the population of the $a \ ^3\Sigma^+ \ v^{\prime \prime}=13$ state, causing the REMPI signal to decrease. # RE2PI Measurements We show an example of a RE2PI spectrum in Fig. [\[fig:v11progression\]](#fig:v11progression){reference-type="ref" reference="fig:v11progression"}. Transitions observed in this spectrum are $d \ ^3\Pi_{\Omega} \ v^{\prime} \leftarrow a \ ^3\Sigma^+ \ v^{\prime \prime}=11$. We have marked the transitions to the $\Omega=2$ progression with black solid lines, $\Omega=1$ with blue dashed lines, and $\Omega=0$ with green dot-dashed lines. $\Omega$ is the total electronic angular momentum, orbital $L$ + spin $S$, projected onto the internuclear axis. The numerical label for each peak is the vibrational number $v^{\prime}$ of the $d \ ^3\Pi_{\Omega}$ state. We have also marked three lines in this spectrum corresponding to transitions to the $D \ ^1\Pi$ state with red dotted lines. From the spectrum of these $d \ ^3\Pi$ states, we observe the typical hierarchy of line spacings: the vibrational splitting is large (on the order of 100 cm$^{-1}$ for low vibrational quantum number $v^\prime$, and decreasing with increasing $v^\prime$), and the spin-orbit splitting between different $\Omega$ states is smaller (on the order of 30 cm$^{-1}$). The rotational splitting for low $J^\prime$ (on the order of 0.1 cm$^{-1}$) is too small to be resolved in these RE2PI spectra since the spectral resolution of the PDL is $\sim$`<!-- -->`{=html}0.5 cm$^{-1}$. The appearance of transitions belonging to vibrational levels of the $D \ ^1\Pi$ electronic state in Fig. [\[fig:v11progression\]](#fig:v11progression){reference-type="ref" reference="fig:v11progression"} is evidence of mixing between the $D \ ^1\Pi$ and the $d \ ^3\Pi_{1}$ potentials near an avoided crossing between the two states. The energy of these $D \ ^1\Pi$ states is known from Refs. . State mixing gives these states partial character of each electronic state, which in this case manifests itself through strong transitions from a triplet state (i.e. $a \ ^3\Sigma^+ \ v^{\prime \prime}=11$) to singlet states ($D \ ^1\Pi \ v^{\prime}$ = 6, 7, and 8). This state mixing also adds $D ^1\Pi$ character to the $d \ ^3\Pi_{1}$ states, so one should expect the nearby $d \ ^3\Pi_{1}$ states to appear in the singlet spectra. This expectation is borne out in the spectrum shown in Fig. [\[fig:v10progression\]](#fig:v10progression){reference-type="ref" reference="fig:v10progression"}. This spectrum is a REMPI scan generated in our system after photoassociating ultracold LiRb molecules through a 2(1)-4(1) mixed state at $\nu_a = 12574.85$ cm$^{-1}$ , which spontaneously decays to vibrational levels of the $X \ ^1\Sigma^+$ ground electronic state. The spectrum in Fig. [\[fig:v10progression\]](#fig:v10progression){reference-type="ref" reference="fig:v10progression"} primarily shows transitions to low-lying $D \ ^1\Pi$ vibrational levels from $X \ ^1\Sigma^+ \ v^{\prime \prime}=10$. We also observe in this spectrum $d \ ^3\Pi_{1} \ v^{\prime}=4$, 5, and 6 $\leftarrow X \ ^1\Sigma^+ \ v^{\prime \prime}=10$ transitions. We chose $X \ ^1\Sigma^+ \ v^{\prime \prime}=10$ because it is strongly populated by spontaneous decay of the 2(1)-4(1) PA resonance and transitions to deeply bound $D \ ^1\Pi$ vibrational levels are clearly identified. We can estimate the degree of mixing based on the relative strength of the different REMPI peaks. The $D \ ^1\Pi \ v^\prime = 7 \leftarrow X \ ^1\Sigma^+ \ v^{\prime \prime}=10$ transition is twice as strong as the $d \ ^3\Pi_1 \ v^\prime = 5 \leftarrow X \ ^1\Sigma^+ \ v^{\prime \prime}=10$ transition so there is twice as much singlet character to $D \ ^1\Pi \ v^\prime = 7$ as $d \ ^3\Pi_1 \ v^\prime = 5$. Following the same procedure that we have used in the past , we can estimate the interaction strength to be about 7 cm$^{-1}$. Interestingly, this rough estimate is consistent with the following simple perturbative argument. The spin-orbit interaction responsible for the state mixing can be estimated as about one half the spin orbit mixing in atomic rubidium , or about 120 cm$^{-1}$. The Franck-Condon factor (FCF) between $D \ ^1\Pi \ v^\prime = 7$ and $d \ ^3\Pi_1 \ v^\prime = 5$, as calculated by LEVEL 8.0 using the PEC from , is about 0.08 and thus the strength of interaction between these states should be approximately 10 cm$^{-1}$. We have applied this perturbative analysis to each of the vibrational levels of the $d \ ^3\Pi_{1}$ state (not including $v^{\prime}=5$), and find that each contains some small component of $D \ ^1\Pi$ perturber state, on the order of 10% or smaller. This is too small to be seen in the spectra of Fig. [\[fig:v11progression\]](#fig:v11progression){reference-type="ref" reference="fig:v11progression"}, but could be sufficient to be useful in a Raman or STIRAP transfer of population to low lying levels of the electronic ground state in the future. Many of the RE2PI spectra that we collected are less clear than that shown in Fig. [\[fig:v11progression\]](#fig:v11progression){reference-type="ref" reference="fig:v11progression"}. In particular, the peaks in the RE2PI spectra near the Rb 5S + Li 2P asymptote are strong, but line congestion becomes significant, and clear identification of the lines in this region becomes difficult. These assignments could probably be improved using a spectroscopic technique that is capable of higher spectral resolution, such as photoassociative spectroscopy, but this was beyond the scope of the present work. Assigning peaks in the RE2PI spectra was equally difficult for deeply bound vibrational states (i.e. $v^{\prime}$ $\le$ 4). In fact we were unable to observe a clear cut off in our RE2PI data corresponding to $v^{\prime}=0$. We attribute this to difficulties with the LDS 750 dye, specifically the large spontaneous emission content in the pulse. To rectify this problem, we turned to a second set of measurements, based upon depletion spectroscopy. # Depletion Measurements We used depletion spectroscopy to identify the lowest two vibrational levels of the $d \ ^3\Pi$ state. Compared to RE2PI, depletion spectra are sparse and have narrow peak widths, in this case $\sim$`<!-- -->`{=html}1 GHz, a typical linewidth for this type of measurement at this intensity . Additionally, depletion spectra allow us to extract some rotational constants of the lines, a very useful tool for comparing experiment to theory. We show depletion spectra for $v^{\prime}=0$ in Fig. [\[fig:depletion\]](#fig:depletion){reference-type="ref" reference="fig:depletion"}. To assign these data care must be taken with selection rules for radiative transitions in molecules. The two that apply here are: $\Delta J= 0, \ \pm1$ and $-\leftrightarrow +$, that is positive symmetry states (with respect to coordinate inversion) must transition to negative symmetry states and vise versa. The initial state in this depletion transition is a $^3\Sigma^+$ state, which is a strict Hund's case (b) state. As such, its rotational energy is determined by quantum number $K$, which designates the total angular momentum of the molecule apart from spin, rather than the total angular momentum (including spin) quantum number $J$. For this $a ^3\Sigma^+$state, the electronic spin is $S$=1, and levels with $J=K+1, K$ and $K-1$, are nearly degenerate for $K\ge 1$. Additionally $K$ determines the symmetry of the state. This is summarized in Fig. [\[fig:parity\]](#fig:parity){reference-type="ref" reference="fig:parity"}, adapted and modified from Ref. . The logic that leads to our assignments in Fig. [\[fig:depletion\]](#fig:depletion){reference-type="ref" reference="fig:depletion"} goes as follows. We start with the depletion data on transitions to $\Omega=1$ shown in Fig. [\[fig:depletion\]](#fig:depletion){reference-type="ref" reference="fig:depletion"}(c), which is extensive enough to show that we do not populate $J^{\prime \prime}=2$ of the $a \ ^3\Sigma^+$ state; that is, we see no depletion signal corresponding to a transition to $J^{\prime} = 3$. Since we do see transitions to $J^{\prime} = 1$ and $2$, we know that we populate some mix of $J^{\prime \prime}=0$ and $1$ belonging to either $K^{\prime \prime}=0$, 1, or 2. $K^{\prime \prime}=0$ and $K^{\prime \prime}=2$ have the same symmetry and some of the same rotational numbers; this implies that any PA state that could decay to $K^{\prime \prime}=0$ or 2 could decay to the other as well. Because the energies of $K^{\prime \prime}=0$ and $K^{\prime \prime}=2$ differ from one another, but we see no additional structure in the depletion spectra, we infer that the transitions seen in Fig. [\[fig:depletion\]](#fig:depletion){reference-type="ref" reference="fig:depletion"}(c) originate from a single state, that is $K^{\prime \prime}=1$. The spacing between the two peaks in this spectrum should be $4B_0$, where $B_0$ is the rotational constant of the $d \ ^3\Pi \ v^\prime = 0$ state, allowing us to determine $B_0 = 4.59$ GHz. This rotational constant agrees with the prediction from LEVEL 8.0 with PECs from Ref. , which confirms our assignments and guides our interpretation of the $\Omega=0$ data shown in Fig. [\[fig:depletion\]](#fig:depletion){reference-type="ref" reference="fig:depletion"}(d), in which we see three transitions. The first two are spaced by $6B_0$ (27.5 GHz), implying that these peaks are transitions to the positive symmetry levels of the $\Omega=0^+$ electronic state, $J^{\prime}=0 \leftarrow K^{\prime \prime}=1$ and $J^{\prime}=2 \leftarrow K^{\prime \prime}=1$. The absence of a peak for $J^{\prime}=1$ is consistent with the selection rule $-\not\leftrightarrow-$. The remaining transition in $\Omega=0$ must be the only allowed transition to $\Omega=0^-$, that is $J^{\prime}=1 \leftarrow K^{\prime \prime}=1$. There is only one transition in $\Omega=2$, which is trivial to identify as $J^{\prime}=2 \leftarrow K^{\prime \prime}=1$. This picture is consistent with what we know about our PA state. That is, we PA into the $J=1$ level of a $2(0^-)$ state, which has positive symmetry, and the allowed decay paths are to $K^{\prime \prime}=1$ (with $J^{\prime \prime}=0$, $1$, and $2$, each with negative symmetry, of which we populate $0$ and $1$). Of note, we chose to PA to $J=1$ because it is the strongest PA resonance in the $2(0^-) \ v=-5$ progression. Our depletion data confirms that the parity of this PA state is even, which is suggestive of a scattering state $p$-wave shape resonance . Similar shape resonances have been seen in other bi-alkali's . We used these data to determine the spin-orbit splitting between the different $\Omega$ progressions deep in the $d \ ^3\Pi$ well, and followed these progressions back to the asymptote in our RE2PI data. Most importantly, these data provide accurate locations of $v^{\prime}=0$ and $1$, $J^{\prime}$ for future work on short range PA . The depletion data is limited by the shot noise in our ion counting. For each data point we integrate 200 shots and usually count $\sim$`<!-- -->`{=html}350 ions. To get 2$\sigma$ resolution the smallest depletion feature that we can be confident in has to be a 10% decrease. Here our on-resonance depletion signal drops by around 30% which is 6$\sigma$ and statistically significant. Unfortunately, our Ti:Sapphire struggles to tune to wavelengths much shorter than 740 nm, which limited our depletion spectra to the lowest two vibrational levels only. Despite these short comings, depletion spectroscopy gives us accurate identifications of these bottom two vibrational levels of the $d \ ^3\Pi$ electronic state and unambiguously determines the spin orbit splitting. # Discussion We determine the vibrational binding energies and rotational constants of the states seen in depletion spectroscopy, which we tabulate in Table. [\[tab:depletionAssignments\]](#tab:depletionAssignments){reference-type="ref" reference="tab:depletionAssignments"}. We list in Table [\[tab:masterAssignments\]](#tab:masterAssignments){reference-type="ref" reference="tab:masterAssignments"} the assignments and energy of each of the $d \ ^3\Pi$ states that we observe through RE2PI and depletion spectroscopy. We also include in this table the energy difference between adjacent states, which aids in the assignment of the lines. The theoretical vibrational levels and spin-orbit splittings that we used to guide our work and for comparison of results come from ab initio calculations by Korek et al. with aid from LEVEL 8.0 . We found good overall agreement with these ab initio results. The spin-orbit splittings for $\Omega=0$ to $\Omega=1$, predicted to be 21 cm$^{-1}$, are measured here to be 37 cm$^{-1}$; for $\Omega=1$ to $\Omega=2$, they are predicted to be 38 cm$^{-1}$, and we found them to be 33 cm$^{-1}$. For the spin-orbit splitting between the $\Omega=0^{+}$ to $\Omega=0^{-}$ states, however, we observe 0.9 cm$^{-1}$, significantly less than the predicted 36 cm$^{-1}$. Our depletion data is unambiguous in establishing the $\Omega=0^{+}$ to $\Omega=0^{-}$ splitting, and a small $\Omega=0^{+}$ to $\Omega=0^{-}$ splitting is consistent with our observations in the $(3) \ ^3\Pi$ state . We show the vibrational spacing, $\Delta E = E_{v+1}-E_v$ vs $v$, of the different series in Fig. [\[fig:vibrationalSplitting\]](#fig:vibrationalSplitting){reference-type="ref" reference="fig:vibrationalSplitting"}. These data are in reasonable agreement with the predicted vibrational splittings although there appears to be a nearly uniform difference of a few cm$^{-1}$. We found that the depth of the $d \ ^3\Pi$ potential (exp. value) is less than that predicted (th. value). We looked extensively for another vibrational level below our assigned $v^{\prime}=0$ level. We covered $\pm$ 10 cm$^{-1}$ around the expected vibrational location with our depletion spectra, but found no indication of a depletion resonance. Our extracted molecular constants are listed in Tab. [\[tab:molecularConstants\]](#tab:molecularConstants){reference-type="ref" reference="tab:molecularConstants"}. These provide an easy estimation of the spectral structure of the $d \ ^3\Pi$ states as well as a quick comparison to theoretical predictions. As borne out in Fig. [\[fig:vibrationalSplitting\]](#fig:vibrationalSplitting){reference-type="ref" reference="fig:vibrationalSplitting"}, there is good agreement between our fitted harmonic constant, $\omega_e$, and the predictions. However, there is considerably less agreement between our extracted term energy, $T_e$, and the predictions for reasons discussed previously. Additionally, it is important to note that when we fitted the experimental data to determine $T_e$, $\omega_e$, $x_e$ and $y_e$ we used only $v^{\prime}$=0-19. This increased the accuracy of the fit so that for these vibrational levels our molecular constants reproduce our data with a standard deviation of 2 cm$^{-1}$. We believe most of the deviations are caused by experimental uncertainties on determining the frequencies of the peaks, as well as small perturbations to state locations caused by spin-orbit mixing. # Conclusion In this work we have found and identified the $v=0$ to $v=22$ states of the $d \ ^3\Pi_{\Omega}$ electronic state in LiRb. We explored singlet-triplet mixing to evaluate the possibility of using the $d \ ^3\Pi$-$D \ ^1\Pi$ mixed states as the intermediate state in a STIRAP transfer from a weakly bound triplet state to deeply bound singlet states. We know from heat pipe spectra and other REMPI data that the transitions from deeply bound $D \ ^1\Pi \ v^{\prime}$ to deeply bound $X \ ^1\Sigma^+$ states are strong. Our depletion data shows that transitions from weakly bound triplet molecules to deeply bound $d \ ^3\Pi_1 \ v^\prime$ states are also strong. Finally, our RE2PI data demonstrate that there is about 10% mixing between most of these singlet and triplet states. We suggest using the $d \ ^3\Pi_1 \ v^\prime = 0 \ J^\prime = 1$ state as the intermediate state for a STIRAP transfer in future work. The laser wavelengths for this transfer would be 740 nm, achievable with Ti:Sapphire or diode lasers, and 516 nm, which is accessible with green laser diodes. Using calculated FCFs and an estimate of the transition dipole moment of a few times $e a_0$, the 'up' transition would be around $4 \times 10^{-2}$ $e a_0$ and the 'down' transition would be around $10^{-2}$ $e a_0$ which is competitive with the transfer strength used by the KRb JILA team . Using the binding energy of the lowest several vibrational levels we have assessed their potential for short-range photoassociation. There are only six possible PA transitions we can observe with our current Ti:Sapphire laser, the same six we saw in depletion. However, our depletion data provides the perfect stepping stone for this type of work, providing very trusted locations of vibrational levels (to within 0.5 cm$^{-1}$, the uncertainty in the binding energy of $a \ ^3\Sigma^+ \ v^{\prime \prime}=13$). Of particular interest to us is the combination of short range PA and mixing between $d \ ^3\Pi$-$D \ ^1\Pi$. This provides an avenue through simple spontaneous decay to the rovibronic ground state although we will need a different laser to access these PA resonances. We are happy to acknowledge useful conversations with Jesús Pérez-Rı́os, and university support of this work through the Purdue OVPR AMO incentive grant. And we would like to acknowledge the work done by S. Dutta and J. Lorenz in building the LiRb machine.	{'timestamp': '2016-07-18T02:09:11', 'yymm': '1607', 'arxiv_id': '1607.04608', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04608'}
# Introduction {#sec:introduction} The use of GPS for localizing sensor nodes in a sensor network is considered to be excessively expensive and wasteful, also in some cases intractable,. Instead many solutions for the localization problem tend to use inter-sensor distance or range measurements. In such a setting the localization problem is to find unknown locations of say $N$ sensors using existing noisy distance measurements among them and to sensors with known locations, also referred to as anchors. This problem is known to be NP hard, and there have been many efforts to approximately solve this problem,. One of the major approaches for approximating the localization problem, has been through the use of convex relaxation techniques, namely semidefinite, second-order and disk relaxations, see e.g.,. Although the centralized algorithms based on the these approximations reduce the computational complexity of solving the localization problem, they are still not scalable for solving large problems. Also centralized algorithms are generally communication intensive and more importantly lack robustness to failures. Furthermore, the use of these algorithms can become impractical due to certain structural constraints resulting from, e.g., privacy constraints and physical separation. These constraints generally prevent us from forming the localization problem in a centralized manner. One of the approaches to evade such issues is through the use of scalable and/or distributed algorithms for solving large localization problems. These algorithms enable us to solve the problem through collaboration and communication of several computational agents, which could correspond to sensors, without the need for a centralized computational unit. The design of distributed localization algorithms is commonly done by first reformulating the problem by exploiting or imposing structure on the problem and then employing efficient optimization algorithms for solving the reformulated problem, see e.g., some recent papers. For instance, authors in put forth a solution for the localization problem based on minimization the discrepancy of the squared distances and the range measurements. They then propose a second-order cone relaxation for this problem and apply a Gauss-Seidel scheme to the resulting problem. This enables them to solve the problem distributedly. The proposed algorithm does not provide a guaranteed convergence and at each iteration of this algorithm, each agent is required to solve a second-order cone program, SOCP, which can potentially be expensive. Furthermore, due to the considered formulation of the localization problem, the resulting algorithm is prone to amplify the measurement errors and is sensitive to outliers. In, the authors consider an SDP relaxation of the maximum likelihood formulation of the localization problem. They further relax the problem to an edge-based formulation as suggested in. This then allows them to devise a distributed algorithm for solving the reformulated problem using alternating direction method of multipliers (ADMM). Even though this algorithm has convergence guarantees, each agent is required to solve an SDP at every iteration of the algorithm. In order to alleviate this, authors in and consider a disk relaxation of the localization problem and which correspond to an under-estimator of the original problem. They then use projection-based methods and Nestrov's optimal gradient method, respectively, for devising distributed algorithms for solving the resulting problem. These algorithms rely on finding a solution that lies in the intersection of the disks or spheres defined by the range measurements. Consequently, the computational demand on each agent for these algorithms is far less than the aforementioned algorithms. These algorithms commonly work well when there are many range measurements available and their performance is adversely affected if the number of measurements are decreased. Moreover, for the case of low quality, particularly biased, measurements, the convergence of the algorithms can be interrupted as the intersection can be empty. The proposed algorithms in the aforementioned papers have been shown to be effective in analyzing large-scale localization problems. However, all these methods rely on first-order optimization algorithms and hence can require many iterations and communications to converge to an accurate enough solution. Furthermore, the number of iterations can vary significantly with different realizations of range measurements and changing topology of the sensor network. In this paper we show that in case it is possible to provide a tree representation of the inter-sensor range measurement graph of the sensor network (which is the case in many scenarios with few available range measurements), it is possible to alleviate these issues by devising far more efficient distributed localization algorithms that purely rely on second-order methods. ## Outline {#outline .unnumbered} In Section [2](#sec:MLL){reference-type="ref" reference="sec:MLL"} we review a maximum-likelihood formulation of the localization problem. Section [3](#sec:scattered){reference-type="ref" reference="sec:scattered"} provides a formal description of tree-structured scattered sensor networks and describes how the structure in the problem can be reflected in the localization optimization problem. Section [4](#sec:chordal){reference-type="ref" reference="sec:chordal"} reviews how certain structure in nonlinear SDPs enable us to utilize domain-space decomposition to decompose them. This decomposition technique is then used in Section [5](#sec:decompositionloc){reference-type="ref" reference="sec:decompositionloc"} to decompose the localization optimization problem. In this section we also describe how the decomposed problem can be written as coupled SDP. We then put forth a generic description of primal-dual interior-point methods in Section [6](#sec:DPDIPM){reference-type="ref" reference="sec:DPDIPM"} and show how they, in combination with message-passing, can be used to devise efficient distributed solvers for the localization problems. In this section we also discuss the computational and communication complexity of the proposed distributed algorithm. The numerical experiments are presented in Section [7](#sec:numerical){reference-type="ref" reference="sec:numerical"}, and we conclude the paper with final remarks in Section [8](#sec:conclusions){reference-type="ref" reference="sec:conclusions"}. ## Notations and Definitions {#sec:notation .unnumbered} We denote by $\mathbb R$ the set of real scalars and by $\mathbb R^{n\times m}$ the set of real $n\times m$ matrices. The set of $n \times n$ symmetric matrices are represented by $\mathbf S^n$. The transpose of a matrix $A$ is denoted by $A^T$ and the column and null space of this matrix is denoted by $\mathcal{C}(A)$ and $\mathcal N(A)$, respectively. We denote the set of positive integers $\{1,2,\ldots,p\}$ with $\mathbb{N}_p$. Given a set $J \subset \mathbb{N}_n$, the matrix $E_J \in \mathbb{R}^{\|J\|\times n}$ is the $0$-$1$ matrix that is obtained by deleting the rows indexed by $\mathbb{N}_n \setminus J$ from an identity matrix of order $n$, where $\|J\|$ denotes the number of elements in set $J$. This means that $E_Jx$ is a $\|J\|$-dimensional vector with the components of $x$ that correspond to the elements in $J$, and we denote this vector with $x_J$. Also $e_j$ denotes a 0--1 $n$-dimensional vector with only a nonzero element at the $j$th component. Similarly, given $J \subset \mathbb N_n$, $e_J$ denotes a 0--1 $n$-dimensional vector with ones at elements specified by $J$. With $x^{i,(k)}_l$ we denote the $l$th element of vector $x^i$ at the $k$th iteration. Also given vectors $x^i$ for $i= 1, \dots, N$, the column vector $(x^1, \dots, x^N)$ is all of the given vectors stacked. For a vector $x$, with $\diag(x)$ we denote a diagonal matrix with its diagonal elements given by $x$. Similarly, given matrices $X^i$ for $i = 1, \dots, N$, with $\blkdiag(X^1, \dots, X^N)$ we denote a block-diagonal matrix with diagonal blocks given by each of the given matrices. For a matrix $X \in \mathbb R^{m\times n}$, $\vectorize(X)$ is an $mn$-dimensional vector that is obtained by stacking all columns of $X$ on top of each other. Given a symmetric matrix $X \in \mathbf S^n$ \[\begin{gathered} \svec(X) := (X_{11}, \sqrt{2}X_{21}, \dots, \sqrt{2} X_{n1}, X_{22},\\ \sqrt{2} X_{32}, \dots, \sqrt{2} X_{n2}, \dots, X_{nn}). \end{gathered}\] Also for a square matrix $X \in \mathbb R^{n\times n}$ we denote with $\vectri(X)$ a column vector which includes all elements on the upper triangle of $X$ stacked. Given two matrices $X$ and $Y$ by $X\otimes Y$ we denote the standard Kronecker product. Given $X \in \mathbf S^n$, define $U$ as an $n(n + 1)/2 \times n^2$ matrix such that $U \vectorize(X) = \svec(X)$. Then for two matrices $X, Y \in \mathbb R^{n\times n}$, $\otimes_s$ denotes the symmetrized Kronecker product that is defined as \[\begin{aligned} X \otimes_s Y := \frac{1}{2} U(X \otimes Y + Y \otimes X)U^T. \end{aligned}\] For properties of the symmetrized Kronecker product refer to. A graph is denoted by $Q(V,\mathcal E)$ where $V = \{1, \dots, n\}$ is its set of vertices or nodes and $\mathcal E \subseteq V\times V$ denotes its set of edges. Vertices $i, j \in V$ are adjacent if $(i, j) \in E$, and we denote the set of adjacent vertices of $i$ by $\Ne(i) = \{ j \in V \| (i, j) \in\mathcal E \}$. A graph is said to be complete if all its vertices are adjacent. An induced graph by $V^\prime \subseteq V$ on $Q(V,\mathcal E)$, is a graph $Q_I(V^\prime,\mathcal E^\prime)$ where $\mathcal E^\prime = \mathcal E\cap V^\prime \times V^\prime$. A clique $C_i$ of $Q(V,\mathcal E)$ is a maximal subset of $V$ that induces a complete subgraph on $Q$, i.e., no clique is properly contained in another clique,. Assume that all cycles of length at least four of $Q(V,\mathcal E)$ have a chord, where a chord is an edge between two non-consecutive vertices in a cycle. This graph is then called chordal. It is possible to make graphs chordal by adding edges to the graph. The resulting graph is then referred to as a chordal embedding. Let $\mathbf C_Q = \{ C_1, \dots, C_q \}$ denote the set of its cliques, where $q$ is the number of cliques of the graph. Then there exists a tree defined on $\mathbf C_Q$ such that for every $C_i, C_j \in\mathbf C_Q$ where $i \neq j$, $C_i \cap C_j$ is contained in all the cliques in the path connecting the two cliques in the tree. This property is called the clique intersection property,. Trees with this property are referred to as clique trees. # Maximum Likelihood Localization {#sec:MLL} In this paper we consider a localization problem for a network of $N$ sensors distributed in an area in presence of $m$ anchors. The exact locations of these sensors, $x^i_s$, are deemed to be unknown however we assume that the positions of the anchors, $x^i_a$, are given. Furthermore, the sensors are capable of performing computations and some can measure their distance to certain sensors and some of the anchors. We assume that if sensor $i$ can measure its distance to sensor $j$ so can sensor $j$ measure its distance to sensor $i$. This then allows us to describe the range measurement availability among sensors using an undirected graph $G_r(V_r, \mathcal E_r)$ with vertex set $V_r = \{ 1, \dots, N \}$ and edge set $\mathcal E_r$. An edge $(i,j) \in \mathcal E_r$ if and only if a range measurement between sensors $i$ and $j$ is available. We refer to this graph as inter-sensor measurement graph and assume that it is connected. Let us define the set of neighbors of each sensor $i$, $\text{Ne}_r(i)$, as the set of sensors to which this sensor has an available range measurement. In a similar fashion let us denote the set of anchors to which sensor $i$ can measure its distance to by $\text{Ne}_a(i) \subseteq \{ 1, \dots, m\}$. Let us describe the inter-sensor range measurements for each sensor, $i \in \mathbb N_N$, as \[\begin{aligned} \label{eq:InterSensorMeasurement} \mathcal R_{ij} = \mathcal D_{ij} + E_{ij}, \quad j \in \text{Ne}_r(i), \end{aligned}\] where $\mathcal D_{ij} = \\| x_s^i-x_s^j \\|_2$ defines the noise-free sensor distance, $E_{ij}$ is the inter-sensor measurement noise and $E_{ij} \sim P_{ij}^s(\mathcal D_{ij}\|\mathcal R_{ij})$ with $P_{ij}^s(\cdot)$ being the so-called inter-sensor sensing probability density function (PDF). We here make the standard assumption that $\mathcal R_{ij} = \mathcal R_{ji}$, see e.g.,. Similarly we can describe the anchor range measurements for each sensor $i$ as \[\begin{aligned} \label{eq:AnchorSensorMeasurement} \mathcal Y_{ij} = \mathcal Z_{ij} + V_{ij}, \quad j \in \text{Ne}_a(i), \end{aligned}\] where $\mathcal Z_{ij} = \\| x_s^i-x_a^j \\|_2$ defines the noise-free anchor-sensor distance, $V_{ij}$ is the anchor-sensor measurement noise and $V_{ij} \sim P_{ij}^a(\mathcal Z_{ij}\|\mathcal Y_{ij})$ with $P_{ij}^a(\cdot)$ being the so-called anchor-sensor sensing PDF. Here we assume that the inter-sensor and anchor-sensor measurement noise PDFs, i.e., $P_{ij}^s(\cdot)$ and $P_{ij}^a(\cdot)$, respectively, are Gaussian. Particularly, we assume that the inter-sensor and anchor-sensor measurement noises are independent and that $E_{ij} \sim \mathcal N(0, \Sigma^r_{ij})$ and $V_{ij} \sim \mathcal N(0, \Sigma^a_{ij})$. Notice that this assumption can be relaxed to any distribution that is a log-concave function of distances $\mathcal D_{ij}$ and $\mathcal Z_{ij}$, however, for the sake of brevity we limit ourselves to the case of Gaussian distributions. Having defined the setup of the sensor network, we can write the localization problem in a maximum likelihood setting as \[\begin{gathered} \label{eq:MLOriginal} X_{\text{ML}}^\ast = \argmin_{X} \Bigg\{ \sum_{i = 1}^N\Bigg( \sum_{\tiny\begin{split}j \in &\text{Ne}_r(i)\\ i &< j\end{split}\normalsize} \frac{1}{\Sigma^r_{ij}}\left( \mathcal D_{ij}(x^i_s, x^j_s)-\mathcal R_{ij}\right)^2 \\ + \sum_{j \in \text{Ne}_a(i)} \frac{1}{\Sigma^a_{ij}}\left(\mathcal Z_{ij}(x^i_s, x^j_a)-\mathcal Y_{ij}\right)^2 \Bigg) \Bigg\}, \end{gathered}\] where $X = \begin{bmatrix} x^1_s & \dots & x^N_s \end{bmatrix}\in \mathbb R^{d\times N}$ with $d = 2$ or $d = 3$. This problem can be formulated as a constrained optimization problem, as was described in, which is discussed next. First let us define the function \[\begin{gathered} f(\Lambda, \Xi, D, Z): = \sum_{i = 1}^N \Bigg(\sum_{\tiny\begin{split}j \in &\text{Ne}_r(i)\\ i &< j\end{split}\normalsize}\frac{1}{\Sigma^r_{ij}} (\Lambda_{ij}-2D_{ij}\mathcal R_{ij} +\mathcal R_{ij}^2) \\ + \sum_{j \in \text{Ne}_a(i)}\frac{1}{\Sigma^a_{ij}} (\Xi_{ij}-2Z_{ij}\mathcal Y_{ij} +\mathcal Y_{ij}^2) \Bigg). \end{gathered}\] Then the problem in [\[eq:MLOriginal\]](#eq:MLOriginal){reference-type="eqref" reference="eq:MLOriginal"} can be equivalently rewritten as the following constrained optimization problem So far we have reviewed a way to formulate the localization problem over general sensor networks as a constrained optimization problem. In this paper, however, we are particularly interested in localization problem pertaining to sensor networks with an inherent tree structure which relies on the assumption that the graph $G_r(V_r, \mathcal E_r)$ can be represented using a tree. We describe the localization problem of such networks in the next section. # Localization of Tree-structured Scattered Sensor Networks {#sec:scattered} Let the graph $G_r(V_r, \mathcal E_r)$ be connected with few edges. Also assume that a chordal embedding $\bar G_r(V_r, \bar{\mathcal E}_r)$ of this graph can be achieved by adding only a few edges. This graph can then be represented using its clique tree. Furthermore, given the set of its cliques $\mathbf C_{\bar G_r} = \{ C_1, \dots, C_q \}$, we have $\| C_i \| \ll N$. We refer to such sensor networks as tree-structured scattered. The localization problem of these sensor networks can also be formulated as a constrained optimization problem using the approach discussed in Section [\[sec:notation\]](#sec:notation){reference-type="ref" reference="sec:notation"}. However, the formulation of the problem in [\[eq:MLConstrained\]](#eq:MLConstrained){reference-type="eqref" reference="eq:MLConstrained"} is not fully representative of the structure in the problem. In order to exploit the structure in our localization problem we modify [\[eq:MLConstrained\]](#eq:MLConstrained){reference-type="eqref" reference="eq:MLConstrained"}, and equivalently rewrite it as Note that, here, we have modified the constraint in [\[eq:MLConstrained-d\]](#eq:MLConstrained-d){reference-type="eqref" reference="eq:MLConstrained-d"} so that the structure in the problem is more explicit. This modification is based on the observation that not all the elements of $S$ are used in [\[eq:MLConstrained-b\]](#eq:MLConstrained-b){reference-type="eqref" reference="eq:MLConstrained-b"} and [\[eq:MLConstrained-c\]](#eq:MLConstrained-c){reference-type="eqref" reference="eq:MLConstrained-c"}, and hence we only have to specify the ones that are needed and can leave the rest free. In, , the authors first conduct a semidefinite relaxation on [\[eq:MLConstrained\]](#eq:MLConstrained){reference-type="eqref" reference="eq:MLConstrained"}. They then exploit the structure as we did in [\[eq:MLConstrainedScattered\]](#eq:MLConstrainedScattered){reference-type="eqref" reference="eq:MLConstrainedScattered"} and use the ideas in to devise efficient centralized solvers for the localization problem. Here, however, we stick to the formulation in [\[eq:MLConstrainedScattered\]](#eq:MLConstrainedScattered){reference-type="eqref" reference="eq:MLConstrainedScattered"} which is a nonlinear SDP, and use scheme in to decompose this problem directly. We then perform a semidefinite relaxation on the resulting problem and rewrite the problem as a coupled SDP. This in turn facilitates the use of efficient scalable or distributed solvers. The use of the so-called domain-space decomposition presented in is at the heart of this reformulation approach. We review this decomposition scheme next, for the sake of completeness. # Chordal Sparsity in Semidefinite Programs {#sec:chordal} In this section we first briefly review some of important properties of sparse semidefinite matrices and then discuss how these can be used for reformulating semidefinite programs with chordal sparsity suitable to be solved distributedly. ## Chordal Sparsity {#sec:sparsity} Graphs can be used to characterize partial symmetric matrices. Partial symmetric matrices correspond to symmetric matrices where only a subset of their elements are specified and the rest are free. We denote the set of all $n \times n$ partially symmetric matrices on a graph $Q(V,\mathcal E)$ by $\mathbf S_Q^n$, where only elements with indices belonging to $\mathbf I_s = \mathcal E \cup \{ (i,i) \ \| \ i \in \mathbb N_n\}$ are specified. Now consider a matrix $X \in \mathbf S_Q^n$. Then $X$ is positive semidefinite completable if by manipulating its free elements, i.e., elements with indices belonging to $\mathbf I_f = (V \times V) \setminus \mathbf I_s$, we can generate a positive semidefinite matrix. The following theorem states a fundamental result on positive semidefinite completion. Note that the matrices $X_{C_iC_i}$ for $i \in \mathbb N_q$, are the fully specified principle submatrices of $X$. Hence, Theorem [\[thm:NSDC\]](#thm:NSDC){reference-type="ref" reference="thm:NSDC"} states that a chordal matrix $X \in \mathbf S_{Q}^n$ is positive semidefinite completable if and only if all its fully specified principle submatices are positive semidefinite. As we will see next this property can be used for decomposing SDPs with this structure. ## Domain-space Decomposition {#sec:decomposition} Consider a chordal graph $Q(V,\mathcal E)$, with $\{ C_1, \dots, C_q \}$ the set of cliques such that the clique intersection property holds. Let us define sets $J_i \subset \mathbb N_n$ such that the sparsity pattern graph for $\sum_{i=1}^N e_{J_i}e_{J_i}^T$ is $Q(V,\mathcal E)$. Then for the following nonlinear SDP the only elements of $X$ that affect the cost function in [\[eq:SDP1\]](#eq:SDP1){reference-type="eqref" reference="eq:SDP1"} and the constraint in [\[eq:SDP2\]](#eq:SDP2){reference-type="eqref" reference="eq:SDP2"} are elements specified by indices in $\mathbf I_s$. Using Theorem [\[thm:NSDC\]](#thm:NSDC){reference-type="ref" reference="thm:NSDC"}, the optimization problem in [\[eq:SDP\]](#eq:SDP){reference-type="eqref" reference="eq:SDP"} can then be equivalently rewritten as where notice that the constraints in [\[eq:MSDP1-3\]](#eq:MSDP1-3){reference-type="eqref" reference="eq:MSDP1-3"} are coupled semidefinite constraints,. It is possible to explicitly describe the coupling using consistency constraints and rewrite [\[eq:MSDP1\]](#eq:MSDP1){reference-type="eqref" reference="eq:MSDP1"} as where $X^i \in \mathbf S^{\|C_i\|}$. This method of reformulating [\[eq:SDP\]](#eq:SDP){reference-type="eqref" reference="eq:SDP"} as [\[eq:MSDP\]](#eq:MSDP){reference-type="eqref" reference="eq:MSDP"} is referred to as the domain-space decomposition,. The structure in the localization of tree-structured scattered sensor networks enable us to use this technique for reformulating the problem in such a way that would better facilitate the use of efficient distributed solvers. This is discussed in the next section. # Decomposition and Convex Formulation of Localization of Tree-structured Scattered Sensor Networks {#sec:decompositionloc} Consider the inter-sensor measurement graph $G_r(V_r, \mathcal E_r)$, and assume that it is chordal. In case this graph is not chordal the upcoming discussions hold for any of its chordal embeddings. Let $\mathbf C_{G_r} = \{ C_1, \dots, C_q \}$ and $T(V_t,\mathcal E_t)$ be a clique tree. Based on the discussion in Section [4.2](#sec:decomposition){reference-type="ref" reference="sec:decomposition"}, then for the problem in [\[eq:MLConstrainedScattered\]](#eq:MLConstrainedScattered){reference-type="eqref" reference="eq:MLConstrainedScattered"} we have $S \in \mathbf S^N_{G_r}$. Hence, we can rewrite [\[eq:MLConstrainedScattered\]](#eq:MLConstrainedScattered){reference-type="eqref" reference="eq:MLConstrainedScattered"} as Notice that even though the cost function for this problem is convex, the constraints in [\[eq:MLConstrainedScattered1-b\]](#eq:MLConstrainedScattered1-b){reference-type="eqref" reference="eq:MLConstrainedScattered1-b"}--[\[eq:MLConstrainedScattered1-d\]](#eq:MLConstrainedScattered1-d){reference-type="eqref" reference="eq:MLConstrainedScattered1-d"} are non-convex and hence the problem is non-convex. Consequently, we next address the localization problem by considering a convex relaxation of this problem. This allows us to solve the localization problem approximately. One of the ways to provide a convex approximation of the problem in [\[eq:MLConstrainedScattered1\]](#eq:MLConstrainedScattered1){reference-type="eqref" reference="eq:MLConstrainedScattered1"} is to relax the quadratic equality constraints in [\[eq:MLConstrainedScattered1-b\]](#eq:MLConstrainedScattered1-b){reference-type="eqref" reference="eq:MLConstrainedScattered1-b"}--[\[eq:MLConstrainedScattered1-d\]](#eq:MLConstrainedScattered1-d){reference-type="eqref" reference="eq:MLConstrainedScattered1-d"} using Schur complements, which results in where \[\begin{aligned} f_{ij}(\Lambda_{ij}, D_{ij}) & = \frac{1}{\sigma_{ij}^2} (\Lambda_{ij}-2D_{ij}R_{ij} + R_{ij}^2),\\ g_{ij}(\Xi_{ij}, Z_{ij}) & = \frac{1}{\delta_{ij}^2} (\Xi_{ij}-2Z_{ij}Y_{ij} + Y_{ij}^2), \end{aligned}\] and \[\begin{aligned} \Omega_{ij} &= \Bigg \{ (S_{ii}, S_{jj}, S_{ij}, \Lambda_{ij}, D_{ij}, \Gamma^{ij}) \ \Bigg \| \ S_{ii} + S_{jj}-2S_{ij} = \Lambda_{ij}, \\ & \quad \quad \quad \quad \quad \quad \quad \quad \quad \ \begin{bmatrix} 1 & D_{ij} \\ D_{ij} & \Lambda_{ij} \end{bmatrix} = \Gamma^{ij}, \ \Gamma^{ij} \succeq 0, \ D_{ij} \geq 0 \Bigg\},\\ \Theta_{ij} &= \Bigg \{ (S_{ii}, x_s^i, \Xi_{ij}, Z_{ij}, \Phi^{ij}) \ \Bigg \| \ S_{ii}-2(x_s^i)^Tx_a^j + \\| x_a^j \\|^2_2 = \Xi_{ij}, \\ & \quad \quad \quad \ \ \ \begin{bmatrix} 1 & Z_{ij} \\ Z_{ij} & \Xi_{ij} \end{bmatrix} = \Phi^{ij}, \ \Phi^{ij} \succeq 0, Z_{ij} \geq 0, \ \ j \in \text{Ne}_a(i) \Bigg\}, \end{aligned}\] with the variables $\Gamma^{ij}$, $\Phi^{ij}$ and $T^i$ as slack variables. The addition of the slack variables enable us to make the description of the semidefinite constraints simpler. This problem is a coupled SDP and can be solved distributedly using $q$ computational agents. In order to see this with more ease, let us introduce a grouping of the cost function terms and constraints in [\[eq:MLConstrainedScattered2-a\]](#eq:MLConstrainedScattered2-a){reference-type="eqref" reference="eq:MLConstrainedScattered2-a"}--[\[eq:MLConstrainedScattered2-c\]](#eq:MLConstrainedScattered2-c){reference-type="eqref" reference="eq:MLConstrainedScattered2-c"}. To this end we first describe a set of assignment rules. It is possible to assign 1. the constraint $(S_{ii}, S_{jj}, S_{ij}, \Lambda_{ij}, D_{ij}, \Gamma^{ij}) \in \Omega_{ij}$ and the cost function term $f_{ij}$ to agent $k$ if $(i,j) \in C_k \times C_k$; 2. the set of constraints $(S_{ii}, x_s^i, \Xi_{ij}, Z_{ij}, \Phi^{ij}) \in \Theta_{ij}, \ \ j \in \text{Ne}_a(i)$ and the cost function terms $g_{ij}, \ \ j \in \text{Ne}_a(i)$ to agent $k$ if $i \in C_k$. We denote the indices of the constraints and cost function terms assigned to agent $k$ through Rule 1 above as $\phi_{k}$, and similarly we denote the set of constraints and cost function terms that are assigned to agent $k$ through Rule 2 by $\bar \phi_k$. Using the mentioned rules and the defined notations, we can now group the constraints and the cost function terms and rewrite the problem in [\[eq:MLConstrainedScattered2\]](#eq:MLConstrainedScattered2){reference-type="eqref" reference="eq:MLConstrainedScattered2"} as Notice that this problem can now be seen as a combination of $q$ coupled subproblems, each defined by a term in the cost function together with its corresponding set of constraints in [\[eq:MLConstrainedScattered3-b\]](#eq:MLConstrainedScattered3-b){reference-type="eqref" reference="eq:MLConstrainedScattered3-b"}. It is possible to decompose this problem by introducing additional local variables and consistency constraints and use any proximal point splitting method, e.g., ADMM, to solve this problem distributedly. However, there are major disadvantages for the resulting distributed solution, such as - the local subproblems that needs to be solved by each agent is a semidefinite program that are computationally expensive to solve; - inexact solutions for semidefinite programs can be far away from the optimal solution; - the algorithm generally requires many iterations to converge to an accurate solution that particularly satisfies the consistency constraints; - the number of consistency constraints are generally big for such problems which can even further adversely affect the convergence and numerical properties of such algorithms. In order to evade the aforementioned issues, we next put forth an alternative distributed algorithm based on primal-dual interior-point methods that fully takes advantage of the structure in the problem and yields an accurate solution within much lower number of iterations and with far less computational demands from each agent. ## A Simple Assignment Strategy Before we continue, let us first put forth an assignment strategy that is simple and satisfies the assignment rules discussed above. Recall that in order to form the problem in [\[eq:MLConstrainedScattered3\]](#eq:MLConstrainedScattered3){reference-type="eqref" reference="eq:MLConstrainedScattered3"}, we first need to cluster the sensor nodes. Based on this clustering, we use the assignment strategy described in Algorithm [\[alg:Ass\]](#alg:Ass){reference-type="ref" reference="alg:Ass"}. # Distributed Primal-dual Interior-point Method for Coupled SDPs {#sec:DPDIPM} The problem in [\[eq:MLConstrainedScattered3\]](#eq:MLConstrainedScattered3){reference-type="eqref" reference="eq:MLConstrainedScattered3"} can be written in the following standard form where the variables $X^i_j$ and $y$ are matrix and linear variables, respectively. This problem can be written more compactly as with $Q^i = \blkdiag(Q^i_1, \dots, Q^i_{m_i})$, $W^i = \begin{bmatrix} (W^i_1)^T & \dots & (W^i_{m_i})^T\end{bmatrix}^T$, $b^i = (b^i_1, \dots, b^i_{m_i})$, $x^i = \left(\svec(X^i_1), \dots, \svec(X^i_{m_i})\right)$ and $X^i = \blkdiag(X^i_1, \dots, X^i_{m_i})$. It is possible to solve this problem using a primal-dual interior-point method, ,. Next we briefly discuss the main stages of such a method. The Karush-Kuhn-Tucker, KKT, optimality conditions for this problem are given as together with $D^i x \leq g^i$ and $X^i \succeq 0$, where $Z^i = \blkdiag(Z^i_1, \dots, Z^i_{m_i})$ and $z^i = \left(\svec(Z^i_1), \dots, \svec(Z^i_{m_i})\right)$. Any solution to this set of nonlinear equations is optimal for [\[eq:GeneralSDP\]](#eq:GeneralSDP){reference-type="eqref" reference="eq:GeneralSDP"}. Within a primal-dual interior-point method, we set out to compute a solution to [\[eq:GeneralSDP\]](#eq:GeneralSDP){reference-type="eqref" reference="eq:GeneralSDP"}, by considering a sequence of perturbed KKT conditions where [\[eq:KKT-c\]](#eq:KKT-c){reference-type="eqref" reference="eq:KKT-c"} and [\[eq:KKT-d\]](#eq:KKT-d){reference-type="eqref" reference="eq:KKT-d"} are modified as \[\begin{aligned} X^iZ^i &= \delta I, \quad i \in \mathbb N_q,\\ \diag(\lambda^i)\left( D^iy-g^i \right) &=-\delta \mathbf 1, \quad i \in \mathbb N_q. \end{aligned}\] where $\delta >0$ is the perturbation parameter. Particularly at each iteration, given feasible iterates $\lambda^i > 0$, $y$ so that $D^i y > g^i$ and $X^i \succ 0$ for $i = 1, \dots, q$, the primal-dual search directions are computed by solving a linearized version of the perturbed KKT conditions, given as with $U^i = \blkdiag(U^i_1, \dots, U^i_{m_i})$, $F^i = \blkdiag(F^i_1, \dots, F^i_{m_i})$, where given \[\label{eq:scaling} \begin{split} W^i_j :&= (X^i_j)^{\frac{1}{2}} \left( (X^i_j)^{\frac{1}{2}} Z^i_j (X^i_j)^{\frac{1}{2}} \right)^{-\frac{1}{2}} (X^i_j)^{\frac{1}{2}}\\ & = (Z^i_j)^{-\frac{1}{2}} \left( (Z^i_j)^{\frac{1}{2}} X^i_j (Z^i_j)^{\frac{1}{2}} \right)^{\frac{1}{2}} (Z^i_j)^{-\frac{1}{2}}, \end{split}\] $W^i_j =: G^i_j(G^i_j)^T$ and $D^i_j = (G^i_j)^{-1}$, we have $U^i_j = D^i_j \otimes_s (D^i_j)^{-T}Z^i_j$ and $F^i_j = D^i_jX^i_j \otimes_s (D^i_j)^{-T}$. Furthermore, the residuals are given as where $H_D(M) = 1/2(DMD^{-1} + D^{-T}MD^{T})$. Having computed the search directions, suitable primal and dual step sizes, i.e., $t_d$ and $t_p$, are calculated so as to guarantee feasibility of the iterates with respect to inequality constraints and persistent reduction of residual norms, see e.g., and references therein, which then allows us to update the iterates. This process is then repeated until certain stopping criteria are satisfied, which commonly depend on the residual norms and the size of the perturbation parameter. A generic description of a primal-dual interior-point method is given in Algorithm [\[alg:PDIPM\]](#alg:PDIPM){reference-type="ref" reference="alg:PDIPM"}. which is possible since $F$ and $E$ are both invertible, see e.g.,. This then allows us to rewrite [\[eq:KKTCompact\]](#eq:KKTCompact){reference-type="eqref" reference="eq:KKTCompact"} as \[\begin{gathered} \begin{bmatrix}-D^TE^{-1}D & & W^T & A^T \\ & F^{-1}U & Q^T & \\ W & Q & & \\A & & & \end{bmatrix} \begin{bmatrix} \Delta y \\ \Delta x \\ \Delta v \\ \Delta \bar v \end{bmatrix} = \begin{bmatrix} r_{\textrm{lin}}\\ r \\ r_{\textrm{p}} \\ r_{\textrm{p,lin}} \end{bmatrix} \end{gathered}\] where $r_{\textrm{lin}} = r_{\textrm{d,lin}}-D^T E^{-1} r_{\textrm{c,lin}}$ and $r = r_{\textrm{d}}-F^{-1} r_{\textrm{c}}$. Notice that this set of linear equations also defines the optimality conditions for the convex quadratic program (QP) For the localization problem, this QP has a particular structure which enables us to solve it distributedly and efficiently, using message-passing. Next we briefly discuss this algorithm for the sake of completeness and to provide a better understanding of the presented material. ## Solving Coupled Optimization Problems Using Message-passing Consider the following coupled optimization problem \[\begin{aligned} \label{eq:UncostrainedProblem} \minimize & \quad F_1(x) + F_2(x) + \dots + F_q(x), \end{aligned}\] where $x \in \mathbb R^n$ and and the functions $F_i \: \ \mathbb R^n \rightarrow \mathbb R$ for $i \in \mathbb N_q$ are convex. Also we assume that each term in the objective function (each subproblem) only depends on a few variables. Let us denote the indices of the variables that appear in the $i$th term, $F_i$, by $J_i$. This definition allows us to rewrite the problem in [\[eq:UncostrainedProblem\]](#eq:UncostrainedProblem){reference-type="eqref" reference="eq:UncostrainedProblem"} as \[\begin{aligned} \label{eq:UncostrainedProblemReform} \minimize & \quad \bar F_1(x_{_{J_1}}) +\bar F_2(x_{_{J_2}}) + \dots +\bar F_q(x_{_{J_q}}), \end{aligned}\] where $x_{_{J_i}} = E_{J_i} x$. The functions $\bar F_i \: \ \mathbb R^{\|J_i\|} \rightarrow \mathbb R$ are lower dimensional descriptions of $F_i$s such that $F_i(x) = \bar F_i(E_{J_i}x)$ for all $x \in \mathbb R^n$ and $i \in \mathbb N_N$. We also define $\mathcal I_j$ as the set of indices of terms in the cost function that depend on $x_j$, i.e., $\{ i \ \| \ j \in J_i \}$. The sets $J_i$ for $i \in \mathbb N_q$ and $\mathcal I_j$ for $j \in \mathbb N_n$ provide a clear mathematical description of the coupling structure in the problem. It is also possible to describe the coupling structure in the problem graphically, using graphs. For this purpose, we introduce the sparsity graph. The sparsity graph $G_s(V_s, \mathcal E_s)$ of a coupled problem is an undirected graph with the vertex set $V_s = \left\{ 1, \dots, n\right\}$ and the edge set $\mathcal E_s = \left\{ (i, j) \ \| \ i, j \in V_s, \ \mathcal I_i \cap \mathcal I_j \neq \emptyset \right\}$. As an example consider the following problem \[\begin{gathered} \label{eq:Example} \minimize_x \quad \bar F_1(x_1, x_3, x_4) + \bar F_2(x_1, x_2, x_4) + \bar F_3(x_4, x_5) +\\ \bar F_4(x_3, x_6, x_7) + \bar F_5(x_3, x_8). \end{gathered}\] The sparsity graph for this problem are illustrated in Figure [\[fig:Example\]](#fig:Example){reference-type="ref" reference="fig:Example"}. It is possible to devise scalable or distributed algorithms for solving the problem in [\[eq:UncostrainedProblem\]](#eq:UncostrainedProblem){reference-type="eqref" reference="eq:UncostrainedProblem"}. In this paper we focus on message-passing. Consider the problem in [\[eq:UncostrainedProblemReform\]](#eq:UncostrainedProblemReform){reference-type="eqref" reference="eq:UncostrainedProblemReform"}, and assume that its sparsity graph is chordal. Let its set of cliques be given as $\mathbf C_{G_s} = \{ J_1, \dots, J_q\}$ and $T_s(V_t, \mathcal E_t)$ be a clique tree over the cliques. It is possible to solve the problem in [\[eq:UncostrainedProblemReform\]](#eq:UncostrainedProblemReform){reference-type="eqref" reference="eq:UncostrainedProblemReform"} distributedly, using an algorithm with the clique tree as its computational graph. That means each node in the tree corresponds to a computational agent and they communicate/collaborate with one another if there is an edge between them. Recall that each node in the clique tree is assigned a clique of the sparsity graph, i.e., $J_i$. In such a setting, we also assign each term in the objective function (each subproblem), i.e., $\bar F_i$, to each agent $i$. We can now describe how the problem in [\[eq:UncostrainedProblemReform\]](#eq:UncostrainedProblemReform){reference-type="eqref" reference="eq:UncostrainedProblemReform"} can be solved using message-passing by performing an upward-downward pass through the clique tree. The message-passing algorithm starts from the agents at the leaves of the tree, i.e., all $i \in \leaves(T)$, where every such agent computes the following message \[\begin{aligned} \label{eq:mijLeaves} m_{i\parent(i)}(x_{_{S_{i\parent(i)}}}) = \minimum_{x_{_{R_{i\parent(i)}}}} \left\{ \bar F_i(x_{_{J_i}})\right\}, \end{aligned}\] with $S_{i\parent(i)} := J_i \cap J_{\parent(i)}$ and $R_{i\parent(i)} := J_i \setminus S_{i\parent(i)}$ are the so-called separators and residuals, respectively, and communicates it to its corresponding parent, denoted by $\parent(i)$. Notice that this message is a functional and not a scalar value, and hence agent $i$ needs to communicate the functional form. Then every parent $j$ that has received these messages from its children, denoted by $\children(j)$, computes its corresponding message to its parent as \[\begin{aligned} \label{eq:mij} m_{j\parent(j)}(x_{_{S_{j\parent(j)}}}) = \minimum_{x_{_{R_{j\parent(j)}}}} \left\{ \bar F_j(x_{_{J_j}}) + \sum_{k \in \children(j)} m_{kj}(x_{_{S_{kj}}}) \right\}. \end{aligned}\] This procedure is then continued until we arrive at the agent at the root. At this point, the agent at the root, indexed $r$, having received all messages from its children can compute the optimal solution for its corresponding variables specified by $J_r$ as \[\begin{aligned} \label{eq:RLocalProblem} x^\ast_{_{J_r}} = \argmin_{x_{_{J_r}}} \left\{ \bar F_k(x_{_{J_r}}) + \sum_{k \in \children(r)} m_{kr}(x_{_{S_{rk}}}) \right\}. \end{aligned}\] This agent then having computed its optimal solution, communicates this solution to its children, at which point every such agent $i \in \children(r)$ computes its optimal solution as \[\begin{gathered} \label{eq:LocalProblempar} x^\ast_{_{J_i}} = \argmin_{x_{_{J_i}}} \left\{ \bar F_i(x_{_{J_i}}) + \sum_{k \in \children(i)} m_{ki}(x_{_{S_{ik}}}) + \right. \\ \left. \frac{1}{2} \left\\| x_{_{S_{\parent(i)i}}}-\left( x_{_{S_{\parent(i)i}}}^\ast \right)^{\parent(i)} \right\\|^2 \right\}, \end{gathered}\] where $\left( x_{_{S_{\parent(i)i}}}^\ast \right)^{\parent(i)}$ is the the computed optimal solution by the parent $\parent(i)$. This procedure is continued until we reach the agents at the leaves. At this point all agents have computed their corresponding optimal solution and the algorithm can be terminated, and hence, we have convergence after one upward-downward pass through the tree, ,. Let us now illustrate this procedure using an example. Consider the example given in [\[eq:Example\]](#eq:Example){reference-type="eqref" reference="eq:Example"}. The sparsity graph of this problem is chordal and its cliques are marked in Figure [\[fig:ExampleClique\]](#fig:ExampleClique){reference-type="ref" reference="fig:ExampleClique"} on the left. A clique tree for this graph is illustrated in the same figure on the right, where also a valid subproblem assignment is presented. As was discussed above we start the message-passing from the leaves of the tree, particularly agents 3, 4 and 5. These agents compute and communicate their messages to their corresponding parents as \[\begin{aligned} m_{32}(x_4) &= \minimum_{x_5}\left\{ \bar F_3(x_4, x_5)\right\}\\ m_{41}(x_3) & = \minimum_{x_6, x_7}\left\{\bar F_4(x_3, x_6, x_7)\right\} \\ m_{51}(x_3) & = \minimum_{x_8}\left\{ \bar F_5(x_3, x_8)\right\}. \end{aligned}\] At this point agent 2 has received all messages from its children and can in turn compute and communicate its message to its parent as \[\begin{aligned} m_{12}(x_1, x_4) &= \minimum_{x_3}\left\{ m_{41}(x_3) + m_{51}(x_3) + \bar F_1(x_1, x_3, x_4)\right\}. \end{aligned}\] This completes the upward pass and now the agent at the root, i.e., agent 2, can compute its optimal solution as \[\begin{gathered} (x_1^, x_2^, x_4^) = \argmin_{x_1, x_2, x_4} \\\quad\left\{ m_{12}(x_1, x_4) + m_{32}(x_4) + \bar F_2(x_1, x_2, x_4)\right\}, \end{gathered}\] which initiates the downward pass. Agent 2 will then communicate $x_1^, x_4^$ and $x_4^$ to agents 2 and 3 respectively, where they compute their corresponding optimal solution for the remainder of their variables as \[\begin{aligned} x_3^* &= \argmin_{x_3} \quad\left\{ m_{41}(x_3) + m_{51}(x_3) + \bar F_1(x_1^, x_3, x_4^)\right\}\\ x_5^* &= \argmin_{x_5} \quad\left\{ \bar F_3(x_4^, x_5)\right\}. \end{aligned}\] The last step of the downward pass is then accomplished by agent 2 communicating $x_3^$ to agents 4 and 5, and these agents computing their optimal solution as \[\begin{aligned} (x_6^, x_7^) &= \argmin_{x_6, x_7} \quad\left\{ \bar F_4( x_3^, x_6, x_7)\right\}\\ x_8^ &= \argmin_{x_8} \quad\left\{ \bar F_5(x_3^, x_8)\right\}, \end{aligned}\] which finishes the algorithm. Notice that the message-passing algorithm described in this section can be viewed as dynamic programming over trees. Next we discuss how message-passing can be used within the primal-dual method. ## Distributed Computations In Primal-dual methods The problem in [\[eq:KKTQP\]](#eq:KKTQP){reference-type="eqref" reference="eq:KKTQP"} can be written as where $H^i =-(D^i)^T(E^i)^{-1}D^i$ and $r^i_{\textrm{lin}} = r^i_{\textrm{d,lin}}-(D^i)^T (E^i)^{-1} r^i_{\textrm{c,lin}}$ and $r^i = r^i_{\textrm{d}}-(F^i)^{-1} r^i_{\textrm{c}}$. This problem can be viewed as a combination of $q$ subproblems, where each of which is defined by a term in the objective function and its corresponding equality constraints. Notice that the coupling among the subproblems does not stem from the matrix variables and on the surface all subproblems seem to be coupled to one another through the linear variables directions $\Delta y$. However, for the localization problem in [\[eq:MLConstrainedScattered3\]](#eq:MLConstrainedScattered3){reference-type="eqref" reference="eq:MLConstrainedScattered3"}, each subproblem only relies on a certain elements of $\Delta y$. This can be seen by first noticing that the linear variables for each subproblem $k$ is given by $\vectri(S_{C_kC_k})$, $\Lambda_{ij}, D_{ij}$ for $(i,j) \in \phi_k$ and $x^i_s, \Xi_{ij}, Z_{ij}$ for $j \in \textrm{Ne}_a(i)$ and $i \in \bar \phi_k$. Let us assume that the indices of elements of $\Delta y$ that correspond to these variables be given by set $J_k$. We can then rewrite the problem in [\[eq:KKTQPSum\]](#eq:KKTQPSum){reference-type="eqref" reference="eq:KKTQPSum"} as where $\bar H^i = E_{J_i}H^iE_{J_i}^T$, $\bar r^i_{\textrm{lin}} = E_{J_i}r^i_{\textrm{lin}}$, $\bar A^i = A^iE_{J_i}^T$ and $\bar W^i = W^i E_{J_i}^T$. Through the use of indicator functions, this problem can be written as \[\begin{aligned} \minimize \quad& \sum_{i = 1}^q \begin{bmatrix}\Delta y_{_{J_i}} \\ \Delta x^i \end{bmatrix}^T \begin{bmatrix} \bar H^i & \\ & (F^i)^{-1}U^i \end{bmatrix} \begin{bmatrix} \Delta y_{_{J_i}} \\ \Delta x^i \end{bmatrix}-\notag\\ & \hspace{25mm}\begin{bmatrix}\bar r^i_{\textrm{lin}}\\ r^i \end{bmatrix}^T \begin{bmatrix} \Delta y_{_{J_i}} \\ \Delta x^i \end{bmatrix} + \mathcal I_{\mathcal C_i}(\Delta y_{_{J_i}}, \Delta x^i) \end{aligned}\] where $\mathcal C_i = \{ (\Delta y_{_{J_i}},\Delta x^i)\ \| \ \bar W^i\Delta y_{_{J_i}} + Q^i \Delta x^i = r^i_{\textrm{p}}, \ \bar A^i \Delta y_{_{J_i}} = r^i_{\textrm{p,lin}}\}$ and \[\begin{aligned} \mathcal I_{\mathcal C_i}\left(x\right) = \begin{cases} 0 \hspace{6mm} x \in \mathcal C_i \\ \infty \hspace{4mm} \text{Otherwise} \end{cases} \end{aligned}\] This problem is in the same format as [\[eq:UncostrainedProblemReform\]](#eq:UncostrainedProblemReform){reference-type="eqref" reference="eq:UncostrainedProblemReform"}. It is now possible to see that the coupling comes from the fact that for some $C_i$ and $C_j$, $C_i \cap C_j \neq \emptyset$. Recall that one way to describe the intersection among the cliques of the inter-sensor measurement graph can be described using its clique tree, $T(V_t, \mathcal E_t)$. The sparsity graph of this problem is in fact chordal with cliques defined by the variables that appear in each subproblem. Furthermore, the clique tree for the sparsity graph of this problem has the same structure as that of the inter-sensor measurement graph. This is the case since the ordering defined by this tree defines perfect elimination ordering for the sparsity graph, see for more details. Consequently, this problem can be solved distributedly using message-passing as discussed above. As a result, we can compute the primal-dual search directions for the problem in [\[eq:MLConstrainedScattered3\]](#eq:MLConstrainedScattered3){reference-type="eqref" reference="eq:MLConstrainedScattered3"} distributedly, by an upward-downward pass through the clique tree. Notice that the messages for solving this problem are quadratic functions, and hence the hessian and linear term that describes this function need to be communicated. The remaining stages of a primal-dual interior-point method can also be done distributedly over the clique tree. For the sake of brevity, we here do not discuss the details any further, for more info see, and. A summary of our proposed distributed localization method is given in Algorithm [\[alg:Local\]](#alg:Local){reference-type="ref" reference="alg:Local"}. Next we investigate the performance of our proposed algorithm, using two sets of numerical experiments. # Numerical Experiments {#sec:numerical} In this section we compare the performance of our proposed distributed algorithm with that of presented in. We refer to this algorithm as distributed disk relaxation algorithm (DDRA). To this end, we conduct two sets of experiments, one that relies on simulated data and one that is based on real data from. Notice that we do not conduct a comparison with other algorithms, since a thorough comparison with DDRA has been conducted in, which illustrated the superiority of their proposed algorithm to high performance algorithms in and both in accuracy and number of communications among agents. ## Experiments Using Simulated Data {#sec:numerical-A} Our experiments based on simulated data concern networks of sensors with connected inter-sensor measurement graphs. In all experiments there are 9 anchors in the network which are uniformly distributed in the area. The experiments in this section are divided into two setups. In both setups, we consider a network of several sensors which are placed in a two-dimensional area, with their locations randomly generated using a uniform distribution. The noisy range measurements are generated as \[\begin{aligned} \mathcal R_{ij} &= \left\|\\| (x^_s)^i-(x^_s)^j \\|_2 + E_{ij}\right\|, \quad j \in \text{Ne}_r(i),\\ \mathcal Y_{ij} &= \left\|\\| (x^_s)^i-(x_a)^j \\|_2 + V_{ij}\right\|, \quad j \in \text{Ne}_a(i), \end{aligned}\] where $(x^_s)^i$ denotes the true location of the $i$th sensor. Furthermore we assume that all noises are gaussian and mutually independent, see also. In the first setup we conduct experiments using a network 50 sensors in a $0.8 \times 0.8$ area. We consider four different measurement noise standard deviations, namely $0.01, 0.05$, $0.1$ and $0.3$, and for each noise level we generate 50 problem instances. In order to ensure that the generated inter-sensor measurement graph is loosely connected, we assume there exist a measurement between two sensors or between a sensor and an anchor if the distance between them is less than the communication range $r_c =0.2$. The resulting sensor network is depicted in Figure [\[fig:network\]](#fig:network){reference-type="ref" reference="fig:network"}. In this figure, the sensor nodes are marked with red crosses and the anchors are marked with green circles. As can be seen from the figure the inter-sensor measurement graph is connected. The performance of distributed algorithms are quantified using three measures. Namely (i) their accuracy based on the root mean squared error (RMSE) defined as \[\begin{aligned} \textrm{RMSE} = \sqrt{\frac{1}{MN}\sum_{i=1}^M\sum_{j=1}^N \\| (x^_s)^j-x_s^j(m) \\|^2} \end{aligned}\] where $M$ is the number of experiments and the argument $m$ marks the computed estimate for the $m$th experiment, (ii) number of required iterations and communications to converge to a solution with a given accuracy and (iii) the computational time. Notice that both algorithms are run in a centralized manner. The algorithm in is terminated if the norm of the gradient of its considered cost function is below $10^{-6}$. This threshold was chosen based on the authors experience, so as to guarantee DDRA generates accurate enough solutions. Figures [\[fig:ex11\]](#fig:ex11){reference-type="ref" reference="fig:ex11"}--[\[fig:ex13\]](#fig:ex13){reference-type="ref" reference="fig:ex13"} illustrate the achieved results. In these figures and the ones to come the $*$-marked curves illustrate the results from DPDLA, whereas the o-marked curves show the results from DDRA. As can be seen from Figure [\[fig:ex11\]](#fig:ex11){reference-type="ref" reference="fig:ex11"}, DPDLA outperforms or provides comparable accuracy with respect to DDRA for different levels of measurement noise. This shows the superiority of semidefinite relaxation to disk relaxation. The number of communications that each agent is required to conduct for each algorithm to converge to a solution is depicted in Figure [\[fig:ex12\]](#fig:ex12){reference-type="ref" reference="fig:ex12"}. For these experiments, the considered clique tree for the inter-sensor measurement graph in Figure [\[fig:network\]](#fig:network){reference-type="ref" reference="fig:network"}, has height 8, and the primal-dual method converged within around 10 iterations. As can be seen from the figure, DPDLA requires roughly two orders of magnitude less number of communications for computing a solution. The shaded areas depict the maximum and minimum values within the 50 instances for each of these quantities. Notice that this area for the results corresponding to DPDLA is not even visible. We can hence deduce that in comparison DDRA, the number of communications for DPDLA seems to be much less sensitive to the noise level and also to data realizations. The computational time for the considered algorithms are presented in Figure [\[fig:ex13\]](#fig:ex13){reference-type="ref" reference="fig:ex13"}. As can be seen from the figure DDRA is at least twice as fast as DPDLA, owing to very simple computations required from each agent at every iteration. This is the case if both algorithms are executed in a centralized manner and if we neglect the communication cost or delay. Based on the presented results, our proposed algorithm provides more accurate estimates, and even though slower when implemented in a centralized manner, it provides a better distributed algorithm as it requires far less amount of communications. In the second simulation setup, we test the performance of the considered algorithms, when applied to networks with varying number of sensors, namely, 10, 20, 30, 40 and 50. In this setup we assume that the measurement noise standard deviation is 0.01, and we consider 50 instances for each network size. Furthermore the size of the considered area and the communication range, $r_c$, for each network size are chosen such that the resulting inter-sensor measurement graphs are connected but loosely. Figure [\[fig:ex21\]](#fig:ex21){reference-type="ref" reference="fig:ex21"} illustrates the RMSE results for this experiment. As before, as can be seen from the figure, DPDLA provides more accurate estimates for all network sizes. Also as can be seen from Figure [\[fig:ex22\]](#fig:ex22){reference-type="ref" reference="fig:ex22"}, the estimates are computed using far fewer communications among agents. The primal-dual method converged within around 11 iterations and the heights of the clique trees for the different sensor networks ware between 3 to 8. As can be seen from the figure, the number of required communications for the DDRA to converge grows much faster with network size than that of DPDLA which seems to be far less sensitive to this change. Figure [\[fig:ex23\]](#fig:ex23){reference-type="ref" reference="fig:ex23"} illustrates the total computational time of both algorithms when implemented in a centralized manner. As can be seen from this figure, our proposed algorithm requires similar or less amount of time to converge to a solution for networks of up to 30 sensors. Consequently, for networks with less than 30 sensors, our proposed algorithm outperforms DDRA in all the performance criteria. It is also worth mentioning that, the performance of our algorithm can be improved considerably, if the clustering of the sensors and generation of a clique tree are done using more sophisticated and tailored approaches. However, since we did not discuss such approaches, we abstained from any manipulation of the cliques and the clique tree and simply relied on standard and simple heuristics for this purpose, see e.g., and references therein. ## Experiments Using Real Data In this section, we present the results from conducted experiments based on real data. This data was taken from, that includes time of arrival (TOA) measurements among 44 sensors, 4 of which are deemed to be anchors. The sensors are spread out in a $14 \times 13$ area. We extract the range measurements from the available TOA measurements. This provides us with biased range measurements with a standard deviation of 1.82 meters, see. We here study the performance of DDRA and DPDRA for different levels of connectivity of the inter-sensor graph. To this end, we gradually change the communication range from 4 to 6.5 meters. Figures [\[fig:ex31\]](#fig:ex31){reference-type="ref" reference="fig:ex31"} and [\[fig:ex33\]](#fig:ex33){reference-type="ref" reference="fig:ex33"} illustrate the results. Notice that due to biasedness and quality of the measurements, the intersection of the range measurement disks can be empty and hence DDRA fails to converge. This is because the gradient of the cost function of the disk relaxation problem does not vanish. Consequently, this algorithm has been terminated after 5000 iterations. Figure [\[fig:ex31\]](#fig:ex31){reference-type="ref" reference="fig:ex31"} illustrates the RMSE results from the experiment, which clearly depicts that DPDLA outperforms DDRA. Furthermore, DPDLA required each agent was required to communicate with its neighbors around 100 times which seemed to be robust with respect to the level of connectivity of the inter-sensor range measurement graph. The primal-dual method for all these instances converged within roughly 17 iterations and the height of the clique tree varied between 7 to 9. Figure [\[fig:ex33\]](#fig:ex33){reference-type="ref" reference="fig:ex33"} illustrates the computational time for DDRA and DPDLA. As was also observed from the experiments in Section [7.1](#sec:numerical-A){reference-type="ref" reference="sec:numerical-A"}, DDRA clearly outperforms DPDLA when implemented in a centralized manner. # Conclusions {#sec:conclusions} In this paper we proposed a distributed localization algorithm for tree-structured scattered sensor networks founded on semidefinite relaxation of the localization problem. This algorithm is based on state-of-the-art primal-dual interior-point methods and relies on message-passing or dynamic programming over trees to distribute the computations. Due to this, the resulting algorithm requires far fewer steps and even fewer communications among computational agents to converge to an accurate solution, and it achieves this by putting a moderate computational burden on the agents. Furthermore, the proposed distributed algorithm is robust to biases in the measurements, or in general bad quality of the measurements. This stems from the power of semidefinite relaxation for localization problems. Despite these advantages, the proposed algorithm is much more complicated than algorithms that rely on first-order methods. This is largely due to the fact that generally second-order methods are far more complicated than their first-order counter parts. The choice of clustering of the sensors and the strategy for assigning the available measurements to computational agents can have a significant effect on the performance of our proposed algorithm. Also smart clustering of the sensors, may even enable us to use the computational infrastructure at the anchors and utilize them as computational agents. In this paper, we briefly discussed the importance of this and provided some suggestions on how the used heuristic strategies for this purpose can be improved. We did not investigate this topic in detail, however, we believe that further exploration of this matter can result in interesting results. Furthermore, distributed approaches for computing cliques and clique trees of the inter-sensor measurement were not covered in this paper, although, complementing the proposed algorithm with such methods can enhance the practicality of the algorithm.	{'timestamp': '2016-07-19T02:05:25', 'yymm': '1607', 'arxiv_id': '1607.04798', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04798'}
# Introduction Quantum theory of gravity is one of the most sought-after goals in physics. Despite continuous efforts to tackle this important problem, resulting in interesting proposals such as superstring theory and loop quantum gravity, there is still no clear sign of the theory of quantum gravity. However, when the back-action of matter on the gravitation field is neglected, one can write down a theory of quantum fields in a background curved spacetime by extending quantum field theory in Minkowski metric to a general metric. This is analogous to treating external fields as c-numbers and predicts interesting new phenomena that are valid in appropriate regimes. Similarly to the prediction of pair creation in external electric fields, external gravitational fields (as described by a background spacetime) induce particle creation. In the latter, particle creation is caused by the change in the vacuum state itself under quite generic conditions. In this work, we propose a classical optical simulation of particle creation in binary waveguide arrays. There have been many proposals and experimental demonstrations of a plethora of interesting physics in coupled waveguide arrays. In particular, optical simulation of the 1+1 dimensional Dirac equation in binary waveguide arrays has been proposed and experimentally demonstrated. We show that the setup can be generalised to also simulate the Dirac equation in 2 dimensional curved spacetime. Particle creation is by definition a multi-particle phenomenon and the full simulation of the result requires quantum fields as a main ingredient (for example, see for a simulation of the Dirac equation in curved spacetime with cold atoms on optical lattices). However, light propagation in a waveguide array is an intrinsically classical phenomena, so how can we simulate particle creation in a binary waveguide array? The short answer is that we will be looking at a single-particle analog of particle creation. As we will see, the fundamental reason behind particle creation is the difference in the vacuum state, which in turn is captured by different mode-expansions of quantum fields. We can thus concentrate on a single-mode at a time and simulate the effect. In fact, the well-known Klein paradox shows that the single-particle Dirac equation contains subtle hints of multi-particle effects, and the phenomenon of pair production in strong electric fields has been studied within the single particle picture. We study the time evolution of spinor wave packets and demonstrate that an analog of particle creation can be visualised in the light evolution in a binary waveguide array. Here, we stress that by using the phenomenon of 'zitterbewegung' (the jittering motion of a Dirac particle), one can bypass the quantitative checks in 'proving' the simulation of particle creation. This article is organised as follows. In section II, we provide a pedagogical introduction to the Dirac equation in curved spacetime, assuming familiarity with the conventional Dirac equation. In section III, we specialise to the 1+1 dimensions and provide a few examples of spacetime metrics. Particle creation in curved spacetime is explained, using scalar fields for simplicity, and the single-particle analogs for the Dirac spinors is discussed. Section IV shows the wave packet evolution both in flat and curved spacetimes, using time-dependent gravitational fields as an example. A single-particle analog of particle creation in a particular case is explicitly demonstrated. Section V explains the optical simulation of the Dirac equation in a binary waveguide array and proposes a generalisation to curved spacetimes. We conclude in section VI. We have tried to be as pedagogical and self-contained as we could. We have tried to collect and present essential ideas to understand quantum fields in curved spacetime and how it predicts particle creation. It is our hope that the reader will find this helpful in understanding the essential ideas quickly and develop further interesting analogies. # Dirac equation in curved spacetime Let us start with the derivation of the Dirac equation in curved spacetime. This requires the notion of spin-connection, which will be discussed at a pedagogical level. We closely follow ref. . ## General covariance Special theory of relativity has taught us that time and space are observer-dependent concepts in that observers moving relative to each other have different notions of time and space intervals. This comes about because the laws of physics are covariant under a Lorentz transformation, meaning that all observers agree on the form of physical laws in their own coordinate frames. General theory of relativity takes this one step further and states that the laws should be covariant under general coordinate transformations. Equations of motion are written in terms of tensors--quantities that are independent of local coordinate systems used to describe them. To write down the equations of motion one also requires the notion of a covariant derivative, which basically is the correct way of differentiating tensors to yield another tensor (of higher rank). To understand this, consider a vector in Minkowski spacetime $V^\mu$, which transforms under the Lorentz transformation as $V^{\mu'} = \Lambda \indices{^{\mu' }_{\mu}} V^\mu$. We use the Einstein convention where the repeated indices are summed over. Now let's see how the partial derivative of a vector transforms in flat spacetime: \[\begin{aligned} \partial_{\nu '}V^{\mu '} = \Lambda \indices{_{\nu '}^{\nu}}\partial_{\nu} \left( \Lambda \indices{^{\mu '}_{\mu}} V^\mu \right) = \Lambda \indices{_{\nu '}^{\nu}}\Lambda \indices{^{\mu '}_{\mu}} \partial_{\nu} V^\mu. \end{aligned}\] In tensorial language, one says that $\partial_{\nu}V^{\mu}$ transforms as a tensor of rank (1,1), where $(n,m)$ denote $n$ indices on the top and $m$ indices on the bottom. In curved spacetime, this is no longer true because $\Lambda \indices{_{\nu '}^{\nu}} \rightarrow \tfrac{\partial x^{\nu '}}{\partial x^{\nu}}$ is generically a spacetime dependent quantity. The partial differential operator must therefore be generalised (to a covariant derivative), so that the 'differentiated' object is also a tensor. This requires the notion of parallel transport. ## Parallel transport and affine connection Mathematically, one has to be careful when taking a derivative of a vector because the definition of a vector along a curve defined by $\lambda$ \[\begin{aligned} \frac{d V^\mu}{d\lambda} = {\rm lim}_{\delta\lambda \rightarrow 0} \frac{V^\mu(Q)-V^\mu(P)}{\delta\lambda}, \end{aligned}\] where $Q$ and $P$ are spacetime points at $\lambda + \delta\lambda$ and $\lambda$ respectively, require comparison between two vectors in different tangent spaces. This is okay in flat spaces, but in curved spaces there is no intrinsic way to do it. What we need is a concept of parallel transport that moves a vector--or more generally a tensor--along a curve while keeping it 'constant'. $V^\mu(P)$ can be parallel transported to $Q$ with the help of the 'affine connection' $\Gamma$ such that \[\begin{aligned} \label{paralleltransport} V^\mu(P\rightarrow Q) = V^\mu(P)-\delta\Lambda \Gamma^\mu_{\nu\sigma}(P) V^\nu(P)\frac{dx^\sigma}{d\lambda}. \end{aligned}\] The corresponding covariant derivative is written as \[\begin{aligned} \nabla_\sigma V^\mu = \frac{V^\mu(Q)-V^\mu(P\rightarrow Q)}{\delta x^\sigma} = \partial_\sigma V^\mu + \Gamma^\mu_{\nu\sigma}V^\nu, \end{aligned}\] with which we can define the parallel transport condition as \[\begin{aligned} \frac{dx^\sigma}{d\lambda}\nabla_\sigma V^\mu = 0. \end{aligned}\] Note that this is a generalisation of the condition in flat space $\tfrac{\partial V^\mu}{\partial x^\sigma} = 0$. Covariant derivative of a one-form, i.e. a (0,1) tensor $\omega_\mu$ is found to be \[\begin{aligned} \nabla_\sigma \omega_{\nu} = \partial_\sigma \omega_\mu-\Gamma^\lambda_{\sigma\nu}\omega_\lambda. \end{aligned}\] The transformation properties of a connection can be found by demanding that $\nabla_\sigma V^\mu$ transforms as a tensor, i.e., \[\begin{aligned} \nabla_{\sigma '} V^{\mu '} = \frac{\partial x^\sigma}{\partial x^{\sigma '}} \frac{\partial x^{\mu '}}{\partial x^{\mu}} \nabla_\sigma V^\mu. \end{aligned}\] This yields \[\begin{aligned} \Gamma^{\mu'}_{\nu'\sigma'} = \frac{\partial x^{\nu}}{\partial x^{\nu'}}\frac{\partial x^{\sigma}}{\partial x^{\sigma'}}\frac{\partial x^{\mu'}}{\partial x^{\mu}}\Gamma^{\mu}_{\nu\sigma}-\frac{\partial x^{\nu}}{\partial x^{\nu'}}\frac{\partial x^{\sigma}}{\partial x^{\sigma'}}\frac{\partial}{\partial x^\sigma}\left( \frac{\partial x^{\mu'}}{\partial x^{\nu}}\right), \end{aligned}\] showing that $\Gamma$ is not a tensor. ## The metric connection The above consideration shows that the concept of parallel transport requires an affine connection to be defined. This is a very general property irrespective of the detailed structure of the manifold, allowing many distinct definitions of parallel transport. In general relativity, however, it is possible define a unique connection compatible with the metric $g_{\mu\nu}$ (remember, $ds^2 = g_{\mu\nu}dx^\mu dx^\nu$) as follows. First, the connection is assumed to be torsion-free, meaning that $\Gamma^\lambda_{\mu\nu} = \Gamma^\lambda_{\nu\mu}$. Second, the metric is assumed to obey the parallel transport condition: $\nabla_\sigma g_{\mu\nu} = 0$. The latter guarantees that the scalar product of two parallel-transported vectors is constant. That is, if $\tfrac{dx^\sigma}{d\lambda}\nabla_\sigma U^\mu = \tfrac{dx^\sigma}{d\lambda}\nabla_\sigma V^\nu = 0$, then $\tfrac{dx^\sigma}{d\lambda}\nabla_\sigma(g_{\mu\nu}U^\mu V^\nu) = 0$. From the second assumption, we have the following three relations \[\begin{aligned} \nabla_\rho g_{\mu\nu} = \partial_\rho g_{\mu\nu}-\Gamma^{\lambda}_{\rho\mu}g_{\lambda\nu}-\Gamma^{\lambda}_{\rho\nu}g_{\mu\lambda} = 0, \nonumber \\ \nabla_\mu g_{\nu\rho} = \partial_\mu g_{\nu\rho}-\Gamma^{\lambda}_{\mu\nu}g_{\lambda\rho}-\Gamma^{\lambda}_{\mu\rho}g_{\nu\lambda} = 0, \nonumber \\ \nabla_\nu g_{\rho\mu} = \partial_\nu g_{\rho\mu}-\Gamma^{\lambda}_{\nu\rho}g_{\lambda\mu}-\Gamma^{\lambda}_{\nu\mu}g_{\rho\lambda} = 0. \end{aligned}\] Subtracting the second and third lines from the first and using the first assumption, we obtain \[\begin{aligned} \partial_\rho g_{\mu\nu}-\partial_\mu g_{\nu\rho}-\partial_\nu g_{\rho\mu} + 2\Gamma^{\lambda}_{\mu\nu}g_{\lambda\rho} = 0, \end{aligned}\] which after some rearranging yields \[\begin{aligned} \Gamma^\sigma_{\mu\nu} = \frac{1}{2}g^{\sigma\rho}\left( \partial_\mu g_{\nu\rho} + \partial_\nu g_{\rho\mu}-\partial_\rho g_{\mu\nu} \right). \end{aligned}\] This is the metric connection, called the Christoffel symbol. ## Spin connection So far, we have seen how to take covariant derivatives of tensors. However, this is not enough to write down the Dirac equation in a curved spacetime. We also need to know how to take the covariant derivative of a spinor. For this purpose, we will use the fact that locally it is always possible (due to the equivalence principle) to find an inertial coordinate system in which the metric becomes Minkowskian. Suppose that $y^a$ are such local coordinates at point $x^\mu = X^\mu$ (we use the convention that latin indices $a,b...$ are used to label local inertial coordinates and greek indices $\mu,\nu...$ for general coordinates). Then defining \[\begin{aligned} e\indices{^\mu_a}(X) &= \frac{\partial x^\mu}{\partial y^a} \Bigg \|_{x^\mu = X^\mu} \;\;\;\;\; {\rm and} \nonumber \\ e\indices{^a_\mu}(X) &= \frac{\partial y^a}{\partial x^\mu} \Bigg \|_{x^\mu = X^\mu} \end{aligned}\] at each point in spacetime, we get the vielbein ('many-legs') $e^\mu_a(x)$ that diagonalises the metric, as well as its inverse \[\begin{aligned} e\indices{^\mu_a}(x)e\indices{^\nu_b}(x)g_{\mu\nu}(x) &= \eta_{ab}, \nonumber \\ e\indices{^a_\mu}(x)e\indices{^b_\nu}(x)\eta_{ab}(x) &= g_{\mu\nu}. \end{aligned}\] What we want is to know what the parallel transport equation ([\[paralleltransport\]](#paralleltransport){reference-type="ref" reference="paralleltransport"}) looks like in the local inertial coordinate system. Basically, we want \[\begin{aligned} \label{spintransport} V^a(x \rightarrow x + dx) = V^a(x)-\omega\indices{^a_b_\nu}V^b(x)dx^\nu, \end{aligned}\] where $\omega\indices{^a_b_\nu}$ is a generalisation of the affine connection called the spin connection. Noting that \[\begin{aligned} V^\mu(x) &= e\indices{^\mu_a}(x)V^a(x), \nonumber \\ V^\mu(x\rightarrow x+dx) &= e\indices{^\mu_a}(x+dx)V^a(x\rightarrow x+dx), \end{aligned}\] and $e\indices{^\mu_a}(x+dx) \approx e\indices{^\mu_a}(x) + \partial_\nu (e\indices{^\mu_a}(x))dx^\nu$, we can obtain the spin connection in terms of the affine connection \[\begin{aligned} \label{spinconnection} \omega\indices{^a_b_\nu} = e\indices{^a_\mu} \partial_\nu (e\indices{^\mu_b}) + e\indices{^a_\mu}e\indices{^\sigma_b}\Gamma^\mu_{\sigma\nu}. \end{aligned}\] This allows one to take the covariant derivative of a tensor with mixed indices. For example \[\begin{aligned} \nabla_\nu e\indices{^\mu_a} = \partial_\nu e\indices{^\mu_a} + \Gamma^\mu_{\sigma\nu}e\indices{^\sigma_a}-\omega \indices{^b_a_\nu} e\indices{^\mu_b} = 0, \end{aligned}\] where the last equality follows from ([\[spinconnection\]](#spinconnection){reference-type="ref" reference="spinconnection"}). The spin connection is antisymmetric in the first two indices, i.e., $\omega_{ab\nu} =-\omega_{ba\nu}$, so that the magnitude of the Lorentz vector remains constant upon parallel transport. Using the spin connection we can derive the covariant derivative operator for spinors. The latter is defined through the parallel transport equation \[\begin{aligned} \psi(x\rightarrow x + dx) = \psi(x)-\Omega_\nu(x)\psi(x)dx^\nu, \end{aligned}\] where $\Omega_\nu$ is an n-by-n matrix for each index $\nu$ with n=2 for 2 and 3 spacetime dimensions and 4 for 4-dimensional spacetime. To determine $\Omega_\nu$, we use the fact that $S(x) = \bar{\psi}(x)\psi(x)$ and $V^a(x) = \bar{\psi}(x)\gamma^a\psi(x)$ transform as a scalar and a vector, respectively. Firstly, \[\begin{aligned} S(x\rightarrow & x + dx)-S(x) \nonumber \\ &= \bar{\psi}(x)\left[ \gamma^0\Omega^\dagger_\nu(x)\gamma^0 + \Omega_\nu(x)\right]\psi(x) dx^\nu, \end{aligned}\] yielding \[\begin{aligned} \label{cond1} \gamma^0\Omega^\dagger_\nu(x)\gamma^0 =-\Omega_\nu(x). \end{aligned}\] Secondly, \[\begin{aligned} V^a(x\rightarrow & x + dx)-V^a(x) \nonumber \\ &=-\bar{\psi}(x)\left[ \gamma^a\Omega_\nu +\gamma^0\Omega^\dagger_\nu\gamma^0\gamma^a\right]\psi(x) dx^\nu, \nonumber \\ &=-\bar{\psi}(x)\left[ \gamma^a\Omega_\nu-\Omega_\nu\gamma^a\right]\psi(x) dx^\nu, \nonumber \\ &=-\omega\indices{^a_b_\nu}\gamma^bdx^\nu. \end{aligned}\] The second equality results from ([\[cond1\]](#cond1){reference-type="ref" reference="cond1"}), while the third equality results from the definition of the spin connection. From the second and third lines we conclude that \[\begin{aligned} \label{cond2} [\gamma^a,\Omega_\nu] = \omega\indices{^a_b_\nu}\gamma^b. \end{aligned}\] Noting that \[\begin{aligned} = 2i(\eta^{ab}\gamma^c-\eta^{ac}\gamma^b), \end{aligned}\] where $\sigma^{bc} = i[\gamma^b,\gamma^c]/2$, one can verify by direct substitution that \[\begin{aligned} \Omega_\nu(x) =-\frac{i}{4}\omega_{ab\nu}(x)\sigma^{ab} \end{aligned}\] satisfies ([\[cond2\]](#cond2){reference-type="ref" reference="cond2"}), while ([\[cond1\]](#cond1){reference-type="ref" reference="cond1"}) can be verified using the relationship $\gamma^0\gamma^{a\dagger}\gamma^0 =-\gamma^a$. Using the covariant derivative $\nabla_\nu = \partial_\nu + \Omega_\nu$, the flat spacetime Dirac equation (c=1) \[\begin{aligned} \left[i\gamma^a\partial_a\psi-m \right] \psi(x) = 0, \end{aligned}\] generalises to the curved spacetime Dirac equation \[\begin{aligned} \label{curveddiraceqn} \left[ i\gamma^\mu\nabla_\mu-m\right]\psi(x) = 0, \end{aligned}\] where the vielbein was used to transform the local $\gamma$ matrices: $\gamma^\mu = e\indices{^\mu_a}(x)\gamma^a$. Note that $\{ \gamma^\mu(x),\gamma^\nu(x) \} = 2g^{\mu\nu}(x)$. From here on, we will use the notation $\tilde{\gamma}^a \equiv \gamma^a$ to avoid confusion when numerical indices are substituted. # Dirac equation in 2 dimensional curved spacetime A special feature of the 2 dimensional spacetime is that the metric can always be reduced to the conformally flat form \[ds^2=\Omega^2(dt^2-dx^2)\] for some function $\Omega(x,t)$. To derive the Dirac equation in this metric we first need the Christoffel symbols which are readily calculated to be: $\Gamma^0_{00}=\Gamma^0_{11}=\Gamma^1_{10}=\Gamma^1_{01}=\dot{\Omega}/\Omega$ and $\Gamma^0_{01}=\Gamma^0_{10}=\Gamma^1_{00}=\Gamma^1_{11} =\Omega'/\Omega$. The dot denotes a derivative with respect to time and the prime a spatial derivative. Using these, the vielbein can be readily calculated: $e\indices{^0_0}= e\indices{^1_1}= 1/\Omega$. These lead to non-vanishing spin connections $\omega\indices{^0_1_0}=\omega\indices{^1_0_0}= \Omega'/\Omega$ and $\omega\indices{^0_1_1}=\omega\indices{^1_0_1}= \dot{\Omega}/\Omega$, which in turn lead to $\Omega_0= \Omega'/(4\Omega) [\tilde{\gamma}^0,\tilde{\gamma}^1]$ and $\Omega_1= \dot{\Omega}/(4\Omega) [\tilde{\gamma}^0,\tilde{\gamma}^1]$. That this spacetime is curved can be verified by calculating the Ricci curvature $R = g^{\mu\nu}R_{\mu\nu} = 2\left( (\dot{\Omega}/\Omega)^2-\ddot{\Omega}/\Omega\right)/\Omega^2$, where $R_{\mu\nu} = \partial_\lambda \Gamma^{\lambda}_{\mu\nu}-\partial_\nu \Gamma^{\lambda}_{\mu\lambda} + \Gamma^{\lambda}_{\mu\nu}\Gamma^{\sigma}_{\lambda\sigma}-\Gamma^{\lambda}_{\mu\sigma}\Gamma^{\sigma}_{\nu\lambda}$ is the Ricci tensor. We can thus write down the Dirac equation in a general 1+1 dimensional spacetime. Inserting the above results into Eq. ([\[curveddiraceqn\]](#curveddiraceqn){reference-type="ref" reference="curveddiraceqn"}), multiplying with $\Omega \tilde{\gamma}^0$ from the left, and rearranging, we get \[\begin{split} i(\partial_t+\tilde{\gamma}^0\tilde{\gamma}^1\frac{[\tilde{\gamma}^0,\tilde{\gamma}^1]}{4}\frac{\dot{\Omega}}{\Omega})\psi &= i[\tilde{\gamma}^0\tilde{\gamma}^1\partial_x + \frac{[\tilde{\gamma}^0,\tilde{\gamma}^1]}{4} \frac{\Omega'}{\Omega} ]\psi \\ &\quad+\tilde{\gamma}^0\Omega m\psi. \end{split}\] Choosing $\tilde{\gamma}^0 = \sigma_z$ and $\tilde{\gamma}^1= i\sigma_y$, the Dirac equation becomes \[i\left(\partial_t+\frac{\dot{\Omega}}{2\Omega}\right)\psi= \left[-i\sigma_x\left(\partial_x + \frac{\Omega'}{2\Omega} \right)+\sigma_z\Omega m\right]\psi(x).\] Which, upon defining $\sqrt{\Omega} \begin{equation} \label{curvedDirac} i\partial_t(\Omega \end{equation}\[$\begin{aligned} \label{flateqn} i\partial_t \psi=-i\sigma_x\partial_x\psi+\sigma_z m \psi. \end{aligned}\] ## Specific examples Here we provide examples of specific spacetimes and corresponding Dirac equations. ### Static spacetime Static spacetimes are described by a metric that has a time-like variable $x_0$ such that $g_{0i} = 0$ and $\partial_0 g_{\mu \nu} = 0$. In 1+1 dimensions the line element for these spacetimes may be written as \[ds^2=e^{2\Phi}dt^2-e^{2\Psi}dx^2,\] where $\Phi$ and $\Psi$ are independent of $t$. Non-vanishing Christoffel symbols for this metric are $\Gamma^0_{10}=\Gamma^0_{01}= \Phi'$, $\Gamma^1_{00}= \Phi'e^{2(\Phi-\Psi)}$, and $\Gamma^1_{11}= \Psi'$, whereas the vielbein is easily found to be $e\indices{^0_0}= e^{-\Phi}$, $e\indices{^1_1}= e^{-\Psi}$. Then the non-vanishing spin connections are $\omega\indices{^1_0_0}=\omega\indices{^0_1_0}= \Phi'e^{\Phi-\Psi}$, leaving us with the non-vanishing element of $\Omega_\nu$: $\Omega_0 = \frac14\Phi'e^{\Phi-\Psi}[\tilde{\gamma}^0,\tilde{\gamma}^1]$. Inserting the results and using the same gamma-matrices as above, we can write the Dirac equation as \[i\partial_t \psi =-ie^{\Phi-\Psi}\sigma_x\left(\partial_x+\frac{\Phi'}{2} \right)\psi + e^{\Phi}\sigma_z m\psi.\] ### FRW Metric The FRW metric in 1+1 D reads \[ds^2=dt^2-a^2(t) dx^2.\] The non-vanishing spin connections, Christoffel symbols, and $\Omega_{\nu}$ are $e\indices{^0_0}= 1$, $e\indices{^1_1}= 1/a$; $\Gamma^0_{11}= a\dot{a}$; $\Gamma^1_{10}=\Gamma^1_{01}= \dot{a}/{a}$; $\omega\indices{^0_1_1}=\omega\indices{^1_0_1}= \dot{a}$; $\Omega_1= \frac{\dot{a}}{4}[\tilde{\gamma}^0,\tilde{\gamma}^1]$. The Dirac equation reads \[i(\tilde{\gamma}^0 \partial_0 \psi + \tilde{\gamma}^1\frac1a \partial_x \psi+\tilde{\gamma}^1\frac{\dot{a}}{4a}[\tilde{\gamma}^0,\tilde{\gamma}^1] \psi)-m\psi = 0.\] After plugging in $\tilde{\gamma}^a$, we obtain \[i\partial_t \psi =-i\frac{\dot{a}}{2a} \psi-i\frac{\sigma_x}{a} \partial_x \psi+\sigma_z m\psi.\] Note that the FRW metric can be converted to the conformally flat form by setting $\eta(t) = \int^t \frac{dt}{a(t)}$. ### Rindler spacetime Rindler metric describes the dynamics of a uniformly accelerating observer in flat spacetime. Even though the spacetime is flat, Unruh showed that spontaneous particle creation occurs in the frame of the accelerating observer analogously to the famous Hawking radiation. The vielbein formalism described above proves useful for deriving the wave equation in this metric. By introducing the coordinate $t = u\sinh v$ and $x = u \cosh v$, the Minkowski metric is converted to \[\begin{aligned} ds^2 = u^2dv^2-du^2, \end{aligned}\] when $u$ is constrained to be positive. Because this is a static spacetime, we can use the formula derived above to obtain the Dirac equation in the Rindler spacetime: \[\begin{aligned} i\partial_v\psi =-iu\sigma_x\left( \partial_u +\frac{1}{2u} \right)\psi + mu\sigma_z\psi. \end{aligned}\] The metric can also be put in a conformally flat form by introducing new coordinates \[\begin{aligned} t = \frac{1}{a}e^{a \end{aligned}\] in the region $x>\|t\|$. In these coordinates, the metric takes the form \[\begin{aligned} ds^2 = e^{2a \end{aligned}\] and the Dirac equation reads \[\begin{aligned} i\partial_\eta\psi = \left[-i\sigma_x\left( \partial_ \end{aligned}\] ## Particle creation in curved spacetime {#sect3b} It is well established that there is particle creation in curved spacetime in general. The background metric acts in a similar manner to an external field such as, for example, the electromagnetic field, and something akin to the Schwinger effect (creation of electron-positron pair in strong electric fields) occur. Consider as an example an expanding-universe scenario where the expansion is asymptotically turned on. The field is initially in the vacuum state and as expansion is gradually turned on, electrons and positrons pop out in pairs. The essence of understanding this phenomena is that a vacuum state is not unique. It is defined as an eigen-state of the field operator $\psi$ with eigenvalue 0. To define the latter more precisely, one has to expand the field operator in terms of mode functions and assign creation and annihilation operators that create and annihilate these modes. The vacuum is then the state with eigenvalue 0 for all the mode-annihilation operators. Now, these mode functions depend on the background spacetime, which means that the vacuum state of one spacetime need not be the vacuum state of another. Below, we provide a more detailed and pedagogical explanation of this effect for scalar fields in FRW spacetime, before we come back to Dirac fermions. ### Scalar field in FRW spacetime A scalar field obeying the Klein-Gordon equation is one of the simplest quantum fields to deal with and a spatially-flat, time-dependent metric is one of the simplest examples of curved spacetime metrics. As an example of such a metric, we choose the FRW metric and explain creation of scalar fields in it. We follow closely the exposition by Mukhanov and Winitzki, but work in 2 dimensional spacetime instead of the usual 4 dimensional one. Remember that a real scalar field in Minkowski (or flat) spacetime obeys the Klein-Gordon equation \[\begin{aligned} \partial_\mu\partial^\mu \phi + m^2\phi = 0. \end{aligned}\] In curved spacetime this becomes (in the minimal coupling scheme) \[\begin{aligned} g^{\mu\nu}\nabla_\mu\nabla_\nu \phi + m^2\phi = 0. \end{aligned}\] To work out the explicit form of this equation, it is helpful to convert it to the following form \[\begin{aligned} g^{\mu\nu}\partial_\mu\partial_\nu\phi + \frac{1}{\sqrt{-g}}\left( \partial_\nu\phi\right) \partial_\nu \left( g^{\mu\nu}\sqrt{-g}\right) + m^2\phi = 0, \end{aligned}\] where $g$ is the determinant of the metric. Note that $g$ depends on the spacetime dimension: for the case of conformal spacetime described by $ds^2 = \Omega^2(dt^2-dx^2 \cdots )$, $g = \Omega^2d$, where $d$ is the dimension of the spacetime. Working with the conformal version of the FRW metric, the wave equation evaluates to \[\begin{aligned} \ddot{\phi}-\phi '' + \Omega(\eta)^2m^2\phi = 0 \end{aligned}\] in 1+1D and \[\begin{aligned} \ddot{\phi} + 2\frac{\dot{\Omega}}{\Omega}\dot{\phi}-\nabla\phi + \Omega(\eta)^2m^2\phi = 0 \end{aligned}\] in 3+1D. As before, the dot denotes derivative with respect to the conformal time $t$, whereas the prime denotes a spatial derivative. Going to the momentum space by defining \[\begin{aligned} \phi(x,t) = \frac{1}{\sqrt{2\pi}}\int dk \phi_k(t) e^{ikx}, \end{aligned}\] the wave equation reduces to \[\begin{aligned} \ddot{\phi}_k-k^2\phi_k + \Omega^2(t) m^2\phi_k \equiv \ddot{\phi}_k + \omega_k^2(t)\phi_k = 0. \end{aligned}\] The general solution of this time-dependent oscillator equation can be written in terms of complex mode functions $v_k(t)$ with the normalisation condition $-iW[v_k,v_k^]/2 \equiv-i(\dot{v}_k v_k^-v_k\dot{v}_k^)./2= 1$, where $W$ is called the Wronskian. The field operator $\phi_k$ can be expanded as \[\begin{aligned} \phi_k(t) = \frac{1}{2}\left[ a_k v_k^(t) + a^\dagger_k v_k(t) \right], \end{aligned}\] where $a_k$ and $a^\dagger_k$ are the bosonic annihilation and creation operators for mode $k$. It is interesting to note that the (bosonic) commutation relation is crucial for preserving the normalisation defined by the Wronskian. The anticommutation relation (for fermions) is not consistent with the latter. Once the mode expansion is defined, the vacuum state $\|0\rangle$ is defined by the condition $a_k\|0\rangle = 0$ for all $k$. Now consider two different mode functions $u_k(t)$ and $v_k(t)$. Since $u_k$ and $u_k^$ form a basis, we can expand $v_k$ in terms of them, $v_k^(t) = \alpha_k u_k^(t) + \beta_k u_k(t)$, with $t$-independent coefficients $\alpha_k$ and $\beta_k$ that obey the normalisation condition $\|\alpha_k\|^2-\|\beta_k\|^2 = 1$. If $b_k$ and $b_k^\dagger$ are the annihilation and creation operators corresponding to the mode functions $u_k$, the following relations hold \[\begin{aligned} b_k = \alpha_ka_k + \beta_k^ a_{-k}^\dagger, \;\;\;\;\; b_k^\dagger = \alpha_k^a_k^\dagger + \beta_k a_{-k}. \end{aligned}\] This is called the Bogolyubov transformation in the literature. What is interesting is that the vacuum state for $a$ is not the vacuum state for $b$. Indeed a simple calculation reveals that the expectation value of the number of $b$-particles in $a$'s vacuum state is generically non-zero: \[\begin{aligned} \langle 0_a \| b_k^\dagger b_k \|0_a\rangle &= \langle 0_a \| (\alpha_k^a_k^\dagger + \beta_k a_{-k}) (\alpha_k a_k + \beta_k^a_{-k}^\dagger) \|0_a\rangle \nonumber \\ &= \|\beta_k\|^2 \langle 0_a \| a_{-k}a_{-k}^\dagger \|0_a\rangle = \|\beta_k\|^2. \end{aligned}\] To repeat, if $\beta_k$ is non-zero, the vacuum state defined with respect to the $a_k$ modes contains a finite density of $b_k$ particles. It is generally impossible to find a unique vacuum state in curved spacetime unlike in flat (Minkowski) spacetime. To discuss particle creation, it is therefore a good idea to work with an asymptotically flat spacetime that has the Minkowski metric both in far-enough past and far-enough future. As a simple example, we here consider a specially chosen FRW spacetime with $\Omega^2(t) = 1$ if $t<0$ or $t>t_0$ and $\Omega^2(t)=-1$ if $0<t<t_0$. The 'in' and 'out' vacuum modes at $t<0$ and $t>t_0$ are described by the standard Minkowski mode functions \[\begin{aligned} v_k^{(in)}(t) = \frac{1}{\sqrt{\omega_k}}e^{i\omega_kt},\;\;\; v_k^{(out)}(t) = \frac{1}{\sqrt{\omega_k}}e^{i\omega_k(t-t_0)}, \end{aligned}\] with $\omega_k = \sqrt{k^2 + m^2}$. To find the Bogolyubov coefficients, we need to find the relationship between the input and output modes. Solving the mode equations in all regions, one obtains the following relation for $t>t_0$ \[\begin{aligned} v_k^{(in)}(t) = \frac{1}{\sqrt{\omega_k}}\left[ \alpha_k^ e^{i\omega_k(t-t_0)}+\beta_k^* e^{-i\omega_k(t-t_0)}\right], \end{aligned}\] where \[\begin{aligned} \alpha_k &= \tfrac{e^{-i\Omega_k t_0}}{4}\left( \sqrt{\tfrac{\omega_k}{\Omega_k}}+\sqrt{\tfrac{\Omega_k}{\omega_k}}\right)^2-\tfrac{e^{i\Omega_k t_0}}{4}\left( \sqrt{\tfrac{\omega_k}{\Omega_k}}-\sqrt{\tfrac{\Omega_k}{\omega_k}}\right)^2, \nonumber \\ \beta_k &= \frac{1}{2}\left(\tfrac{\Omega_k}{\omega_k}-\tfrac{\omega_k}{\Omega_k}\right)\sin(\Omega_kt_0), \end{aligned}\] with $\Omega = \sqrt{k^2-m^2}$. This tells us that the field mode that has evolved from $\|0_{in}\rangle$ is different from the vacuum state at $t>t_0$ and has a finite particle number density $n_k = \|\beta_k\|^2 = \tfrac{m^4}{\|k^4-m^4\|}\|\sin\left( t_0(k^2-m^2)\right)\|^2$ ### Dirac field in curved spacetime, a single particle analog of particle creation The above procedure applies in exactly the same way for Dirac fields, except that one has to use the anticommutation relation instead of the commutation relation when quantizing the field operators. One other difference is that particles are produced in pairs so that the total charge is conserved. Because the mode expansion is a little more involved, and we are actually interested in the single-particle manifestation of pair creation, we will not go into the details here. Interested readers are referred to the original article by Parker. Instead let us discuss the single-particle manifestation of pair creation. In order to see how this is possible, consider the pair creation process due to an electric field (Schwinger effect) in the Dirac sea picture. Initially, the negative-energy sea is fully occupied. Then upon applying a strong electric field, a negative energy electron is kicked (or tunnels) to occupy a positive energy state, leaving a hole (positron) behind. Therefore, in terms of single-particle dynamics, what we will see is a conversion of a negative energy wave packet to a positive energy wave packet. In the next section we verify this claim, by studying the dynamics of a wave packet in curved spacetime. # Dynamics of spinor wave packet {#dynamics} In this section, we provide a basic background on the wave packet dynamics, both in flat and curved spacetime. In particular, we show a single-particle analog of particle creation in an FRW spacetime. The results from this section will be used in the next section, where we discuss the implementation of wave packet dynamics in binary waveguide arrays. ## Flat spacetime The Dirac equation in flat 2 dimensional space time was written in Eq. ([\[flateqn\]](#flateqn){reference-type="ref" reference="flateqn"}), which we repeat here for the reader's convenience \[\begin{aligned} i\partial_t \psi=-i\sigma_x\partial_x\psi+\sigma_z m \psi. \end{aligned}\] Using the plane wave ansatz, the eigen-solutions can be easily found. The two solutions with positive and negative energies can be written as \[\begin{aligned} \label{eestates} \psi_+^k(x,t) &= \frac{1}{\sqrt{2 \pi}}\frac{\exp [-iE_kt + ikx]}{\sqrt{2E_k(E_k + m)}}\begin{pmatrix} E_k+m \\ k \end{pmatrix}, \nonumber \\ \psi_-^k(x,t) &= \frac{1}{\sqrt{2 \pi}}\frac{\exp [ iE_kt + ikx]}{\sqrt{2E_k(E_k + m)}}\begin{pmatrix}-k \\ E_k + m \end{pmatrix}. \end{aligned}\] The time evolution of an initial spinor wave packet, $(\phi_1(k),\phi_2(k))$, is given by \[\begin{aligned} \psi(x,t) = \int dk \tfrac{\alpha (E_k+m)\phi_1(k) + \beta k\phi_2(k) }{2 \sqrt{2\pi} E_k(E_k + m)}\begin{pmatrix} E_k +m \\ k \end{pmatrix} e^{i(kx-E_kt)} \nonumber \\ +\int dk \tfrac{-\alpha k \phi_1(k) + \beta (E_k+m)\phi_2(k) }{2 \sqrt{2\pi} E_k(E_k + m)}\begin{pmatrix}-k \\ E_k +m \end{pmatrix} e^{i(kx+E_kt)}. \end{aligned}\] Let us first look at the dynamics of a Gaussian wave packet $\psi_0(x) \propto \exp[-x^2/18](1,1)^T$. Figure [\[flat1\]](#flat1){reference-type="ref" reference="flat1"}(a) shows the time evolution of the wave packet for $m=1$. Notice the zig-zag motion of the centre of mass as exemplified in Fig. [\[flat1\]](#flat1){reference-type="ref" reference="flat1"}(b). This phenomenon is called zitterbewegung and occurs because of the interference between the positive-and negative-energy components of the spinor. Properties of zitterbewegung can be seen from the position operator in the Heisenberg picture. Noting that $v(t) =-i[x,H] = \sigma_x(t)$ and \[\begin{aligned} \frac{\partial\sigma_x}{\partial t} =-i\left[ \sigma_x,H\right] =-2i\left[p-H\sigma_x \right], \end{aligned}\] we obtain \[\begin{aligned} \sigma_x(t) = H^{-1}p + e^{2iHt}\left[ \sigma_x(0)-H^{-1}p\right]. \end{aligned}\] $x(t) = \int v(t)$ is then given by \[\begin{aligned} x(t) = x(0) + \frac{p}{H}t-\frac{i}{2H}\left( e^{2iHt}-1\right) \left[ \sigma_x(0)-\frac{p}{H}\right]. \end{aligned}\] From the last term of this expression we notice several things. 1) the frequency of the oscillation is $2E$, 2) the amplitude is proportional to 1/(2E), and 3) zitterbewegung is non-zero only if there is a superposition between positive and negative energy states with equal momentum. The last observation follows because $\sigma_x-p/H$ anti-commutes with the Hamiltonian, which means that the matrix element of it is non-zero only between the eigenstates with equal momentum and opposite energies. Note that for a small initial momentum, the amplitude and frequency of ZB is thus 2m and 1/(2m) respectively. This is demonstrated in Fig. [\[ZB\]](#ZB){reference-type="ref" reference="ZB"}. Absence of zitterbewegung in a wave packet composed of positive-energy spinors only is shown in Fig. [\[flat2\]](#flat2){reference-type="ref" reference="flat2"}. The positive energy spinor is constructed by superposing the positive-energy spinor in momentum space with a Gaussian weight $\exp[-k^2/2]$. ## Curved spacetime {#sect:curvedDirac} We have seen that in curved spacetime, the Dirac equation is given by Eq. ([\[curvedDirac\]](#curvedDirac){reference-type="ref" reference="curvedDirac"}): \[i\partial_t(\Omega\] which we write as \[i\partial_t\psi=-i\sigma_x\partial_x\psi+\sigma_z m_{eff}(t)\psi,\] Here we concentrate on the FRW spacetime with a time-dependent conformal factor $a(t) \equiv \Omega(t)$. Neglecting the overall factor $\Omega(t)$, this equation is simply the Dirac equation with a time-dependent mass term. Analogously to a real scalar field in an FRW spacetime, Dirac fermions are produced (in pairs) in generic cases, which we will show by demonstrating conversion of a negative energy wavepacket to a positive energy wave packet. Quantitatively, the conversion can be proven by calculating the norm of the positive-energy-projected spinor wave packet, but we can do better: We can make use of the fact that ZB only occurs when positive-energy spinors are superposed with negative-energy spinors. So our aim is to show how ZB is induced in an asymptotically flat spacetime with an initial negative-energy wave packet. It turns out that physically interesting scenarios such as expanding spacetime (anti-de Sitter spacetime) produce only tiny effects that are unobservable, so instead we employ the 'inverted square hat' profile of $m_{eff}(t)$ already introduced earlier in Sect. [3.2](#sect3b){reference-type="ref" reference="sect3b"}. Actually, we will use the smoothed version of this, by replacing the 'inverted square hat' by an inverted Gaussian profile. Evidence of particle creation in this scenario is evident in Fig. [\[curved1\]](#curved1){reference-type="ref" reference="curved1"}(a) where the induction of ZB by the excursion of $m_{eff}$ from it asymptotic value 1, is clearly visible. The initial state was constructed by superposing $\psi_-^k(x,0)$ from Eq. ([\[eestates\]](#eestates){reference-type="ref" reference="eestates"}) with a Gaussian weight $\propto \exp[-4(k-0.1)^2]$. # Simulation of the Dirac equation in binary waveguide arrays In this section, we show how to simulate the 2 dimensional Dirac equation in binary waveguide arrays. We start off with a description of the coupled-mode theory of light propagation in a waveguide array and then demonstrate an equivalence between the discretised Dirac equation and the coupled-mode equations for a binary waveguide array. In particular, we show that the Dirac equation in 2 dimensional FRW spacetime can be straightforwardly implemented, paving the way towards experimental demonstration of the single particle analog of gravitational pair creation. ## Coupled-mode theory of waveguide arrays Propagation of an electromagnetic field $E$ in a medium is described by \[\nabla^2 E-\frac{n^2}{c^2} \frac{\partial^2 E}{\partial t^2} = 0,\] where $n$ is the refractive index of the medium, spatially varying in general. For a monochromatic field $E$ propagating predominantly in the $z-$direction, i.e., \[E= Re[E_0({\bf r})e^{i(kz-wt)}],\] one can make the so-called paraxial approximation which amounts to assuming $\|\nabla^2E_0\| \ll \|k_z \partial_z E_0\|$. In this case, the wave equation becomes the paraxial Helmholtz equation \[k \frac{\partial E_0}{\partial z} + \frac{i}{2}\left( k^2-\omega^2 \frac{n^2(r)}{c^2} \right) E_0-\frac{i}{2} \nabla^2_{\perp} E_0 = 0,\] $\nabla^2_\perp$ denoting the Laplacian in $x$ and $y$ directions. Let us choose $\omega/k = v_0 \equiv c/n_0$, where $v_0$ is the mean velocity in the medium and $n_0$ the mean refractive index. Then the above equation becomes \[\begin{aligned} 2ik\partial_z E_0 + \nabla_\perp^2E_0-\frac{2k^2}{n_0}\Delta n E_0 = 0, \end{aligned}\] where we have assumed that $n_0 \approx n$ and $\Delta n\equiv n_0-n$. Lastly, define the reduced wavelength $\lambdabar = n_0/k$ to rewrite the equation as \[\begin{aligned} i\lambdabar \partial_z E_0 =-\frac{\lambdabar^2}{2n_0}\nabla^2_\perp E_0 + \Delta n E_0. \end{aligned}\] Written this way, resemblance to the Schrödinger equation is obvious, the role of time being played by the spatial dimension $z$. Now imagine periodically modulating the index of refraction in the plane perpendicular to the z-direction. An EM field is attracted to regions of increased index of refraction and mainly stays in the vicinity of these regions during propagation. The field, however, is not completely confined to these regions but leaks into the area between the waveguides, with evanescent tails. From the similarity with the Schrödinger equation, it is clear that one can apply the tight-binding approximation to describe the propagation of the EM field in this case. In optics, this is called the coupled mode approximation (see, which we are closely following). In a one dimensional lattice we get for the amplitude $c_n$ in the $n$th waveguide: \[i\frac{\partial c_n}{\partial z}= k_n \left( c_{n+1} + c_{n-1} \right),\] where $k_n$ is the coupling strength, determined by the overlap between the transverse components of the modes in adjacent guides. Introducing yet another modulation such that alternating lattice sites have deep and shallow 'potentials', one obtains the following coupled mode equations \[\label{cmeqn} i\frac{\partial c_n}{\partial z}= k_n \left( c_{n+1} + c_{n-1} \right) + (-1)^n \sigma c_n.\] In the next section we show that this equation is equivalent to the discretised Dirac equation. ## Discretising the Dirac equation ### Flat spacetime To simulate the Dirac equation in flat 2D spacetime, \[i\frac{\partial \psi}{\partial t}=\left[-i\sigma_x{\partial_x}+m\sigma_z\right]\psi,\] in a waveguide array, we need to first change the differential equation to a difference equation by discretising the spatial coordinates. Choosing a discretisation length $d$, we define \[\begin{aligned} \psi(x,t) &\rightarrow \psi(nd,t) \equiv \tilde{\psi}(n,t), \nonumber \\ {\partial_x}\psi_1(x,t) &\rightarrow \frac{ \tilde{\psi}_1(n,t)-\tilde{\psi}_1(n-1,t)}{d}, \nonumber \\ {\partial_x}\psi_2(x,t) &\rightarrow \frac{ \tilde{\psi}_2(n+1,t)-\tilde{\psi}_2(n,t)}{d}, \end{aligned}\] where $n \in \mathbb{Z}$. To put this in the form of coupled mode equations, let us first define $\tilde{c}_{2n}(t) \equiv \tilde{\psi}_1(n,t)$ and $\tilde{c}_{2n-1} \equiv \tilde{\psi}_2(n,t)$. Then the Dirac equation becomes \[\begin{aligned} i\dot{\tilde{c}}_{2n}(t) =-\frac{i}{d}\left( \tilde{c}_{2n+1}(t)-\tilde{c}_{2n-1}(t) \right) + m\tilde{c}_{2n}(t), \nonumber \\ i\dot{\tilde{c}}_{2n-1}(t) =-\frac{i}{d}\left( \tilde{c}_{2n}(t)-\tilde{c}_{2n-2}(t) \right)-m\tilde{c}_{2n-1}(t). \end{aligned}\] To change the difference into the sum, we further define $c_{2n} = (-1)^n\tilde{c}_{2n} = (-1)^n\tilde{\psi}_1(n,t)$ and $c_{2n-1} =-i(-1)^n\tilde{c}_{2n-1} =-i(-1)^n\tilde{\psi}_2(n,t)$. The above equation changes to \[\begin{aligned} i\dot{c}_{2n}(t) = \frac{1}{d}\left( c_{2n+1}(t) + c_{2n-1}(t) \right) + mc_{2n}(t), \nonumber \\ i\dot{c}_{2n-1}(t) = \frac{1}{d}\left( c_{2n}(t) + c_{2n-2}(t) \right)-mc_{2n-1}(t), \end{aligned}\] which can be combined to \[i\dot{c_l}=-\frac{1}{d}(c_{l-1}+c_{l+1})+(-1)^{l} m c_l.\] Figure [\[assignspinorcomp\]](#assignspinorcomp){reference-type="ref" reference="assignspinorcomp"} provides a schematic illustration of how the spinor components are assigned to waveguides. The above equation is equivalent to Eq. ([\[cmeqn\]](#cmeqn){reference-type="ref" reference="cmeqn"}) given that $k_n = 1/d$ and $\sigma = m$, as first noted by Longhi in his proposal to simulate zitterbewegung and Klein's paradox. Subsequently, these effects, which are predicted by the single-particle Dirac equation, have been observed in laser-written waveguide arrays. ### Curved spacetime {#curved-spacetime} To facilitate the simulation of the Dirac equation in 2 dimensional curved spacetime, we need to discretise Eq. ([\[curvedDirac\]](#curvedDirac){reference-type="ref" reference="curvedDirac"}). In an optical simulation of single particle physics, the overall factor $\Omega$ is irrelevant, which means that we have at hand the Dirac equation with a spacetime-dependent mass as already noted in Sect. [4.2](#sect:curvedDirac){reference-type="ref" reference="sect:curvedDirac"}. The generalisation of the coupled-mode equations Eq. ([\[cmeqn\]](#cmeqn){reference-type="ref" reference="cmeqn"}) is then simple: \[\begin{aligned} i\frac{\partial c_n}{\partial z}= k_n \left( c_{n+1} + c_{n-1} \right) + (-1)^n \sigma_n (z) c_n. \end{aligned}\] ## Optical simulation Let us start with flat spacetime. Figure [\[simflat1\]](#simflat1){reference-type="ref" reference="simflat1"} shows the evolution of a Gaussian spinor wave packet $\propto \exp[-x^2/18](1,1)^T$ when $m=1$. Simulation results for $N=502$ (requiring $k \approx 6.2$) waveguides are displayed in Fig. [\[simflat1\]](#simflat1){reference-type="ref" reference="simflat1"}(a) and (b). They are in excellent agreement with the results in Sect. [4](#dynamics){reference-type="ref" reference="dynamics"}. However 502 is quite a large number and experiments are usually implemented with a much smaller number. To show the effects of discretisation we depict analogous results for $N=50$ ($k \approx 0.63$) in Fig. [\[simflat1\]](#simflat1){reference-type="ref" reference="simflat1"}(c) and (d). Apart from coarse graining effects in visualisation, the wave packet evolution is seen to be remarkably accurate as exemplified by the average position. Next, we simulate the conversion of a negative energy wave packet into a mixture of positive and negative energy wave packets. The latter is made by an arbitrary superposition of a negative energy eigen-spinor \[\begin{aligned} \phi_-^k = \frac{1}{\sqrt{2E_k(E_k + m)}}\begin{pmatrix}-k \\ E_k + m \end{pmatrix}. \end{aligned}\] In Fig. [\[simflat2\]](#simflat2){reference-type="ref" reference="simflat2"}, we consider the evolution of a Gaussian-averaged spinor, $\propto \int dk \exp[-(k-k_0)^2/2\sigma_k^2] \phi_-^k$ with the width $\sigma_k = 1/(2\sqrt{2})$, $k_0 = 0.1$, and mass $m = 1$. There is a small amount of ZB as a result of discretisation (that goes away with the increasing number of waveguides), but the magnitude is quite small. Finally, we show the evolution of a negative energy spinor in a FRW spacetime in Fig. [\[simcurved\]](#simcurved){reference-type="ref" reference="simcurved"}, with the inverted Gaussian conformal factor as used in Fig. [\[curved1\]](#curved1){reference-type="ref" reference="curved1"}. We see a good agreement with the exact numerical result shown in Fig. [\[curved1\]](#curved1){reference-type="ref" reference="curved1"}, signifying the feasibility of optical simulation of particle creation. # Conclusion We gave a pedagogical introduction to the Dirac equation in background curved spacetime and particle creation. Using the fact that the Dirac equation in the FRW metric is equivalent to the flat-spacetime Dirac equation with a time-dependent mass term, we demonstrated that a single-particle analog of particle creation can be observed in the dynamical evolution of spinor wave packets. In particular, we showed how a negative energy spinor gets mixed with a positive energy spinor when the conformal factor changes in time. Finally, we demonstrated that the Dirac equation in curved spacetime can be simulated in binary waveguide arrays, allowing direct experimental simulation of particle creation in curved spacetime. Although our example was for a time-dependent conformal factor, a general spacetime dependence can be easily simulated in waveguide arrays. Acknowledgments. D.G.A would like to acknowledge the financial support provided by the National Research Foundation and Ministry of Education Singapore (partly through the Tier 3 Grant "Random numbers from quantum processes" (MOE2012-T3-1-009)), and travel support by the EU IP-SIQS.	{'timestamp': '2017-05-17T02:02:43', 'yymm': '1607', 'arxiv_id': '1607.04821', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04821'}
null	null
# Introduction The concept of synchronistion is based on the adjustment of rhythms of oscillating systems due to their interaction. Synchronisation phenomenon was recognised by Huygens in the 17th century, time when he performed experiments to understand this phenomenon. To date, several kinds of synchronisation among coupled systems were reported, such as complete, phase, lag, and collective almost synchronisation. Neuronal synchronous rhythms have been observed in a wide range of researches about cognitive functions. Electroencephalography and magnetoencephalography studies have been suggested that neuronal synchronization in the gamma frequency plays a functional role for memories in humans. Steinmetz et al. investigated the synchronous behaviour of pairs of neurons in the secondary somatosensory cortex of monkey. They found that attention modulates oscillatory neuronal synchronisation in the somatosensory cortex. Moreover, in the literature it has been proposed that there is a relationship between conscious perception and synchronisation of neuronal activity. We study spiking and bursting synchronisation between neuron in a neuronal network model. A spike refers to the action potential generated by a neuron that rapidly rises and falls, while bursting refers to a sequence of spikes that are followed by a quiescent time. It was demonstrated that spiking synchronisation is relevant to olfactory bulb and is involved in motor cortical functions. The characteristics and mechanisms of bursting synchronisation were studied in cultured cortical neurons by means of planar electrode array. Jefferys $\&$ Haas discovered synchronised bursting of CA1 hippocampal pyramidal cells. There is a wide range of mathematical models used to describe neuronal activity, such as the cellular automaton, the Rulkov map, and differential equations. One of the simplest mathematical models and that is widely used to depict neuronal behaviour is the integrate-and-fire, which is governed by a linear differential equation. A more realistic version of it is the adaptive exponential integrate-and-fire (aEIF) model which we consider in this work as the local neuronal activity of neurons in the network. The aEIF is a two-dimensional integrate-and-fire model introduced by Brette $\&$ Gerstner . This model has an exponential spike mechanism with an adaptation current. Touboul $\&$ Brette studied the bifurcation diagram of the aEIF. They showed the existence of the Andronov-Hopf bifurcation and saddle-node bifurcations. The aEIF model can generate multiple firing patterns depending on the parameter and which fit experimental data from cortical neurons under current stimulation. In this work, we focus on the synchronisation phenomenon in a randomly connected network. This kind of network, also called Erdös-Rényi network, has nodes where each pair is connected according to a probability. The random neuronal network was utilised to study oscillations in cortico-thalamic circuits and dynamics of network with synaptic depression. We built a random neuronal network with unidirectional connections that represent chemical synapses. We show that there are clearly separated ranges of parameters that lead to spiking or bursting synchronisation. In addition, we analyse the robustness to external perturbation of the synchronisation. We verify that bursting synchronisation is more robustness than spiking synchronisation. However, bursting synchronisation requires larger chemical synaptic strengths, and larger voltage potential relaxation reset to appear than those required for spiking synchronisation. This paper is organised as follows: in Section II we present the adaptive exponential integrate-and-fire model. In Section III, we introduce the neuronal network with random features. In Section IV, we analyse the behaviour of spiking and bursting synchronisation. In the last Section, we draw our conclusions. # Adaptive exponential integrate-and-fire As a local dynamics of the neuronal network, we consider the adaptive exponential integrate-and-fire (aEIF) model that consists of a system of two differential equations given by \[\begin{aligned} \label{eqIF} C \frac{d V}{d t} & = &-g_L (V-E_L) + {\Delta}_T \exp \left(\frac{V-V_T}{{\Delta}_T} \right) \nonumber \\ & & +I-w , \nonumber \\ \tau_w \frac{d w}{d t} & = & a (V-E_L)-w, \end{aligned}\] where $V(t)$ is the membrane potential when a current $I(t)$ is injected, $C$ is the membrane capacitance, $g_L$ is the leak conductance, $E_L$ is the resting potential, $\Delta_T$ is the slope factor, $V_T$ is the threshold potential, $w$ is an adaptation variable, $\tau_w$ is the time constant, and $a$ is the level of subthreshold adaptation. If $V(t)$ reaches the threshold $V_{\rm{peak}}$, a reset condition is applied: $V\rightarrow V_r$ and $w\rightarrow w_r=w+b$. In our simulations, we consider $C=200.0$pF, $g_L=12.0$nS, $E_L=-70.0$mV, ${\Delta}_T=2.0$mV, $V_T=-50.0$mV, $I=509.7$pA, $\tau_w=300.0$ms, $a=2.0$nS, and $V_{\rm{peak}}=20.0$mV. The firing pattern depends on the reset parameters $V_r$ and $b$. Table [1](#table1){reference-type="ref" reference="table1"} exhibits some values that generate five different firing patterns (Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}). In Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"} we represent each firing pattern with a different colour in the parameter space $b\times V_r$: adaptation in red, tonic spiking in blue, initial bursting in green, regular bursting in yellow, and irregular in black. In Figs. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}a, [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}b, and [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}c we observe adaptation, tonic spiking, and initial burst pattern, respectively, due to a step current stimulation. Adaptation pattern has increasing inter-spike interval during a sustained stimulus, tonic spiking pattern is the simplest regular discharge of the action potential, and the initial bursting pattern starts with a group of spikes presenting a frequency larger than the steady state frequency. The membrane potential evolution with regular bursting is showed in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}d, while Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}e displays irregular pattern. [\[table1\]]{#table1 label="table1"} As we have interest in spiking and bursting synchronisation, we separate the parameter space into a region with spike and another with bursting patterns (Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}). To identify these two regions of interest, we use the coefficient of variation (CV) of the neuronal inter-spike interval (ISI), that is given by \[\begin{aligned} \label{CV} {\rm CV}=\frac{{\sigma}_{\rm{ISI}}}{\rm{\overline{ISI}}}, \end{aligned}\] where ${\sigma}_{\rm{ISI}}$ is the standard deviation of the ISI normalised by the mean $\bar{\rm ISI}$. Spiking patterns produce $\rm{CV}<0.5$. Parameter regions that represent the neurons firing with spiking pattern are denoted by gray colour in Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}. Whereas, the black region represents the bursting patterns, which results in $\rm{CV} \geq 0.5$. # Spiking or bursting synchronisation In this work, we constructed a network where the neurons are randomly connected. Our network is given by \[\begin{aligned} \label{eqIFrede} C \frac{d V_i}{d t} & = &-g_L (V_i-E_L) + {\Delta}_T \; \rm{exp} \left(\frac{V_i-V_T}{{\Delta}_T} \right) \nonumber \\ & + & I_i-w_i + g_{\rm{ex}} (V_{\rm{ex}}-V_i) \sum_{j=1}^N A_{ij} s_j + \Gamma_i, \nonumber \\ \tau_w \frac{d w_i}{d t} & = & a_i (V_i-E_L)-w_i, \nonumber \\ \tau_{\rm{ex}} \frac{d s_i}{d t} & = &-s_i. \end{aligned}\] where $V_i$ is the membrane potential of the neuron $i$, $g_{\rm{ex}}$ is the synaptic conductance, $V_{\rm{ex}}$ is the synaptic reversal potential, $\tau_{\rm{ex}}$ is the synaptic time constant, $s_i$ is the synaptic weight, $A_{ij}$ is the adjacency matrix, $\Gamma_i$ is the external perturbation, and $a_i$ is randomly distributed in the interval $[1.9,2.1]$. The schematic representation of the neuronal network that we have considered is illustrated in Fig [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}. Each neuron is randomly linked to other neurons with a probability $p$ by means of directed connections. When $p$ is equal to 1, the neuronal network becames an all-to-all network. A network with this topology was used by Borges et al. to study the effects of the spike timing-dependent plasticity on the synchronisation in a Hodgkin-Huxley neuronal network. A useful diagnostic tool to determine synchronous behaviour is the complex phase order parameter defined as \[z(t)=R(t)\exp({\rm i}\Phi(t))\equiv\frac{1}{N}\sum_{j=1}^{N}\exp({\rm i}\psi_{j}),\] where $R$ and $\Phi$ are the amplitude and angle of a centroid phase vector, respectively, and the phase is given by \[\psi_{j}(t)=2\pi m+2\pi\frac{t-t_{j,m}}{t_{j,m+1}-t_{j,m}},\] where $t_{j,m}$ corresponds to the time when a spike $m$ ($m=0,1,2,\dots$) of a neuron $j$ happens ($t_{j,m}< t < t_{j,m+1}$). We have considered the beginning of the spike when $V_j>-20$mV. The value of the order parameter magnitude goes to 1 in a totally synchronised state. To study the neuronal synchronisation of the network, we have calculated the time-average order-parameter, that is given by \[\overline{R}=\frac{1}{t_{\rm fin}-{t_{\rm ini}}}\sum_{t_{\rm ini}}^{t_{\rm fin}}R(t),\] where $t_{\rm fin}-t_{\rm ini}$ is the time window for calculating $\bar{R}$. Figs. [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}a, [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}b, and [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}c show the raster plots for $g_{\rm ex}=0.02$nS, $g_{\rm ex}=0.19$nS, and $g_{\rm ex}=0.45$nS, respectively, considering $V_r=-58$mV, $p=0.5$, and $b=70$pA, where the dots correspond to the spiking activities generated by neurons. For $g_{\rm ex}=0.02$nS (Fig. [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}a) the network displays a desynchonised state, and as a result, the order parameter values are very small (black line in Fig. [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}d). Increasing the synaptic conductance for $g_{\rm ex}=0.19$nS, the neuronal network exhibits spike synchronisation (Fig. [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}b) and the order parameter values are near unity (red line in Fig. [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}d). When the network presents bursting synchronisation (Fig. [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}c), the order parameter values vary between $R\approx 1$ and $R\ll 1$ (blue line in Fig. [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}d). $R\ll 1$ to the time when the neuron are firing. In Fig. [\[fig5\]](#fig5){reference-type="ref" reference="fig5"}a we show ${\bar R}$ as a function of $g_{\rm ex}$ for $p=0.5$, $b=50$pA (black line), $b=60$pA (red line), and $b=70$pA (blue line). The three results exhibit strong synchronous behaviour (${\bar R}>0.9$) for many values of $g_{\rm ex}$ when $g_{\rm ex}\gtrsim 0.4$nS. However, for $g_{\rm ex}\lesssim 0.4$nS, it is possible to see synchronous behaviour only for $b=70$pA in the range $0.15{\rm nS}<g_{\rm ex}<0.25{\rm nS}$. In addition, we calculate the coefficient of variation (CV) to determine the range in $g_{\rm ex}$ where the neurons of the network have spiking or bursting behaviour (Fig. [\[fig5\]](#fig5){reference-type="ref" reference="fig5"}b). We consider that for CV$<0.5$ (black dashed line) the neurons exhibit spiking behaviour, while for CV$\geq 0.5$ the neurons present bursting behaviour. We observe that in the range $0.15{\rm nS}<g_{\rm ex}<0.25{\rm nS}$ for $b=70$pA there is spiking sychronisation, and bursting synchronisation for $g_{\rm ex}\gtrsim 0.4$nS. # Parameter space of synchronisation The synchronous behaviour depends on the synaptic conductance and the probability of connections. Fig. [\[fig6\]](#fig6){reference-type="ref" reference="fig6"} exhibits the time-averaged order parameter in colour scale as a function of $g_{\rm ex}$ and $p$. We verify a large parameter region where spiking and bursting synchronisation is strong, characterised by ${\bar R}>0.9$. The regions I and II correspond to spiking and bursting patterns, respectively, and these regions are separated by a white line with circles. We obtain the regions by means of the coefficient of variation (CV). There is a transition between region I and region II, where neurons initially synchronous in the spike, loose spiking synchronicity to give place to a neuronal network with a regime of bursting synchronisation. We investigate the dependence of spiking and bursting synchronisation on the control parameters $b$ and $V_r$. To do that, we use the time average order parameter and the coefficient of variation. Figure [\[fig7\]](#fig7){reference-type="ref" reference="fig7"} shows that the spike patterns region (region I) decreases when $g_{\rm ex}$ increases. This way, the region I for $b<100$pA and $V_r=-49$mV of parameters leading to no synchronous behaviour (Fig. [\[fig7\]](#fig7){reference-type="ref" reference="fig7"}a), becomes a region of parameters that promote synchronised bursting (Fig. [\[fig7\]](#fig7){reference-type="ref" reference="fig7"}b and [\[fig7\]](#fig7){reference-type="ref" reference="fig7"}c). However, a large region of desynchronised bursting appears for $g_{\rm ex}=0.25$nS about $V_r=-45$mV and $b>100$pA in the region II (Fig. [\[fig7\]](#fig7){reference-type="ref" reference="fig7"}b). For $g_{\rm ex}=0.5$nS, we see, in Fig. [\[fig7\]](#fig7){reference-type="ref" reference="fig7"}c, three regions of desynchronous behaviour, one in the region I for $b<100$pA, other in region II for $b<200$pA, and another one is located around the border (white line with circles) between regions I and II for $b>200$pA. It has been found that external perturbations on neuronal networks not only can induce synchronous behaviour, but also can suppress synchronisation. Aiming to study the robustness to perturbations of the synchronous behaviour, we consider an external perturbation $\Gamma_i$ ([\[eqIFrede\]](#eqIFrede){reference-type="ref" reference="eqIFrede"}). It is applied on each neuron $i$ with an average time interval of about $10$ms and with a constant intensity $\gamma$ during $1$ms. Figure [\[fig8\]](#fig8){reference-type="ref" reference="fig8"} shows the plots $g_{\rm ex} \times p$ for $\gamma>0$, where the regions I and II correspond to spiking and bursting patterns, respectively, separated by white line with circles, and the colour bar indicates the time-average order parameter values. In this Figure, we consider $V_r=-58$mV, $b=70$pA, (a) $\gamma=250$pA, (b) $\gamma=500$pA, and (c) $\gamma=1000$pA. For $\gamma=250$pA (Fig. [\[fig8\]](#fig8){reference-type="ref" reference="fig8"}a) the perturbation does not suppress spike synchronisation, whereas for $\gamma=500$pA the synchronisation is completely suppressed in region I (Fig. [\[fig8\]](#fig8){reference-type="ref" reference="fig8"}b). In Fig. [\[fig8\]](#fig8){reference-type="ref" reference="fig8"}c, we see that increasing further the constant intensity for $\gamma=1000$pA, the external perturbation suppresses also bursting synchronisation in region II. Therefore,the synchronous behavior in region II is more robustness to perturbations than in the region I, due to the fact that the region II is in a range with high $g_{\rm ex}$ and $p$ values, namely strong coupling and high connectivity. In order to understand the perturbation effect on the spike and bursting patterns, we consider the same values of $g_{\rm ex}$ and $p$ as Fig. [\[fig7\]](#fig7){reference-type="ref" reference="fig7"}a. Figure [\[fig9\]](#fig9){reference-type="ref" reference="fig9"} exhibits the space parameter $b\times V_r$, where $\gamma$ is equal to $500$pA. The external perturbation suppresses synchronisation in the region I, whereas we observe synchronisation in region II. The synchronous behaviour in region II can be suppressed if the constant intensity $\gamma$ is increased. Therefore, bursting synchronisation is more robustness to perturbations than spike synchronisation. # Conclusion In this paper, we studied the spiking and bursting synchronous behaviour in a random neuronal network where the local dynamics of the neurons is given by the adaptive exponential integrate-and-fire (aEIF) model. The aEIF model can exhibit different firing patterns, such as adaptation, tonic spiking, initial burst, regular bursting, and irregular bursting. In our network, the neurons are randomly connected according to a probability. The larger the probability of connection, and the strength of the synaptic connection, the more likely is to find bursting synchronisation. It is possible to suppress synchronous behaviour by means of an external perturbation. However, synchronous behaviour with higher values of $g_{\rm ex}$ and $p$, which typically promotes bursting synchronisation, are more robust to perturbations, then spike synchronous behaviour appearing for smaller values of these parameters. We concluded that bursting synchronisation provides a good environment to transmit information when neurons are strongly perturbed (large $\Gamma$).	{'timestamp': '2016-07-18T02:09:37', 'yymm': '1607', 'arxiv_id': '1607.04618', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04618'}
null	null
null	null
# Introduction To write high-quality program code for a Multi-Processor System-on-Chip (MPSoC), software developers must fully understand how their code will be executed on-chip. Debugging and tracing tools can help developers to gain this understanding. They are a keyhole through which developers can peek and observe the software execution. Today, and even more in the future, this keyhole narrows as MPSoCs integrate more functionalities, while at the same time the amount of software increases dramatically. Furthermore, concurrency and deep interaction of software with hardware components beyond the instruction set architecture (ISA) boundary are on the rise. Therefore more, not less, insight into the system is needed to keep up or even increase developer productivity. Many of today's MPSoCs are executing concurrent code on multiple cores, interact with the physical environment (cyber-physical systems), or must finish execution in a bounded amount of time (hard real-time). In these scenarios, a non-intrusive observation of the software execution is required, like it is provided by tracing. Instead of stopping the system for observation, as done in run-control debugging, the observed data is transferred off-chip for analysis. Unfortunately, observing a full system execution would generate data streams in the range of petabits per second . This is the most significant drawback of tracing: the system insight is limited by the off-chip bottleneck. Today's tracing systems, like ARM CoreSight or NEXUS 5001 are designed to efficiently capture the operation of a functional unit (like a CPU) as compressed trace stream. With filters and triggers it is possible to configure which and when a functional unit is traced (observed). The trace streams (or short, traces) are then transported across an off-chip interface (and possibly other intermediate devices) to a host PC. Upon arrival the compressed trace streams are first decompressed (reconstructed) using the program binary and other static information which was removed before. The reconstructed trace streams are then fed to a data analysis application, which extracts information out of the data. This information can then be presented to a developer or it can be used by other tools, e.g. for runtime verification. The main idea in this work is to move the data analysis (at least partially) from the host PC into the chip. Bringing the computation closer to the data sources reduces the off-chip bandwidth requirements, and ultimately increases insight into software execution. To realize this idea, we introduce DiaSys, a replacement for the tracing system in an MPSoC. DiaSys does not stream full execution traces off-chip for analysis. Instead, it first creates events from observations on the chip. Events can signal any interesting state change of the observed system, like the execution of a function in the program code, a change in interconnect load beyond a threshold, or the read of a data word from a certain memory address. A diagnosis application then processes the observed events to give them "meaning." Given an appropriate diagnosis application, a software developer might not be presented with any more a sequence of events like "a read/write request was issued", but with the more meaningful output of the diagnosis application "a race condition bug was observed." Analyzing the data on-chip is not only beneficial to reduce the off-chip bandwidth requirements, but also enables new use cases in the future, such as self-adapting or self-healing systems. However, doing all this processing on-chip would, in some cases (and markets), be too costly in terms of chip area. Therefore, we describe the diagnosis applications so that they can be transparently split into multiple parts: one part executing on-chip in dedicated, distributed hardware components close to the data source, and another part running on a host PC. In summary, our key contributions are: - an architecture and component library of on-chip infrastructure to collect and analyze diagnosis data created during the software execution, and - a model of computation which allows developers to describe data analysis tasks (the "diagnosis application") in a way which is independent of the specific hardware implementation of the diagnosis system. Combining these two contributions, we show that - DiaSys is a viable alternative to tracing systems in the two major fields where tracing is employed today: hypothesis testing (debugging) and the collection of runtime statistics. Two case studies explore these use cases (Section [\[sec:usage\]](#sec:usage){reference-type="ref" reference="sec:usage"}). - the diagnosis applications introduced by DiaSys are a beneficial representation of a data analysis task: they abstract from the implementation through a clearly defined model of computation to foster re-use and portability (Section [3.4](#sec:method:diagnosis_applications){reference-type="ref" reference="sec:method:diagnosis_applications"}). - DiaSys is implementable in hardware with reasonable system cost (Section [4](#sec:hwimpl){reference-type="ref" reference="sec:hwimpl"}). In the following, we explore our diagnosis system in depth. We start with a thorough analysis of the state of the art in tracing systems in Section [2](#sec:related_work){reference-type="ref" reference="sec:related_work"}, based on which we developed our concept of the diagnosis system presented in Section [3](#sec:method){reference-type="ref" reference="sec:method"}. We include a detailed model of a diagnosis application and a discussion of its semantics in Section [3.4](#sec:method:diagnosis_applications){reference-type="ref" reference="sec:method:diagnosis_applications"}. The architecture of our diagnosis system is presented next in Section [3.5](#sec:arch){reference-type="ref" reference="sec:arch"}, followed by a discussion of its possible limitations. In Section [4](#sec:hwimpl){reference-type="ref" reference="sec:hwimpl"} we present our hardware implementation. Combining all parts, we show two usage examples in Section [\[sec:usage\]](#sec:usage){reference-type="ref" reference="sec:usage"}, one to find a multi-core race condition bug, and one to create an application profile. #### A word on terminology We use the terms "diagnosis" and "diagnosis system" to stress the fact that we integrate the on-chip observation of the software execution with the data analysis. We use "tracing" to refer to the method of obtaining insight into the SoC by transferring a stream of observations from a functional unit off-chip. "Debugging" is used as a synonym for "hypothesis testing," the process of (usually manually) checking (by various means) if the software behaves as expected. # Background and Related Work {#sec:related_work} Our approach touches and integrates two usually separated topics: obtaining a software execution trace from a SoC, and processing the obtained information in order to generate useful information. In this section we present background and related work on both topics. ## Gaining Insight into SoCs {#sec:related_work:insight} Today's tracing solutions for SoCs are designed to capture and transfer as much as possible of the SoC's internal state to an external observer. They are generally structured as shown in Figure [\[fig:diagnosis_applications_existing\]](#fig:diagnosis_applications_existing){reference-type="ref" reference="fig:diagnosis_applications_existing"}. First, trace data streams are obtained from the observation of various functional units in the system, like CPUs, buses and memories. Then, this data is spatially and temporally reduced through the use of filters and triggers. Finally, the redundancy in the data is removed by the use of compression algorithms. The resulting trace data stream is then transferred off-chip (live or delayed through an on-chip memory buffer). On a host PC, the original trace streams are reconstructed using information from the program binary and other static information, which was discarded as part of the compression process. All major commercial SoC vendors offer tracing solutions based on this template. ARM provides its licensees the CoreSight intellectual property (IP) blocks . They are used in SoCs from Texas Instruments, Samsung and STMicroelectronics, among others. Vendors such as Qualcomm (formerly Freescale) include tracing solutions based on the IEEE-ISTO 5001 (Nexus) standard , while Infineon integrates the Multi-Core Debug Solution (MCDS) into its automotive microcontrollers . Since 2015 Intel also includes a tracing solution in their desktop, server and embedded processors called Intel Processor Trace (PT) . The main differentiator between the solutions is the configurability of the filter and trigger blocks. Driven by the off-chip bottleneck, a major research focus are lossless trace compression schemes. Program trace compression available in commercial solutions typically requires 1 to 4 bit per executed instruction , while solutions proposed in academia claim compression ratios down to $0.036$ bit per instruction . Even though data traces contain in general no redundancy, in practice compression rates of about 4:1 have been achieved. ## Analyzing System Behavior {#sec:related_work:analysis} A human is easily overwhelmed when asked to analyze multiple gigabits of trace data each second. Instead, automated analysis tools are used to extract useful information out of the vast amount of trace data. Such tools have one common goal: to help a developer better understand the software execution on the target system. The means to achieve this goal, however, vary widely. Diagnosis applications which analyze non-functional issues such as performance bugs often generate results in the form of ordered lists. They list for example applications which consume most processing or memory resources, or which generate most I/O traffic. This report can then be a starting point for a more fine-grained analysis of the problem. Diagnosis applications which target functional bugs are usually more specialized; in many cases, a diagnosis application is created just to confirm or negate one single hypothesis about the software execution on the chip. For example, a developer might want to confirm that a certain variable stays within defined bounds, e.g. to check if an array overflow occurred. Most analysis tools for SoCs are not stand-alone applications, but part of debugging and tracing software packages from vendors like Lauterbach, Green Hills or ARM. They are usually controlled through a graphical user interface. Of course, analysis applications used to understand software execution are not only developed for SoCs and other embedded systems. Most tools in this domains are intrusive: they run as part of the analyzed system and obtain the required system state through instrumentation. However, the general concepts are also relevant for the diagnosis of SoCs. This is especially true for scriptable or programmable debugging, which applies the concept of event-driven programming to debugging. Whenever a defined probe point is hit, an event is triggered and an event handler executes. Common probe points are the execution of a specific part of the program (like entering a certain program function), or the access to a given memory location. The best-known current implementations of this concept are DTrace and SystemTap, which run on, or are part of, BSDs, Linux, and macOS (where DTrace is integrated into the "Apple Instruments" product). The concept, however, is much older. Dalek is built on top of the GNU Debugger (GDB) and uses a dataflow approach to combine events and generate higher-level events out of primitive events. Marceau et al. extend the dataflow approach and apply it to the debugging of Java applications . Coca , on the other hand, uses a language based on Prolog to define conditional breakpoints as a sequence of events described through predicates for debugging C programs. In a work targeting early multi-processor systems, but otherwise closely related to our approach, Lumpp et. al. present a debugging system which is based on an event/action model . A specification language is used to describe events in the system trigger which an action, and hardware units can be used to identify these events. None of the presented works directly tackle the observability problem in SoCs by moving the data analysis partially on-chip. However, they form a strong foundation of ideas, which inspired us in the design of the diagnosis system. It is presented in the following sections. # DiaSys, Our Diagnosis System {#sec:method} We have designed our diagnosis system to address the shortcomings of today's tracing systems. Based on a set of requirements, we discuss the design of the diagnosis system in depth, followed by a hardware/software architecture implementing the diagnosis system. First, however, we define some terms used in the following discussion. ## Definitions #### Observed system The part of the SoC which is observed or monitored by the diagnosis system. In other works, the term "target system" is used. #### Functional unit A subset of the observed system which forms a logical unit to provide a certain functionality. Examples for functional units are CPUs, memories, or interconnect resources such as a bus or NoC routers. #### State The state of a system is the unity of all stored information in that system at a given point in time which is necessary to explain its future behavior. In a sequential circuit, the state is equal to the memory contents of the system. ## Design Requirements for the Diagnosis System {#sec:method:requirements} A set of requirements guides the design of the diagnosis system. #### Distributed The diagnosis system must be able to reduce the amount of observation data as close to the source, i.e. the functional units, as possible. Since the data sources are distributed across the chip, the diagnosis system must also be distributed appropriately. #### Non-Intrusive The diagnosis system must be non-intrusive (passive). Non-intrusive observation preserves the event ordering and temporal relationships in concurrent executions, a requirement for debugging multi-core, real-time, or cyber-physical systems . Non-intrusiveness also gives a developer the confidence that he or she is observing a bug in the program code, not chasing a problem caused by the observation (a phenomenon often called "Heisenbug" ). #### Flexible On-Chip/Off-Chip Cost Split The diagnosis system must be flexible to implement. The implementation of the diagnosis system involves a trade-off between the provided level of observability and the system cost. The two main cost contributions are the off-chip interface and the chip area spent on diagnosis extensions. The diagnosis system concept must be flexible enough to give the chip designer the freedom to configure the amount of chip resources, the off-chip bandwidth and the pin count in a way that fits the chip's target market. At the same time, to provide flexibility on the observation, the system must be able to adapt to a wide range of bugs. #### Relaxed Timing Constraints The diagnosis system must not assume a defined timing relationship between the individual distributed components. Today's larger SoCs are designed as globally asynchronous, locally synchronous (GALS) systems with different power and clock domains, where no fixed time relationship between components can be given. ## The Concept of the Diagnosis System {#sec:method:concept} Based on the discussed requirements this section gives an overview on the diagnosis system as shown in Figure [\[fig:diasys_model\]](#fig:diasys_model){reference-type="ref" reference="fig:diasys_model"}. The input to the diagnosis system is the state of the observed system over time, the output are the diagnosis results, which can be represented in various forms. With respect to the input and output interfaces, the diagnosis system is identical to a traditional tracing system. The difference lies in the components which generate the output from the input. Three main components are responsible for this processing: the event generators, the diagnosis application together with its execution platform, and the event sinks. Between these components, data is exchanged as diagnosis events. Diagnosis events are the container for data exchanged in the diagnosis system. In the general case, an event consists of a type identifier and a payload. Events are self-contained, i.e. they can be decoded without the help from previous or subsequent events. Event generators produce primary events based on state changes in the observed system. They continuously compare the state of the observed system with a trigger condition. If this condition holds, they trigger the generation of a primary event. A primary event is a specialized diagnosis event in which the type identifier is equal to a unique identifier describing the event generator. The payload contains a partial snapshot of the state of the observed system at the same instant in time as the event was triggered. Which parts of the state are attached to the event is specified by the event generator. For example, a CPU event generator might produce primary events when it observes a function call and attach the current value of a CPU register as payload. A primary event answers two questions: why was the event generated, and in which state was the observed system at this moment in time. The diagnosis application analyzes the software execution on the observed system. It is modeled as transformational dataflow application, which transforms primary events into output events. The goal of this transformation is to interpret the state changes represented in primary events in a way that yields useful information for a developer or an automated tool. We describe diagnosis applications in more detail in Section [3.4](#sec:method:diagnosis_applications){reference-type="ref" reference="sec:method:diagnosis_applications"}. The diagnosis application execution platform executes diagnosis applications. The execution platform can span (transparent to the diagnosis application developer) across the chip boundary. On the chip, it consists of specialized hardware blocks which are able to execute (parts of) the diagnosis application. On the host PC, a software runtime environment enables execution of the remaining parts of the diagnosis application. The on-and off-chip part of the execution platform are connected through the off-chip interface. This split design of the execution platform allows hardware designers to trade off chip area with the bandwidth provided for the off-chip interface, while retaining the same level of processing power, and in consequence, system observability. Event sinks consume output events produced by the diagnosis application. Their purpose is to present the data either to a human user in a suitable form (e.g. as a simple log of events, or as visualization), or to format the events in a way that makes them suitable for consumption by an automated tool, or possibly even for usage by an on-chip component. An example usage scenario for an automated off-chip user is runtime validation, in which data collected during the runtime of the program is used to verify properties of the software. Together, event generators, the diagnosis application and the event sink build a processing chain which provides a powerful way to distill information out of observations in the SoC. ## Diagnosis Applications {#sec:method:diagnosis_applications} Diagnosis applications are the heart of the diagnosis system, as they perform the "actual work" of interpreting what happens on the observed system during the software execution. Diagnosis applications are transformational dataflow applications. We chose this model to enable the transparent mapping of the diagnosis application to an execution platform spanning across the chip boundary. Our goal is that the developer of the diagnosis application does not need to explicitly partition the diagnosis application into an on-chip and an off-chip part; instead, this mapping could be performed in an automated way. (Currently, however, we do not perform an automated mapping.) No matter how the diagnosis application is mapped onto the execution platform, the behavior of the application follows identical rules, i.e. the semantics of the application stay the same. The diagnosis application is a transformational application, in contrast to reactive or interactive applications . This means, starting from a given set of inputs, the application eventually produces an output. The application code only describes the functional relationship between the input and the output, not the timing when the output is generated. The application also does not influence or interact in another way with the observed system from which its inputs are derived. The diagnosis application is structured as dataflow application. Its computation is represented by a directed graph, in which the nodes model the computation, and the edges model communication links. In diagnosis applications we call the graph nodes transformation actors, and the graph edges channels. Each transformation actor reads events from $n \in \mathbb{N}_0$ input channels, and writes events to $m \in \mathbb{N}_0$ output channels, as shown in Figure [\[fig:transformation_actor\]](#fig:transformation_actor){reference-type="ref" reference="fig:transformation_actor"}. A transformation actor starts its processing, it "fires," if a sufficient number of events are available at its inputs. The definition of "sufficient" depends on the individual transformation actor. For example, one transformation actor might always read one event from each input before starting the processing, while another one might always read two events from input 1 and one event from input 2. When firing, the transformation actor applies an arbitrary transformation function $f$ to the input events. The generated output depends on - the read input events, - the ordering of the input events, and - the internal state of the transformation actor. Transformation actors may communicate only through the input and output channels, but not through additional side channels (e.g. shared variables). Diagnosis applications built out of such transformation actors are nondeterministic, as defined by Kahn . This means, the output not only depends on the history of inputs (i.e. the current input and the state of the actor), but also on the relative timing (the ordering) of events. Nondeterministic behavior of diagnosis applications is, in most cases, the expected and wanted behavior; it gives its authors much needed flexibility. An example of nondeterministic diagnosis applications are applications which aggregate data over time, like the lock contention profiling presented in Section [\[sec:usage:lockprofiling\]](#sec:usage:lockprofiling){reference-type="ref" reference="sec:usage:lockprofiling"}. These applications consume an unspecified amount of input events and store an aggregate of these inputs. After a certain amount of time, they send a summary of the observations to an output channel. But at the same time, nondeterministic diagnosis applications prevent the static analysis of event rates, bandwidth and processing requirements. If wanted, application authors can therefore create deterministic diagnosis applications, if they restrict themselves to - always reading the input channels in the same order without testing for data availability first (instead, block and wait until the data arrives), - connecting one channel to exactly one input and one output of an actor, and - using only transformation functions which are deterministic themselves. Note that we only describe the diagnosis application itself as nondeterministic. Its execution, after being mapped to an execution platform, can be deterministic, i.e. multiple identical runs produce the same diagnosis result. ## Diagnosis System Architecture {#sec:arch} In the previous sections we presented the diagnosis system from a functional perspective. We continue now by presenting an implementation architecture of the diagnosis system. It consists of extensions to the SoC, as well as software on a host PC; an exemplary architecture is shown in Figure [\[fig:diasys_hwarch\]](#fig:diasys_hwarch){reference-type="ref" reference="fig:diasys_hwarch"}. In the SoC, different functional units (FU) can be observed, like a CPU, a memory, or a bus. Each functional unit can be attached to one or multiple event generators (EG). The resulting events are transmitted over a diagnosis interconnect to on-chip processing nodes. The processing nodes form the execution platform for the diagnosis application; they will be discussed in depth in Section [3.6](#sec:arch:execution_platform){reference-type="ref" reference="sec:arch:execution_platform"}. Invisible to the diagnosis application, the execution platform spans across the chip boundary. Through an I/O interface, a host PC is connected to the SoC. The PC contains a software runtime environment to provide further processing nodes as part of the diagnosis application execution platform. If all processing has been accomplished, the output events are sent to an event sink application on the host PC, which formats the output events for developers or automation. Depending on the features and computational power provided by the processing nodes, a diagnosis application can be mapped to the execution platform in a flexible way. ## Diagnosis Application Execution Platform {#sec:arch:execution_platform} The heart of the diagnosis system are the diagnosis applications, which are executed on the diagnosis application execution platform. As discussed before, this platform spans across the SoC and the host PC. On the SoC, it consists of processing nodes of different types which are connected by a shared interconnect (such as a NoC or a bus). Each processing node has an input and output interface to receive and send out events on the interconnect. Different types of processing nodes can offer different degrees of flexibility regarding their computation. Some might be able to perform only a single functionality specified at hardware design time, like a counter or a statistical aggregator, while others might be freely programmable. As an example, we present in Section [3.6.1](#sec:arch:diagnosis_processor){reference-type="ref" reference="sec:arch:diagnosis_processor"} the Diagnosis Processor, a programmable general-purpose processing node. As the chip area (economically) available for on-chip diagnosis processing is limited, the diagnosis application execution platform extends to the host PC. Connected through an arbitrary off-chip interface, a runtime layer in software provides a virtually unlimited number of "soft" processing nodes. Such PNs are implemented in software on the host PC and accept, like their on-chip counterpart, events as input and produce events as output. By being executed on a host PC, they provide more compute and memory resources. The transformation or computation in a diagnosis application is represented by transformation actors. For execution, they are mapped to the available processing nodes, as shown exemplary in Figure [\[fig:diasys_hwarch\]](#fig:diasys_hwarch){reference-type="ref" reference="fig:diasys_hwarch"}. An $n$:1 mapping of transformation actors to processing nodes is possible, if the combined transformation of all $n$ transformation actors can be executed by the processing node. To achieve the greatest possible reduction in off-chip traffic, as much computation as possible should be mapped to on-chip processing nodes. The remainder of processing is then mapped to processing nodes on a host PC, where significantly more processing power is available. ### The Diagnosis Processor: A Multi-Purpose Processing Node {#sec:arch:diagnosis_processor} The diagnosis processor is a freely programmable general-purpose processing node. Like any processor design, it sacrifices computational density for flexibility. Its design is inspired by existing scriptable debugging solutions, like SystemTap or DTrace, which have shown to provide a very useful tool for software developers in a growingly complex execution environment. The usage scenario for this processing node are custom or one-off data analysis tasks. This scenario is very common when searching for a bug in software. First, a hypothesis is formed by the developer why a problem might have occurred. Then, this hypothesis must be validated in the running system. For this validation, a custom data analysis script must be written, which is highly specific to the problem (or the system state is manually inspected). This process is repeated multiple times, until the root cause of the problem is found. As this process is approached differently by every developer (and often also influenced by experience and luck), a very flexible processing node is required. We present the hardware design of our diagnosis processor implementation in Section [\[sec:hwimpl:diagnosis_processor\]](#sec:hwimpl:diagnosis_processor){reference-type="ref" reference="sec:hwimpl:diagnosis_processor"}. We envision the programming of the diagnosis processor being done through scripts similar to the ones used by SystemTap or DTrace. They allow to write trace analysis tasks on a similar level of abstraction as the analyzed software itself, leading to good developer productivity. ## Discussion The presented diagnosis system is designed to fulfill the requirements outlined in Section [3.2](#sec:method:requirements){reference-type="ref" reference="sec:method:requirements"}. In the following, we discuss the consequences of the design decisions, which can limit the applicability of the diagnosis system approach in some cases. By transforming the observed system state close to the source into denser information, the off-chip bottleneck can be circumvented. As a downside, this lossy transformation thwarts a usage scenario of today's tracing systems. In many of these systems, it is possible to capture a trace once, store it, and run different analysis tasks on it. If major parts of the captured data are dismissed early, this is not possible any more. Instead, the analysis task must be defined (as diagnosis application) before the system is run. If the problem hypothesis changes and a different diagnosis application is required, the system must be run again. The severity of this limitation strongly depends on how hard it is to reproduce a bug or behavior across runs. Another feature present in many of today's tracing systems, which is explicitly not supported by the diagnosis system, are cross-triggers. Cross-triggers are a mechanism in the tracing system to start or stop the observation, or to observe different components, based on another observation in the system. For example, memory accesses could be traced only after a CPU executed a certain program counter. Cross-triggers are most useful if their timing behavior is predictable. For example, memory accesses are traced "in the next cycle" after the specified program counter was executed. In GALS SoCs, such timing guarantees cannot be given; for a diagnosis application spanning across a SoC and a host PC, it is equally impossible to give (reasonably low bounded) timing guarantees. We make this property explicit by modeling the diagnosis system as a transformational system, not a reactive system. The commercially available tracing systems today are less specific about this. For example, ARM CoreSight uses a handshaking protocol for cross-triggers delivered across clock boundaries, which guarantees save delivery of the signal, but does not guarantee any latency. Instead of relying on cross-triggers to collect data from different sources at the same instant in time, we capture this data already when creating primary events through event generators. The payload of primary events is the only way to pass multiple state observations with a defined timing relation to the diagnosis system. For example, an event generator attached to a CPU can trigger an event based on a program counter value, and attach current contents of certain CPU registers or stack contents to it. Using this method, it is possible to generate for example an event which informs about a function being called, and which function arguments (stored in CPU registers or on the stack) have been passed to it. We show an example of such an event generator as part of our hardware implementation. Finally, we discuss the system behavior in overload situations, i.e. if more input data is received than the diagnosis system can process. Given the generally unknown input data, and the generally nondeterministic behavior of the diagnosis application, it is not possible to statically dimension the diagnosis system to be able to handle all possible input sequences. Therefore, overload situations are unavoidable in the general (and most common) case. If an overload situation is detected, the diagnosis system can react in multiple ways. First, it could temporarily stall the observed system. This gives the diagnosis system time to process outstanding events without new events being produced. This approach is only feasible in a synchronous non-realtime system. A more common approach is to discard incoming data until further processing resources are available. Depending on the diagnosis application, a recovery strategy needs to be formulated. Some applications can deal easily with incomplete input data, e.g. diagnosis applications creating statistics. Others are not able to work with an incomplete input sequence and in consequence fail to be executed properly. This ends the discussion of DiaSys in general. In the following, we present a hardware implementation of our approach, and then continue then with two usage examples how DiaSys can be put to work. # Implementing DiaSys in Hardware {#sec:hwimpl} DiaSys, as presented in the previous section, can be implemented in various ways in hardware. Our implementation, which we present in the following, is one such implementation. It was created to answer two questions: first, to show that DiaSys can be implemented in hardware, and second, to give resource usage numbers for one specific implementation. As dimensioning and optimization for speed or area usage strongly depends on how DiaSys is used, a general answer to this question must remain out of scope for this work. The diagnosis system extends a $2 \times 2$ tiled multi-core system as shown in Figure [\[fig:prototype_sys\]](#fig:prototype_sys){reference-type="ref" reference="fig:prototype_sys"}. Our implementation runs on an FPGA and uses the OpTiMSoC framework . The observed system consists of four mor1kx CPU cores (an implementation of the OR1K or "OpenRISC" ISA), each connected to a distributed memory and a mesh NoC interconnect (components with white background). This system is representative of the multi-and many-core architecture template currently in research and available early products, such as the Intel SCC or the Mellanox (formerly Tilera and EZchip) Tile processors. The diagnosis system, depicted in blue, consists of the following components. - Four event generators attached to the CPUs (marked "EG"). - A single diagnosis processor. - A 16 bit wide, unidirectional ring NoC, the "diagnosis NoC," to connect the components of the diagnosis system. It carries both the event packets as well as the configuration and control information for the event generators and processing nodes. - A USB 2.0 off-chip interface. - Software support on the host PC to control the diagnosis system, and to display the results. All components connected to the diagnosis NoC follow a common template to increase reusability. Common parts are the NoC interface and a configuration module, which exposes readable and writable configuration registers over the NoC. In the following, we explain the implementation of the main components in detail. ## CPU Event Generator {#sec:hwimpl:implementation:eventgen_cpu} The CPU event generator is attached to a single CPU core. Its main functionality is implemented in the trigger and the state capture modules. The trigger unit of the CPU event generator triggers on two types of conditions: either the value of the program counter (PC), or the return from a function call (the jump back to the caller). At each point in time, 12 independent trigger conditions can be monitored. The number of monitored trigger conditions is proportional to the used hardware resources. Our dimensioning was determined by statistical analysis of large collections of SystemTap and DTrace scripts: $\leq 9$ concurrent probes are used in 95 % of SystemTap scripts, and $\leq 12$ concurrent probes cover 92 % of the DTrace scripts. The partial system state snapshot can be configured to contain the CPU register contents and the function arguments passed to the function. A block diagram of the CPU event generator is shown in Figure [\[fig:eventgen_cpu\]](#fig:eventgen_cpu){reference-type="ref" reference="fig:eventgen_cpu"}. The PC trigger is implemented as simple comparator. The "function return" trigger requires a special implementation, because no unique point in the program flow, i.e. no single PC value, describes the return from a function (a function can have multiple call sites and can return to the caller from different points in the function body). Instead, we use the following method: 1. A PC trigger is set to the first instruction of the called function. 2. If the trigger fires, the link (a.k.a. return) address is pushed to a memory structure inside the return monitor, the "return address stack." The link address is the program counter to jump to if the function has finished its execution and the execution returns to the caller. On OR1K (as common on RISC architectures, including ARM and MIPS) the link address is stored in a CPU register. On other architectures and calling conventions (such as x86 and x86_64), the link address is pushed to the stack. 3. Now the system monitors the program flow for the topmost PC value in the return address stack. If this PC is executed, a function returned to its caller and the function return monitor triggers the generation of a primary event. To capture the values inside CPU registers, the register writeback signal of the mor1kx CPU is observed, and a copy of the register file is created. This copy can then be included in the event packet if a trigger fires. The passing of function arguments to functions depends on the calling convention. On OR1K, the first six data words passed to a function are available in CPU registers, all other arguments are pushed to the stack before calling the function. This is common for RISC architectures; other architectures and calling conventions might pass almost all arguments on the stack (such as x86). To record the function arguments as part of the primary event we therefore need to create a copy of the stack memory that can be accessed non-intrusively. We do this by observing CPU writes to the stack memory location. In our implementation for the mor1kx CPU we create a copy of the stack memory by monitoring the instruction stream for a store word (`l.sw`) instruction with a target address `rA` equal to the stack pointer `R1`. The data in the source register `(rB)`, together with a write offset `I` (with $\texttt{I} \geq 0$, i.e. targeting the previous stack frame) can be then used to recreate the stack frame.	{'timestamp': '2017-01-19T02:06:50', 'yymm': '1607', 'arxiv_id': '1607.04549', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04549'}
# Introduction {#sec:intro} The Type Ia supernova (SN Ia) SN 2011fe was discovered on 2011 August 24, just 11 hr after explosion. It is among the nearest ($\sim 6.9$ Mpc) and youngest ($\sim 11$ hr) SNe Ia ever discovered. Extensive spectroscopic and photometric studies of SN 2011fe indicate that it is "normal" in nearly every sense: in luminosity, spectral and color evolution, abundance patterns, etc.. Its unremarkable nature coupled with the wealth of observations made over its lifetime render it an ideal laboratory for understanding the physical processes which govern the evolution of normal SNe Ia. Indeed, these data have allowed observers to place numerous and unprecedented constraints on the progenitor system of a particular SN Ia. Equally as information-rich as observations taken at early times are those taken much later, when the supernova's photosphere has receded and spectrum formation occurs deep in the SN core. For example, used late-time spectra to further constrain the progenitor system of SN 2011fe, namely that the amount of hydrogen stripped from the putative companion must be $< 0.001~M_\odot$. found that the luminosity from SN 2011fe in the 3.6 $\mu$m channel of Spitzer/IRAC fades almost twice as quickly as in the 4.5 $\mu$m channel, which they argue is a consequence of recombination from doubly ionized to singly ionized iron peak elements. In addition, used photometric observations near 930 d post-maximum light to construct a late-time quasi-bolometric light curve, and showed that the luminosity continues to trace the radioactive decay rate of $^{56}$Co quite closely, suggesting that positrons are fully trapped in the ejecta, disfavoring a radially combed or absent magnetic field in this SN. presented an optical spectrum at 981 d post-explosion and used constraints on both the mass of hydrogen as well as the luminosity of the putative secondary star as evidence against a single-degenerate explosion mechanism. presented an optical spectrum at 1034 d post-explosion, and speculated about the presence of \[\] lines near 6300 Å, which, if confirmed, would provide strong constraints on the mass of unburned material near the center of the white dwarf progenitor of SN 2011fe. Non-detections of the H$\alpha$ line at both of these very late epochs also strengthened the constraints on the presence of hydrogen initially posed by. Finally, used spectrum synthesis models of SN 2011fe from 192 to 364 days post-explosion to argue for a large central mass of stable iron and a small mass of stable nickel--about 0.23 $M_\odot$ and 0.01 $M_\odot$, respectively. We complement these various late-time analyses with a series of radiative transfer models corresponding to a series of optical and ultraviolet (UV) spectra of SN 2011fe. # Observations {#sec:obs} We obtained optical spectra of SN 2011fe at days +100, +205, +311, +349, and +594 (Dec 19, 2011, Apr 2, 2012, Jul 17, 2012, Aug 23, 2012, Mar 27, 2013); the observations are shown in Figure [\[fig:all_optical_spectra_11fe\]](#fig:all_optical_spectra_11fe){reference-type="ref" reference="fig:all_optical_spectra_11fe"} and described in Table [1](#tab:obs){reference-type="ref" reference="tab:obs"}. The day +205 and +311 spectra were presented in. We also obtained an ultraviolet spectrum with Hubble Space Telescope at day +360 (GO 12948; PI: R. Thomas). This latter observation consisted of ten orbits, the first nine using the STIS/NUV-MAMA configuration, and the last with the STIS/CCD G430L and G750L configurations. The data from one of the NUV-MAMA orbits was unrecoverable, and so the final spectrum, shown in Figure [\[fig:sn11fe_combined_hst_spectra_ft_smoothed_no_model\]](#fig:sn11fe_combined_hst_spectra_ft_smoothed_no_model){reference-type="ref" reference="fig:sn11fe_combined_hst_spectra_ft_smoothed_no_model"}, represents co-addition of the nine remaining observations. Also shown in Figure [\[fig:sn11fe_combined_hst_spectra_ft_smoothed_no_model\]](#fig:sn11fe_combined_hst_spectra_ft_smoothed_no_model){reference-type="ref" reference="fig:sn11fe_combined_hst_spectra_ft_smoothed_no_model"} is a smoothed version of the HST spectrum. We used the algorithm presented in, which consists of applying a low-pass filter to the signal. The motivation for this approach is the notion that the physical features in SN Ia spectra are broad, while most noise in the spectrum is narrow. Therefore, if one can suppress the high-"frequency" features (the noise), what will remain will be the pure signal from the SN. To accomplish this task, one calculates the power spectrum of the original spectrum using a Fourier transformation, suppresses the power spectrum at all high "frequencies" in which information is deemed to be noise, and applies an inverse Fourier transformation to recover the smoothed spectrum. An especially useful feature of this technique is its insensitivity to spikes in HST spectra due to cosmic rays. # Radiative transfer models We used the `PHOENIX/1D` code with the same modifications discussed in to capture the most important physical processes at late times in SNe Ia. The underlying explosion model was a spherically symmetric delayed-detonation model presented in. For each observation presented in §[2](#sec:obs){reference-type="ref" reference="sec:obs"} we calculated a corresponding synthetic spectrum, assuming a 16 day rise time for the model. # Discussion {#sec:discussion} The theory of spectrum formation at late times in normal SNe Ia has broadly converged to a scenario in which electron configurations of atoms in the rarefied ejecta are primarily in their ground state, and are excited by collisions with free electrons to low-lying metastable levels, which in turn emit forbidden lines as they return to the ground state. Little to no continuum emission is present in SN Ia spectra at these epochs. This stands in contrast to the spectrum formation mechanism at early times, near maximum light, in which the optical depth to Thomson scattering on free electrons is large, leading to the formation of a photosphere on top of which atoms undergo line scattering via strong permitted lines, giving rise to P Cygni spectral features. argued that the P Cygni mechanism is no longer active at late times because the photosphere has disappeared (indicated by the absence of continuum) and there are no longer enough photons for these strong lines to scatter. Curiously, good spectral fits have been obtained for relatively late SN Ia spectra with the parameterized spectrum synthesis code `SYNOW`, which treats only line scattering by permitted lines. Such fits require only a few ions--, , and --and fit optical spectra fairly well, especially blueward of $\sim 6000$ Å. Examples include the normal SN 1994D at day +115, the normal SN 2003du at day +84, the subluminous SN 1991bg at day +91, and the peculiar SN 2002cx at day +227. While the parameterized approach of `SYNOW` to solving the radiative transfer equation restricts analysis of those fits to putative line identifications and velocity measurements, they nevertheless demonstrate that either SN Ia spectra exhibit a remarkable degeneracy with respect to forbidden and permitted line formation, or that permitted lines continue to drive emergent spectrum formation at late times, regardless of what physical mechanisms generate the underlying flux. These two competing analyses of late-time SN Ia spectra agree that the majority of the spectrum is formed by Fe lines, but they predict dramatically different velocities of the line-forming regions in the ejecta. For example, argue that Fe, Ca, and Na are located at 7000 km s$^{-1}$ and higher in the day +84 spectrum of SN 2003du. In contrast, argue for velocities from 1000 -- 3000 km s$^{-1}$ in the +95 spectrum of the same object. Identifying the correct velocity of the line-forming region has important consequences for constraining the structure of the inner regions of SN Ia ejecta, which in turn constrain the as-yet unknown explosion mechanism. ## Day +100 {#subsec:11fe_p100} The day +100 spectrum of SN 2011fe and the corresponding synthetic spectrum from `PHOENIX` are shown in Figure [\[fig:pah_std_d116_delta_t\_no_forb_lines_vs_11fe_p100\]](#fig:pah_std_d116_delta_t_no_forb_lines_vs_11fe_p100){reference-type="ref" reference="fig:pah_std_d116_delta_t_no_forb_lines_vs_11fe_p100"}. Overall the fit is good, although a few features in our model do not match those in the observed spectrum, namely the emission feature near 5900 Å. In addition, the peak at 4700 Å in the synthetic spectrum is too weak, the blue side of the broad emission at 7200 Å is absent in the model, and the flux in the blue and near-UV is too high. Most other features are well reproduced, both in strength and in shape. In the context of much of the literature which concerns late-time SN Ia spectra, the fidelity of this fit is peculiar because the calculation used the most current atomic database of, which includes no forbidden line data for any ions. This stands in contrast to the most common interpretations of spectra of "old" SN Ia, which were discussed earlier. If the purely permitted line identifications are correct they are difficult to reconcile with kinematic analyses such as that of, which assume that the emission peaks correspond to forbidden lines forming within a few 100 km s$^{-1}$ of $v \simeq 0$ km s$^{-1}$. Rather, the strong emission peaks at $4700$ Å and $5200$ Å, which each have previously been identified as a mixture of \[\] and \[\] lines, may instead be formed by the handful of permitted lines of whose upper levels are among the crowded $3d^6(^5D)4p$ configuration, with energies between 5 eV and 6 eV, and whose lower levels are, coincidentally, the handful of metastable levels around 3 eV which are purportedly responsible for the aforementioned forbidden emission features. However, it is important to note that the analysis of only requires that the ejecta is optically thin at the rest wavelength of the line and as we show below in the redder parts of the spectra that condition is met at later epochs. In the line-scattering interpretation of spectrum formation, this would imply that the dips just to the blue of these two strong emission features are the corresponding absorptions, rather than regions lacking in emission. These absorptions would correspond to line velocities of $\sim 6000$ km s$^{-1}$, similar to that found in the +115 spectrum of SN 1994D by. Although below we illustrate some complications with this permitted-line-only model, it is instructive first to entertain the idea that this is, in fact, representative of late-time spectrum formation physics in SNe Ia. Given the contrast of this result with those found elsewhere in the literature, it is important to evaluate the late-time line scattering scenario within the context of other analyses of SN 2011fe, in order to determine whether or not it is copacetic with what is already known about the spectral evolution of this object. We consider three such pieces of analysis. First, traced the velocity evolution of in the early spectra of SN 2011fe using the automated spectrum code `SYNAPPS` and found that at day +15 the minimum velocity of that ion was $\sim 8000$ km s$^{-1}$ (see their Figure 3). Furthermore, after maximum light, the rate of change of line velocities in SN 2011fe, and in most SNe Ia in general, slows dramatically. Second, in the hydrodynamical model used in our calculation, Fe remains the most abundant species in the ejecta from the center of the ejecta out to $\sim 12 000$ km s$^{-1}$; our putative line velocity estimate of $\sim 6000$ km s$^{-1}$ falls well within this range. Finally, show that the optical depth of computed in local thermodynamic equilibrium (LTE) peaks between 5000 K and $10 000$ K, roughly the same temperature range as that of the ejecta in our models. (One would be remiss to read too much into this corroboration, as the radiation field in the SN Ia ejecta at this epoch is far from LTE.) Although none of these results offer conclusive evidence that the strong features in the +100 spectrum of SN 2011fe are indeed P Cygni profiles, they do show that it is quite reasonable to consider that possibility. We caution that it is unlikely that the entire optical spectrum consists of overlapping P Cygni line profiles due to resonance-scattering, as is the case at very early (photospheric) epochs in SNe Ia. attempted to fit the day +84 spectrum of SN 2003du with the resonance-scattering code `SYNOW`, and found that P Cygni lines fit the blue part of the spectrum (blueward of 6600 Å) quite well, but failed quite severely redward of that. As they discuss, the likely explanation is that resonance-scattering near this epoch is very influential at blue wavelengths, but forbidden emission is prominent in the red. (We find a similar result in our attempts to fit the optical and UV spectra at +349 and +360, respectively, which we discuss below.) In fact, to argue that spectrum formation consists of either resonance-scattering by optically thick permitted lines or emission from optically thin forbidden lines is somewhat of a false dilemma, as both scenarios assume a degree of locality in the radiative transfer which is probably unphysical. In particular, the former assumes that the source function $S$ depends only on the local mean intensity $J$, while the latter assumes that emitting lines are well separated in wavelength such that they act independently of each other. Each of these approximations is valid in certain regimes, i.e., resonance-scattering at photospheric epochs and forbidden emission at very late times ($> 1$ yr) and in wavelength regions far from the forest of iron-peak lines, such as the infrared, but there exists a wide range between those extremes, in which all of these effects compete to form the emergent spectrum. addressed this topic in detail by calculating a grid of `PHOENIX` spectra using the hydrodynamical model W7 at 20 days post-explosion. They found that even at very low velocities and high optical depth ($\tau > 3$), the "spectrum"[^1] at those velocities is already highly distorted from that of a blackbody, due to line and continuum interactions of the radiation field with iron-peak elements deep in the core of the SN. The ions found at higher velocities, near the photosphere, then further distort this underlying spectrum through additional absorption, emission, and line scattering, leading to an emergent spectrum containing a complicated mixture of P Cygni, continuum, and thermal components which are difficult to disentangle. That our spectral model which explicitly omits forbidden line data fits the day +100 spectrum of SN 2011fe reasonably well, suggests that day +100 is simply too early for collisionally excited forbidden emission to be the primary driver of spectrum formation. It appears that line scattering processes continue to contribute significantly, even this late in the lifetime of this SN. In short, there are many physically-motivated reasons to suspect that permitted lines play an important role in SN Ia spectrum formation at this epoch. However, since forbidden lines are frequently identified in spectra of SNe Ia of this age, we tested this theory by expanding our atomic database to include collisional and radiative data of forbidden lines, as described in. We then repeated the radiative transfer calculation with this expanded database, and compare the two results in Figure [\[fig:pah_std_d116_delta_t\_forb_lines_vs_no_forb_lines_vs_11fe_p100\]](#fig:pah_std_d116_delta_t_forb_lines_vs_no_forb_lines_vs_11fe_p100){reference-type="ref" reference="fig:pah_std_d116_delta_t_forb_lines_vs_no_forb_lines_vs_11fe_p100"}. The results are quite similar, except that the model with forbidden lines has a lower UV flux and stronger emission at 7300 Å and 8600 Å. The most notable shortcoming of both synthetic spectra is the lack of emission at 5900 Å. This feature has been identified alternatively as D or \[\] $\lambda 5888$ ÅṪhe explosion model used in these calculations contains little , so it is not surprising that we do not recover a strong Na I D emission feature there. However, at day 116 the model contains several 0.1 $M_\odot$ of $^{56}$Co, and yet the forbidden emission at 5900 Å does not appear. This discrepancy may be related to underestimating the gas temperature in the model at this epoch (see §[4.2](#subsec:11fe_p205){reference-type="ref" reference="subsec:11fe_p205"}). Identifying whether a feature is an "emission" or "absorption" is not a straightforward task in `PHOENIX` calculations. This is because the algorithm calculates emissivities and opacities of NLTE species by adding up all contributions to each at each wavelength point, with no regard to the underlying atomic processes which produced them. Such an approach captures naturally the notion that spectrum formation is inherently multi-layered in supernovae: one region deep in the ejecta may be strongly in emission at one wavelength, but a region above it may be optically thick at that wavelength, absorbing much of the underlying emission. The emergent spectrum is then a convolution of both processes, and such classifications as "absorption" or "emission," while relevant to single interactions, no longer describe adequately the complete process of spectrum formation. We are therefore relegated to using more indirect methods for isolating the sources of features in synthetic spectra. The single-ion spectrum method can help one identify the particular ion or ions which influence particular parts of a synthetic spectrum, but it cannot, e.g., isolate the effects of permitted lines from forbidden lines, which is desirable in this context. We found that the only useful way to accomplish this was to remove the forbidden lines from the calculation entirely and re-compute the entire model. This can unfortunately broaden the parameter space of the model, since forbidden lines affect the temperature structure by acting as coolants. Unfortunately we are aware of no more targeted method of accomplishing this goal. We computed single-ion spectra for both synthetic spectra (with and without forbidden lines), for the most influential ions. For the spectrum without forbidden lines, these are shown in Figure [\[fig:pah_std_d116_delta_t\_no_forb_lines_vs_11fe_p100_single_ion_spectra\]](#fig:pah_std_d116_delta_t_no_forb_lines_vs_11fe_p100_single_ion_spectra){reference-type="ref" reference="fig:pah_std_d116_delta_t_no_forb_lines_vs_11fe_p100_single_ion_spectra"}. For the spectrum with forbidden lines, these are shown in Figure [\[fig:pah_std_d116_delta_t\_forb_coll_vs_11fe_p100_single_ion_spectra\]](#fig:pah_std_d116_delta_t_forb_coll_vs_11fe_p100_single_ion_spectra){reference-type="ref" reference="fig:pah_std_d116_delta_t_forb_coll_vs_11fe_p100_single_ion_spectra"}. Both with and without forbidden lines, the synthetic spectra indicate that most of the optical spectrum at day +100 is formed by . In addition, the H & K doublet at $\lambda \lambda 3934, 3968$ Å, a pair of strong resonance lines, contributes significantly to the emission at 4000 Å. However, comparison between the two sets of single-ion spectra indicate a fascinating result which found highly improbable: it appears that entirely different combinations of atomic lines can conspire to produce similar optical spectra. Furthermore, the synthetic spectra in Figure [\[fig:pah_std_d116_delta_t\_no_forb_lines_vs_11fe_p100\]](#fig:pah_std_d116_delta_t_no_forb_lines_vs_11fe_p100){reference-type="ref" reference="fig:pah_std_d116_delta_t_no_forb_lines_vs_11fe_p100"} and Figure [\[fig:pah_std_d116_delta_t\_forb_lines_vs_no_forb_lines_vs_11fe_p100\]](#fig:pah_std_d116_delta_t_forb_lines_vs_no_forb_lines_vs_11fe_p100){reference-type="ref" reference="fig:pah_std_d116_delta_t_forb_lines_vs_no_forb_lines_vs_11fe_p100"} are the natural endpoints of calculations subject to otherwise identical parameters. For example, when forbidden lines are included, the emission at 4000 Å is due entirely to the H & K doublet; when they are absent, it is a combination of that same doublet with contributions also from and . The double-horned emissions at 7250 Å and 7500 Å in the synthetic spectra lacking forbidden lines are due to emission from ; but the double-horned features at 7150 Å and 7400 Å in the spectra containing forbidden lines are due to (possibly \[\] $\lambda \lambda 7155, 7171$ Å)and \[\] $\lambda 7374, 7412$ Å. This is likely stable $^{58}$Ni, since the radioactive $^{56}$Ni has mostly decayed by this point. The degeneracy among these various features is the likely explanation for the conflicting results of, e.g., and. One would expect that adding forbidden lines is always favorable over neglecting them: if a calculation captures all relevant atomic processes and if forbidden lines are truly unimportant in some scenario, they will naturally "deactivate". And in fact, adding in the forbidden atomic data did address some problems in the synthetic spectra which lacked them. The lower UV flux and the 8600 Å emission in the synthetic spectrum forbidden lines can be explained by their cooling effects: lower temperatures generally lead to lower opacities in the UV, and a lower temperature allows more to recombine to . However, the cooling effects also introduced a new problem: the infrared triplet (IR3) of $\lambda \lambda 8498, 8542, 8662$ Å (a trio of strong permitted lines) is responsible for the emission at 8600 Å, but the model overestimates the strength of the emission at this wavelength. In the observation there is a pair of weaker emission features at the same location, and it is possible that at least one of these two emissions is due to the IR3, although probably not both, since they are spread too far apart in wavelength. It is therefore not entirely clear which of the two synthetic spectra are "better," and it is possible therefore that both permitted and forbidden lines affect the optical spectra of SNe Ia at this epoch. ## Day +205 {#subsec:11fe_p205} The day +205 spectrum of SN 2011fe and the corresponding `PHOENIX` spectrum are shown in Figure [\[fig:pah_std_d221_delta_t\_forb_coll_vs_11fe_p205\]](#fig:pah_std_d221_delta_t_forb_coll_vs_11fe_p205){reference-type="ref" reference="fig:pah_std_d221_delta_t_forb_coll_vs_11fe_p205"}. Attempts to calculate a spectrum at this epoch without forbidden lines, as was done in §[4.1](#subsec:11fe_p100){reference-type="ref" reference="subsec:11fe_p100"}, led to unrecoverable numerical instabilities in the code. It seems, then, that by this age forbidden lines play an important role. The single-ion spectra are shown in Figure [\[fig:pah_std_d221_delta_t\_forb_coll_vs_11fe_p205_single_ion_spectra\]](#fig:pah_std_d221_delta_t_forb_coll_vs_11fe_p205_single_ion_spectra){reference-type="ref" reference="fig:pah_std_d221_delta_t_forb_coll_vs_11fe_p205_single_ion_spectra"}. The emission feature at 4700 Å is primarily \[\] $\lambda \lambda 4607, 4658$ Å and \[\] $\lambda \lambda 4640, 4664$ Å. The weak but clearly separate features around 4300 Å are \[\] $\lambda \lambda 4287, 4359$ Å, and the emission at 5300 Å is \[\] $\lambda 5300$ Å. The double-horned feature in the synthetic spectrum centered around 7300 Å consists of \[\] $\lambda \lambda 7155, 7172$ Å on the left, and \[\] $\lambda 7412$ Å on the right; the shape of this pair of features is too exaggerated in the synthetic spectrum compared to the day +100 spectrum of SN 2011fe, but at later epochs the shape is a good match to the observations. At this epoch the ratio of to is well reproduced, with the strength of the 4700 Å emission from improved over that from the day +100 spectrum. However, the 5200 Å emission, also from , is somewhat weak. The strong IR3 emission which was overestimated in strength in they day +100 synthetic spectrum is now absent entirely. Coincidentally, the forbidden line \[\] $\lambda 8617$ Å, at nearly the same wavelength as IR3, has grown in strength at day +205, and fits quite well to the observation. The H & K doublet, which was quite strong at day +100, has diminished in strength and is replaced mostly by \[\] $\lambda 4008$ Å. It seems, then, that the serendipitous degeneracy among permitted and forbidden lines which found unlikely, may actually be realized, at least for some features in the optical spectra of SN 2011fe. The double-horned pair of emissions centered at 7200 Å is better reproduced at this epoch as well. Curiously, the emission at 5900 Å is now quite well fit with \[\] $\lambda 5888$ Å, while at day +100, when most of the $^{56}$Ni had decayed to $^{56}$Co, the feature was absent entirely. It is possible that the temperature in the day +100 model was too low, which would explain the underabundance of emitting at 4700 Å and emitting at 5900 Å. A higher temperature would also reduce the abundance of in favor of , explaining the reduced strength of both the H & K doublet and the IR3. ## Day +311 The observed and synthetic spectra at day +311 are shown in Figure [\[fig:pah_std_d331_delta_t\_forb_coll_vs_11fe_p311\]](#fig:pah_std_d331_delta_t_forb_coll_vs_11fe_p311){reference-type="ref" reference="fig:pah_std_d331_delta_t_forb_coll_vs_11fe_p311"}, and the single-ion spectra are shown in Figure [\[fig:pah_std_d331_delta_t\_forb_coll_vs_11fe_p311_single_ion_spectra\]](#fig:pah_std_d331_delta_t_forb_coll_vs_11fe_p311_single_ion_spectra){reference-type="ref" reference="fig:pah_std_d331_delta_t_forb_coll_vs_11fe_p311_single_ion_spectra"}. At this epoch the spectra look similar to those at day +205. The \[\] emission at 4700 Å is still strong, although at day +311 the \[\] emission at 4400 Å has grown in strength, and continues to do so at later epochs. This is likely a reflection of some (but not much) recombination from to at this age. The H & K emission is still present at 4000 Å, but somewhat weak, just as at day +205. ## Day +349 {#subsec:11fe_p349} The model and observations at day +349 are displayed in Figure [\[fig:pah_std_d376_delta_t\_vs_11fe_p349\]](#fig:pah_std_d376_delta_t_vs_11fe_p349){reference-type="ref" reference="fig:pah_std_d376_delta_t_vs_11fe_p349"}. The corresponding single-ion spectra are shown in Figure [\[fig:pah_std_d376_delta_t\_forb_coll_vs_11fe_p349_single_ion_spectra\]](#fig:pah_std_d376_delta_t_forb_coll_vs_11fe_p349_single_ion_spectra){reference-type="ref" reference="fig:pah_std_d376_delta_t_forb_coll_vs_11fe_p349_single_ion_spectra"}. The H & K doublet emission at 4000 Å is of similar strength as at day +205, and it may still contribute to that feature in SN 2011fe, although the \[\] appears to be stronger at that wavelength. At this and later epochs the model begins to exhibit some problems. A sharp emission feature forms in the near-UV in the synthetic spectrum which is not observed in SN 2011fe. In addition, in the synthetic spectrum the emission at 4400 Å (due mostly to ) is nearly as strong as the 4700 Å feature (mostly ), while in the observed spectrum the latter remains considerably stronger. This problem is likely not one of radiative transfer effects in the model, but rather one of atomic physics. The recombination rate for ions scales with the free electron density $n_e$, which dilutes geometrically roughly as $t^{-3}$. Thus at these very late times the recombination time scale for, e.g., , can be of the same order as the dynamical time scale, i.e., the age of the SN. In this case, time-dependent effects of ion recombination can become influential on spectrum formation. In the calculations used to generate the above figures, we neglected time-dependence in both the radiation field and the ion populations: both are assumed to be in steady-state. Assuming a steady-state radiation field is a valid approximation at late times---since at most wavelengths the optical depths in the ejecta are low, the radiative transfer time scale is effectively the light-crossing time, which is many orders of magnitude shorter than the dynamical time scale. Thus the radiation field equilibrates with the ejecta almost instantaneously at any given time $t$. However, by assuming steady-state in the ion populations, we overestimate the rate of recombination from, e.g., to . This manifests in synthetic spectra as features which are too strong, as in Figure [\[fig:pah_std_d376_delta_t\_vs_11fe_p349\]](#fig:pah_std_d376_delta_t_vs_11fe_p349){reference-type="ref" reference="fig:pah_std_d376_delta_t_vs_11fe_p349"}. Presumably the same pathology affects the day +578 synthetic spectrum more severely (see Figure [4.6](#subsec:11fe_p578){reference-type="ref" reference="subsec:11fe_p578"}). Time-dependent effects in the ion populations may not be the only source of the discrepant features in the synthetic spectra at very late times. The density profile of the ejecta in the explosion model also strongly affect $n_e$. Thus the -to- population ratio may provide a constraint on the initial conditions of the explosion model. For example, Figure [\[fig:pah_std_d376_delta_t\_vs_11fe_p349\]](#fig:pah_std_d376_delta_t_vs_11fe_p349){reference-type="ref" reference="fig:pah_std_d376_delta_t_vs_11fe_p349"} may indicate that the density of the iron-rich core of the model is too high, leading to an $n_e$ which is too high, inducing recombination from to too soon. ## The UV spectrum at day +360 {#subsec:11fe_p360} The UV spectrum from HST at day +360, as well as the best fitting spectrum from `PHOENIX`, are shown in Figure [\[fig:sn11fe_combined_hst_spectra_ft_smoothed_vs_pah_std_d376_fixed_tcor\]](#fig:sn11fe_combined_hst_spectra_ft_smoothed_vs_pah_std_d376_fixed_tcor){reference-type="ref" reference="fig:sn11fe_combined_hst_spectra_ft_smoothed_vs_pah_std_d376_fixed_tcor"}. The single-ion spectra are shown in Figure [\[fig:sn11fe_combined_hst_spectra_ft_smoothed_vs_pah_std_d376_fixed_tcor_single_ion\]](#fig:sn11fe_combined_hst_spectra_ft_smoothed_vs_pah_std_d376_fixed_tcor_single_ion){reference-type="ref" reference="fig:sn11fe_combined_hst_spectra_ft_smoothed_vs_pah_std_d376_fixed_tcor_single_ion"}. From these one finds that is responsible for most of the spectral features in the UV at day +360, just as it was at day +100. However, , , and all contribute significantly to the bluest portion of the spectrum as well. The most interesting result, however, is that the H & K doublet continues to contribute significantly to the emission around 4000 Å, despite being over 1 yr since explosion. It seems, then, that the extreme strength of this line overcomes both the small total abundance of in the ejecta, as well as the large amount of geometric dilution which accompanies a year of expansion. A second notable feature of this result is that nearly all of the features are permitted lines, not forbidden; the bluest forbidden lines for any ion in this version of the `PHOENIX` atomic database is about 3200 Å. Identifying these features is no simple feat, however, because has thousands of lines between $\sim 1600-4000$ Å. Furthermore, the contributions from several other ions in the UV at late times, each with several thousands of lines themselves, are at some wavelengths of similar strength as . The convolution of all of these lines from different species renders the identification of individual features in the UV a difficult task. However, even without identifying particular lines, we can nevertheless learn a great deal about the UV line forming region using other methods (see Figure [6](#sec:opacity_late_times){reference-type="ref" reference="sec:opacity_late_times"}). ## Day +578 {#subsec:11fe_p578} Figure [\[fig:pah_std_d594_vs_11fe_p578\]](#fig:pah_std_d594_vs_11fe_p578){reference-type="ref" reference="fig:pah_std_d594_vs_11fe_p578"} shows the observed and synthetic spectra of SN 2011fe at day +578. The single-ion spectra are shown in Figure [\[fig:pah_std_d594_delta_t\_forb_coll_vs_11fe_p578_single_ion_spectra\]](#fig:pah_std_d594_delta_t_forb_coll_vs_11fe_p578_single_ion_spectra){reference-type="ref" reference="fig:pah_std_d594_delta_t_forb_coll_vs_11fe_p578_single_ion_spectra"}. The fit of the `PHOENIX` synthetic spectrum to the observation is poor, and has resisted improvement even experimenting with a variety of different temperature-correction algorithms. Possible culprits for this include time-dependent effects as discussed in §[4.4](#subsec:11fe_p349){reference-type="ref" reference="subsec:11fe_p349"}, as well as other physical processes which `PHOENIX` currently does not treat, including dielectric recombination and charge-exchange reactions. However, analysis of this result nevertheless reveals some useful information. For example, the unphysical spike in flux around 3250 Å is due entirely to , although to which line in particular is not clear. In addition, the emission at 8600 Å formerly produced by IR3, has been replaced by \[\] $\lambda 8617$ Å. This is yet another example of a truly remarkable degeneracy among permitted lines and forbidden lines at similar wavelength, but which become active at very different times. Although our model spectra predict the recombination to too early, the event eventually does happen in SN 2011fe. In particular, in the day +594 spectrum, the strong emission peak at 4700 Å has disappeared entirely, with only a handful of features remaining. Indeed, and have tentatively identified features of in $\sim 1000$ d spectrum of SN 2011fe, perhaps heralding a concurrent recombination transition from to . # Velocity Shifts In a recent paper, examined the wavelength shifts of prominent features at late times in a series of SNe Ia spectra, including the spectra of SN 2011fe. They found, in particular, a redward shift of of the prominent 4700 Åfeature with no signs of the redward drift slowing down at epochs up to day +400. Figure [\[fig:feature_shifts\]](#fig:feature_shifts){reference-type="ref" reference="fig:feature_shifts"} shows that our models show no such general trend. In fact, the 4700 Åfeature after a strong redward shift, begins to move back to the blue. Since our models do not show a strong fidelity with the observations it is difficult to draw firm conclusions. suggest that the redward shift is primarily caused by temporal variations due to lines of \[Fe III\] and \[Co II\]. Since these lines are very temperature sensitive, this could be indicative of our general model uncertainties. # Opacity at late times {#sec:opacity_late_times} Our `PHOENIX` calculations have been fairly successful at reproducing the late-time optical and UV spectra of the normal SN 2011fe. In addition, one of the great advantages of using first-principles codes such as `PHOENIX` is that one may glean a great deal of information about the underlying physics which drives the formation of the synthetic spectra. For example, in Figure [\[fig:optical_depths_optical_p349\]](#fig:optical_depths_optical_p349){reference-type="ref" reference="fig:optical_depths_optical_p349"} we show the optical depths along the $\mu =-1$ (radially inward) ray in the day +349 model whose spectrum was shown in Figure [\[fig:pah_std_d376_delta_t\_vs_11fe_p349\]](#fig:pah_std_d376_delta_t_vs_11fe_p349){reference-type="ref" reference="fig:pah_std_d376_delta_t_vs_11fe_p349"}. For reference, the black dashed line shows $\tau = 1$, the division between optically thick and optically thin. The dashed red line shows the optical depth due only to Thomson scattering; at early times this is the dominant opacity source and gives rise to the photosphere in SNe. At late times, however, the geometric dilution of the free electron density $n_e$ leads to a very low Thomson scattering opacity, falling well below $\tau = 1$ even all the way to the center of the ejecta. From this alone we may infer that there is likely very little continuum radiation present this late in a SN Ia's lifetime, as reflected in the spectra. Furthermore, at only a select few wavelengths--mostly on the blue edge of the optical band--does the optical depth reach $\tau = 1$ at all; at most wavelengths $\lambda \gtrsim 4500$ Å, the ejecta are quite optically thin, in agreement with previous studies. Our calculations therefore indicate that most of the optical spectrum at day +349 in SN 2011fe consists of blended emission features from collisionally excited, optically thin forbidden lines. The UV portion of the spectrum of SN 2011fe behaves entirely differently than the optical, however. One may suspect as much simply by noticing the significant degree of structure and complexity in the observed UV spectrum (Figure [\[fig:sn11fe_combined_hst_spectra_ft_smoothed_no_model\]](#fig:sn11fe_combined_hst_spectra_ft_smoothed_no_model){reference-type="ref" reference="fig:sn11fe_combined_hst_spectra_ft_smoothed_no_model"}). These suspicions are confirmed by analogous calculations of optical depths at various UV wavelengths, shown in Figure [\[fig:optical_depths_UV_p349\]](#fig:optical_depths_UV_p349){reference-type="ref" reference="fig:optical_depths_UV_p349"}. The black and red dashed lines are the same as in Figure [\[fig:optical_depths_optical_p349\]](#fig:optical_depths_optical_p349){reference-type="ref" reference="fig:optical_depths_optical_p349"}. Unlike the optical band, however, most UV wavelengths are extremely optically thick, with many reaching $\tau \sim 10^5$ at the center of the ejecta. Another surprising result is that many UV wavelengths become optically thick at quite high velocity, crossing the $\tau = 1$ threshold at $v \sim 10 000-15000$ km s$^{-1}$. This result is corroborated by the presence of the emission component of the H & K doublet near 4000 Å in Figure [\[fig:sn11fe_combined_hst_spectra_ft_smoothed_vs_pah_std_d376_fixed_tcor_single_ion\]](#fig:sn11fe_combined_hst_spectra_ft_smoothed_vs_pah_std_d376_fixed_tcor_single_ion){reference-type="ref" reference="fig:sn11fe_combined_hst_spectra_ft_smoothed_vs_pah_std_d376_fixed_tcor_single_ion"}. If the UV remains as optically thick as Figure [\[fig:optical_depths_optical_p349\]](#fig:optical_depths_optical_p349){reference-type="ref" reference="fig:optical_depths_optical_p349"} and Figure [\[fig:optical_depths_UV_p349\]](#fig:optical_depths_UV_p349){reference-type="ref" reference="fig:optical_depths_UV_p349"} suggest, then it appears that the transition from the "photospheric" to the "nebular" phase in SNe Ia is far more complex than expected. Specifically, there are likely few or no forbidden emission lines which are active in the UV; this precludes the possibility of measuring asymmetric bulk motion of the inner regions of the SN ejecta, since the assumption behind such measurements is that the emission lines are optically thin and centered at the line rest wavelengths. On the other hand, since the UV is optically thick, the spectrum may consist of the same overlapping P Cygni line scattering profiles which characterize maximum-light spectra of SNe; if this is the case we may be able to infer ejecta velocities of the iron-rich core of SN 2011fe by measuring the location of the absorption minima of the features in Figure [\[fig:sn11fe_combined_hst_spectra_ft_smoothed_no_model\]](#fig:sn11fe_combined_hst_spectra_ft_smoothed_no_model){reference-type="ref" reference="fig:sn11fe_combined_hst_spectra_ft_smoothed_no_model"}. However, as discussed in §[4.5](#subsec:11fe_p360){reference-type="ref" reference="subsec:11fe_p360"}, the multitude of UV lines, as well as likely blending among several atomic species, make this challenging. If we entertain the possibility that some of the UV spectrum is forming at the UV photosphere at velocities of order $10 000$ km s$^{-1}$, the rest wavelengths of such lines would still lie within a crowded space of UV and optical transitions of iron-peak elements. The problem with the later fits may be related to time-dependent effects in the ionization of the gas, as the recombination time scale at late times is likely of order of the age of the SN. By assuming steady-state in our calculations, we likely overestimated the recombination rate, leading to an overabundance of with respect to . Re-computing these models with time-dependence in the NLTE rate equations is possible in principle, but the inherent "noisiness" of the root-finding algorithm we used to calculate the temperatures becomes amplified with each time step, resulting in a large amount of spurious temperature oscillations in the model at very late times, and a poorly fitting spectrum. # Conclusions We extended `PHOENIX` to calculate radiative transfer models well into the late-time epochs of SNe Ia, with an eye toward obtaining good fits to the high-quality optical and UV spectra of SN 2011fe. Doing so required similar methods to those discussed in, in particular using an alternative method to that of Unsöld-Lucy for calculating the temperature structure of the gas, as well as accounting for the collisional and radiative rate data for forbidden lines, which behave quite differently than permitted lines. The resulting synthetic spectra, ranging from +100 to +578 days post-maximum light, vary in degrees of fidelity to corresponding observed spectra of SN 2011fe, with the earlier epochs fitting quite well and the later epochs less so. At day +100 we found that radiative transfer calculations which neglect forbidden lines and those which include them can produce remarkably similar optical spectra, but with quite different atomic species and combinations of lines forming the various features. We found that, at least as late as day +360, permitted lines such as H & K and IR3 continue to influence spectrum formation in the optical, and permitted lines of form much of the spectrum in the UV. In addition, these models indicate that some emission features from permitted lines are replaced by other emission features of forbidden lines at nearly the same wavelength as the SN evolves. For example, the emission from H & K at 4000 Å is replaced around day +205 by \[\] $\lambda 4008$ Å, and the emission from IR3 at around 8600 Å is replaced by \[\] $\lambda 8617$ Å.	{'timestamp': '2016-07-19T02:05:12', 'yymm': '1607', 'arxiv_id': '1607.04784', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04784'}
# High-field instability from spin-wave theory In the asymptotic high-field limit all spins are aligned along the field axis. Magnon excitations are suppressed by a large energy gap. By lowering the field strength the magnon gap decreases and eventually vanishes at some critical field strength $h_{\mathrm{c}0}$. Below $h_{\mathrm{c}0}$ the high-field state becomes unstable, indicating a transition towards one of the various intermediate-field phases. (This is true as long as the continuous transition is not preempted by a first-order transition at some higher field $h_{\mathrm c} > h_{\mathrm{c}0}$.) The wavevector at which the magnon gap closes then determines the ordering wavevector of the intermediate-field phase. We parameterize the magnon excitations above the polarized ground state at high field by Holstein-Primakoff bosons $a_i$ and $b_i$ on the A and B sublattices of the honeycomb lattice. It is convenient to use a spin-space frame obtained by rotating the cubic-axes basis $\vec e_x$, $\vec e_y$, $\vec e_z$ such that the magnetic field lies in the $3$-direction, \[\begin{aligned} \label{eq:spin-frame} \vec e_1 & = \frac{(\vec e_z \times \vec h) \times \vec h}{\lvert (\vec e_z \times \vec h) \times \vec h \rvert}, & \vec e_2 & = \frac{\vec e_z \times \vec h}{\lvert \vec e_z \times \vec h \rvert}, & \vec e_3 & = \frac{\vec h}{\lvert \vec h \rvert}. \end{aligned}\] E.g., for field in the diagonal $[111]$ direction we choose the new spin-basis vectors $\vec e_1 = (\vec e_x + \vec e_y-2 \vec e_z)/\sqrt{6}$, $\vec e_2 = (-\vec e_x + \vec e_y)/\sqrt{2}$, and $\vec e_3 = (\vec e_x + \vec e_y + \vec e_z)/\sqrt{3}$. To leading order in the $1/S$ expansion the spin operators in this basis read: Here, $\vec \delta_x$, $\vec \delta_y$, and $\vec \delta_z$ are the nearest-neighbor vectors on $x$, $y$, and $z$ bonds, respectively, of the honeycomb lattice. The leading-order piece of $\mathcal H_\mathrm{SW}$ is quadratic in boson operators and can be diagonalized analytically by means of a Bogoliubov transformation. Resulting magnon spectra with $h \equiv \lvert \vec h \rvert$ tuned to the instability field strength $h_{\mathrm{c}0}$ are depicted for different coupling parameters $\varphi$ and field directions $\hat h \equiv \vec h / h$ in Figs. [\[fig:magnon-spectrum-001\]](#fig:magnon-spectrum-001){reference-type="ref" reference="fig:magnon-spectrum-001"}--[\[fig:magnon-spectrum-1107\]](#fig:magnon-spectrum-1107){reference-type="ref" reference="fig:magnon-spectrum-1107"}. For large $h/S \gg \|J\|, \|K\|$ the minimum of the magnon dispersion is always at the $\Gamma$ point in the Brillouin zone. As long as $\|K\| \ll \|J\|$ (i.e., when the zero-field ground state is a simple Néel or FM state), it remains at the $\Gamma$ point upon decreasing $h \searrow h_{\mathrm{c}0}$ at which it eventually vanishes; cf. right panel of Fig. [\[fig:magnon-spectrum-111\]](#fig:magnon-spectrum-111){reference-type="ref" reference="fig:magnon-spectrum-111"}. Above the zigzag and stripy zero-field ground states, however, the minimum of the dispersion shifts discontinuously from $\Gamma$ towards a finite wavevector as a function of field. For field in the $[001]$ direction ($[110]$ direction) the instability wavevector at which the magnon gap eventually vanishes is at two (one) of the three inequivalent $\mathrm M$ points in the Brillouin zone, indicating a direct continuous transition towards the canted zigzag or stripy phase; see Fig. [\[fig:magnon-spectrum-001\]](#fig:magnon-spectrum-001){reference-type="ref" reference="fig:magnon-spectrum-001"}. For field in the $[111]$ direction (left and middle panel of Fig. [\[fig:magnon-spectrum-111\]](#fig:magnon-spectrum-111){reference-type="ref" reference="fig:magnon-spectrum-111"}), by contrast, the instability wavevector is at the $\mathrm K$ points, forbidding a direct continuous transition towards a simple canted deformation of the zero-field ground state. Remarkably, intermediate field directions lead to magnon softening at incommensurate wavevectors, see Fig. [\[fig:magnon-spectrum-1107\]](#fig:magnon-spectrum-1107){reference-type="ref" reference="fig:magnon-spectrum-1107"} for field in the $[11 \frac{7}{10}]$ direction; we leave a detailed study of the resulting ordered states for future work. We note that there is a linear band crossing point for field in the $[001]$ direction when $-0.148\pi < \varphi < 0.687\pi$. (An analogous band crossing point occurs when $0.852\pi < \varphi < 1.687\pi$.) This can be understood as a "Dirac magnon" that is located at the K point (and finite energy) in the Heisenberg limit $\varphi = 0$, and shifted from K towards the $\Gamma$ point (M$_2$ point) for finite $\varphi>0$ (finite $\varphi < 0$). Another such bosonic Dirac point is located at the opposite \[with respect to the $\Gamma$ point (M$_2$ point) for $\varphi > 0$ ($\varphi < 0$)\] wavevector. At $\varphi = 0.687\pi$ ($\varphi =-0.148\pi$) both merge and annihilate at the $\Gamma$ point (M$_2$ point). Explicit values for the instability field strength $h_{\mathrm{c}0}$ and corresponding instability wavevectors are given in Table [\[tab:instability-field\]](#tab:instability-field){reference-type="ref" reference="tab:instability-field"}. There, we have also indicated the special cases when the instability of the high-field magnon is preempted by a discontinuous transition, as obtained from the analytical parameterization of phases (Sec. [3](#sec:parametrization){reference-type="ref" reference="sec:parametrization"}). In all other cases, the transition from polarized towards intermediate-field phases is continuous, and we have checked that the instability field strength $h_{\mathrm{c}0}$ indeed then always coincides with the critical field strength $h_\mathrm{c}$ as obtained from the parametrization (Sec. [3](#sec:parametrization){reference-type="ref" reference="sec:parametrization"}), as well as with $h_\mathrm{c}$ from the MC data (Sec. [2](#sec:mc){reference-type="ref" reference="sec:mc"}). This also serves as an independent verification of the numerics. We also note that the magnetization process in the HK model on the 3D hyperhoneycomb lattice in $[111]$ field appears to be similarly complex as found here. This is because the magnon instability at $h_{\mathrm{c}0}$ (which happens to coincide with $h_{\mathrm{c}0}$ for the 2D honeycomb lattice) occurs above the metamagnetic first-order transitions found in the MC simulations. This has apparently been overlooked in the previous analysis. # Monte-Carlo simulations {#sec:mc} To identify the intermediate-field phases, we study the large-$S$ limit of the HK model by employing a combination of classical Monte Carlo (MC) simulations and energy minimization. We work on honeycomb lattices of size $L\times L$ with periodic boundary conditions. The lattices are spanned by the primitive lattice vectors $\vec{a}_{1\left(2\right)}=\left(3/2,\pm\sqrt{3}/2\right)$, with each unit cell containing two sites amounting to a total number of spins of $N=2L^{2}$. We perform equilibrium MC simulations using single-site updates with a combination of the heat-bath and microcanonical (or over-relaxation) algorithms, with typically $10^{7}$ MC steps per spin. We combine these updates with the parallel-tempering algorithm in order to efficiently equilibrate the MC configurations at very low $T$. From the MC data, we compute the uniform magnetization in the field direction (Fig. 4 in the main text) \[\label{eq:mag} \frac{\vec m \cdot \hat h}{S}=\left \langle \frac{1}{N}\sum_{i}\frac{\vec{S}_{i}}{S}\cdot\frac{\vec h}{\lvert \vec h \rvert}\right\rangle,\] where $\langle \cdots \rangle$ denotes MC average, as well as the static spin structure factor (Fig. 2 in the main text) \[\label{eq:sfactor} S_{\vec{k}} = \left\langle \vec{S}(\vec{k})\cdot \vec{S}(-\vec{k})\right\rangle,\] where \[\label{eq:sfourier} \vec{S}(\vec{k}) = \frac{1}{\sqrt{N}}\sum_{i}\vec{S}_{i}e^{-i\vec{k}\cdot\vec{R}_{i}},\] is the Fourier transform of a given spin configuration and $\vec R_i$ is the lattice vector at site $i$. To find the classical ground state, we start from a MC spin configuration obtained at low $T$ (typically $T/\lvert J S^2 \rvert \sim 0.005$) and then iteratively align the spins with their local fields $\vec{h}_{i}^\text{loc}$, \[\label{eq:min} \vec{S}_{i} = \frac{\vec{h}_{i}^\text{loc}}{\left\|\vec{h}_{i}^\text{loc}\right\|} S.\] Convergence is reached after the largest update in a lattice sweep, $\|\vec{S}_{i}^\text{new}-\vec{S}_{i}^\text{old}\|_\text{max}/S$, is smaller than $10^{-12}$. Because of the several competing ground states, it is important to start from unbiased MC configurations in order to obtain the correct classical ground state. We performed extensive field scans at $\varphi = 0.57\pi$, $0.62\pi$, $0.733\pi$, $0.83\pi$, $1.578\pi$, $1.687\pi$, $1.813\pi$, and $1.922\pi$ for the $[111]$ field direction, $\varphi = 0.62\pi$ and $1.687\pi$ for the $[001]$ direction, as well as $\varphi = 1.687\pi$ for the $[110]$ field direction, with system sizes up to $L=24$. For field in the $[111]$ direction we find a total of 10 phases at finite $h$. Spin-configuration snapshots from cooled MC data are depicted for selected parameter values in Fig. [\[fig:mc-snapshots\]](#fig:mc-snapshots){reference-type="ref" reference="fig:mc-snapshots"}. # Parametrization of phases {#sec:parametrization} The low-$T$ spin configurations obtained from the MC simulations allows the deduction of the symmetries, unit-cell sizes, and sublattice structure of the different phases. We make use of this information by parametrizing the spin configurations in terms of a set of angles, which then are optimized at fixed model parameters $\varphi$ and $\vec h$ to determine the state of lowest energy. Doing this for all phases enables a comparison of energies from which we deduce the classical phase diagram in the low-temperature limit. Using the rotated basis as defined in Eq. [\[eq:spin-frame\]](#eq:spin-frame){reference-type="eqref" reference="eq:spin-frame"}, the spin $\vec S_i$ at site $i$ can be parametrized as \[\label{eq:spin-ansatz} \vec S_i = S \left( \vec e_1 \sin \theta_i \cos \phi_i + \vec e_2 \sin \theta_i \sin \phi_i + \vec e_3 \cos \theta_i \right).\] In the polarized phase we have $\theta_i \equiv 0$, while $\theta_i > 0$ defines a canted state. For given coupling parameter $\varphi$ and magnetic field $\vec h = h \vec e_3$ our ansätze for the angles $\theta_i$ and $\phi_i$ as obtained from the MC simulations are given in Table [\[tab:angles\]](#tab:angles){reference-type="ref" reference="tab:angles"}. Except for the vortex, AF vortex, and zigzag star phases the spin projections onto the plane perpendicular to $\vec h$ ($\vec e_1$-$\vec e_2$ plane) are locked on the directions of the cubic-axes projections $\vec e_3 \times (\vec e_x \times \vec e_3)$, $\vec e_3 \times (\vec e_y \times \vec e_3)$, and $\vec e_3 \times (\vec e_z \times \vec e_3)$, see Fig. [\[fig:mc-snapshots\]](#fig:mc-snapshots){reference-type="ref" reference="fig:mc-snapshots"}. For these phases we therefore have $\phi_i \in \{0, \frac{\pi}{3}, \frac{2\pi}{3}, \pi, \frac{4\pi}{3}, \frac{5\pi}{3}\}$, and we may minimize with respect to the field-dependent canting angles $\theta_i$ only. In each case we in fact find that there are at most only two different possible $\theta$ angles \[indicated by the at most two different lengths of the spin projections in Fig. [\[fig:mc-snapshots\]](#fig:mc-snapshots){reference-type="ref" reference="fig:mc-snapshots"}(a), (d), and (e)\]. This makes the computation of the minimized energy of a given classical state and their comparison among different states numerically cheap. In the cases of the vortex and AF vortex phases the $\phi$ angles are not locked onto the projection of the cubic-axes direction. However, we find that the classical energy in these cases in fact becomes independent of the angle $\delta$ that determines the (uniform) deviation from the cubic-axes locking. (The MC data show that thermal fluctuations lift this degeneracy by an order-from-disorder mechanism.) By contrast, for the zigzag star phase we do not use any particular ansatz for the configuration, except for the fact (again as obtained from the cooled MC data) that the magnetic unit cell spans $2 \times 9$ crystallographic unit cells. The explicit assumptions for $\theta_i$ and $\phi_i$ for all states are summarized in Table [\[tab:angles\]](#tab:angles){reference-type="ref" reference="tab:angles"}. The parametrization allows the straightforward comparison of the minimized energies of the various states and the deduction of the phase boundaries for arbitrary coupling parameter $\varphi$ and field strength $h$ under the assumption that no further states (not parametrized in Table [\[tab:angles\]](#tab:angles){reference-type="ref" reference="tab:angles"}, and missed by the MC scans) are stabilized somewhere in the phase diagram. The result is depicted in Fig. 1 in the main text. In addition to the quadruple point at $(\varphi, h/(AS)) = (7\pi/4, 3/\sqrt{2})$ we find 8 triple points at ($\pi/2$, 4), ($0.55\pi$, 2.17), ($0.698\pi$, 1.38), ($0.715\pi$, 1.25), ($0.722\pi$, 1.01), ($1.673\pi$, 1.35), ($1.812\pi$, 2.76), and ($1.825\pi$, 2.62). The total magnetization in field direction $\hat h = \vec h / \lvert \vec h \rvert$ is given by $\vec m \cdot \hat h/S = N^{-1} \sum_i \cos \theta_i =-N^{-1} (\partial E/\partial h)$ with $E\equiv E(\varphi,h)$ as the ground-state energy for given $\varphi$ and $h$. The magnetization curves agree very well with the MC measurements, see Fig. 4 in the main text. Exceptions are a few data points very close to first-order transitions; we attribute these deviations to hysteresis effects in the MC simulations. We have explicitly checked that the minimized energy from the analytical parametrization is always less than or equal the one from the cooled MC configuration for the same parameters. We visualize several magnetic unit cells of the spin configurations for all canted phases in Fig. [\[fig:spin-configs\]](#fig:spin-configs){reference-type="ref" reference="fig:spin-configs"}. Fig. 3 in the main text analogously shows one respective magnetic unit cell for selected phases. In Fig. [\[fig:spin-configs\]](#fig:spin-configs){reference-type="ref" reference="fig:spin-configs"}, we display also the projections of the spin configurations onto the plane perpendicular to $\vec h$ (to be compared with the cooled MC spin configurations in Fig. [\[fig:mc-snapshots\]](#fig:mc-snapshots){reference-type="ref" reference="fig:mc-snapshots"}). We can also use the parametrized spin configurations to compute static spin structure factors, allowing a comparison with the MC structure factors (Fig. 2 in the main text). However, an efficient MC simulation (in our case with parallel tempering) averages over the full ground-state manifold. It consequently does not allow the direct distinction between single-and multi-$Q$ states. For example, while a pure "$+z$ zigzag" state with the spins of a particular zigzag line on the honeycomb lattice pointing along the $+z$ direction would exhibit a Bragg peak at only one out of the three inequivalent M points in the first Brillouin zone (M$_2$), the simulations always average over $\pm x$, $\pm y$, and $\pm z$ zigzag states (as long as these are degenerate), and the MC structure factors exhibit Bragg peaks at all three M points. Experimentally, this is equivalent to having multiple magnetic domains in a large sample. Using the analytical parametrization, by contrast, we can compute "single-domain" structure factors for fixed states without averaging over the ground-state manifold, allowing us to distinguish between single-$Q$ and multi-$Q$ phases in a direct way. In Fig. [\[fig:strctr-fctr_ana\]](#fig:strctr-fctr_ana){reference-type="ref" reference="fig:strctr-fctr_ana"} we show examples for the canted zigzag phase with a Bragg peak at only one out of the three M points in the first Brillouin zone (a), to be compared with the AF star phase which exhibits Braggs peaks at all three M points (b). Fig. [\[fig:strctr-fctr_ana\]](#fig:strctr-fctr_ana){reference-type="ref" reference="fig:strctr-fctr_ana"}(c) shows the single-domain structure factor of the zigzag star phase, with a total of 18 inequivalent Bragg peaks in the first Brillouin zone, to be compared with the MC averaged structure factor of Fig. 2 (b) in the main text. # Klein duality and star vs. zigzag/stripy phases We explain how the fact that the AF star and zigzag states, and analogously the FM star and stripy states, are classically degenerate for all $\varphi$ can be understood in terms of the Klein duality . This will also allow us to gain useful insight into the quantum-fluctuation effects on the phase diagram for $S=1/2$. We introduce the dual spins $\vec{S}_i'$ by dividing the honeycomb lattice into four sublattices $\mathrm A$, $\mathrm B$, $\mathrm C$, $\mathrm D$ and identifying \[\label{eq:duality} \vec{S}_i' \equiv \begin{cases} \vec S_i & \text{for } i \in \mathrm A,\\ \diag(1,-1,-1)\, \vec S_i & \text{for } i \in \mathrm B,\\ \diag(-1,1,-1)\, \vec S_i & \text{for } i \in \mathrm C,\\ \diag(-1,-1,1)\,\vec S_i & \text{for } i \in \mathrm D. \end{cases}\] In terms of the dual spins the Heisenberg-Kitaev Hamiltonian \[Eq. (1) in the main text\] can be written as \[\label{eq:hk-dual} \mathcal{H} =-J \sum_{\left\langle ij\right\rangle} \vec{S}_{i}' \cdot \vec{S}_{j}' + 2(K+J) \sum_{\left\langle ij\right\rangle_{\gamma}} S_{i}^{\gamma}{}' S_{j}^{\gamma}{}'-\sum_{i} \vec h_i' \cdot \vec{S}_i'.\] with $\vec h_i'$ denoting the dual magnetic field, obtained by a duality transformation that is analogous to the spin transformation in Eq. [\[eq:duality\]](#eq:duality){reference-type="eqref" reference="eq:duality"}. Eq. [\[eq:hk-dual\]](#eq:hk-dual){reference-type="eqref" reference="eq:hk-dual"} describes a Heisenberg-Kitaev model in a nonuniform field $\vec{h}_i'$. For $\vec{h}_i' = 0$ and $K=-J$, i.e., $\varphi \in \{3\pi/4, 7\pi/4\}$, it features a spin $\mathrm{SU}(2)$ symmetry that is hidden in the original basis. For finite field, a $\mathrm U(1)$ part of the hidden symmetry is left intact, if and only if $\vec h$ points along one of the cubic axes $\vec e_x$, $\vec e_y$, or $\vec e_z$ (and thus $\vec{h}_i'$ is parallel or antiparallel to this axis). For other field directions, no continuous spin symmetry remains at finite $\vec h$. Consider the exactly solvable "stripy Klein point" for $K=-J < 0$ (i.e., $\varphi = 7\pi/4$) and $\vec h = 0$. The quantum ground state is a ferromagnet with the dual spins pointing along a fixed, but arbitrary direction $\vec{S}_i' = S \vec n$ in spin space. Only the six states with $\vec n \in \{\pm \vec e_x, \pm \vec e_y, \pm \vec e_z\}$ out of this $\mathrm{SU}(2)$ degenerate ground-state manifold correspond to collinear spin configurations in the original basis. These are precisely the six possible stripy quantum ground states of the Heisenberg-Kitaev model. States for which only one (no) component of $\vec n$ in the cubic-axes basis vanishes correspond to coplanar (noncoplanar) spin textures in the original basis. Due to the hidden $\mathrm{SU}(2)$ symmetry an order-from-disorder mechanism can lift this quantum-ground-state degeneracy only away from the Klein point, e.g., when we consider a different set of couplings with $\varphi \notin \{3\pi/4, 7\pi/4\}$ or switch on an external field $\vec h \neq 0$. In fact, these states belong to the highly-degenerate ground-state manifold of the classical Kitaev model, and thus have the same classical energy for all $\varphi$. For $7\pi/4 \leq \varphi < 1.85\pi$, we find classically that an infinitesimally small field $\vec h \parallel [111]$ lifts the degeneracy in favor of a state in which also $\vec n \parallel [111]$. In the original spin basis this state corresponds to the FM star configuration. Upon inclusion of quantum fluctuations, one may expect that an order-from-disorder mechanism will shift the phase boundary between the stripy phase and FM star phase from zero field for $S \to \infty$ to finite values of the field for $S = 1/2$ if $\varphi > 7\pi/4$. Directly at $\varphi = 7\pi/4$, however, the degeneracy survives in the quantum case because of the presence of the hidden $\mathrm{SU}(2)$ symmetry. We infer (in the sense of degenerate perturbation theory in small $h$) that the FM star phase reaches all the way down to $h \searrow 0$, with the triple point at $(\varphi,h) = (7\pi/4,0)$ staying at zero field also for $S=1/2$. Note that the above argument does not rely on the fact that quantum fluctuations are absent in the zero-field ground state of the dual FM model. An analogous mechanism should therefore be expected at the "zigzag Klein point" for $K=-J > 0$, i.e., $\varphi = 3\pi/4$. Here, the zero-field ground state in the dual basis is a Néel antiferromagnet with $\vec{S}_i' = (-1)^i S \vec n$ along an arbitrary direction $\vec n$. $\vec n \in \{\pm \vec e_x, \pm \vec e_y, \pm \vec e_z\}$ corresponds to one of the six possible zigzag states in the original basis. $\vec n \parallel [111]$ corresponds to the noncoplanar AF star phase. Again, we find that a finite $\vec h \parallel [111]$ lifts the degeneracy in favor of the state with $\vec n \parallel [111]$. Due to the absence of an order-from-disorder mechanism in the hidden-$\mathrm{SU}(2)$-symmetric model when $\varphi = 3\pi/4$ and $\vec h=0$, we expect that the degenerate zero-field ground state gives way to an AF star ground state at infinitesimal field in the $[111]$ direction also in the quantum limit when $S=1/2$. We conclude that finite regions of both FM star and AF star phases exist for field in the $[111]$ direction not only classically, but also in the quantum phase diagram for $S=1/2$, at least in the vicinity of the Klein points at $\varphi = 3\pi/4$ or $7\pi/4$.	{'timestamp': '2016-11-07T02:07:19', 'yymm': '1607', 'arxiv_id': '1607.04640', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04640'}
# Introduction The black hole information puzzle is the puzzle of whether black hole formation and evaporation is unitary, and debate on this issue has continued for more than 36 years, since Hawking radiation was discovered. Hawking originally used local quantum field theory in the semiclassical spacetime background of an evaporating black hole to deduce that part of the information about the initial quantum state would be destroyed or leave our Universe at the singularity or quantum gravity region at or near the centre of the black hole, so that what remained outside after the black hole evaporated would not be given by unitary evolution from the initial state. However, this approach does not fully apply quantum theory to the gravitational field itself, so it was objected that the information-loss conclusion drawn from it might not apply in quantum gravity. Maldacena's AdS/CFT conjecture has perhaps provided the greatest impetus for the view that quantum gravity should be unitary within our Universe and give no loss of information. If one believes in local quantum field theory outside a black hole and also that one would not experience extreme harmful conditions ('drama') immediately upon falling into any black hole sufficiently large that the curvature at the surface would not be expected to be dangerous, then recent papers by Almheiri, Marolf, Polchinski, and Sully (AMPS), and by them and Stanford (AMPSS), give a new challenge to unitarity, as they argued that unitarity, locality, and no drama are mutually inconsistent. It seems to us that locality is the most dubious of these three assumptions. Nevertheless, locality seems to be such a good approximation experimentally that we would like a much better understanding of how its violation in quantum gravity might be able to preserve unitarity and yet not lead to the drama of firewalls or to violations of locality so strong that they would be inconsistent with our observations. Giddings (occasionally with collaborators) has perhaps done the most to investigate unitary nonlocal models for quantum gravity. For other black hole qubit models, see. Here we present a qubit toy model for how a black hole might evaporate unitarily and without firewalls, but with nonlocal gravitational degrees of freedom. We model radiation modes emitted by a black hole as localized qubits that interact locally with these nonlocal gravitational degrees of freedom. Similar models were first investigated by Giddings in his previously referred papers, particularly in. Nomura and his colleagues also have a model with some similarities to ours. In this way we can go from modes near the horizon that to an infalling observer appear to be close to a vacuum state (and hence without a firewall), and yet the modes that propagate outward can pick up information from the nonlocal gravitational field they pass through so that they transfer that information out from the black hole. # Qualitative Description of Our Qubit Model Using Planck units in which $\hbar = c = G = k_\mathrm{Boltzmann} = 1$, a black hole that forms of area $A$ and Bekenstein-Hawking entropy $S_\mathrm{BH} = A/4$ may be considered to have $e^{S_\mathrm{BH}} = 2^{S_\mathrm{BH}/(\ln{2})}$ orthonormal states, which is the same number as the number of orthonormal states of $n = S_\mathrm{BH}/(\ln{2}) = A/(4\ln{2})$ qubits if this is an integer, which for simplicity we shall assume. We shall take the state of these $n$ qubits as being the state of the gravitational field of the black hole. We assume that this state is rapidly scrambled by highly complex unitary transformations, so that generically a black hole formed by collapse, even if it is initially in a pure state, will have these $n$ qubits highly entangled with each other. However, in our model we shall assume that there are an additional $n$ qubits of outgoing radiation modes just outside the horizon, and a third set of $n$ qubits of outgoing but infalling radiation modes just inside the horizon. We shall assume that these two sets of qubits have a unique pairing (as partner modes in the beginning of the Hawking radiation) and further that each pair is in the singlet Bell state that we shall take to represent the vacuum state as seen by an infalling observer, so that all of these $2n$ qubits of radiation modes near the black hole horizon are in the vacuum pure state and hence give no contribution to the Bekenstein-Hawking entropy $S_\mathrm{BH} = n\ln{2}$. We thus explicitly assume that the infalling observer sees only the vacuum and no firewall in crossing the event horizon. See for one argument for justifying this assumption. Now we assume that the Hawking emission of one mode corresponds to one of the $n$ outgoing radiation modes from just outside the horizon propagating to radial infinity. However, the new assumption of this model is that the radiation qubit that propagates outward interacts (locally) with one of the $n$ nonlocal qubits representing the black hole gravitational field, in just such a way that when the mode gets to infinity, the quantum state of that radiation qubit is interchanged with the quantum state of the corresponding black hole gravitational field qubit. This is a purely unitary transformation, not leading to any loss of information. Assume for simplicity that the black hole forms in a pure state that becomes highly scrambled by a unitary transformation. Therefore, as an early outgoing radiation qubit propagates out to become part of the Hawking radiation, when it interchanges its state with that of the corresponding gravitational field qubit, it will become nearly maximally entangled with the black hole state and will have von Neumann entropy very nearly $\ln{2}$, the maximum for a qubit. So the early Hawking radiation qubits will each have nearly the maximum entropy allowed, and there will be very little entanglement between the early radiation qubits themselves. Meanwhile, the black hole qubit corresponding to each outgoing radiation qubit will have taken on the state that the outgoing radiation qubit had when it was just outside the horizon and hence be in the unique singlet Bell state with the infalling radiation qubit just inside the horizon that was originally paired with the outgoing qubit. This vacuum singlet Bell state can then be omitted from the analysis without any loss of information. In this way we can model the reduction in the size of the black hole as it evaporates by the reduction of the number of black hole qubits. We might say that each such vacuum Bell pair falls into the singularity, but what hits the singularity in this model is a unique quantum state, similar to the proposal of Horowitz and Maldacena. Therefore, if we start with $n$ black hole gravitational field qubits, $n$ outgoing radiation qubits just outside the horizon, and $n$ infalling radiation qubits just inside the horizon, after the emission of $n_r$ outgoing radiation qubits, $n_r$ of the infalling radiation qubits will have combined into a unique quantum state with the $n_r$ black hole qubits that were originally interacting with the $n_r$ outgoing radiation qubits that escaped, so that we can ignore them as what we might regard as merely vacuum fluctuations. This leaves $n-n_r$ pairs of outgoing radiation qubits just outside the horizon and infalling qubits just inside the horizon (each pair being in the singlet Bell state), and $n-n_r$ black hole gravitational field qubits. Eventually the number of Hawking radiation qubits, $n_r$, exceeds the number of black hole qubits remaining, $n-n_r$, when $n_r > n/2$, and the black hole becomes 'old.' At this stage, the remaining black hole qubits all become nearly maximally entangled with the Hawking radiation qubits, so that the von Neumann entropy of the black hole becomes very nearly $(n-n_r)\ln{2}$, which we shall assume is very nearly $A/4$ at that time. Since the whole system is assumed to be in a pure state, and since we have assumed unitary evolution throughout, the von Neumann entropy of the Hawking radiation at this late stage is also very nearly $(n-n_r)\ln{2}$, but now this is less than the maximum value, which is $n_r\ln{2}$. Thus each of the $n_r$ Hawking radiation qubits can no longer be maximally entangled with the remaining $n-n_r$ black hole qubits, and significant entanglement begins to develop between the Hawking radiation qubits themselves. Nevertheless, for any collection of $n' < n/2$ qubits of the Hawking radiation, the von Neumann entropy of that collection is expected to be very nearly $n'\ln{2}$, so one would still find negligible quantum correlations between any collection of $n'$ Hawking radiation qubits. Finally, when all $n$ of the original outgoing radiation qubits have left the black hole and propagated to infinity to become Hawking radiation qubits, there are no qubits left for the black hole; hence it has completely evaporated away. The $n$ Hawking radiation qubits now form a pure state, just as the original quantum state that formed the black hole was assumed to be. Of course, the unitary scrambling transformation of the black hole qubits means that the pure state of the final Hawking radiation can look quite different from the initial state that formed the black hole, but the two are related by a unitary transformation. The net effect is that the emission of one outgoing radiation qubit gives the transfer of the information in one black hole qubit to one Hawking radiation qubit. But rather than simply saying that this transfer is nonlocal, from the inside of the black hole to the outside, we are saying that the black hole qubit itself is always nonlocal, and that the outgoing radiation qubit picks up the information in the black hole qubit locally, as it travels outward through the nonlocal gravitational field of the black hole. Therefore, in this picture in which we have separated the quantum field theory qubits of the radiation from the black hole qubits of the gravitational field, we do not need to require any nonlocality for the quantum field theory modes, but only for the gravitational field. In this way the nonlocality of quantum gravity might not have much observable effect on experiments in the laboratory focussing mainly on local quantum field theory modes. # Mathematics of Qubit Transport Before the black hole forms, we assume that we have a Hilbert space of dimension $2^n$ in which each state collapses to form a black hole whose gravitational field can be represented by $n$ nonlocal qubits. We assume that we have a pure initial state represented by the set of $2^n$ amplitudes $A_{q_1q_2\ldots q_n}$, where for each $i$ running from 1 to $n$, the corresponding $q_i$ can be 0 or 1, representing the two basis states of the $i$th qubit. Once the black hole forms, without changing the Hilbert space dimension, we can augment this Hilbert space by taking its tensor product with a 1-dimensional Hilbert space for the vacuum state of $n$ infalling and $n$ outgoing radiation modes just inside and just outside the event horizon. We shall assume that this vacuum state is the tensor product of vacuum states for each pair of modes, with each pair being in the singlet Bell state that we shall take to represent the vacuum for that pair of modes. That is, once the black hole forms, we assume that we have $n$ nonlocal qubits for the gravitational field of the black hole, labeled by $a_i$, where $i$ runs from 1 to $n$, $n$ localized qubits for the infalling radiation modes just inside the horizon, labeled by $b_i$, and $n$ localized qubits for the outgoing radiation modes just outside the horizon, labeled by $c_i$. Suppose that each qubit has basis states $\ket{0}$ and $\ket{1}$, where subscripts (either $a_i$, $b_i$, or $c_i$) will label which of the $3n$ qubits one is considering. We assume that each pair of infalling and outgoing radiation qubits is in the vacuum singlet Bell state \[\ket{B}_{b_i c_i} = \frac{1}{\sqrt{2}}\Bigl(\ket{0}_{b_i}\ket{1}_{c_i}-\ket{1}_{b_i}\ket{0}_{c_i}\Bigr). \label{Bell}\] Initially the quantum state of the black hole gravitational field and radiation modes is \[\ket{\Psi_0}=\sum_{q_1=0}^1\sum_{q_2=0}^1\cdots\sum_{q_n=0}^1 A_{q_1q_2\ldots q_n}\prod_{i=1}^n\ket{q_i}_{a_i}\prod_{i=1}^n\ket{B}_{b_ic_i}, \label{initial state}\] where the $A_{q_1q_2\ldots q_n}$ are the amplitudes for the $2^n$ product basis states for the black hole gravitational field. Note that the entire quantum state is the product of a state of all the black hole gravitational qubits and a single pure vacuum state for the radiation modes. During the emission of the $i$th radiation mode to become a mode of Hawking radiation at radial infinity, the basis state for the subsystem of the $i$th black hole, infalling radiation, and outgoing radiation qubits changes as \[\ket{q_i}_{a_i}\ket{B}_{b_ic_i} \mapsto-\ket{B}_{a_ib_i}\ket{q_i}_{c_i}, \label{transfer}\] where $\ket{B}_{a_ib_i}$ is the analogue of $\ket{B}_{b_ic_i}$ given by Eq. ([\[Bell\]](#Bell){reference-type="ref" reference="Bell"}) with $b_i$ replaced by $a_i$ and $c_i$ replaced by $b_i$. As is obvious from the expressions on the right hand sides, this just interchanges the state of the $i$th black hole qubit with the state of the $i$th outgoing radiation qubit. If $P_{a_ic_i} = \ket{B}_{a_ic_i}\bra{B}_{a_ic_i}$ multiplied by the identity operator in the $b_i$ subspace, then for $\theta = \pi$ the continuous sequence of unitary transformations \[U(\theta)=\exp\Bigl(-i\theta P_{a_ic_i}\Bigr)={\rm I}+(e^{-i\theta}-1)P_{a_ic_i} \label{Unitary operator for qubit transfer}\] becomes $U(\pi) = {\rm I}-2P_{a_ic_i}$, which gives the unitary transformation [\[transfer\]](#transfer){reference-type="eqref" reference="transfer"}, interchanging the states of the $i$th black hole qubit with the state of the outgoing radiation qubit. We might suppose that as the radiation qubit moves outward, the $\theta$ parameter of the unitary transformation is a function of the radius $r$ that changes from 0 at the horizon to $\pi$ at radial infinity. For example, one could take $\theta = \pi(1-K/K_h)$, where $K$ is some curvature invariant (such as the Kretschmann invariant, $K = R^{\mu\nu\rho\sigma}R_{\mu\nu\rho\sigma}$) that decreases monotonically from some positive value at the horizon (where its value is $K_h$) to zero at infinity. We now assume that after the emission of the $i$th mode, the vacuum Bell state of the $i$th black hole qubit and the $i$th infalling radiation qubit can be dropped from the analysis, so that one only has the Hawking radiation qubit remaining for that $i$. Then the state of the subsystem for that $i$ goes from $-\ket{B}_{a_ib_i}\ket{q_i}_{c_i}$ given by Eq. ([\[transfer\]](#transfer){reference-type="ref" reference="transfer"}) to simply $\ket{q_i}_{c_i}$ for the qubit representing the Hawking radiation mode. Therefore, after all of the $n$ outgoing radiation modes propagate out to infinity while interacting with the black hole gravitational field, and after all the Bell vacua left inside the black hole are omitted, one is left with no black hole and the Hawking radiation in the final pure state \[\ket{\Psi_1}=\sum_{q_1=0}^1\sum_{q_2=0}^1\cdots\sum_{q_n=0}^1 A_{q_1q_2\ldots q_n}\prod_{i=1}^n\ket{q_i}_{c_i}. \label{final state}\] As a note we require that nonlocal gravitational qubits $a_i$ do not create firewalls by themselves. That is, even though the vacuum states on the horizon $b_i,c_i$ are in the range of nonlocal effects, they remain to be constrained in the singlet state unless systems $c_i$ are propagating away to infinity as Hawking radiation by Eq. [\[Unitary operator for qubit transfer\]](#Unitary operator for qubit transfer){reference-type="eqref" reference="Unitary operator for qubit transfer"}. This is consistent with the above assumption that the parameter $\theta$ in Eq. [\[Unitary operator for qubit transfer\]](#Unitary operator for qubit transfer){reference-type="eqref" reference="Unitary operator for qubit transfer"} is a function of the radius $r$. Conversely, it seems plausible to assume that any incoming mode gradually drops off some of its information during propagation through this nonlocal gravitational field. ## Mining Issue AMPSS, whose Eq. (3.3) is essentially the same as our [\[transfer\]](#transfer){reference-type="eqref" reference="transfer"}, raised the following issue with subsystem transfer models as resolutions of the firewall paradox. Suppose there exists an ideal mining equipment that can approach arbitrarily close to the horizon without falling into it, and then the equipment interacts with one of systems $c_i$ just outside the horizon. Note that this can be done without any exchange of energy due to the infinite redshift, and it is assumed that there is no entangling either. For example, the mining equipment can unitarily acts on the system $c_i$ as \[\begin{aligned} U_{\text{mine}}&:&\ket{0}_{c_i}\mapsto e^{i\phi}\ket{0}_{c_i},\;\;\;\ket{1}_{c_i}\mapsto e^{-i\phi}\ket{1}_{c_i}.\\ U_{\text{mine}}&:&\ket{B}_{b_ic_i}\mapsto\frac{\cos\phi}{\sqrt{2}}\Bigl(\ket{0}_{b_i}\ket{1}_{c_i}-\ket{1}_{b_i}\ket{0}_{c_i}\Bigr)+\frac{i\sin\phi}{\sqrt{2}}\Bigl(\ket{0}_{b_i}\ket{1}_{c_i} +\ket{1}_{b_i}\ket{0}_{c_i}\Bigr).\label{mine} \end{aligned}\] Thus the system on the horizon has one bit of information after this mining process and is thus no longer in the vacuum state. First of all, it seems implausible that such an ideal equipment can be physically realistic. Since the equipment is accelerating in order to stay outside the horizon without falling into the black hole, it has an Unruh temperature that becomes very high near the horizon. Then the equipment and the modes it interacts with, $c_i$ in this case, should strongly couple and would be expected to be approximately in a thermal state. As a consequence it seems plausible that energy must be transferred between the mining equipment and the modes $c_i$. Also, notice that the AMPSS mining argument does not take nonlocality into account. That is, the mining equipment would interact with the nonlocal gravitational degrees of freedom even if it could avoid the objection of the previous paragraph. As discussed previously, interactions with nonlocal gravitational degrees of freedom transfer part of the quantum information of the mining system into the gravitational degrees of freedom as the equipment approaches to the horizon. We can think of this transferred part as now being a part of the temporarily enlarged nonlocal gravitational degrees of freedom when the equipment is very near to the horizon. Then in this picture the mining equipment can still produce the phase change Eq. [\[mine\]](#mine){reference-type="eqref" reference="mine"} on the system just outside the horizon, but this excitation will be eventually absorbed into the nonlocal gravitational degrees of freedom. This absorption is possible regardless of how old the black hole is, because the nonlocal degrees of freedom are temporarily enlarged by the partially transferred degrees of freedom of the mining equipment. In summary, the AMPSS mining argument is not problematic for our model. # Giddings' Physical Conditions Giddings has proposed a list of physical constraints on models of black hole evaporation. We shall write each constraint in italics below and then follow that with comments on how our qubit model can satisfy the proposed constraint. (i) Evolution is unitary. Our model explicitly assumes unitary evolution. (ii) Energy is conserved. Our model is consistent with a conserved energy given by the asymptotic behavior of the gravitational field. The unitary transformation $U(\theta(r))$ during the propagation of each radiation qubit can be written in terms of a radially dependent Hamiltonian without any explicit time dependence, so there is nothing in our model that violates energy conservation. (iii) The evolution should appear innocuous to an infalling observer crossing the horizon; in this sense the horizon is preserved. We explicitly assume that the radiation modes are in their vacuum states when they are near the horizon, so there is no firewall or other drama there. (iv) Information escapes the black hole at a rate $dS/dt\sim1/R$. Although we did not discuss the temporal rates above, if one radiation qubit propagates out through some fiducial radius, such as $r = 3M$, during a time period comparable to the black hole radius $R$, since during the early radiation each qubit carries an entropy very nearly $\ln{2}$, indeed one would have $dS/dt\sim1/R$. (v) The coarse-grained features of the outgoing radiation are still well-approximated as thermal. Because of the scrambling of the black hole qubits so that each one is very nearly in a maximally mixed state, when the information is transferred from the black hole qubits to the Hawking radiation qubits, each one of these will also be very nearly in a maximally mixed state, which in the simplified toy model represents thermal radiation. Furthermore, one would expect that any collection of $n' < n/2$ qubits of the Hawking radiation also to be nearly maximally mixed, so all the coarse-grained features of the radiation would be well-approximated as thermal. (vi) Evolution of a system ${\cal H}_A\otimes{\cal H}_B$ saturates the subadditivity inequality $S_A+S_B \geq S_{AB}$. Here it is assumed that $A$ and $B$ are subsystems of $n_A$ and $n_B$ qubits respectively of the black hole gravitational field and of the Hawking radiation, not including any of the infalling and outgoing radiation qubits when they are near the horizon. Then for $n_A + n_B < n/2$, $A$, $B$, and $AB$ are all nearly maximally mixed, so $S_A \approx n_A\ln{2}$, $S_B \approx n_B\ln{2}$, and $S_{AB} \approx (n_A+n_B)\ln{2}$, thus approximately saturating the subadditivity inequality. (Of course, for any model in which the total state of $n$ qubits is pure and any collection of $n' < n/2$ qubits has nearly maximal entropy, $S \approx n'\ln{2}$, then if $n_A < n/2$, $n_B < n/2$, but $n_A + n_B > n/2$, then $S_A \approx n_A\ln{2}$ and $S_B \approx n_B\ln{2}$, but $S_{AB} \approx (n-n_A-n_B)\ln{2}$, so $S_A+S_B-S_{AB} \approx 2n_A+2n_B-n > 0$, so that the subadditivity inequality is generically not saturated in this case.) # Conclusions We have given a toy qubit model for black hole evaporation that is unitary and does not have firewalls. It does have nonlocal degrees of freedom for the black hole gravitational field, but the quantum field theory radiation modes interact purely locally with the gravitational field, so in some sense the nonlocality is confined to the gravitational sector. The model has no mining issue and also satisfies all of the constraints that Giddings has proposed, though further details would need to be added to give the detailed spectrum of Hawking radiation. The model is in many ways ad hoc, such as in the details of the qubit transfer, so one would like a more realistic interaction of the radiation modes with the gravitational field than the simple model sketched here. One would also like to extend the model to include possible ingoing radiation from outside the black hole.	{'timestamp': '2018-04-04T02:05:58', 'yymm': '1607', 'arxiv_id': '1607.04642', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04642'}
null	null
# Introduction Currents of spin angular momentum play a central role in the field of spintronics. Significant contributions have been made by spin currents, such as control of magnetizations by spin transfer torque, transmission of electric signals through insulators, thermoelectric conversion, and electric probing of insulator magnetization. In order to detect and utilize these spin based-phenomena, conversion between spin and charge currents is necessary. For realizing efficient spin-to-charge current conversion, a wide range of materials are currently being investigated, including metals, semiconductors, organic materials, carbon-based materials, and topological insulators. Finding materials suitable to the spin-to-charge conversion is thus indispensable to making spintronic devices. One of the popular methods of the spin-to-charge conversion is the inverse spin Hall effect (ISHE), which is the reciprocal effect of the spin Hall effect caused by spin-orbit interaction. In the ISHE, a spin current generates a transverse charge current in a conductor such as Pt. Since the first demonstration of the ISHE in Pt and Al, it has been extensively studied because of its versatility. Dynamical generation of spin currents can be achieved by the spin pumping (SP). At the interface between a normal conductor (N) and a ferromagnet (F), the SP causes emission of spin currents into the N layer from magnetization dynamics in the adjacent F layer. Such the magnetization dynamics is typically triggered by applying a microwave field; at the ferromagnetic resonance (FMR) or the spin wave resonance (SWR) condition, the magnetization resonantly absorbs the microwave power and exhibits a coherent precessional motion. A part of the angular momenta stored in this precessional motion is the source of the spin current generated by the SP. The combination of the ISHE and the SP enables electric detection and generation of spin currents. This is the setup commonly used to study the properties of spin-to-charge current conversion and spin transport in materials. A spin current is injected into an N layer by the SP and is converted to a measurable electromotive force by the ISHE. The conversion efficiency between spin and charge currents in this process can be determined by estimating the density of the injected spin current by analyzing microwave spectra. The spin transport property of a material can be investigated by constructing a heterostructure in which the material of interest is placed between a spin-current injector and detector layers. It should be noted that the voltage signal from the ISHE can be contaminated by other contributions in practice experiments. We thus need to extract the ISHE contribution by separating or minimizing the unwanted signals in order to ensure the validity of the measurements. The ISHE signal is characterized by Lorentzian spectral shape and sign change under the magnetization reversal, and some of the unwanted signals show the same spectral shape and the sign change in configurations commonly used in the SP experiments. Such signals can be induced by a temperature gradient via thermoelectric effects and by an rf current via rectification effects (See Fig. [\[fig:A-schematic-illustration\]](#fig:A-schematic-illustration){reference-type="ref" reference="fig:A-schematic-illustration"}). The heat emitted by excited magnetization dynamics induces a thermal gradient in a sample, resulting in an electromotive force due to the conventional Seebeck effect. This type of heat induced signals can be eliminated if one designs experimental conditions appropriately. Rectification effects comes from interplay between stray rf currents induced by an incident microwave and galvanomagnetic effects coming from oscillating magnetizations. The direction, magnitude, and phase of the stray rf current contain uncertainty because they depend on the details of an experimental setting, so the signals from the rectification effects tend to be complicated. Since the rectification effects were first observed in 1963 in a Ni film with the anisotropic magnetoresistance (AMR) and the anomalous Hall effect (AHE), we can not make light of the effects. A number of works regarding to the extraction of the ISHE signal from other contributions in electric measurements on the SP have been reported. In this article, we review the previous studies of the voltage signals induced by the SP and further introduce some methods to analyze the signals on the FMR. Here, we will focus on the experiments in a microwave cavity, so that small density of induced rf currents and thus small rectification contribution can be expected. The original field distribution in a cavity is minimally disturbed by placing a sample in it because the empty region in the cavity does not carry rf currents. In experiments, it is often difficult to identify the origin of the stray rf currents because it depends on an individual setup: the sample structure including wires for electric measurements. In this paper, we describe methods to separate the SP contribution from other artifacts in the electric measurements by introducing parameter dependence to the voltages in the presence of the stray rf currents. The methods can be applied to systems with a microstrip line or coplanar waveguide, but the rf-current-induced magnetization excitation due to the SHE, or the spin transfer torque FMR (STT-FMR), is neglected. This article is organized as follows. In Section [\[sec:signals-due-to\]](#sec:signals-due-to){reference-type="ref" reference="sec:signals-due-to"}, we show the analytical descriptions of the signals from the SP, rectification effects, and heating effects based on the magneto-circuit theory and the Landau-Lifshitz-Gilbert (LLG) equation. In Section [\[sec:Separation-methods\]](#sec:Separation-methods){reference-type="ref" reference="sec:Separation-methods"}, we discuss the dependence of the signals on the FMR spectrum, sample geometry, magnetization orientation, and excitation frequency. Here, we point out the similarity between the voltage signals due to the SP and rectification effects. Finally, in Section [\[sec:Summary\]](#sec:Summary){reference-type="ref" reference="sec:Summary"}, we give a summary of the methods based on the dependences described in Section [\[sec:Separation-methods\]](#sec:Separation-methods){reference-type="ref" reference="sec:Separation-methods"} and emphasize that the voltage measurement of the in-plane magnetization angular dependence with out-of-plane microwave magnetic field in a properly designed system is the most reliable way and thus enables quantitative studies of the SP on various kinds of materials. # Signals due to spin pumping and microwave effects in bilayer systems[\[sec:signals-due-to\]]{#sec:signals-due-to label="sec:signals-due-to"} In this section, we discuss the analysis of the voltage signals induced by the SP, rectification effects, and heating effects. For the calculation of the SP-induced ISHE signal, we consider a spin current generated by magnetization dynamics via the SP in a bilayer film consisting of a normal conductor (N) and a ferromagnet (F). The rectification signals are expressed using the derived magnetization dynamics for the SP. For the heating-induced signals, we discuss thermoelectric effects due to the heating at the FMR and SWR of magnetostatic surface spin waves. ## Spin current induced by spin pumping The spin pumping (SP) is the phenomenon that a magnetically excited F layer induces a spin current into the N layer placed adjacent to it. Here, let us suppose the N and the F layers span $xy$ plane, and they are stacked in z direction \[Fig. [\[fig:pumped-and-back\]](#fig:pumped-and-back){reference-type="ref" reference="fig:pumped-and-back"}(a)\]. The spin current density through the F/N interface due to the SP is given by \[\mathbf{j}_{{\rm s}}^{{\rm {\rm pump}}}=\frac{\hbar}{4\pi}g_{{\rm r}}^{\uparrow\downarrow}\left(\mathbf{m}\times\dot{\mathbf{m}}\right)+\frac{\hbar}{4\pi}g_{{\rm i}}^{\uparrow\downarrow}\dot{\mathbf{m}},\label{eq:sp}\] where $\mathbf{m}$ denotes the unit vector along the magnetization in the F layer, $\mathbf{\dot{m}}$ the time derivative of $\mathbf{m}$, and $g_{{\rm r}}^{\uparrow\downarrow}$ ($g_{{\rm i}}^{\uparrow\downarrow}$) the real (imaginary) part of mixing conductance per unit area, $\hbar$ the Planck constant. The spin polarization of $\mathbf{j}_{{\rm s}}^{{\rm pump}}$ is represented by its vector direction, and its flow direction is the interface normal $\boldsymbol{z}$. In diffusive N layers, the spin accumulation $\boldsymbol{\mu}_{{\rm s}}$ is formed owing to the pumped spin current. This spin accumulation induces a back-flow spin current into the F layer, and it renormalizes the mixing conductance in Eq. ([\[eq:sp\]](#eq:sp){reference-type="ref" reference="eq:sp"}) to an effective one denoted by $g_{{\rm eff}}^{\uparrow\downarrow}$. The back-flow spin current density from the magneto-circuit theory is given by \[\mathbf{j}_{{\rm s}}^{{\rm back}}=\frac{1}{4\pi}g_{{\rm r}}^{\uparrow\downarrow}\mathbf{m}\times\left(\boldsymbol{\mu}_{{\rm s}}^{{\rm F/N}}\times\mathbf{m}\right)+\frac{1}{4\pi}g_{{\rm i}}^{\uparrow\downarrow}\boldsymbol{\mu}_{{\rm s}}^{{\rm F/N}}\times\mathbf{m}\label{eq:back}\] when the spin relaxation is fast enough and $\boldsymbol{\mu}_{{\rm s}}^{{\rm F/N}}\propto\mathbf{j_{{\rm s}}^{{\rm pump}}}$ holds, where $\boldsymbol{\mu}_{{\rm s}}^{{\rm F/N}}$ denotes the spin accumulation at the F/N interface ($z=0$). The solution without the approximation can be found in Ref. [@jiao2013spin]. Then, combining Eqs. ([\[eq:sp\]](#eq:sp){reference-type="ref" reference="eq:sp"}) and ([\[eq:back\]](#eq:back){reference-type="ref" reference="eq:back"}), one can find the net dc spin current density, \[\mathbf{j}_{{\rm s}}^{{\rm {\rm F/N}}}=\mathbf{j}_{{\rm s}}^{{\rm pump}}-\mathbf{j}_{{\rm s}}^{{\rm back}}=\frac{\hbar}{4\pi}g_{{\rm r,eff}}^{\uparrow\downarrow}\mathbf{m}\times\dot{\mathbf{m}}+\frac{\hbar}{4\pi}g_{{\rm i,eff}}^{\uparrow\downarrow}\dot{\mathbf{m}},\label{eq:jsdc0}\] where $g_{{\rm r\left(i\right),eff}}^{\uparrow\downarrow}$ represents the real (imaginary) part of the effective mixing conductance per unit area. The mixing conductance at the F/N interfaces has been widely investigated in many combinations of materials. Hereafter $g_{{\rm i}{\rm ,eff}}^{\uparrow\downarrow}$ is omitted for simplicity because it is much smaller than the real part. The expression of $g_{{\rm r,eff}}^{\uparrow\downarrow}$ can be obtained in terms of the parameters of the N layer. $\boldsymbol{\mu}_{{\rm s}}^{{\rm F/N}}$ is calculated from the spin accumulation profile $\boldsymbol{\mu}_{{\rm s}}\left(z\right)$ determined by the spin diffusion equation \[\frac{\partial}{\partial t}\boldsymbol{\mu}_{{\rm s}}\left(z\right)=-\gamma_{{\rm N}}\boldsymbol{\mu}_{{\rm s}}\left(z\right)\times\mu_{0}\mathbf{H}_{0}+D\nabla^{2}\boldsymbol{\mu}_{{\rm s}}\left(z\right)-\frac{\boldsymbol{\mu}_{{\rm s}}\left(z\right)}{\tau_{{\rm s}}}\label{eq:diffeq}\] with the boundary conditions: $-\frac{\hbar\sigma_{{\rm N}}}{4e^{2}}\nabla\boldsymbol{\mu}_{{\rm s}}\left(0\right)=\mathbf{j}_{{\rm s}}^{{\rm F/N}}$ at the interface ($z=0$) and $-\frac{\hbar\sigma_{{\rm N}}}{4e^{2}}\nabla\boldsymbol{\mu}_{{\rm s}}\left(d_{{\rm N}}\right)=0$ at the outer boundary of the N layer ($z=d_{{\rm N}}$). $\sigma_{{\rm N}\left({\rm F}\right)}$ and $d_{{\rm N}\left({\rm F}\right)}$ are the conductivity and thickness of the N (F) layer. $\gamma_{{\rm N}}$ denotes the gyromagnetic ratio of electrons in the N layer, $e$ the electron charge, $\mu_{0}$ the permittivity of a vacuum, $\mathbf{H}_{0}$ an external field, $D$ the diffusion constant, and $\tau_{{\rm s}}$ the spin-relaxation time. We focus on the regime where the Hanle effect is negligibly small (a rigorous treatment can be found in Ref. [@ando2011electrically]). Then, $\boldsymbol{\mu}_{{\rm s}}\left(z\right)$ is obtained as \[\boldsymbol{\mu}_{{\rm s}}\left(z\right)=\frac{4e^{2}}{\hbar\sigma_{{\rm N}}}\lambda\frac{\cosh\left[\left(z-d_{{\rm N}}\right)/\lambda\right]}{\sinh\left(d_{{\rm N}}/\lambda\right)}\mathbf{j}_{{\rm s}}^{{\rm {\rm F/N}}},\label{eq:mus}\] where $\lambda\equiv\sqrt{D\tau_{{\rm s}}}$ denotes the spin diffusion length. Using Eqs. ([\[eq:back\]](#eq:back){reference-type="ref" reference="eq:back"}), ([\[eq:jsdc0\]](#eq:jsdc0){reference-type="ref" reference="eq:jsdc0"}), and [\[eq:mus\]](#eq:mus){reference-type="eqref" reference="eq:mus"}, we find that the effective mixing conductance is given by \[g_{{\rm r,eff}}^{\uparrow\downarrow}=\left(\frac{1}{g_{r}}+\frac{\pi\hbar\sigma_{{\rm N}}}{e^{2}\lambda}\text{tanh}\frac{d_{{\rm N}}}{\lambda}\right)^{-1}.\label{eq:geff}\] ## Magnetization dynamics and spin current Next, let us examine the effect of the SP on magnetization dynamics. We will calculate the effective mixing conductance and the magnitude of a pumped spin current in terms of observable parameters in experiments. The SP affects magnetization dynamics. When the incoming and outgoing spin currents, given by $\gamma{\rm div}\left({\rm \mathbf{j}_{s}^{{\rm F/N}}}\right)/\left(d_{{\rm F}}I_{{\rm s}}\right)$, are included, the LLG equation is modified as \[\begin{aligned} \dot{\mathbf{m}} & = &-\gamma\mathbf{m}\times\mu_{0}\left[\mathbf{H}_{{\rm eff}}\left(\mathbf{m}\right)+\mathbf{h}_{{\rm rf}}\left(t\right)\right]\nonumber \\ & & +\left(\alpha_{0}+\frac{\gamma}{d_{{\rm F}}I_{{\rm s}}}\frac{\hbar}{4\pi}g_{{\rm r,eff}}^{\uparrow\downarrow}\right)\mathbf{m}\times\dot{\mathbf{m}}\label{eq:LLG-1} \end{aligned}\] where $\gamma$ denotes the gyromagnetic ratio of the ferromagnet F, $\mathbf{H}_{{\rm eff}}\left(\mathbf{m}\right)$ the effective field, $\mathbf{h}_{{\rm rf}}\left(t\right)$ an microwave field, $\alpha_{0}$ the Gilbert damping constant without spin current exchange, and $I_{{\rm s}}$ the saturation magnetization. The effective field is given by $\mathbf{H}_{{\rm eff}}\left(\mathbf{m}\right)=-\nabla_{\mathbf{m}}F_{m}/I_{{\rm s}}$, where the magnetostatic energy $F_{m}$ includes magnetocrystalline anisotropy. The second term in the second line of Eq. ([\[eq:LLG-1\]](#eq:LLG-1){reference-type="ref" reference="eq:LLG-1"}) indicates that the SP acts as an additional damping term. The reverse process is also demonstrated; a spin current injected from the N layer reduces the damping of the F layer. Finding the enhanced damping thus directly relates to $g_{{\rm r,eff}}^{\uparrow\downarrow}$. In practical experiments, the total damping term represented by $\alpha\mathbf{m}\times\dot{\mathbf{m}}$ is observed. By measuring $\alpha$ from the FMR spectra of F/N and F films, the change in the Gilbert damping constant can be found, i.e. $\Delta\alpha=\alpha_{{\rm F/N}}-\alpha_{{\rm F}}$. Figure [\[fig:pumped-and-back\]](#fig:pumped-and-back){reference-type="ref" reference="fig:pumped-and-back"}(b) shows an example of the comparison of the FMR microwave spectrum for ${\rm Ni}_{81}{\rm Fe}_{19}$/Pt (F/N) bilayer and ${\rm Ni}_{81}{\rm Fe}_{19}$ (F) single-layer samples, where broadening of the F/N spectral peak can be seen. By fitting the spectral peak using Lorentzian, $\alpha$ is obtained from the full-width at the half-maximum (FWHM) which has the relation $\Delta H=\left(\partial\omega_{{\rm r}}/\partial H_{0}\right)^{-1}\alpha\left(\omega_{\theta}+\omega_{\phi}\right)$ for field strength $H_{0}$ swept measurements, where $\omega$ and $\omega_{{\rm r}}$ respectively denote the angular frequency of the magnetization precession and that at the resonance determined by $\omega_{\theta}$ and $\omega_{\phi}$ derived below. Finding the value of $\Delta\alpha$, one obtains \[g_{{\rm r,eff}}^{\uparrow\downarrow}=4\pi\frac{I_{{\rm s}}}{\gamma\hbar}\Delta\alpha d_{{\rm F}}.\label{eq:geffexp}\] The measurement of $\Delta\alpha$ requires some care. Since the magnetic properties of a ferromagnetic film in a heterostructure are affected by the other part of the structure, there can be contributions on $\Delta\alpha$ other than the SP through inhomogeneous broadening and two-magnon scattering. Thus careful comparison of $\alpha$ is required. In order to find the spectral shape and magnitude of the spin current $\mathbf{j}_{{\rm s}}^{{\rm F/N}}$, we calculate the magnetization dynamics $\mathbf{m}\times\dot{\mathbf{m}}$ from Eq. ([\[eq:LLG-1\]](#eq:LLG-1){reference-type="ref" reference="eq:LLG-1"}). Here, we consider precessing magnetization $\mathbf{m}_{{\rm rf}}\equiv{\rm Re}\left[\left(m_{\theta}\mathbf{e}_{\theta}+m_{\phi}\mathbf{e}_{\phi}\right)\exp\left(i\omega t\right)\right]$ around the equilibrium magnetization vector $\mathbf{m}_{0}$ excited by an microwave field $\mathbf{h}_{{\rm rf}}\equiv{\rm Re}\left[\left(h_{\theta}\mathbf{e}_{\theta}+h_{\phi}\mathbf{e}_{\phi}\right)\exp\left(i\omega t\right)\right]$ as shown in Fig. [\[fig:(a)-Spatial-coordinate\]](#fig:(a)-Spatial-coordinate){reference-type="ref" reference="fig:(a)-Spatial-coordinate"}. $\mathbf{m}_{0}$ is determined by the condition $\mathbf{m}_{0}\times\mathbf{H}_{{\rm eff}}=0$. The unit vectors of the polar coordinate $\mathbf{e}_{p}\left(p=r,\theta,\phi\right)$ in which $\mathbf{e}_{r}$ points $\mathbf{m}_{0}$ have a relation to the unit vectors of the Cartesian coordinate $\mathbf{e}_{i}\left(i=x,y,z\right)$, $\mathbf{e}_{p}=\sum_{i=x,y,z}u_{pi}\mathbf{e}_{i}$ with \[u=\left(\begin{array}{ccc} \sin\theta_{{\rm M}}\cos\phi_{{\rm M}} & \sin\theta_{{\rm M}}\sin\phi_{{\rm M}} & \cos\theta_{{\rm M}}\\ \cos\theta_{{\rm M}}\cos\phi_{{\rm M}} & \cos\theta_{{\rm M}}\sin\phi_{{\rm M}} &-\sin\theta_{{\rm M}}\\-\sin\phi_{{\rm M}} & \cos\phi_{{\rm M}} & 0 \end{array}\right),\label{eq:u}\] where $\theta_{{\rm M}}$ and $\phi_{{\rm M}}$ denote the polar angle between the $z$ axis and $\mathbf{m}_{0}$, and the azimuthal angle measured from the $x$ axis, respectively. Taking the time average of the pumped spin current in Eq. [\[eq:jsdc0\]](#eq:jsdc0){reference-type="eqref" reference="eq:jsdc0"} yields \[\left\langle \mathbf{j}_{{\rm s}}^{{\rm F/N}}\right\rangle _{t}=\frac{\hbar\omega}{4\pi}g_{r,{\rm eff}}^{\uparrow\downarrow}{\rm Im}\left[m_{\theta}m_{\phi}^{}\right]\mathbf{m}_{0},\label{eq:spin pump}\] where $\left\langle \cdots\right\rangle _{t}$ and $a^{}$ mean temporal average and complex conjugate of $a$, respectively. The relation between the magnitude of $\mathbf{m}_{{\rm rf}}$ and $\mathbf{h}_{{\rm rf}}$ obtained from the Eq. [\[eq:LLG-1\]](#eq:LLG-1){reference-type="eqref" reference="eq:LLG-1"} is reduced to $(m_{\theta},m_{\phi})=\boldsymbol{\chi}.(h_{\theta},h_{\phi})$, where the susceptibility $\boldsymbol{\chi}$ is given by \[\boldsymbol{\chi}\approx\frac{\gamma\mu_{0}}{\alpha\omega}\frac{S\left(\omega,\omega_{{\rm r}}\right)}{\omega_{\theta}+\omega_{\phi}}\left(\begin{array}{cc} \omega_{\theta} &-\omega_{\theta\phi}+i\omega\\-\omega_{\theta\phi}-i\omega & \omega_{\phi} \end{array}\right)\label{eq:chi-1}\] with $\omega_{\theta}\equiv\omega_{\phi\phi}-\frac{\partial}{\partial m_{r}}F_{m}$, $\omega_{\phi}\equiv\omega_{\theta\theta}-\frac{\partial}{\partial m_{r}}F_{m}$, and \[\omega_{pq}\equiv\frac{\gamma\mu_{0}}{I_{{\rm s}}}\frac{\partial}{\partial m}_{p}\frac{\partial}{\partial m}_{q}F_{m}\ \left(p,q\in\{\theta,\phi\}\right).\label{eq:omegath}\] $S\left(\omega,\omega_{{\rm r}}\right)$ represents a spectrum function, \[S\left(\omega,\omega_{{\rm r}}\right)\equiv\frac{\alpha\omega}{\left(\omega_{{\rm r}}-\omega\right)^{2}+\left(\alpha\omega\right)^{2}}\left[\left(\omega_{{\rm r}}-\omega\right)-i\alpha\omega\right],\] which real part represents asymmetric spectrum, ${\rm Asym\left(\omega,\omega_{{\rm r}}\right)\equiv{\rm Re}}\left[S\left(\omega,\omega_{{\rm r}}\right)\right]$ (known as asymmetric Lorentzian), and which imaginary part represents symmetric spectrum, ${\rm Lor}\left(\omega,\omega_{{\rm r}}\right)\equiv{\rm Im}\left[S\left(\omega,\omega_{{\rm r}}\right)\right]$ (Lorentzian). The resonance angular frequency $\omega_{{\rm r}}$ is given by \[\omega_{{\rm r}}\equiv\sqrt{\omega_{\theta}\omega_{\phi}-\omega_{\theta\phi}^{2}}.\label{eq:omegares}\] Calculation using Eq. ([\[eq:chi-1\]](#eq:chi-1){reference-type="ref" reference="eq:chi-1"}) yields \[\left\langle \mathbf{j}_{{\rm s}}^{{\rm F/N}}\right\rangle _{t}=\frac{\hbar g_{r,{\rm eff}}^{\uparrow\downarrow}}{4\pi}\frac{\gamma^{2}\mu_{0}^{2}\left(\omega_{\theta}\left\|h_{\theta}\right\|^{2}+\omega_{\phi}\left\|h_{\phi}\right\|^{2}+{\rm Im}\left[h_{\theta}h_{\phi}^{}\left\{ \omega_{\theta}\omega_{\phi}-\left(\omega_{\theta\phi}+i\omega\right)^{2}\right\} /\omega\right]\right)}{\alpha^{2}\left(\omega_{\theta}+\omega_{\phi}\right)^{2}}{\rm Lor}\left(\omega,\omega_{{\rm r}}\right)\mathbf{m}_{0},\label{eq:jsdc}\] where we used a relation, $\left\|S\left(\omega,\omega_{{\rm r}}\right)\right\|^{2}={\rm Lor}\left(\omega,\omega_{{\rm r}}\right)$. Equation [\[eq:jsdc\]](#eq:jsdc){reference-type="eqref" reference="eq:jsdc"} means that the spectral shape of $\left\langle \mathbf{j}_{{\rm s}}^{{\rm F/N}}\right\rangle _{t}$ is Lorentzian. Equation [\[eq:jsdc\]](#eq:jsdc){reference-type="eqref" reference="eq:jsdc"} can be also expected in terms of the elliptic precession with the cone angle $\Theta$ in Refs. [@mosendz2010quantifying; @Ando:2009gh; @czeschka2011scaling]: at resonance $\mathbf{m}_{{\rm rf}}={\rm Re}\left[\sin\Theta\exp\left(i\omega t\right)\mathbf{e}_{\theta}+iA\sin\Theta\exp\left(i\omega t\right)\mathbf{e}_{\phi}\right]$ and thus ${\rm Im}\left[m_{\theta}m_{\phi}^{}\right]=A\sin^{2}\Theta$, where $A$ is an correction factor for the elliptical precession motion. The magnetization excitation leads to decreased power of the incident microwave and thus microwave measurements are useful for determining $\alpha$ and other parameters related to the resonance condition. The microwave power absorption on the FMR per unit volume is calculated by \[\begin{aligned} \Delta P & = &-\frac{\omega}{2\pi}\int_{0}^{\frac{2\pi}{\omega}}\mathbf{h}\cdot\dot{\mathbf{m}}dt\nonumber \\ & = & \frac{\gamma\mu_{0}}{2}\frac{\left(\omega_{\theta}\left\|h_{\theta}\right\|^{2}+\omega_{\phi}\left\|h_{\phi}\right\|^{2}-2{\rm Re}\left[h_{\theta}h_{\phi}^{}\left(\omega_{\theta\phi}+i\omega\right)\right]\right)}{\alpha\left(\omega_{\theta}+\omega_{\phi}\right)}{\rm Lor}\left(\omega,\omega_{{\rm r}}\right).\label{eq:delP} \end{aligned}\] The transferred energy from the microwave to the magnetization dynamics finally results in heat via damping processes of the dynamics, which causes thermoelectric signals. In this calculation, we assumed that $\mathbf{h}_{{\rm rf}}$ is only induced by an applied microwave field and is not affected by an induced rf current. Rf currents are known to trigger the FMR through generation of rf spin currents via the SHE. This contribution can be included to $\mathbf{h}_{{\rm rf}}$ by calculating the spin-transfer torque due to the absorption of the rf spin current. Such a contribution will result in a phase shift between the actual rf current and the rf current determined by the analysis provided in this article. ## Voltage generated by inverse spin Hall effect Here, we describe the electromotive force generated by the SP and ISHE in the N layer by taking account of the spin current profile in the N layer. The pumped spin current in the N layer gives rise to an electromotive force due to the ISHE. The ISHE induces a charge current density transverse to both the spin polarization ($\propto\mathbf{j}_{{\rm s}}$) and its flow direction $\left(\propto\mathbf{z}\right)$, which can be expressed as \[\mathbf{j}_{{\rm c,ISHE}}=\frac{2e}{\hbar}\theta_{{\rm SHE}}\mathbf{j}_{{\rm s}}\times\mathbf{z}.\] Therefore, as $\left\langle \mathbf{j}_{{\rm s}}^{{\rm F/N}}\right\rangle _{t}\propto\mathbf{m}_{0}$ holds in the SP, the direction of the ISHE current is reversed under the magnetization reversal ($\mathbf{m}_{{\rm 0}}\rightarrow-\mathbf{m}_{{\rm 0}}$), which is an important feature of the ISHE induced by the SP. According to the short circuit model, the electromotive force is calculated by the sum of the induced current. The total dc current induced by the ISHE is given by \[\mathbf{J}_{{\rm ISHE}}=w\frac{2e}{\hbar}\theta_{{\rm SHE}}\int_{0}^{d_{{\rm N}}}\mathbf{z}\times\left\langle \mathbf{j}_{{\rm s}}\left(z\right)\right\rangle _{t}dz\label{eq:JISHE}\] with $w$ being the width of a sample \[See Fig. [\[fig:Typical-configuration-of\]](#fig:Typical-configuration-of){reference-type="ref" reference="fig:Typical-configuration-of"}(a)\]. One can observe the electromotive force $\mathbf{E}_{{\rm ISHE}}$ induced by $\mathbf{J}_{{\rm ISHE}}$, which satisfies the relation $\tilde{R}_{{\rm tot}}\mathbf{J}_{{\rm ISHE}}+\mathbf{E}_{{\rm {\rm ISHE}}}=0$ for an open circuit condition, where $\tilde{R}_{{\rm tot}}$ is the total resistance per unit length of the system, e.g. $\tilde{R}_{{\rm tot}}^{-1}=w\left(\sigma_{{\rm N}}d_{{\rm N}}+\sigma_{{\rm F}}d_{{\rm F}}\right)$ for F/N bilayer systems. The spin current profile is obtained from Eqs. ([\[eq:jsdc0\]](#eq:jsdc0){reference-type="ref" reference="eq:jsdc0"}) and ([\[eq:mus\]](#eq:mus){reference-type="ref" reference="eq:mus"}) as \[\begin{aligned} \mathbf{j}_{{\rm s}}\left(z\right) & = &-\frac{\hbar}{2e}\frac{\sigma_{{\rm N}}}{2}\nabla\boldsymbol{\mu}_{{\rm s}}\left(z\right)\\ & = & \mathbf{j}_{{\rm s}}^{{\rm {\rm F/N}}}\frac{\sinh\left(\left[d_{{\rm N}}-z\right]/\lambda\right)}{\sinh\left(d_{{\rm N}}/\lambda\right)}.\label{eq:jsz} \end{aligned}\] Then we yields \[\mathbf{E}_{{\rm ISHE}}=w\tilde{R}_{{\rm tot}}\theta_{{\rm SHE}}\frac{2e}{\hbar}\lambda\tanh\frac{d_{N}}{2\lambda}\left\langle j_{{\rm s}}^{{\rm {\rm F/N}}}\right\rangle _{t}\mathbf{z}\times\mathbf{m}_{0}\label{eq:EISHE}\] from Eqs. ([\[eq:jsdc0\]](#eq:jsdc0){reference-type="ref" reference="eq:jsdc0"}), ([\[eq:JISHE\]](#eq:JISHE){reference-type="ref" reference="eq:JISHE"}), and ([\[eq:jsz\]](#eq:jsz){reference-type="ref" reference="eq:jsz"}). A typical experimental setup for observing the ISHE voltage is depicted in Fig. [\[fig:Typical-configuration-of\]](#fig:Typical-configuration-of){reference-type="ref" reference="fig:Typical-configuration-of"}(b). A bilayer sample is placed on a cutout of a quartz rod, which has a through hole in the center for wires connected to a voltmeter. The wires are covered by an insulating polymer and connected to the electrodes in the $y$ axis. These wires measure the $y$ component of Eq. [\[eq:EISHE\]](#eq:EISHE){reference-type="eqref" reference="eq:EISHE"}, so that the maximized ISHE voltage is detected when the magnetization points along the $x$ axis. $\mathbf{m}_{0}$ is controlled by an applied field, $\mathbf{H}_{0}$. During the measurements, while $\omega$ is fixed, the field strength $H_{0}$ is swept so that $\omega_{{\rm r}}$ is changed. ## Voltage generated by rectification effects Rectification effects at the FMR are caused by an rf current in a sample possessing a galvanomagnetic effect, i.e. magnetization dependent resistivity $\rho\left(\mathbf{m}\right)$. Describing the resistivity which oscillates due to the precessing magnetization as $\tilde{\rho}\left(\mathbf{m}\right)\propto m_{p}\cos\left(\omega t\right)$ and an rf current as $j_{{\rm rf}}=j_{{\rm rf}}^{0}\cos\left(\omega t+\psi\right)$, a rectified dc electromotive force, $E_{{\rm rect}}$, is given by \[\begin{aligned} E_{{\rm rect}} & \propto & m_{p}\cos\left(\omega t\right)\cdot j_{{\rm rf}}^{0}\cos\left(\omega t+\psi\right)\nonumber \\ & \propto & \frac{jm_{p}}{2}\left[\cos\psi+O\left(t\right)\right], \end{aligned}\] where $\psi$ denotes the phase difference between the precessing magnetization and the rf current, as depicted in Fig. [\[fig:possible-origins-of\]](#fig:possible-origins-of){reference-type="ref" reference="fig:possible-origins-of"}(a). Here, we will consider the anisotropic magnetoresistance (AMR), the anomalous Hall effect (AHE), and the spin Hall magnetoresistance (SMR), which are common examples of galvanomagnetic effects in the bilayer systems used in the SP measurements. Note that any effect leading to magnetization dependent resistance gives a dc voltage signal. Therefore the tunnel magnetoresistance, the colossal magnetoresistance, and the spin accumulation Hall effect can give the rectified dc voltages which can be calculated in the way described below. The induced voltage is calculated from the Ohm's law $E_{i}=\rho_{ij}(\mathbf{m})j_{{\rm rf}}^{j}$, where $j_{{\rm rf}}^{i}$ denotes the $i$-th component of an rf current passing through a sample. For an F layer with the AMR and AHE, the resistivity tensor can be expanded in terms of the magnetization by \[\rho_{ij}^{{\rm F}}\left(\mathbf{m}\right)=\rho_{ij}^{0}+\Delta\rho_{{\rm AMR}}m_{i}m_{j}+\rho_{{\rm AHE}}\sum_{l=x,y,z}\epsilon_{ijl}m_{l},\label{eq:restensor}\] where $m_{i}$ denotes the component in the $i$ axis and $\epsilon$ denotes the Levi-Civita tensor. $\rho_{ij}^{0}$ denotes the resistivity part insensitive to $\mathbf{m}$, $\Delta\rho_{{\rm AMR}}$ the AMR coefficient, and $\rho_{{\rm AHE}}$ the AHE coefficient. For an N layer with the SMR, whose magnitude is $\Delta\rho_{{\rm SMR}}$, the resistivity tensor is given by \[\rho_{ij}^{{\rm N}}\left(\mathbf{m}\right)=\rho_{ij}^{0}-\Delta\rho_{{\rm SMR}}\sum_{k=x,y}\epsilon_{ikz}m_{k}\sum_{l=x,y}\epsilon_{jlz}m_{l}+\rho_{{\rm AHE}}^{{\rm N}}\epsilon_{ijz}m_{z}.\label{eq:restensor-1}\] Substituting $m_{i}=\sum_{p=\theta,\phi}u_{pi}m_{p}$ with $u_{pi}$ in Eq. ([\[eq:u\]](#eq:u){reference-type="ref" reference="eq:u"}) and extracting the components proportional to precessing magnetization, $m_{\theta}$ and $m_{\phi}$, yields the time-dependent resistivity tensor $\tilde{\rho}^{{\rm F}\left({\rm N}\right)}$ for the F (N) layer \[cf. Ref. [@iguchi2013effect]\]. Then the $i$-th component of the rectified dc current density from the $j$-th component of the rf current is given by $\sigma_{{\rm F(N)}}{\rm Re}\left[\tilde{\rho}_{ij}^{{\rm F(N)}}j_{{\rm rf}}^{j}\right]/2$ for the F(N) layer. Considering the short circuit model, the electromotive force $\mathbf{E}_{{\rm rect}}$ is obtained as \[\mathbf{E}_{{\rm rect}}=\tilde{R}_{{\rm tot}}\left(\int_{-d_{{\rm F}}}^{0}\frac{\sigma_{{\rm F}}}{2}{\rm Re}\left[\tilde{\rho}^{{\rm F}}\cdot\mathbf{j}_{{\rm rf}}^{}\left(z\right)\right]dz+\int_{0}^{d_{{\rm N}}}\frac{\sigma_{{\rm N}}}{2}{\rm Re}\left[\tilde{\rho}^{{\rm N}}\cdot\mathbf{j}_{{\rm rf}}^{}\left(z\right)\right]dz\right).\label{eq:Erect}\] An origin of $z$ dependence of $\mathbf{j}_{{\rm rf}}\left(z\right)$ comes from the skin effect (which appears as the Dyson effect for the microwave resonance experiments). $\mathbf{j}_{{\rm rf}}\left(z\right)$ is localized within the skin depth: $\delta_{{\rm skin}}=2\left(\sigma\mu\omega\right)^{-1/2}$, where $\mu$ denotes the permeability. $\delta_{{\rm skin}}$ is typically smaller than the thickness of systems used for the SP experiments; for example, $\delta_{{\rm skin}}$ for Cu at $\omega/2\pi$=10 GHz is estimated to be 0.6 $\mu$m, and the thickness scale used in the experiments is less than a few hundred nanometers. The in-plane electromotive force induced by an rf current which lies in the film plane ($xy$ plane) and homogeneous over $z$ is \[\begin{aligned} \left(\begin{array}{c} E_{{\rm rect}}^{x}\\ E_{{\rm rect}}^{y} \end{array}\right)= & \frac{\tilde{R}_{{\rm tot}}}{2}\sin\theta_{{\rm M}}\sum_{L={\rm F},{\rm N}}\sigma_{L}d_{L}\times\nonumber \\ & \left\{ \left[\cos\theta_{{\rm M}}\rho_{L}^{\theta}+\rho_{{\rm AHE}}^{L}\left(\begin{array}{cc} 0 & 1\\-1 & 0 \end{array}\right)\right]{\rm Re}\left[m_{\theta}\left(\begin{array}{c} j_{{\rm rf}}^{x}\\ j_{{\rm rf}}^{y} \end{array}\right)\right]+\rho_{L}^{\phi}{\rm Re}\left[m_{\phi}\left(\begin{array}{c} j_{{\rm rf}}^{x}\\ j_{{\rm rf}}^{y} \end{array}\right)\right]\right\} \label{eq:Erectex} \end{aligned}\] with \[\rho_{{\rm F(N)}}^{\theta}=\Delta\rho_{{\rm AMR(SMR)}}\left(\begin{array}{cc} +(-)1+\cos\left(2\phi_{{\rm M}}\right) & \sin\left(2\phi_{{\rm M}}\right)\\ \sin\left(2\phi_{{\rm M}}\right) & +(-)1-\cos\left(2\phi_{{\rm M}}\right) \end{array}\right),\] \[\rho_{{\rm F(N)}}^{\phi}=\left(\begin{array}{cc}-\sin\left(2\phi_{{\rm M}}\right) & \cos\left(2\phi_{{\rm M}}\right)\\ \cos\left(2\phi_{{\rm M}}\right) & \sin\left(2\phi_{{\rm M}}\right) \end{array}\right),\] where + (-) sign on the first term of the diagonal parts corresponds to the F (N) layer. The magnitude and spectral shape are determined by ${\rm Re}\left[m_{p}j_{{\rm rf}}^{i{\rm }}\right]$ through Eq. ([\[eq:chi-1\]](#eq:chi-1){reference-type="ref" reference="eq:chi-1"}). Figure [\[fig:(c)-the-spectral\]](#fig:(c)-the-spectral){reference-type="ref" reference="fig:(c)-the-spectral"} shows possibly induced spectra with various $j_{{\rm rf}}^{}$ direction in response to the dynamic magnetizations, which exhibits both the Lorentzian and asymmetric Lorentzian. The rectification signals show linear dependence to the incident power because $m_{p}\propto h_{{\rm rf}}$ and $j_{{\rm rf}}\propto h_{{\rm rf}}$, which is same as the ISHE signal since $\left\langle j_{{\rm s}}^{{\rm F/N}}\right\rangle _{t}\propto{\rm Im}\left[m_{\theta}m_{\phi}^{}\right]\propto h_{{\rm rf}}^{2}$. In the following, we discuss the origin of the stray rf currents causing the extrinsic signals in the previous experiments. One possible origin of the rf current in a cavity is the generation due to the Ohm's law because of a non-zero rf electric field, $\mathbf{E}_{{\rm rf}}$, which is considered in most of the previous studies. However, in an open circuit condition, a conductor smaller than the microwave wavelength, typically $\sim1$ cm for the X-band microwaves, can screen $\mathbf{E}_{{\rm rf}}$ below the plasma frequency, which results in a displacement current $\mathbf{j_{{\rm dis}}}\approx i\varepsilon\omega\mathbf{E}_{{\rm rf}}$ rather than the rf current $\mathbf{j}_{{\rm rf}}=\sigma\mathbf{E}_{{\rm rf}}$ induced by the Ohm's law \[See Fig. [\[fig:possible-origins-of\]](#fig:possible-origins-of){reference-type="ref" reference="fig:possible-origins-of"}(b)\]. Here $\varepsilon$ is the permittivity. A simulation in Ref. [@peligrad1998cavity] shows that the original cavity field distribution is modified so that $\mathbf{E}_{{\rm rf}}$ in a conductor is zero. $\mathbf{j}_{{\rm dis}}$ is usually much smaller than the $\mathbf{j}_{{\rm rf}}$, e.g. by a factor of $10^{8}$ for Cu at $\omega/2\pi=10$ GHz, and the phase of $\mathbf{j}_{{\rm dis}}$ is $90^{\circ}$ different from the $\mathbf{j}_{{\rm rf}}$, which indicates other consideration is necessary to explain previous results. The Faraday's law causes an rf current from a microwave magnetic field. An in-plane microwave magnetic field induces an eddy current in the cross section of the bilayer systems \[See Fig. [\[fig:possible-origins-of\]](#fig:possible-origins-of){reference-type="ref" reference="fig:possible-origins-of"}(c)\]. Though the current directions are opposite in the F and N layers, the microwave current can induce a non-zero rectification signal due to the difference between $\tilde{\rho}^{{\rm F}}\left(\mathbf{m}\right)$ and $\tilde{\rho}^{{\rm N}}\left(\mathbf{m}\right)$. This is one of plausible contributions to the experiments. For out-of-plane microwave field, the voltage appears only when the electrodes are placed off center from the rf-eddy-current distribution \[See Fig. [\[fig:possible-origins-of\]](#fig:possible-origins-of){reference-type="ref" reference="fig:possible-origins-of"}(d)\]. Such an effect was studied in Ref. [@EuO], which remarks that this effect can be eliminated by making the sample structure symmetric. Consideration on wires connected to samples for the ISHE measurements leads to two additional contributions by the Ohm's law and Faraday's law, which are shown in Fig. [\[fig:possible-origins-of\]](#fig:possible-origins-of){reference-type="ref" reference="fig:possible-origins-of"}(e,f). In one case, $\mathbf{j}_{{\rm rf}}$ can be proportional to $\sigma\mathbf{E}_{{\rm rf}}$ with a factor considering the modification of the field due to a conducting sample. $\mathbf{j}_{{\rm rf}}$ is generated and is transmitted through the paired or twisted wires forming an microwave transmission line. In the other case, $\mathbf{j}_{{\rm rf}}$ is generated by the induction around a sample forming a pick-up coil with wires. In both the cases, rf currents only appear along the electrode direction. As we discussed above, there are several mechanisms for the stray rf-current generation in cavities. The suppression of the rf current is not straightforward and thus the analysis based on $\mathbf{j}_{{\rm rf}}$ by leaving its magnitude, direction and phase undefined parameters, which are to be fitted with experimental data, is appropriate. The consideration of in-plane rf currents is enough for the analysis because the aforementioned mechanisms do not induce an rf current along the $z$ direction unless a pathway for the rf current is formed in the $z$ direction. This also holds for rf currents originating from the magnonic charge pumping, the ac ISHE current due to the ac SP, and other generation effects. ## Voltage generated by heating[\[subsec:Induced-voltage-by\]]{#subsec:Induced-voltage-by label="subsec:Induced-voltage-by"} Damping processes of magnetization dynamics generate heat, which can result in an electromotive force via thermoelectric effects. Thermoelectric effects, such as the Seebeck and Nernst effects, are seen in conducting materials by themselves regardless of the presence of the F layer. When a magnetization emits the absorbed power $\Delta P$ to phonons, a temperature gradient, $\nabla T\propto\Delta P\propto{\rm Lor}\left(\omega,\omega_{{\rm r}}\right)$, can be induced. $\nabla T$ is then formed into an electromotive force, e.g. $E_{{\rm SE}}\propto\nabla T$ via the Seebeck effect. While the spectral shape of $E_{{\rm SE}}$ on the FMR is Lorentzian, it becomes complicated on the SWR in thick ferromagnetic films, which is also used to drive the SP. The SWR gives a different heat profile on each resonance, so that the sign of $\nabla T$ easily changes. When the neighboring peaks have different signs, the signal should look like the asymmetric Lorentzian, as shown in Fig. [\[fig:Electromotive-force-due\]](#fig:Electromotive-force-due){reference-type="ref" reference="fig:Electromotive-force-due"}(b). As a result, the total thermoelectric voltage becomes the superposition of the symmetric and asymmetric Lorentzian for the SWR case. The Seebeck effect contributes to the voltage signal when the heat profile produced by the FMR or SWR is not symmetrically distributed with respect to the electrodes for the voltage detection.In most systems, it does not have explicit dependence on the magnetization direction. However, there can be a case where $\nabla T$ is sensitive to the magnetization direction as is seen in a system comprised of a thick ferromagnetic film. In such films, spin waves known as magnetostatic surface spin waves (MSSWs) localize on the top and bottom surfaces of the F layer and propagate non-reciprocally. MSSWs are demonstrated to convey heat to an arbitrary direction controlled by the magnetization polarization utilizing the non-reciprocity by unbalanced excitation of spin waves with wavevectors $+\mathbf{k}$ and $-\mathbf{k}$. This results in $\nabla T\propto\mathbf{k}$ and thus induces a Seebeck voltage in the N layer, $\mathbf{E}_{{\rm SE}\left({\rm MSSW}\right)}\propto\Delta P\mathbf{k}$. In microstrip antenna excitation, $\mathbf{k}$ of the dominant MSSWs is reversed under the magnetization reversal, $\mathbf{m}_{0}\rightarrow-\mathbf{m}_{0}$. This is because $\mathbf{k}$ of the MSSWs localized at a surface with its normal $\mathbf{n}$ is determined by $\mathbf{m}_{0}\times\mathbf{n}$. Consequently, the sign of the induced heat current by the MSSW is reversed, resulting in a thermoelectric signal with the same symmetry as the ISHE, i.e. \[\mathbf{E}_{{\rm SE}\left({\rm MSSW}\right)}=A_{{\rm SE\left(MSSW\right)}}\Delta P\mathbf{z}\times\mathbf{m}_{0},\label{eq:Ese}\] where $A_{{\rm SE\left(MSSW\right)}}$ is a constant determined by the Seebeck coefficient and the temperature profile due to the MSSW. In the cavity experiments on the SWR, though the spin waves with $+\mathbf{k}$ and $-\mathbf{k}$ are equally excited, asymmetries between the surfaces that these two modes are localized on can give the thermoelectric voltage in the same form as Eq. ([\[eq:Ese\]](#eq:Ese){reference-type="ref" reference="eq:Ese"}); as we place a N layer on top of the F layer, the inversion symmetry between the top and bottom layers is broken, so that the contributions from $+\mathbf{k}$ and $-\mathbf{k}$ give unequal contributions to $\nabla T$ in the N layer, which results in $\nabla T\propto\mathbf{k}$. Moreover, the existence of the substrate at the bottom surface may promote the asymmetry of the thermal conduction, possibly growing $\nabla T$ \[See Fig. [\[fig:Electromotive-force-due\]](#fig:Electromotive-force-due){reference-type="ref" reference="fig:Electromotive-force-due"}(c)\]. Thus, the MSSW heating effect can appear regardless of the excitation methods in the F/N bilayer systems. This effect can be significant in materials with high thermoelectric conversion efficiency, such as low carrier density conductors. Other contributions come from the transverse thermoelectric effects reflecting field or magnetization direction, such as the Nernst-Ettingshausen, anomalous Nernst effect (ANE) and spin Seebeck effect (SSE). Neglecting the angular difference between an applied field and $\mathbf{m}_{0}$, the induced voltage is proportional to $\mathbf{m}_{0}\times\nabla T$. Thus when $\nabla T$ is formed in the thickness direction, it gives an in-plane electromotive force \[\mathbf{E}_{{\rm TTE}}=A_{{\rm TTE}}\Delta P\mathbf{z}\times\mathbf{m}_{0},\] where $A_{{\rm TTE}}$ denotes a proportionality constant determined by the magnitude of the transverse thermoelectric effects and $\nabla T$ along $z$ direction. Importantly, $\mathbf{E}_{{\rm TTE}}$ shows the same symmetry as $\mathbf{E}_{{\rm ISHE}}$. # Separation methods of SP-induced ISHE signal[\[sec:Separation-methods\]]{#sec:Separation-methods label="sec:Separation-methods"} In this section, we will introduce a guideline to select proper materials for the F layer and four methods to extract the spin current contribution from observed signals. Here, we discuss the microwave contribution to the voltage signals in terms of spectral shape, thickness, magnetization angle, and excitation frequency dependences. Understanding these dependences, the rectification effects can be isolated by a measurement of magnetization angular dependence, and the heating effects can be isolated by that of frequency dependence. ## Suitable sample design for measurements of ISHE driven by SP[\[subsec:Suitable-materials-for\]]{#subsec:Suitable-materials-for label="subsec:Suitable-materials-for"} For electric measurements of the SP, an appropriate choice of materials for reducing the rectification and heating effects can improve the performance of the experiments. The first step of experiments of the ISHE induced by the SP thus is to consider the right choice of materials for the spin injector. The rectification effects can be suppressed by using a material with low galvanomagnetic coefficients. The coefficients represented by $\Delta\rho_{{\rm AMR}}$, $\rho_{{\rm AHE}}^{{\rm F}}$, $\Delta\rho_{{\rm SMR}}$, and $\rho_{{\rm AHE}}^{{\rm N}}$ in Eq. ([\[eq:restensor\]](#eq:restensor){reference-type="ref" reference="eq:restensor"}) are proportional to the signals. For the metallic spin injector, ${\rm Ni}_{81}{\rm Fe}_{19}$, so called Permalloy, is often used but other materials such as CoFe alloys with the low AMR ratio is a good candidate for the SP. Similarly, the SMR is known to be small compared to the AMR, and thus the use of a ferrimagnetic insulator is effective. The FMR and MSSW heating effects due to the conventional Seebeck effect can be minimized by making the sample structure symmetric about the electrodes and by reducing the thickness of the ferromagnetic layer $d_{{\rm F}}$, respectively. The feature due to magnetostatic interaction is dominant around $\left\|\mathbf{k}\right\|d_{{\rm F}}\sim1$. When $d_{{\rm F}}$ is decreased, such a value of $\mathbf{k}$ increases and eventually reaches the exchange regime where the magnetostatic feature is lost. Figure [\[fig:Electromotive-force-due\]](#fig:Electromotive-force-due){reference-type="ref" reference="fig:Electromotive-force-due"} illustrates the dispersion relation of spin waves for thick and thin ferromagnetic films. The manifold of the dispersion shrinks as the film thickness reduces. The group velocity of the MSSWs correspondingly becomes smaller, and the heat conveyer effect eventually disappears. Depending on the strength of the Seebeck effect, the appropriate thickness is below $100$ nm for the measurements free from the MSSWs heating, which can be fabricated by pulsed laser deposition, sputtering, or metal-organic decomposition. The MSSW contribution can also be confirmed by a control experiment with the insertion of a thin nonmagnetic insulator layer between the N and F layers because the nonmagnetic insulator cuts the spin transport but allows heat transport. ## Spectral shape dependence First, let us introduce a way to separate a measured electric signal into symmetric and antisymmetric parts with respect to reversal of magnetization. We will explain why this simple method does not work for isolating the microwave effects. In addition to the spectral shape separation introduced here, measurements on the other dependences are strongly recommended. As it is derived in Eqs. ([\[eq:EISHE\]](#eq:EISHE){reference-type="ref" reference="eq:EISHE"}) and ([\[eq:Erect\]](#eq:Erect){reference-type="ref" reference="eq:Erect"}), the dc electromotive force on the FMR $\mathbf{E}_{{\rm tot}}=\mathbf{E}_{{\rm ISHE}}+\mathbf{E}_{{\rm rect}}+\mathbf{E}_{{\rm SE}{\rm \left(MSSW\right)}}+\mathbf{E}_{{\rm TTE}}$ has the two distinct parts proportional to ${\rm Lor}\left(\omega,\omega_{r}\right)$ and ${\rm Asym}\left(\omega,\omega_{r}\right)$. By fitting an observed signal by \[\mathbf{E}_{{\rm tot}}=\mathbf{E}_{{\rm sym}}{\rm Lor}\left(\omega,\omega_{{\rm r}}\right)+\mathbf{E}_{{\rm asym}}{\rm Asym}\left(\omega,\omega_{{\rm r}}\right),\] the separation can be done, where $E_{{\rm sym}\left({\rm asym}\right)}$ is the magnitude of the ${\rm Lor}\left(\omega,\omega_{{\rm r}}\right)$ (${\rm Asym}\left(\omega,\omega_{{\rm r}}\right)$) part (See Fig. [\[fig:Analysis-of-obtained\]](#fig:Analysis-of-obtained){reference-type="ref" reference="fig:Analysis-of-obtained"}). The earlier naive discussions attribute the whole $E_{{\rm sym}}$ due to the ISHE, but this assumption does not hold as is discussed below. Generally, $E_{{\rm sym}}$ includes not only the ISHE component but also the rectification contribution. For example, in a ${\rm TE}_{011}$ cavity with the rf field $h_{{\rm rf}}$ in the $y$ direction and an rf current along the $x$ direction with $j_{{\rm rf}}\propto ih_{{\rm rf}}$, the $y$-component of Eq. [\[eq:Erectex\]](#eq:Erectex){reference-type="eqref" reference="eq:Erectex"} leads to \[\begin{aligned} E_{{\rm sym},{\rm ISHE}}^{y}\propto & \frac{\omega_{\phi}}{\left(\omega_{\theta}+\omega_{\phi}\right)^{2}}\sin\theta_{{\rm M}},\nonumber \\ E_{{\rm sym},{\rm rect}}^{y}\propto & \frac{\left(\rho_{{\rm AHE}}\omega_{\theta\phi}-\Delta\rho_{{\rm AMR}}\omega_{\phi}\right)}{\omega\left(\omega_{\theta}+\omega_{\phi}\right)}\sin\theta_{{\rm M}}\label{eq:ESMRy} \end{aligned}\] at $\phi_{{\rm M}}=0$. The configuration is depicted in Fig. [\[fig:Typical-configuration-of\]](#fig:Typical-configuration-of){reference-type="ref" reference="fig:Typical-configuration-of"}(b). $E_{{\rm sym}}$ includes signals due to the rectification effects and, importantly, possesses the same symmetry as the ISHE signal, i.e. $\sin\theta_{{\rm M}}$. Therefore, the part of the Lorentzian signal with the sign change following to the magnetization reversal cannot be attributed to only the ISHE without further examinations. In Ref. [@chen2013direct], a separation of the contributions based on the difference in the pre-factors, i.e. $\omega_{\theta\phi}$, $\omega_{\phi}$, $\omega$, darling the $\theta_{{\rm M}}$ scan is suggested and will be introduced in Sect. [\[subsec:Angular-dependence-of\]](#subsec:Angular-dependence-of){reference-type="ref" reference="subsec:Angular-dependence-of"}. Note that Eq. ([\[eq:ESMRy\]](#eq:ESMRy){reference-type="ref" reference="eq:ESMRy"}) holds only when the rf current is constant during the scan. However, the stray rf current often shows angular dependence. The heating effect due to transverse thermoelectric effects can also induce the similar signal to the ISHE, which is given by \[E_{{\rm sym,T{\rm TE}}}^{y}\propto\frac{\omega_{\phi}}{\omega_{\theta}+\omega_{\phi}}{\rm sin\theta_{{\rm M}}.}\] The FMR heating contribution discussed in Sect. [\[subsec:Induced-voltage-by\]](#subsec:Induced-voltage-by){reference-type="ref" reference="subsec:Induced-voltage-by"} can be extracted by the frequency dependence measurement. The MSSW heating contribution discussed in Sect. [\[subsec:Induced-voltage-by\]](#subsec:Induced-voltage-by){reference-type="ref" reference="subsec:Induced-voltage-by"} is unable to be removed until the F layer thickness is reduced. One possible solution for handling this difficulty is finding this contribution from the the calculation based on the Seebeck coefficient following a temperature profile measurement as is done in Ref. [@qiu2015spin]. An asymmetric component is a sign of contribution from the rectification effects although the reverse is not true because rectification signals can have only the Lorentzian component. If one knows the direction of the rf current, then the asymmetric component might be a good measure to determine its magnitude and thus the rectification contribution. The direction may be estimated by measuring the voltage along other directions as shown in Ref. [@lustikova2015vector; @tsukahara2014self]. ## Thickness dependence[\[subsec:Thickness-dependence-of\]]{#subsec:Thickness-dependence-of label="subsec:Thickness-dependence-of"} The voltage signals from the ISHE and rectification effects have different dependence on the thickness of the F and N layers. For a bilayer where the galvanomagnetic effects in the F layer is dominant, the symmetric Lorentzian signal after taking the difference between $\mathbf{m}_{0}$ and $-\mathbf{m}_{0}$ is expressed in the following form, \[E_{{\rm sym},{\rm ISHE}}^{y}=\frac{E_{{\rm ISHE}}^{0}}{\sigma_{{\rm N}}d_{{\rm N}}+\sigma_{{\rm F}}d_{{\rm F}}}\tanh\frac{d_{{\rm N}}}{2\lambda},\label{eq:thEyishe}\] \[E_{{\rm sym},{\rm rect}}^{y}=\frac{d_{{\rm F}}E_{{\rm rect}}^{{\rm 0}}}{\sigma_{{\rm N}}d_{{\rm N}}+\sigma_{{\rm F}}d_{{\rm F}}},\] where $E_{{\rm ISHE}}^{0}$ and $E_{{\rm rect}}^{0}$ are respectively determined by Eqs. ([\[eq:EISHE\]](#eq:EISHE){reference-type="ref" reference="eq:EISHE"}) and ([\[eq:Erectex\]](#eq:Erectex){reference-type="ref" reference="eq:Erectex"}). Regarding $E_{{\rm ISHE}}^{0}$ (or $\left\langle j_{{\rm s}}^{{\rm F/N}}\right\rangle _{t}$) and $E_{{\rm rect}}^{0}$ as constants, the $d_{{\rm N}}$ dependence of Eq. ([\[eq:thEyishe\]](#eq:thEyishe){reference-type="ref" reference="eq:thEyishe"}) reads a tangent hyperbolic function divided by the total resistance, which shows a positive peak around $d_{{\rm N}}\approx2\lambda$ \[See Fig. [\[fig:Thickness-dependence-of\]](#fig:Thickness-dependence-of){reference-type="ref" reference="fig:Thickness-dependence-of"}(a)\]. The $d_{{\rm N}}$ dependence of the rectification contribution shows monotonic decrease as $d_{{\rm N}}$ increases. The distinction between these two becomes difficult for $d_{{\rm N}}>\lambda_{{\rm N}}$, because both show just decrease as $d_{{\rm N}}$ increases \[See Fig. [\[fig:Thickness-dependence-of\]](#fig:Thickness-dependence-of){reference-type="ref" reference="fig:Thickness-dependence-of"}(b)\]. On the other hand, the $d_{{\rm F}}$ dependence shows clear difference; as $d_{{\rm F}}$ increases, the ISHE and rectification contributions respectively show monotonic decrease and increase before saturation. This is because the SP (the rectification effects) induced voltage is electrically-shorted by the additional F (N) layer. This contrary dependence of the SP and rectification contributions is very useful to separate them. For a practice analysis, thickness dependence of parameters should be considered, such as $g_{{\rm eff,r}}^{\uparrow\downarrow}$, $\tilde{\rho}^{{\rm F}}$, and $j_{{\rm rf}}\left(z\right)$. The FMR and MSSW heating effects can complicate the thickness dependence of the signal because the induced temperature gradient depends on sample structure and environment. Therefore, it is highly recommended to carefully consider an appropriate design of the samples as described in Sect. [\[subsec:Suitable-materials-for\]](#subsec:Suitable-materials-for){reference-type="ref" reference="subsec:Suitable-materials-for"}. ## Angular dependence[\[subsec:Angular-dependence-of\]]{#subsec:Angular-dependence-of label="subsec:Angular-dependence-of"} The magnetization angular dependence of $\mathbf{E}_{{\rm ISHE}}$ is derived in Ref. [@ando2008angular] and that of $\mathbf{E}_{{\rm rect}}$ due to the AMR (PHE), AHE and SMR can be found in Refs. [@chen2013direct; @iguchi2013effect; @Egan:1963gr; @mecking2007microwave; @yamaguchi2008broadband]. Here, we will focus on the ISHE and rectification effects, because the heating effects can not be isolated by the angular dependence. The angular dependence can be studied by two types of rotation: out-of-plane angular dependence (OP) and in-plane angular dependence (IP), which are shown in Figs. [\[fig:OP-angular-dependence\]](#fig:OP-angular-dependence){reference-type="ref" reference="fig:OP-angular-dependence"} and [\[fig:IP-angular-dependence\]](#fig:IP-angular-dependence){reference-type="ref" reference="fig:IP-angular-dependence"}. In addition, there are two different means to excite the ferromagnet: applying an rf excitation field in OP or IP. Here, these four configurations are considered in a film system with uniaxial anisotropy perpendicular to the film plane, which is relevant to most of experiments. Such systems are described by the magnetostatic energy: \[F_{m}=-\mathbf{m}\cdot\mathbf{H}_{0}+\frac{I_{{\rm s}}^{2}}{2\mu_{0}}m_{z}^{2}-K_{{\rm u}}m_{z}^{2},\label{eq:Um}\] where $\mathbf{H}_{0}$ is an applied external field, given by $\mathbf{H}_{0}=H_{0}\left(\sin\theta_{{\rm H}}\cos\phi_{{\rm H}}\mathbf{e}_{x}+\sin\theta_{{\rm H}}\sin\phi_{{\rm H}}\mathbf{e}_{y}+\cos\theta_{{\rm H}}\mathbf{e}_{z}\right)$, and $K_{{\rm u}}$ denotes the perpendicular anisotropy constant. Equation ([\[eq:Um\]](#eq:Um){reference-type="ref" reference="eq:Um"}) reduces Eq. ([\[eq:omegath\]](#eq:omegath){reference-type="ref" reference="eq:omegath"}) to \[\begin{aligned} \omega_{\theta} & =\gamma\left[\mu_{0}H_{0}\cos\left(\theta_{{\rm H}}-\theta_{{\rm M}}\right)-I_{{\rm eff}}\cos^{2}\theta_{{\rm M}}\right],\\ \omega_{\phi} & =\gamma\left[\mu_{0}H_{0}\cos\left(\theta_{{\rm H}}-\theta_{{\rm M}}\right)-I_{{\rm eff}}\cos\left(2\theta_{{\rm M}}\right)\right], \end{aligned}\] and $\omega_{\phi\theta}=0$, where $I_{{\rm eff}}$ denotes the effective magnetization $I_{{\rm eff}}=I_{{\rm s}}-2\mu_{0}K_{{\rm u}}/I_{{\rm s}}$, and $\theta_{{\rm H}}$ $\left(\phi_{{\rm H}}\right)$ denotes the polar (azimuthal) angle in the spherical coordinate as shown in Fig. [\[fig:OP-angular-dependence\]](#fig:OP-angular-dependence){reference-type="ref" reference="fig:OP-angular-dependence"}(a). The resonance field $H_{{\rm r}}$ is determined by $H_{0}$ which simultaneously satisfies $\omega_{{\rm r}}=\sqrt{\omega_{\theta}\omega_{\phi}}$ and $\mathbf{m}_{0}\times\mathbf{H}_{{\rm eff}}=0$, which is reduced to \[2\mu_{0}H_{0}\sin\left(\theta_{{\rm H}}-\theta_{{\rm M}}\right)+I_{{\rm eff}}\sin\left(2\theta_{{\rm M}}\right)=0\] and $\phi_{{\rm H}}=\phi_{{\rm M}}$. Note that in the following calculation we still use $\omega_{{\rm \theta}\left(\phi\right)}$ for simple notation but impose $\omega_{{\rm \theta\phi}}=0$. In the setup for the OP angular dependence measurements, the magnetization is rotated in the $xz$ plane ($\phi_{{\rm M}}=0$) and electrodes for detecting the ISHE voltage are placed in the $y$ axis \[See Fig. [\[fig:OP-angular-dependence\]](#fig:OP-angular-dependence){reference-type="ref" reference="fig:OP-angular-dependence"}(a)\]. An IP excitation field, $h_{{\rm rf}}$ along the $y$ axis, induces \[E_{{\rm sym},{\rm ISHE}}^{y}=A_{{\rm ISHE}}\frac{\omega_{\phi}}{\left(\omega_{\theta}+\omega_{\phi}\right)^{2}}\sin\theta_{{\rm M}},\] \[E_{{\rm sym},{\rm rect}}^{y}=A_{{\rm rect}}\frac{\left(\omega\rho_{{\rm AHE}}{\rm Re}\left[j_{{\rm rf}}^{x}\right]+\omega_{\phi}\Delta\rho_{{\rm AMR}}{\rm Im}\left[j_{{\rm rf}}^{x}\right]\right)}{\omega\left(\omega_{\theta}+\omega_{\phi}\right)}\sin\theta_{{\rm M}},\] where the coefficients $A$ are given by \[A_{{\rm ISHE}}=w\tilde{R}_{{\rm tot}}\theta_{{\rm SHE}}\frac{2e}{\hbar}\lambda\tanh\frac{d_{{\rm N}}}{2\lambda}\frac{\hbar g_{r,{\rm eff}}^{\uparrow\downarrow}}{4\pi}\frac{\gamma^{2}\mu_{0}^{2}h_{{\rm rf}}^{2}}{\alpha^{2}},\] \[A_{{\rm rect}}=\tilde{R}_{{\rm tot}}\frac{\sigma_{{\rm F}}d_{{\rm F}}}{2}\frac{\gamma\mu_{0}h_{{\rm rf}}}{\alpha},\] both of which are proportional to $\sin\theta_{{\rm M}}$. In Ref. [@chen2013direct], the separation is done based on the difference in the dependence on $\omega_{\theta}\left(\theta_{{\rm M}}\right)$ and $\omega_{\phi}\left(\theta_{{\rm M}}\right)$ between the ISHE and rectification voltages, which is because they have different responses to the effective field sweeping over $\theta_{{\rm H}}$. Figure [\[fig:OP-angular-dependence\]](#fig:OP-angular-dependence){reference-type="ref" reference="fig:OP-angular-dependence"}(b) shows $H_{{\rm r}}$ and $\theta_{{\rm M}}$ as a function of $\theta_{{\rm H}}$ calculated with $\gamma=1.79\times10^{11}$ ${\rm T^{-1}s^{-1}}$ and $I_{{\rm eff}}=1$ T. Figure [\[fig:OP-angular-dependence\]](#fig:OP-angular-dependence){reference-type="ref" reference="fig:OP-angular-dependence"}(c) shows the $\theta_{{\rm M}}$ dependence of $E_{{\rm ISHE}}$ and $E_{{\rm rect}}$, which possess similar form, but slight difference seen in the solid and dashed curves. This similarity can be a large source of error in data fitting. An OP excitation field, $h_{z}$, induces \[E_{{\rm sym},{\rm ISHE}}^{y}=A_{{\rm ISHE}}\frac{\omega_{\theta}}{\left(\omega_{\theta}+\omega_{\phi}\right)^{2}}\sin^{3}\theta_{{\rm M}},\] \[E_{{\rm sym},{\rm rect}}^{y}=A_{{\rm rect}}\frac{\left(\omega\Delta\rho_{{\rm AMR}}{\rm Re}\left[j_{{\rm rf}}^{x}\right]-\omega_{\theta}\rho_{{\rm AHE}}{\rm Im}\left[j_{{\rm rf}}^{x}\right]\right)}{\omega\left(\omega_{\theta}+\omega_{\phi}\right)}\sin^{2}\theta_{{\rm M}},\] \[See Fig. [\[fig:OP-angular-dependence\]](#fig:OP-angular-dependence){reference-type="ref" reference="fig:OP-angular-dependence"}(d)\]. At a glance, the difference in the two signals are clear, namely $E_{{\rm rect}}^{y}\propto\sin^{2}\theta_{{\rm M}}$ and $E_{{\rm ISHE}}\propto\sin^{3}\theta_{{\rm M}}$. However, when $j_{{\rm rf}}\propto\sin\theta_{{\rm M}}$ holds, the contributions cannot be separated by the harmonic functions. Similarly to the OP angular dependence in the OP excitation, the IP angular dependence excited by an IP rf field has the same difficulty. In the setup, the magnetization is rotated in the $xy$ plane ($\theta_{{\rm M}}=90^{\circ}$) and electrodes for detecting the ISHE voltage are placed in the $y$ axis. The electromotive forces in the $y$ direction are \[E_{{\rm sym},{\rm ISHE}}^{y}=A_{{\rm ISHE}}\frac{\omega_{\phi}}{\omega\left(\omega_{\theta}+\omega_{\phi}\right)^{2}}\cos^{3}\phi_{{\rm M}},\] \[E_{{\rm sym},{\rm rect}}^{y}=A_{{\rm rect}}\cos\phi_{{\rm M}}\frac{\omega_{\phi}\Delta\rho_{{\rm AMR}}\left(\cos\left(2\phi_{{\rm M}}\right){\rm Im}\left[j_{{\rm rf}}^{x}\right]+\sin\left(2\phi_{{\rm M}}\right){\rm Im}\left[j_{{\rm rf}}^{y}\right]\right)+\omega\rho_{{\rm AHE}}{\rm Re}\left[j_{{\rm rf}}^{x}\right]}{\omega\left(\omega_{\theta}+\omega_{\phi}\right)}\label{eq:EsymIPIP}\] \[See Fig. [\[fig:IP-angular-dependence\]](#fig:IP-angular-dependence){reference-type="ref" reference="fig:IP-angular-dependence"}(a)\]. In this configuration, the ISHE and AMR contributions mix even for the simplest condition that the rf current is angular-independent. The IP angular dependence excited by an OP rf field has an advantage to the previous three angular dependences because the AMR and AHE show different symmetric angular dependences in the IP configuration. An OP excitation field, $h_{z}$, induces \[E_{{\rm sym},{\rm ISHE}}^{y}=A_{{\rm ISHE}}\frac{\omega_{\theta}}{\left(\omega_{\theta}+\omega_{\phi}\right)^{2}}\cos\phi_{{\rm M}},\label{eq:IP-1}\] \[E_{{\rm sym},{\rm rect}}^{y}=A_{{\rm rect}}\frac{\omega\Delta\rho_{{\rm AMR}}\left(\cos\left(2\phi_{{\rm M}}\right){\rm Re}\left[j_{{\rm rf}}^{x}\right]+\sin\left(2\phi_{{\rm M}}\right){\rm Re}\left[j_{{\rm rf}}^{y}\right]\right)-\omega_{\theta}\rho_{{\rm AHE}}{\rm Im}\left[j_{{\rm rf}}^{x}\right]}{\omega\left(\omega_{\theta}+\omega_{\phi}\right)}.\label{eq:IP-2}\] For the simplest case where $j_{x\left(y\right)}^{{\rm rf}}$ is constant during the rotation, importantly, the ISHE, AMR and AHE show the different angular dependences, $\cos\phi_{{\rm M}}$, $\cos\left(2\phi_{{\rm M}}\right)$, and constant \[Fig. [\[fig:IP-angular-dependence\]](#fig:IP-angular-dependence){reference-type="ref" reference="fig:IP-angular-dependence"}(b)\]. Thus, fitting the result using the harmonic functions gives the ISHE contribution directly. Figure [\[fig:IP-angular-dependence\]](#fig:IP-angular-dependence){reference-type="ref" reference="fig:IP-angular-dependence"} shows the calculation result based on Eqs. ([\[eq:IP-1\]](#eq:IP-1){reference-type="ref" reference="eq:IP-1"}) and ([\[eq:IP-2\]](#eq:IP-2){reference-type="ref" reference="eq:IP-2"}). When $j_{x\left(y\right)}^{{\rm rf}}$ has an angular dependence, an analysis method based on the Fourier series coefficient is effective in the measurements on the IP angular dependence with OP excitation. The $n$-th Fourier cosine coefficient of the voltage is given by \[\begin{aligned} F_{{\rm sym}}^{{\rm cos}}\left(n\right) & =\frac{\omega_{\theta}A{}_{{\rm ISHE}}}{\left(\omega_{\theta}+\omega_{\phi}\right)^{2}}\delta_{n-1}+\frac{A_{{\rm rect}}}{\omega\left(\omega_{\theta}+\omega_{\phi}\right)}\left(\omega\Delta\rho_{{\rm AMR}}{\rm Re}\left[J_{{\rm rf}}^{x}\left(n-2\right)\right]-\omega_{\theta}\rho_{{\rm AHE}}{\rm Im}\left[J_{{\rm rf}}^{x}\left(n\right)\right]\right),\\ F_{{\rm asym}}^{{\rm cos}}\left(n\right) & =\frac{A_{{\rm rect}}}{\omega\left(\omega_{\theta}+\omega_{\phi}\right)}\left(-\omega\Delta\rho_{{\rm AMR}}{\rm Im}\left[J_{{\rm rf}}^{x}\left(n-2\right)\right]-\omega_{\theta}\rho_{{\rm AHE}}{\rm Re}\left[J_{{\rm rf}}^{x}\left(n\right)\right]\right),\label{eq:Fsymasym} \end{aligned}\] where $\delta_{n}$ denotes the Kronecker delta function and $J_{{\rm rf}}^{x}\left(n\right)$ is the $n$-th Fourier cosine coefficient of $j_{{\rm rf}}^{x}$. Figure [\[fig:Analysis-scheme-based\]](#fig:Analysis-scheme-based){reference-type="ref" reference="fig:Analysis-scheme-based"}(a) shows an expected intensity of the coefficients for $J_{{\rm rf}}^{x}\left(n\right)\propto\delta_{n}$. The contributions are clearly separated. Figure [\[fig:Analysis-scheme-based\]](#fig:Analysis-scheme-based){reference-type="ref" reference="fig:Analysis-scheme-based"}(b) shows a calculation of the coefficients for $J_{{\rm rf}}^{x}\left(n\right)$ with $J_{{\rm rf}}^{x}\left(1\right)\neq0$. $J_{{\rm rf}}^{x}\left(1\right)$ can induces an ISHE-like signal via the AHE. However, this contribution can be removed by comparing symmetric and antisymmetric components because there is a relation, \[F_{{\rm sym}}^{{\rm cos}}\left(1\right)=-\frac{\omega_{\theta}}{\omega}\frac{\rho_{{\rm AHE}}}{\Delta\rho_{{\rm AMR}}}F_{{\rm asym}}^{{\rm cos}}\left(3\right),\] in the absence of the ISHE. Therefore, the factor above gives the upper limit of the Lorentzian part due to the AHE. When $\omega_{{\rm \theta}}\left(\omega_{\phi}\right)$ possesses $\phi_{{\rm M}}$ dependence because of a magnetic anisotropy field in-plane, Eq. ([\[eq:Fsymasym\]](#eq:Fsymasym){reference-type="ref" reference="eq:Fsymasym"}) should be recalculated by considering the Fourier coefficients of $\omega_{\theta}$ and $\omega_{{\rm \phi}}$. Note that when $\Delta\rho=0$ but $\rho_{{\rm AHE}}\neq0$, this method can not be applied, and it is better to change a material for the F layer or to try a measurement on the ferromagnetic layer thickness dependence described in Sect. [\[subsec:Thickness-dependence-of\]](#subsec:Thickness-dependence-of){reference-type="ref" reference="subsec:Thickness-dependence-of"}. ## Frequency dependence Here, we focus on the difference in the frequency dependence of the signals from the ISHE and the microwave effects. The frequency dependence has attracted much attention for its nonlinear physics coming from magnon-magnon interactions. Our interest is the linear excitation regime in which the derived equations for the FMR can be used. The heating effects show a clear difference from the ISHE and rectification effects, so that this method works effectively for removing the heating contribution. In the calculation we consider a system with field along the $x$ and rf field along the $y$ axis, i.e. the IP excitation at $\theta_{{\rm M}}=90^{\circ}$ and $\phi_{{\rm M}}=0^{\circ}$. This configuration is often used in measurements with a microstrip line or a coplanar waveguide. The frequency dependence of the ISHE and rectification signals are respectively given by \[E_{{\rm sym},{\rm ISHE}}^{y}=A_{{\rm ISHE}}\frac{\omega_{\phi}}{\left(\omega_{\theta}+\omega_{\phi}\right)^{2}}\] and \[E_{{\rm sym},{\rm rect}}^{y}=A{}_{{\rm rect}}\frac{\left(\omega\rho_{{\rm AHE}}{\rm Re}\left[j_{{\rm rf}}^{x}\right]+\omega_{\phi}\Delta\rho_{{\rm AMR}}{\rm Im}\left[j_{{\rm rf}}^{x*}\right]\right)}{\omega\left(\omega_{\theta}+\omega_{\phi}\right)}.\] The electromotive force due to the heating, proportional to microwave absorption $\Delta P$, is given by \[E_{{\rm sym},{\rm TE}}^{y}=A{}_{{\rm TE}}\frac{\omega_{\phi}}{\left(\omega_{\theta}+\omega_{\phi}\right)}\] with $A_{{\rm TE}}=\frac{\gamma\mu_{0}h_{{\rm rf}}^{2}}{2\alpha}\left(A_{{\rm TTE}}+A_{{\rm SE}\left({\rm MSSW}\right)}\right)$. Figure [\[fig:frequency-dependence\]](#fig:frequency-dependence){reference-type="ref" reference="fig:frequency-dependence"} shows the calculated frequency dependence of the ISHE, rectification effects, and heating effects. Since the ISHE and AHE are similar, the separation between the ISHE and rectification signals based on the frequency will not be so accurate. The difference between the SP and the rectification effects comes from the ellipticity of magnetization precession due to demagnetizing and anisotropic fields. At frequencies $\omega>\gamma I_{{\rm eff}}$,the large external field necessary for the FMR makes the precession trajectory circular, so that both of them become proportional to $1/\omega$. The heating is proportional to $\Delta P$ \[Eq. ([\[eq:delP\]](#eq:delP){reference-type="ref" reference="eq:delP"})\] and is proportional to $\omega_{\theta\left(\phi\right)}/\omega$, which reaches constant at high frequencies. This feature is distinct from the others. The studies on the frequency dependence indicates that the thermoelectric contribution in the ferrimagnetic insulator/Pt bilayer systems is not dominant for the microwave spin pumping experiments. # Summary[\[sec:Summary\]]{#sec:Summary label="sec:Summary"} In this article, we reviewed voltage generation by the SP-induced ISHE, the rectification effects due to galvanomagnetic effects, and the heating effects due to thermoelectric effects. The electric detection of a spin current induced by the SP using the ISHE is a strong method to study spin physics in a material of interest. The key for the study is clear separation between the ISHE and the other extrinsic contributions. In some configurations, they look similar to each other, but by employing the separation methods introduced here, one can perform a reliable measurement with high accuracy. Figure [\[fig:Summary-of-how\]](#fig:Summary-of-how){reference-type="ref" reference="fig:Summary-of-how"} summarizes the recommended method by which the accuracy can be easily obtained. For a better experiment, a material of interest should be on top of a thin ferrimagnetic insulator film, which reduces both the rectification and heating effects. For systems with metallic ferromagnet, the IP magnetization angular dependence is the best configuration to clarify the differences among the voltage signals of the ISHE and the rectification effects, because the ISHE is sensitive to spin polarization while the AMR is sensitive to the magnitude than the polarization. The separation schemes discussed in this article provides a better way to extract the SP-originated signals. We thus anticipate that the experimental schemes help further material investigations and contribute to the development of novel spintronic devices.	{'timestamp': '2016-07-19T02:03:17', 'yymm': '1607', 'arxiv_id': '1607.04716', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04716'}
null	null
null	null
# Introduction {#sintro} # Background: CUR and low-rank approximation {#sbcgr} Low-rank approximation of an $m\times n$ matrix $W$ having a small numerical rank $r$, that is, having a well-conditioned rank-$r$ matrix nearby, is one of the most fundamental problems of numerical linear algebra with a variety of applications to highly important areas of modern computing, which range from the machine learning theory and neural networks, to numerous problems of data mining and analysis. One of the most studied approaches to the solution of this problem is given by $CUR$ approximation where $C$ and $R$ are a pair of $m\times l$ and $k\times n$ submatrices formed by $l$ columns and $k$ rows of the matrix $W$, respectively, and $U$ is a $k\times l$ matrix such that $W\approx CUR$. Every low-rank approximation allows very fast approximate multiplication of the matrix $W$ by a vector, but CUR approximation is particularly transparent and memory efficient. The algorithms for computing it are characterized by the two main parameters: (i) their complexity and (ii) bounds on the error norms of the approximation. We assume that $r\ll \min\{m,n\}$, that is, the integer $r$ is much smaller than $\min\{m,n\}$, and we seek algorithms that use $o(mn)$ flops, that is, much fewer than the information lower bound $mn$. # State of the art and our progress {#ssartpr} The algorithms of and compute CUR approximations by using order of $mn\min\{m,n\}$ flops.[^1] do this in $O(mn\log(mn))$ flops by using randomization. These are record upper bounds for computing a CUR approximation to any input matrix $W$, but the user may be quite happy with having a close CUR approximations to many matrices $W$ that make up the class of his/her interest. The information lower bound $mn/2$ (a flop involves at most two entries) does not apply to such a restricted input classes, and we go well below it in our paper (we must refer to that paper for technical details because of the limitation on the size of this submission). We first formalize the problem of CUR approximation of an average $m\times n$ matrix of numerical rank $r\ll \min\{m,n\}$, assuming the customary Gaussian (normal) probability distribution for its $(m+n)r$ i.i.d. input parameters. Next we consider a two-stage approach: (i) first fix a pair of integers $k\le m$ and $l\le n$ and compute a CUR approximation (by using the algorithms of or ) to a random $k\times l$ submatrix and then (ii) extend it to computing a CUR approximation of an input matrix $W$ itself. We must keep the complexity of Stage (i) low and must extend the CUR approximation from the submatrix to the matrix $W$. We prove that for a specific class of input matrices $W$ these two tasks are in conflict (see Example 11 of ), but such a class of hard inputs is narrow, because we prove that our algorithm produces a close approximation to the average $m\times n$ input matrix $W$ having numerical rank $r$. (We define such an average matrix by assuming the standard Gaussian (normal) probability distribution.) By extending our two-stage algorithms with the technique of, which we call cross-approximation, we a little narrow the class of hard inputs of Example 11 of to the smaller class of Example 14 of and moreover deduce a sharper bounds on the error of approximation by maximizing the volume of an auxiliary $k\times l$ submatrix that defines a CUR approximation In our extensive tests with a variety of real world input data for regularization of matrices from Singular Matrix Database, our fast algorithms consistently produce close CUR approximation. Since our fast algorithms produce reasonably accurate CUR approximation to the average input matrix, the class of hard input matrices for these algorithms must be narrow, and we studied a tentative direction towards further narrowing this input class. We prove that the algorithms are expected to output a close CUR approximation to any matrix $W$ if we pre-process it by applying Gaussian multipliers. This is a nontrivial result of independent interest (proven on more than three pages), but its formal support is only for application of Gaussian multipliers, which is quite costly. We hope, however, that we can still substantially narrow the class of hard inputs even if we replace Gaussian multipliers with the products of reasonable numbers of random bidiagonal matrices and if we partly curb the permutation of these matrices. If we achieve this, then preprocessing would become non-costly. This direction seems to be quite promising, but still requires further work. Finally, our algorithms can be extended to the acceleration of various computational problems that are known to have links to low-rank approximation, but in our concluding Section [4](#scncl){reference-type="ref" reference="scncl"} we describe a novel and rather unexpected extensions to the acceleration of the Fast Multipole Method and Conjugate Gradient Algorithms,[^2] both being among the most celebrated achievements of the 20th century in Numerical Linear Algebra. ## Some related results on matrix algorithms and our progress on other fundamental subjects of matrix computations {#srltwr} A huge bibliography on CUR and low-rank approximation, including the known best algorithms, which we already cited, can be accessed from the papers, , and . Our main contribution is dramatic acceleration of the known algorithms. Some of our techniques extend the ones of , , and, where we also show duality of randomization and derandomization and apply it to fundamental matrix computations. In we prove that preprocessing with almost any well-conditioned multiplier of full rank is as efficient on the average for low-rank approximation as preprocessing with a Gaussian one, and then we propose some new highly efficient sparse and structured multipliers. Besides providing a new insight into the subject, this motivates the design of more efficient algorithms and shows specific direction to this goal. We obtain similar progress in for and for preprocessing Gaussian elimination with no pivoting and block Gaussian elimination. We recall that Gaussian elimination with partial pivoting is performed millions time per day, where pivoting, required for numerical stabilization, is frequently a bottleneck because it interrupts the stream of arithmetic operations with foreign operations of comparison, involves book-keeping, compromises data locality, and increases communication overhead and data dependence. Randomized preprocessing is a natural substitution for pivoting, and in we show that Gaussian elimination with no pivoting as well as block Gaussian elimination (which is another valuable algorithm and which also requires protection against numerical problems) are efficient on the average input with preprocessing by any nonsingular and well-conditioned multipliers. obtains similar progress for the important subject of the approximation of trailing singular spaces associated with the $\nu$ smallest singular values of a matrix having numerical nullity $\nu$. Our current progress greatly supersedes these earlier results, however, in terms of the scale of the acceleration of the known algorithms. Our technique of representing random Gaussian multipliers as a product of random bidiagonal factors, our extension of CUR approximation to FMM and CG algorithms, and our analysis of CUR approximation for the average input are new and can have some independent interest. # Conclusions {#scncl} We dramatically accelerated the known algorithms for the fundamental problems of CUR and low-rank approximation in the case of the average input matrix and then pointed out a direction towards heuristic extension of the resulting fast algorithm to a wider class of inputs by applying quasi Gaussian preprocessing. Our extensive tests for benchmark matrices of discretized PDEs have consistently supported the results of our formal analysis. Our study can be extended to a variety of important subjects of matrix computations. Some of such extensions have been developed in papers , and , and there are various challenging directions for further progress. In particular our accelerated CUR and low-rank approximation enables faster solution of some new important computational problems, thus extending the long list of the known applications. In the concluding section of , we add two new highly important subjects to this long list. [^1]: Here and hereafter "flop\" stands for "floating point arithmetic operation\". [^2]: Hereafter we use the acronyms FMM and CG.	{'timestamp': '2016-12-20T02:08:35', 'yymm': '1607', 'arxiv_id': '1607.04825', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04825'}
null	null
null	null
null	null
null	null
null	null
# Menger co-analytic groups We shall assume all spaces are completely regular. Arhangel'skiı̆ proved that Menger analytic spaces are $\sigma$-compact, generalizing Hurewicz's classic theorem that Menger completely metrizable spaces are $\sigma$-compact. Menger's conjecture was disproved in, where Miller and Fremlin also showed it undecidable whether Menger co-analytic sets of reals are $\sigma$-compact. In we proved that Menger Čech-complete spaces are $\sigma$-compact and obtained various sufficient conditions for Menger co-analytic topological spaces to be $\sigma$-compact. We continue that study here. In we observed that $\mathbf{\Pi}_{1}^1$-determinacy--which we also call CD: the Axiom of Co-analytic Determinacy--implies Menger co-analytic sets of reals are $\sigma$-compact. Indeed, PD (the Axiom of Projective Determinacy) implies Menger projective sets of reals are $\sigma$-compact ,. When one goes beyond co-analytic spaces in an attempt to generalize Arhangel'skiı̆'s theorem, one runs into ZFC counterexamples, but it is not clear whether there is a ZFC co-analytic counterexample. Assuming $V=L$, there is a counterexample which is a subset of $\mathbb{R}$ ,. Here we prove: We first slightly generalize the CD result quoted above. In order to prove this, we need some general facts about analytic spaces and perfect maps. Since analytic spaces are Lindelöf, a co-analytic group is a perfect pre-image of a metrizable space. Since Menger spaces are Lindelöf, a Menger co-analytic topological group $G$ is a perfect pre-image of a separable metrizable space $M$. In, we proved perfect images of co-analytic spaces are co-analytic, so $M$ is co-analytic and Menger and therefore $\sigma$-compact by CD and Lemma [\[lem1\]](#lem1){reference-type="ref" reference="lem1"}. Then $G$ is $\sigma$-compact as well.0◻After hearing about Theorem [\[thm1\]](#thm1){reference-type="ref" reference="thm1"}, S. Tokgöz proved: # Productively Lindelöf co-analytic spaces We have extensively studied productively Lindelöf spaces, as have other authors. Since productively Lindelöf spaces consistently are Menger it is natural to ask: It is consistent that there is a Michael space, but it is not known whether there is one from ZFC. If there is no Michael space, then the space $\mathbb{P}$ of irrationals is productively Lindelöf, co-analytic, nowhere locally compact, but not $\sigma$-compact. We shall prove: I do not know whether the unwanted "nowhere locally compact\" clause can be removed. It assures us that $\beta X \setminus X$ is dense in $\beta X$. Laying the groundwork for proving Theorem [\[thm4\]](#thm4){reference-type="ref" reference="thm4"}, we need some definitions and previous results. We can now prove Theorem [\[thm4\]](#thm4){reference-type="ref" reference="thm4"}. Let $X$ be productively Lindelöf, co-analytic, and nowhere locally compact. $\beta X \setminus X$ is analytic and hence Lindelöf and separable. It is dense in $\beta X$, so $w(\beta X)$ and hence $w(X) \leq 2^{\aleph_0}=\aleph_1$. Then $X$ is Alster. Since $\beta X \setminus X$ is Lindelöf, $X$ has countable type, so it is $\sigma$-compact. 0◻For metrizable spaces, Repovś and Zdomskyy proved: We would like to drop the metrizability assumption, using: As in, we run up against the unsolved problem: However, we can apply the various partial results in the previous section and to obtain: The two hypotheses of Theorem [\[thm10\]](#thm10){reference-type="ref" reference="thm10"} are compatible, since it is well-known that CH is compatible with the existence of a measurable cardinal, and that CH implies the existence of a Michael space. Various other hypotheses about cardinal invariants of the continuum also imply the existence of a Michael space--see e.g.. These are all compatible with CD. We also have: Repovś and Zdomskyy prove: # K-analytic and K-Lusin spaces We take the opportunity to make some observations about K-analytic, K-Lusin, absolute Borel, Frolík, and what Arhangel'skiı̆ calls Borelian of the first type spaces. These are all attempts to generalize concepts of Descriptive Set Theory beyond separable metrizable spaces. Frolík showed that each Frolík space is absolute $K_{\sigma\delta}$ (and therefore Lindelöf), i.e. the intersection of countably many $\sigma$-compact subspaces of its Čech-Stone compactification (and conversely), and also is the continuous image of a Čech-complete Frolík space, so that Frolík spaces are absolute Borel and $K$-analytic. $K$-Lusin spaces are clearly $K$-analytic; $K$-Lusin spaces are also Frolík. Since $K$-analytic metrizable spaces are analytic and analytic Menger spaces are $\sigma$-compact, we see that Menger K-analytic spaces are projectively $\sigma$-compact. In we proved that projectively $\sigma$-compact Lindelöf spaces are Hurewicz, so we conclude: Hurewicz is a property strictly between $\sigma$-compact and Menger. A space is Hurewicz if every Čech-complete space including it includes a $\sigma$-compact subspace including it (This is equivalent to the usual definition--see ). This theorem may give some inkling as to why it seems to be hard to find topological properties that imply Hurewicz spaces are $\sigma$-compact which don't in fact imply Menger spaces are $\sigma$-compact. There are, however, Hurewicz subsets of $\mathbb{R}$ which are not $\sigma$-compact---see e.g.. There is a projectively $\sigma$-compact Frolík space which is not $\sigma$-compact (Okunev's space--see ). Okunev's space is also not Čech-complete, since Menger Čech-complete spaces are $\sigma$-compact. There is a Frolík subspace of $\mathbb{R}$ which is not Čech-complete, since "Čech-complete\" translates into being a $G_{\delta}$, and we know the Borel hierarchy is non-trivial. There are of course analytic subsets of $\mathbb{R}$ which are not absolute Borel and hence not Frolík. Moore's L-space is projectively countable but not K-analytic. The reason is that all its points are $G_\delta$'s, which contradicts projectively countable for K-analytic spaces. Since K-Lusin spaces are Frolík, it is worth mentioning that: We could add to this "if and only if every K-Lusin space is productively Lindelöf". Also of interest is: Compare with [\[prop8\]](#prop8){reference-type="ref" reference="prop8"}. The Corollary follows from [\[lem9\]](#lem9){reference-type="ref" reference="lem9"}. In order to prove [\[thm32\]](#thm32){reference-type="ref" reference="thm32"} we need to know: We also need to know about p-spaces and $\mathbf{\mathop{\pmb{\sum}}}$-spaces, but do not need their internal characterizations. What we need are: To see this, we introduce: Rogers and Jayne prove: Theorem [\[thm310\]](#thm310){reference-type="ref" reference="thm310"} now follows from [\[thm32\]](#thm32){reference-type="ref" reference="thm32"}. Arhangel'skiı̆ proved that Borelian sets of the first type are s-spaces. This is interesting because: Note Okunev's space is Lindelöf absolute $F_{\sigma\delta}$ but is not s, since it is not of countable type, while s-spaces are. By Theorem [\[thm32\]](#thm32){reference-type="ref" reference="thm32"}, Okunev's space is not co-analytic. Borel sets of reals are of course analytic; Okunev's space shows that Lindelöf absolute Borel spaces need not be analytic, since it is Menger but not $\sigma$-compact. Compact spaces are Borelian of the first type, so the latter spaces need not be analytic. A somewhat smaller class of spaces than the $K$-analytic ($K$-Lusin) ones is comprised of what Rogers and Jayne call the proper $K$-analytic (proper $K$-Lusin) spaces. Rogers and Jayne prove that a space is proper K-Lusin if and only if both it and its remainder are K-analytic. It follows that a space is proper K-Lusin if and only if it and its remainder are K-Lusin. They also prove that K-Lusin spaces are absolute $K_{\sigma\delta}$, i.e. what we have called Frolík. It follows that proper K-Lusin spaces are both $K_{\sigma\delta}$ and $G_{\delta\sigma}$, i.e. countable unions of Čech-complete spaces. We shall provide a large number of equivalences for "proper K-Lusin" below. Proper K-analytic spaces are p-spaces, and their continuous real-valued images are analytic, so: Note that zero-sets are closed $G_\delta$'s, so that the absolute Baire sets, i.e. the elements of the $\sigma$-algebra generated by the zero-sets, are both Lindelöf Borelian of the first type and absolute Borel. Mixing Rogers and Jayne with Arhangel'skiı̆, we have: We know that Menger proper K-analytic (a fortiori, proper K-Lusin) spaces are $\sigma$-compact, but Menger K-analytic spaces may not be. An interesting fact about K-Lusin spaces is that: From this, we can conclude that Okunev's space is not K-Lusin, since it is not $\sigma$-compact but doesn't include a compact perfect set. Indeed we have: Thus Rothberger is a strengthening of Menger. Here are some more problems we have not been able to solve: Note $V=L$ implies there is a co-analytic Hurewicz group of reals that is not $\sigma$-compact.	{'timestamp': '2016-07-19T02:05:10', 'yymm': '1607', 'arxiv_id': '1607.04781', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04781'}
# Introduction There are many open questions regarding the strength and geometry of the magnetic field in radio galaxies and their relation to other properties of the radio source. The observed degree of polarization depends on the intrinsic properties, such as the regularity and orientation of the source magnetic fields as well as the Faraday effects from the intervening regions of ionized gas along the line of sight. The largest current sample of polarized sources is the NRAO/VLA all sky survey, NVSS, at 1.4 GHz. It shows that the majority of extragalactic radio sources are only a few percent polarized. Polarization studies of small samples of extragalactic radio sources at other frequencies also show a similar weak average polarization, and suggest the fractional polarization increases at frequencies higher than 1.4 GHz. It is not clear which mechanism is dominant in reducing the fractional polarization at lower frequencies and depolarizing the sources, although several models have been suggested. One key cause for depolarization is Faraday rotation, which can be characterized to first order by a change in the angle of the linear polarization: \[\Delta \chi=\left(0.812 \int \frac{n_e{\bf B}}{(1+z)^2}\cdot \frac{d{\bf l}}{dz} \,dz\right) \lambda^2 \equiv \phi \lambda^2\] where $\Delta \chi$ is the amount of the rotation of the polarization vector in rad, $\lambda$ is the observation wavelength in m, $z$ is the redshift of the Faraday screen, ${\bf B}$ is the ionized medium magnetic field vector in $\mu$G, $n_e$ is the number density of electrons in the medium in cm$^{-3}$ and $\,d{\bf l}$ is the distance element along the line of sight in pc. The term in parentheses is called the Faraday depth, $\phi$. For a single line of sight through a thin ionized screen, this is equivalent to the rotation measure, $\textrm{RM}$, defined by $\textrm{RM} \equiv \frac{\Delta \chi}{\Delta \lambda^2}$ which can be measured observationally. Different lines of sight to the source all within the observing beam can have different values of $\phi$. Typically, this progressively depolarizes the source at longer wavelengths, but it can also lead to constructive interference and re-polarization, i.e., higher fractional polarizations at longer wavelengths. There are at least three separate possible Faraday screens with different $\textrm{RM}$ distributions along the line of sight: the Galactic component, intervening extragalactic ionized gas, and material local to the source. Multiple studies such as have tried to identify and distinguish these separate components and study the evolution of the magnetic field of galaxies through cosmic time. When many lines of sight each have independent single Faraday depths, this problem is approached statistically. Another long standing puzzle is the anti-correlation between the total intensity of radio sources and their degree of polarization, as observed by many groups such as, ,, ,, and. The physical nature of this relation has been a mystery for almost a decade, and is confused by the dependency on other source properties. found that most of their highly polarized sources are steep spectrum, show signs of resolved structure on arc-second scales, and are lobe dominated. However, they found no further correlation between the spectral index and fractional polarization. The anti-correlation between total intensity and fractional polarization seems to become weak for very faint objects with 1.4 GHz total intensities between 0.5 mJy $< I <$ 5 mJy as suggested in, based on a small sample of polarized radio galaxies in the GOODS-N field. Recently, studied a sample of 796 radio-loud AGNs with $z < 0.7$. They found that low-excitation radio galaxies have a wide range of fractional polarizations up to $\sim$ 30 %, and are more numerous at faint Stokes I flux densities while high-excitation radio galaxies are limited to polarization degrees less than 15%. They suggest that the ambient gas density and magnetic fields local to the radio source might be responsible for the difference. Using WISE colors, suggested that the observed anti-correlation primarily reflects the difference between infrared AGN and star-dominated populations. Large samples of polarization data at multiple frequencies are required to understand the magnetic field structures and depolarization mechanisms responsible for the low observed polarization fractions. have showed the polarization fraction of compact sources decreases significantly at 189 MHz compared to 1.4 GHz. They studied a sample of 137 sources brighter than 4 mJy and only detected one polarized source with probably a depolarization mechanism intrinsic to the source. Recently, used the (hereafter TSS09) catalog, and assembled polarization spectral energy distributions for 951 highly polarized extragalactic sources over the broad frequency range, 0.4 GHz to 100 GHz. They showed that objects with flat spectra in total intensity have complicated polarization spectral energy distributions (SEDs), and are mostly re-polarized somewhere in the spectrum, while steep spectrum sources show higher average depolarization. As a result, they claimed that the dominant source of depolarization should be the local environment of the source, since the spectral index is an intrinsic property of these highly polarized sources. The current work follows up on their discovery, using a sample selected only on the basis of total intensity at 2.3 GHz. In this work, we use the data from the S-PASS survey, conducted by the Australian Parkes single dish radio telescope at 2.3 GHz. We cross match the data with the NVSS catalog and generate a new independent depolarization catalog of bright extragalactic radio sources. Unlike other polarization studies such as and our catalog is not selected based on high polarized intensity which enables us to include objects with low fractional polarizations as well. We study the evolution and possible correlation between quantities such as depolarization, spectral indices and $\textrm{RM}$s. We will tackle the nature of the well-known observed anti-correlation between total intensity and fractional polarization as well as the origin of the dominant component of depolarization. Section [2](#sec:obs){reference-type="ref" reference="sec:obs"} presents the 1.4 GHz and 2.3 GHz observations. Section [3.1](#sec:mapanalysis){reference-type="ref" reference="sec:mapanalysis"} explains the steps in our analysis of the S-PASS total intensity and polarization maps as well as the cross matching with the NVSS catalog. In Section [3.2](#quantities){reference-type="ref" reference="quantities"} we derive quantities such as spectral index, residual rotation measure, fractional polarization and depolarization. The main results and their implications are discussed in sections [4](#result){reference-type="ref" reference="result"} and [5](#discussion){reference-type="ref" reference="discussion"} respectively. At the end, Section [6](#summary){reference-type="ref" reference="summary"} summarizes the main findings and conclusions. Throughout this paper we employ the $\Lambda$CDM cosmology with parameters of H$_0=70$ km.s$^{-1}$Mpc$^{-1}$, $\Omega_m=0.3$ and $\Omega_{\Lambda}=0.7$. # Observations {#sec:obs} ## The 2.3 GHz Data {#spass} The S-PASS is a project to map the southern sky at Dec $<-1.0$ deg in total intensity and linear polarization. The observations were conducted with the 64-m Parkes Radio Telescope, NSW Australia. A description of S-PASS is given in and; here we report a summary of the main details. The S-band receiver used is a circular polarization package with system temperature T$_{sys}$ = 20 K, and beam width FHWM= 8.9 arcmin at 2300 MHz. Data were collected with the digital correlator Digital Filter Banks mark 3 (DFB3) recording the two autocorrelation products (RR\* and LL\) and their complex cross-correlation (RL\). The sources PKS B1934-638 and PKS B0407-658 were used for flux density calibration and PKS B0043-424 for polarization calibration. After Radio Frequency Interference (RFI) flagging, frequency channels were binned together covering the ranges 2176-2216 and 2256-2400 MHz, for an effective central frequency of 2307 MHz and bandwidth of 184 MHz. As described in, the observing strategy is based on long azimuth scans taken at the elevation of the south celestial pole at Parkes covering the entire Dec range (-89 deg to-1 deg) in each scan. For the current work, the spatial large scale component has been removed from each Stokes parameter, applying a high pass spatial filter to optimize for compact source finding and analysis. A median filter with a window of 45 arc-min was used. The final product was a set of 15$\times$`<!-- -->`{=html}15 deg$^2$ zenithal projection maps covering the entire sky observed by S-PASS. Final maps are convolved to a beam of FWHM = 10.75 arcmin. Stokes I, Q, and U sensitivity is better than 1.0 mJy beam$^{-1}$. Details of scanning strategy, map-making, and final maps are in and and will be presented in full details in a forthcoming paper (Carretti et al. 2016, in preparation). The confusion limit is 6 mJy in Stokes I and much lower in polarization (average polarization fraction in compact sources is around 2%, see this work). The instrumental polarization leakage is 0.4% on-axis and less than 1.5% off-axis. ## The 1.4 GHz Data The NVSS is a 1.4 GHz radio survey with the Very Large Array (VLA) covering the entire sky north of-40 degrees declination at a resolution of 45 arcsec (FWHM). The rms brightness fluctuations are approximately uniform across the sky at $\sim$`<!-- -->`{=html}0.45 mJy per beam in Stokes I and $\sim$`<!-- -->`{=html}0.29 mJy per beam in Stokes Q and U. The astrometry is accurate to within $<1$ arcsec for point sources with flux densities $>15$ mJy, and to $<7$ arcsec for the faintest detectable sources ($\sim$`<!-- -->`{=html}2.3 mJy in Stokes I). The survey has a completeness limit of 2.5 mJy, which resulted in a catalog of over 1.8 million discrete sources in Stokes I. More details about the NVSS can be found in. # Creating the new sample ## Cross-matching and selection criteria {#sec:mapanalysis} We first attempted to construct a joint S-PASS/NVSS catalog using NVSS I,Q, and U images convolved to the processed S-PASS resolution of $\sim$`<!-- -->`{=html}11'. However, upon convolution, the resulting NVSS images were very heavily mottled because of the lack of short interferometer baselines, and the noise level increased dramatically above the full resolution images. We therefore followed an alternative approach, viz., measuring the contributions of all individual NVSS sources at the position of each NVSS source, as described further below. There are rare situations where very-low level diffuse NVSS emission could also have contributed significantly to the S-PASS flux [@2001ApJ...548..639K], and would be missed by our procedure, but this very minor possible contribution to our strong total intensity sources has been ignored. We constructed the initial S-PASS catalog by searching the S-PASS maps at the position of all NVSS sources with $I_\mathrm{NV}~>~$`<!-- -->`{=html}10 mJy in the overlap region between the two surveys, and fitting Gaussian functions to the S-PASS total intensity images. For sources with a spectral index of-0.7 (-0.3) this would correspond to a 4(5) $\sigma$ detection in S-PASS. However, in order to have adequate sensitivity to sources with low fractional polarizations in S-PASS, we adopted a much higher threshold of $I_\mathrm{SP} >$ 420 mJy for the catalog. Duplicate sources were eliminated. Additional sources were eliminated from the catalog if they had either of these data quality issues:\ a) Excess noise ($>$`<!-- -->`{=html}0.75 mJy per beam rms, 1.5 $\times$ the mode calculated in bins of 0.01 mJy) in the 7.5'-11.25' annulus around the total intensity NVSS source;\ b) Excess noise ($>$`<!-- -->`{=html}3 mJy per beam rms. 2$\times$ the average rms value) in the 45'-90' annulus in either Q or U maps in S-PASS.\ We verified by visual inspection that the above selection criteria have successfully eliminated the NVSS and S-PASS regions with instrumental artifacts. At the processed S-PASS resolution of $\sim$`<!-- -->`{=html}11', many sources identified by the above procedure are actually blends of multiple NVSS sources. In order to derive meaningful information from the sample, we therefore needed to eliminate sources with significant contributions from blending. To do this, we defined a search radius of 16' (i.e., to the 3.5$\sigma$, 2$\times 10^{-3}$ level of the S-PASS beam) around each S-PASS source, and calculated the I,Q, and U contributions of each NVSS source (with $I_\mathrm{NV}>$`<!-- -->`{=html}10 mJy) at the position of the S-PASS source. Thus, for the NVSS portion of the catalog, we have two values for each Stokes parameter: X$_{Ntarget}$, the flux (I,Q, or U) of the NVSS source with the largest Stokes I contribution at the S-PASS position, and X$_{Ncont}$, the I,Q, or U flux from all other NVSS sources within the 16' search radius, scaled by their distance from the S-PASS peak position using a Gaussian kernel representing the S-PASS beam. The final values for comparison with S-PASS are then X$_{Ntotal} \equiv$ X$_{Ntarget}$ + X$_{Ncont}$. Figure [\[contI2\]](#contI2){reference-type="ref" reference="contI2"} shows the distribution of the percent contamination of the target source in NVSS total intensity, $\frac{I_{\text{cont}}}{I_{\text{target}}}$, and polarization, $\frac{P_{\text{cont}}}{P_{\text{target}}}$. We then adopted a 10% polarization contamination threshold, and only selected sources with $\frac{P_{\text{cont}}}{P_{\text{target}}} < 0.1$. The joint S-PASS/NVSS catalog contains 533 sources meeting all of the above criteria. A description of the biases that could result from our contamination threshold is discussed in Section [3.3](#bias){reference-type="ref" reference="bias"}. ## Derived quantities {#quantities} ### NVSS and S-PASS polarized flux density, fractional polarization and depolarization We calculated the polarization intensity (averaged over the entire bandwidth) for the NVSS and S-PASS surveys separately. The effect of bandwidth depolarization is discussed in section [3.2.3](#bwdepol){reference-type="ref" reference="bwdepol"}. We used Stokes $Q$ and $U$ to calculate the polarized intensity, $P$, in both NVSS and S-PASS as following: \[P=\sqrt{Q^2+U^2}\] where for NVSS the $Q$ and $U$ include both the target and contamination flux density, $Q=Q_{\text{target}}+Q_{\text{cont}}$ and $U=U_{\text{target}}+U_{\text{cont}}$. The bias corrected polarized flux density, $P_{bc}$, is approximated as follows: \[P_{\text{bc}}=\sqrt{Q^2+U^2-\sigma_{p}^2-\sigma_{\text{cont}}^2}\] where $\sigma_{p}$ is the global rms of $U$ or $Q$ maps ($\sigma^{NV}_{U,Q} \approx 0.3$ mJy per beam and $\sigma^{SP}_{U,Q} \approx 1.7$ mJy per 3-arcmin pixel), measured through the entire $Q$ and $U$ maps, and $\sigma_{\text{cont}}$ is the total contribution of the contaminant apertures rms noise to the bias in NVSS, scaled for their separation from the target. We also calculated the fractional polarization, $\pi$, \[\pi_\mathrm{SP}=\frac{P^{SP}_{\text{bc}}}{I_\mathrm{SP}}\] \[\pi_\mathrm{NV}=\frac{P^{NV}_{\text{bc}}}{I_\mathrm{NV}}\] where the NVSS total intensity is equal to the target plus the contamination flux density, $I_\mathrm{NV}=I_{\text{target}}+I_{\text{cont}}$. The NVSS residual instrumental polarization percentage peaks at $\epsilon_\mathrm{NV} \approx 0.12\%$ for a sample of strong and unpolarized sources. We used this value as a cutoff for the NVSS fractional polarization; for any sources below this threshold we report upper limits as the maximum of $\left(\frac{3\sigma_p}{I}, \epsilon_\mathrm{NV} \right)$. To estimate the S-PASS residual instrumental polarization we selected the 27 objects with $\pi_\mathrm{NV} < 0.12\%$ in our final sample, and plotted the distribution of their $\pi_\mathrm{SP}$ values (Figure [\[fig:sprespi\]](#fig:sprespi){reference-type="ref" reference="fig:sprespi"}). The median of the distribution, $\bar{\pi}_\mathrm{SP}=0.55\%$, which we assumed to be a good estimator of the S-PASS residual instrumental polarization percentage, $\epsilon_\mathrm{SP}$. Note that if the residual instrumental polarizations were zero, then the rms noise of 1.7 mJy per beam would result in much smaller fractional polarizations than $0.55\%$ for the 27 mentioned objects. On the other hand, objects with $\pi_\mathrm{NV} < 0.12\%$ can potentially be more polarized at higher frequencies, so we could be overestimating the instrumental contribution. We ignored this possibility, and chose the more conservative approach of assuming $\pi_\mathrm{SP}=0.55\%$ is only due to instrument leakage. Out of $533$ objects, 416 objects are successfully detected ($P_{\text{bc}} > 3\sigma_p$ and $\pi > \epsilon$) in both NVSS and S-PASS polarized flux densities. There are 90 sources that are not detected in polarization in S-PASS but are detected in NVSS, whereas 12 objects with no detection in NVSS polarization are detected in S-PASS. There are 15 objects that do not have polarization above our threshold in either survey. The depolarization, D, is defined to be the ratio between S-PASS and NVSS fractional polarizations: \[D \equiv \frac{\pi_\mathrm{SP}}{\pi_\mathrm{NV}}\] We calculated the depolarization of all objects with $3\sigma_p$ polarization detection and $\pi > \epsilon$ in both S-PASS and NVSS. Upper/lower limits on $D$ are also calculated for sources as appropriate. ### Polarization angle and rotation measure Assuming that the contaminating sources have very little impact on the polarization angle of the target source, we used NVSS and S-PASS $Q$ and $U$ flux densities to derive the polarization angles $\chi_\mathrm{NV}$ and $\chi_\mathrm{SP}$; where \[\chi=\frac{1}{2}\tan^{-1}\frac{U}{Q}\] These angles are used to estimate the amount of the rotation measure, $\textrm{RM}_\mathrm{NS}$, between the NVSS and S-PASS. The median uncertainty on the derived rotation measures is on the order of 1.6 rad m$^{-2}$. The polarization angle can be wrapped by a positive or negative integer coefficient, $n$, of $\pi$ radians from the true angle, the so-called $n\pi$ ambiguity. In this case, the true rotation measure is $\textrm{RM}_\mathrm{NS}=\textrm{RM}_0 \pm n\pi/\lambda^2$ rad m$^{-2}$. We used the TSS09 rotation measure catalog ( $\textrm{RM}_\mathrm{T}$) to fix $n$ by minimizing the absolute values of $\Delta \textrm{RM}$, where $\Delta\textrm{RM} \equiv \textrm{RM}_\mathrm{T}-\textrm{RM}_\mathrm{NS}-n\pi/\lambda^2$ for the 364 sources in common. These are not necessarily the correct $\textrm{RM}$s, since TSS09 has its own $n\pi$ ambiguity of 653 rad m$^{-2}$, while this ambiguity for $\textrm{RM}_\mathrm{NS}$ is about 108 rad m$^{-2}$. However, they provide the most conservative estimate of $\Delta \textrm{RM}$, the inferred non-linearity in the Faraday rotation as a function of $\lambda$. The parameter $n$ took values of $-1, 0, 1$ for all objects except one with $n=-2$. Note that, including the polarization contamination and recalculating the RMs based on the two NVSS sub-bands could introduce offsets as large as 42 rad m$^{-2}$. As a result, using the uncontaminated NVSS $\textrm{RM}_\mathrm{T}$ is appropriate. ### Bandwidth depolarization {#bwdepol} When the $\textrm{RM}$ is high, the rotation of the polarization angle across a fixed bandwidth reduces the net degree of polarization, which is called bandwidth depolarization. To evaluate the importance of this effect for our sample, we used the 364 sources overlapping with TSS09. We predicted the NVSS and S-PASS bandwidth depolarizations for our objects based on the measured TSS09 $\textrm{RM}_\mathrm{T}$ and our $\textrm{RM}_\mathrm{NS}$, respectively. As shown in Figure [\[fig:BWdep2\]](#fig:BWdep2){reference-type="ref" reference="fig:BWdep2"} the ratio between the observed fractional polarization and the true degree of polarization $\pi_\mathrm{obs}/\pi_\mathrm{true}$ never gets smaller than 0.95 for S-PASS, and only 3% of objects have NVSS $\pi_\mathrm{obs}/\pi_\mathrm{true}$ smaller than 0.9. The median $\pi_\mathrm{obs}/\pi_\mathrm{true}$ for both S-PASS and NVSS are 0.999 and 0.996 respectively, and therefore, bandwidth depolarization will not affect our analysis throughout this work. ### Spectral index We used I$_{NV}$ and peak S-PASS (11' beam) intensities, and calculated the power law spectral index, $\alpha$, where $I \propto \nu^{\alpha}$. Figure [\[fig:sphst\]](#fig:sphst){reference-type="ref" reference="fig:sphst"} shows the distribution of spectral indices for our 533 objects. The median is $\bar\alpha\sim-0.83$. The contaminating flux contributing to the NVSS intensities can be a small source of uncertainty in the calculated spectral indices; we estimated its median to be $\sigma_{\alpha,Cont} \sim 0.01$ while total uncertainties on the derived spectral indices has median value of $\sigma_{\alpha,Tot}=0.05$. ### Surface area of the object We used the NVSS catalog de-convolved minor, $\theta_m$, and major $\theta_J$ axes of the target object, and calculated the effective area, $A$ as follows.: \[A \equiv \frac{1}{4}\pi\theta_m\theta_J\] One must note that almost all sources remain unresolved in S-PASS due to the very large beam size. ### Uncertainties We used the measured local rms values as uncertainties of the NVSS $Q$, $U$ and S-PASS $I$, $Q$ and $U$ flux densities. The uncertainty of the NVSS total intensities are extracted from the NVSS catalog. Error propagation is used to approximate the uncertainty on all the other derived quantities such as polarized flux density and rotation measure. We note that showed that the rotation measure uncertainties reported in the TSS09 catalog might be underestimated. As a result, we multiplied all the $\textrm{RM}_\mathrm{T}$ uncertainties by 1.22 as described in . ## Selection Bias {#bias} We do not select objects based on their polarization intensities or fractional polarizations. However, we apply a threshold cut on the contribution of polarized contaminants. There is a higher probability for objects with low polarized intensity, either intrinsic or due to depolarization, to suffer from contaminating neighbors, and to be dropped from our final sample. To investigate a possible missing population, we compared two different sub-samples a) sources in our catalog with $\frac{P_{\text{cont}}}{P_{\text{target}}} <$ 0.1 (533 sources, 416 detected in both NVSS and S-PASS) and b) objects rejected from our catalog with 0.1 $\le \frac{P_{\text{cont}}}{P_{\text{target}}} <$ 0.25 (75 sources, 40 detected in both NVSS and S-PASS). We compared the fractional polarization and the depolarization properties of these two sub-samples. If we were not creating a selection bias, then they should have similar properties. Figure [\[polcont\]](#polcont){reference-type="ref" reference="polcont"} shows the results. Objects with larger polarization contamination have on average lower 2.3 GHz fractional polarization (median $\bar \pi_\mathrm{SP} = 1.5\%$) while less contaminated sources have $\bar \pi_\mathrm{SP}= 2.5\%$. Moreover, the fraction of sources with $\pi_\mathrm{SP} < 1\%$ is 2.5 times higher (50%) among objects with 0.1 $\le \frac{P_{\text{cont}}}{P_{\text{target}}} <$ 0.25 than sources with $\frac{P_{\text{cont}}}{P_{\text{target}}} <$ 0.1. The Spearman rank test between $D$ and $\frac{P_{\text{cont}}}{P_{\text{target}}}$ with $r= 0.22$ and $p<0.00001$ rejects the null hypothesis of no correlation. Thus, we are likely to be missing a population of highly depolarized sources. Figure [\[polcont\]](#polcont){reference-type="ref" reference="polcont"} suggests that around 30% of sources with polarized contamination 0.1 $\le \frac{P_{\text{cont}}}{P_{\text{target}}} <$ 0.25 have depolarizations $\log(D)>0.47$. Assuming this fraction is also valid for sources with contaminations larger than 25% we estimate that we have missed $\sim 50$ depolarized objects in our final sample due to the polarized contamination threshold cut. Therefore, our final sample of 533 sources is missing a population ($\sim 50$ objects) of heavily depolarized sources due to our contamination threshold cut. However, we can not correct for such an effect since the amount of contamination in our 2.3 GHz polarization intensities can not be measured. As a result, one should treat the number of depolarized sources in our sample as a strong lower limit and consider this in interpreting all the other related conclusions. In addition, it is possible that our total intensity and polarization contamination thresholds have resulted in a bias toward less dense regions of the sky. We measured the surface number density of the contaminating neighbors in our sample and the parent NVSS--S-PASS overlap sample with $I_\textrm{NV} > 10$ mJy. We used the same aperture with a radius of 16 arcmin and found that the contaminant surface number density in our final sample ($4\times 10^{-3}$ arcmin$^{-2}$) is on average 20% less than our parent sample ($5\times 10^{-3}$ arcmin$^{-2}$). It is unlikely that the results of this work are affected by such a bias. ## Statistical tests {#stat} Throughout this work we adopted two nonparametric statistical tests. We calculated the Spearman rank correlation coefficient ($r_s$) to measure the strength of any possible correlation. The two-sample Kolmogorov-Smirnov (KS) test is also used to check the null hypothesis that two sub-samples, divided by a parameter of interest, are drawn from the same parent distribution. The significance of each test is estimated by performing bootstrap sampling simulations and constructing $10^5$ random samples from the initial distribution. We have assigned two-tail p-values based on the results of our simulations. Table [\[table2\]](#table2){reference-type="ref" reference="table2"} summarizes the result of all the statistical tests performed in this work. In the case of a single hypothesis test we would reject the null hypothesis if the p-value $\le0.01$. However, we have performed a total 90 tests, counting both KS and Spearman. To avoid the multiple hypothesis testing problem, we adopted the Bonferroni correction as discussed in and chose a conservative significance level threshold of p-value $\le 10^{-4}$. We therefore rejected the null hypothesis of the KS or the Spearman rank tests only if the corresponding p-value is less than or equal to $10^{-4}$. In addition, to test the robustness of correlations with p-value less than $10^{-4}$ and to identify any possible influence of the total intensity and polarization contaminations on the results we repeated the relevant statistical tests on smaller (by a factor of $\sim$`<!-- -->`{=html}0.4) but clean samples of objects with less than 1% contamination. Although the strength of some correlations became stronger or did not change, their p-values increased up to $2\times 10^{-3}$ due to the much smaller sample size. We therefore, adopted the robustness probability p$_\textrm{robust}$ of $2\times 10^{-3}$ as a second threshold and treated the correlations with original p-value$\le10^{-4}$ and $2\times 10^{-3} <$p$_\textrm{robust}< 0.05$ as suggestive relations only, and did not draw any conclusion based on them. These are marked in Table [\[table2\]](#table2){reference-type="ref" reference="table2"} for completeness and potentially future work. The correlations with original p-value$<10^{-4}$ but p$_\textrm{robust}> 0.05$ are rejected. # Results {#result} We have derived a polarization catalog of 533 extragalactic radio sources, which can be downloaded for public use through the VizieR catalog access tool. The description of the entries in the online catalog is listed in table [\[table3\]](#table3){reference-type="ref" reference="table3"}. ## Rotation measures {#rm0} The distribution of $\textrm{RM}_\mathrm{NS}$ calculated based on NVSS and S-PASS (black) and the Taylor et al. rotation measures, $\textrm{RM}_\mathrm{T}$, (red) for the same objects are shown in Figure [\[fig:rmhist\]](#fig:rmhist){reference-type="ref" reference="fig:rmhist"}. Both distributions are very similar in shape. Their medians are $3.6\pm 2.0$ and $0.5 \pm 1.9$ rad m$^{-2}$, respectively, while their standard deviations are 38.4 and 36.4 rad m$^{-2}$. Some of the scatter in the $\textrm{RM}$ distributions could be due to the uncertainty of the measurements. . However, the median error on the $\textrm{RM}_\mathrm{T}$ for the small bright sample of 364 objects in this work is only $\sigma_\mathrm{T}=3.5$ rad m$^{-2}$. The median measurement uncertainty estimated for $\textrm{RM}_\mathrm{NS}$ is even smaller, $\sigma_\mathrm{NS}=1.6$ rad m$^{-2}$. Subtracting the median errors from the observed standard deviation of the $\textrm{RM}$ distributions in quadrature result in residual standard deviations of $36.2$ rad m$^{-2}$ and $38.36$ rad m$^{-2}$ for $\textrm{RM}_\mathrm{T}$ and $\textrm{RM}_\mathrm{NS}$ respectively, and largely represent the spread in Galactic foregrounds. ## Distribution of fractional polarization and depolarization {#D_distribution} The median NVSS (S-PASS) fractional polarization of all 533 objects is $\bar{\pi}=0.017$ $(0.020)$ including the upper limits. There are $505$ ($428$) objects with detected NVSS (S-PASS) polarization ($P > 3\sigma_p$ and $\pi > \epsilon$). However, 416 of these objects are detected in both NVSS and S-PASS. The distributions of NVSS and S-PASS fractional polarization of these 416 objects are shown in Figure [\[fig:fphist\]](#fig:fphist){reference-type="ref" reference="fig:fphist"}. The median (and standard deviation) of NVSS and S-PASS fractional polarization of these common objects are 0.022 (.022) and 0.025 (0.023), respectively. Although the median values of $\pi_\mathrm{SP}$ and $\pi_\mathrm{NV}$ are very close, the median value of their ratio (the median depolarization) is not necessarily equal to one. The TSS09 catalog was limited to sources with sufficient signal:noise in polarization, and is thus biased towards much higher fractional polarizations (median $\bar{\pi}_T \sim 0.06$) than our catalog, which is $\sim$`<!-- -->`{=html}3.5 times lower, including both measurements and upper limits. Figure [\[fig:depol\]](#fig:depol){reference-type="ref" reference="fig:depol"} shows the normalized distribution of $\log(D)$ for steep and flat spectrum sources separately. Objects with both S-PASS and NVSS detected polarizations are shown in solid black, and have median depolarizations of $\bar{D}=1.4$ and $\bar{D}=0.9$ for 315 steep and 101 flat sources respectively. The depolarization distribution of steep spectrum sources is skewed toward large values of $D$. Almost 28% of steep spectrum (24% of all) objects have $D \ge 2$, and only 2% have $D \le 0.5$. On the other hand, flat spectrum sources include both depolarized and re-polarized objects. There are 17% and 13% of flat spectrum sources with $D \ge 2$ and $D \le 0.5$ respectively. The results of the statistical tests presented in Table [\[table2\]](#table2){reference-type="ref" reference="table2"} confirm that steep and flat spectrum sources do not have the same depolarization distributions. The red dashed histogram in Figure [\[fig:depol\]](#fig:depol){reference-type="ref" reference="fig:depol"} shows the normalized distribution of 58 steep spectrum and 31 flat spectrum objects with upper limits on the depolarization. These sources have S-PASS polarizations less than $3\sigma$ or $\pi_\mathrm{SP} < \epsilon_\mathrm{SP}$ but are detected in NVSS polarization. The 12 steep spectrum objects with NVSS $P < 3\sigma_p$ or $\pi_\mathrm{NV} < \epsilon_\mathrm{NV}$ and detected S-PASS polarization are treated as lower limits on the depolarization. The dotted dashed blue line show the distribution of the lower limits in Figure [\[fig:depol\]](#fig:depol){reference-type="ref" reference="fig:depol"}. In total, 16 objects are detected in neither NVSS nor in S-PASS polarizations and we do not show them in Figure [\[fig:depol\]](#fig:depol){reference-type="ref" reference="fig:depol"}. used their multi wavelength polarization spectra and derived an equivalent power law polarization spectral index $\beta$, where $\pi \propto \lambda^{\beta}$. As long as the power law model is assumed our depolarization parameter $D$ and $\beta$ are related such that $\log(D)=\log(\frac{\lambda_\mathrm{SP}}{\lambda_\mathrm{NV}}) \beta$ where the $\lambda_\mathrm{SP}$ and $\lambda_\mathrm{NV}$ are the average wavelengths of the S-PASS and NVSS surveys respectively. found weak evidence of a bimodal distribution for $\beta$ of steep spectrum objects. We do not see any sign of bimodal depolarization within objects with $\alpha <-0.5$, as shown in Figure [\[beta\]](#beta){reference-type="ref" reference="beta"}. The $\beta$ distribution of steep spectrum objects is single-peaked but asymmetric with a longer tail toward depolarized objects. As will be discussed later, the majority of steep spectrum sources in our sample can be classified as IR AGNs according to their infrared colors. A more complete sample which also includes radio galaxies with infrared colors of normal ellipticals can confirm if the weak bimodal depolarization observed by is real. We also looked at the combined sample of steep and flat spectrum sources and classified them into three depolarization categories. The choice of the depolarization boundaries is somewhat arbitrary. However, we designed the three depolarization categories to isolate the peak observed in Figure [\[fig:depfp\]](#fig:depfp){reference-type="ref" reference="fig:depfp"}, as is discussed below. Sources with $0.6 < D < 1.7$ have median spectral index of $\bar{\alpha} \sim-0.82$ while sources with $D \ge 1.7$ shows a slightly steeper median spectrum with $\bar{\alpha} \sim-0.9$. The spectral slope is mostly flat for re-polarized objects with $D \le 0.6$, with a median $\bar{\alpha} \sim-0.1$. However, there are 14 re-polarized objects with steep spectral indices, $\alpha <-0.5$. This is consistent with who also found a small population of steep spectrum re-polarized sources. Figure [\[sphstrep\]](#sphstrep){reference-type="ref" reference="sphstrep"} shows the distribution of the spectral indices of re-polarized objects. We also included 24 objects with detection in $\pi_\mathrm{NV}$ but only upper limits on $\pi_\mathrm{SP}$. Figure [\[sphstrep\]](#sphstrep){reference-type="ref" reference="sphstrep"} suggests there are two separate populations of re-polarized sources with flat and steep spectra. Including the mentioned upper limits on $D$, 61% of re-polarized sources have $\alpha \ge-0.5$ (i.e., flat). To understand the relation between fractional polarization and depolarization, we plotted $\pi_\mathrm{SP}$ versus $\log(D)$, and calculated the running medians in bins of 30 objects (Figure [\[fig:depfp\]](#fig:depfp){reference-type="ref" reference="fig:depfp"}). There is an apparent peak for S-PASS fractional polarization at $log(D )\sim 0$, while both depolarized and re-polarized sources show weaker $\pi_\mathrm{SP}$ than sources with fractional polarization above 6%. Both KS and Spearman rank coefficient tests on the $\|\log(D)\|$ and $\pi_\mathrm{SP}$ confirm this anti-correlation. We also used two subsamples with $\log(D) >0$ and $\log(D) <0$ and performed the two KS and Spearman rank tests on each subsample separately. The results confirmed that fractional polarizations are higher in the vicinity of $\log(D)\sim 0$ in each subsample. However, the correlation between $\|\log(D)\|$ and $\pi_\mathrm{SP}$ of the subsample with $\log(D) >0$ became uncertain when only including the contamination clean sample of the robustness test. Table [\[table2\]](#table2){reference-type="ref" reference="table2"} summarizes the results of these statistical tests. Figure [\[fig:fphist23\]](#fig:fphist23){reference-type="ref" reference="fig:fphist23"} shows the S-PASS (top) and NVSS (bottom) fractional polarization distributions for three sub-samples with $\|\log(D)\| \le 0.23$, $\log(D) > 0.23$ and $\log(D)<-0.23$. Objects with $\log(D) \sim 0$ have almost the same distribution in both S-PASS and NVSS (by construction) with median fractional polarizations of $\bar{\pi}_\mathrm{SP}=0.030$ and $\bar{\pi}_\mathrm{NV}=0.028$ while depolarized sources have smaller medians, $\bar{\pi}_\mathrm{SP}=0.024$ and $\bar{\pi}_\mathrm{NV}=0.009$ with an offset between NVSS and S-PASS as expected. Objects with re-polarization show more complicated behavior. They have a median $\bar{\pi}_\mathrm{SP}=0.015$ and $\bar{\pi}_\mathrm{NV}=0.035$. By definition the median degree of polarization of a sample of re-polarized sources is expected to be higher at 1.4 GHz than 2.3 GHz. It is possible that the true median $\bar \pi_\textrm{SP}$ and $\bar \pi_\textrm{NV}$ are lower than the above values because we would have systematically excluded re-polarized objects with $\pi_\mathrm{SP}$ less than the detection limit. This results in over estimating the median fractional polarization of re-polarized sources in both NVSS and S-PASS. ## Total intensity and fractional polarization {#ip} Our sample includes total intensities from 0.42 to 10 Jy, which gives us the opportunity to study possible correlations between the fractional polarization and total intensity. As listed in Table [\[table2\]](#table2){reference-type="ref" reference="table2"} both KS and Spearman tests suggest there is a weak anti-correlation between $\pi_\mathrm{SP}$ and $I_\mathrm{SP}$ of the whole sample of sources at 2.3 GHz. More investigation revealed that is true for steep spectrum ($\alpha <-0.5$) sources alone, while it disappears for flat spectrum ($\alpha \ge-0.5$) objects. The anti-correlation among steep spectrum sources became weaker and more uncertain when only including the contamination clean sample of the robustness test, and thus should be treated as a suggestive trend only. Figure [\[fig:Ifp\]](#fig:Ifp){reference-type="ref" reference="fig:Ifp"} shows the S-PASS $\pi_\mathrm{SP}$ of only steep spectrum sources versus their logarithm of total intensity. The calculated running medians (including the upper limits on $\pi_\mathrm{SP}$ to avoid any selection bias due to our total intensity cut) are shown as well. Objects with $\alpha <-0.5$ and $\log(I_\mathrm{SP}) < 2.9$ have median of $\bar{\pi}_\mathrm{SP} \sim 0.03$ while sources with larger total intensity are less polarized with medians of $\bar{\pi}_\mathrm{SP} \sim 0.02$. To shed light on a possible physical origin of the observed anti-correlation we calculated the luminosities, based on the 261 objects in our sample which have redshifts in the catalog. 222 of these sources are detected in both NVSS and S-PASS polarization. Using our spectral indices, we calculated the K-corrected 2.3 GHz luminosities. The 141 steep spectrum objects have median luminosity of $L_{steep}=1.7\times 10^{27}$ WHz$^{-1}$. Although there is a nominal difference between $\bar{\pi}_\mathrm{SP}$ for higher and lower luminosities (2.6% and 2.2%, respectively), these do not appear statistically significant. There is also no statistically significant difference in $\|\log(D)\|$ for the high and low luminosity steep spectrum sources. The 81 flat spectrum sources are at higher redshifts, on average, and have a median luminosity of $\bar L_{flat}=3.0 \times 10^{27}$ W Hz$^{-1}$. ## Correlation between $\textrm{RRM}$, $\Delta \textrm{RM}$, $\pi$ and $D$ {#deltarm} There are two measures to characterize the Faraday effects that are either local to the source or in the intervening IGM medium, the residual rotation measure $\textrm{RRM}$, which takes out the Galactic foreground contribution to the observed $\textrm{RM}$, and $\Delta \textrm{RM} \equiv \textrm{RM}_\mathrm{T}-\textrm{RM}_\mathrm{NS}$, which sheds light on the frequency dependency of the $\textrm{RM}$. The absolute value of $\|\Delta RM\|$ is an indicator of the Faraday complexity of the source and its environment. As explained in the following, we found that $\Delta \textrm{RM}$ is anti-correlated with $\pi$ and correlated with $\|\log(D)\|$. Faraday complex sources, i.e, those with multiple $\textrm{RM}$ components should be both depolarized and have polarization angles which may not vary linearly with $\lambda^2$. We therefore examined the possible correlation between depolarization and $\|\Delta \textrm{RM}\|$. Figure [\[fig:drm2new\]](#fig:drm2new){reference-type="ref" reference="fig:drm2new"} shows $\|\Delta \textrm{RM}\|$ versus $\|\log(D)\|$ for all objects with detected polarization in both NVSS and S-PASS. The running medians of the $\|\Delta \textrm{RM}\|$ calculated in bins of $\|\log(D)\|$ show an evolution. To quantify this, we calculated the Spearman rank, which yielded a correlation coefficient of $r_s=0.23$ and p-value of $p = 0.00003$ establishing that depolarization and non-$\lambda^2$ polarization angle behavior are related. A large $\textrm{RM}$ beyond the Galactic foreground $\textrm{RM}$ screen could also indicate the presence of Faraday complexity and depolarization. To estimate this, we removed the Galactic contribution by subtracting the median $\bar{\textrm{RM}}$ within 3 degrees of each target (excluding the target itself), using the TSS09 catalog. This yields the residual rotation measure, $\textrm{RRM}_\mathrm{T} \equiv \textrm{RM}-\bar{\textrm{RM} }$. Subtracting the median $\bar{\textrm{RM}}$ is not the best method to estimate the extragalactic component of the $\textrm{RM}$ as discussed in. However, for objects above the Galactic latitude of $\|b\| > 20$ degrees, which is true for most of our sample, the difference between the recipe and our method is small. As shown in Figure [\[rrmdep\]](#rrmdep){reference-type="ref" reference="rrmdep"}, we find the Spearman rank coefficient of $r_s=0.21$ and p-value of $7\times10^{-5}$ which suggests a correlation between $\|\textrm{RRM}_\mathrm{T}\|$ and $\|$log(D)$\|$. However, our robustness test on the clean sample failed to confirm such a trend. Thus, only $\|\Delta \textrm{RM}\|$ shows a clear sign of a correlation with depolarization. We also found, the 1.4 GHz and 2.3 GHz fractional polarizations show moderate anti-correlations with $\|\Delta \textrm{RM} \|$, as shown in Figure [\[drmp\]](#drmp){reference-type="ref" reference="drmp"} and listed in Table [\[table2\]](#table2){reference-type="ref" reference="table2"}. Thus, depolarization does reduce the fractional polarizations at these frequencies, although the dominant role of field disorder is discussed in Section [5.1](#obs){reference-type="ref" reference="obs"}. Moreover, the Spearman rank test with $r_s=-0.25$ and p-value of $<10^{-5}$ suggest an anti-correlation between $\|\textrm{RRM}_\mathrm{T}\|$ and $\pi_\textrm{NV}$. However, our robustness test failed to confirm this significance. ## Polarization, depolarization and the object angular extent {#extent} To study how the morphology of a system affects the depolarization, we used total intensity deconvolved areas ($A$) derived from the NVSS catalog. Flat spectrum objects in our sample are unresolved in the NVSS synthesized beam while steep spectrum objects include both resolved and unresolved sources. For the steep spectrum sources, Figure [\[fig:dsize\]](#fig:dsize){reference-type="ref" reference="fig:dsize"} shows the distributions of the absolute $\|\log(D)\|$ for two sub-samples-unresolved and resolved sources with the dividing line at $\log(A) = 2.5$ arcsec$^2$. On average, resolved sources have smaller $\|\log(D)\|$ with median of $0.12$ compared to $0.20$ for unresolved sources. The scatter of the two samples is almost the same with standard deviation of $0.21$. Beam depolarization should only play a small role, because most resolved sources are only slightly resolved. We also looked at the dependence of fractional polarization on size. Figure [\[fig:sizefp\]](#fig:sizefp){reference-type="ref" reference="fig:sizefp"} shows the distributions of the S-PASS fractional polarization for the unresolved and extended samples of steep spectrum objects. On average, resolved and extended steep spectrum objects have 2.3 GHz fractional polarizations, $\bar \pi_\mathrm{SP} \sim 4\%$, two times larger than their unresolved counterparts. Both KS and Spearman tests confirm a strong strong positive correlation between $A$ and $\pi_\mathrm{SP}$ of steep spectrum objects. ## Spatial distribution of depolarization in the sky {#galD} We carried out a brief investigation to see if the depolarization properties in our sample were related to their position in Galactic coordinates. Figure [\[fig:galactic\]](#fig:galactic){reference-type="ref" reference="fig:galactic"} shows the distribution of 533 objects in the sky, color coded with respect to their depolarization. Visual inspection does not reveal any obvious over-density of depolarized or re-polarized objects. We also calculated the auto correlation between depolarization and angular separation, and the two point angular correlation function for the most depolarized and least depolarized sources. None of these showed any evidence for clustering of depolarization in space. Similarly, the two point angular correlation functions for the highest and lowest fractional polarizations at 2.3 GHz revealed no clustering. Other work has identified some positional dependence to polarizations in the NVSS catalog. discovered regions with angular scales of $\sim 10$ degrees in which the density of the polarized sources drops by a factor of 2-4. They named these regions the "polarization shadows," and found that some of them are associated with the Galactic HII regions while the rest are related to the depolarized areas in the diffuse Galactic radio emission. All polarization shadows in are located within the Galactic plane at $\|b\| < 20$ degrees except one which is at Galactic ($l=5$, $b=+24$). Almost all of the objects in our sample have Galactic latitudes of $\|b\| > 20$ degrees, and none are located around ($l=5$, $b=+24$), so the Galactic polarization shadows likely do not affect the current work. However, it is interesting to search for high latitude Galactic diffuse emissions in smaller scales and their probable signature on the depolarization of the extragalactic sources in future surveys and larger samples with higher number density. ## WISE colors and polarization {#spwise} We matched our catalog to the Wide-field Infrared Survey Explorer, WISE, catalog,, with a search radius of five arc-seconds. Out of 533 objects, 455 have WISE counterparts. All of them are detected with at least $5\sigma$ in the WISE $3.4 \mu m$ band, W1, while 445 (323) have $> 5\sigma$ detection in $4.6 \mu m$, W2, ($12 \mu m$, W3) band. $W1-W2$ and $W2-W3$ colors can be used to separate different galaxy populations such as AGNs and ellipticals. Recently, studied WISE colors of a large sample of resolved radio galaxies from the Radio Galaxy Zoo project, and found that most radio objects can be classified as ellipticals, AGNs and LIRGs. Figure [\[fig:wise5arc2\]](#fig:wise5arc2){reference-type="ref" reference="fig:wise5arc2"} shows the WISE color-color diagram of objects in our sample for which we have depolarization measurements and WISE counterparts. All objects used in Figure [\[fig:wise5arc2\]](#fig:wise5arc2){reference-type="ref" reference="fig:wise5arc2"} have W1 and W2 detections larger than $5\sigma$ and with small errors in the $W_2-W_3$ colors $\sigma_{(W2-W3)} < 0.4$. We investigated the possible dependence of the polarization and depolarization on WISE colors. The WISE dependence is difficult to isolate, since flat and steep spectrum objects have different WISE and different depolarization distributions. We therefore looked at steep spectrum objects only, and found that neither $\pi_\mathrm{SP}$ or $\|\log(D)\|$ were significantly correlated with WISE colors (Table [\[table2\]](#table2){reference-type="ref" reference="table2"}). We do not sample the "elliptical" region of WISE color space, which makes up a distinct population in the study. ## Redshift Dependence {#zev} There are 222 objects in our sample that are detected in both NVSS and S-PASS polarization maps and have redshifts in catalog. Figure [\[zhist\]](#zhist){reference-type="ref" reference="zhist"} shows the redshift distribution of the 222 matched sources, as well as the separated distributions of steep and flat spectrum objects. Steep spectrum objects are located within $0 \le z \le 2.34$ with median redshift of $\bar z=0.64$ while flat spectrum sources, as expected for a flux limited sample, tend to have larger redshifts, $0.22 \le z \le 2.81$, with median of $\bar z=1.18$. As discussed earlier, steep and flat spectrum objects have different depolarization distributions and therefore, we studied their redshift evolution separately. We examined the redshift dependence of only depolarized steep spectrum sources ($D \ge 1.5$), since we expected to see a change in polarization properties due to the change in rest frame wavelength. We used the threshold $D=1.5$ to choose as many highly depolarized sources as possible while excluding the scattered objects that are in the vicinity of the observed peak at $D\sim1$ in Figure [\[fig:depfp\]](#fig:depfp){reference-type="ref" reference="fig:depfp"}. We found weak evidence for a decrease in depolarization of these sources as redshift increases (Spearman $r_s=-0.36$, p=0.011), which does not cross our conservative detection threshold. The average $\pi_\mathrm{NV}$ of 49 steep spectrum sources with $D \ge 1.5$ seems to increase from $\bar\pi_\mathrm{NV}=0.46\%$ at $z \le 0.5$ to $\bar\pi_\mathrm{NV}=1.02\%$ at $z \ge 0.5$, while their observed depolarization decreases and $\pi_\mathrm{SP}$ stays almost fixed. Figure [\[depolsample\]](#depolsample){reference-type="ref" reference="depolsample"} shows the running median of $\pi_\mathrm{NV}$ and $D$ calculated in bins of redshift as well as the expected evolutionary behavior of the three depolarizing scenarios. We will discuss this more in Section [5.5](#redshift){reference-type="ref" reference="redshift"}. On the other hand, we do not find any change with redshift in depolarization of separate samples of steep or flat spectrum sources which include all re-polarized and depolarized sources. The median, $\log(D) \approx 0.1$, and standard deviation $\sigma_{\log(D)} \approx 0.26$, of steep spectrum objects stay almost constant with increasing redshift. Flat spectrum sources appear to be mostly re-polarized at $z<1$ while at higher redshifts the number of re-polarized and depolarized flat spectrum objects are almost the same. However, as listed in Table [\[table2\]](#table2){reference-type="ref" reference="table2"}, none of the KS and Spearman tests could confirm such a redshift dependence among flat spectrum sources. We also performed both KS and Spearman rank tests on $\|\textrm{RRM}\|$ and $\|\Delta \textrm{RM}\|$, and did not detect any noticeable redshift dependence (Figure [\[zrrmt\]](#zrrmt){reference-type="ref" reference="zrrmt"}). The 2.3 GHz fractional polarization of steep and flat spectrum sources also stays fixed at all cosmic times, although have different average values for populations of steep and flat objects. ## Summary of major results 1. The majority of extragalactic radio objects with $I_\mathrm{SP} \ge 420$ mJy have degrees of polarization on the order of 2% to 3% at both 1.4 GHz and 2.3 GHz. 2. $\pi_\mathrm{SP}$ and $\|\log(D)\|$ are anti-correlated. On average, objects that are not depolarized ($\|\log(D)\| \le 0.23$), have median fractional polarizations of $\bar \pi_\mathrm{SP} \approx \bar \pi_\mathrm{NV} \approx 3\%-4\%$, with $\bar \pi_\mathrm{SP} \approx 2\%$ for more depolarized objects and $\bar \pi_\mathrm{SP} \approx 1\%$ for re-polarized sources. Objects with high fractional polarizations ($\pi_\mathrm{SP}\approx \pi_\mathrm{NV} \approx 10\%$) are not depolarized ($\|\log(D)\|\approx 0$). 3. Flat and steep spectrum objects have different polarization properties. 55% of flat spectrum sources are re-polarized, compared to only 24% for steep spectrum sources. Steep spectrum sources have larger degrees of polarization as well as stronger average depolarization. 4. Extended objects ($>20''$) have higher fractional polarizations ($\bar{\pi}_\mathrm{SP}=4\%$) and smaller depolarizations ($\|\log(D)\| \sim 0.13$) than compact sources ($\bar{\pi}_\mathrm{SP} \sim 2 \%$, $\|\log(D)\| \sim 0.20$). 5. Almost 24% of the objects with detected polarization have $D > 2$. An additional 10% of all sources may be too depolarized to be included in our sample. 6. On average, sources with large $\|\log(D)\|$ (depolarized or re-polarized) show larger changes in $\textrm{RM}$ with wavelength ($\Delta \textrm{RM}$). 7. We find weak evidence for a redshift dependence of the depolarization in a sub-sample of sources, those with steep spectra and $D\ge1.5$. 8. We do not find any evidence for changes of the observed 2.3 GHz fractional polarization, depolarization, $\|\textrm{RRM}_\mathrm{T}\|$ and $\Delta \textrm{RM}$ from $z=0$ to $z=2$ when all sources are considered. The median degree of polarization of both steep (141) and flat (81) spectrum sources with known redshift remain almost constant at $\pi_\mathrm{SP} \approx 2.5 \%$ and $\pi_\mathrm{SP} \approx 2.0 \%$ respectively. 9. A large scatter in both depolarization and fractional polarization is seen at all redshifts. 10. We did not find any evidence for angular clustering in the distribution of the depolarized sources. 11. Both $\pi$ and $\|\log(D)\|$ of steep spectrum sources are independent of WISE $W_1-W_2$ color. # Discussion ## Radio source field disorder {#obs} While radio synchrotron radiation can potentially be highly polarized, the NVSS and S-PASS fractional polarizations of most objects in our sample are around $2\%-3\%$, and very rarely exceed $10\%$ (Figure [\[fig:fphist\]](#fig:fphist){reference-type="ref" reference="fig:fphist"}). Depolarization due to the presence of an irregular Faraday screen between the source and the observer, e.g., can potentially reduce the initial degree of the polarization, generally leading to higher fractional polarizations at higher frequencies. However, between 1.4 GHz and 2.3 GHz we find that the majority of extragalactic objects experience only small depolarizations, with $60\%$ of the objects have $0.6< D < 1.7$. Moreover, objects with the strongest fractional polarizations ($\pi\approx 10\%$) have little depolarization. The reduction from a theoretical maximum of $\sim$`<!-- -->`{=html}40-70% to either $\approx$`<!-- -->`{=html}10% with no depolarization, or $\approx$`<!-- -->`{=html}3%, with modest depolarization, must therefore be due to field disorder. To approximate the necessary number of randomly oriented magnetic field patches within an unresolved source, we performed a simple simulation. We considered a uniform brightness two dimensional source, with equal fractional polarizations $\pi_0=50\%$ in each patch. By randomizing the polarization angles, we estimated that sources currently unresolved in our beam should contain approximately 70 to 80 independent magnetic patches to reduce the observed fractional polarization to $\sim$`<!-- -->`{=html}4%. There is a subset of sources where depolarization does play a significant role. Almost, 24% of sources with detected polarizations have $D>2$. Moreover, we estimated a missing $\approx$`<!-- -->`{=html}10% population of heavily depolarized sources. It is not clear how strong an effect field disorder has for that subset. ## Prospects for high frequency surveys One important implication of these results is for surveys at higher frequencies, where one might expect to increase number counts by a large factor because of less depolarization. However, changing the frequency of observation from L to S band will not result in a major increase in the number of polarized detections. the number of polarized objects. As an example, the number of sources with polarized flux densities larger than 10 mJy in our sample is almost equal at both 2.3 GHz and 1.4 GHz (368 in S band and 363 in L band). Future polarization surveys and the Square Kilometer Array, SKA precursors such as Polarization Sky Survey of the Universe's Magnetism, POSSUM, Westerbork Observations of the Deep APERTIF Northern sky, WODAN, MeerKAT International GigaHertz Tiered Extragalactic Exploration survey, MIGHTEE, Very Large Array Sky Survey, VLASS and VLASS Deep will detect hundreds of thousands of polarized sources in different frequencies. The VLASS will operate at S band from 2 to 4 GHz and has angular resolution and sensitivity of $\sim 3.5$ arcsec and 0.7 mJy per beam respectively. The number density of flat spectrum sources is expected to be similar in L and S bands since their flux density is almost independent of the frequency, and their median depolarization is $\bar D \sim 1$ as shown in Figure [\[fig:depol\]](#fig:depol){reference-type="ref" reference="fig:depol"}. On the other hand, steep spectrum, $\alpha <-0.5$, sources in our sample with median $\bar{\alpha}=-0.9$ are on average fainter at S band by a factor of 1.4. Therefore, their number density at a fixed signal to noise reduces. However, the median polarization of steep spectrum objects in our sample is approximately 1.3 times higher at 2.3 GHz than 1.4 GHz at resolutions as low as S-PASS, $\sim 9$ arcmin. This indicates that the median polarization flux density of these objects should have been reduced by $\sim$`<!-- -->`{=html}17%. showed at 1.6 arcsec resolution there are $\sim 6$ polarized sources per squared degree at 0.7 mJy per beam and S:N $>$ 10 in L band, and the integrated number density of objects with polarization flux density larger than $p$ goes as $N_p\propto p^{-0.6}$. As a result, one can expect to detect roughly 11% less polarized objects at S band compared to L band at 1.6 arcsec resolution. All in all, considering the larger beam size of the VLASS all sky survey one can expect to detect approximately the same number of polarized sources in S band as the calculation of in L band. This is already a factor of six above the existing surface density of polarized sources from the NVSS catalog in L band. ## Prospects for $\textrm{RM}$ grid experiments There is strong interest in measuring and estimating the intergalactic magnetic field in clusters of galaxies or in cosmic filaments through $\textrm{RM}$ analysis and tomography, e.g. . In the presence of a single Faraday screen along the line of sight, the rotation angle of the radio polarization vector of extra-galactic sources depends linearly on $\lambda^2$. This simple relation makes it possible to estimate the magnetic field of the medium with some assumptions for the electron density, after subtracting out a Galactic component. However, any complication in the structure of the Faraday screen within the observation beam or along the line of sight through the emitting source will result in non-$\lambda^2$ behavior, and an inability to isolate the foreground screen of interest. We have measured the non-$\lambda^2$ behavior using $\Delta \textrm{RM}$. As shown in Figure [\[drmp\]](#drmp){reference-type="ref" reference="drmp"}, large $\Delta \textrm{RM}$s occur preferentially at low fractional polarizations. In order to avoid large values of $\Delta \textrm{RM}$, which would compromise any foreground experiment, it is necessary to use only fractional polarizations ($\ge 3-4\%$). This will cause a reduction in the number of available sources; only 33% of sources in our sample have $\pi_{SP} >$ 3%. However, if reliable $\chi(\lambda^2)$ were available for some subset of sources, then it might be possible to increase this number. ## Origins of depolarization As shown in Section [4.6](#galD){reference-type="ref" reference="galD"} we did not detect any angular clustering of sources by fractional polarization or depolarization, that would have implied a Galactic origin. We can not rule out the possibility of Galactic $\textrm{RM}$ fluctuations on arcsec scales, but these are likely to be extremely small and we do not consider them further here. The dependence of depolarization on spectral index shows that it must primarily occur local to the source. If depolarization is local to the environment of the source, then it may show signs of dependence to some intrinsic characteristics of the source such as spectral index or the luminosity. The results found here on the spectral behavior are consistent with who did a multi-wavelength polarization study on sources selected from the TSS09 catalog. The dependence of polarization properties of objects on their angular extent (Section [4.5](#extent){reference-type="ref" reference="extent"}) also supports the local depolarization scenario. As shown in Figure [\[fig:dsize\]](#fig:dsize){reference-type="ref" reference="fig:dsize"}, compact sources seem to have larger depolarizations ( $\|\log(D)\| \sim 0.20$ vs. $\sim 0.13$) and smaller fractional polarizations ($\bar{\pi}_\mathrm{SP}=4\%$ vs. 2%) than sources extended in NVSS. This is inconsistent with irregular screens either Galactic or extragalactic, which should yield higher fractional polarizations and less depolarization for compact sources. Thus, the depolarization must arise in a Faraday component directly related to the source. If Galactic or intervening Faraday screens were the dominant depolarizing components then we expect to see larger depolarization in a sample of extended sources. ### The origin of the total intensity and fractional polarization anti-correlation {#origin} The anti-correlation between total intensity and fractional polarization at 1.4 GHz has been extensively discussed (such as @2002A&A...396..463M [@2004MNRAS.349.1267T; @2007ApJ...666..201T; @2010ApJ...714.1689G; @2010MNRAS.402.2792S; @2014ApJ...787...99S]). Recently, used WISE colors to suggest that the anti-correlation was due to the difference in environments between WISE-AGNs (IR colors dominated by AGN) and WISE-Ellipticals (IR colors dominated by starlight). These effects are likely confused by the fact that the anti-correlation is found only among steep-spectrum sources, as discussed in Section [4.3](#ip){reference-type="ref" reference="ip"}. The WISE-AGN class contains a large fraction of flat spectrum objects, for which we find no anti-correlation, while the WISE-Ellipticals are largely steep-spectrum. The dependence we found on the spectral index is also consistent with and the stacking analysis of. The limited range of $I_\textrm{SP}$ in our sample makes it difficult to study these effects. However, to illuminate the underlying issues, we note that the suggestive anti-correlation between $I_\mathrm{SP}$ and $\pi_\mathrm{SP}$ of steep spectrum sources must arise from some physical difference in properties between the bright and faint sources that are not expected in fair, uniform samples. We have not been able to identify this underlying parameter. We find no statistically significant anti-correlation between $L_\mathrm{SP}$ and $\pi_\mathrm{SP}$. We attempted to correct for the size dependence, in case that was a confounding variable, but the anti-correlation remained. Size could still be an important factor, since the resolution of even the NVSS is much larger than the typical source size. Higher resolution observations of this sample could reveal, e.g., that the bright sources are much more compact and dominated by central AGN, as opposed to fainter, lobe-dominated structures with more ordered fields. Depolarization might also be playing a role, since $\pi_\mathrm{SP}$ is correlated with the $\|\log(D)\|$. However, again, the anti-correlation breaks down when we look at $L_\mathrm{SP}$ and $\|\log(D)\|$. This leaves us back, again, at some as yet undetermined physical difference between the faint and bright sources. ### Re-polarized objects We showed that most re-polarized objects have flat spectra ($\alpha \ge-0.5$), and are therefore concentrated in the WISE-AGN population (Figure [\[fig:wise5arc2\]](#fig:wise5arc2){reference-type="ref" reference="fig:wise5arc2"}). This makes it likely that they contain a high proportion of compact nuclei with polarization SEDs influenced by self-absorbed, and perhaps Faraday thick components. This is consistent with who also found flat spectrum objects have complex polarization behaviors. While 61% of re-polarized objects have flat spectra and are optically thick sources, the remaining 39% have steep spectra. The nature of these objects is not clear. However, there are few proposed models in the literature. re-polarization can occur when there is interference between two (or a few) unresolved and separate Faraday patches in the beam of the telescope. This can result in an oscillatory behavior of the fractional polarization with changing frequency as discussed in and. studied the AGN jet structure of 191 extragalactic radio objects, and found multiple regions along the jets of a few objects show signs of re-polarization. As discussed in they argue that both internal Faraday rotation in the jet medium as well as the configuration of the magnetic fields can explain the observed re-polarization in these optically thin jets. In Faraday thick regions the rotation of the polarization angles might align the polarization vectors from the far and near sides along the line of sight which can potentially result in re-polarization. ## Redshift Evolution {#redshift} The evolution of the magnetic properties of galaxies with time has been subject of multiple studies (such as @2012arXiv1209.1438H [@2008ApJ...676...70K; @2008Natur.454..302B; @2005MNRAS.359.1456G; @1995ApJ...445..624O; @1984ApJ...279...19W]). We distinguish here between two different quantities, an observed redshift dependence and an inferred redshift evolution, based on applying the polarization equivalent of a K-correction (redshift dilution). As discussed in Section [4.8](#zev){reference-type="ref" reference="zev"}, we found weak evidence that the average observed depolarization of steep spectrum depolarized sources with $D \ge 1.5$ decreases with increasing redshift, while the 1.4 GHz fractional polarization increases (the 2.3 GHz fractional polarization shows no change). The detected redshift variations are weak, compared to the scatter, and their probability (0.011) does not cross our conservative detection threshold. However, given the importance of this issue, we discuss the causes and consequences of redshift dependencies to help clarify the underlying issues. Polarization SEDs are often complex, especially for flat spectrum sources. This is seen in our numerous detections of re-polarization, and the broad wavelength SEDs cataloged by. In such cases, it is impossible to predict the trends of depolarization and fractional polarization with redshift expected from the K-correction. In the case where $D \sim 1$, no redshift dependence is expected, since there is no wavelength dependence to the fractional polarization. Therefore, the fact that we observe decreasing depolarization and increasing 1.4 GHz fractional polarization at increased redshift only for steep-spectrum sources with $D > 1.5$ is consistent with K-corrections only, without any physical redshift evolution. We now look at this more quantitatively, assuming the simplest case of an unresolved source with an irregular depolarizing Faraday screen (B66), external to, but at the same redshift as the source. The expected fractional polarization behavior is then \[\label{B66} \pi=\pi_0 \exp(-C\lambda_{rest}^4)\] where $\pi_0$ is the initial fractional polarization and $C \propto \sigma_{\phi}^2$ is a function of the dispersion in the Faraday depth. For a region with electron density $n$ and magnetic field component parallel to the line of sight $B_z$, fluctuations in the parameter $nB_z$ over the extent of the region is represented by $\sigma_{\phi}$. Assuming no physical change in $\sigma_{\phi}$ with time, the redshift dilution effect results in an increase in the observed fractional polarization, $\pi\propto exp\left(-C\lambda^4(1+z)^{-4}\right)$. The observed depolarization also decreases with redshift since $D\propto exp\left(C(\lambda_\mathrm{NV}^4-\lambda_\mathrm{SP}^4)(1+z)^{-4}\right)$. This simplest picture (Model 1), however, is not quantitatively consistent with our observations (Figure [\[depolsample\]](#depolsample){reference-type="ref" reference="depolsample"}). We therefore considered two additional models based on the B66 screen. Model 2: A combination of two depolarizing components, one Galactic or relatively local to us, and one at the redshift of the source, and Model 3: A physical change in $\sigma_{\phi}$ of the depolarizing screen at the source redshift. As shown in Figure [\[depolsample\]](#depolsample){reference-type="ref" reference="depolsample"}, the general behavior of the observed $\pi_\mathrm{NV}$, and $D$ as well as $\pi_\mathrm{SP}$ (not shown) of the depolarized steep spectrum sources and their evolution with redshift can be explained by models 2 and 3. However, a single depolarizing component, local to the source, with no evolution in $\sigma_{\phi}$ does not seem to be consistent with the observation. Larger samples, and resolved polarization maps where the Faraday structure can be directly seen, are needed to clarify these results. As an alternative to the B66 screen, suggested depolarization can be modeled as power law $\pi \propto \lambda^{-4/m}$ at wavelengths larger than $\lambda_{1/2}$, at which the degree of polarization is equal to half of its maximum value. The above relation only holds under certain condition in which the Faraday screen $\textrm{RM}$ structure function varies as a power law across the source $S(\delta x) \propto \delta x^m$ where $S(\delta x)\equiv <[\textrm{RM}(x+\delta x)-\textrm{RM}(x)]^2>$ and $x$ is the angular coordinate. If we assume the fractional polarization of unresolved objects follows any power law model with arbitrary exponent $-4/m$ and a constant related to the $\textrm{RM}$ dispersion, $\pi =C \lambda^{-4/m}$, then the observed depolarization, $D=\pi_{SP}/\pi_{NV}$, and both the redshift and the $\sigma_{\phi}$ dependences cancel out. Therefore, one can expect to observe no evolution in the average $D$ even if $\sigma_{\phi}$ changes with redshift, contrary to what we observe. ### Comparisons to previous work Earlier work has been based on samples including sources with both flat and steep spectra, and without selections based on depolarization. For our full sample, we find no redshift trends in fractional polarizations or depolarization, or measures of increased Faraday structure such as $\|\textrm{RRM}_\mathrm{T}\|$ and $\|\Delta \textrm{RM}\|$. This is consistent with the negative results from and. In addition, their samples were taken from the TSS09 catalog, which is biased towards high fractional polarizations, and thus, towards depolarizations D$\sim$`<!-- -->`{=html}1, for which no redshift evolution is expected. Our data are inconsistent with the analysis of, who claimed that the rotation measure of galaxies at redshifts larger than $z=1$ are on average larger (by $\sim 10$ rad m$^{-2}$) than the low redshift objects, despite the redshift dilution effect. In Figure [\[zrrmt\]](#zrrmt){reference-type="ref" reference="zrrmt"} we show $\|\textrm{RRM}_\mathrm{T}\|$ versus the redshift of objects in our sample and overlay the median $\|\textrm{RRM}\|$ values from their Figure 3. Our data are consistent with theirs, and show no evidence for the claimed increase in $\textrm{RRM}$. It is possible that a physical increase in $\sigma_{\phi}$ and depolarization as a function of redshift could mask the redshift dilution effect, leaving no observed redshift dependence to fractional polarization, $\textrm{RRM}$, $\Delta \textrm{RM}$ or depolarization. This is discussed with more details in, ,, and. studied the redshift evolution of the depolarization of 26 resolved, powerful radio galaxies and quasars over the cosmic time. They applied corrections to the measured depolarizations based on models of the wavelength and resolution effects at different redshifts. They claim a physical evolution in $\sigma_{\phi}$ and depolarization as a function of redshift, but we cannot compare their results to ours, since neither the original data nor the details of the models are shown. # Conclusions {#summary} We constructed a depolarization ($D=\pi_{2.3}/\pi_{1.4}$) catalog of extragalactic radio sources brighter than $420$ mJy at 2.3 GHz including total intensities, spectral indices, observed and residual rotation measures, fractional polarization, depolarization as well as the redshift, 2.3 GHz luminosity and WISE magnitudes for almost half of the objects. We looked for possible correlations between these quantities and found that the fractional polarization of extragalactic radio sources depends on the spectral index, morphology, the intrinsic magnetic field disorder as well as the depolarization of these sources. We summarize our main conclusions as follows:\ Consistent with previous studies over half of flat spectrum sources in our sample are re-polarized while the majority of steep spectrum objects are depolarized. There is also a significant population of steep-spectrum sources that are repolarized; their underlying physical structure is currently unknown. Although steep objects are more polarized at 2.3 GHz, they are fainter in total intensity, and therefore future surveys at higher frequencies will result in approximately the same number of sources at fixed sensitivity as the lower frequencies. Depolarization, and thus fractional polarizations, are related to the presence of Faraday structures indicated by the non-$\lambda^2$ behavior of polarization angles ($\Delta \textrm{RM}$). Future studies using polarized sources as background probes need to minimize $\textrm{RM}$ structures intrinsic to the sources. Such clean samples require high fractional polarizations ($\pi \ge 4\%$), which will severely limit the number of available sources. Sources with little or no depolarization between 1.4 GHz and 2.3 GHz have fractional polarizations ranging from a few to 10%. This is much lower than the theoretical maximum, and therefore shows the dominant role of field disorder in creating low polarizations. Compact steep spectrum objects in the NVSS catalog have more Faraday structure, and are $\sim 2$ times less polarized at 2.3 GHz than the extended sources. We found suggestive evidence for a decrease in the depolarization from $z=0$ to $z=2.3$, but only when the sample is restricted to the steep spectrum, $\alpha <-0.5$, depolarized, $D \ge 1.5$ objects. More investigation is needed to confirm the depolarization trend. Assuming that it's real, it is likely the result of the redshift dilution effect (at least partially) but requires more than a simple depolarizing screen local to the source.\ The National Radio Astronomy Observatory is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc. Partial support for ML and LR comes from National Science Foundation grant AST-1211595 to the University of Minnesota. B.M.G. has been supported by the Australian Research Council through the Centre for All-sky Astrophysics (grant CE110001020) and through an Australian Laureate Fellowship (grant FL100100114). The Dunlap Institute is funded through an endowment established by the David Dunlap family and the University of Toronto. We would like to thank G. Bernardi and D. H. F. M. Schnitzeler and the referee for a number of useful conversations and comments on the manuscript.	{'timestamp': '2016-07-19T02:08:45', 'yymm': '1607', 'arxiv_id': '1607.04914', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04914'}
# Introduction {#sec:intro} The constraint satisfaction problem (CSP) provides a framework in which it is possible to express, in a natural way, many combinatorial problems encountered in computer science and AI . An instance of the CSP consists of a set of variables, a domain of values, and a set of constraints on combinations of values that can be taken by certain subsets of variables. The basic aim is then to find an assignment of values to the variables that satisfies the constraints (decision version) or that satisfies the maximum number of constraints (optimization version). Since CSP-related algorithmic tasks are usually hard in full generality, a major line of research in CSP studies how possible algorithmic solutions depend on the set of relations allowed to specify constraints, the so-called constraint language, (see, e.g. ). The constraint language is denoted by $\Gamma$ and the corresponding CSP by $\CSP\Gamma$. For example, when one is interested in polynomial-time solvability (to optimality, for the optimization case), the ultimate sort of results are dichotomy results , pioneered by , which characterise the tractable restrictions and show that the rest are NP-hard. Classifications with respect to other complexity classes or specific algorithms are also of interest (e.g. ). When approximating (optimization) CSPs, the goal is to improve, as much as possible, the quality of approximation that can be achieved in polynomial time, see e.g. surveys . Throughout the paper we assume that P$\ne$NP. The study of almost satisfiable CSP instances features prominently in the approximability literature. On the hardness side, the notion of approximation resistance (which, intuitively, means that a problem cannot be approximated better than by just picking a random assignment, even on almost satisfiable instances) was much studied recently, e.g. . Many exciting developments in approximability in the last decade were driven by the Unique Games Conjecture (UGC) of Khot, see survey . The UGC states that it is NP-hard to tell almost satisfiable instances of $\CSP\Gamma$ from those where only a small fraction of constraints can be satisfied, where $\Gamma$ is the constraint language consisting of all graphs of permutations over a large enough domain. This conjecture (if true) is known to imply optimal inapproximability results for many classical optimization problems . Moreover, if the UGC is true then a simple algorithm based on semidefinite programming (SDP) provides the best possible approximation for all optimization problems $\CSP\Gamma$ , though the exact quality of this approximation is unknown. On the positive side, Zwick initiated the systematic study of approximation algorithms which, given an almost satisfiable instance, find an almost satisfying assignment. Formally, call a polynomial-time algorithm for CSP robust if, for every $\eps>0$ and every $(1-\eps)$-satisfiable instance (i.e. at most a $\eps$-fraction of constraints can be removed to make the instance satisfiable), it outputs a $(1-g(\eps))$-satisfying assignment (i.e. that fails to satisfy at most a $g(\eps)$-fraction of constraints). Here, the loss function $g$ must be such that $g(\eps)\rightarrow 0$ as $\eps\rightarrow 0$. Note that one can without loss of generality assume that $g(0)=0$, that is, a robust algorithm must return a satisfying assignment for any satisfiable instance. The running time of the algorithm should not depend on $\eps$ (which is unknown when the algorithm is run). Which problems $\CSP\Gamma$ admit robust algorithms? When such algorithms exist, how does the best possible loss $g$ depend on $\Gamma$? ## Related Work {#related-work .unnumbered} In , Zwick gave an SDP-based robust algorithm with $g(\eps)=O(\eps^{1/3})$ for -Sat and an LP-based robust algorithm with $g(\eps)=O(1/\log(1/\eps))$ for Horn $k$-Sat. Robust algorithms with $g(\eps)=O(\sqrt{\eps})$ were given in for -Sat, and in for Unique Games($q$) where $q$ denotes the size of the domain. For Horn-2-Sat, a robust algorithm with $g(\eps)=2\eps$ was given in . These bounds for Horn $k$-Sat ($k\ge 3$), Horn $2$-Sat, -Sat, and Unique Games($q$) are known to be optimal , assuming the UGC. The algebraic approach to CSP has played a significant role in the recent massive progress in understanding the landscape of complexity of CSPs. The key to this approach is the notion of a polymorphism, which is an $n$-ary operation (on the domain) that preserves the constraint relations. Intuitively, a polymorphism provides a uniform way to combine $n$ solutions to a system of constraints (say, part of an instance) into a new solution by applying the operation component-wise. The intention is that the new solution improves on the initial solutions in some problem-specific way. Many classifications of CSPs with respect to some algorithmic property of interest begin by proving an algebraic classification stating that every constraint language either can simulate (in a specific way, via gadgets,--see e.g. for details) one of a few specific basic CSPs failing the property of interest or else has polymorphisms having certain nice properties (say, satisfying nice equations). Such polymorphisms are then used to obtain positive results, e.g. to design and analyze algorithms. Getting such a positive result in full generality in one step is usually hard, so (typically) progress is made through a series of intermediate steps where the result is obtained for increasingly weaker algebraic conditions. The algebraic approach was originally developed for the decision CSP , and it was adapted for robust satisfiability in . One such algebraic classification result gives an algebraic condition (referred to as $\mathrm{SD}(\wedge)$ or "omitting types 1 and 2"-- see for details) equivalent to the [in]{.ul}ability to simulate Lin-$p$--systems of linear equations over $Z_p$, $p$ prime, with 3 variable per equation. Håstad's celebrated result implies that Lin-$p$ does not admit a robust algorithm (for any $g$). This result carries over to all constraint languages that can simulate (some) Lin-$p$ . The remaining languages are precisely those that have the logico-combinatorial property of CSPs called "bounded width" or "bounded treewidth duality" . This property says, roughly, that all unsatisfiable instances can be refuted via local propagation--see for a survey on dualities for CSP. Barto and Kozik used $\mathrm{SD}(\wedge)$ in , and then in they used their techniques from to prove the Guruswami-Zhou conjecture that each bounded width CSP admits a robust algorithm. The general bound on the loss in is $g(\eps)=O((\log\log(1/\eps))/\log(1/\eps))$. It is natural to ask when a better loss can be achieved. In particular, the problems of characterizing CSPs where linear loss $g(\eps)=O(\eps)$ or polynomial loss $g(\eps)=O(\eps^{1/k})$ (for constant $k$) can be achieved have been posed in . Partial results on these problems appeared in . For the Boolean case, i.e. when the domain is $\{0,1\}$, the dependence of loss on $\Gamma$ is fully classified in . ## Our Contribution {#our-contribution .unnumbered} We study CSPs that admit a robust algorithm with polynomial loss. As explained above, the bounded width property is necessary for admitting any robust algorithm. Horn 3-Sat has bounded width, but does not admit a robust algorithm with polynomial loss (unless the UGC fails) . The algebraic condition that separates Lin-$p$ and Horn 3-Sat from the CSPs that can potentially be shown to admit a robust algorithm with polynomial loss is known as $\mathrm{SD}(\vee)$ or "omitting types 1, 2 and 5" , see Section [2.2](#sec:algebra){reference-type="ref" reference="sec:algebra"} for the description of $\mathrm{SD}(\vee)$ in terms of polymorphisms. The condition $\mathrm{SD}(\vee)$ is also a necessary condition for the logico-combinatorial property of CSPs called "bounded pathwidth duality" (which says, roughly, that all unsatisfiable instances can be refuted via local propagation in a linear fashion), and possibly a sufficient condition for it too . It seems very hard to obtain a robust algorithm with polynomial loss for every CSP satisfying $\mathrm{SD}(\vee)$ all in one step. From the algebraic perspective, the most general natural condition that is (slightly) stronger than $\mathrm{SD}(\vee)$ is the near-unanimity (NU) condition . CSPs with a constraint language having an NU polymorphism received a lot of attention in the literature (e.g. ). Bounded pathwidth duality for CSPs admitting an NU polymorphism was established in a series of papers , and we use some ideas from in this paper. We prove that any CSP with a constraint language having an NU polymorphism admits a randomized robust algorithm with loss $O(\eps^{1/k})$, where $k$ depends on the size of the domain. It is an open question whether this dependence on the size of the domain is necessary. We prove that, for the special case of a ternary NU polymorphism known as dual discriminator (the corresponding CSP is a common generalisation of Unique Games with a fixed domain and -Sat), we can always choose $k=2$. Our algorithms use the standard SDP relaxation for CSPs. The algorithm for the general NU case follows the same general scheme as : 1. Solve the LP/SDP relaxation for a $(1-\eps)$-satisfiable instance $\inst I$. 2. Use the LP/SDP solution to remove certain constraints in $\inst I$ with total weight $O(g(\eps))$ (in our case, $O(\eps^{1/k})$) so that the remaining instance satisfies a certain consistency condition. 3. Use the appropriate polymorphism (in our case, NU) to show that any instance of $\CSP\Gamma$ with this consistency condition is satisfiable. Steps 1 and 2 in this scheme can be applied to any CSP instance, and this is where essentially all work of the approximation algorithm happens. Polymorphisms are not used in the algorithm, they are used in Step 3 only to prove the correctness. While the above general scheme is rather simple, applying it is typically quite challenging. Obviously, Step 2 prefers weaker conditions (achievable by removing not too many constraints), while Step 3 prefers stronger conditions (so that they can guarantee satisfiability), so reaching the balance between them is the main (and typically significant) technical challenge in any application of this scheme. Our algorithm is somewhat inspired by , but it is also quite different from the algorithm there. That algorithm is designed so that Steps 1 and 2 establish a consistency condition that, in particular, includes the 1-minimality condition, and establishing 1-minimality alone requires removing constraints with total weight $O(1/\log{(1/\eps)})$ , unless UGC fails. To get the right dependency on $\eps$ we introduce a new consistency condition somewhat inspired by . The proof that the new consistency condition satisfies the requirements of Steps 2 and 3 of the above scheme is one of the main technical contributions of our paper. ### Organization of the paper {#organization-of-the-paper .unnumbered} After some preliminaries, we formulate the two main results of this paper in Section [3](#sec:main){reference-type="ref" reference="sec:main"}. Section [4](#sec:SDP){reference-type="ref" reference="sec:SDP"} then contains a description of SDP relaxations that we will use further on. Sections [5](#sec:overview1){reference-type="ref" reference="sec:overview1"} and [6](#sec:overview-thm2){reference-type="ref" reference="sec:overview-thm2"} contain the description of the algorithms for constraint languages compatible with NU polymorphism and dual discriminator, respectively; the following chapters prove the correctness of the two algorithms. # Preliminaries ## CSPs Throughout the paper, let $D$ be a fixed finite set, sometimes called the domain. An instance of the $\csp$ is a pair $\inst I=(V,{\mathcal C})$ with $V$ a finite set of variables and ${\mathcal C}$ is a finite set of constraints. Each constraint is a pair $(\overline{x},R)$ where $\overline{x}$ is a tuple of variables (say, of length $r>0$), called the scope of $C$ and $R$ an $r$-ary relation on $D$ called the constraint relation of $C$. The arity of a constraint is defined to be the arity of its constraint relation. In the weighted optimization version, which we consider in this paper, every constraint $C\in{\mathcal C}$ has an associated weight $w_C\geq 0$. Unless otherwise stated we shall assume that every instance satisfies $\sum_{C\in{\mathcal C}} w_C=1$. An assignment for $\inst I$ is a mapping $s:V\rightarrow D$. We say that $s$ satisfies a constraint $((x_1,\dots,x_r),R)$ if $(s(x_1),\dots,s(x_r))\in R$. For $0\leq \beta\leq 1$ we say that assignment $s$ $\beta$-satisfies $\inst I$ if the total weight of the constraints satisfied by $s$ is at least $\beta$. In this case we say that $\inst I$ is $\beta$-satisfiable. The best possible $\beta$ for $\inst I$ is denoted by $\mathrm{Opt}(\inst I)$. A constraint language on $D$ is a finite set $\Gamma$ of relations on $D$. The problem $\csp(\Gamma)$ consists of all instances of the CSP where all the constraint relations are from $\Gamma$. Problems $k$-Sat, Horn $k$-Sat, Lin-$p$, Graph $H$-colouring, and Unique Games$(\|D\|)$ are all of the form $\CSP\Gamma$. The decision problem for $\csp(\Gamma)$ asks whether an input instance $\inst I$ of $\csp(\Gamma)$ has an assignment satisfying all constraints in $\inst I$. The optimization problem for $\csp(\Gamma)$ asks to find an assignment $s$ where the weight of the constraints satisfied by $s$ is as large as possible. Optimization problems are often hard to solve to optimality, motivating the study of approximation algorithms. ## Algebra {#sec:algebra} An $n$-ary operation $f$ on $D$ is a map from $D^n$ to $D$. We say that $f$ preserves (or is a polymorphism of) an $r$-ary relation $R$ on $D$ if for all $n$ (not necessarily distinct) tuples $(a^i_1,\dots,a_r^i)\in R$, $1\leq i\leq n$, the tuple $(f(a_1^1,\dots,a_n^1),\dots,f(a_1^r,\dots,a_n^r))$ belongs to $R$ as well. Say, if $R$ is the edge relation of a digraph $H$, then $f$ is a polymorphism of $R$ if and only if, for any list of $n$ (not necessarily distinct) edges $(a_1,b_1),\ldots,(a_n,b_n)$ of $H$, there is an edge in $H$ from $f(a_1,\ldots,a_n)$ to $f(b_1,\ldots,b_n)$. If $f$ is a polymorphism of every relation in a constraint language $\Gamma$ then $f$ is called a polymorphism of $\Gamma$. Many algorithmic properties of $\CSP\Gamma$ depend only on the polymorphisms of $\Gamma$, see survey , also . An $(n+1)$-ary ($n\ge 2$) operation $f$ is a near-unanimity (NU) operation if, for all $x,y\in D$, it satisfies \[\begin{gathered} f(x,x,\ldots,x,x,y)=f(x,x,\ldots,x,y,x)=\dots =f(y,x,\ldots,x,x,x)=x. \end{gathered}\] Note that the behaviour of $f$ on other tuples of arguments is not restricted. An NU operation of arity 3 is called a majority operation. We mentioned in the introduction that (modulo UGC) only constraint languages satisfying condition $\mathrm{SD}(\vee)$ can admit robust algorithms with polynomial loss. The condition $\mathrm{SD}(\vee)$ can be expressed in many equivalent ways: for example, as the existence of ternary polymorphisms $d_0,\ldots, d_t$, $t\ge 2$, satisfying the following equations : \[\begin{aligned} d_0(x,y,z) &= x,\quad d_t(x,y,z) = z, \\ d_i(x,y,x) &= d_{i+1}(x,y,x) \text{ for all even $i<t$}, \label{sdjoin3}\\ d_i(x,y,y) &= d_{i+1}(x,y,y) \text{ for all even $i<t$},\\ d_i(x,x,y) &= d_{i+1}(x,x,y) \text{ for all odd $i<t$}. \end{aligned}\] If line ([\[sdjoin3\]](#sdjoin3){reference-type="ref" reference="sdjoin3"}) is strengthened to $d_i(x,y,x)=x$ for all $i$, then, for any constraint language, having such polymorphisms would be equivalent to having an NU polymorphism of some arity (this is true only when constraint languages are assumed to be finite). NU polymorphisms appeared many times in the CSP literature. For example, they characterize the so-called "bounded strict width" property , which says, roughly, that, after establishing local consistency in an instance, one can always construct a solution in a greedy way, by picking values for variables in any order so that constraints are not violated. Every relation with an $(n+1)$-ary NU polymorphism is $n$-decomposable (and in some sense the converse also holds) . We give a formal definition only for the majority case $n=2$. Let $R$ be a $r$-ary ($r\ge 2$) relation. For every $i,j\in\{1,\dots,r\}$, let $\pr_{i,j} R$ be the binary relation $\{(a_i,a_j)\mid (a_1,\dots,a_r)\in R\}$. Then $R$ is called $2$-decomposable if the following holds: a tuple $(a_1,\dots,a_r)\in D^r$ belongs to $R$ if and only if $(a_i,a_j)\in \pr_{i,j} R$ for every $i,j\in\{1,\dots,r\}$. The dual discriminator is a majority operation $f$ such that $f(x,y,z)=x$ whenever $x,y,z$ are pairwise distinct. Binary relations preserved by the dual discriminator are known as implicational or 0/1/all relations. Every such relation is of one of the four following types: 1. $(\{a\}\times D)\cup (D\times\{b\})$ for $a,b\in D$, 2. $\{(\pi(a),a)\mid a\in D\}$ where $\pi$ is a permutation on $D$, 3. $P\times Q$ where $P,Q\subseteq D$, 4. a intersection of a relation of type 1 or 2 with a relation of type 3. The relations of the first kind, when $D=\{0,1\}$, are exactly the relations allowed in -Sat, while the relations of the second kind are precisely the relations allowed in Unique Games $(\|D\|)$. We remark that having such an explicit description of relations having a given polymorphism is rare beyond the Boolean case. # Main result {#sec:main} It was stated as an open problem in whether every CSP that admits a robust algorithm with loss $O(\eps^{1/k})$ admits one where $k$ is bounded by an absolute constant (that does not dependent on $D$). In the context of the above theorem, the problem can be made more specific: is dependence of $k$ on $\|D\|$ in this theorem avoidable or there is a strict hierarchy of possible degrees there? The case of a majority polymorphism is a good starting point when trying to answer this question. As mentioned in the introduction, robust algorithms with polynomial loss and bounded pathwidth duality for CSPs seem to be somehow related, at least in terms of algebraic conditions. The condition $\mathrm{SD}(\vee)$ is the common necessary condition for them, albeit it is conditional on UGC for the former and unconditional for the latter. Having an NU polymorphism is a sufficient condition for both. Another family of problems $\CSP\Gamma$ with bounded pathwidth duality was shown to admit robust algorithms with polynomial loss in , where the parameter $k$ depends on the pathwidth duality bound (and appears in the algebraic description of this family). This family includes languages not having an NU polymorphism of any arity--see . It is unclear how far connections between the two directions go, but consistency notions seem to be the common theme. Returning to the discussion of a possible hierarchy of degrees in polynomial loss in robust algorithms--there was a similar question about a hierarchy of bounds for pathwidth duality, and the hierarchy was shown to be strict , even in the presence of a majority polymorphism. # SDP relaxation {#sec:SDP} Associated to every instance $\inst I=(V,{\mathcal C})$ of CSP there is a standard SDP relaxation. It comes in two versions: maximizing the number of satisfied constraints and minimizing the number of unsatisfied constraints. We use the latter. We define it assuming that all constraints are binary, this will be sufficient for our purposes. The SDP has a variable $\mathbf{x}_a$ for every $x\in V$ and $a\in D$. It also contains a special unit vector $\mathbf{v}_0$. The goal is to assign $(\|V\\|D\|)$-dimensional real vectors to its variables minimizing the following objective function: \[\sum_{C=((x,y),R)\in {\mathcal C}} w_C\sum_{(a,b)\not\in R} \mathbf{x}_a\mathbf{y}_b \label{sdpobj}\] subject to: \[\begin{aligned} &\mathbf{x}_a\mathbf{y}_b\geq 0 & x,y\in V, a,b\in D \label{sdp1} \\ &\mathbf{x}_a\mathbf{x}_b= 0 & x\in V, a,b\in D, a\neq b \label{sdp2} \\ &\textstyle\sum_{a\in D} \mathbf{x}_a=\mathbf{v}_0 & x\in V \label{sdp3} \\ &\\|\mathbf{v}_0\\|=1 & \label{sdp4} \end{aligned}\] In the intended integral solution, $x=a$ if $\mathbf{x}_a=\mathbf{v}_0$. In the fractional solution, we informally interpret $\\|\mathbf{x}_a\\|^2$ as the probability of $x=a$ according to the SDP (the constraints of the SDP ensure that $\sum_{a\in D} \\|\mathbf{x}_a\\|^2 =1$). If $C=((x,y),R)$ is a constraint and $a,b\in D$, one can think of $\mathbf{x}_a\mathbf{y}_b$ as the probability given by the solution of the SDP to the pair $(a,b)$ in $C$. The optimal SDP solution, then, gives as little probability as possible to pairs that are not in the constraint relation. For a constraint $C=((x,y),R)$, conditions ([\[sdp3\]](#sdp3){reference-type="ref" reference="sdp3"}) and ([\[sdp4\]](#sdp4){reference-type="ref" reference="sdp4"}) imply that $\sum_{(a,b)\in R} \mathbf{x}_a\mathbf{y}_b$ is at most $1$. Let $\loss(C)=\sum_{(a,b)\not\in R} \mathbf{x}_a\mathbf{y}_b$. For a subset $A\subseteq D$, let $\mathbf{x}_A=\sum_{a\in A} \mathbf{x}_a$. Note that $\mathbf{x}_D=\mathbf{y}_D(=\mathbf{v}_0)$ for all $x,y\in D$. Let $\mathrm{SDPOpt}(\inst I)$ be the optimum value of ([\[sdpobj\]](#sdpobj){reference-type="ref" reference="sdpobj"}). It is clear that, for any instance $\inst I$, we have $\mathrm{Opt}(\inst I)\ge \mathrm{SDPOpt}(\inst I)\ge 0$. There are algorithms that, given an SDP instance $\inst I$ and some additive error $\delta>0$, produce in time $\operatorname{\textit{poly}}\, (\|\inst I\|, \log(1/\delta))$ an output vector solution whose value is at most $\mathrm{SDPOpt}(\inst I)+\delta$. There are several ways to deal with the error $\delta$. In this paper we deal with it by introducing a preprocessing step which will also be needed to argue that the algorithm described in the proof of Theorem [\[the:main\]](#the:main){reference-type="ref" reference="the:main"}(1) runs in polynomial time. Preprocessing step 1.Assume that ${\mathcal C}=\{C_1,\dots,C_m\}$ and that $w_{C_1}\ge w_{C_2}\ge \ldots \ge w_{C_m}$. Using the algorithm from Theorem [\[the:nu\]](#the:nu){reference-type="ref" reference="the:nu"}, find the largest $j$ such that the subinstance $\inst I_j=(V,\{C_1,\dots,C_j\})$ is satisfiable. If the total weight of the constraints in $\inst I_j$ is at least $1-1/m$ then return the assignment $s$ satisfying $\inst I_j$ and stop. If the preprocessing step returns an assignment then we are done. So assume that it did not return an assignment. Then we know that $\epsilon\ge 1/m^2$. We then solve the SDP relaxation with $\delta=1/m^2$ obtaining a solution with objective value at most $2\epsilon$ which is good enough for our purposes. # Overview of the proof of Theorem [\[the:main\]](#the:main){reference-type="ref" reference="the:main"}(1) {#sec:overview1} We assume throughout that $\Gamma$ has a near-unanimity polymorphism of arity $n+1$ ($n\ge 2$). It is sufficient to prove Theorem [\[the:main\]](#the:main){reference-type="ref" reference="the:main"}(1) for the case when $\Gamma$ consists of binary relations and $k=6\|D\|+7$. The rest will follow by Proposition 4.1 of (see also Theorem 24 in ), which shows how to reduce the general case to constraint languages consisting of unary and binary relations in such a way that the domain size increases from $\|D\|$ to $\|D\|^r$ where $r$ is the maximal arity of a relation in $\Gamma$. Note that every unary constraint $(x,R)$ can be replaced by the binary constraint $((x,x),R')$ where $R'=\{(a,a)\mid a\in R\}$. Throughout the rest of this section, let $\inst I=(V,{\mathcal C})$ be a $(1-\eps)$-satisfiable instance of $\CSP\Gamma$. ## Patterns and realizations {#sect:pattern} A pattern in $\inst I$ is defined as a directed multigraph $p$ whose vertices are labeled by variables of $\inst I$ and edges are labeled by constraints of $\inst I$ in such a way that the beginning of an edge labeled by $((x,y),R)$ is labeled by $x$ and the end by $y$. Two of the vertices in $p$ can be distinguished as the beginning and the end of $p$. If these two vertices are labeled by variables $x$ and $y$, respectively, then we say that $p$ is a pattern is from $x$ to $y$. For two patterns $p$ and $q$ such that the end of $p$ and the beginning of $q$ are labeled by the same variable, we define $p+q$ to be the pattern which is obtained from the disjoint union of $p$ and $q$ by identifying the end of $p$ with the beginning of $q$ and choosing the beginning of $p+q$ to be the beginning of $p$ and the end of $p+q$ to be the end of $q$. We also define $jp$ to be $p + \dots + p$ where $p$ appears $j$ times. A pattern is said to be a path pattern if the underlying graph is an oriented path with the beginning and the end being the two end vertices of the path, and is said to be an $n$-tree pattern if the underlying graph is an orientation of a tree with at most $n$ leaves, and both the beginning and the end are leaves. A path of $n$-trees pattern is then any pattern of the form $t_1+\dots+t_j$ for some $n$-tree patterns $t_1,\dots,t_j$. A realization of a pattern $p$ is a mapping $r$ from the set of vertices of $p$ to $D$ such that if $(v_x,v_y)$ is an edge labeled by $((x,y),R)$ then $(r(v_x),r(v_y)) \in R$. Note that $r$ does not have to map different vertices of $p$ labeled with same variable to the same element in $D$. A propagation of a set $A\subseteq D$ along a pattern $p$ whose beginning vertex is $b$ and ending vertex is $e$ is defined as follows. For $A\subseteq D$, define $A + p=\{r(e) \mid \text{ $r$ is a realization of $p$ with $r(b) \in A$}\}$. Also for a binary relation $R$ we put $A+R=\{b\mid (a,b)\in R \mbox{ and } a\in A\}$. Observe that we have $(A+p)+q = A+(p+q)$. Further, assume that we have non-empty sets $D_x^\ell$ where $1\leq \ell\leq \|D\|+1$ and $x$ runs through all variables in an instance $\inst I$. Let $p$ be a pattern in $\inst I$ with beginning $b$ and end $e$. We call a realization $r$ of $p$ an $\ell$-realization (with respect to the family $\{D_x^\ell\}$) if, for any vertex $v$ of $p$ labeled by a variable $x$, we have $r(v)\in D_x^{\ell+1}$. For $A\subseteq D$, define $A +^\ell p=\{r(e) \mid r \text{ is an $\ell$-realization of $p$ with $r(b) \in A$}\}$. Also, for a constraint $((x,y),R)$ or $((y,x),R^{-1})$ and sets $A,B\subseteq D$, we write $B=A+^\ell (x,R,y)$ if $B=\{b\in D_y^{\ell+1} \mid (a,b)\in R \mbox{ for some } a\in A\cap D_x^{\ell+1}\}$. ## The consistency notion Recall that we assume that $\Gamma$ contains only binary relations. Before we formally introduce the new consistency notion, which is the key to our result, as we explained in the introduction, we give an example of a similar simpler condition. We mentioned before that -Sat is a special case of a CSP that admits an NU polymorphism (actually, the only majority operation on $\{0,1\}$). There is a textbook consistency condition characterizing satisfiable -Sat instances, which can be expressed in our notation as follows: for each variable $x$ in a -Sat instance $\inst I$, there is a value $a_x$ such that, for any path pattern $p$ in $\inst I$ from $x$ to $x$, we have $a_x\in \{a_x\}+p$. Let $\inst I$ be an instance of $\CSP\Gamma$ over a set $V$ of variables. We say that $\inst I$ satisfies condition if the following holds: Note that $+$ between $p$ and $q$ is the pattern addition and thus independent of $\ell$. Note also that $a$ in the above condition belongs to $D_x^\ell$, while propagation is performed by using $\ell$-realizations, i.e., inside sets $D_y^{\ell+1}$. The following theorem states that this consistency notion satisfies the requirements of Step 3 of the general scheme (for designing robust approximation algorithms) discussed in the introduction. ## The algorithm Let $k = 6\|D\| + 7$. We provide an algorithm which, given a $(1-\epsilon)$-satisfiable instance $\inst I$ of $\CSP\Gamma$, removes $O(\epsilon^{1/k})$ constraints from it to obtain a subinstance $\inst I'$ satisfying condition (IPQ)$_n$. It then follows from Theorem [\[thm:nicelevels\]](#thm:nicelevels){reference-type="ref" reference="thm:nicelevels"} that $\inst I'$ is satisfiable, and we can find a satisfying assignment by Theorem [\[the:nu\]](#the:nu){reference-type="ref" reference="the:nu"}. ### More preprocessing {#sec:preprocessing} By Lemma [\[lem:prep1\]](#lem:prep1){reference-type="ref" reference="lem:prep1"} we can assume that $\epsilon\geq 1/m^2$. We solve the SDP relaxation with error $\delta=1/m^2$ and obtain an solution $\{\vec x_a\}$ $(x\in V,a\in D)$ whose objective value $\epsilon'$ is at most $2\epsilon$. Let us define $\alpha$ to be $\max\{\epsilon',1/m^2\}$. It is clear that $\alpha=O(\epsilon)$. Furthermore, this gives us that $1/\alpha\le m^2$. This will be needed to argue that the main part of the algorithm runs in polynomial time. Let $\kappa=1/k$ (we will often use $\kappa$ to avoid overloading formulas). Preprocessing step 2.For each $x\in V$ and $1\leq \ell\leq \|D\|+1$, compute sets $D_x^{\ell}\subseteq D$ as follows. Set $D_x^{\|D\|+1}=D$ and, for $1\leq \ell\leq \|D\|$, set $D_x^{\ell}=\{a\in D\mid \\|\vec x_a\\|\ge r_{x,\ell}\}$ where $r_{x,\ell}$ is the smallest number of the form $r=\alpha^{3\ell \kappa}(2\|D\|)^{i/2}$, $i\ge 0$ integer, with $\{b\in D\mid r(2\|D\|)^{-1/2}\le \\|\vec x_b\\|< r\}=\emptyset$. It is easy to check that $r_{x,\ell}$ is obtained with $i\le \|D\|$. It is clear that the sets $D_x^{\ell}\subseteq D$, $x\in V$, $1\leq \ell\leq \|D\|$, can be computed in polynomial time. The sets $D_x^\ell$ are chosen such that $D_x^\ell$ contains relatively "heavy" elements ($a$'s such that $\\|\vec x_a\\|^2$ is large). The thresholds are chosen so that there is a big gap (at least by a factor of $2\|D\|$) between "heaviness" of an element in $D_x^\ell$ and outside. ### Main part {#sec:algorithm} Given the preprocessing is done, we have that $1/\alpha \leq m^2$, and we precomputed sets $D_x^\ell$ for all $x\in V$ and $1\leq \ell \leq \|D\|+1$. The description below uses the number $n$, where $n+1$ is the arity of the NU polymorphism of $\Gamma$. Step 0.Remove every constraint $C$ with $\loss(C) > \alpha^{1-\kappa}$. Step 1.For every $1 \leq \ell \leq \|D\|$ do the following. Pick a value $r_\ell \in (0,\alpha^{(6\ell +4)\kappa})$ uniformly at random. Here we need some notation: for $x,y \in V$ and $A,B\subseteq D$, we write $\vec x_A \preceq^\ell \vec y_B$ to indicate that there is no integer $j$ such that $\\| \vec y_B \\|^2 < r_\ell +j\alpha^{(6\ell +4)\kappa} \leq \\| \vec x_A \\|^2.$ Then, remove all constraints $((x,y),R)$ such that there are sets $A,B \subseteq D$ with $B = A +^\ell (x,R,y)$ and $\vec x_A \not\preceq^\ell \vec y_B$, or with $B = A +^\ell (y,R^{-1},x)$ and $\vec y_A \not\preceq^\ell \vec x_B$. Step 2.For every $1 \leq \ell \leq \|D\|$ do the following. Let $m_0 = \lfloor \alpha^{-2\kappa} \rfloor$. Pick a value $s_\ell \in \{0,\dots, m_0-1\}$ uniformly at random. We define $\vec x_A \preceq^{\ell}_w \vec y_B$ to mean that there is no integer $j$ such that $\\|\vec y_B\\|^2 < r_\ell + (s_\ell +jm_0)\alpha^{(6\ell +4)\kappa} \leq \\|\vec x_A\\|^2.$ Obviously, if $\vec x_A \preceq^{\ell} \vec y_B$ then $\vec x_A \preceq^{\ell}_w \vec y_B$. Now, if $A\subseteq B \subseteq D_x^{\ell+1}$ are such that $\\|\vec x_B-\vec x_A\\|^2 \leq (2n-3)\alpha^{(6\ell +4)\kappa}$ and $\vec x_B \not\preceq^{\ell}_w \vec x_A$, then remove all the constraints in which $x$ participates. Step 3.For every $1\leq \ell \leq \|D\|$ do the following. Pick $m_\ell = \lceil \alpha ^ {-(3\ell +1)\kappa} \rceil$ unit vectors independently uniformly at random. For $x,y\in V$ and $A,B \subseteq D$, say that $\vec x_A$ and $\vec y_B$ are cut by a vector $\vec u$ if the signs of $\vec u \cdot (\vec x_A-\vec x_{D\setminus A})$ and $\vec u \cdot (\vec y_B-\vec y_{D\setminus B})$ differ. Furthermore, we say that $\vec x_A$ and $\vec y_B$ are $\ell$-cut if there are cut by at least one of the chosen $m_\ell$ vectors. For every variable $x$, if there exist subsets $A,B \subseteq D$ such that $A \cap D_x^\ell \neq B \cap D_x^\ell$ and the vectors $\vec x_A$ and $\vec x_B$ are not $\ell$-cut, then remove all the constraints in which $x$ participates. Step 4.For every $1 \leq \ell \leq \|D\|$, remove every constraint $((x,y),R)$ such that there are sets $A,B \subseteq D$ with $B = A +^\ell (x,R,y)$, and $\vec x_A$ and $\vec y_B$ are $\ell$-cut, or with $B = A +^\ell (y,R^{-1},x)$, and $\vec y_A$ and $\vec x_B$ are $\ell$-cut. Step 5.For every $1 \leq \ell \leq \|D\|$ do the following. For every variable $x$, If $A,B\subseteq D_x^{\ell +1}$ such that $\\| \vec x_B-\vec x_A \\|^2 \leq (2n-3)\alpha^{(6\ell+4)\kappa}$ and $\vec x_A$ and $\vec x_B$ are $\ell$-cut, remove all constraints in which $x$ participates. Step 6.By Proposition [\[prop:consistence\]](#prop:consistence){reference-type="ref" reference="prop:consistence"} and Theorem [\[thm:nicelevels\]](#thm:nicelevels){reference-type="ref" reference="thm:nicelevels"}, the remaining instance $\inst I'$ is satisfiable. Use the algorithm given by Theorem [\[the:nu\]](#the:nu){reference-type="ref" reference="the:nu"} to find a satisfying assignment for $\inst I'$. Assign all variables in $\inst I$ that do not appear in $\inst I'$ arbitrarily and return the obtained assignment for $\inst I$. Note that we chose to define the cut condition based on $\mathbf{x}_A-\mathbf{x}_{D\setminus A}$, rather than on $\mathbf{x}_A$, because the former choice has the advantage that $\\|\mathbf{x}_A-\mathbf{x}_{D\setminus A}\\|=1$, which helps in some calculations. In Step 0 we remove constraints such that, according to the SDP solution, have a high probability to be violated. Intuitively, steps 1 and 2 ensure that the loss in $\\|\vec x_A\\|$ after propagating $A$ by a path of $n$-trees is not too big. This is achieved first by ensuring that by following a path we do not lose too much (step 1) which also gives a bound on how much we can lose by following an $n$-tree pattern (see Lemma [\[lem:tree-loss\]](#lem:tree-loss){reference-type="ref" reference="lem:tree-loss"}). Together with the removal of constraints in step 2, this guarantees that following a path of $n$-trees we do not lose too much. This ensures that $\{a\} +^\ell (j(p+q) + p)$ is non-vanishing as $j$ increases. Steps 3--5 ensure that if $A$ and $B$ are connected by paths of $n$-trees in both directions (i.e. $\vec x_A = \vec x_B +^{\ell} p_1$ and $\vec x_B = \vec x_A +^{\ell} p_2$), then $\vec x_A$ and $\vec x_B$ do not differ too much (i.e. $A \cap D_x^\ell = B \cap D_x^\ell$). This is achieved by separating the space into cones by cutting it using the $m_\ell$ chosen vectors, removing the variables which have two different sets that are not $\ell$-cut (step 3), and then ensuring that if we follow an edge (step 4), or if we drop elements that do not extend to an $n$-tree (step 5) we do not cross a border to another cone. This gives us both that the sequence $A_j = \{a\} +^\ell (j(p+q) +p)$ stabilizes and that, after it stabilizes, $A_j$ contains $a$. This provides condition for the remaining instance $\inst I'$. The algorithm runs in polynomial time. Since $D$ is fixed, it is clear that the steps 0--2 can be performed in polynomial time. For steps 3--5, we also need that $m_\ell$ is bounded by a polynomial in $m$, which holds because $\alpha \geq 1/m^2$. The correctness of the algorithm is given by (Theorem [\[thm:nicelevels\]](#thm:nicelevels){reference-type="ref" reference="thm:nicelevels"} and) the two following propositions whose proof can be found in Section [8](#sec:correctness-of-algorithm){reference-type="ref" reference="sec:correctness-of-algorithm"}. These propositions show that our new consistency notion satisfies the requirements of Step 2 of the general scheme for designing robust approximation algorithms discussed in the introduction. # Overview of the proof of Theorem [\[the:main\]](#the:main){reference-type="ref" reference="the:main"}(2) {#sec:overview-thm2} Since the dual discriminator is a majority operation, every relation in $\Gamma$ is 2-decomposable. Therefore, it follows, e.g. from Lemma 3.2 in , that to prove that $\CSP\Gamma$ admits a robust algorithm with loss $O(\sqrt\eps)$, it suffices to prove this for the case when $\Gamma$ consists of all unary and binary relations preserved by the dual discriminator. Such binary constraints are of one of the four kinds described in Section [2.2](#sec:algebra){reference-type="ref" reference="sec:algebra"}. Using this description, it follows from Lemma 3.2 of that it suffices to consider the following three types of constraints: 1. Disjunction constraints of the form $x= a \vee y = b$, where $a,b\in D$; 2. Unique game (UG) constraints of the form $x = \pi(y)$, where $\pi$ is any permutation on $D$; 3. Unary constraints of the form $x \in P$, where $P$ is an arbitrary non-empty subset of $D$. We present an algorithm that, given a $(1-\eps)$-satisfiable instance $\inst I=(V,{\mathcal C})$ of the problem, finds a solution satisfying constraints with expected total weight $1-O(\sqrt{\eps\log{\|D\|}})$ (the hidden constant in the $O$-notation does not depend on $\eps$ and $\|D\|$). We now give an informal and somewhat imprecise sketch of the algorithm and its analysis. We present details in Section [9](#sec:thm2){reference-type="ref" reference="sec:thm2"}. We use the SDP relaxation from Section [4](#sec:SDP){reference-type="ref" reference="sec:SDP"}. Let us call the value $\\|\mathbf{x}_a\\|^2$ the SDP weight of the value $a$ for variable $x$. ### Variable Partitioning Step {#variable-partitioning-step .unnumbered} The algorithm first solves the SDP relaxation. Then, it partitions all variables into three groups $\VV{0}$, $\VV{1}$, and $\VV{2}$ using a threshold rounding algorithm with a random threshold. If most of the SDP weight for $x$ is concentrated on one value $a\in D$, then the algorithm puts $x$ in the set $\VV{0}$ and assigns $x$ the value $a$. If most of the SDP weight for $x$ is concentrated on two values $a,b\in D$, then the algorithm puts $x$ in the set $\VV{1}$ and restricts the domain of $x$ to the set $D_x = \{a,b\}$ (thus we guarantee that the algorithm will eventually assign one of the values $a$ or $b$ to $x$). Finally, if the SDP weight for $x$ is spread among 3 or more values, then we put $x$ in the set $\VV{2}$; we do not restrict the domain for such $x$. After we assign values to $x\in \VV{0}$ and restrict the domain of $x\in \VV{1}$ to $D_x$, some constraints are guaranteed to be satisfied (say, the constraint $(x=a)\vee (y=b)$ is satisfied if we assign $x$ the value $a$ and the constraint $x\in P$ is satisfied if $D_x\subseteq P$). Denote the set of such constraints by ${\cal C}_s$ and let ${\mathcal C}'={\mathcal C}\setminus{\mathcal C}_s$. We then identify a set ${\cal C}_v\subseteq {\mathcal C}'$ of constraints that we conservatively label as violated. This set includes all constraints in ${\mathcal C}'$ except those belonging to one of the following 4 groups: 1. disjunction constraints $(x = a) \vee (y = b)$ with $x, y \in {\cal V}_1$ and $a\in D_x$, $b\in D_y$; 2. UG constraints $x = \pi (y)$ with $x, y \in {\cal V}_1$ and $D_x = \pi(D_y)$; 3. UG constraints $x = \pi (y)$ with $x, y \in {\cal V}_2$; 4. unary constraints $x \in P$ with $x\in{\cal V}_2$. Our construction of sets $\VV{0}$, $\VV{1}$, and $\VV{2}$, which is based on randomized threshold rounding, ensures that the expected total weight of constraints in ${\cal C}_v$ is $O(\eps)$ (see Lemma [\[lem:preproc-violated\]](#lem:preproc-violated){reference-type="ref" reference="lem:preproc-violated"}). The constraints from the 4 groups above naturally form two disjoint sub-instances of $\inst I$: $\inst I_1$ (groups 1 and 2) on the set of variables $\VV{1}$, and $\inst I_2$ (groups 3 and 4) on $\VV{2}$. We treat these instances independently as described below. ### Solving Instance $\inst I_1$ {#solving-instance-inst-i_1 .unnumbered} The instance $\inst I_1$ with the domain of each $x$ restricted to $D_x$ is effectively an instance of Boolean 2-CSP (i.e. each variable has a 2-element domain and all constraints are binary). A robust algorithm with quadratic loss for this problem was given by Charikar et al. . This algorithm finds a solution violating an $O(\sqrt{\varepsilon})$ fraction of all constraints if the optimal solution violates at most $\varepsilon$ fraction of all constraints or $\mathrm{SDPOpt}(\inst I_1)\leq \eps$. However, we cannot apply this algorithm to the instance $\inst I_1$ as is. The problem is that the weight of violated constraints in the optimal solution for $\inst I_1$ may be greater than $\omega(\varepsilon)$. Note that the unknown optimal solution for the original instance $\inst I$ may assign values to variables $x$ outside of the restricted domain $D_x$, and hence it is not a feasible solution for $\inst I_1$. Furthermore, we do not have a feasible SDP solution for the instance $\inst I_1$, since the original SDP solution (restricted to the variables in $\VV{1}$) is not a feasible solution for the Boolean 2-CSP problem (because $\sum_{a\in D_x}\mathbf{x}_a$ is not necessarily equal to $\mathbf{v}_0$ and, consequently, $\sum_{a\in D_x}\\|\mathbf{x}_a\\|^2$ may be less than 1). Thus, our algorithm first transforms the SDP solution to obtain a feasible solution for $\inst I_1$. To this end, it partitions the set of vectors $\{\mathbf{x}_a: x\in \VV{1}, a\in D_x\}$ into two sets $H$ and $\bar{H}$ using a modification of the hyperplane rounding algorithm by Goemans and Williamson . In this partitioning, for every variable $x$, one of the two vectors $\{\mathbf{x}_a: a\in D_x\}$ belongs to $H$ and the other belongs to $\bar H$. Label the elements of each $D_x$ as $\alpha_x$ and $\beta_x$ so that so that $\mathbf{x}_{\alpha_x}$ is the vector in $H$ and $\mathbf{x}_{\beta_x}$ is the vector in $\bar H$. For every $x$, we define two new vectors $\mathbf{\tilde{x}}_{\alpha_x} = \mathbf{x}_{\alpha_x}$ and $\mathbf{\tilde{x}}_{\beta_x} = \mathbf{v}_0-\mathbf{x}_{\alpha_x}$. It is not hard to verify that the set of vectors $\{\mathbf{\tilde{x}}_{a}: x\in\VV{1}, a\in D_x\}$ forms a feasible SDP solution for the instance $\inst I_1$. We show that for each disjunction constraint $C$ in the instance $\inst I_1$, the cost of $C$ in the new SDP solution is not greater than the cost of $C$ in the original SDP solution (see Lemma [\[lem:new-SDP-feasible\]](#lem:new-SDP-feasible){reference-type="ref" reference="lem:new-SDP-feasible"}). The same is true for all but $O(\sqrt{\eps})$ fraction of UG constraints. Thus, after removing UG constraints for which the SDP value has increased, we get an SDP solution of cost $O(\varepsilon)$. Using the algorithm for Boolean 2-CSP, we obtain a solution for $\inst I_1$ that violates constraints of total weight at most $O(\sqrt{\eps})$. ### Solving Instance $\inst I_2$ {#solving-instance-inst-i_2 .unnumbered} The instance $\inst I_2$ may contain only unary and UG constraints as all disjunction constraints are removed from $\inst I_2$ at the variable partitioning step. We run the approximation algorithm for Unique Games by Charikar et al. on $\inst I_2$ using the original SDP solution restricted to vectors $\{\mathbf{x}_a: x\in \VV{2}, a\in D\}$. This is a valid SDP relaxation because in the instance $\inst I_2$, unlike the instance $\inst I_1$, we do not restrict the domain of variables $x$ to $D_x$. The cost of this SDP solution is at most $\eps$. As shown in , the weight of constraints violated by the algorithm is at most $O(\sqrt{\eps\log \|D\|})$. We get the solution for $\inst I$ by combining solutions for $\inst I_1$ and $\inst I_2$, and assigning values chosen at the variable partitioning step to the variables from the set $\VV{0}$. # Proof of Theorem [\[thm:nicelevels\]](#thm:nicelevels){reference-type="ref" reference="thm:nicelevels"} {#sect:proofof3} In this section we prove Theorem [\[thm:nicelevels\]](#thm:nicelevels){reference-type="ref" reference="thm:nicelevels"}. The proof will use constraint languages with relations of arity greater than two. In order to talk about such instances we need to extend the definition of a pattern. Note that patterns (in the sense of Section [5.1](#sect:pattern){reference-type="ref" reference="sect:pattern"}) are instances (with some added structure) and the realizations of patterns are solutions. We use the pattern/instance and solution/realization duality to generalize the notion of a pattern. Moreover we often treat patterns as instances and (whenever it makes sense) instances as patterns. We will often talk about path/tree instances; they are defined using the incidence multigraph. The incidence multigraph of an instance $\inst J$ is bipartite, its the vertex set consists of variables and constraints of $\inst J$ (which form the two parts), and if a variable $x$ appears $j$ times in a constraint $C$ then the vertices corresponding to $x$ and $C$ have $j$ edges between them. An instance is connected if its incidence multigraph is connected; an instance is a tree instance if it is connected and its incidence multigraph has no multiple edges and no cycles. A leaf variable in a tree instance is a variable which corresponds to a leaf in the incidence multigraph, and we say that two variables are neighbours if they appear together in a scope of some constraint (i.e., the corresponding vertices are connected by a path of length 2 in the incidence multigraph). Note that the incidence multigraph of a path pattern in a binary instance (treated as an instance, as described in the first paragraph of this section) is a path, and of an $n$-tree pattern is a tree with $n$ leaves. The next definition captures, among other things, the connection between the pattern (treated as an instance) and the instance in which the pattern is defined. Let $\inst J_1$ and $\inst J_2$ be two instances over the same constraint language. An (instance) homomorphism $e\colon \inst J_1 \to \inst J_2$ is a mapping that maps each variable of $\inst J_1$ to a variable of $\inst J_2$ and each constraint of $\inst J_1$ to constraint of $\inst J_2$ in such a way that every constraint $((y_1,\dots,y_k),R)$ in $\inst J_1$ is mapped to $((e(y_1),\dots,e(y_k)),R)$. Using these new notions, a path pattern in an instance $\inst I$ (see the definition in Section [5.1](#sect:pattern){reference-type="ref" reference="sect:pattern"}) can alternatively be defined as an instance, with beginning and end chosen among the leaf variables, whose incidence graph is a path from beginning to end, together with a homomorphism into $\inst I$. Similarly we define a path pattern in a (not necessarily binary) instance $\inst I$ as an instance $\inst J$, with chosen beginning/end leaf variables, whose incidence graph, after removing all the other vertices of degree one, is a path from beginning to end, together with a homomorphism $e\colon \inst J\rightarrow\inst I$. Addition of path patterns and propagation are defined in an analogous way as for patterns with binary constraints (see Section [5.1](#sect:pattern){reference-type="ref" reference="sect:pattern"}). For a $k$-ary relation $R$, let $\pr_i(R)=\{a_i\mid (a_1,\ldots,a_i,\ldots,a_k)\in R\}$. A CSP instance $\inst J$ is called arc-consistent in sets $D_x$ ($x$ ranges over variables of $\inst J$) if, for any variable $x$ and any constraint $((x_1,\ldots,x_k),R)$ in $\inst J$, if $x_i=x$ then $\pr_i(R)=D_x$. We say that a CSP instance $\inst J$ satisfies condition (PQ) in sets $D_x$ if 1. $\inst J$ is arc-consistent in these sets and [\[cond:pqac\]]{#cond:pqac label="cond:pqac"} 2. for any variable $x$, [\[cond:pqpq\]]{#cond:pqpq label="cond:pqpq"} any path patterns $p,q$ from $x$ to $x$, and any $a\in D_x$ there exists $j$ such that $a\in \{a\} + (j(p+q) + p)$. Note that if the instance $\inst J$ is binary then (PQ) implies for all $n$ (setting $D_x^i=D$ if $i=\|D\|+1$ and $D_x^i=D_x$ if $i<\|D\|+1$). The following fact, a special case of Theorem A.2 in , provides solutions for (PQ) instances. Finally, a standard algebraic notion has not been defined yet: having fixed $\Gamma$ over a set $D$, a subset $A\subseteq D$ is a subuniverse if, for any polymorphism $g$ of $\Gamma$, we have $g(a_1,a_2,\ldots)\in A$ whenever $a_1,a_2,\ldots \in A$. For any $S\subseteq D$, the subuniverse generated by $S$ is defined as \[\{g(a_1,\ldots,a_r)\mid r\ge 1, a_1,\ldots,a_r\in S, \text{$g$ is an $r$-ary polymorphism of $\Gamma$} \}\] ## Into the proof We begin the proof of Theorem [\[thm:nicelevels\]](#thm:nicelevels){reference-type="ref" reference="thm:nicelevels"}. We fix a binary language $\Gamma$ compatible with a $(n+1)$-ary NU polymorphism and an instance $\inst I$ of in $\CSP\Gamma$ which satisfies $\IPQ{n}$ with sets $D_x^\ell$. Note that we can assume that all $D_x^\ell$'s are subuniverses. If this is not the case, we replace each $D_x^\ell$ with the subuniverse generated by it. It is easy to check that (after the change) the instance $\inst I$ still satisfies $\IPQ{n}$ with such enlarged $D_x^\ell$'s. For each variable $x$, choose and fix an arbitrary index $i$ such that $D_x^i=D_x^{i+1}$ and call it the level of $x$. Note that each variable has a level (since the sets $D_x^\ell$ are non-empty and $\ell$ ranges from 1 to $\|D\|+1$). Let $V^i$ denote the set of variables of level $i$ and $V^{< i}, V^{\le i},\dots$ be defined in the natural way. Our proof of Theorem [\[thm:nicelevels\]](#thm:nicelevels){reference-type="ref" reference="thm:nicelevels"} will proceed by applying Theorem [\[thm:nice\]](#thm:nice){reference-type="ref" reference="thm:nice"} to $\inst I$ restricted to $V^1$, then to $V^2$ and so on. However, in order to obtain compatible solutions, we will add constraints to the restricted instances. ## The instances in levels Let $\inst I^i$ (for $i\leq \|D\|$) be an instance such that: 1. $V^i$ is the set of variables of $\inst I^i$; 2. $\inst I^i$ contains, for every every $n$-tree pattern $t$ of $\inst I$, the constraint $((x_1,\dots,x_k),R)$ defined in the following way: [\[enum:newconstraints\]]{#enum:newconstraints label="enum:newconstraints"} let $v_1,\dotsc,v_k$ be all the vertices of $t$ labeled by variables from $V^i$, then $x_1,\dots,x_k$ are the labels of $v_1,\dots,v_k$ respectively and \[R = \{ (r(v_1),\dots,r(v_k)) \mid r \text{ is a $i$-realization of $t$ (i.e. inside sets $D_x^{i+1}$)} \}.\] This definition has a number of immediate consequences: First, every binary constraint between two variables from $V^i$ is present in $\inst I^i$ (as it defines a two-element $n$-tree). Second, note that if for some $n$-tree contains a vertex $v_j$ in $V^i$ which is not a leaf then by splitting the tree $t$ at $v_j$ (with $v_j$ included in both parts) we obtain two trees defining constraints which together are equivalent to the constraint defined by $t$. This implies that by including only the constraints defined by $n$-trees $t$ such that only the leaves can be from $V^i$, we obtain an equivalent (i.e. having the same set of solutions) instance. Throughout most of the proof we will be working with such a restricted instance. In this instance the arity of constraints is bounded by $n$. Since the arity of a constraint in $\inst I^i$ is bounded and the size of the universe is fixed, $\inst I^i$ is a finite instance, even though some constraints in it can be defined via infinitely many $n$-tree patterns. It is easy to see that all the relations in the constraints are preserved by all the polymorphisms of $\Gamma$. The instance $\inst I^i$ is arc-consistent with sets $D_x^i(=D_x^{i+1})$: Let $((x_1,\dotsc,x_k),R)$ be a constraint defined by $v_1,\dotsc, v_k$ in $t$ and let $a\in D_{x_j}^i$. By there is a realization of $t$ in $D_x^{i+1}$ mapping $v_j$ to $a$ and thus $D_{x_j}^i\subseteq \pr_j R$. On the other hand, as $D_{x_j}^i=D_{x_j}^{i+1}$ and every tuple in $R$ comes from a realization inside the sets $D_x^{i+1}$'s, we get $\pr_j R\subseteq D_{x_j}^i$. Next we show that $\inst I^i$ has property (PQ). Part [\[cond:pqac\]](#cond:pqac){reference-type="ref" reference="cond:pqac"} of the definition was established in the paragraph above. For part [\[cond:pqpq\]](#cond:pqpq){reference-type="ref" reference="cond:pqpq"}, let $p$ and $q$ to be arbitrary path patterns from $x$ to $x$ in $\inst I^i$. Define $p'$ and $q'$ to be the paths of trees in $\inst I$ obtained, from $p$ and $q$, respectively, by replacing (in the natural way) each constraint in $p$ and $q$ with the tree that defines it (we use the fact that each constraint is defined by leaves of a tree). We apply property for $\inst I$ with $\ell=i$ and patterns $p'$ and $q'$ to get that, for any $x\in V^i$ and any $a\in D_x^i$, there is a number $j$ such that $a\in \{a\}+^i(j(p'+q')+p')$. The property (PQ) follows immediately. Since $\inst I^i$ has the property (PQ) then, by Theorem [\[thm:nice\]](#thm:nice){reference-type="ref" reference="thm:nice"}, it has a solution. The solution to $\inst I$ will be obtained by taking the union of appropriately chosen solutions to $\inst I^1,\dotsc, \inst I^{\|D\|}$. ## Invariant of the iterative construction A global solution, denoted $\sol\colon V\to D$, is constructed in steps. At the start, we define it for the variables in $V^1$ by choosing an arbitrary solution to $\inst I^1$. In step $i$ we extend the definition of $\sol$ from $V^{<i}$ to $V^{\le i}$, using a carefully chosen solution to $\inst I^i$. Our construction will maintain the following condition: Note that, after the first step, the condition $(E_1)$ is guaranteed by the constraints of $\inst I^1$. Assume that we are in step $i$: we have already defined $\sol$ on $V^{<i}$ and condition $(E_{i-1})$ holds. Our goal is to extend $\sol$ by a solution of $\inst I^i$ in such a way that $(E_{i})$ holds. The remainder of Section [7](#sect:proofof3){reference-type="ref" reference="sect:proofof3"} is devoted to proving that such a solution exists. Once we accomplish that, we are done with the proof: Condition $(E_{i})$ implies that $\sol$ is defined on $V^{\le i}$, and for every constraint $((x,y),R)$ between $x,y\in V^{\le i}$ the pattern from $x$ to $y$ containing a single edge labeled by $((x,y),R)$ is an $n$-tree. This implies that $\sol$ satisfies $((x,y),R)$ i.e. it is a solution on $V^{\le i}$. After establishing $(E_{\|D\|})$ we obtain a solution to $\inst I$. ## Restricting $\inst I^i$ We begin by defining a new instance $\inst K^i$: it is defined almost identically to $\inst I^i$, but in part [\[enum:newconstraints\]](#enum:newconstraints){reference-type="ref" reference="enum:newconstraints"} of the definition we require that the realization $r$ sends vertices from $V^{<i}$ according to $\sol$. As in the case of $\inst I^i$ we can assume that all the constraints are defined by leaves of the tree. Thus every $n$-tree pattern with no internal vertices in $V^{i}$ defines one constraint in $\inst I^i$ and another in $\inst K^i$. Just like $\inst I^i$, the instance $\inst K^i$ is finite. Note that we yet need to establish that constraints of $\inst K^i$ are non-empty, but the following claim, where $f$ is the fixed $(n+1)$-ary near unanimity polymorphism, holds independently. In order to proceed we need to show that the instance $\inst K^i$ contains a non-empty, arc-consistent subinstance, i.e. an arc-consistent instance (in some non-empty sets $D_x$) obtained from $\inst K^i$ by restricting every constraint in it so that each coordinate can take value only in the appropriate set $D_x$. A proof of this claim is the subject of the next section. ## Arc-consistent subinstance of $\inst K^i$ In order to proceed with the proof we need an additional definition. Let $e\colon \inst J_1 \to \inst J_2$ be an instance homomorphism. If for any variable $y$ of $\inst J_1$ and any constraint $((x_1,\dots,x_k),R)$ of $\inst J_2$ with $e(y) = x_i$ (for some $i$) the constraint $((x_1,\dots,x_k),R)$ has exactly one preimage $((y_1,\dots,y_k),R)$ with $y = y_i$, we say that $e$ is a covering. A universal covering tree instance $\uct(\inst J)$ of a connected instance $\inst J$ is a (possibly countably infinite) tree instance $\inst T$ together with a covering $e\colon\inst T\to \inst J$ satisfying some additional properties. If $\inst J$ is a tree instance, then one can take $\uct(\inst J)=\inst J$, otherwise $\uct(\inst J)$ is always infinite. If an instance $\inst J$ is disconnected then $\uct(\inst J)$ is a disjoint union of universal covering tree instances for connected components of $\inst J$. Several equivalent (precise) definitions of $\UCT$ can be found in Section 5.4 of or Section 4 of . For our purposes, it is enough to mention that, for any $\inst J$, the instance $\uct(\inst J)$ (with covering $e$) has the following two properties. For any two variables $v,v'$ satisfying $e(v)=e(v')$ there exists an endomorphism $h$ of $\uct(\inst I)$ (i.e a homomorphism into itself) sending $v$ to $v'$ and such that $e\circ h =e$. Similarly for constraints $C$ and $C'$ if $e(C)=e(C')$ then there is an endomorphism $h$ such that $h(C)=C'$ and $e\circ h = e$. It is well known that $\UCT(\inst J)$ has a solution if and only if $\inst J$ has an arc-consistent subinstance. Consider $\UCT(\inst K^i)$ and fix a covering $e': \UCT(\inst K^i) \rightarrow \inst K^i$. Let $\mathcal{T}^i$ be an instance obtained from $\UCT(\inst K^i)$ by replacing each constraint $C$ in it by a tree that defines $e'(C)$, each time introducing a fresh set of variables for the internal vertices of the trees. Let $e$ be the instance homomorphism from $\mathcal{T}^i$ to $\inst I$ defined in the natural way. We call a solution (or a partial solution) to $\mathcal{T}^i$ nice if it maps each $v$ into $D_{e(v)}^{i+1}$ and moreover if $e(v)\in V^{<i}$ then $v$ is mapped to $\sol(e(v))$. It should be clear that nice solutions to $\mathcal{T}^i$ correspond to solutions of $\UCT(\inst K^i)$ (although the correspondence is not one-to-one). We will denote the arc-consistent subinstance of $\inst K^i$ (which is about to be constructed) by $\inst L^i$. The variables of $\inst L^i$ and $\inst K^i$ (or indeed $\inst I^i$) are the same. For every constraint $(\tuple x, R)$ in $\inst K^i$ we introduce a constraint $(\tuple x, R')$ into $\inst L^i$ where \[R' = \{\tuple{a}\colon \tuple a = s(\tuple y) \text{ where $s$ is a solution to $\UCT(\inst K^i)$ and $e'((\tuple y,R)) = (\tuple x, R)$}\}\] where $e'$ is an instance homomorphism mapping $\UCT(\inst K^i)$ to $\inst K^i$. In other words we restrict a relation in a constraint of $\inst K^i$ by allowing only the tuples which appear in a solution of the $\UCT(\inst K^i)$ (at this constraint). All the relations of $\inst L^i$ are preserved by all the polymorphisms of $\Gamma$, and are non-empty (by Claim 2). The fact that $\inst L^i$ is arc-consistent is an easy consequence of the endomorphism structure of universal covering trees. Finally Claim [\[claim:absorbing\]](#claim:absorbing){reference-type="ref" reference="claim:absorbing"} holds for $\inst L^i$: ## A solution to $\inst K^i$ In order to find a solution to $\inst L^i$, we will use Corollary B.2 from . We state it here in a simplified form using the following notation: for subuniverses $A'\subseteq A$, we say that $A'$ nu-absorbs $A$ if, for some NU polymorphism $f$, $f(a_1,\ldots,a_n)\in A'$ whenever $a_1,\ldots,a_n\in A$ and at most one $a_i$ is in $A\setminus A'$. Similarly, if $R'\subseteq R$ are relations preserved by all polymorphisms of $\Gamma$ we say $R'$ nu-absorbs $R$, if for some near-unanimity operation $f$ taking all arguments from $R'$ except for one which comes from $R$ produces a result in $R'$. We will apply the corollary above using $\inst I^i$ for $\inst I$ and $\inst L^i$ for $\inst I'$. By our construction, $\inst I^i$ satisfies condition (PQ), and the sets $D_x^i$ (which play the role of $A_x$) are subuniverses of $D$. On the other hand $\inst L^i$ is arc-consistent and all the relations involved in it are closed under the polymorphisms of $\Gamma$. Claim [\[claim:absorbingL\]](#claim:absorbingL){reference-type="ref" reference="claim:absorbingL"} shows that each relation $R'$ nu-absorbs the corresponding $R$. By arc-consistency, the projection of $R'$ on a variable $x$ is the same for each constraint $((x_1,\ldots,x_n),R')$ containing $x$, call the corresponding sets $A'_x$. Since each $R'$ nu-absorbs $R$, it follows that each $A'_x$ nu-absorbs the corresponding $A_x$. The corollary implies that we can restrict the instance $\inst L^i$ to obtain an instance satisfying (PQ). By Theorem [\[thm:nice\]](#thm:nice){reference-type="ref" reference="thm:nice"} such an instance, and thus both $\inst K^i$ and $\inst L^i$, has a solution. ## Finishing the proof We choose any solution to $\inst K^i$ and extend the global solution $\sol$ to $V^i$ according to it. There exists a solution on $V^{\le i}$, because every constraint between two variables from this set is either in $V^{<i}$ or defines a two-variable $n$-tree which was used to define a constraint in $\inst K^i$. It remains to prove that, with such an extension, condition $(E_{i})$ holds. Let $t$ be an $n$-tree pattern in $\inst I$. If it has no variables mapped to $V^{i}$, then $(E_i)$ follows from $(E_{i-1})$. Assume that it has such variables. By splitting $t$ at internal vertices mapped to $V^i$, it is enough to consider the case when only leaves of $t$ are mapped to $V^i$. Then $t$ defines a constraint $(\tuple x,R)$ in $\inst K^i$. The solution to $\inst K^i$ mapping $\tuple x$ to $\tuple a\in R$ and the evaluation of $t$ witnessing that $\tuple a$ belongs to $R$ can be taken to satisfy $(E_i)$ for $t$. Theorem [\[thm:nicelevels\]](#thm:nicelevels){reference-type="ref" reference="thm:nicelevels"} is proved. # Full proof of Theorem [\[the:main\]](#the:main){reference-type="ref" reference="the:main"}(1) {#sec:correctness-of-algorithm} In this subsection we prove Propositions [\[prop:expected-lost-weight\]](#prop:expected-lost-weight){reference-type="ref" reference="prop:expected-lost-weight"} and [\[prop:consistence\]](#prop:consistence){reference-type="ref" reference="prop:consistence"}. The following equalities, which can be directly verified, are used repeatedly in this section: for any subsets $A,B$ of $D$ and any feasible solution $\{\mathbf{x}_a\}$ of the SDP relaxation of $\inst I$ it holds that $\\|\mathbf{x}_A\\|^2=\mathbf{x}_A\mathbf{y}_D$ and $\\|\mathbf{y}_B-\mathbf{x}_A\\|^2=\mathbf{x}_{D\setminus A}\mathbf{y}_B+\mathbf{x}_A\mathbf{y}_{D\setminus B}$. ## Analysis of Preprocessing step 2 {#sec:preproc} In some of the proofs it will be required that $\alpha\leq c_0$ for some constant $c_0$ depending only on $\|D\|$. This can be assumed without loss of generality, since we can adjust constants in $O$-notation in Theorem [\[the:main\]](#the:main){reference-type="ref" reference="the:main"}(1) to ensure that $\eps\le c_0$ (and we know that $\alpha\le \eps$). We will specify the requirements on the choice of $c_0$ as we go along. ## Proof of Proposition [\[prop:expected-lost-weight\]](#prop:expected-lost-weight){reference-type="ref" reference="prop:expected-lost-weight"} We will prove that the total weight of constraints removed in each step 0-5 of the algorithm in Section [5.3.2](#sec:algorithm){reference-type="ref" reference="sec:algorithm"} is $O(\alpha^\kappa)$. ## Proof of Proposition [\[prop:consistence\]](#prop:consistence){reference-type="ref" reference="prop:consistence"} All patterns appearing in this subsection are in $\inst I'$. The following notion will be used several times in our proofs: Let $t$ be a tree and let $y$ be one of its nodes. We say that a subtree $t'$ of $t$ is separated by vertex $y$ if $t'$ is maximal among all the subtrees of $t$ that contain $y$ as a leaf. Finally, setting $A=\{a\}$ in Lemma [\[lem:consist\]](#lem:consist){reference-type="ref" reference="lem:consist"} gives Proposition [\[prop:consistence\]](#prop:consistence){reference-type="ref" reference="prop:consistence"}. # Full proof of Theorem [\[the:main\]](#the:main){reference-type="ref" reference="the:main"}(2) {#sec:thm2} In this section, we prove Theorem [\[the:main\]](#the:main){reference-type="ref" reference="the:main"}(2). A brief outline of the proof is given in Section [6](#sec:overview-thm2){reference-type="ref" reference="sec:overview-thm2"}. Throughout this section, $\inst I=(V,{\mathcal C})$ is a $(1-\eps)$-satisfiable instance of $\CSP\Gamma$ where $\Gamma$ consists of implicational constraints. ## SDP Relaxation {#sec:thm2:SDP} We use SDP relaxation ([\[sdpobj\]](#sdpobj){reference-type="ref" reference="sdpobj"})--([\[sdp4\]](#sdp4){reference-type="ref" reference="sdp4"}) from Section [4](#sec:SDP){reference-type="ref" reference="sec:SDP"}. For convinience, we write the SDP objective function as follows. \[\begin{gathered} \sum_{C \in {\mathcal C}\text{ equals } (x = a) \vee (y = b)} w_C \vprod{(\mathbf{v}_0-\mathbf{x}_{a})}{(\mathbf{v}_0-\mathbf{y}_{b})} \\ {}+\frac{1}{2}\sum_{C \in {\mathcal C}\text{ equals } x = \pi(y)} \,\sum_{a\in D} w_C \\| \mathbf{x}_{\pi(a)}-\mathbf{y}_{a}\\|^2 \\ {}+ \sum_{C \in {\mathcal C}\text{ equals }x \in P} w_C \left(\sum_{a\in D\setminus P} \\|\mathbf{x}_{a}\\|^2 \right). \label{SDP} \end{gathered}\] This expression equals ([\[sdpobj\]](#sdpobj){reference-type="ref" reference="sdpobj"}) because of SDP constraint ([\[sdp3\]](#sdp3){reference-type="ref" reference="sdp3"}). Minimize \[\begin{gathered} \sum_{C \in {\mathcal C}\text{ equals } (x = a) \vee (y = b)} w_C \vprod{(\mathbf{v}_0-\mathbf{x}_{a})}{(\mathbf{v}_0-\mathbf{y}_{b})} \\ {}+\frac{1}{2}\sum_{C \in {\mathcal C}\text{ equals } x = \pi(y)} \,\sum_{a\in D} w_C \\| \mathbf{x}_{\pi(a)}-\mathbf{y}_{a}\\|^2 \\ {}+ \sum_{C \in {\mathcal C}\text{ equals }x \in P} w_C \left(\sum_{a\in D\setminus P} \\|\mathbf{x}_{a}\\|^2 \right) \label{SDP} \end{gathered}\] subject to \[\begin{aligned} &\vprod{ \mathbf{x}_{a}}{ \mathbf{y}_{b}} \geq 0 & x,y\in V,\ a,b\in D \label{sdp:pos} \\ &\vprod{ \mathbf{x}_{a}}{ \mathbf{x}_{b}} = 0 & x\in V,\ a,b\in D,\ a\neq b \\ &\sum_{a\in D} \mathbf{x}_{a} = \mathbf{v}_0 & x\in V \label{sdp:sum-to-one}\\ &\\|\mathbf{x}_{a}-\mathbf{z}_{c}\\|^2 \leq \\|\mathbf{x}_{a}-\mathbf{y}_{b}\\|^2 + \\|\mathbf{y}_{b}-\mathbf{z}_{c}\\|^2 \hskip-8em\label{sdp:triangle-ineq}\\ && x,y,z\in V,\ a,b,c\in D\notag\\ &\\|\mathbf{v}_0\\|^2 = 1.\label{sdp:unitnorm} \end{aligned}\] This SDP and the SDP we presented in Section [4](#sec:SDP){reference-type="ref" reference="sec:SDP"} are almost identical. Their objective functions are equal, because of constraints ([\[sdp3\]](#sdp3){reference-type="ref" reference="sdp3"}) and ([\[sdp:sum-to-one\]](#sdp:sum-to-one){reference-type="ref" reference="sdp:sum-to-one"}). For convinience, we write the objective function differently in SDP ([\[SDP\]](#SDP){reference-type="ref" reference="SDP"})--([\[sdp:unitnorm\]](#sdp:unitnorm){reference-type="ref" reference="sdp:unitnorm"}). The only difference between the SDPs is the presence of the "triangle inequalities" ([\[sdp:triangle-ineq\]](#sdp:triangle-ineq){reference-type="ref" reference="sdp:triangle-ineq"}). We introduce them, because we use the algorithm from for Unique Games, which assumes that the SDP has triangle inequalities. As discussed before (Lemma [\[le:prep1\]](#le:prep1){reference-type="ref" reference="le:prep1"}) we can assume that $\eps\geq 1/m^2$ where $m$ is the number of constraints. We solve SDP with error $\delta=1/m^2$ obtaining a solution, denoted $\mathsf{SDP}$, with objective value $O(\epsilon)$. Note that every feasible SDP solution satisfies the following conditions. \[\begin{aligned} &\\| \mathbf{x}_{a} \\|^2 = \vproddot{ \mathbf{x}_{a}}{\bigl(\mathbf{v}_0-\sum_{b\neq a} \mathbf{x}_{b}\bigr)} = \vproddot{ \mathbf{x}_{a}}{\mathbf{v}_0}-\sum_{b\neq a}\vproddot{\mathbf{x}_a}{\mathbf{x}_b} = \vprod{ \mathbf{x}_{a}}{ \mathbf{v}_0 },\label{eq:length}\\[1mm] &\vprod{ \mathbf{x}_{a} }{ \mathbf{y}_{b}} = \vproddot{ \mathbf{x}_{a}}{(\mathbf{v}_0-\sum_{b'\neq b}\mathbf{y}_{b'})} \label{eq:2SATrequirement} = \\|\mathbf{x}_{a}\\|^2-\sum_{b'\neq b} \vprod{ \mathbf{x}_{a}}{ \mathbf{y}_{b'}} \leq \\|\mathbf{x}_{a}\\|^2, \\[1mm] & \\|\mathbf{x}_{a}\\|^2-\\|\mathbf{y}_{b}\\|^2 = \\|\mathbf{x}_{a}-\mathbf{y}_{b}\\|^2 + 2(\vprod{\mathbf{x}_{a}}{\mathbf{y}_{b}}-\\|\mathbf{y}_{b}\\|^2) \leq \\|\mathbf{x}_{a}-\mathbf{y}_{b}\\|^2, \label{eq:triangle}\\[1mm] & \vprod{(\mathbf{v}_0-\mathbf{x}_{a})}{( \mathbf{v}_0-\mathbf{y}_{b})} = \vprod{ \sum_{a'\neq a} \mathbf{x}_{a'}} {\sum_{b'\neq b} \mathbf{y}_{b'}} \geq 0.\label{eq:positivity} \end{aligned}\] ## Variable Partitioning Step {#sec:2pre} In this section, we describe the first step of our algorithm. In this step, we assign values to some variables, partition all variables into three groups ${\cal V}_0$, ${\cal V}_1$ and ${\cal V}_2$, and then split the instance into two sub-instances ${\cal I}_1$ and ${\cal I}_2$. #### Vertex Partitioning Procedure. Choose a number $r \in (0, 1/6)$ uniformly at random. Do the following for every variable $x$. 1. Let $D_x = \{a: 1/2-r < \vprod{ \mathbf{x}_{a}}{ \mathbf{v}_0}\}$. 2. Depending on the size of $D_x$ do the following: 1. If $\|D_x\| = 1$, add $x$ to ${\cal V}_0$ and assign $x = a$, where $a$ is the single element of $D_x$. 2. If $\|D_x\| > 1$, add $x$ to ${\cal V}_1$ and restrict $x$ to $D_x$ (see below for details). 3. If $D_x = \varnothing$, add $x$ to ${\cal V}_2$. Note that each variable in ${\cal V}_0$ is assigned a value; each variable $x$ in ${\cal V}_1$ is restricted to a set $D_x$; each variable in ${\cal V}_2$ is not restricted. We say that an assignment is admissible if it assigns a value in $D_x$ to every $x \in {\cal V}_1$ and it is consistent with the partial assignment to variables in ${\cal V}_0$. From now on we restrict our attention only to admissible assignments. We remove those constraints that are satisfied by every admissible assignment (our algorithm will satisfy all of them). Specifically, we remove the following constraints: 1. UG constraints $x=\pi(y)$ with $x, y \in {\cal V}_0$ that are satisfied by the partial assignment; 2. disjunction constraints $(x = a) \vee (y = b)$ such that either $x \in {\cal V}_0$ and $x$ is assigned value $a$, or $y \in {\cal V}_0$ and $y$ is assigned value $b$; 3. unary constraints $x \in P$ such that either $x \in {\cal V}_0$ and the value assigned to $x$ is in $P$, or $x \in {\cal V}_1$ and $D_x \subseteq P$. We denote the set of satisfied constraints by ${\cal C}_s$. Let ${\cal C}'={\cal C}\setminus {\cal C}_s$ be the set of remaining constraints. We now define a set of violated constraints---those constraints that we conservatively assume will not be satisfied by our algorithm (even though some of them might be satisfied by the algorithm). We say that a constraint $C\in {\cal C}'$ is violated if at least one of the following conditions holds: 1. $C$ is a unary constraint on a variable $x \in {\cal V}_0 \cup {\cal V}_1$. 2. $C$ is a disjunction constraint $(x = a) \vee (y = b)$ and either $x \notin {\cal V}_1$, or $y \notin {\cal V}_1$ (or both). 3. $C$ is a disjunction constraint $(x = a) \vee (y = b)$, and $x, y \in {\cal V}_1$, and either $a\notin D_x$, or $b\notin D_y$ (or both). 4. $C$ is a UG constraint $x = \pi(y)$, and at least one of the variables $x$, $y$ is in ${\cal V}_0$. 5. $C$ is a UG constraint $x = \pi(y)$, and one of the variables $x$, $y$ is in ${\cal V}_1$ and the other is in ${\cal V}_2$. 6. $C$ is a UG constraint $x = \pi(y)$, $x, y\in {\cal V}_1$ but $D_x \neq \pi(D_y)$. We denote the set of violated constraints by ${\cal C}_v$ and let ${\cal C}'' = {\cal C}'\setminus {\cal C}_v$. We restrict our attention to the set ${\cal C}''$. There are four types of constraints in ${\cal C}''$. 1. disjunction constraints $(x = a) \vee (y = b)$ with $x, y \in {\cal V}_1$ and $a\in D_x$, $b\in D_y$; 2. UG constraints $x = \pi (y)$ with $x, y \in {\cal V}_1$ and $D_x = \pi(D_y)$; 3. UG constraints $x = \pi (y)$ with $x, y \in {\cal V}_2$; 4. unary constraints $x \in P$ with $x\in{\cal V}_2$. Denote the set of type 1 and 2 constraints by ${\cal C}_1$, and type 3 and 4 constraints by ${\cal C}_2$. Let ${\cal I}_1$ be the sub-instance of ${\cal I}$ on variables ${\cal V}_1$ with constraints ${\cal C}_1$ in which every variable $x$ is restricted to $D_x$, and ${\cal I}_2$ be the sub-instance of ${\cal I}$ on variables ${\cal V}_2$ with constraints ${\cal C}_2$. In Sections [9.3](#solve-i1){reference-type="ref" reference="solve-i1"} and [9.4](#solve-i2){reference-type="ref" reference="solve-i2"}, we show how to solve ${\cal I}_1$ and ${\cal I}_2$, respectively. The total weight of constraints violated by our solution for ${\cal I}_1$ will be at most $O(\sqrt{\eps})$; The total weight of constraints violated by our solution for ${\cal I}_2$ will be at most $O(\sqrt{\eps\log{\|D\|} })$. Thus the combined solution will satisfy a subset of the constraints of weight at least $1-O(\sqrt{\eps \log{\|D\|}})$. ## Solving Instance ${\cal I}_1$ {#solve-i1} In this section, we present an algorithm that solves instance ${\cal I}_1$. The algorithm assigns values to variables in ${\cal V}_1$ so that the total weight of violated constraints is at most $O(\sqrt{\eps})$. We remove all constraints ${\cal C}_{bad}$ from ${\cal I}_1$ and obtain an instance ${\cal I}_1'$ (with the domain for each variable $x$ now restricted to $D_x$). We construct an SDP solution $\{\mathbf{\tilde x}_{a}\}$ for ${\cal I}_1'$. We let \[\mathbf{\tilde x}_{\alpha_x} = \mathbf{x}_{\alpha_x} \quad\text{and}\quad \mathbf{\tilde x}_{\beta_x} = \mathbf{v}_0-\mathbf{x}_{\alpha_x}.\] We define $S_{x\alpha_x} = \{\alpha_x\}$ and $S_{x\beta_x} = D\setminus S_{x\alpha_x}$. Since $\mathbf{\mathbf{\tilde x}}_{\beta_x} = \mathbf{v}_0-\mathbf{x}_{\alpha_x} = \mathbf{x}_{S_{x\beta_x}}$, we have, \[\mathbf{\tilde x}_{a} = \mathbf{x}_{S_{xa}} \quad \text{for every } a\in D_x.\label{eq:redef-x-a}\] Note that $a\in S_{xa}$ for every $a\in D_x$. Finally, we note that ${\cal I}_1'$ is a Boolean 2-CSP instance. We round solution $\{\mathbf{\tilde x}_{a}\}$ using the rounding procedure by Charikar et al. for Boolean 2-CSP (when $\|D\| = 2$, the SDP relaxation used in is equivalent to SDP ([\[sdpobj\]](#sdpobj){reference-type="ref" reference="sdpobj"})--([\[sdp4\]](#sdp4){reference-type="ref" reference="sdp4"})). We get an assignment of variables in ${\cal V}_1$. The weight of constraints in ${\cal C}_{1}\setminus {\cal C}_{bad}$ violated by this assignment is at most $O(\sqrt{\eps})$. Since $w({\cal C}_{bad}) = O(\sqrt{\eps})$, the weight of constraints in ${\cal C}_{1}$ violated by the assignment is at most $O(\sqrt{\eps})$. ## Solving Instance ${\cal I}_2$ {#solve-i2} Instance ${\cal I}_2$ is a unique games instance with additional unary constraints. We restrict the SDP solution for $\cal I$ to variables $x \in {\cal V}_2$ and get a solution for the unique game instance ${\cal I}_2$. Note that since we do not restrict the domain of variables $x\in \VV{2}$ to $D_x$, the SDP solution we obtain is feasible. The SDP cost of this solution is at most $\mathsf{SDP}$. We round this SDP solution using a variant of the algorithm by Charikar et al. that is presented in Section 3 of the survey ; this variant of the algorithm does not need $\ell_2^2$-triangle-inequality SDP constraints. Given a $(1-\eps)$-satisfiable instance of Unique Games, the algorithm finds a solution with the weight of violated constraints at most $O(\sqrt{\eps \log{\|D\|}})$. We remark that paper considers only unique game instances. However, in , we can restrict the domain of any variable $x$ to a set $S_x$ by setting $\mathbf{x}_a = 0$ for $a\in D\setminus S_x$. Hence, we can model unary constraints as follows. For every unary constraint $x\in P$, we introduce a dummy variable $z_{x,P}$ and restrict its domain to the set $P$. Then we replace each constraint $x\in P$ with the equivalent constraint $x = z_{x,P}$. The weight of the constraints violated by the obtained solution is at most $O(\sqrt{\eps \log{\|D\|}})$. Finally, we combine results proved in Sections [9.2](#sec:2pre){reference-type="ref" reference="sec:2pre"}, [9.3](#solve-i1){reference-type="ref" reference="solve-i1"}, and [9.3](#solve-i1){reference-type="ref" reference="solve-i1"} and obtain Theorem [\[the:main\]](#the:main){reference-type="ref" reference="the:main"}(2).	{'timestamp': '2018-01-09T02:13:26', 'yymm': '1607', 'arxiv_id': '1607.04787', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04787'}
# Introduction {#sec:Introduction} Topological insulators (TIs) form a class of materials with unique properties, associated with a non-trivial topology of their quasiparticle band structure (for a review, see Refs. ). The key feature of two-dimensional (2D) and three-dimensional (3D) TIs is the existence of special gapless edge and surface states, respectively, while the bulk states of those materials are gapped. The hallmark property of the surface states is their topological protection. Mathematically, the nontrivial topological properties of time-reversal (TR) invariant TIs are generally described by multiple copies of the $Z_2$ invariants found by Kane and Mele. This implies that the energy band gap should close at the boundary between topological and trivial insulator (e.g., vacuum) giving rise to the occurrence of the gapless interface states and the celebrated bulk-boundary correspondence. The discovery of the $Z_2$ topology in TIs is an important breakthrough because it showed that nontrivial topology can be embedded in the band structure and that the presence of an external magnetic field is not mandatory for the realization of topological phases. Another distinctive feature of the 3D TIs is a relativistic-like energy spectrum of the surface states, whose physical origin is related to a strong spin-orbit coupling. Indeed, the surface states on each of the surfaces are described by 2D massless Dirac fermions in an irreducible 2$\times$`<!-- -->`{=html}2 representation, with a single Dirac point in the reciprocal space. For comparison, quasiparticles in graphene demonstrate similar properties, but have four inequivalent Dirac cones due to a spin and valley degeneracy that makes certain aspects of their physics very different from those of the surface states in TIs. In our study below, we will concentrate only on the case of the strong 3D TIs whose surface states are protected by the topology of the bulk bands in combination with the TR symmetry. This leads to the locking of momenta and spin degrees of freedom and, consequently, to the formation of a helical Dirac (semi)metal state. Such a state is characterized by the electron antilocalization and the absence of backscattering. The phenomenon of antilocalization has deep mathematical roots and is usually explained by an additional Berry's phase $\pi$ that is acquired when an electron circles a Dirac point. From the physical viewpoint, when scattering on an impurity, an electron must change its spin in order to preserve its chirality. Such a process is possible only in the case of magnetic impurities which break explicitly the TR symmetry. Experimentally, a linear relativistic-like dispersion law of the surface states is observed in Bi$_{1-x}$Sb$_x$, Bi$_2$Se$_3$, Bi$_2$Te$_3$, Sb$_2$Te$_3$, Bi$_2$Te$_2$Se, and other materials by using angle resolved photoemission spectroscopy (ARPES). Furthermore, scanning tunneling microscopy and scanning tunneling spectroscopy provide additional information about the topological nature of the surface states, such as the quasiparticles interference patterns around impurities and defects. The Fourier analysis of these patterns has shown that the backscattering between $\mathbf{k}$ and $-\mathbf{k}$ is highly suppressed in Bi$_{1-x}$Sb$_x$ and Bi$_2$Te$_3$ in accord with the TR symmetry protection. The existence of an odd number of Dirac nodes leads to other exotic properties associated with surface states of TIs, e.g., an axion electromagnetic response, an unusual surface Hall conductance, etc. It is well known that electrons confined to two dimensions can form numerous interaction-induced phases. By using numerical calculations, it was shown in Ref. that it is energetically favorable for the 2D electron liquid in a weak magnetic field to form domains with empty and fully filled higher Landau levels. Depending on the number of Landau levels filled, the corresponding charge density wave (CDW) phase is realized with a "stripe" or "bubble" pattern. By using the simplest model of the surface states in a magnetic field with strong local repulsion and a long-range Coulomb interaction included perturbatively, it was suggested that a similar CDW phase with a "stripe" or "bubble" pattern can be also realized on the surface of 3D TIs for supercritical values of a local repulsion constant. For subcritical local repulsion, the composite Fermi liquid (CFL) phase is expected. It is interesting that composite fermions in conventional 2D electron gas at half-filling were recently suggested to be massless Dirac (quasi)particles similar to the surface quasiparticles of TIs. This result was also checked numerically in Ref. , where it was shown that at the half-filling the particle-hole symmetry for composite fermions plays the same role as the TR symmetry for the 2D Dirac fermions and, consequently, the backscattering off symmetry preserving impurities is also forbidden. The influence of an external electric field on the exciton condensation in thin films of TIs was studied in Refs. , where it was shown that the electron condensate effectively joins the surfaces of a thin film and leads to the formation of a pairing gap. However, this is important only in thin ($l_z \lesssim 8~\mbox{nm}$) films of TIs and can be ignored in sufficiently thick slabs. The exciton condensate exhibits unusual properties including a stable zero mode and a fractional charge $\pm e/2$ carried by a singly quantized vortex in the exciton condensate. The dynamical gap generation in a simple model of TIs was also considered in Ref. . Just like a magnetic field, an external electric field may play an important role in the dynamics of the surface states in a 3D TI slab. In this paper, we study the dynamical gap generation and the phase diagram of a TI slab placed in external magnetic and electric fields perpendicular to the slab surfaces. (Note that the case of the parallel fields is rather trivial. While a parallel magnetic field does not affect the orbital motion, a parallel electric field produces a current on the surface.) We argue that a uniform phase with both dynamically generated Dirac and Haldane gaps is realized in sufficiently strong (weak) electric (magnetic) fields. Although the explicit calculations performed in this paper use the model parameters suitable for Bi$_2$Se$_3$, the main qualitative conclusions should be valid for all similar TIs. The paper is organized as follows. The effective Hamiltonian of the surface states in the simplest model of a topological insulator with short-and long-range interactions is described in Sec. [2](#sec:model-all){reference-type="ref" reference="sec:model-all"}. The set of gap equations at finite temperature is derived in Sec. [3](#sec:gap-equation-Coulomb){reference-type="ref" reference="sec:gap-equation-Coulomb"} and its solutions in electric and magnetic fields are obtained numerically in Sec. [4](#sec:gap-equation-Coulomb-results-kappa){reference-type="ref" reference="sec:gap-equation-Coulomb-results-kappa"}. The qualitative description of the inhomogeneous phase with two stripes is given in Sec. [5](#sec:stripe){reference-type="ref" reference="sec:stripe"}. The main results are discussed and summarized in Secs. [6](#sec:Discussion){reference-type="ref" reference="sec:Discussion"} and [7](#sec:Conclusion){reference-type="ref" reference="sec:Conclusion"}, respectively. For convenience, throughout the paper, we set $\hbar=c=1$. # Model {#sec:model-all} By projecting the 3D bulk Hamiltonian onto the subspace of surface states (see Refs. for a detailed consideration), the following effective Hamiltonian for the top surface of a 3D TI is obtained: \[\begin{aligned} \mathcal{H}_{\mathrm{top}\,\,\mathrm{surf}}(\mathbf{k})=C+v_F\left(\bm{\sigma}\times\mathbf{k}\right)_z +\mathcal{O}\left(\mathbf{k}^2\right) = C+\left( \begin{array}{cc} 0 & v_F(ik_x+k_y) \\ v_F(-ik_x+k_y) & 0 \\ \end{array} \right)+ \mathcal{O}\left(\mathbf{k}^2\right), \label{model-H-2D} \end{aligned}\] where $C$ is a constant, $\bm{\sigma}=\left(\sigma_x, \sigma_y\right)$ are the Pauli matrices, $v_F=4.1~\mbox{eV\AA}=6.2\times10^5~\mbox{m/s}$ is the Fermi velocity, and $\mathbf{k}=\left(k_x, k_y\right)$ is the surface momentum. The effective surface Hamiltonian for the bottom surface is obtained from the Hamiltonian of the top surface by the inversion $\mathbf{k} \to-\mathbf{k}$ (see Sec. III.C in Ref. ). It is worth noting that the effective surface Hamiltonian is valid only at sufficiently small chemical potentials, when the bulk states are gapped. Therefore, the corresponding energy cutoff can be approximated by the bulk band gap, i.e., $\Lambda\simeq \Delta_{\rm bulk}$. In the case of Bi$_2$Se$_3$, for example, $\Delta_{\rm bulk} \approx 0.35~\mbox{eV}$. The resulting model Hamiltonian, describing quasiparticle states on the top and bottom surfaces of the 3D TI in constant electric and magnetic fields applied perpendicular to the surfaces of the slab, is given by $H^{(0)}=H^{(0)}_{+}\oplus H^{(0)}_{-}$, where \[H^{(0)}_{\lambda}=\int d^2\mathbf{r}\,\,\psi^{\dag}_{\lambda}(\mathbf{r})\left( \begin{array}{cc} m^{(0)}-\mu_{\lambda}^{(0)} & iv_F(\pi_x-i\pi_y) \\ -iv_F(\pi_x +i\pi_y) &-m^{(0)}-\mu_{\lambda}^{(0)} \\ \end{array} \right)\psi_{\lambda}(\mathbf{r}). \label{model-H-s-0}\] Here $\lambda=\pm$ denotes the top and bottom surfaces, respectively, $\mu_{\lambda }^{(0)}$ is the surface electrochemical potential, $\bm{\pi} \equiv-i \bm{\nabla} + e\mathbf{A}$ is the canonical momentum, $\mathbf{A}=\left(0, Bx\right)$ is the vector potential that describes the constant magnetic field $\mathbf{B}$ pointing in the $z$ direction, and $e$ is the electron charge. Note that in Eq. ([\[model-H-s-0\]](#model-H-s-0){reference-type="ref" reference="model-H-s-0"}) we redefined the wave function on the bottom surface by replacing $\psi_{-}\to\sigma_z\psi_{-}$. As the notation suggests, the value of $\mu^{(0)}_{\lambda}$ may depend on the surface index $\lambda$. Indeed, this is quite natural in the model at hand since fixing charge densities on the top and bottom surfaces requires an introduction of the corresponding local electrochemical potentials. In view of a large surface $g$-factor, $g_{s}=18\pm4$, the Zeeman splitting is important in TIs. This spin splitting is included in Hamiltonian ([\[model-H-s-0\]](#model-H-s-0){reference-type="ref" reference="model-H-s-0"}) as the bare gap parameter $m^{(0)} =g_s\mu_{B} B/2$, where $\mu_{B}=5.788 \times 10^{-5}~\mbox{eV/ T}$ is the Bohr magneton. Before proceeding with the analysis of the model, it is convenient to rewrite the model Hamiltonian ([\[model-H-s-0\]](#model-H-s-0){reference-type="ref" reference="model-H-s-0"}) in terms of the Dirac matrices. It is well known that there are two irreducible representations of the Clifford-Dirac algebra in (2+1)-dimensions, e.g., see Ref. . One of them is \[\gamma^0=\sigma_z, \quad \gamma^1=i\sigma_x, \quad \gamma^2=i\sigma_y \label{model-gamma-matrices}\] and the other irreducible representation is obtained by changing $\gamma^{\mu}\to-\gamma^{\mu}$ with $\mu=0, 1, 2$ in Eq. ([\[model-gamma-matrices\]](#model-gamma-matrices){reference-type="ref" reference="model-gamma-matrices"}). In terms of the Dirac matrices ([\[model-gamma-matrices\]](#model-gamma-matrices){reference-type="ref" reference="model-gamma-matrices"}), the free Hamiltonian ([\[model-H-s-0\]](#model-H-s-0){reference-type="ref" reference="model-H-s-0"}) takes the following form: \[H^{(0)}_{\lambda}=\int d^2\mathbf{r}\,\,\bar{\psi}_{\lambda}(\mathbf{r})\left( -\mu^{(0)}_{\lambda}\gamma^0+v_F(\bm{\pi}\cdot\bm{\gamma})+m^{(0)} \right)\psi_{\lambda}(\mathbf{r}), \label{model-H-0-matrices}\] where $\bar{\psi}_{\lambda}(\mathbf{r})=\psi^{\dag}_{\lambda}(\mathbf{r})\gamma^0$. When an external electric field is applied perpendicularly to the surfaces of the TI slab, the gapless surface states will tend to completely screen the field out. Indeed, from a physics viewpoint, the TI slab is like a Faraday cage made of gapless (metallic) surface states enclosing a gapped (insulating) interior. This implies that there should be no electric field inside a (sufficiently thick) TI slab. Enforcing this condition allows one to determine the charge densities and electrochemical potentials on the surfaces. In terms of the charge densities on the top and bottom surfaces, one has \[\rho_{\lambda }=\lambda \epsilon_0 \mathcal{E}, \label{model-DOS-zero-field}\] where $\mathcal{E}$ is the external electric field pointing in the $z$-direction, $\epsilon_0 \approx 8.854 \times10^{-12}~\mbox{F/m}$ is the permittivity of free space, and $2\epsilon_0 \mathcal{E}$ corresponds to the difference of the charge densities of the top and bottom surfaces needed to compensate the external electric field. Under the parity transformation $P$ in (2+1) dimensions, which changes the sign of a spatial coordinate, i.e., $(x,y) \to (-x,y)$, the two-component spinors transform as follows: $P\psi(t,x,y)P^{-1}=\sigma_x\psi(t,-x,y)$. Clearly, the last term in Hamiltonian ([\[model-H-0-matrices\]](#model-H-0-matrices){reference-type="ref" reference="model-H-0-matrices"}) breaks parity, as well as the TR symmetry $T\psi(t,x,y)T^{-1}=\sigma_y\psi^{}(-t,x,y)$. This mass term is known in the literature as the Haldane mass $\sum_{\lambda}m_H\,\bar{\psi}_{\lambda}\psi_{\lambda}$. A parity and TR invariant mass is also possible in the model with two irreducible representations. It is given by the Dirac mass term $\sum_{\lambda}m_D\,\lambda\bar{\psi}_{\lambda}\psi_{\lambda}$ with the parity transformation defined by $\psi_{\lambda=+1} \to \sigma_x \psi_{\lambda=-1}$ and $\psi_{\lambda=-1} \to \sigma_x \psi_{\lambda=+1}$. (Note that, in the TI slab model, this transformation interchanges the states on the different spatially separated surfaces.) While a Chern--Simons mass term for the gauge field is induced via one-loop polarization when the Haldane mass is present, the Chern-Simons term is absent in the case of the Dirac mass. The spontaneous breaking of parity in $(2+1)$-dimensional QED was studied in Ref. . In this study, the model interaction Hamiltonian $H_{\rm int}$ includes both a long-range Coulomb and a short-range local four-fermion interactions \[H_{\rm int} = \frac{e^2}{8\pi \epsilon_0 \kappa_{\rm surf}}\int d^2\mathbf{r}d^2\mathbf{r}^{\prime}\,\frac{\Psi^{\dagger}(\mathbf{r}) \Psi(\mathbf{r}) \Psi^{\dagger}(\mathbf{r}^{\prime})\Psi(\mathbf{r}^{\prime})}{\|\mathbf{r}-\mathbf{r}^{\prime}\|} +\frac{G_{\rm int}}{2}\int d^2\mathbf{r}\, \Psi^{\dagger}(\mathbf{r})\Psi(\mathbf{r}) \Psi^{\dagger}(\mathbf{r})\Psi(\mathbf{r}), \label{model-H-int}\] where $\Psi(\mathbf{r})=\left(\psi_{\lambda=+1}(\mathbf{r}),\psi_{\lambda=-1}(\mathbf{r})\right)^T$. The first term in $H_{\rm int}$ describes the long-range Coulomb interaction and takes into account the effective surface dielectric constant $\kappa_{\rm surf}=\left(1+\kappa_{\rm bulk}\right)/2 \approx 56$, where the bulk dielectric constant $\kappa_{\rm bulk}\approx113$ for Bi$_2$Se$_3$. The second term captures the on-site local repulsion, parametrized by the dimensionful coupling constant $G_{\rm int}$. In view of the large bulk dielectric constant and assuming a large slab thickness, we neglect the intersurface interaction and the possible formation of an intersurface exciton condensate. Thus, the full Hamiltonian of our model is given by the sum of the free and interaction Hamiltonians in Eqs. ([\[model-H-0-matrices\]](#model-H-0-matrices){reference-type="ref" reference="model-H-0-matrices"}) and ([\[model-H-int\]](#model-H-int){reference-type="ref" reference="model-H-int"}). # Gap equations {#sec:gap-equation-Coulomb} In this section, we study the gap generation in the effective model of a sufficiently thick TI slab described in the previous section. The inverse free surface fermion propagator is given by \[iS^{-1}_{\lambda}(u,u^\prime) = \left[(i\partial_t+\mu^{(0)}_{\lambda })\gamma^0-v_F(\bm{\pi}\cdot\bm{\gamma})-m^{(0)}\right]\delta^{3}(u-u^\prime), \label{gapEq-Coulomb-sinverse}\] where $u=\left(t, \mathbf{r}\right)$ denotes a space-time coordinate. By using this as a guide, we assume the following rather general ansatz for the inverse full surface fermion propagator: \[\begin{aligned} iG^{-1}_{\lambda}(u,u^\prime)= \left[(i\partial_t+\mu_{\lambda })\gamma^0-v_F(\bm{\pi}\cdot\bm{\gamma})-m_{\lambda}\right]\delta^{3}(u-u^\prime), \label{gapEq-Coulomb-ginverse} \end{aligned}\] where $m_{\lambda }$ is a dynamically generated gap (mass) which, in general, includes both Haldane and Dirac gaps and $\mu_{\lambda }$ denotes the dynamical electrochemical potential. Note that all dynamical parameters in the full propagator are assumed to be functions of $(\bm{\pi}\cdot\bm{\gamma})^2l^2$, where $l = 1/\sqrt{\|eB\|}$ is the magnetic length. Therefore, in the end, they all depend on the Landau level index $n$. Because of the long-range interaction, in principle, the renormalization of the wave function should be included in the full propagator ([\[gapEq-Coulomb-ginverse\]](#gapEq-Coulomb-ginverse){reference-type="ref" reference="gapEq-Coulomb-ginverse"}). This can be formally done by replacing the Fermi velocity $v_F$ with a dynamical function $F_{\lambda}$. It is well justified, however, to neglect the renormalization of the Fermi velocity and replace it with $v_F$. Indeed, even in the case of graphene with an unscreened Coulomb interaction, the renormalized Fermi velocity is generically $10\%$ to $30\%$ larger than the corresponding bare value $v_F$. Because of a much larger surface dielectric constant and, consequently, a much smaller coupling constant, the Coulomb interaction will play a minor role in the Fermi velocity renormalization and, as we will show below, in the generation of dynamical gaps in TIs. In order to represent the inverse propagator in the form of a Landau-level expansion, we use the following complete set of eigenstates (for details, see Appendix A in Ref. ): \[\psi_{n, k_y}(\mathbf{r})=\frac{1}{\sqrt{2^{n+1} \pi l n!}} H_n\left(k_yl+\frac{x}{l}\right)e^{-\frac{1}{2l^2}\left(x+k_yl^2\right)^2} e^{is_Bk_yy}, \label{gapEq-Coulomb-psi-nky}\] where $H_n(x)$ are the Hermite polynomials and $s_{B}=\mbox{sign}(eB)$. By making use of the results in Appendix [\[sec:wf-Green\]](#sec:wf-Green){reference-type="ref" reference="sec:wf-Green"}, we derive the following inverse fermion propagators in the mixed frequency-momentum representation: \[\begin{aligned} \label{gapEq-Coulomb-sinverse-LL-no-phase} S^{-1}_{\lambda}(\omega,\mathbf{r},\mathbf{r}^\prime)&=& e^{i\Phi(\mathbf{r},\mathbf{r}^\prime)}\tilde{S}_{\lambda}^{-1}(\omega,\mathbf{r}-\mathbf{r}^\prime) ,\\ \label{gapEq-Coulomb-ginverse-LL-no-phase} G^{-1}_{\lambda}(\omega,\mathbf{r},\mathbf{r}^\prime)&=& e^{i\Phi(\mathbf{r},\mathbf{r}^\prime)}\tilde{G}_{\lambda}^{-1}(\omega,\mathbf{r}-\mathbf{r}^\prime). \end{aligned}\] Here $\Phi(\mathbf{r}, \mathbf{r}^{\prime})=-eB(x+x^{\prime})(y-y^{\prime})/2$ is the famous Schwinger phase and the translation invariant parts of the inverse propagators are given by \[\begin{aligned} \label{gapEq-Coulomb-sinverse-LL} iS^{-1}_{\lambda}(\omega,\mathbf{r}-\mathbf{r}^\prime)&=& \frac{e^{-\eta/2}}{2\pi l^2}\sum_{n=0}^{\infty} \Big\{ s_B (\omega+\mu^{(0)}_{\lambda })\left[P_{+} L_{n-1}(\eta)-P_{-} L_{n}(\eta)\right]-m^{(0)}\left[P_{+} L_{n-1}(\eta)+P_{-} L_{n}(\eta) \right] \nonumber \\ &-&\frac{i}{l^2}v_F (\bm{\gamma}\cdot\mathbf{r}) L^1_{n-1}(\eta)\Big\},\\ \label{gapEq-Coulomb-ginverse-LL} iG^{-1}_{\lambda}(\omega,\mathbf{r}-\mathbf{r}^\prime)&=& \frac{e^{-\eta/2}}{2\pi l^2}\sum_{n=0}^{\infty} \Big\{ s_B (\omega+\mu_{n,\lambda })\left[P_{+} L_{n-1}(\eta)-P_{-} L_{n}(\eta)\right]-m_{n,\lambda }\left[P_{+} L_{n-1}(\eta)+P_{-} L_{n}(\eta) \right]\nonumber \\ &-&\frac{i}{l^2}v_F (\bm{\gamma}\cdot\mathbf{r}) L^1_{n-1}(\eta)\Big\}, \end{aligned}\] where $P_{\pm} =\left(1 \pm s_B \gamma^0\right)/2$, $\eta= (\mathbf{r}-\mathbf{r}^{\prime})^2/(2l^2)$, and $L^{j}_{n} \left(x\right)$ are the generalized Laguerre polynomials (by definition $L_n\equiv L_n^0$). In order to study the dynamical gap generation, we utilize the Baym--Kadanoff (BK) formalism, which leads to a self-consistent Schwinger-Dyson equation for the fermion propagator. In contrast to a perturbative analysis, the BK formalism can capture nonperturbative effects such as spontaneous symmetry breaking. To leading order in coupling, the BK effective action in the model under consideration is given by Eq. ([\[app-action-BK\]](#app-action-BK){reference-type="ref" reference="app-action-BK"}) in Appendix [\[sec:App-action-true\]](#sec:App-action-true){reference-type="ref" reference="sec:App-action-true"}. In view of the geometry of conducting states of our TI system, it should not be too surprising that the effective action ([\[app-action-BK\]](#app-action-BK){reference-type="ref" reference="app-action-BK"}) has a form similar to that in bilayer graphene (compare with Eq. (9) in Ref. ). The extremum of the effective action $\frac{\delta \Gamma(G)}{\delta G_{\lambda}} =0$ defines the following Schwinger-Dyson equation for the full fermion propagator (for details, see Appendix [\[sec:App-action-true\]](#sec:App-action-true){reference-type="ref" reference="sec:App-action-true"}): \[iG^{-1}_{\lambda}(u,u^\prime) = iS^{-1}_{\lambda}(u,u^\prime)-e^2 \gamma^0 G_{\lambda}(u,u^{\prime}) \gamma^0 D(u^{\prime}-u)-G_{\rm int}\left\{ \gamma^0 G_{\lambda}(u,u) \gamma^0-\gamma^0\, \mbox{tr}[\gamma^0G_{\lambda}(u,u)]\right\}\delta^{3}(u-u^{\prime}), \label{gapEq-Coulomb-gap}\] where the trace in the last term is taken over the spinor indices and the Hartree term due to the Coulomb interaction is absent. This is justified because of the overall neutrality of the sample, i.e., \[Q_b-e\, \sum_{\lambda=\pm}\mbox{tr}[\gamma^0G_{\lambda}(u,u)]=0, \label{gapEq-Coulomb-Hartree}\] where $Q_b$ denotes the background charge due to the external gates. We note, however, that it does not make sense to drop the Hartree-type term due to the contact interaction. Therefore, the corresponding term is kept in curly brackets in Eq. ([\[gapEq-Coulomb-gap\]](#gapEq-Coulomb-gap){reference-type="ref" reference="gapEq-Coulomb-gap"}). The propagator mediating the Coulomb interaction is denoted by $D(u)$. Its explicit expression is given by \[D(u)=\int\frac{d\omega d^2\mathbf{k}}{(2\pi)^3}D(\omega, \mathbf{k})e^{-i\omega t+i\mathbf{k}\cdot\mathbf{r}}\approx\delta(t) \frac{1}{4\pi\epsilon_0\kappa_{\rm surf}} \int \frac{dk}{2\pi} \frac{kJ_0(kr)}{k+\Pi(0,k)}, \label{gapEq-Coulomb-D}\] where $J_0(x)$ is the Bessel function. In the last expression, we neglected the dependence of the polarization function $\Pi(\omega,k)$ on $\omega$, which corresponds to an instantaneous approximation. Such an approximation may be reasonable for the TI surfaces, where charge carriers propagate much slower than the speed of light and, thus, the retardation effects are negligible. It is worth noting, however, that the instantaneous approximation has a tendency to underestimate the strength of the Coulomb interaction. Just like the inverse propagators in Eqs. ([\[gapEq-Coulomb-sinverse-LL-no-phase\]](#gapEq-Coulomb-sinverse-LL-no-phase){reference-type="ref" reference="gapEq-Coulomb-sinverse-LL-no-phase"}) and ([\[gapEq-Coulomb-ginverse-LL-no-phase\]](#gapEq-Coulomb-ginverse-LL-no-phase){reference-type="ref" reference="gapEq-Coulomb-ginverse-LL-no-phase"}), the propagators themselves have the same Schwinger phase. The full propagator, in particular, takes the following explicit form: \[\begin{aligned} \label{gapEq-Coulomb-G-phase} G_{\lambda}(\omega, \mathbf{r}, \mathbf{r}^{\prime}) &=& e^{i\Phi(\mathbf{r},\mathbf{r}^\prime)}\tilde{G}_{\lambda}(\omega, \mathbf{r}-\mathbf{r}^{\prime}) ,\\ \label{gapEq-Coulomb-G-no-phase} \tilde{G}_{\lambda}(\omega, \mathbf{r}-\mathbf{r}^{\prime})&=& \frac{e^{-\eta/2}}{2\pi l^2}\sum_{n=0}^{\infty} \Bigg\{\frac{s_B\left(\omega +\mu_{n,\lambda } \right)\left[ L_{n-1}(\eta)P_{+}-L_{n}(\eta)P_{-}\right]}{\left(\omega+\mu_{n,\lambda } +i0\,\mbox{sign}{(\omega)}\right)^2-M_n^2} \nonumber\\ &+&\frac{m_{n,\lambda } \left[L_{n-1}(\eta)P_{+} +L_{n}(\eta)P_{-}\right]-i\frac{v_F}{l^2}L_{n-1}^1(\eta) \left(\bm{\gamma}\cdot(\mathbf{r}-\mathbf{r}^{\prime})\right) }{\left(\omega+\mu_{n,\lambda } +i0\,\mbox{sign}{(\omega)}\right)^2-M_n^2} \Bigg\}, \end{aligned}\] where $M_n=\sqrt{\left(m_{n, \lambda }\right)^2+\epsilon_{B}^2n}$ and $\epsilon_{B}= \sqrt{2v_F^2\|eB\|}$ is the Landau energy scale. The inverse and full fermion propagators at finite temperature are easily obtained through the standard replacement $\omega\to i\omega_{m^{\prime}}=i\pi T(2m^{\prime}+1)$. By factorizing the Schwinger phase on both sides of Eq. ([\[gapEq-Coulomb-gap\]](#gapEq-Coulomb-gap){reference-type="ref" reference="gapEq-Coulomb-gap"}), we arrive at the following gap equation for the translation invariant part of the full propagator: \[i\tilde{G}^{-1}_{\lambda}(\omega, \mathbf{r}) = i\tilde{S}^{-1}_{\lambda}(\omega, \mathbf{r})-\alpha v_F \!\! \int\frac{d\Omega}{2\pi} \frac{dk}{2\pi} \frac{kJ_0(kr)}{k+\Pi(0,k)} \gamma^0 \tilde{G}_{\lambda}(\Omega, \mathbf{r}) \gamma^0-G_{\rm int}\!\int\frac{d\Omega}{2\pi} \delta^2(\mathbf{r}) \left\{ \gamma^0 \tilde{G}_{\lambda}(\Omega, \mathbf{r}) \gamma^0-\gamma^0\, \mbox{tr}[\gamma^0\tilde{G}_{\lambda}(\Omega, \mathbf{r})]\right\}. \label{gapEq-Coulomb-gap-1}\] Here we introduced the following notation $\alpha=e^2/(4\pi \epsilon_0 v_F \kappa_{\rm surf})$. In the case of Bi$_2$Se$_3$, in particular, $\alpha\approx0.062$. Although it is hard to estimate $G_{\rm int}$ reliably, its origin is the Coulomb repulsion on distance scales comparable to the lattice spacing. Thus, it may be reasonable to use the following approximate model value: \[G_{\rm int}=\frac{\alpha v_F^2\kappa_{\rm surf}}{\Delta_{\rm bulk}} \approx 168.7~\mbox{eV\AA}^2, \label{gapEq-Coulomb-Gint}\] where the factor $\kappa_{\rm surf}$ was introduced in order to compensate for polarization effects in the definition of $\alpha$. Indeed, polarization effects should be negligible at small distances. It is worth noting that the corresponding dimensionless constant \[g_{\rm int}= \frac{G_{\rm int}\Delta_{\rm bulk}}{8\sqrt{2\pi} v_F^2}\approx 0.18 \label{gapEq-Coulomb-g-def}\] is rather small. In fact, it is an order of magnitude smaller than the critical value $g_{cr}=\sqrt{\pi}$ needed for generating a gap in a (2+1)-dimensional model in the absence of a magnetic field. Because of this and because of the strong suppression of the Coulomb interaction by the large dielectric constant, no dynamical generation of a gap is expected in such a TI material in the absence of an external magnetic field. Consequently, the magnetic catalysis will play a crucial role in the generation of dynamical gaps in TIs. (For a recent review on magnetic catalysis, see Ref. .) By using the explicit form of the fermion propagator ([\[gapEq-Coulomb-G-no-phase\]](#gapEq-Coulomb-G-no-phase){reference-type="ref" reference="gapEq-Coulomb-G-no-phase"}) on the right-hand side of Eq. ([\[gapEq-Coulomb-gap-1\]](#gapEq-Coulomb-gap-1){reference-type="ref" reference="gapEq-Coulomb-gap-1"}), we can easily calculate the integral over $\Omega$ (or the sum over the Matsubara frequency at nonzero temperature, see Appendix [\[sec:Greens-function\]](#sec:Greens-function){reference-type="ref" reference="sec:Greens-function"}). Afterwards, by multiplying both sides of the gap equation ([\[gapEq-Coulomb-gap-1\]](#gapEq-Coulomb-gap-1){reference-type="ref" reference="gapEq-Coulomb-gap-1"}) by $e^{-\eta/2}L_{n^{\prime}}(\eta)$ or $e^{-\eta/2}(\bm{\gamma}\cdot\mathbf{r})L_{n^{\prime}}^1(\eta)$ and then integrating over $\mathbf{r}$, the complete set of equations for the dynamical parameters can be straightforwardly obtained. In particular, the gap equations for the lowest Landau level (LLL) parameters are given by \[\begin{aligned} \Delta_{\rm eff, \lambda }&=&\mu_{\lambda}^{(0)}+s_Bm^{(0)}+ \alpha \frac{v_F}{2l} \Bigg\{ \mathcal{K}^{(0)}_{0,0}\left[ 1-2n_F\left(\Delta_{\mathrm{eff}, \lambda}\right) \right]-\sum_{n^{\prime}=1}^{\infty}\mathcal{K}^{(0)}_{n^{\prime},0}\Big[ n_F(M_{n^{\prime}}+\mu_{n^{\prime}, \lambda })-n_F(M_{n^{\prime}}-\mu_{n^{\prime}, \lambda })\nonumber\\ &-& s_B m_{n^{\prime}, \lambda } \frac{1-n_F(M_{n^{\prime}}+\mu_{n^{\prime}, \lambda })-n_F(M_{n^{\prime}}-\mu_{n^{\prime}, \lambda })}{M_{n^{\prime}}} \Big] \Bigg\} +\frac{G_{\rm int}}{4\pi l^2}\Bigg\{ \sum_{n^{\prime}=1}^{\infty} \left[n_F\left(M_{n^{\prime}}+\mu_{n^{\prime}, \lambda }\right)-n_F\left(M_{n^{\prime}}-\mu_{n^{\prime}, \lambda }\right)\right]\nonumber\\ &+& \sum_{n^{\prime}=1}^{\infty} s_B m_{n^{\prime}, \lambda } \frac{1-n_F\left(M_{n^{\prime}}+\mu_{n^{\prime}, \lambda }\right)-n_F \left(M_{n^{\prime}}-\mu_{n^{\prime}, \lambda } \right)}{M_{n^{\prime}}} \Bigg\}, \label{gapEq-Coulomb-gap-Delta} \end{aligned}\] where $n_{F}(x)=1/\left(e^{x/T}+1\right)$ is the Fermi-Dirac distribution. Notice that we introduced an effective LLL electrochemical potential $\Delta_{\mathrm{eff}, \lambda} =\mu_{0, \lambda}+ s_{B}m_{0, \lambda}$ because the LLL parameters $\mu_{0,\lambda}$ and $m_{0,\lambda}$ cannot be unambiguously defined separately and only their combination $\Delta_{\mathrm{eff}, \lambda}$ has a well-defined physical meaning. Similarly, the equations for the dynamical parameters associated with higher Landau levels read as \[\begin{aligned} m_{n, \lambda }&=&m^{(0)}+ s_B \alpha \frac{v_F}{4l} \Bigg\{ \mathcal{K}^{(0)}_{0,n}\left[ 1-2n_F\left(\Delta_{\mathrm{eff}, \lambda}\right) \right]+ \sum_{n^{\prime}=1}^{\infty}\mathcal{K}^{(0)}_{n^{\prime}-1,n-1} \Big[ n_F(M_{n^{\prime}}+\mu_{n^{\prime}, \lambda })-n_F(M_{n^{\prime}}-\mu_{n^{\prime}, \lambda }) \nonumber\\ &+&s_B m_{n^{\prime},\lambda } \frac{1-n_F(M_{n^{\prime}}+\mu_{n^{\prime},\lambda })-n_F(M_{n^{\prime}}-\mu_{n^{\prime},\lambda })} {M_{n^{\prime}}} \Big] \nonumber\\ &-& \sum_{n^{\prime}=1}^{\infty}\mathcal{K}^{(0)}_{n^{\prime},n}\left[ n_F(M_{n^{\prime}}+\mu_{n^{\prime}, \lambda })-n_F(M_{n^{\prime}}-\mu_{n^{\prime}, \lambda })-s_B m_{n^{\prime}, \lambda } \frac{1-n_F(M_{n^{\prime}}+\mu_{n^{\prime}, \lambda })-n_F(M_{n^{\prime}}-\mu_{n^{\prime}, \lambda })}{M_{n^{\prime}}} \right] \Bigg\} \nonumber\\ &+& \frac{G_{\rm int}}{8\pi l^2}\left\{ s_B\left[1-2n_F\left(\Delta_{\mathrm{eff}, \lambda}\right)\right] +2\sum_{n^{\prime}=1}^{\infty} m_{n^{\prime}, \lambda } \frac{1-n_F\left(M_{n^{\prime}}+\mu_{n^{\prime}, \lambda }\right)-n_F\left(M_{n^{\prime}}-\mu_{n^{\prime}, \lambda } \right)}{M_{n^{\prime}}}\right\}, \label{gapEq-Coulomb-gap-m} \end{aligned}\] \[\begin{aligned} \mu_{n, \lambda }&=&\mu^{(0)}_{\lambda }+\alpha \frac{v_F}{4l} \Bigg\{ \mathcal{K}^{(0)}_{0,n} \left[ 1-2n_F\left(\Delta_{\mathrm{eff}, \lambda}\right) \right]-\sum_{n^{\prime}=1}^{\infty}\mathcal{K}^{(0)}_{n^{\prime}-1,n-1} \Big[ n_F(M_{n^{\prime}}+\mu_{n^{\prime}, \lambda })-n_F(M_{n^{\prime}}-\mu_{n^{\prime}, \lambda }) \nonumber\\ &+&s_B m_{n^{\prime}, \lambda } \frac{1-n_F(M_{n^{\prime}}+\mu_{n^{\prime}, \lambda })-n_F(M_{n^{\prime}}-\mu_{n^{\prime}, \lambda })} {M_{n^{\prime}}} \Big] -\sum_{n^{\prime}=1}^{\infty}\mathcal{K}^{(0)}_{n^{\prime},n}\Big[ n_F(M_{n^{\prime}}+\mu_{n^{\prime}, \lambda })-n_F(M_{n^{\prime}}-\mu_{n^{\prime}, \lambda }) \nonumber\\ &-&s_B m_{n^{\prime}, \lambda } \frac{1-n_F(M_{n^{\prime}}+\mu_{n^{\prime}, \lambda })-n_F(M_{n^{\prime}}-\mu_{n^{\prime}, \lambda })} {M_{n^{\prime}}} \Big] \Bigg\} \nonumber\\ &-&\frac{G_{\rm int}}{8\pi l^2}\left\{ \left[1-2n_F\left(\Delta_{\mathrm{eff}, \lambda}\right)\right]-2\sum_{n^{\prime}=1}^{\infty} \left[n_F\left(M_{n^{\prime}}+\mu_{n^{\prime}, \lambda }\right)-n_F\left(M_{n^{\prime}}-\mu_{n^{\prime}, \lambda }\right)\right] \right\}. \label{gapEq-Coulomb-gap-mu} \end{aligned}\] The kernel coefficients $\mathcal{K}^{(0)}_{m,n}$ that capture the long-range interaction effects in the gap equations are defined in Eq. ([\[app-K-def\]](#app-K-def){reference-type="ref" reference="app-K-def"}). In this study, for simplicity, we neglect all screening effects, i.e., we set $\Pi(0,k)=0$. Then, the numerical analysis greatly simplifies because the coefficients $\mathcal{K}^{(0)}_{m,n}$ can be calculated analytically, see Eq. ([\[app-K-Pi0\]](#app-K-Pi0){reference-type="ref" reference="app-K-Pi0"}). In addition to the gap equation ([\[gapEq-Coulomb-gap-1\]](#gapEq-Coulomb-gap-1){reference-type="ref" reference="gapEq-Coulomb-gap-1"}), the constraints for the surface charge densities in Eq. ([\[model-DOS-zero-field\]](#model-DOS-zero-field){reference-type="ref" reference="model-DOS-zero-field"}) should be also satisfied. In terms of the model parameters, the explicit form of the constraint reads as \[-\frac{e}{4\pi l^2}\left\{ \left[1-2n_F\left(\Delta_{\mathrm{eff}, \lambda}\right)\right]-2\sum_{n=1}^{\infty} \left[n_F\left(M_n+\mu_{n, \lambda }\right)-n_F\left(M_n-\mu_{n, \lambda }\right)\right] \right\} =\lambda \epsilon_0 \mathcal{E}, \label{gapEq-Coulomb-DOS}\] where we used the definition for the surface charge densities in terms of the fermion propagator, i.e., $\rho_{\lambda }=e\,\mbox{tr}[G_{\lambda}(u,u) \gamma^0]$. Because the surface charge density is fixed by the external electric field, the electrochemical potential $\mu^{(0)}_{\lambda}$ is not an independent parameter. It is determined together with the other dynamical parameters by solving the system of Eqs. ([\[gapEq-Coulomb-gap-Delta\]](#gapEq-Coulomb-gap-Delta){reference-type="ref" reference="gapEq-Coulomb-gap-Delta"}) through ([\[gapEq-Coulomb-DOS\]](#gapEq-Coulomb-DOS){reference-type="ref" reference="gapEq-Coulomb-DOS"}). # Numerical results {#sec:gap-equation-Coulomb-results-kappa} In this section, we present our numerical solutions of gap equations ([\[gapEq-Coulomb-gap-Delta\]](#gapEq-Coulomb-gap-Delta){reference-type="ref" reference="gapEq-Coulomb-gap-Delta"}) through ([\[gapEq-Coulomb-gap-mu\]](#gapEq-Coulomb-gap-mu){reference-type="ref" reference="gapEq-Coulomb-gap-mu"}), together with the constraint in Eq. ([\[gapEq-Coulomb-DOS\]](#gapEq-Coulomb-DOS){reference-type="ref" reference="gapEq-Coulomb-DOS"}). Before proceeding to the analysis, it is convenient to give the formal definition of the Dirac and Haldane gaps in the TI model at hand. While the original surface gaps $m_{n,+}$ and $m_{n,-}$ (with $n \ge 1$) have a straightforward physical meaning, the symmetry properties of the ground state can be better understood in terms of Dirac and Haldane gaps, i.e., \[m_{n, D}= \frac{m_{n,+}-m_{n,-}}{2}, \qquad m_{n, H}= \frac{m_{n,+}+m_{n,-}}{2}. \label{gap-Coulomb-results-Dirac-Haldanem-tmu}\] Strictly speaking, these gaps cannot be associated with the usual Dirac and Haldane masses in (2+1)-dimensional QED (see, Ref. ), because $m_{\pm}$ in TIs correspond to spatially separated surfaces. Since the free Hamiltonian ([\[model-H-s-0\]](#model-H-s-0){reference-type="ref" reference="model-H-s-0"}) contains the bare Haldane gap $m^{(0)}$ due to the Zeeman interaction, it is also convenient to define the dynamical part of the total Haldane gap $\Delta m_{n, H}\equiv m_{n,H}-m^{(0)}$. In order to provide an insight into relation ([\[gap-Coulomb-results-Dirac-Haldanem-tmu\]](#gap-Coulomb-results-Dirac-Haldanem-tmu){reference-type="ref" reference="gap-Coulomb-results-Dirac-Haldanem-tmu"}) between $m_{\pm}$ and Dirac and Haldane gaps, let us recall the reducible 4$\times$`<!-- -->`{=html}4 representation for $\mbox{QED}_{2+1}$ considered in Ref. \[\begin{aligned} \tilde{\gamma}^0=\left( \begin{array}{cc} \gamma^0 & 0 \\ 0 &-\gamma^0 \\ \end{array} \right), \quad \tilde{\gamma}^1=\left( \begin{array}{cc} \gamma^1 & 0 \\ 0 &-\gamma^1 \\ \end{array} \right), \quad \tilde{\gamma}^2=\left( \begin{array}{cc} \gamma^2 & 0 \\ 0 &-\gamma^2 \\ \end{array} \right). \label{gap-Coulomb-results-reducible} \end{aligned}\] In this representation, in addition to the $\gamma$-matrices in Eq. ([\[gap-Coulomb-results-reducible\]](#gap-Coulomb-results-reducible){reference-type="ref" reference="gap-Coulomb-results-reducible"}), there exist two other matrices, \[\begin{aligned} \tilde{\gamma}^3=i\left( \begin{array}{cc} 0 & 1 \\ 1 & 0 \\ \end{array} \right), \quad \tilde{\gamma}^5=i\left( \begin{array}{cc} 0 & 1 \\ -1 & 0 \\ \end{array} \right), \label{gap-Coulomb-results-reducible-35} \end{aligned}\] which anticommute with $\tilde{\gamma}^0$, $\tilde{\gamma}^1$, and $\tilde{\gamma}^2$. In terms of these $4\times 4$ matrices, the existence of a $U(2)$ symmetry in the model of a TI slab, defined by Eqs. ([\[model-H-0-matrices\]](#model-H-0-matrices){reference-type="ref" reference="model-H-0-matrices"}) and ([\[model-H-int\]](#model-H-int){reference-type="ref" reference="model-H-int"}), is transparent. The corresponding group generators are given by \[1, \quad i{\cal R}_{\mu}\tilde{\gamma}^3, \quad {\cal R}_{\mu}\tilde{\gamma}^5, \quad \mbox{and} \quad \tilde{\gamma}^3\tilde{\gamma}^5, \label{gap-Coulomb-results-generators}\] where ${\cal R}_{\mu}$ is the operator which interchanges $\mu_{+}^{(0)}\leftrightarrow\mu_{-}^{(0)}$ in the low-energy free Hamiltonian ([\[model-H-0-matrices\]](#model-H-0-matrices){reference-type="ref" reference="model-H-0-matrices"}). (Note that, in the absence of an external electric field, there is no need in the operator ${\cal R}_{\mu}$.) As is easy to check, the Dirac gap $m_D\bar{\Psi}\Psi$ breaks the $U(2)$ symmetry down to $U_{+}(1)\times U_{-}(1)$, where $\bar{\Psi}=\Psi^{\dagger}\tilde{\gamma}^0$ and the subscript $\lambda=\pm$ labels the two irreducible representations or the surfaces of the TI slab. The Haldane gap $m_H\bar{\Psi}\tilde{\gamma}^3\tilde{\gamma}^5\Psi$ is invariant with respect to the $U(2)$ symmetry, but, unlike the Dirac gap, it breaks the parity $P$ and $T$ symmetries. Since external electric and magnetic fields break $P$ and $T$ symmetries, the generation of the Haldane gap has no effect on symmetry breaking. Therefore, only the dynamically generated Dirac gap will spontaneously break the symmetry of our model. As we will see below, such a gap is indeed generated due to the electrified magnetic catalysis. For numerical calculations, it is useful to estimate energy scales in the problem at hand \[\begin{aligned} \Delta_{\rm bulk}\approx 350~\mbox{meV},\quad\quad \frac{g_s\mu_B B}{2}\approx 0.5B[T]~\mbox{meV},\nonumber \\ \epsilon_B=\sqrt{2v_F^2\|eB\|}\approx 22.6\sqrt{B[T]}~\mbox{meV},\quad\quad l\approx25.7~\mbox{nm}/\sqrt{B[T]}. \label{gap-Coulomb-results-energy-scales} \end{aligned}\] By solving numerically the gap equations ([\[gapEq-Coulomb-gap-Delta\]](#gapEq-Coulomb-gap-Delta){reference-type="ref" reference="gapEq-Coulomb-gap-Delta"}), ([\[gapEq-Coulomb-gap-m\]](#gapEq-Coulomb-gap-m){reference-type="ref" reference="gapEq-Coulomb-gap-m"}) and ([\[gapEq-Coulomb-gap-mu\]](#gapEq-Coulomb-gap-mu){reference-type="ref" reference="gapEq-Coulomb-gap-mu"}) together with constraint ([\[gapEq-Coulomb-DOS\]](#gapEq-Coulomb-DOS){reference-type="ref" reference="gapEq-Coulomb-DOS"}), we straightforwardly obtain the electrochemical potentials $\mu_{n,\pm}$ and the gaps $m_{n,\pm}$ as functions of the magnetic field. The results for the lowest and first Landau level parameters are shown in Fig. [\[fig:gapEq-Coulomb-results-surface-tmu0-HLL-B-kappa\]](#fig:gapEq-Coulomb-results-surface-tmu0-HLL-B-kappa){reference-type="ref" reference="fig:gapEq-Coulomb-results-surface-tmu0-HLL-B-kappa"} for fixed values of the electric field and temperature, $\mathcal{E}=1~\mbox{mV/\AA}$ and $T=5\times10^{-3}\Delta_{\rm bulk} \approx20~\mbox{K}$, respectively. In the calculation, we truncated the system of equations by including only $n_{\rm max}=26$ Landau levels. As we see from the left panel in Fig. [\[fig:gapEq-Coulomb-results-surface-tmu0-HLL-B-kappa\]](#fig:gapEq-Coulomb-results-surface-tmu0-HLL-B-kappa){reference-type="ref" reference="fig:gapEq-Coulomb-results-surface-tmu0-HLL-B-kappa"}, the absolute values of the electrochemical potentials $\Delta_{\mathrm{eff}, \pm}$ and $\mu_{1,\pm}$ experience a large jump around $\|B\|\approx5~\mbox{T}$. The jump corresponds to the point at which the filling of the first Landau level starts. We checked that the position of the jump shifts to larger values of the magnetic field with increasing the external electric field. Of course, this behavior is expected, since larger electric fields require higher charge densities on the TI surfaces. In addition to the large jump around $\|B\|\approx5~\mbox{T}$, we also observe additional features in the dependence of $\Delta_{\mathrm{eff}, \pm}$ and $\mu_{1,\pm}$ at smaller values of the magnetic field. They generically correspond to the onset of filling of higher Landau levels. Here it is appropriate to mention that, in all regimes studied, the electrochemical potentials $\mu^{(0)}_{\pm}$ are very similar quantitatively to $\mu_{1,\pm}$ and, therefore, we do not show them in our figures. Let us now turn to the discussion of the dynamically generated gaps. The results in the middle panel of Fig. [\[fig:gapEq-Coulomb-results-surface-tmu0-HLL-B-kappa\]](#fig:gapEq-Coulomb-results-surface-tmu0-HLL-B-kappa){reference-type="ref" reference="fig:gapEq-Coulomb-results-surface-tmu0-HLL-B-kappa"} clearly demonstrate that the surface gaps $m_{1,\pm}$ monotonically increase with the magnetic field. More interestingly, however, we find that the values of the gaps on the two surfaces, $m_{1,+}$ and $m_{1,-}$, remain comparable, although not identical to each other for sufficiently weak electric fields. The importance of this observation becomes obvious in the context of the $U(2)$ symmetry discussed earlier. Indeed, if the values of $m_{1,+}$ and $m_{1,-}$ were exactly the same, they would describe a pure Haldane solution. As is clear from the definition in Eq. ([\[gap-Coulomb-results-Dirac-Haldanem-tmu\]](#gap-Coulomb-results-Dirac-Haldanem-tmu){reference-type="ref" reference="gap-Coulomb-results-Dirac-Haldanem-tmu"}), a small difference between $m_{1,+}$ and $m_{1,-}$ implies the existence of a dynamically generated Dirac gap. Such a gap is induced by the applied electric field. This conclusion is further supported by the dependence of the dynamical gaps on the electric field, shown in Fig. [\[fig:gapEq-Coulomb-results-surface-tmu0-HLL-E-kappa\]](#fig:gapEq-Coulomb-results-surface-tmu0-HLL-E-kappa){reference-type="ref" reference="fig:gapEq-Coulomb-results-surface-tmu0-HLL-E-kappa"} and discussed below Eq. ([\[gapEq-results-Coulomb-DOS\]](#gapEq-results-Coulomb-DOS){reference-type="ref" reference="gapEq-results-Coulomb-DOS"}). The results in the right panel of Fig. [\[fig:gapEq-Coulomb-results-surface-tmu0-HLL-B-kappa\]](#fig:gapEq-Coulomb-results-surface-tmu0-HLL-B-kappa){reference-type="ref" reference="fig:gapEq-Coulomb-results-surface-tmu0-HLL-B-kappa"} demonstrate that the absolute value of the Dirac gap $m_{1, D}$ increases with the magnetic field at sufficiently small fields, $\|B\|\lesssim 5~\mbox{T}$, when the LLL is fully filled. At larger magnetic fields, $\|B\|\gtrsim5~\mbox{T}$, when the LLL is not fully filled, the Dirac gap remains nearly constant (or increases very slowly). In contrast, the dynamical part of the Haldane gap $\Delta m_{1, H}$ increases approximately as $B^2$. Note that, because of the linear dependence of $m^{(0)}$ on the magnetic field, the total Haldane gap $m_{1, H}$ grows almost linearly with $B$. The dependencies of the lowest and first Landau level parameters on the external electric field are presented in Fig. [\[fig:gapEq-Coulomb-results-surface-tmu0-HLL-E-kappa\]](#fig:gapEq-Coulomb-results-surface-tmu0-HLL-E-kappa){reference-type="ref" reference="fig:gapEq-Coulomb-results-surface-tmu0-HLL-E-kappa"} for fixed values of the magnetic field ($B=5~\mbox{T}$) and temperature ($T=5\times10^{-3}\Delta_{\rm bulk}\approx20~\mbox{K}$). As we see from the left panel in Fig. [\[fig:gapEq-Coulomb-results-surface-tmu0-HLL-E-kappa\]](#fig:gapEq-Coulomb-results-surface-tmu0-HLL-E-kappa){reference-type="ref" reference="fig:gapEq-Coulomb-results-surface-tmu0-HLL-E-kappa"}, the absolute values of the electrochemical potentials slowly increase with electric field at first, and then experience a substantial jump at $\|\mathcal{E}\|\approx1~\mbox{mV/\AA}$. The jump corresponds to the field at which the filling of the first Landau level begins. From the middle panel in Fig. [\[fig:gapEq-Coulomb-results-surface-tmu0-HLL-E-kappa\]](#fig:gapEq-Coulomb-results-surface-tmu0-HLL-E-kappa){reference-type="ref" reference="fig:gapEq-Coulomb-results-surface-tmu0-HLL-E-kappa"}, we see that the surface gaps $m_{1,\pm}$ have a linear dependence at weak fields (i.e., in the regime of a partially filled LLL) and stay approximately constant at higher electric fields. As might have been expected, the Haldane gap $m_{1,H}$, which is shown in the right panel of Fig. [\[fig:gapEq-Coulomb-results-surface-tmu0-HLL-E-kappa\]](#fig:gapEq-Coulomb-results-surface-tmu0-HLL-E-kappa){reference-type="ref" reference="fig:gapEq-Coulomb-results-surface-tmu0-HLL-E-kappa"}, depends very weakly on the applied electric field. This is in contrast to the behavior of the Dirac gap $m_{1, D}$ (see the right panel in Fig. [\[fig:gapEq-Coulomb-results-surface-tmu0-HLL-E-kappa\]](#fig:gapEq-Coulomb-results-surface-tmu0-HLL-E-kappa){reference-type="ref" reference="fig:gapEq-Coulomb-results-surface-tmu0-HLL-E-kappa"}), which is linear in $\mathcal{E}$ at small fields and stays approximately constant at large fields. It should be also emphasized that the Dirac gap vanishes at $\mathcal{E}=0$. As we argue below, this fact is important from the viewpoint of symmetry properties in the model. As already suggested earlier, the generation of the Dirac gap is directly connected with the applied external electric field. In order to demonstrate this in the simplest possible setting, it is instructive to consider an approximate form of the gap equation ([\[gapEq-Coulomb-gap-m\]](#gapEq-Coulomb-gap-m){reference-type="ref" reference="gapEq-Coulomb-gap-m"}) in the limit of a large magnetic field. By rewriting it in terms of the Haldane and Dirac gaps, we obtain \[\begin{aligned} m_{n, H}&\approx &m^{(0)}+\frac{1}{4l} \sum_{n^{\prime}=1}^{\infty} \left(\alpha v_F \mathcal{K}^{(0)}_{n^{\prime}-1,n-1} + \alpha v_F \mathcal{K}^{(0)}_{n^{\prime},n} +\frac{G_{\rm int}}{\pi l} \right) \frac{m_{n^{\prime},H} }{M_{n^{\prime}}}, \label{gapEq-results-Coulomb-gap-mH} \\ m_{n, D}&\approx&-\pi l s_B \epsilon_0\left(\alpha v_F \mathcal{K}^{(0)}_{0,n}+\frac{G_{\rm int}}{2\pi l}\right)\frac{\mathcal{E}}{e} + \frac{1}{4l} \sum_{n^{\prime}=1}^{\infty} \left(\alpha v_F \mathcal{K}^{(0)}_{n^{\prime}-1,n-1} + \alpha v_F \mathcal{K}^{(0)}_{n^{\prime},n} +\frac{G_{\rm int}}{\pi l} \right) \frac{m_{n^{\prime},D} }{M_{n^{\prime}}}, \label{gapEq-results-Coulomb-gap-mD} \end{aligned}\] where we took into account that $n_F(M_{n}\pm \mu_{n, \lambda }) \ll 1$ for $n\geq 1$ and assumed that $M_{n^{\prime}}$ is almost independent of the small Dirac gap. Note that in order to rewrite the LLL contributions in the gap equations in terms of the electric field $\mathcal{E}$, we used the following approximate expression for the surface charge densities: \[\rho_{+}=-\rho_{-}=\frac{\rho_{+}-\rho_{-}}{2}\approx\frac{e}{4\pi l^2} \left[n_F\left(\Delta_{\mathrm{eff}, +}\right)-n_F\left(\Delta_{\mathrm{eff},-}\right) \right] =\epsilon_0\mathcal{E}. \label{gapEq-results-Coulomb-DOS}\] By comparing the gap equations ([\[gapEq-results-Coulomb-gap-mH\]](#gapEq-results-Coulomb-gap-mH){reference-type="ref" reference="gapEq-results-Coulomb-gap-mH"}) and ([\[gapEq-results-Coulomb-gap-mD\]](#gapEq-results-Coulomb-gap-mD){reference-type="ref" reference="gapEq-results-Coulomb-gap-mD"}), we see that the external electric field plays the role of a "seed\" for the Dirac gap $m_{n, D}$, just as the bare gap parameter $m^{(0)}$ for the Haldane gap $m_{n,H}$. This explains why the external electric field is the key factor in generating the Dirac gap and breaking the $U(2)$ symmetry in the slab of 3D TIs. It may be instructive to study the dependence of the electrochemical potentials $\mu_{n,\pm}$ and gaps $m_{n,\pm}$ on the Landau level index $n$. The corresponding results for two different values of the electric field are presented in Fig. [\[fig:gapEq-Coulomb-results-surface-tmu0-HLL-E-n-kappa\]](#fig:gapEq-Coulomb-results-surface-tmu0-HLL-E-n-kappa){reference-type="ref" reference="fig:gapEq-Coulomb-results-surface-tmu0-HLL-E-n-kappa"} for $B=5~\mbox{T}$ and $T=5\times10^{-3}\Delta_{\rm bulk}\approx20~\mbox{K}$. As we see, all dynamical parameters depend very weakly on the Landau level index $n$. In view of the large surface dielectric constant and, consequently, weak Coulomb interaction, this result is not surprising. Moreover, it strongly suggests that the long-range interaction indeed plays a minor role compared to the local interaction. By using the above results, we can also obtain the quasiparticle energy levels as functions of the magnetic and electric fields \[\begin{aligned} \omega_{0,\lambda}=-\Delta_{\rm eff, \lambda}, \quad \omega_{n>0, \lambda}=-\mu_{n, \lambda}\pm M_{n}, \label{gap-Coulomb-results-spectrum} \end{aligned}\] where $M_n$ were given below Eq. ([\[gapEq-Coulomb-G-no-phase\]](#gapEq-Coulomb-G-no-phase){reference-type="ref" reference="gapEq-Coulomb-G-no-phase"}). The corresponding numerical results are summarized in Fig. [\[fig:gapEq-Coulomb-results-surface-spectrum\]](#fig:gapEq-Coulomb-results-surface-spectrum){reference-type="ref" reference="fig:gapEq-Coulomb-results-surface-spectrum"}. As we see from the left panel in Fig. [\[fig:gapEq-Coulomb-results-surface-spectrum\]](#fig:gapEq-Coulomb-results-surface-spectrum){reference-type="ref" reference="fig:gapEq-Coulomb-results-surface-spectrum"}, there is a rather large splitting between the energy levels on the top and bottom surfaces at small values of the magnetic field. This corresponds to the regime with higher Landau levels being occupied. With increasing the magnetic field, the magnitude of splitting quickly diminishes and becomes rather small when the LLL regime is reached. In contrast, the increase of the electric field tends to amplify the splitting between the Landau levels. The existence of such a splitting may lead to an observation of new plateaus in the Hall conductivity. The large jumps in the energy spectrum at $\|B\|\approx5~\mbox{T}$ and $\|\mathcal{E}\|\approx1~\mbox{mV/\AA}$ correspond to the onset of and the exit from the LLL regime, respectively. As is clear, these features are directly connected with the corresponding jumps in the electrochemical potentials, seen in the left panels of Figs. [\[fig:gapEq-Coulomb-results-surface-tmu0-HLL-B-kappa\]](#fig:gapEq-Coulomb-results-surface-tmu0-HLL-B-kappa){reference-type="ref" reference="fig:gapEq-Coulomb-results-surface-tmu0-HLL-B-kappa"} and [\[fig:gapEq-Coulomb-results-surface-tmu0-HLL-E-kappa\]](#fig:gapEq-Coulomb-results-surface-tmu0-HLL-E-kappa){reference-type="ref" reference="fig:gapEq-Coulomb-results-surface-tmu0-HLL-E-kappa"}. Before concluding this section, let us briefly discuss the role of finite temperature in our solution. As expected, the main results remain qualitatively the same for a whole range of sufficiently small values of the temperature. With increasing (decreasing) the temperature, however, the jumps that correspond to the onset of and the exit from the LLL regime become smoother (sharper) in the dependence of the electrochemical potentials on the fields, shown in the left panels of Figs. [\[fig:gapEq-Coulomb-results-surface-tmu0-HLL-B-kappa\]](#fig:gapEq-Coulomb-results-surface-tmu0-HLL-B-kappa){reference-type="ref" reference="fig:gapEq-Coulomb-results-surface-tmu0-HLL-B-kappa"} and [\[fig:gapEq-Coulomb-results-surface-tmu0-HLL-E-kappa\]](#fig:gapEq-Coulomb-results-surface-tmu0-HLL-E-kappa){reference-type="ref" reference="fig:gapEq-Coulomb-results-surface-tmu0-HLL-E-kappa"}. It is also worth pointing that a weak dependence of electrochemical potentials on the fields in the regions between the jumps is caused by thermal broadening of Landau levels. It vanishes in the limit $T\to0$. # Inhomogeneous phase with two stripes: qualitative approach {#sec:stripe} In the previous section, we advocated the homogeneous phase with dynamically generated gaps as the ground state of 3D TIs in a sufficiently strong external electric field. On the other hand, the inhomogeneous CDW phase considered in Ref. is likely to be more favorable in weak electric fields. In order to provide a qualitative analytic description of the inhomogeneous CDW phase with a "stripe" pattern, in this section we consider a simple configuration of two stripes with an infinitely thin transition region, or a domain wall at $x=0$. This is modeled by an inhomogeneous gap $m(x)=\|m\|\,\mbox{sign}{(x)}$ for the top surface, i.e., $\lambda=+1$ and $s_{B}=+1$. (Note that such a gap function with asymptotes of opposite sign at $x\to\pm\infty$, but without a magnetic field, is qualitatively similar to the famous Jackiw-Rebbi solution in 1D.) The solution to the Dirac equation with the gap function in the form $m(x)=\|m\|\,\mbox{sign}{(x)}$ is discussed in Appendix [\[sec:domain-wall\]](#sec:domain-wall){reference-type="ref" reference="sec:domain-wall"}. The corresponding numerical results for the quasiparticle energy spectrum as a function of $k_y$, as well as the chiral condensate and charge density as functions of the spatial coordinate $x$, are shown in Fig. [\[fig:stripe-zero-modes-chiral-condensate\]](#fig:stripe-zero-modes-chiral-condensate){reference-type="ref" reference="fig:stripe-zero-modes-chiral-condensate"}. In order to plot the results, we fixed the model parameters as follows: $m=5~\mbox{meV}$, $\mu=0$, and $B=5~\mbox{T}$. In the calculation, we also limited the sum over Landau levels ($n_{\rm max}=26$) and cut off the integration over $k_y$ ($-6/l \leq k_y \leq 6/l$). As we see from Fig. [\[fig:stripe-zero-modes-chiral-condensate\]](#fig:stripe-zero-modes-chiral-condensate){reference-type="ref" reference="fig:stripe-zero-modes-chiral-condensate"}, the chiral condensate and the charge density have a kink and antikink structure, respectively. Therefore, the existence of zero energy states on the domain wall agrees with the inhomogeneous form of the gap function. The chiral condensate and the charge density can be fitted well by the following functions: \[\begin{aligned} \label{stripe-zero-modes-chircond-fit} \mbox{tr}[G(u,u)]=\sum_{n} \frac{\,\mbox{sign}{(\omega_n)}}{2} \bar{\psi}_{\omega_n}(x)\psi_{\omega_n}(x) \approx \frac{1}{4\pi l^2} 2.5 \tanh{\left(2.6\frac{x}{l}\right)}, \\ \label{stripe-zero-modes-DOS-fit} \mbox{tr}[\gamma^0 G(u,u)]=\sum_{n} \frac{\,\mbox{sign}{(\omega_n)}}{2} \psi^{\dag}_{\omega_n}(x)\psi_{\omega_n}(x) \approx-\frac{1}{4\pi l^2} 1.0 \tanh{\left(1.3\frac{x}{l}\right)}, \end{aligned}\] where the sum runs over the complete set of eigenstates, given by the solutions to the spectral equation ([\[stripe-zero-modes-spectral-eq\]](#stripe-zero-modes-spectral-eq){reference-type="ref" reference="stripe-zero-modes-spectral-eq"}). These fits are shown in Fig. [\[fig:stripe-zero-modes-chiral-condensate\]](#fig:stripe-zero-modes-chiral-condensate){reference-type="ref" reference="fig:stripe-zero-modes-chiral-condensate"} alongside with the numerical solutions. Let us now consider the case of a nonzero external electric field, applied perpendicularly to the slab. By considering a sufficiently thick slab, we will assume that the average electric field inside the slab vanishes. In the homogeneous case, the field is screened by the uniform surface charge densities. This generically requires specific nonzero electrochemical potentials $\mu_{\pm}$ on the top and bottom TIs surfaces. In the inhomogeneous striped phase, however, the simplest way to achieve a nonzero average surface charge densities is to vary the width of stripes by $\Delta l_x$. The value of $\Delta l_x$ can be estimated from the following expression: \[\epsilon_0\mathcal{E} = \rho(l_x+\Delta l_x)-\rho(l_x)+ \rho(-l_x+\Delta l_x)-\rho(-l_x) \approx \frac{e}{4\pi l^2} \left[\tanh{\left(1.3\frac{l_x+\Delta l_x}{l}\right)}-\tanh{\left(1.3\frac{l_x-\Delta l_x}{l}\right)}\right]. \label{stripe-energy-diff-stripes-DOS}\] where we used the standard definition for the surface charge density $\rho=e\,\mbox{tr}[\gamma^0 G(u,u)]$ together with Eq. ([\[stripe-zero-modes-DOS-fit\]](#stripe-zero-modes-DOS-fit){reference-type="ref" reference="stripe-zero-modes-DOS-fit"}). Our numerical result for the ratio of the correction $\Delta l_x$ to the stripe half-width $l_x$ is plotted in Fig. [\[fig:stripe-energy-diff-nonzero-field-Lky-kappa\]](#fig:stripe-energy-diff-nonzero-field-Lky-kappa){reference-type="ref" reference="fig:stripe-energy-diff-nonzero-field-Lky-kappa"}. We see that the correction to the stripe width $\Delta l_x$ becomes significant at sufficiently strong electric fields. In fact, it is comparable to $l_x$ already for $\mathcal{E} \gtrsim 1~\mbox{mV/\AA}$. This suggests that the stripe phase is unstable when the electric field exceeds a certain critical value $\mathcal{E}_{\rm cr}$. Quantitatively, the critical value roughly corresponds to the beginning of the first Landau level filling, \[\mathcal{E}_{\rm cr}\approx\frac{e^2B}{4\pi \epsilon_{0} \hbar c}, \label{stripe-energy-diff-critical-E}\] where we used Eq. ([\[gapEq-Coulomb-DOS\]](#gapEq-Coulomb-DOS){reference-type="ref" reference="gapEq-Coulomb-DOS"}) and restored Plank's constant $\hbar$ and the speed of light $c$. This relation implies that $\mathcal{E}_{\rm cr}\approx0.22B[\mbox{T}]~\mbox{mV}/\mbox{\AA}$, which is in a good agreement with our previous numerical estimate $\mathcal{E} \gtrsim 1~\mbox{mV/\AA}$ at $B=5~\mbox{T}$. Another way to estimate the critical value of the electric field is to compare the free energy density in the homogeneous phase, which is given by Eq. ([\[app-free-energy-T\]](#app-free-energy-T){reference-type="ref" reference="app-free-energy-T"}), with the energy density of the stripe phase estimated in Ref. . In other words, the value of $\mathcal{E}_{\rm cr}$ is given by the solution to the following equation: \[0=\sum_{\lambda}\Omega_{\lambda}-\left[\frac{\sqrt{\alpha \gamma}}{l^2}-\frac{m}{l^2}\right], \label{stripe-energy-diff-critical-field-Eq}\] where the term in square brackets corresponds to the energy cost of creating the stripe phase. The latter is characterized by the domain wall tension $\gamma\sim1/l^2$ and the magnetic mass $m\sim 1/l$. The value of $\alpha$ is given below Eq. ([\[gapEq-Coulomb-gap-1\]](#gapEq-Coulomb-gap-1){reference-type="ref" reference="gapEq-Coulomb-gap-1"}). The solution to Eq. ([\[stripe-energy-diff-critical-field-Eq\]](#stripe-energy-diff-critical-field-Eq){reference-type="ref" reference="stripe-energy-diff-critical-field-Eq"}) can be easily obtained numerically and appears to agree quite well with the estimate in Eq. ([\[stripe-energy-diff-critical-E\]](#stripe-energy-diff-critical-E){reference-type="ref" reference="stripe-energy-diff-critical-E"}). One can also obtain an approximate analytical solution to Eq. ([\[stripe-energy-diff-critical-field-Eq\]](#stripe-energy-diff-critical-field-Eq){reference-type="ref" reference="stripe-energy-diff-critical-field-Eq"}) by using the LLL approximation for the free energy density ([\[app-free-energy-T\]](#app-free-energy-T){reference-type="ref" reference="app-free-energy-T"}), i.e., \[\Omega_{\lambda} \approx \frac{\lambda \epsilon_0 \mathcal{E} \Delta_{\rm eff}}{e}, \label{stripe-energy-diff-free-energy-LLL-1}\] where we used the LLL approximation for the charge density ([\[gapEq-Coulomb-DOS\]](#gapEq-Coulomb-DOS){reference-type="ref" reference="gapEq-Coulomb-DOS"}). Then, by substituting this into Eq. ([\[stripe-energy-diff-critical-field-Eq\]](#stripe-energy-diff-critical-field-Eq){reference-type="ref" reference="stripe-energy-diff-critical-field-Eq"}) and estimating the effective electrochemical potential as $\Delta_{\rm eff}\sim \lambda/l$, we find the following critical value of the electric field: \[\mathcal{E}_{\rm cr} \approx \frac{e^2B (1-\sqrt{\alpha})}{2\epsilon_0 \hbar c}. \label{stripe-energy-diff-critical-E-2}\] This result is qualitatively the same as the estimate in Eq. ([\[stripe-energy-diff-critical-E\]](#stripe-energy-diff-critical-E){reference-type="ref" reference="stripe-energy-diff-critical-E"}), although quantitatively appears to be somewhat larger, $\mathcal{E}_{\rm cr} \approx 1.03B[\mbox{T}]~\mbox{mV}/\mbox{\AA}$. We conclude, therefore, that the critical electric field scales linearly with the magnetic field, $\mathcal{E}_{\rm cr} \sim e^2B/(\epsilon_0 \hbar c)$, but the coefficient of proportionality is determined only up to an overall factor of order $1$. # Discussion {#sec:Discussion} In this section, we discuss the range of validity and limitations of our study, and compare our main results with those existing in the literature. Let us start by pointing the limitation of our model used for the description of the TI surface states. While the model captures the Dirac nature of the low-energy quasiparticles, it does not describe the hexagonal warping of the Fermi surface that occurs away from the Dirac point. The corresponding effect was taken into account in the study of gap generation in Ref. and could play an essential role in some TIs. For example, this may be the case in Bi$_2$Te$_3$ \[e.g., see Fig. 1(c) in Ref. \], in which the band gap is about three times smaller than in Bi$_2$Se$_3$ and the trigonal potential $\sim k^3$ is rather strong. In the case of Bi$_2$Se$_3$, however, the hexagonal warping could be safely neglected, except for the case of rather high values of the chemical potential \[e.g., see Fig. 8 in Ref. and Fig. 3(b) in Ref. \]. The model used in this study also ignores a Schrödinger-type term $\sim k^2$, which describes an asymmetry between the electron and hole bands (see also Fig. 1 in Ref. ). When the quadratic term is sufficiently small, it is not expected to substantially affect the dynamics of the gap generation. The study here did not include the effects of the intersurface tunneling on the dynamical generation of gaps. According to Ref. , tunneling between the opposite surfaces may be quite important only for sufficiently thin ($l_z\lesssim8~\mbox{nm}$) TI slabs. Therefore, neglecting the intersurface tunneling is expected to be a good approximation in the case of thick samples. It would be interesting, however, to rigorously study the corresponding effects in thin TI films in external electric and magnetic fields. One of the uncertainties of the model Hamiltonian used in this study is the strength of the local interaction $G_{\rm int}$. Although the order of magnitude of this coupling constant could be estimated by using general arguments, its precise value is unknown. Despite this, we argue that the simplified model ([\[model-H-0-matrices\]](#model-H-0-matrices){reference-type="ref" reference="model-H-0-matrices"}) that includes both short-and long-range interactions ([\[model-H-int\]](#model-H-int){reference-type="ref" reference="model-H-int"}) is sufficient for the qualitative analysis of the electrified magnetic catalysis in 3D TIs. Moreover, we might even suggest that, irrespective of the specific value of the coupling constants, the qualitative features established here should be rather universal. It is interesting to compare our results with those obtained in Ref. , where the phase diagram was studied in 3D TIs in a magnetic field, but without an external electric field. The authors of Ref. argued that, depending on the strength of local interaction, the CFL or CDW ("stripe" or "bubble") phases can be realized. Our results here suggest that neither of those two phases describe the ground state of the TI slab in a sufficiently strong external electric field. The CFL phase with the half-filled LLL on each TI surface cannot be easily deformed to screen out the external electric field from penetrating into the TI bulk. This would imply a large energy cost and disfavor the CFL phase. The CDW phase could perhaps survive when a relatively weak electric field is applied. In this case, the average charge densities on the TI surfaces, which are needed to screen the electric field out from the bulk, could be simply obtained by the formation of positive and negative stripes (or bubbles) of unequal size. (Note that it is energetically favorable to have either completely filled or empty LLL inside the stripes.) As we showed in Sec. [5](#sec:stripe){reference-type="ref" reference="sec:stripe"}, a simple estimate suggests that the charge imbalance obtained by the variation of the stripe widths can compensate only relatively weak electric fields. Therefore, a sufficiently strong electric field $\mathcal{E}>\mathcal{E}_{\rm cr}$ also destroys the CDW phase. Our parametric estimate for the critical electric field strength is $\mathcal{E}_{\rm cr} \sim e^2B/(\epsilon_0 \hbar c)$. In view of the above arguments, we claim that the ground state of the TI slab in a nonzero magnetic field and a sufficiently strong electric field is a homogeneous phase with equal in magnitude, but opposite in sign surface charge densities. It is also characterized by the presence of both Dirac and Haldane gaps. While our qualitative conclusion seems rather rigorous, this study is insufficient to establish the precise structure of the phase diagram in the plane of the applied electric and magnetic fields. It would be very interesting to clarify the details of the corresponding phase diagram either experimentally or numerically. It may be instructive to note that the thermal broadening of Landau levels plays a relatively important technical role in our analysis and in the description of the electrified homogeneous phases. Indeed, by using a nonzero temperature, we were able to unambiguously describe the surface ground states with adjustable partial fillings of Landau levels, needed to screen the external electric field. Certainly, the corresponding ground states allow a well defined zero temperature limit, but their description may become more subtle. By noting that surface impurities also broaden Landau levels, we suggest that their presence could lead to a realization of the electrified magnetic catalysis similar to that in Sec. [4](#sec:gap-equation-Coulomb-results-kappa){reference-type="ref" reference="sec:gap-equation-Coulomb-results-kappa"}. As is clear from our study, the low-energy model for the surface states of 3D TIs is essentially a (2+1)-dimensional QED, supplemented by certain constraints. The generation of different types of gaps, such as those describing spontaneous parity breaking and chiral symmetry breaking were studied in QED$_{2+1}$ without background electromagnetic fields in Refs. a long time ago. Moreover, it was shown that the Dirac mass can be spontaneously generated, while the Haldane mass is energetically disfavored. Clearly, this is not the case in the problem at hand, where both types of gaps are generated on the TI surfaces. This is due to the fact that the TR and inversion symmetries are explicitly broken by the external magnetic and electric fields. Furthermore, we find that the Haldane gap dominates at small values of the electric field. This situation is reminiscent of the dynamically enhanced Zeeman splitting in graphene. # Conclusion {#sec:Conclusion} In this study, we considered the dynamical generation of gaps in a slab of a 3D TI, such as Bi$_2$Se$_3$, placed in the magnetic and electric fields perpendicular to its surfaces. (Although we used the model parameters for Bi$_2$Se$_3$, the main conclusions should be valid for all similar TIs.) Note that the conducting states on the TI slab surfaces and the overall geometry of the system are rather similar to bilayer graphene. On the other hand, the degeneracy connected with the valley and spin degrees of freedom, which is responsible for a variety of quantum Hall states in bilayer graphene, is absent in a TI slab. Still, there are notable similarities in the dynamics of gap generation in these two physical systems. For example, the valley quantum Hall (or layer polarized) state, which is realized in a sufficiently strong external electric field in bilayer graphene, resembles the homogeneous state considered in this paper. By solving the gap equations for the surface quasiparticle propagators in a simple model with short-and long-range interactions, we found that both* the Dirac and Haldane gaps are dynamically generated in the TI slab in external electric and magnetic fields. The underlying mechanism is a different version of the magnetic catalysis. Because of a large surface dielectric constant, the Coulomb interaction appears to play a minor role in the dynamics. Unlike the Dirac gap, the Haldane gap respects the $U(2)$ symmetry with the generators given in Eq. ([\[gap-Coulomb-results-generators\]](#gap-Coulomb-results-generators){reference-type="ref" reference="gap-Coulomb-results-generators"}), but breaks the parity and TR symmetries. Since both discrete symmetries are explicitly broken by external electric and magnetic fields, the generation of the Haldane gap does not break any symmetries. The Dirac gap, on the other hand, is generated only in the presence of an electric field. The result of such an electrified magnetic catalysis is a spontaneous breaking of the $U(2)$ symmetry. By comparing our results with the findings in Ref. , we argued that the homogeneous phase with dynamically generated Dirac and Haldane gaps is the true ground state in the TI slab in nonzero magnetic and sufficiently strong electric fields. The precise structure of the phase diagram in the plane of applied electric and magnetic fields remains to be clarified, however. The authors are grateful to V.P. Gusynin for useful discussions. The work of E.V.G. was supported partially by the Ukrainian State Foundation for Fundamental Research. The work of V.A.M. and P.O.S. was supported by the Natural Sciences and Engineering Research Council of Canada. The work of I.A.S. was supported in part by the U.S. National Science Foundation under Grant No. PHY-1404232.	{'timestamp': '2016-09-29T02:02:31', 'yymm': '1607', 'arxiv_id': '1607.04649', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04649'}
# Gibbs' Canonical Ensemble From Gibbs' 1902 text Elementary Principles in Statistical Mechanics, page 183: > "If a system of a great number of degrees of freedom is microcanonically distributed in phase, any very small part of it may be regarded as canonically distributed." Thus J. Willard Gibbs pointed out that the energy states of a "small" system weakly coupled to a larger "heat reservoir" with a temperature $T$ have a "canonical" distribution: \[f(q,p) \propto e^{-{\cal H}(q,p)/kT} \.\] with the Hamiltonian ${\cal H}(q,p)$ that of the small system. Here $(q,p)$ represents the set of coordinates and momenta of that system. " Canonical " means simplest or prototypical. The heat reservoir coupled to the small system and responsible for the canonical distribution of energies is best pictured as an ideal-gas thermometer characterized by an unchanging kinetic temperature $T$. The reservoir gas consists of many small-mass classical particles engaged in a chaotic and ergodic state of thermal and mechanical equilibrium with negligible fluctuations in its temperature and pressure. Equilibrium within this thermometric reservoir is maintained by collisions as is described by Boltzmann's equation. His "H Theorem" establishes the Maxwell-Boltzmann velocity distribution found in the gas. See Steve Brush's 1964 translation of Boltzmann's 1896 text Vorlesungen über Gastheorie. Prior to fast computers texts in statistical mechanics were relatively formal with very few figures and only a handful of numerical results. In its more than 700 pages Tolman's 1938 tome The Principles of Statistical Mechanics includes only two Figures. \[ The more memorable one, a disk colliding with a triangle, appears on the cover of the Dover reprint volume. \] Today the results-oriented graphics situation is entirely different as a glance inside any recent issue of Science confirms. # Nosé-Hoover Canonical Dynamics--Lack of Ergodicity In 1984, with the advent of fast computers and packaged computer graphics software already past, Shuichi Nosé set himself the task of generalizing molecular dynamics to mimic Gibbs' canonical distribution. In the end his approach was revolutionary. It led to a new form of heat reservoir described by a single degree of freedom with a logarithmic potential, rather than the infinitely-many oscillators or gas particles discussed in textbooks. Although the theory underlying Nosé's approach was cumbersome Hoover soon pointed out a useful simplification: Liouville's flow equation in the phase space provides a direct proof that the "Nosé-Hoover" motion equations are consistent with Gibbs' canonical distribution. Here are the motion equations for the simplest interesting system, a single one-dimensional harmonic oscillator: \[\dot q = (p/m) \; \ \dot p =-\kappa q-\zeta p \; \ \dot \zeta = [ \ (p^2/mkT)-1 \ ]/\tau^2 \.\] The "friction coefficient" $\zeta$ stabilizes the kinetic energy $(p^2/2m)$ through integral feedback, extracting or inserting energy as needed to insure a time-averaged value of precisely $(kT/2)$. The parameter $\tau$ is a relaxation time governing the rate of the thermostat's response to thermal fluctuations. In what follows we will set all the parameters and constants $(m,\kappa,k,T,\tau)$ equal to unity, purely for convenience. Then the Nosé-Hoover equations have the form: \[\dot q = p \; \ \dot p =-q-\zeta p \; \ \dot \zeta = p^2-1 \ [ \ {\rm NH} \ ] \.\] Liouville's phase-space flow equation, likewise written here for a single degree of freedom, is just the usual continuity equation for the three-dimensional flow of a probability density in the ($q,p,\zeta$) phase space: \[\dot f = (\partial f/\partial t) + \dot q(\partial f/\partial q) + \dot p(\partial f/\partial p) + \dot \zeta(\partial f/\partial \zeta) =-f(\partial \dot q/\partial q)-f(\partial \dot p/\partial p)-f(\partial \dot \zeta/\partial \zeta) \.\] This approach leads directly to the simple \[ NH \] dynamics described above. It is easy to verify that Gibbs' canonical distribution needs only to be multiplied by a Gaussian distribution in $\zeta$ in order to satisfy Liouville's equation. \[e^{-q^2/2}e^{-p^2/2}e^{-\zeta^2/2} \propto f_{NH} \propto f_Ge^{-\zeta^2/2} \longrightarrow (\partial f_{NH}/\partial t) \equiv 0 \ .\] Hoover emphasized that the simplest thermostated system, a harmonic oscillator, does not fill out the entire Gibbs' distribution in $(q,p,\zeta)$ space. It is not "ergodic" and fails to reach all of the oscillator phase space. In fact, with all of the parameters ( mass, force constant, Boltzmann's constant, temperature, and relaxation time $\tau$ ) set equal to unity only six percent of the Gaussian distribution is involved in the chaotic sea. See Figure 1 for a cross section of the Nosé-Hoover sea in the $p=0$ plane. The complexity in the figure, where the "holes" correspond to two-dimensional tori in the three-dimensional $(q,p,\zeta)$ phase space, is due to the close relationship of the Nosé-Hoover thermostated equations to conventional chaotic Hamiltonian mechanics with its infinitely-many elliptic and hyperbolic points. # More General Thermostat Ideas New varieties of thermostats, some of them Hamiltonian and some not, appeared over the ensuing 30-year period following Nosé's work[@b6; @b7; @b8; @b9; @b10; @b11; @b12; @b13; @b14; @b15; @b16; @b17; @b18]. This list is by no means complete. Though important, simplicity is not the sole motivation for abandoning purely-Hamiltonian thermostats. Relatively recently we pointed out that Hamiltonian thermostats are incapable of generating or absorbing heat flow. The close connection between changing phase volume and entropy production guarantees that Hamiltonian mechanics is fundamentally inconsistent with irreversible flows. At equilibrium Brańka, Kowalik, and Wojciechowski followed Bulgac and Kusnezov in emphasizing that cubic frictional forces, $-\zeta^3p$, which also follow from a novel Hamiltonian, promote a much better coverage of phase space, as shown in Figure 2. The many small holes in the $p=0$ cross section show that this approach also lacks ergodicity. ## Joint Control of Two Velocity Moments Attempts to improve upon this situation led to a large literature with the most useful contributions applying thermostating ideas with two or more thermostat variables. An example, applied to the harmonic oscillator, was tested by Hoover and Holian and found to provide all of Gibbs' distribution: \[\dot q = p \; \ \dot p =-q-\zeta p-\dot \zeta = p^2-1 \; \ \dot\] The two thermostat variables $(\zeta, and the fourth moments of the velocity distribution have their Maxwell-Boltzmann values [ 1 and 3 ]. Notice that two-dimensional cross sections like those in the Figures are no longer useful diagnostics for ergodicity once the phase-space dimensionality exceeds three. \subsection{Joint Control of Coordinates and Velocities} In 2014 Patra and Bhattacharya suggested thermostating both the coordinates and the momenta:\[\dot q = p-\dot \zeta = p^2-1 \; \ \dot\]an approach already tried by Sergi and Ezra in 2001. A slight variation of the Sergi-Ezra-Patra-Bhattacharya thermostat takes into account Bulgac and Kusnezov's observation that cubic terms favor ergodicity:\[\dot q = p-\dot \zeta = p^2-1 \; \ \dot\]These last two-thermostat equations appear to be a good candidate for ergodicity, reproducing the second and fourth moments of$(q,p,, carried out the thorough investigation that would be required to establish their ergodicity as the single-thermostat models are not only simpler but also much more easily diagnosed because their sections are two-dimensional rather than three-dimensional. # Single-Thermostat Ergodicity Combining the ideas of "weak control" and the successful simultaneous thermostating of coordinates and momenta led to further trials attempting the weak control of two different kinetic-energy moments. One choice out of the hundreds investigated turned out to be successful for the harmonic oscillator: \[\dot q = p \; \ \dot p =-q-\zeta( 0.05p + 0.32p^3) \; \ \dot \zeta = 0.05(p^2-1) + 0.32(p^4-3p^2) \ [ \ {\rm ``0532 \ Model''} \ ] \.\] These three oscillator equations passed all of the following tests for ergodicity: The moments $\langle \ p^2 \ \rangle = 1 \; \ \langle \ p^4\ \rangle = 3 \; \ \langle \ p^6 \ \rangle = 15$ were confirmed. The independence of the largest Lyapunov exponent to the initial conditions indicated the absence of the toroidal solutions. The separation of two nearby trajectories had an average value of 6 :\ $\langle \ (q_1-q_2)^2 + (p_1-p_2)^2 + (\zeta_1-\zeta_2)^2 \ \rangle = 2 + 2 + 2 = 6$. The times spent at positive and negative values of $\{ \ q,p,\zeta \ \}$ were close to equal. The times spent in regions with each of the 3! orderings of the three dependent variables were equal for long times. These five criteria were useful tools for confirming erogidicity. Evidently weak control is the key to efficient ergodic thermostating of oscillator problems. # A Fly in the Ointment, the Quartic Potential The success in thermostating the harmonic oscillator led to like results for the simple pendulum but not for the quartic potential. See Figure 3. This somewhat surprising setback motivates the need for more work and is the subject of the Ian Snook Prize for 2016. This Prize will be awarded to the author(s) of the most interesting original work exploring the ergodicity of single-thermostated statistical-mechanical systems. The systems are not at all limited to the examples of the quartic oscillator and the Mexican Hat potential but are left to the imagination and creativity of those entering the competition. # Conclusions--Ian Snook Prize for 2016 It is our intention to reward the most interesting and convincing entry submitted for publication to Computational Methods in Science and Technology ( www.cmst.eu ) prior to 31 January 2017. The 2016 Ian Snook prize of \$500 dollars will be presented to the winner in early 2017. An Additional Prize of the same amount will likewise be presented by the Institute of Bioorganic Chemistry of the Polish Academy of Sciences ( Poznan Supercomputing and Networking Center ). We are grateful for your contributions. This work is dedicated to the memories of our colleagues, Ian Snook ( 1945-2013 ) and Shuichi Nosé ( 1951-2005 ), shown in Figure 4.	{'timestamp': '2016-07-18T02:08:57', 'yymm': '1607', 'arxiv_id': '1607.04595', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04595'}
null	null
null	null
# Introduction In his book "Proximal Flows" Glasner defines the notion of a strongly amenable group: A group is strongly amenable if each of its proximal actions on a compact space has a fixed point. A continuous action $G \curvearrowright X$ of a topological group on a compact Hausdorff space is proximal if for every $x, y \in X$ there exists a net $\{g_n\}$ of elements of $G$ such that $\lim_n g_n x = \lim_n g_n y$. Glasner shows that virtually nilpotent groups are strongly amenable and that non-amenable groups are not strongly amenable. He also gives examples of amenable---in fact, solvable---groups that are not strongly amenable. Glasner and Weiss construct proximal minimal actions of the group of permutations of the integers, and Glasner constructs proximal flows of Lie groups . To the best of our knowledge there are no other such examples known. Furthermore, there are no other known examples of minimal proximal actions that are not also strongly proximal. An action $G \curvearrowright X$ is strongly proximal if the orbit closure of every Borel probability measure on $G$ contains a point mass measure. This notion, as well as that of the related Furstenberg boundary [@furstenberg1963poisson; @furstenberg1973boundary; @furman2003minimal], have been the object of a much larger research effort, in particular because a group is amenable if and only if all of its strongly proximal actions on compact spaces have fixed points. Richard Thompson's group $F$ has been alternatively "proved" to be amenable and non-amenable (see, e.g., ), and the question of its amenability is currently unresolved. In this paper we pursue the less ambitious goal of showing that is it not strongly amenable, and do so by directly constructing a proximal action that has no fixed points. This action does admit an invariant measure, and thus does not provide any information about the amenability of $F$. It is a new example of a proximal action which is not strongly proximal. The authors would like to thank Eli Glasner and Benjamin Weiss for enlightening and encouraging conversations. # Proofs Let $F$ denote Thompson's group $F$. In the representation of $F$ as a group of piecewise linear transformations of $\mathbb{R}$ (see, e.g., ), it is generated by $a$ and $b$ which are given by \[\begin{aligned} a(x) &= x-1\\ b(x) &= \begin{cases} x& x \leq 0\\ x/2& 0 \leq x \leq 2\\ x-1& 2 \leq x. \end{cases} \end{aligned}\] The set of dyadic rationals $\Gamma =\mathbb{Z}[\frac{1}{2}]$ is the orbit of $0$. The Schreier graph of the action $G \curvearrowright \Gamma$ with respect to the generating set $\{a,b\}$ is shown in Figure [\[fig:schreier\]](#fig:schreier){reference-type="ref" reference="fig:schreier"} (see ). The solid lines denote the $a$ action and the dotted lines denote the $b$ action; self-loops (i.e., points stabilized by a generator) are omitted. This graph consists of a tree-like structure (the blue and white nodes) with infinite chains attached to each node (the red nodes). Equipped with the product topology, $\{-1,1\}^\Gamma$ is a compact space on which $F$ acts continuously by shifts: \[\begin{aligned} \label{shift-action} [f x](\gamma) = x(f^{-1}\gamma). \end{aligned}\] Given $x_1,x_2 \in \{-1,1\}^{\Gamma}$, let $d$ be their pointwise product, given by $d(\gamma) = x_1(\gamma) \cdot x_2(\gamma)$. By Proposition [\[prop:pre_proximal\]](#prop:pre_proximal){reference-type="ref" reference="prop:pre_proximal"} there exists a sequence $\{f_n\}$ of elements in $F$ such that either $\lim_n f_n d = c_{+1}$ or $\lim_n f_n d = c_{-1}$. In the first case $\lim_n f_n x_1 = \lim_n f_n x_2$, while in the second case $\lim_n f_n x_1 =-\lim_n f_n x_2$, and so this action resembles a proximal action. In fact, by identifying each $x \in \{-1,1\}^{\Gamma}$ with $-x$ one attains a proximal action, and indeed we do this below. However, this action has a fixed point---the constant functions---and therefore does not suffice to prove our result. We spend the remainder of this paper in deriving a new action from this one. The new action retains proximality but does not have fixed points. Consider the path $(\rfrac{1}{2}, \rfrac{1}{4},\rfrac{1}{8},\ldots,\rfrac{1}{2^n},\ldots)$ in the Schreier graph of $\Gamma$ (Figure [\[fig:schreier\]](#fig:schreier){reference-type="ref" reference="fig:schreier"}); it starts in the top blue node and follows the dotted edges through the blue nodes on the rightmost branch of the tree. The pointed Gromov-Hausdorff limit of this sequence of rooted graphs[^1] is given in Figure [\[fig:schreier2\]](#fig:schreier2){reference-type="ref" reference="fig:schreier2"}, and hence is also a Schreier graph of some transitive $F$-action $F \curvearrowright F/K$. In terms of the topology on the space $\mathrm{Sub}_F \subset \{0,1\}^F$ of the subgroups of $F$, the subgroup $K$ is the limit of the subgroups $K_n$, where $K_n$ is the stabilizer of $\rfrac{1}{2^n}$. It is easy to verify that $K$ is the subgroup of $F$ consisting of the transformations that stabilize $0$ and have right derivative $1$ at $0$ (although this fact will not be important). Let $\Lambda = F/K$. We can naturally identify with $\mathbb{Z}$ the chain black nodes at the top of $\Lambda$ (see Figure [\[fig:schreier2\]](#fig:schreier2){reference-type="ref" reference="fig:schreier2"}). Let $\Lambda'$ be the subgraph of $\Lambda$ in which the dotted edges connecting the black nodes have been removed. Given a black node $n \in \mathbb{Z}$, denote by $T_n$ the connected component of $n$ in $\Lambda'$; this includes the black node $n$, the chain that can be reached from it using solid edges, and the entire tree that hangs from it. Each graph $T_n$ is isomorphic to the Schreier graph of $\Gamma$, and so the graph $\Lambda$ is a covering graph of $\Gamma$ (in the category of Schreier graphs). Let \[\begin{aligned} \Psi \colon \Lambda \to \Gamma \end{aligned}\] be the covering map. That is, $\Psi$ is a graph isomorphism when restricted to each $T_n$, with the black nodes in $\Lambda$ mapped to the black node $0 \in \Gamma$. Using the map $\Psi$ we give names to the nodes in $\Lambda$. Denote the nodes in $T_0$ as $\{(0, \gamma) \,:\, \gamma \in \Gamma\}$ so that $\Psi(0,\gamma) = \gamma$. Likewise, in each $T_n$ denote by $(n,\gamma)$ the unique node in $T_n$ that $\Psi$ maps to $\gamma$. Hence we identify $\Lambda$ with \[\begin{aligned} \mathbb{Z} \times \Gamma = \{(n, \gamma)\,:\, n \in \mathbb{Z}, \gamma \in \Gamma\} \end{aligned}\] and the $F$-action is given by \[\begin{aligned} \label{a-action-on-Lambda} a (n,\gamma) &= (n, a \gamma)\\ \label{b-action-on-Lambda} b (n,\gamma) &= \begin{cases} (n, b \gamma)&\mbox{if }\gamma \neq 0\\ (n+1, 0)&\mbox{if }\gamma= 0 \end{cases} \end{aligned}\] Equip $\{-1,1\}^\Lambda$ with the product topology to get a compact space. As usual, the $F$-action on $\Lambda$ (given explicitly in [\[a-action-on-Lambda\]](#a-action-on-Lambda){reference-type="ref" reference="a-action-on-Lambda"} and [\[b-action-on-Lambda\]](#b-action-on-Lambda){reference-type="ref" reference="b-action-on-Lambda"}) defines a continuous action on $\{-1,1\}^\Lambda$. Consider $\pi:\{-1,1\}^\Gamma \to \{-1,1\}^\Lambda$, given by $\pi(x)(n, \gamma) = (-1)^n x(\gamma)$. Let $Y = \pi(\{-1,1\}^\Gamma) \subseteq \{-1,1\}^\Lambda$. The last $F$-space we define is $Z$, the set of pairs of mirror image configurations in $Y$: \[\begin{aligned} \label{the-space-Z} Z = \left\{\{y,-y\}\,:\,y\in Y \right\}. \end{aligned}\] Now it is clear that equipped with the quotient topology, $Z$ is a compact and Hausdorff $F$-space. Furthermore, we now observe that $Z$ admits an invariant measure. Consider the i.i.d. Bernoulli $1/2$ measure on $\{-1,1\}^\Gamma$, i.e. the unique Borel measure on $\{-1,1\}^\Gamma$, for which \[\begin{aligned} X_\gamma \colon & \{-1,1\}^\Gamma \to \{0, 1\},\quad x\mapsto \frac{x(\gamma)+1}{2} \end{aligned}\] are independent Bernoulli $1/2$ random variables for all $\gamma \in \Gamma$. Clearly, it is an invariant measure and hence it is pushed forward to an invariant measure on $Y$, and then on $Z$. In particular, this shows that $Z$ is not strongly proximal. [^1]: The limit of a sequence of rooted graphs $(G_n,v_n)$ is a rooted graph $(G,v)$ if each ball of radius $r$ around $v_n$ in $G_n$ is, for $n$ large enough, isomorphic to the ball of radius $r$ around $v$ in $G$ (see, e.g., ).	{'timestamp': '2017-04-28T02:03:47', 'yymm': '1607', 'arxiv_id': '1607.04915', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04915'}
# Introduction {#sec:1} The influence of social groups in pedestrian dynamics, especially in evacuation scenarios, is an area of recent interest, see e.g. and other contributions in these proceedings. The situations that are considered are widespread and well-known in everyday life. For example, many people visit concerts or soccer matches not alone, but together with family and friends in so-called social groups. In case of emergency, these groups will try to stay together during an evacuation. The strength of this cohesion depends on the composition of the social group. Several adult friends would form a loose group that is mainly connected via eye contact, whereas a mother would take her child's hand and form a strong or even fixed bond. In addition, even the size of the social groups could have an effect on the evacuation behaviour. In order to consider these phenomena in a more detailed way, a cooperation of researchers of the universities of Cologne and Wuppertal and the Forschungszentrum Jülich has performed several experiments aiming at the determination of the general influence of inhomogeneities on pedestrian dynamics. They contained two series of experiments with pupils of different ages in two schools in Wuppertal. The first series focussed on the determination of the fundamental diagram of inhomogeneous groups, i.e. pedestrians of different size. The second series of experiments considered evacuation scenarios. In several runs the parameters of the crowd of evacuating pupils were varied, i.e. the size of the social group and its structure and the interaction between the group members. Here we present first results for these evacuation experiments. # Teaching units {#sec:2} The experiments were accompanied by teaching units for all involved students providing an introduction into the topic of traffic and pedestrian dynamics. In classes of fifth and sixth grade, the focus of the classes was on the important quantities of pedestrian dynamics, especially density, time and bottleneck situations. This introduction to crowd effects and pedestrian behaviour was intended to raise awareness for their relevance for their everyday lives and safety issues. Therefore we arranged little experiments the students could perform themselves, e.g. the panic experiment according to Mintz (see Fig. [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"}). In small groups the pupils had to pull several wooden wedges out of a bottle with a narrow neck as fast as possible and observe the blocking of the wedges when every students pulls at the same time. This experiment was supposed to indicate that coordination can lead to better results compared to selfish behavior. The older pupils of classes 10 and 11 participated in an introduction to cellular automata and the physics of traffic. They received several worksheets on the Game of Life and other cellular automata, especially the Nagel-Schreckenberg model. The aim of these lessons was to obtain a first qualitative and quantitative understanding of the collective effects in traffic systems. This should help to increase the identification with the experiments they later participated in and raise awareness about the relevance of this kind of research for everyday life. # Experimental set-up and procedure {#sec:3} The experiments were performed in two schools in Wuppertal. Students of four classes participated as part of project work. The experimental room was built in the school's assembly hall. ## Experimental set-up The experimental area was a square room of $5\times 5$ m$^2$ bounded by several small buckets. In the center of this area there was a square starting area of $3\times 3$ m$^2$ denoted by the white marks. The students stepped into the room through the entrance that is shown below in Fig. [\[fig:2\]](#fig:2){reference-type="ref" reference="fig:2"} and assembled in the starting area. During the evacuation they had to leave the room using the exit on the left side. The exit door was built by two upstanding platforms and had a variable width changing between 0.8 m and 1.2 m. The area behind the door was connected to the waiting area before the entrance, so the students could walk on a closed path. For the collection of data all experiments were recorded by a camera system. This system was mounted on the hall's ceiling and contained customary digital cameras and GoPros. All students wore caps of different colour. Each colour represented a certain interval of body heights. The body height of each pupil was measured before the experiments started. This information is needed to determine the position accurately, but the different colours can also be used to draw conclusions about the composition of the group of evacuating pupils later in the video. All caps showed also a black point at the middle of the head. That allows to recognize and track each person in the video. ## Experimental procedure In general, the pupils had to perform several evacuation runs. For each run, a group of 32 to 46 persons assembled in the starting area, distributed nearly uniformly. During the evacuation, the students were allowed to use the whole experimental area. After a starting signal, the participants had to leave the room using the exit door. They should walk briskly and evacuate the room as fast as possible. The pupils were told to imagine there would be a kind of danger, like fire or smoke. However, they were not allowed to run, scramble or push each other. After leaving the room they had to assemble again in the waiting area in front of the entrance and to wait for the next run. The group of pupils that was placed into the experiment was compound in different ways to consider different parameters. The first parameter that was varied in the experiment was the composition of the entire group. At all, there were two different age classes allowing for three different group compositions. The crowd could consist only of children aged 10 to 12 years, only of young adults aged 15 to 17, or a mixture of both groups whereby children and youths were equally represented. The second parameter concerned the social group size. In several runs, the students had either to evacuate on their own without regarding the others around them, or to form pairs, or larger social groups. These groups could contain four, six or eight persons. Within one pair or social group the students had to try to stay together during the evacuation run. As a third parameter we considered the interaction within the social group. The nature of the interaction can be specified by (i) its strength and (ii) the hierarchy of group members. Regarding the interaction strength, the group members could either be connected loosely, by just trying to stay together via eye contact, or they could have a fixed bond. A fixed bond was realized by holding each other's hand or some other physical contact. Furthermore, hierarchy of the group members could be different. In the first case, all partners were treated equally. Each group member had to leave the room and to stay together with their partners. In the other case, one student was declared as the \"leader\", the other one as the \"follower\". The leader had to leave the room without regarding its partner or the other students, whereas the follower just had to follow the leader through the room. This leads to four different ways to form pairs during the evacuation run. In the case of age-matched partners, the leader was chosen randomly. For the runs with the mixed crowd, the pairs were composed of one child and one teenager that took the part of the leader in the runs they were needed. All runs with larger social groups were done with loose bonds. In social groups of same age, there was no leader, but in mixed social groups one of the youths was declared as the leader. To analyze the experiments the videos of the camera system that was mounted on the hall's ceiling was available. For each run of the experiment there is a video sequence. Using the `PeTrack` software, it is possible to extract the trajectories for each person and each run. The students were recognized via the black point in the middle of their coloured caps. The position of this point was tracked in each frame, generating the trajectory of each participant. # Analysis {#sec:4} First, we focus on the analysis of the data regarding the influence of group size on the evacuation scenario. Therefore, we use the data of one school and of the runs with larger groups. Most of these experiments were performed only with the older pupils with loose bonds and no leader-follower relationship, to which we restrict our analysis for now. In different runs, the students formed groups of four, six and eight persons. In addition, one run with groups of six students and with an explicitly cooperative behaviour within the group was performed. They should concentrate a bit more on their group members and try to leave the room together. For comparison, we also consider the run with pairs and a loose bond that can be seen as a smaller group of two persons. ## Evacuation times for large groups First we consider the evacuation time. In Fig. [\[fig:3\]](#fig:3){reference-type="ref" reference="fig:3"}, we plot the number of evacuated persons against the time needed to leave the room. The results can be compared between the different runs. The evacuation time for each person is defined as the time difference between the beginning of the evacuation and the moment when the person passes the door, exactly when he/she leaves the aisle that is formed by the two platforms. The beginning of the evacuation can be determined only approximately because the starting signal is not audible in the videos that are used for the extraction of the trajectories. For extracting the evacuation times we set the beginning on the moment of the first movement towards the door. However, for the analysis the influences of this inaccurate definition, the pre-movement time or other delays should be minimized. In doing so, we take the evacuation time of the very first person that left the room and subtract it from all the other times. Thereby, all plots start at zero for the first person and it is easier to compare different runs. For the analysis of the runs with larger groups the evacuation times are shown in Fig. [\[fig:3\]](#fig:3){reference-type="ref" reference="fig:3"}. All graphs show a nearly linear behaviour that could be expected. At the beginning of the evacuation all evacuation times are roughly the same. Between three and six evacuated persons the curves start to split into two groups. After increasing slightly, the difference between the two progresses remains nearly constant until the end of the evacuation. The main insight is that there are several runs that are clearly faster than other ones. The upper two curves represent the evacuation in pairs and in groups of six with very cooperative behaviour. The lower graphs show the runs with larger groups of four, six and eight persons. Within the two groups of curves the differences are not large enough to separate the runs from each other. However, in the lower group, the run with eight participants per group seems to be a bit faster at the end of the evacuation. The run with six participants per group and cooperative behaviour is clearly slower than the run with same group size but without this instruction. These results indicate that forming groups is advantageous for the evacuation, whereas behaving cooperatively inhibits this effect. While looking for reasons for the differences in evacuation times, one first approach could be to determine the density distribution. Therefore, we determined the Voronoi cells within the experimental room for each person at different times. As a measure of density we coloured all cells dependent on their size: smaller cells are coloured in shades of red, larger ones in blue. In Fig. [\[fig:4\]](#fig:4){reference-type="ref" reference="fig:4"} the density distributions for the run with pairs and with groups of four persons are shown. It is clearly seen that the distribution for the run in (b) is a bit narrower than the other one at the same time step. That means when forming groups, the children order rather behind each other than next to each other in front of the door. This behaviour seems to be advantageous for evacuating the room as it leads to a shorter evacuation time. ## First attempts to interpretation {#sec:5} The results obtained so far suggest certain interpretations which, however, need to be substantiated by further experiments with better statistics. It is obvious from the plot of the evacuation times that increasing the group size leads to a decrease in evacuation times. The density distributions show the pupils ordered in different ways for forming groups than for pairs. A possible explanation is that the persons subordinate within the group and just follow the other group members. Because of that, there may be less conflicts between persons that meet at the door in competing for space. A person is just in competition with persons of other groups, not with own group members. Increasing the group size reduces the number of possible competitors. This reduction of conflicts may have a positive influence on the evacuation time. When the children have to show cooperative behaviour, the evacuation is slower than without this instruction. It is a possible explanation that here the effort to stay together is larger and reduces the effort to leave the room. # Summary and outlook {#sec:6} We performed experiments under laboratory conditions to determine the influence of social groups on evacuations. A comparison of evacuation times between runs with different group sizes shows that increasing the group sizes lowers the evacuation time. The participants order in a different way for larger groups. These first preliminary results have to be analyzed in more detail. The statistics need to be improved by further experiments. However, with the help of the density distributions, photographs of the finish and the data of the second school we hope to get more information from the present experiments, e.g. about the microscopic mechanisms especially close to the exit. In addition, there are some few parameters that should also be analyzed, e.g. the effect of body size and age.	{'timestamp': '2016-07-18T02:07:24', 'yymm': '1607', 'arxiv_id': '1607.04518', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04518'}
# Introduction {#sec:intro} Policy search algorithms based on supervised learning from a computational or human "teacher" have gained prominence in recent years due to their ability to optimize complex policies for autonomous flight, video game playing, and bipedal locomotion. Among these methods, guided policy search algorithms are particularly appealing due to their ability to adapt the teacher to produce data that is best suited for training the final policy with supervised learning. Such algorithms have been used to train complex deep neural network policies for vision-based robotic manipulation, as well as a variety of other tasks. However, convergence results for these methods typically follow by construction from their formulation as a constrained optimization, where the teacher is gradually constrained to match the learned policy, and guarantees on the performance of the final policy only hold at convergence if the constraint is enforced exactly. This is problematic in practical applications, where such algorithms are typically executed for a small number of iterations. In this paper, we show that guided policy search algorithms can be interpreted as approximate variants of mirror descent under constraints imposed by the policy parameterization, with supervised learning corresponding to a projection onto the constraint manifold. Based on this interpretation, we can derive a new, simplified variant of guided policy search, which corresponds exactly to mirror descent under linear dynamics and convex policy spaces. When these convexity and linearity assumptions do not hold, we can show that the projection step is approximate, up to a bound that depends on the step size of the algorithm, which suggests that for a small enough step size, we can achieve continuous improvement. The form of this bound provides us with intuition about how to adjust the step size in practice, so as to obtain a simple algorithm with a small number of hyperparameters. The main contribution of this paper is a simple new guided policy search algorithm that can train complex, high-dimensional policies by alternating between trajectory-centric reinforcement learning and supervised learning, as well as a connection between guided policy search methods and mirror descent. We also extend previous work on bounding policy cost in terms of KL divergence to derive a bound on the cost of the policy at each iteration, which provides guidance on how to adjust the step size of the method. We provide empirical results on several simulated robotic navigation and manipulation tasks that show that our method is stable and achieves similar or better performance when compared to prior guided policy search methods, with a simpler formulation and fewer hyperparameters. # Guided Policy Search Algorithms {#sec:gps} We first review guided policy search methods and background. Policy search algorithms aim to optimize a parameterized policy $\pi_\theta(\action_t\|\state_t)$ over actions $\action_t$ conditioned on the state $\state_t$. Given stochastic dynamics $p(\mathbf{x}_{t+1}\|\state_t,\action_t)$ and cost $\ell(\state_t,\action_t)$, the goal is to minimize the expected cost under the policy's trajectory distribution, given by $J(\theta) = \sum_{t=1}^T E_{\pi_\theta(\state_t,\action_t)} [\ell(\state_t,\action_t)]$, where we overload notation to use $\pi_\theta(\state_t,\action_t)$ to denote the marginals of $\pi_\theta(\tau) = p(\mathbf{x}_1)\prod_{t=1}^T p(\mathbf{x}_{t+1}\|\state_t,\action_t) \pi_\theta(\action_t\|\state_t)$, where $\tau =\{\mathbf{x}_1,\mathbf{u}_1,\dots,\mathbf{x}_T,\mathbf{u}_T\}$ denotes a trajectory. A standard reinforcement learning (RL) approach to policy search is to compute the gradient $\nabla_\theta J(\theta)$ and use it to improve $J(\theta)$. The gradient is typically estimated using samples obtained from the real physical system being controlled, and recent work has shown that such methods can be applied to very complex, high-dimensional policies such as deep neural networks. However, for complex, high-dimensional policies, such methods tend to be inefficient, and practical real-world applications of such model-free policy search techniques are typically limited to policies with about one hundred parameters. Instead of directly optimizing $J(\theta)$, guided policy search algorithms split the optimization into a "control phase" (which we'll call the C-step) that finds multiple simple local policies $p_i(\action_t\|\state_t)$ that can solve the task from different initial states $\mathbf{x}_1^i \sim p(\mathbf{x}_1)$, and a "supervised phase" (S-step) that optimizes the global policy $\pi_\theta(\action_t\|\state_t)$ to match all of these local policies using standard supervised learning. In fact, a variational formulation of guided policy search corresponds to the EM algorithm, where the C-step is actually the E-step, and the S-step is the M-step. The benefit of this approach is that the local policies $p_i(\action_t\|\state_t)$ can be optimized separately using domain-specific local methods. Trajectory optimization might be used when the dynamics are known, while local RL methods might be used with unknown dynamics, which still requires samples from the real system, though substantially fewer than the direct approach, due to the simplicity of the local policies. This sample efficiency is the main advantage of guided policy search, which can train policies with nearly a hundred thousand parameters for vision-based control using under 200 episodes, in contrast to direct deep RL methods that might require orders of magnitude more experience. This means that if we can bound the KL-divergence between the policies, then the total variation divergence between their state marginals (given by $D_\text{TV}(p(\state_t)\\| \pi_\theta(\state_t)) = \frac{1}{2}\\|p(\state_t)-\pi_\theta(\state_t)\\|_1$) will also be bounded. This bound allows us in turn to relate the total expected costs of the two policies to each other according to the following lemma, which we prove in Appendix [\[app:cost_bound\]](#app:cost_bound){reference-type="ref" reference="app:cost_bound"}: This bound on the cost of $\pi_\theta(\action_t\|\state_t)$ tells us that if we update $p(\action_t\|\state_t)$ so as to decrease its total cost or decrease its KL-divergence against $\pi_\theta(\action_t\|\state_t)$, we will eventually reduce the cost of $\pi_\theta(\action_t\|\state_t)$. For the MDGPS algorithm, this bound suggests that we can ensure improvement of the global policy within a small number of iterations by appropriately choosing the constraint $\epsilon$ during the C-step. Recall that the C-step constrains $\sum_{t=1}^T \epsilon_t \leq \epsilon$, so if we choose $\epsilon$ to be small enough, we can close the gap between the local and global policies. Optimizing the bound directly turns out to produce very slow learning in practice, because the bound is very loose. However, it tells us that we can either decrease $\epsilon$ toward the end of the optimization process or if we observe the global policy performing much worse than the local policies. We discuss how this idea can be put into action in the next section. ## Step Size Selection {#sec:step} In prior work, the step size $\epsilon$ in the local policy optimization is adjusted by considering the difference between the predicted change in the cost of the local policy $p(\action_t\|\state_t)$ under the fitted dynamics, and the actual cost obtained when sampling from that policy. The intuition is that, because the linearized dynamics are local, we incur a larger cost the further we deviate from the previous policy. We can adjust the step size by estimating the rate at which the additional cost is incurred and choose the optimal tradeoff. Let $\ell_{k-1}^{k-1}$ denote the expected cost under the previous local policy $\bar{p}(\action_t\|\state_t)$, $\ell_{k-1}^k$ the cost under the current local policy $p(\action_t\|\state_t)$ and the previous fitted dynamics (which were estimated using samples from $\bar{p}(\action_t\|\state_t)$ and used to optimize $p(\action_t\|\state_t)$), and $\ell_k^k$ the cost of the current local policy under the dynamics estimated using samples from $p(\action_t\|\state_t)$ itself. Each of these can be computed analytically under the linearized dynamics. We can view the difference $\ell_k^k-\ell_{k-1}^k$ as the additional cost we incur from imperfect dynamics estimation. Previous work suggested modeling the change in cost as a function of $\epsilon$ as following: $\ell_k^k-\ell_{k-1}^{k-1} = a\epsilon^2 + b\epsilon$, where $b$ is the change in cost per unit of KL-divergence, and $a$ is additional cost incurred due to inaccurate dynamics. This model is reasonable because the integral of a quadratic cost under a linear-Gaussian system changes roughly linearly with KL-divergence. The additional cost due to dynamics errors is assumes to scale superlinearly, allowing us to solve for $b$ by looking at the difference $\ell_k^k-\ell_{k-1}^k$ and then solving for a new optimal $\epsilon^\prime$ according to $\epsilon^\prime =-b/2a$, resulting in the update $\epsilon^\prime = \epsilon (\ell_{k-1}^k-\ell_{k-1}^{k-1}) / 2(\ell_{k-1}^k-\ell_k^k)$. In MDGPS, we propose to use two step size adjustment rules. The first rule simply adapts the previous method to the case where we constrain the new local policy $p(\action_t\|\state_t)$ against the global policy $\pi_\theta(\action_t\|\state_t)$, instead of the previous local policy $\bar{p}(\action_t\|\state_t)$. In this case, we simply replace $\ell_{k-1}^{k-1}$ with the expected cost under the previous global policy, given by $\ell_{k-1}^{k-1,\pi}$, obtained using its linearization $\bar{\pi}_\theta(\action_t\|\state_t)$. We call this the "classic" step size: $\epsilon^\prime = \epsilon (\ell_{k-1}^{k}-\ell_{k-1}^{k-1,\pi}) / 2(\ell_{k-1}^k-\ell_k^k)$. However, we can also incorporate intuition from the bound in the previous section to obtain a more conservative step adjustment that reduces $\epsilon$ not only when the obtained local policy improvement doesn't meet expectations, but also when we detect that the global policy is unable to reproduce the behavior of the local policy. In this case, reducing $\epsilon$ reduces the KL-divergence between the global and local policies which, as shown in the previous section, tightens the bound on the global policy return. As mentioned previously, directly optimizing the bound tends to perform poorly because the bound is quite loose. However, if we estimate the cost of the global policy using its linearization, we can instead adjust the step size based on a simple model of global policy cost. We use the same model for the change in cost, given by $\ell_k^{k,\pi}-\ell_{k-1}^{k-1,\pi} = a\epsilon^2 + b\epsilon$. However, for the term $\ell_k^k$, which reflects the actual cost of the new policy, we instead use the cost of the new global policy $\ell_k^{k,\pi}$, so that $a$ now models the additional loss due to both inaccurate dynamics and inaccurate projection: if $\ell_k^{k,\pi}$ is much worse than $\ell_{k-1}^{k}$, then either the dynamics were too local, or S-step failed to match the performance of the local policies. In either case, we decrease the step size.[^3] As before, we can solve for the new step size $\epsilon^\prime$ according to $\epsilon^\prime = \epsilon (\ell_{k-1}^{k}-\ell_{k-1}^{k-1,\pi}) / 2(\ell_{k-1}^{k}-\ell_k^{k,\pi})$. We call this the "global" step size. Details of how each quantity in this equation is computed are provided in Appendix [\[app:step\]](#app:step){reference-type="ref" reference="app:step"}. # Relation to Prior Work {#sec:prior_work} While we've discussed the connections between MDGPS and prior guided policy search methods, in this section we'll also discuss the connections between our method and other policy search methods. One popular supervised policy learning methods is DAGGER , which also trains the policy using supervised learning, but does not attempt to adapt the teacher to provide better training data. MDGPS removes the assumption in DAGGER that the supervised learning stage has bounded error against an arbitrary teacher policy. MDGPS does not need to make this assumption, since the teacher can be adapted to the limitations of the global policy learning. This is particularly important when the global policy has computational or observational limitations, such as when learning to use camera images for partially observed control tasks or, as shown in our evaluation, blind peg insertion. When we sample from the global policy $\pi_\theta(\action_t\|\state_t)$, our method resembles policy gradient methods with KL-divergence constraints. However, policy gradient methods update the policy $\pi_\theta(\action_t\|\state_t)$ at each iteration by linearizing with respect to the policy parameters, which often requires small steps for complex, nonlinear policies, such as neural networks. In contrast, we linearize in the space of time-varying linear dynamics, while the policy is optimized at each iteration with many steps of supervised learning (e.g. stochastic gradient descent). This makes MDGPS much better suited for quickly and efficiently training highly nonlinear, high-dimensional policies. # Experimental Evaluation {#sec:experiments} We compare several variants of MDGPS and a prior guided policy search method based on Bregman ADMM (BADMM). We evaluate all methods on one simulated robotic navigation task and two manipulation tasks. Guided policy search code, including BADMM and MDGPS methods, is available at `https://www.github.com/cbfinn/gps`. #### Obstacle Navigation. In this task, a 2D point mass (grey) must navigate around obstacles to reach a target (shown in green), using velocities and positions relative to the target. We use $N=5$ initial states, with 5 samples per initial state per iteration. The target and obstacles are fixed, but the starting position varies. #### Peg Insertion. This task, which is more complex, requires controlling a 7 DoF 3D arm to insert a tight-fitting peg into a hole. The hole can be in different positions, and the state consists of joint angles, velocities, and end-effector positions relative to the target. This task is substantially more challenging physically. We use $N=9$ different hole positions, with 5 samples per initial state per iteration. #### Blind Peg Insertion. The last task is a blind variant of the peg insertion task, where the target-relative end effector positions are provided to the local policies, but not to the global policy $\pi_\theta(\action_t\|\state_t)$. This requires the global policy to search for the hole, since no input to the global policy can distinguish between the different initial state $\mathbf{x}_1^i$. This makes it much more challenging to adapt the global and local policies to each other, and makes it impossible for the global learner to succeed without adaptation of the local policies. We use $N=4$ different hole positions, with 5 samples per initial state per iteration. The global policy for each task consists of a fully connected neural network with two hidden layers with $40$ rectified linear units. The same settings are used for MDGPS and the prior BADMM-based method, except for the difference in surrogate costs, constraints, and step size adjustment methods discussed in the paper. Results are presented in Figure [\[fig:results\]](#fig:results){reference-type="ref" reference="fig:results"}. On the easier point mass and peg tasks, all of the methods achieve similar performance. However, the MDGPS methods are all substantially easier to apply to these tasks, since they have very few free hyperparameters. An initial step size must be selected, but the adaptive step size adjustment rules make this choice less important. In contrast, the BADMM method requires choosing an initial weight on the augmented Lagrangian term, an adjustment schedule for this term, a step size on the dual variables, and a step size for local policies, all of which have a substantial impact on the final performance of the method (the reported results are for the best setting of these parameters, identified with a hyperparameter sweep). On the harder blind peg task, MDGPS consistently outperforms BADMM when sampling from the local policies ("off policy"), with both the classic and global step sizes. This is particularly apparent in the success rates in Table [1](#tbl:success){reference-type="ref" reference="tbl:success"}, which shows that the MDGPS policies succeed at actually inserting the peg into the hole more often and on more conditions. This suggests that our method is better able to improve global policies particularly in situations where informational or representational constraints make naïve imitation of the local policies insufficient to solve the task. On-policy sampling tends to learn slower, since the approximate projection causes the global policy to lag behind the local policy in performance, but this method is still able to consistently improve the global policies. Sampling from the global policies may be desirable in practice, since the global policies can directly use observations at runtime instead of requiring access to the state. The global step size also tends to be more conservative, but produces more consistent and monotonic improvement. # Discussion and Future Work {#sec:discussion} We presented a new guided policy search method that corresponds to mirror descent under linearity and convexity assumptions, and showed how prior guided policy search methods can be seen as approximating mirror descent. We provide a bound on the return of the global policy in the nonlinear case, and argue that an appropriate step size can provide improvement of the global policy in this case also. Our analysis provides us with the intuition to design an automated step size adjustment rule, and we illustrate empirically that our method achieves good results on a complex simulated robotic manipulation task while requiring substantially less tuning and hyperparameter optimization than prior guided policy search methods. Manual tuning and hyperparameter searches are a major challenge across a range of deep reinforcement learning algorithms, and developing scalable policy search methods that are simple and reliable is vital to enable further progress. As discussed in Section [5](#sec:prior_work){reference-type="ref" reference="sec:prior_work"}, MDGPS has interesting connections to other policy search methods. Like DAGGER, MDGPS uses supervised learning to train the policy, but unlike DAGGER, MDGPS does not assume that the learner is able to reproduce an arbitrary teacher's behavior with bounded error, which makes it very appealing for tasks with partial observability or other limits on information, such as learning to use camera images for robotic manipulation. When sampling directly from the global policy, MDGPS also has close connections to policy gradient methods that take steps of fixed KL-divergence, but with the steps taken in the space of trajectories rather than policy parameters, followed by a projection step. In future work, it would be interesting to explore this connection further, so as to develop new model-free policy gradient methods. [^1]: As mentioned previously, the initial state $\mathbf{x}_1^i$ of each local policy $p_i(\action_t\|\state_t)$ is assumed to be drawn from $p(\mathbf{x}_1)$, hence the outer sum corresponds to Monte Carlo integration of the expectation under $p(\mathbf{x}_1)$. [^2]: Note that we flip the KL-divergence inside the expectation, following . We found that this produced better results. The intuition behind this is that, because $\log p_i(\action_t\|\state_t)$ is proportional to the Q-function of $p_i(\action_t\|\state_t)$ (see Appendix [\[app:cstep\]](#app:cstep){reference-type="ref" reference="app:cstep"}), $D_\text{KL}(\pi_\theta(\action_t\|\mathbf{x}_{t,i,j}) \\| p_i(\action_t\|\mathbf{x}_{t,i,j})$ minimizes the cost-to-go under $p_i(\action_t\|\state_t)$ with respect to $\pi_\theta(\action_t\|\state_t)$, which provides for a more informative objective than the unweighted likelihood in Equation ([\[eqn:mirrorkl\]](#eqn:mirrorkl){reference-type="ref" reference="eqn:mirrorkl"}). [^3]: Although we showed before that the discrepancy depends on $\sum_{t=1}^T \sqrt{2\epsilon}_t$, here we use $\epsilon^2$. This is a simplification, but the net result is the same: when the global policy is worse than expected, $\epsilon$ is reduced.	{'timestamp': '2016-07-18T02:09:29', 'yymm': '1607', 'arxiv_id': '1607.04614', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04614'}
# Introduction Spatio--temporal models are widely used by practitioners. Explaining economic, environmental, social, or biological phenomena, such as peer influence, neighbourhood effects, contagion, epidemics, interdependent preferences, climate change, and so on, are only some of the interesting applications of such models. A widely used spatio--temporal model is the spatial dynamic panel data model (SDPD) proposed and analysed by. See for a survey. To improve adaptivity of SDPD models, recently proposed a generalized model that assigns different coefficients to varied locations and assumes heteroskedastic and spatially correlated errors. The model is \[\label{eqn1} {\mathbf y}_t = D(\boldsymbol{\lambda}_0){\mathbf W}{\mathbf y}_t + D({\boldsymbol{\lambda}_1}){\mathbf y}_{t-1} + D(\boldsymbol{\lambda}_2){\mathbf W}{\mathbf y}_{t-1} + \mbox{\boldmath$\varepsilon$}_t,\] where the vector ${\mathbf y}_t$ is of order $p$ and contains the observations at time $t$ from $p$ different locations; the errors $\mbox{\boldmath$\varepsilon$}_t$ are serially uncorrelated; the spatial matrix ${\mathbf W}$ is a weight matrix with zero main diagonal and is assumed to be known; $D(\boldsymbol{\lambda}_j)$ with $j=0,1,2$ are diagonal matrices, and $\boldsymbol{\lambda}_j$ are the vectors with the spatial coefficients $\lambda_{ji}$ for $i=1,\ldots,p$. The generalized SDPD model in ([\[eqn1\]](#eqn1){reference-type="ref" reference="eqn1"}) guarantees adaptivity by means of its $3p$ parameters. It is characterized by the sum of three terms: the spatial component, driven by matrix ${\mathbf W}$ and the spatial parameter $\boldsymbol{\lambda}_0$; the dynamic component, driven by the autoregressive parameter $\boldsymbol{\lambda}_1$; and the spatial--dynamic component, driven by matrix ${\mathbf W}$ and the spatial--autoregressive parameter $\boldsymbol{\lambda}_2$. If the vectors $\boldsymbol{\lambda}_j$ are scalars for all $j$, then model ([\[eqn1\]](#eqn1){reference-type="ref" reference="eqn1"}) reduces to the classic SDPD of. The errors $\mbox{\boldmath$\varepsilon$}_t$ in model ([\[eqn1\]](#eqn1){reference-type="ref" reference="eqn1"}) are serially uncorrelated and may show heteroskedasticity and cross-correlation over space, so that $\mathop{var}(\mbox{\boldmath$\varepsilon$}_t)$ is a full matrix. This is a novelty compared with the SDPD model of, where the errors must be cross-uncorrelated and homoskedastic in order to get consistency of the estimators. A setup similar to ours for the errors has been also considered by and, but not for panel models. However, their estimators are generally based on the instrumental variables technique, in order to overcome the endogeneity of the zero-lag regressor. For the generalized SDPD model, instead, propose a new estimation procedure based on a generalized Yule--Walker approach. They show the consistency of the estimators under regularity assumptions. They also derive the convergence rate and the conditions under which the estimation procedure does not suffer for high-dimensional setups, notwithstanding the large number of parameters to be estimated (which become infinite with the dimension $p$). In real data applications, it is important to check the validity of the assumptions required for the consistency of the estimation procedure. See, for example, the assumptions and asymptotic analysis in and as well as the references therein. Checking such assumptions on real data is often not easy; at times, they are clearly violated. Moreover, the spatial matrix ${\mathbf W}$ is assumed to be known, but in many cases, this is not true, and it must be estimated. For example, the spatial weights can be associated with "similarities" between spatial units and measured by estimated correlations. Another example is when the spatial weights are zeroes/ones, depending on the "relationships" between the spatial units, but the neighbourhood structure of ${\mathbf W}$ is unknown (i.e., it is not known where the ones must be allocated). In such cases, the performance of the SDPD models needs to be investigated. Readers are advised to refer to recent papers on spatial matrix estimation (see, among others, ). Motivated by the above considerations, we propose a new version of the SDPD model obtained by adding a constraint to the spatial parameters of the generalized SDPD of. New estimators of the parameters are proposed and investigated theoretically and empirically. The new model is called stationary SDPD and has several advantages. First, the structure of the model and the interpretation of the parameters are simpler than the generalized SDPD model, with the consequence that the assumptions underlying the theoretical results are clearer and can be checked easily with real data. Moreover, the estimation procedure is fast and simple to implement. Second, the proposed estimators of the parameters are always unbiased and reach the $\sqrt{T}$ convergence rate (where $T$ is the temporal length of the time series) even in the high-dimensional case, although the number of parameters tends to infinity with the dimension $p$. Last, but not least, our model allows wider application than the classic SDPD model, and it is general enough to represent a wide range of multivariate linear processes that can be implicitly interpreted (when they are not explicitly interpretable) as spatio--temporal processes, with respect to a "latent spatial matrix," which needs to be estimated. A big implication of this is that our model is not necessarily confined to the representation of strict spatio--temporal processes (where the spatial matrix is known), but it can also be considered as a valid alternative to the general VAR models (where there is no spatial matrix), with two relevant advantages: i) more efficient estimation of the model and ii) the possibility of estimating the model even when $p>T$, thus avoiding the curse of dimensionality that characterizes the VAR models. Surprisingly, the simulation results show the remarkably better performance of our model and the new estimation procedure compared with the standard VAR model and the standard estimation procedure, even when the spatial matrix is latent and, therefore, to be estimated (see section [5.3](#matrixA){reference-type="ref" reference="matrixA"}). The rest of the paper is organized as follows. Section [2](#sdpd){reference-type="ref" reference="sdpd"} presents the new model and discusses the issue of identifiability. Section [3](#est_alg){reference-type="ref" reference="est_alg"} describes the estimation procedure. The theoretical results are shown in section [4](#asymptotic){reference-type="ref" reference="asymptotic"}. The empirical performance of the estimation procedure is investigated in section [5](#simulazioni){reference-type="ref" reference="simulazioni"}. Finally, all the proofs are provided in the Appendix. # A constrained spatio--temporal model: the stationary SDPD {#sdpd} In the sequel, we assume that ${\mathbf y}_1, \cdots, {\mathbf y}_T$ are the observations from a stationary process defined by ([\[eqn3\]](#eqn3){reference-type="ref" reference="eqn3"}). The transpose of a matrix ${\mathbf A}$ is denoted with ${\mathbf A}^T$. We assume that the process has mean zero and denote with $\boldsymbol{\Sigma}_j=\mathop{cov}({\mathbf y}_t,{\mathbf y}_{t-j})=E({\mathbf y}_t{\mathbf y}_{t-j}^T)$ the covariance matrix of the process at the lag $j$. The generalized SDPD model in ([\[eqn1\]](#eqn1){reference-type="ref" reference="eqn1"}) can be reformulated as follows. \[\left[{\mathbf I}_p-D(\boldsymbol{\lambda}_0){\mathbf W}\right]{\mathbf y}_t = D({\boldsymbol{\lambda}_1})\left[{\mathbf I}_p-D(\boldsymbol{\lambda}^+_2){\mathbf W}\right]{\mathbf y}_{t-1} + \mbox{\boldmath$\varepsilon$}_t, \label{eqn3}\] where $\boldsymbol{\lambda}^+_2$ is a vector obtained by dividing the elements of $\boldsymbol{\lambda}_2$ by the corresponding elements of $\boldsymbol{\lambda}_1$ (assuming, for now, that all the coefficients in $\boldsymbol{\lambda}_1$ are different from zero). Note that model ([\[eqn3\]](#eqn3){reference-type="ref" reference="eqn3"}) is equivalent to a multivariate (auto)regression between a linear combination of ${\mathbf y}_t$ and a linear combination of the lag ${\mathbf y}_{t-1}$, where the weights of the two linear combinations depend on ${\mathbf W}$ and the coefficients $\boldsymbol{\lambda}_0$ and $\boldsymbol{\lambda}^+_2$, respectively. \[\label{zeta} {\mathbf z}_t^{(\boldsymbol{\lambda}_0,{\mathbf W})} = D({\boldsymbol{\lambda}_1}){\mathbf z}_{t-1}^{(\boldsymbol{\lambda}^+_2,{\mathbf W})} + \mbox{\boldmath$\varepsilon$}_t.\] Some special cases may arise from model ([\[eqn3\]](#eqn3){reference-type="ref" reference="eqn3"}) by adding restrictions on the parameters $\boldsymbol{\lambda}_{j}$. First, if we assume that the spatial parameters are constant over space, that is, $\boldsymbol{\lambda}_{j}$ is scalar for $j=0,1,2$, then we obtain the classic SDPD model of. Another constrained model, proposed and analysed in this paper, may be derived by assuming that $\boldsymbol{\lambda}_0=\boldsymbol{\lambda}^+_2$. The reason underlying the choice of this constraint is a generalized assumption of stationarity. In time series analysis, stationarity means that the dependence structure of the process is constant (in some sense) over time. In particular, second-order stationarity assumes that correlations between the observations $({\mathbf y}_t,{\mathbf y}_{t-j})$ depend on the lag $j$ but not on $t$, implying that $\mathop{var}({\mathbf y}_t)$ is constant for all $t$. However, in spatio--temporal time series, there are two kinds of correlations: temporal correlations, involving observations at different time points, and spatial correlations, involving observations at different spatial units. As we refer to stationarity, it makes sense to assume that spatial correlations are also time-invariant, which means that the weights in ([\[zeta\]](#zeta){reference-type="ref" reference="zeta"}) must not change over time, thus, $\left[{\mathbf I}_p-D(\boldsymbol{\lambda}_0){\mathbf W}\right]=\left[{\mathbf I}_p-D(\boldsymbol{\lambda}^+_2){\mathbf W}\right]$, also implying that $\mathop{var}({\mathbf z}_t)$ is the same for all $t$. Therefore, we add the constraint $\boldsymbol{\lambda}_0=\boldsymbol{\lambda}^+_2$, and the model becomes \[\label{b1} \left[{\mathbf I}_p-D(\boldsymbol{\lambda}_0){\mathbf W}\right]{\mathbf y}_t = D(\boldsymbol{\lambda}_1)\left[{\mathbf I}_p-D(\boldsymbol{\lambda}_0){\mathbf W}\right]{\mathbf y}_{t-1} + \mbox{\boldmath$\varepsilon$}_t.\] We denote the model as stationary SDPD. Model ([\[b1\]](#b1){reference-type="ref" reference="b1"}) has several advantages that will be shown in the following sections. Above all, imposing spatio--temporal stationarity helps gain efficiency while still preserving the spatial adaptability that characterizes the generalized SDPD model of. Moreover, our model allows representation of a wide range of multivariate processes by means of a simple model subject to few assumptions that can be easily checked using real data. Finally, it is worthwhile to stress the difference between the SDPD model of, the generalized SDPD model of, and the stationary SDPD model proposed here. The first model imposes that the spatial relationships be the same for all units, since the coefficients $\lambda_j$ (with $j=0,1,2$) do not change with $i=1,\ldots,p$. Instead, the stationary SDPD model in ([\[b1\]](#b1){reference-type="ref" reference="b1"}) allows varied coefficients for different spatial units, as in the generalized SDPD of, but they are assumed to be time-invariant thanks to a constraint on the time-lagged parameters. Of course, the estimation procedures vary for the three cases in terms of the convergence rates. The constrained model underlying our stationary SDPD allows the estimation procedure to reach the $\sqrt{T}$ convergence rate and to guarantee unbiased estimators, whatever the dimension $p$ and even when $p\rightarrow\infty$ at any rate. This is a big improvement with respect to the other two models. In fact, for the classic SDPD model, the estimators are characterized by a $\sqrt{Tp}$ convergence rate (which is faster than that of our model, since they have only three parameters to estimate instead of $2p$), but a bias of order $T^{-1}$ exists, and it does not vanish when $p/T\rightarrow\infty$ (see Theorem 3 of ). On the other hand, the convergence rates of the estimators in the generalized SDPD model are slower than those of our model and deteriorate when $p\rightarrow\infty$ at a rate faster than $\sqrt{T}$ (see Theorem 2 of ). ## Identification of parameters in the case of cross-uncorrelated errors In this section, we assume, for simplicity, that the matrix $\boldsymbol{\Sigma}_0^{\varepsilon}=\mathop{var}(\mbox{\boldmath$\varepsilon$}_t)$ is diagonal (i.e., there is heteroskedasticity but no cross-correlation in the error process) and discuss the identifiability of the model. In the next section, we generalize the problem by also adding some cross-correlations in the error process. Defining ${\mathbf z}_t^{(0)}=\left[{\mathbf I}_p-D(\boldsymbol{\lambda}_0){\mathbf W}\right]{\mathbf y}_t$, model ([\[b1\]](#b1){reference-type="ref" reference="b1"}) can be reformulated as \[\begin{aligned} {\mathbf z}_t^{(0)} &=& D(\boldsymbol{\lambda}_1){\mathbf z}_{t-1}^{(0)} + \mbox{\boldmath$\varepsilon$}_t, \label{b1ter} \end{aligned}\] which is a transformed VAR process with uncorrelated components, since $D(\boldsymbol{\lambda}_1)$ is diagonal. Given that we assume that $\boldsymbol{\Sigma}_0^{\varepsilon}$ is also diagonal, the coefficients $\lambda_{1i}$ for $i=1,\ldots,p$, represent the slopes of $p$ univariate autoregressive models with respect to the latent variables $z_{it}^{(0)}$. Therefore, $\lambda_{1i}\equiv\mathop{cor}(z_{it}^{(0)}, z_{i,t-1}^{(0)})$. From ([\[b1\]](#b1){reference-type="ref" reference="b1"}), it follows that \[\left[{\mathbf I}_p-D(\boldsymbol{\lambda}_0){\mathbf W}\right]\boldsymbol{\Sigma}_1 = D(\boldsymbol{\lambda}_1)\left[{\mathbf I}_p-D(\boldsymbol{\lambda}_0){\mathbf W}\right]\boldsymbol{\Sigma}_0 \label{seconda}\\ \] and for the $i$-th equation, \[\label{vincolo} ({\mathbf e}^T_i-\lambda_{0i}{\mathbf w}^T_i)\boldsymbol{\Sigma}_{1}=\lambda_{1i}({\mathbf e}^T_i-\lambda_{0i}{\mathbf w}^T_i)\boldsymbol{\Sigma}_0,\] where ${\mathbf e}_i$ is the column vector with its $i$-th component equal to one and all the others equal to zero, while ${\mathbf w}_i$ is the column vector containing the $i$-th row of matrix ${\mathbf W}$. Under general assumptions, ([\[vincolo\]](#vincolo){reference-type="ref" reference="vincolo"}) admits only one solution with respect to $\lambda_{0i}$ and $\lambda_{1i}$ (see Theorem [\[theorem1\]](#theorem1){reference-type="ref" reference="theorem1"}), which can be found among the extreme points of $\lambda_{1i}=\mathop{cor}(z_{it}^{(0)}, z_{i,t-1}^{(0)})$ as a function of $\lambda_{0i}$. To provide insight into this, the first two plots of figure [\[figure2\]](#figure2){reference-type="ref" reference="figure2"} show two examples based on model 1 used in the simulation study. Denote with ($\lambda_{0i}^,\lambda_{1i}^)$ the true values of the coefficients used in model 1 for a given location $i$ (in particular, in figure [\[figure2\]](#figure2){reference-type="ref" reference="figure2"}, the first two plots refer to locations $i=6$ and $i=8$). The solid line shows $\lambda_{1i}=\mathop{cor}(z_{it}^{(0)}, z_{i,t-1}^{(0)})$ as a function of $\lambda_{0i}$. The two dots show the points of this function where the first derivative is zero. The vertical and horizontal dashed lines identify which one of the two points satisfies the sufficient condition in ([\[vincolo\]](#vincolo){reference-type="ref" reference="vincolo"}). As expected, it coincides with the true values ($\lambda_{0i}^,\lambda_{1i}^)$ used to generate the time series. Theorem [\[theorem1\]](#theorem1){reference-type="ref" reference="theorem1"}, shown in the Appendix, formalizes this result. Remark 1: Theorem [\[theorem1\]](#theorem1){reference-type="ref" reference="theorem1"} not only shows that the stationary SDPD model is well identified, because there is a unique solution for $(\boldsymbol{\lambda}_0,\boldsymbol{\lambda}_1)$, but it also suggests a way to estimate such parameters. In fact, we can find all the solutions to equation ([\[nec_cond\]](#nec_cond){reference-type="ref" reference="nec_cond"}) and then check which one satisfies the sufficient condition in ([\[vincolo\]](#vincolo){reference-type="ref" reference="vincolo"}). This estimation procedure is described in section [3](#est_alg){reference-type="ref" reference="est_alg"}. ## Identification of parameters in the case of cross-correlated errors Now, we relax the assumption on the error $\mbox{\boldmath$\varepsilon$}_t$ by letting $\boldsymbol{\Sigma}_0^\varepsilon$ be a full matrix (i.e., there is heteroskedasticity and cross-correlation in the error process). This setup allows the process ${\mathbf y}_t$ to include some spurious cross-correlation not explained by ${\mathbf W}$. In this case, the coefficients $\lambda_{i1}$ still give the correlations between the latent variables $z_{i,t}^{(0)}$ and $z_{i,t-1}^{(0)}$, but now, the $p$ equations in model ([\[b1ter\]](#b1ter){reference-type="ref" reference="b1ter"}) are correlated. The main consequence of this is that the true values $(\lambda_{0i}^, \lambda_{1i}^)$ do not identify an extreme point of the correlation function (see case $i=2$ in figure [\[figure2\]](#figure2){reference-type="ref" reference="figure2"}). Anyway, the sufficient condition in ([\[vincolo\]](#vincolo){reference-type="ref" reference="vincolo"}) is still valid, and the true coefficients $(\lambda_{0i}^, \lambda_{1i}^)$ can be identified by introducing a "constrained" condition. Remark 2: When the errors $\mbox{\boldmath$\varepsilon$}_t$ are not cross-correlated, the matrix $\boldsymbol{\Sigma}_1$ is symmetric by Lemma [\[lemma1\]](#lemma1){reference-type="ref" reference="lemma1"} in the Appendix, so that point 2 in Theorem [\[theorem1bis\]](#theorem1bis){reference-type="ref" reference="theorem1bis"} becomes the same as in Theorem [\[theorem1\]](#theorem1){reference-type="ref" reference="theorem1"}. Therefore, Theorem [\[theorem1bis\]](#theorem1bis){reference-type="ref" reference="theorem1bis"} includes Theorem [\[theorem1\]](#theorem1){reference-type="ref" reference="theorem1"} as a special case. # Estimation procedure {#est_alg} We present here a simple algorithm for the estimation of the parameters $(\lambda_{0i},\lambda_{1i})$ for $i=1,\ldots,p$. First, estimate the matrices $\boldsymbol{\Sigma}_1$ and $\boldsymbol{\Sigma}_0$ through some consistent estimators $\hat\boldsymbol{\Sigma}_0$ and $\hat\boldsymbol{\Sigma}_1$. For example, $\hat\boldsymbol{\Sigma}_0=(n-1)^{-1}{\mathbf Y}_0{\mathbf Y}_0^T$ and $\hat\boldsymbol{\Sigma}_1=(n-2)^{-1}{\mathbf Y}_0{\mathbf Y}_1^T$, where ${\mathbf Y}_l=({\mathbf y}_{1+l}, \cdots, {\mathbf y}_{n-l})$. Alternatively, the threshold estimator analyzed in can be considered in the high dimensional setup. Then, for each location $i=1,\ldots,p$, implement the following steps. 1. Define ${\mathbf e}_i$ as the $i$-th unit vector and ${\mathbf w}^T_i={\mathbf e}^T_i{\mathbf W}$, then compute: \[\begin{aligned} \hat a_{0i} &=& {\mathbf e}_i^T\hat\boldsymbol{\Sigma}_0{\mathbf e}_i, \quad \hat a_{1i} = {\mathbf e}_i^T\hat\boldsymbol{\Sigma}_1 {\mathbf e}_i, \quad \hat a_{2i} = {\mathbf e}_i^T(\hat\boldsymbol{\Sigma}_1^T-\hat\boldsymbol{\Sigma}_1){\mathbf w}_i, \\ \hat b_{0i} &=&-2{\mathbf e}_i^T\hat\boldsymbol{\Sigma}_0{\mathbf w}_i, \quad \hat b_{1i} = -{\mathbf e}^T_i(\hat\boldsymbol{\Sigma}_1+\hat\boldsymbol{\Sigma}_1^T){\mathbf w}_i, \\ \hat c_{0i} &=& {\mathbf w}^T_i\hat\boldsymbol{\Sigma}_0{\mathbf w}_i,\quad \hat c_{1i} = {\mathbf w}^T_i\hat\boldsymbol{\Sigma}_1{\mathbf w}_i. \end{aligned}\] 2. Find the two roots $\lambda_{0i}^{(1)}$ and $\lambda_{0i}^{(2)}$ of the following two-order polynomial equation. \[\label{eqq} \hat t_{0i} + \hat t_{1i}\lambda_{i0} + \hat t_{2i}\lambda_{i0}^2 =0,\] where $\hat t_{0i} = \hat b_{1i}\hat a_{0i}-\hat b_{0i}\hat a_{1i}+\hat a_{0i}\hat a_{2i}$, $\hat t_{1i} = 2(\hat a_{0i}\hat c_{1i}-\hat c_{0i}\hat a_{1i})+\hat a_{2i}\hat b_{0i}$, and $\hat t_{2i} = \hat c_{1i}\hat b_{0i}-\hat c_{0i}\hat b_{1i}+\hat a_{2i}\hat c_{0i}$. 3. Estimate $\lambda_{0i}$ and $\lambda_{1i}$ by \[\begin{aligned} \label{stimatore0} \hat\lambda_{0i}&=&\arg\min_{j=1,2}{\mathbf v}_{ij}^T{\mathbf v}_{ij}, \\ \hat\lambda_{1i}&=&\frac{({\mathbf e}_i^T-\hat\lambda_{0i}{\mathbf w}^T_i)\hat\boldsymbol{\Sigma}_1({\mathbf e}_i-\hat\lambda_{0i}{\mathbf w}_i)}{({\mathbf e}_i^T-\hat\lambda_{0i}{\mathbf w}^T_i)\hat\boldsymbol{\Sigma}_0({\mathbf e}_i-\hat\lambda_{0i}{\mathbf w}_i)}, \label{stimatore1} \end{aligned}\] where ${\mathbf v}^T_{ij} = ({\mathbf e}_i^T-\lambda_{0i}^{(j)}{\mathbf w}^T_i)\hat\boldsymbol{\Sigma}_1-\lambda_{1i}^{(j)}({\mathbf e}_i^T-\lambda_{0i}^{(j)}{\mathbf w}^T_i)\hat\boldsymbol{\Sigma}_0,$, and $\lambda_{1i}^{(j)} = ({\mathbf e}_i^T-\lambda_{0i}^{(j)}{\mathbf w}^T_i)\hat\boldsymbol{\Sigma}_1({\mathbf e}_i-\lambda_{0i}^{(j)}{\mathbf w}_i)/({\mathbf e}_i^T-\lambda_{0i}^{(j)}{\mathbf w}^T_i)\hat\boldsymbol{\Sigma}_0({\mathbf e}_i-\lambda_{0i}^{(j)}{\mathbf w}_i)$. Remark 3: Assumption $A2$ in section [4](#asymptotic){reference-type="ref" reference="asymptotic"} guarantees that matrix $\left[{\mathbf I}_p-D(\boldsymbol{\lambda}_0){\mathbf W}\right]$ has full rank. However, the above estimation procedure may suffer for some locations if matrix $\left[{\mathbf I}_p-D(\boldsymbol{\lambda}_0){\mathbf W}\right]$ is near singularity. Such a case may come about because of the presence of some almost linearly dependent rows in the matrix, which may cause the quantity ${\mathbf w}_i^T[\boldsymbol{\Sigma}_1-\lambda_{1i}\boldsymbol{\Sigma}_0]{\mathbf w}_i$ to be almost zero for those rows (see Lemma [\[lemma2\]](#lemma2){reference-type="ref" reference="lemma2"}). As a result, the procedure loose efficiency for the estimation of $\lambda_{i0}$ for those locations (but it still works for $\lambda_{1i}$). Something similar may happen if there are some (almost) zero rows in ${\mathbf W}$, which is excluded by assumption $A4$. Anyway, it is worthwhile to stress that the estimation procedure works efficiently for all the other locations. In fact, the procedure does not require the inversion of matrix $\left[{\mathbf I}_p-D(\boldsymbol{\lambda}_0){\mathbf W}\right]$, so it is able to isolate and separate the effects of "collinear" locations (or uncorrelated locations) from the other locations and to guarantee consistent and efficient estimations for the last locations. # Theoretical results {#asymptotic} In this section, we show the theoretical foundations of our proposal. In particular, we present the assumptions and show the consistency and the asymptotic normality of the estimators, for the cases of finite dimension and high dimension. Moreover, we show that the stationary SDPD model can be used to represent a wide range of multivariate linear processes with respect to a "latent spatial matrix," and therefore, it is of wider application than classic spatio--temporal contexts. The reduced form of model ([\[b1\]](#b1){reference-type="ref" reference="b1"}) is \[\label{b1bis} {\mathbf y}_t={\mathbf A}^{\mathbf y}_{t-1}+\mbox{\boldmath$\varepsilon$}_t^,\] where $\mbox{\boldmath$\varepsilon$}_t^=\left[{\mathbf I}_p-D(\boldsymbol{\lambda}_0){\mathbf W}\right]^{-1}\mbox{\boldmath$\varepsilon$}_t$ and \[\label{diagonalize} {\mathbf A}^=\left[{\mathbf I}_p-D(\boldsymbol{\lambda}_0){\mathbf W}\right]^{-1}D(\boldsymbol{\lambda}_1)\left[{\mathbf I}_p-D(\boldsymbol{\lambda}_0){\mathbf W}\right].\] Note that the errors $\mbox{\boldmath$\varepsilon$}_t^$ have mean zero and are serially uncorrelated. Model ([\[b1bis\]](#b1bis){reference-type="ref" reference="b1bis"}) has a VAR* representation, so it is stationary when all the eigenvalues of matrix ${\mathbf A}^$ are smaller than one in absolute value. From ([\[diagonalize\]](#diagonalize){reference-type="ref" reference="diagonalize"}), we can note that $\boldsymbol{\lambda}_1$ contains the eigenvalues of ${\mathbf A}^$ whereas $\boldsymbol{\lambda}_0$ only affects its eigenvectors (see the proof of Theorem [\[theorem1\]](#theorem1){reference-type="ref" reference="theorem1"}). Therefore, we must consider the following assumptions: - $\lambda_{1i}\in\mathbb{R}$ and $\|\lambda_{1i}\|<1$, for all $i$, and vector $\boldsymbol{\lambda}_1$ is not scalar; - $\lambda_{0i}\in\mathbb{R}$ for all $i$ and vector $\boldsymbol{\lambda}_0$ is such that matrix $\left[{\mathbf I}_p-D(\boldsymbol{\lambda}_0){\mathbf W}\right]$ has full rank; - the errors $\varepsilon_{it}$ are serially independent and such that $E(\varepsilon_{it})=0$ and $E\|\varepsilon_{it}\|^\delta<\infty$ for all $i,t$, for some $\delta>4$; - the spatial matrix ${\mathbf W}$ is nonsingular and has zero main diagonal. Assumption $A1$ implies stationarity. Moreover, it guarantees that there are at least two distinct values in vector $\boldsymbol{\lambda}_1$ so that model ([\[b1\]](#b1){reference-type="ref" reference="b1"}) is identifiable, as shown in Theorem [\[theorem1\]](#theorem1){reference-type="ref" reference="theorem1"}. Assumption $A2$ is clearly necessary to assure that matrix $\left[{\mathbf I}_p-D(\boldsymbol{\lambda}_0){\mathbf W}\right]$ can be inverted so that the reduced model in ([\[b1bis\]](#b1bis){reference-type="ref" reference="b1bis"}) is well defined (Remark 3 indicates what happens when this assumption is not satisfied). Incidentally, it is worthwhile to note that our setup automatically solves the problem concerning the parameter space of $\boldsymbol{\lambda}_0$, highlighted at the end of section 2.2 by. So, in the empirical applications of our model, it is possible to use any kind of normalization for ${\mathbf W}$ (i.e., row-factor normalization or single-factor normalization), since the vector $\boldsymbol{\lambda}_0$ would automatically rescale accordingly (see section [4.1](#high){reference-type="ref" reference="high"} for more details on this aspect). This means that the coefficients $\lambda_{0i}$ can also take values outside the classic interval $[-1,1]$. Assumption $A3$ assures that the results in can be applied to show the asymptotic normality of the estimators. Assumption $A4$ is classic in spatio--temporal models and guarantees that the model is well defined and identifiable with respect to all the parameters, also for $p\rightarrow\infty$ (see Lemma [\[lemma2\]](#lemma2){reference-type="ref" reference="lemma2"} and Theorem [\[theorem4\]](#theorem4){reference-type="ref" reference="theorem4"}). Under assumptions $A1-A4$, it is immediately evident that the estimators $\hat\boldsymbol{\lambda}_{0}$ and $\hat\boldsymbol{\lambda}_{1}$, presented in section [3](#est_alg){reference-type="ref" reference="est_alg"}, are both consistent following Theorem 11.2.1 in. For asymptotic normality, the following theorem can be stated. Note that the estimators $\widehat\lambda_{ji}$ are unbiased for all $i,j$ and for all $p$. In the high dimension, we have infinite parameters to estimate ($2p$ in total, where $p\rightarrow\infty$). Therefore, we must assure that the consistency of the estimators is still valid in such a case. As expected, the properties of matrix ${\mathbf W}$ influence the consistency and the convergence rates of the estimators $\hat\lambda_{ij}$ when $p\rightarrow\infty$. For example, denote with $k_i$ the number of nonzero elements in vector ${\mathbf w}_i$. If $k_i=O(1)$ as $p\rightarrow\infty$, for all $i$, then the effective dimension of model ([\[b1\]](#b1){reference-type="ref" reference="b1"}) is finite and Theorem [\[theorem3\]](#theorem3){reference-type="ref" reference="theorem3"} can still be applied for the consistency and the asymptotic normality of the estimators $\hat\lambda_{ji}$, even if $p\rightarrow\infty$. The following Theorem [\[theorem4\]](#theorem4){reference-type="ref" reference="theorem4"}, instead, shows the consistency of the estimators under more general vectors ${\mathbf w}_i$, with $k_i\rightarrow\infty$ as $p\rightarrow\infty$. ## Asymptotics for high dimensional setups {#high} In model ([\[b1\]](#b1){reference-type="ref" reference="b1"}), the spatial correlation between a given location $i$-th and the other locations is summarized by $\lambda_{0i}{\mathbf w}_i$. If the vector ${\mathbf w}_i$ is rescaled by a factor $\delta_i$, then we can have an equivalent model by rescaling the spatial coefficient $\lambda_{0i}$ by the inverse of the same factor, since $\lambda_{0i}{\mathbf w}_i=\delta_i^{-1}\lambda_{0i}{\mathbf w}_i\delta_i=\lambda_{0i,\delta}{\mathbf w}_{i,\delta}$. In such a way, we may consider irrelevant a row-normalization of matrix ${\mathbf W}$ if we let the coefficients in $D(\boldsymbol{\lambda}_{0})$ rescale accordingly. Such an approach is not new and follows the idea of. We use this approach here in order to simplify the analysis and the interpretation of the stationary SDPD model in the high dimensional setup. In fact, when $p\rightarrow\infty$ and $k_i=O(p)$, the vectors ${\mathbf w}_i$ may change with $p$ and this may have an influence on the scale order of the process. This happens, for example, if we consider a row-normalized spatial matrix ${\mathbf W}$, since the weights become infinitely small for infinitely large $p$. Looking at the ([\[diagonalize\]](#diagonalize){reference-type="ref" reference="diagonalize"}), model ([\[b1\]](#b1){reference-type="ref" reference="b1"}) appears to become spatially uncorrelated for $p\rightarrow\infty$ because matrix ${\mathbf W}$ tends to be asymptotically diagonal (for $p\rightarrow\infty$ and $T$ given). As a consequence, the model appears to become not identifiable in the high dimension with respect to the parameters $\lambda_{0i}$. To avoid this, we assume here that also the coefficients $\lambda_{0i}$ may depend on the dimension $p$, borrowing the idea of. In such a way, we can derive the conditions for the identifiability of the model in the high dimension and better convergence rates for the estimators. This is shown by the following theorem. As shown by Theorem [\[theorem4\]](#theorem4){reference-type="ref" reference="theorem4"}, cases (i) and (ii), if we consider a row-normalized spatial matrix ${\mathbf W}$, our estimation procedure is consistent for any value of $p$ and with $p\rightarrow\infty$ at any rate. In other words, the convergence rate is not affected by the dimension $p$. However, there are some differences between the two cases of $L_1$ and $L_2$ normalization. In the first case, we need to impose that the spatial coefficients $\lambda_{0i}$ increases in the order $O(p)$ as $p\rightarrow\infty$ (otherwise the model becomes not identifiable in the high dimension), whereas in the last case of $L_2$ norm they can remain constant for $p\rightarrow\infty$. In case (iii), which is more general because it is valid for any ${\mathbf W}$, we need to impose $k_i=o(T^{1/2})$ in order to guarantee the consistency of the estimators. To complete this section, we want to show the class of processes that can be analysed by our stationary SDPD model. Under assumption $A2$, any stationary SDPD model can be equivalently represented as a VAR process as in ([\[b1bis\]](#b1bis){reference-type="ref" reference="b1bis"}), with respect to an autoregressive matrix coefficient ${\mathbf A}^$ defined in ([\[diagonalize\]](#diagonalize){reference-type="ref" reference="diagonalize"}). Now, by exploiting the simple structure of our model, we can show the conditions under which the opposite is true. The following corollary derives from standard results. By corollary [\[corollary1\]](#corollary1){reference-type="ref" reference="corollary1"} and assumptions $A1-A4$, the VAR processes that cannot be represented and consistently estimated by our stationary SDPD* model are those characterized by a matrix ${\mathbf A}^$ with linear dependent eigenvectors (i.e., those with algebraic multiplicities) or those with complex eigenvalues. In order to apply our model to those cases also, we should generalize the estimation procedure using the Jordan decomposition of matrix ${\mathbf A}^$. However, we leave this topic to future study. # Simulation study {#simulazioni} This section contains the results of a simulation study implemented to evaluate the performance of the proposed estimation procedure. In section 5.1, we describe the settings and check the validity of the assumptions for the simulated models. Then, in section 5.2, we evaluate the consistency of the estimation procedure and the convergence rate for the estimators using a known spatial matrix. Finally, section 5.3, we analyse the case when the spatial matrix ${\mathbf W}$ is unknown, and therefore, to be estimated. ## Settings We consider three different spatial matrices. In the first, we randomly generate a matrix of order $p\times p$, and we post-multiply this matrix by its transpose in order to force symmetry. The resulting spatial matrix is denoted with ${\mathbf W}_1$. Note that such a matrix is full, and it may have positive and negative elements. In the other two cases, the spatial matrix is sparse and has only positive entries: ${\mathbf W}_2$ is generated by setting to one only four values in each row while ${\mathbf W}_3$ is generated by setting to one $2\sqrt{p}$ elements in each row. For all three cases, we check the rank to guarantee that the spatial matrix has $p$ linearly independent rows. Moreover, we set to zero the main diagonal, and we rescale the elements so that each row has norm equal to one ($L_2$ row-normalization). For the error process, we generate $p$ independent Gaussian series $e_{ti}$ with mean zero and standard error ${\sigma}_i$, where the values ${\sigma}_i$ are generated randomly from a uniform distribution $U(0.5, 1.5)$ for $i=1,\ldots,p$. Then, we define the cross-correlated error process $\mbox{\boldmath$\varepsilon$}_t=\{\varepsilon_{it},t=1,\ldots,T\}$, where \[\left\{ \begin{array}{ll} \varepsilon_{ti} = e_{ti}-0.7e_{t2} & \mbox{for }i=3,\ldots,p, \\ \varepsilon_{ti} = e_{ti} & \mbox{otherwise}. \\ \end{array} \right.\] We generate all $\lambda_{ji}$ from a uniform distribution $U(-0.7, 0.7)$. The settings above guarantee that assumptions $A1-A4$ hold. We generate different models with dimensions $p = (10, 50, 100, 500)$ and sample sizes $T = (50, 100, 500, 1000)$. Note that we may have $T<<p$. For each configuration of settings, we generate 500 Monte Carlo replications of the model and report the estimation results. All the analyses have been made in R. ## Empirical performance of the estimators when ${\mathbf W}$ is known Figure [\[figure5\]](#figure5){reference-type="ref" reference="figure5"} shows the box plots of the estimations for increasing sample sizes $T = (50, 100, 500, 1000)$ and fixed dimension $p = 100$. The four plots at the top refer to the estimation of $\lambda_{0i}$ while the four plots at the bottom refer to that of $\lambda_{1i}$. Each plot focuses on a different location $i$, where $i = 97,\ldots,100$. The true values of the coefficients $\lambda_{ji}$ are shown through the horizontal lines. Note that we have $T\leq p$ for the first two box plots in each plot, since $p = 100$ for this model. The box plots are centred on the true value of the parameters, and the variance reduces for increasing values of $T$, showing consistency of the estimators and a good performance for small $T$/large $p$ also. To evaluate the estimation error, for each realized time series, we compute the average error $AE$ and the average squared error $ASE$ using the equations below. \[\label{mse} AE(\widehat\boldsymbol{\lambda}_j) = \frac{1}{p}\sum_{i=1}^p{(\hat\lambda_{ji}-\lambda_{ji})}, \qquad ASE(\widehat\boldsymbol{\lambda}_j) = \frac{1}{p}\sum_{i=1}^p{(\hat\lambda_{ji}-\lambda_{ji})^2}, \qquad j=1,2.\] Table [\[tabella1\]](#tabella1){reference-type="ref" reference="tabella1"} reports the mean values of $ASE(\widehat\boldsymbol{\lambda}_j)$ (with the standard deviations in brackets) computed over 500 simulated time series for different values of $T$ and $p$. As shown in the table, the estimation error decreases when the sample size $T$ increases. It is interesting to note that the estimation error does not increase for increasing values of the dimension $p$. This is more evident from figure [\[increasing_p\_global\]](#increasing_p_global){reference-type="ref" reference="increasing_p_global"}, which shows the box plots of the average errors $AE(\boldsymbol{\lambda}_0)$ (at the top) and $AE(\boldsymbol{\lambda}_1)$ (at the bottom) computed for 500 replications of the model, with varying values of $p$, sample sizes $T$, and spatial matrix ${\mathbf W}_1$. We can note from the figure that $\hat\boldsymbol{\lambda}_0$ and $\hat\boldsymbol{\lambda}_1$ are unbiased for all $n$ and $p$. Moreover, the variability of the box plots decreases for $p\rightarrow\infty$ and fixed $T$: this is a consequence of averaging the absolute error over the $p$ locations using equation ([\[mse\]](#mse){reference-type="ref" reference="mse"}). ## Estimation results when the spatial matrix is unknown {#matrixA} In this section, we evaluate the performance of the proposed estimation procedure when the spatial matrix ${\mathbf W}$ is unknown and needs to be estimated. In this case, the estimation error has to be evaluated with respect to matrix ${\mathbf A}^$ in order to include the effects of both $\hat\boldsymbol{\lambda}_{j}$ and $\hat{\mathbf W}$ on the final estimations. So, using ([\[diagonalize\]](#diagonalize){reference-type="ref" reference="diagonalize"}), we define the two estimators \[\begin{aligned} \hat{\mathbf A}_{SDPD}^({\mathbf W}) &=& \left[{\mathbf I}_p-D(\hat\boldsymbol{\lambda}_0){\mathbf W}\right]^{-1}D(\hat\boldsymbol{\lambda}_1)\left[{\mathbf I}_p-D(\hat\boldsymbol{\lambda}_0){\mathbf W}\right] and \label{AW}\\ \hat{\mathbf A}_{SDPD}^(\hat{\mathbf W}) &=& \left[{\mathbf I}_p-D(\hat\boldsymbol{\lambda}_0)\hat{\mathbf W}\right]^{-1}D(\hat\boldsymbol{\lambda}_1)\left[{\mathbf I}_p-D(\hat\boldsymbol{\lambda}_0)\hat{\mathbf W}\right],\label{AWhat} \end{aligned}\] where matrix ${\mathbf W}$ is assumed to be known in the first case and unknown in the second. When ${\mathbf W}$ is unknown, we estimate it by the (row-normalized) correlation matrix at lag zero, but other more efficient estimators of ${\mathbf W}$ can be considered alternatively. For the sake of comparison, remembering the VAR representation of our model in ([\[b1bis\]](#b1bis){reference-type="ref" reference="b1bis"}), we also estimate matrix ${\mathbf A}^$ using the classic Yule--Walker estimator of the VAR model $\hat{\mathbf A}_{VAR}^=\hat\boldsymbol{\Sigma}_0^{-1}\hat\boldsymbol{\Sigma}_1$. To give a measure of the estimation error, we define \[\label{mse2} ASE({\mathbf A}^_{(1)})= \frac{1}{p}\sum_{i=1}^p{(\hat A^_{1i}-A^_{1i})^2},\] where $A^_{1i}$ for $i=1,\ldots,p$ are the true coefficients in the first row of matrix ${\mathbf A}^$, and $\hat A^_{1i}$ are their estimated values. The box plots in figure [\[figure6\]](#figure6){reference-type="ref" reference="figure6"} summarize the results of the estimations from 500 replications of the model with $p=100$ (at the top) and $p=500$ (at the bottom). We report the average squared error computed by ([\[mse2\]](#mse2){reference-type="ref" reference="mse2"}) in three different cases: the classic Yule--Walker estimator of the VAR model $\hat{\mathbf A}_{VAR}^$ on the left, our estimator $\hat{\mathbf A}_{SDPD}^({\mathbf W})$ proposed in ([\[AW\]](#AW){reference-type="ref" reference="AW"}) with the known spatial matrix in the middle, and our estimator $\hat{\mathbf A}_{SDPD}^(\hat{\mathbf W})$ proposed in ([\[AWhat\]](#AWhat){reference-type="ref" reference="AWhat"}) with the estimated spatial matrix on the right. Figure [\[figure6\]](#figure6){reference-type="ref" reference="figure6"} shows interesting results. First, note that the classic estimator $\hat{\mathbf A}_{VAR}^$ cannot be applied when $T\leq p$, and this is a serious drawback of the classic VAR models. On the other hand, the stationary SDPD model is equivalently used to represent the same process but it can always generate an estimation result for all values of $T$ and $p$ regardless of whether ${\mathbf W}$ is known or unknown. Moreover, if we compare the box plots, we can note that both the median and the variability of the estimators $\hat{\mathbf A}_{SDPD}^({\mathbf W})$ and $\hat{\mathbf A}_{SDPD}^(\hat{\mathbf W})$ are remarkably lower than those relative to the classic estimator $\hat{\mathbf A}_{VAR}^$ (when available) for all sample sizes $T$ and dimensions $p$. This deserves a further remark: while it is expected that the estimator $\hat{\mathbf A}_{SDPD}^({\mathbf W})$ performs better than $\hat{\mathbf A}_{VAR}^$ (given that it exploits the knowledge of the true spatial matrix ${\mathbf W}$), it is surprising to also see that the estimator $\hat{\mathbf A}_{SDPD}^(\hat{\mathbf W})$ outperforms the classic estimator $\hat{\mathbf A}_{VAR}^$, notwithstanding the fact that they function under the same conditions (only the time series ${\mathbf y}_t$ is observed and no spatial matrix is known). Of course, the ASE of the estimator $\hat{\mathbf A}_{SDPD}^(\hat{\mathbf W})$ slightly increases compared to that of the estimator $\hat{\mathbf A}_{SDPD}^*(\hat{\mathbf W})$, but its variability remains more or less the same.	{'timestamp': '2016-07-18T02:07:32', 'yymm': '1607', 'arxiv_id': '1607.04522', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04522'}
null	null
# Introduction The purpose of this paper is to provide a convenient operadic framework for the cumulants of free probability theory. In , the author and his collaborators described an operadic framework for so-called Boolean and classical cumulants. In those papers, the fundamental object of study is an algebra $A$ equipped with a linear map $E$, called expectation, to some fixed algebra $B$. The expectation is not assumed to be an algebra homomorphism; rather one measures the degree to which $E$ fails to be an algebra homomorphism with a sequence of multilinear maps $\kappa_n$ from powers of $A$ to $B$, called cumulants. The cumulants, in many cases, can be defined recursively in terms of the expectation map via a formula of the form: \[\label{outline of cumulants} E(x_1\cdots x_n)= \sum \kappa_{i_1}(\cdots)\cdots \kappa_{i_k}(\cdots).\] Depending on what kind of probability theory is under consideration, the summation on the left may be over a different index set. See, e.g., . In , these recursive definitions for the collection of cumulants (in the Boolean and classical regimes, respectively) were reinterpreted as the collection of linear maps determining a coalgebraic map into a cofree object. In the Boolean case, the cofree object is the tensor coalgebra. In the classical case it is the symmetric coalgebra. This reformulation is intended as the background for a homotopical enrichment of probability theory; adding a grading, a filtration, and a differential to this coalgebraic picture leads to a rich theory with applications to quantum field theory . This application is motivational and will play no role in this paper. None of the work mentioned above treats the case of free cumulants, arguably the most important kind of cumulant in noncommutative probability theory. When the target algebra $B$ is commutative, there is a formula similar to those above and the framework outlined above can be used directly, employing a more exotic type of coalgebra than the tensor or symmetric coalgebra. This point of view is taken in . However, there is a flaw in this point of view, which is that assuming the target to be commutative is external to the theory; internally it makes perfect sense for the target itself to be noncommutative. This is called operator-valued free probability theory because the expectation is valued in a noncommutative algebra, such as an operator algebra. Operator-valued free cumulants, as defined by Speicher, are somewhat more cumbersome to describe explicitly than in the commutative case using classical combinatorial methods. Consequently Speicher develops an operator-valued $R$-transform to collect the information concisely. In our setting, there is one evident related obstruction to extending the framework developed in to operator-valued free cumulants. The defining formulas for classical and Boolean cumulants and for free cumulants valued in a commutative algebra share a certain property. Namely, they are string-like, meaning that the right-hand side of Equation ([\[outline of cumulants\]](#outline of cumulants){reference-type="ref" reference="outline of cumulants"}) is a product of cumulants. However, the defining formulas for operator-valued free cumulants (that is, free cumulants valued in a not necessarily commutative algebra) contain terms like \[\kappa_2(x_1\kappa_1(x_2)\otimes x_3).\] or more generally \[\kappa_{n_1}(x_1\kappa_{n_2}(x_2\kappa_{n_3}(\cdots)\kappa_{n_4}(\cdots),\cdots),\cdots).\] In a word, they are not string-like but tree-like. This is precisely the issue that leads Speicher to develop the operator-valued $R$-transform. Here, this tool is avoided by using an operadic reformulation. Tree-like formulas can be obtained by passing from algebras and coalgebras, which have a string-like structure, to nonsymmetric operads and cooperads, which have a tree-like structure. The main result of this paper shows how the relationship between the moments and free cumulants, realized as cooperadic maps $M$ and $K$, is encapsulated quite simply in terms of a canonical twist: \[M=\Phi\circ K.\] As phrased in this paper, the moments and cumulants are defined in some other manner and this is a theorem, but it is probably better to consider this as an alternative definition which is quite simple from the operadic viewpoint. This reformulation is part of a campaign to explore applications of the operadic language in probability theory; the result contained herein is modest and is intended to serve as further advertisment and evidence (following ) of potentially deeper connections between the two areas. It is possible that both this reformulation and those attempted in the author's previous work (cited above) are reflections of a combinatorial relationship between operads and Möbius inversion with respect to a poset. This is not pursued further here, but see for some discussion and further references on this topic. The remainder of the paper is organized as follows. Section [2](#sec: operads){reference-type="ref" reference="sec: operads"} reviews the parts of operadic theory that are used in the paper. Section [3](#sec: partitions and trees){reference-type="ref" reference="sec: partitions and trees"} goes over the combinatorics of non-crossing partitions, and Section [4](#sec:freeprob){reference-type="ref" reference="sec:freeprob"} applies this to define free cumulants. Finally, Section [5](#sec: main result){reference-type="ref" reference="sec: main result"} states and proves the reformulation of free cumulants in operadic terms. ## Conventions Everything linear occurs over a fixed ground ring. Algebras are generally not assumed to be commutative or unital. Every finite ordered set is canonically isomorphic to $[n]\coloneqq\{1,\ldots, n\}$, and this canonical isomorphism will be routinely abused. A graph is a finite set of vertices, a finite set of half-edges, a source map from half-edges to vertices, and an involution on the half-edges; a half-edge is a leaf it is a fixed point of the involution. A graph is connected if every two vertices can be joined by a path of half-edges connected by having the same source or via the involution. A connected graph is a tree if it has more vertices than edges. A root is a choice of leaf of a tree (this is no longer considered a leaf). The root of a vertex is the unique half-edge "closest" to the overall root. The root vertex is the unique vertex whose root is the overall root. A planar tree has a cyclic order on the half-edges of each vertex. # Operads and cooperads {#sec: operads} Aside from some minor changes, conventions of are used for operadic algebra. This section reviews standard definitions (more details can be seen in ). In this paper everything will be nonsymmetric and the adjective will be omitted. Both the cofree and free functor on $M$ can be realized at the collection level as the collection of rooted planar trees with vertices decorated by elements of $M$, denoted $\mathcal{T}(M)$. This implies the following. 1. Fix an operad $\mathbf{P}$ (with underlying collection $P$) and an element of $\mathcal{T}(M)$. That is, take a planar rooted tree $T$ and an element of $\mathbf{P}(n)$ for every vertex of $T$ of valence $n+1$ (collectively called a decoration of $T$ by $\mathbf{P}$). Then there is a canonical element of $\mathbf{P}$ called the composition of the decoration induced by the counit of the forgetful free adjunction $\mathcal{F}(P) \item Dually, given a cooperad\[with underlying collection$C$, the unit of the cofree forgetful adjunction$ 2. There is a canonical linear isomorphism $\psi$ between the free operad on a collection and the conilpotent cofree cooperad on the same collection. This is in precise parallel to the situation with power series, where a power series is invertible if and only if its constant term is invertible. # Partitions and trees {#sec: partitions and trees} Noncrossing partitions are important in combinatorics and there are many bi-indexed sets of combinatorial objects in canonical bijection with them. For our purposes, the following such bijection will be useful. See Figure [\[figure: partitions trees\]](#figure: partitions trees){reference-type="ref" reference="figure: partitions trees"}. Henceforth partitions will be freely identified with the corresponding trees. It will be useful later to modify this construction. See Figure [\[fig: messy tree\]](#fig: messy tree){reference-type="ref" reference="fig: messy tree"}. # Free probability and operator-valued free cumulants {#sec:freeprob} Classically $B$ is the ground field but for a general theory it is necessary to allow more general algebras, in particular, non-commutative algebras. To be precise, a $B$-algebra is a $B$-bimodule $A$ equipped with a product $A\otimes_B A\to A$ and a $B$-linear map $\eta:B\to A$ which respects the product structure. Let $A$ be a $B$-valued probability space and let $f:A^{\otimes_B n}\to B$ be a $B$-multilinear map. For an $(n+1)$-tuple $(i_0,\ldots, i_n)$ of non-negative integers with sum $N$, define a map \]f_{i_0,\ldots, i_n}: Hom(A^{\otimes n}, Hom(B^{\otimes N}, B))\[ whose evaluation on $a_1\otimes a_n$ is given by the composition \]B^{\otimes N}\to B^{\otimes i_0}\otimes A\otimes B^{\otimes i_1}\otimes A\otimes\cdots\otimes A\otimes B^{\otimes i_n} \to A^{\otimes_B n}\[ Where the first map inserts $a_j$ in the $j$th $A$ place and the second map is given by repeated use of the $B$-bimodule structure on $A$. The map $f_{i_0,\ldots, i_n}(a_1,\ldots, a_n)$ can be realized as the composition in $\End B$ along a decoration of the tree $\pi_{\min\{i_0,1\},\ldots, \min\{i_n,1\}}$, where $\pi$ is the trivial partition of $[n]$. Decorate "top" vertices and the root, should it exist, with the product in $B$ and decorate the single vertex corresponding to the single block of $\pi$ with $f_{\min\{i_0,1\},\ldots, \min\{i_n,1\}}(a_1,\ldots, a_n)$. The following definition is Definition 2.1.1 of , restricted to $B$-algebras. It has been reworded to use operadic language. See Figure [\[figure: multiplicative function\]](#figure: multiplicative function){reference-type="ref" reference="figure: multiplicative function"}. The following definition combines Example 1.2.2, Definition 2.1.6, and Proposition 3.2.3 of . It will be useful in the next section to record a version of this defining relationship viewed in $\End B$. The following is a direct application of the definitions. # Main result {#sec: main result} Let $V=\langle \rangle$ be a one-dimensional free module. There is an explicit presentation of the sum of the modules of the underlying collection of $\textsf{coAs}_A$ as $\bigoplus_{n=1}^\infty (A\oplus V)^{\otimes n}$. Here the arity $n$ module consists of those elements that are degree $n$ in the generator $$ of $V$. In this presentation, the cocomposition map is given as follows. Let $w$ be a word in $$ and $A$; let $F(W)$ be the the set of all ways of writing $w$ as the concatenation $b_0a_1b_1\cdots a_nb_n$ where - The words $b_i$ are (possibly empty) words in $A$ - The words $a_i$ are nonempty words in $$ and $A$. Then \]\Delta w =\sum_{F(W)}(b_0b_1\cdots b_n)\circ (a_1\otimes\cdots \otimes a_n).\[ The projection map to $\textsf{coAs}$ is given by projecting to $\bigoplus V^{\otimes n}$ and identifying $^{\otimes n}$ with $1$ in the ground ring. The projection map to $A$ is given by projecting to ${A\oplus V}^{\otimes 1}$, identifying $A$ with itself and $*$ with the image of $I$. See Figure [\[fig: moment cumulant diagram\]](#fig: moment cumulant diagram){reference-type="ref" reference="fig: moment cumulant diagram"}. ## Concluding remarks {#concluding-remarks .unnumbered}	{'timestamp': '2016-07-19T02:09:39', 'yymm': '1607', 'arxiv_id': '1607.04933', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04933'}
# Introduction The Navier-Stokes equations (NSEs) represent a formulation of the Newton's laws of motion for a continuous distribution of matter in a fluid state, characterized by an inability to support shear stresses, see. The NSEs allow to determine the velocity field and the pressure of fluids confined in regions of the space, and they are used to describe many different physics phenomena as weather, water flow in tubes, ocean currents and others. Moreover, these equations are useful in several fields of knowledge such as petroleum industry, plasma physics, meteorology, thermo-hydraulics, among others (see for instance). Due to this fact, these equations have been attracted to the attention of several mathematicians since they play an important role for applications. See and the references therein. On the other hand, the theory of impulsive dynamical systems has been shown to be a powerful tool to model real-world problems in physics, technology, biology, among others. Because of this fact, the interest in the study of impulsive dynamical systems has increasing considerably. For recents trends on this subject we indicate the works and the references therein. However, the study of Navier-Stokes equations with impulse effects is really scarce. Motivated by this fact, in this paper, we investigate existence and uniqueness of mild solutions for the impulsive NSEs \[\label{Eq5} \displaystyle\left\{\begin{array}{ll} \displaystyle\frac{\partial u}{\partial t} + q(t)(u \cdot \nabla)u-\nu\Delta u +\nabla p = \phi(t,u), & (t,x) \in \left((0, +\infty)\setminus \displaystyle\bigcup_{k=1}^{+\infty}\{t_k\}\right) \times \Omega, \vspace{1mm}\\ {\rm div}\, u = 0, & (t,x) \in (0, +\infty) \times \Omega, \vspace{1mm}\\ u = 0, & (t,x) \in (0, +\infty) \times \partial\Omega, \vspace{1mm}\\ u(0, \cdot)= u_0 & x \in \Omega, \vspace{1mm}\\ u(t_k^+, \cdot)-u(t_k^-, \cdot) = I_k (u(t_k, \cdot)), & x\in\Omega, \; k=1, 2,\ldots, \end{array} \right.\] where $\Omega$ is a bounded smooth domain in $\mathbb{R}^2$. Here $u = (u_1,u_2)$ denotes the velocity field of a fluid filling $\Omega$, $p$ is its scalar pressure and $\nu > 0$ is its viscosity. We will assume that $q$ is a bounded function, $\phi$ is a nonlinearity which will be specified later, $\{t_k\}_{k \in \mathbb{N}} \subset (0, +\infty)$ is a sequence of impulse times such that $\displaystyle\lim_{t \rightarrow +\infty} t_k = +\infty$, $u(t_k, \cdot) = u(t_k^+, \cdot) = \displaystyle\lim_{\delta \rightarrow 0+}u(t_k + \delta, \cdot)$, $u(t_k^-, \cdot) = \displaystyle\lim_{\delta \rightarrow 0+}u(t_k-\delta, \cdot)$ and $I_k$, $k \in \mathbb{N}$, are the impulse operators. Besides to impulsive actions in the system [\[Eq5\]](#Eq5){reference-type="eqref" reference="Eq5"}, we also allow that the external force $\phi$ is not continuous and depends on the solution $u$. We point out that the Navier-Stokes equations with impulses make sense physically and allow to describe more precisely the phenomena modeled by these equations, since $u$ represents the velocity of the field of a fluid and moreover, the external force $\phi$ in this case does not need to be continuous. It is well known that the phenomena which occur in the environment have impulsive behavior and the functions which model them have several discontinuities. Therefore, with this impulsive model, we intend to give a more precisely description of the Navier-Stokes equations. The system [\[Eq5\]](#Eq5){reference-type="eqref" reference="Eq5"} without impulse conditions was studied in the classical monograph, where $\phi$ is a function of time $t\in\mathbb{R}$. More precisely, the author studies existence and uniqueness of global mild solutions for the non-impulsive equation \[\displaystyle\frac{\partial u}{\partial t} + q(t)(u \cdot \nabla)u-\nu\Delta u +\nabla p = \phi(t),\] subject to the conditions $\textrm{div}\, u = 0$ and $u\|_{\partial \Omega} = 0$, where $\Omega$ is a bounded smooth domain in $\mathbb{R}^2$. Our goal here is to write a weaker formulation of the system [\[Eq5\]](#Eq5){reference-type="eqref" reference="Eq5"} and then, we intend to investigate the existence and uniqueness of mild solutions. In order to do this, we start by considering some notations which can be found in and, for instance. Let $\mathbb{L}^2 (\Omega) = (L^2 (\Omega))^2$ and $\mathbb{H}_0^1 (\Omega) = (H_0^1 (\Omega))^2$ endowed, respectively, with the inner products \[(u,v) = \displaystyle \sum_{j=1}^2 \int_{\Omega} u_j \cdot v_j \ dx, \ \ \ u = (u_1, u_2), \ v = (v_1, v_2) \in \mathbb L^2 (\Omega),\] and \[((u,v)) = \displaystyle \sum_{j=1}^2 \int_{\Omega} \nabla u_j \cdot \nabla v_j dx, \ \ \ u = (u_1, u_2), \ v = (v_1, v_2) \in \mathbb{H}^1_0 (\Omega)\] and norms $\| \cdot \| = ( \cdot, \cdot)^{1/2}$ and $\\| \cdot \\| = (( \cdot, \cdot))^{1/2}$. Now, we consider the following sets: \[\mathcal{E} = \{ v \in (C_0^{\infty}(\Omega))^2: \; \nabla \cdot v = 0 \ \textrm{in} \ \Omega\},\] \[V = \textrm{closure of} \ \mathcal{E} \ \textrm{in} \ \mathbb{H}_0^1 (\Omega)\] and \[H = \textrm{closure of} \ \mathcal{E} \ \textrm{in} \ \mathbb{L}^2 (\Omega).\] The space $H$ is a Hilbert space with the scalar product $(\cdot, \cdot)$ induced by $\mathbb{L}^2 (\Omega)$ and the space $V$ is a Hilbert space with the scalar product $((u, v))$ induced by $\mathbb{H}_0^1 (\Omega)$. The space $V$ is contained in $H$, it is dense in $H$ and by the Poincare's Inequality, the inclusion $i:V \hookrightarrow H$ is continuous. Denote by $V'$ and $H'$ the dual spaces of $V$ and $H$, respectively. The adjoint operator $i^$ is linear and continuous from $H'$ to $V'$, $i^(H')$ is dense in $V'$ and $i^$ is one to one since $i(V) = V$ is dense in $H$. Moreover, by the Riesz representation Theorem, we can identify $H$ and $H'$ and write \[V \subset H \equiv H' \subset V',\] where each space is dense in the following one and the injections are continuous. As a consequence of the previous identifications, the scalar product in $H$, $(f,u)$, of $f \in H$ and $u \in V$ is the same as the duality product between $V'$ and $V$, $\langle f, u \rangle$, i.e., \[\langle f, u \rangle = (f, u), \ \ \ \text{for all} \; f \in H \; \text{and} \; \text{for all} \; u \in V.\] Also, for each $u \in V$, the form \[v \in V \mapsto \nu((u, v)) \in \mathbb R\] is linear and continuous on $V$. Therefore, there exists an element of $V'$ which we denote by $Au$ such that \[\langle Au, v \rangle = \nu((u, v)), \; \text{for all} \; v \in V.\] Notice that the mapping $u \mapsto Au$ is linear, continuous and it is an isomorphism from $V$ to $V'$. Based on it, we consider the following weak formulation of [\[Eq5\]](#Eq5){reference-type="eqref" reference="Eq5"}: \[\label{weak-Eq} \left\{ \begin{array}{lll} \displaystyle\frac{d}{dt} (u, v) + \nu ((u, v)) + b(t)(u,u,v) = \langle \phi(t,u), v \rangle, \ \ v \in V, \; t > 0, \; t \neq t_k, \vspace{1mm}\\ u(t_k)-u(t_k^-) = I_k(u(t_k^-)), \ \ k \in \mathbb N, \vspace{1mm}\\ u(0) = u_0 \in H, \end{array} \right.\] where $\phi(t,u) \in V'$ and $b(t): V \times V \times V \to \mathbb R$ is given by \[b(t) (u, v, w) = q(t) \displaystyle\sum_{i, j = 1}^2 \displaystyle\int_{\Omega} u_i \displaystyle\frac{\partial v_j}{\partial x_i} w_j dx.\] The weak formulation [\[weak-Eq\]](#weak-Eq){reference-type="eqref" reference="weak-Eq"} is equivalent to the impulsive system \[\label{IS} \left\{ \begin{array}{lll} u'+ Au + B(t)(u, u) = \phi, \ \ \ \textrm{in} \ \ V', \ \ t > 0, \; t \neq t_k,\vspace{1mm}\\ u(t_k)-u(t_k^-) = I_k(u(t_k^-)), \ \ k \in \mathbb{N}, \vspace{1mm}\\ u(0) = u_0 \in H, \end{array} \right.\] where $u'= du/dt$, $A: V \to V'$ is the Stokes operator defined by \[\langle Au, v \rangle = \nu((u,v)), \; \text{for all} \; u, v \in V,\] and $B(t): V \times V \to V'$ is a bilinear operator defined by \[\langle B(t)(u, v), w \rangle = b(t) (u, v, w), \; \text{for all} \; u, v, w \in V.\] In Section 2, we consider the following general impulsive system \[\label{IntroNS1} \left\{ \begin{array}{lll} u' + Au + B(\sigma(\cdot,\omega))(u,u) = f(\cdot, \sigma(\cdot,\omega), u), \quad t > 0, \; t \in I, \; t \neq t_k, \; k\in\mathbb{N}, \vspace{1mm}\\ u(t_k)-u(t_k^-) = I_k (u(t_k^-)), \ \ k \in \mathbb{N},\vspace{1mm}\\ u(0) = u_0 \in H, \end{array} \right.\] where $f: I\times\mathcal{M}\times H\rightarrow H$ is a piecewise continuous function with respect to $t\in\mathbb{R}$, non-stationary and also depends on the solution $u$. All the conditions of system [\[IntroNS1\]](#IntroNS1){reference-type="eqref" reference="IntroNS1"} will be specified later. We prove the existence and uniqueness of global mild solutions for the system [\[IntroNS1\]](#IntroNS1){reference-type="eqref" reference="IntroNS1"} when $\mathcal{M}$ is compact, see Theorems [\[T1.1\]](#T1.1){reference-type="ref" reference="T1.1"} and [\[T1.2\]](#T1.2){reference-type="ref" reference="T1.2"}. The case when $\mathcal{M}$ is not compact is considered in Theorem [\[T1.3\]](#T1.3){reference-type="ref" reference="T1.3"}. In Section 3, we prove existence and uniqueness of global mild solutions for the 2D NSEs with impulses [\[Eq5\]](#Eq5){reference-type="eqref" reference="Eq5"} via system [\[IntroNS1\]](#IntroNS1){reference-type="eqref" reference="IntroNS1"}. We also give sufficient conditions to obtain dissipativity for the system [\[Eq5\]](#Eq5){reference-type="eqref" reference="Eq5"}. All the results from this paper hold for the non-impulsive case. # Preliminaries Let $(\mathcal{M},d)$ be a compact metric space and $(\mathcal{M}, \mathbb R, \sigma)$ be a dynamical system on $\mathcal{M}$, i.e., $\sigma:\mathbb{R} \times \mathcal{M} \to \mathcal{M}$ is a continuous mapping which satisfies the following properties: - $\sigma(0, \omega)= \omega$, $\omega \in \mathcal{M}$; - $\sigma(s,\sigma(t,\omega))=\sigma(s+t,\omega)$, $t,s \in \mathbb{R}$, $\omega \in \mathcal{M}$. Let $H$ be a real or complex Hilbert space and $\mathcal{L}(H) = \{T: H \rightarrow H: \, T \; \text{is linear }$ $\text{and bounded}\}$ equipped with the operational norm. Let $A:D(A) \subset H \to H$ be a self-adjoint operator such that \[\label{Coer} {\rm Re}\, \langle Au, u \rangle \geq a \|u\|^2_H,\] for all $u\in D(A)$ and $a > 0$. It follows by that $\mathbb{C}\setminus (-\infty, a]\subset \rho(A)$ (resolvent of $A$), and there exists a constant $M\geq 1$ such that \[\\|(\lambda-A)^{-1}\\|_{\mathcal{L}(H)} \le \frac{M}{\|\lambda-a \|}, \quad \lambda \in \Sigma_{a,\varphi}, \, \lambda \neq a,\] where $\Sigma_{a,\varphi}=\{\lambda \in \mathbb{C}: \|{\rm arg}\, (\lambda-a)\| \le \varphi\}$, $\varphi < \pi$. In particular, $A$ is sectorial. It follows that $-A$ generates an analytic semigroup $\{e^{-At}: t\geq 0\} \subset \mathcal{L}(H)$ which satisfies \[\label{EQ0} \\|e^{-At}\\|_{\mathcal{L}(H)} \le K e^{-at},\] for some constant $K > 0$, where $a > 0$ comes from [\[Coer\]](#Coer){reference-type="eqref" reference="Coer"}. Assuming that $0 \in \rho(A)$, we consider the scale of Hilbert spaces $X^\alpha=D(A^\alpha)$ of fractional power of the operator $A$ endowed with the norm $\\|\cdot \\|_{X^\alpha}= \\|A^\alpha \cdot \\|_H$ ($X^0=H$). If $\beta > \alpha \geq 0$, it is well known that $X^\beta$ is a dense subspace of $X^\alpha$ with continuous inclusion and \[\label{eq:fractpow} \\|e^{-At}\\|_{\mathcal{L}(X^\alpha, X^\beta)} \le C_{\alpha,\beta} \, t^{\alpha-\beta} e^{-at}.\] Consider $F$ a Hilbert space such that $H \subset F$ with inclusion dense and continuous. Denote by $\mathcal{L}(F, H)=\{T: F \rightarrow H: \, T \; \text{is linear and bounded}\}$ equipped with the operational norm. We will assume that the semigroup $\{e^{-At}: t\geq 0\}$ satisfies: 1. $e^{-At} \in \mathcal{L}(F, H)$, for all $t >0$; 2. There exists $0 \le \alpha_1 <1$, such that \[\label{eq:sg-est-F} \\|e^{-At} \\|_{\mathcal{L}(F, H)} \le K_1 t^{-\alpha_1} e^{-at}, \quad \text{for all}\; t >0.\] We denote by $\mathcal{L}^2 (H, F)$ the space of all continuous bilinear operators $\mathcal{B}: H \times H \to F$ equipped with the norm \[\\|\mathcal{B}\\|_{\mathcal{L}^2 (H, F)} = \sup \{ \|\mathcal{B}(u, v)\|_F: \|u\|_H \leq 1, \ \|v\|_H \leq 1 \}.\] Let $C(\mathcal{M}, \mathcal{L}^2(H, F))$ be the space of all continuous mapping $B: \mathcal{M} \to \mathcal{L}^2 (H, F)$ endowed with the norm \[\\|B\\|_{\infty} = \sup_{\omega \in \mathcal{M}} \\|B(\omega)\\|_{\mathcal{L}^2 (H, F)}.\] The space $\Big( C(\mathcal{M}, \mathcal{L}^2(H, F)), \\|\cdot\\|_{\infty} \Big)$ is a Banach space, see. For all $u, v \in H$ and $\omega \in \mathcal{M}$, we have that \[\label{EQ5} \|B(\omega)(u, u)-B(\omega)(v, v)\|_F \leq \\|B\\|_{\infty}(\|u\|_H + \|v\|_H ) \|u-v\|_H,\] and also \[\label{EQ51} \|B(\omega)(u,u)\|_F \leq \\|B\\|_{\infty}\|u\|_H^2.\] Let $\{t_k\}_{k \in \mathbb{N}}$ be a strictly increasing sequence in $(0, +\infty)$ such that $\displaystyle\lim_{t \rightarrow +\infty} t_k = +\infty$. Let $I \subset \mathbb{R}$ be an interval and $f: I\times \mathcal{M}\times H \rightarrow H$ and $I_k: H \rightarrow H$, $k \in \mathbb{N}$, be functions satisfying the following conditions: 1. For each fixed $t \in I$, $f(t, \cdot, \cdot)$ is continuous on $\mathcal{M}\times H$. 2. Let $\omega \in \mathcal{M}$ and $u \in H$. Then $\displaystyle\lim_{s \rightarrow t} f(s, \omega, u) = f(t, \omega, u)$ for all $t \neq t_k$, $k \in \mathbb{N}$, the limit $\displaystyle\lim_{s \rightarrow t_k-} f(s, \omega, u)$ exists and $\displaystyle\lim_{s \rightarrow t_k+} f(s, \omega, u) = f(t_k, \omega,u)$, for all $k \in \mathbb{N}$. 3. There is a bounded function $M: \mathbb R \to \mathbb R_+$, such that for any interval $[a,b] \subset I$, we have \[\displaystyle \int_a^b \|\phi(s)\| \| f(s, \omega, u)\|_H ds \leq \displaystyle\int_a^b M(s) \|\phi(s)\| ds\] for all $\phi \in L^1 [a,b]$, $\omega \in \mathcal{M}$ and $u \in H$. 4. There is a bounded function $L: \mathbb{R} \rightarrow \mathbb{R}_+$, such that for any interval $[a, b] \subset I$, we have \[\int_a^b \|\phi(s)\| \|f(s, \omega_1, u_1)-f(s, \omega_2, u_2)\|_H \, ds \leq \int_a^b L(s)\|\phi(s)\|(d(\omega_1, \omega_2) + \|u_1-u_2\|_H)ds\] for all $\phi \in L^1[a, b]$, $\omega_1, \omega_2 \in \mathcal{M}$ and $u_1, u_2 \in H$. 5. There exists a constant $K_2>0$ such that \[\sup_{k\in\mathbb{N}}\sup_{u\in H}\|I_k(u)\|_H \le K_2.\] 6. There exists a constant $K_{3}>0$ such that \[\|I_k(u)-I_k(v)\|_H \le K_3 \|u-v\|_{H}\] for all $u,v \in H$ and for all $k\in\mathbb{N}$. Now, given $\omega \in \mathcal{M}$ and assuming all the conditions above, we consider the following impulsive system in the state space $H$: \[\label{NS1} \left\{ \begin{array}{lll} u' + Au + B(\sigma(\cdot,\omega))(u,u) = f(\cdot, \sigma(\cdot,\omega), u), \quad t > 0, \; t \in I, \; t \neq t_k, \; k\in\mathbb{N}, \vspace{1mm}\\ u(t_k)-u(t_k^-) = I_k (u(t_k^-)), \ \ k \in \mathbb{N},\vspace{1mm}\\ u(0) = u_0 \in H. \end{array} \right.\] Given $T>0$, we consider the space $PC^+ ([0,T], H)= \{u:[0, T]\to H: \ u \text{ is continuous at }\linebreak t \neq t_k, \text{ right-continuous at } t=t_k \text{ and the limit } \displaystyle \lim_{t \to t_k^-} u(t) \;\text{exists for all} \, k = 1, \ldots, n_T \}$. It is well known that the space $PC^+ ([0,T], H)$ endowed with the norm $\displaystyle \\|u\\|_{PC^+}= \sup_{t \in [0,T]} \|u(t)\|_{H}$ is a Banach space. In the sequel, we present the definition of a mild solution for the system [\[NS1\]](#NS1){reference-type="eqref" reference="NS1"}. Given $K \subset H$, we consider the following space of functions: \[\begin{aligned} P&C^{+}_1([0,T]\times {K} \times \mathcal{M}, H) =\{ \varphi:[0,T]\times {K}\times \mathcal{M} \to H: \, \text{ for all } (u,\omega) \in {K} \times \mathcal{M}, \\ & \varphi(\cdot,u,\omega) \in PC^+( [0, T], H) \text{ and for all } t \in [0,T], \, \varphi(t,\cdot,\cdot): K \times \mathcal{M} \to H \text{ is continuous} \}. \end{aligned}\] Theorem [\[T1.1\]](#T1.1){reference-type="ref" reference="T1.1"} ensures that the nonautonomous system [\[NS1\]](#NS1){reference-type="eqref" reference="NS1"} admits a unique mild solution in the sense of Definition [\[mild-solution\]](#mild-solution){reference-type="ref" reference="mild-solution"}. Theorem [\[T1.2\]](#T1.2){reference-type="ref" reference="T1.2"} gives sufficient conditions for the mild solution of system [\[NS1\]](#NS1){reference-type="eqref" reference="NS1"} to be prolongated on $\mathbb{R}_+$. By following the proofs of Theorems [\[T1.1\]](#T1.1){reference-type="ref" reference="T1.1"} and [\[T1.2\]](#T1.2){reference-type="ref" reference="T1.2"}, we can state the next result which deals with existence and uniqueness of global mild solutions for the system [\[NS1\]](#NS1){reference-type="eqref" reference="NS1"} when $\mathcal{M}$ is not necessarily compact. # The 2D Navier-Stokes equations with impulses In this section, we present conditions to obtain the existence and uniqueness of global mild solutions for the following 2D Navier-Stokes equations with impulses \[\label{l1} \displaystyle\left\{\begin{array}{ll} \displaystyle\frac{\partial u}{\partial t} + q(t)(u \cdot \nabla)u-\nu\Delta u +\nabla p = \phi(t, u), & (t,x) \in \left((0, +\infty)\setminus \displaystyle\bigcup_{k=1}^{+\infty}\{t_k\}\right) \times \Omega, \vspace{1mm}\\ {\rm div}\, u = 0, & (t,x) \in (0, +\infty) \times \Omega, \vspace{1mm}\\ u = 0, & (t,x) \in (0, +\infty) \times \partial\Omega, \vspace{1mm}\\ u(0, \cdot)= u_0(\cdot) & x \in \Omega, \vspace{1mm}\\ u(t_k, \cdot)-u(t_k^-, \cdot) = I_k(u(t_k^-, \cdot)), & x\in\Omega, \; k=1, 2,\ldots, \end{array} \right.\] where $\Omega$ is an open and bounded set in $\mathbb R^2$ with $\partial \Omega \in C^2$, $u = (u_1,u_2)$ is the velocity field of a fluid, $p$ is the scalar pressure, $\nu > 0$ is the kinematic viscosity of the fluid, $\phi=\phi(t, u) \in \mathbb R^2$ is the external body force, $q(t)$ is a bounded function, $\{t_k\}_{k \in \mathbb{N}} \subset (0, +\infty)$ is a sequence of impulses such that $\displaystyle\lim_{t \rightarrow +\infty} t_k = +\infty$ and $I_k$ is the impulse operator for each $k \in \mathbb N$. Let \[\mathcal{E} = \{u \in (C_0^{\infty}(\Omega))^2: \; \nabla \cdot u = 0 \ \textrm{in} \ \Omega\},\] \[V = \textrm{closure of} \ \mathcal{E} \ \textrm{in} \ \mathbb{H}_0^1 (\Omega)\] and \[H = \textrm{closure of} \ \mathcal{E} \ \textrm{in} \ \mathbb{L}^2 (\Omega),\] where $\mathbb{L}^2 (\Omega) = (L^2 (\Omega))^2$ and $\mathbb{H}_0^1 (\Omega) = (H_0^1 (\Omega))^2$ are endowed, respectively, with the inner products \[(u,v) = \displaystyle \sum_{j=1}^2 \int_{\Omega} u_j \cdot v_j \ dx, \ \ \ u = (u_1, u_2), \ v = (v_1, v_2) \in \mathbb L^2 (\Omega),\] and \[((u,v)) = \displaystyle \sum_{j=1}^2 \int_{\Omega} \nabla u_j \cdot \nabla v_j dx, \ \ \ u = (u_1, u_2), \ v = (v_1, v_2) \in \mathbb{H}^1_0 (\Omega),\] and norms $\| \cdot \| = ( \cdot, \cdot)^{1/2}$ and $\\| \cdot \\| = (( \cdot, \cdot))^{1/2}$. We assume the following general hypotheses throughout this section: 1. $\phi: \mathbb{R}_+\times \mathbb{R}^2 \rightarrow \mathbb{R}^2$ is a bounded function such that for each fixed $t \in \mathbb{R}_+$, $\phi(t, \cdot)$ is continuous on $\mathbb{R}^2$. 2. Let $x \in \mathbb{R}^2$. Then $\displaystyle\lim_{s \rightarrow t} \phi(s, x) = \phi(t, x)$ for all $t \neq t_k$, $k \in \mathbb{N}$, the limit $\displaystyle\lim_{s \rightarrow t_k-} \phi(s, x)$ exists and $\displaystyle\lim_{s \rightarrow t_k+} \phi(s, x) = \phi(t_k, x)$, for all $k \in \mathbb{N}$. 3. There is $C > 0$ such that $\|\phi(s, x)-\phi(s, y)\| \leq C\|x-y\|$ for all $s \in \mathbb{R}_+$ and for all $x, y \in \mathbb{R}^2$. 4. There exists a constant $C_1>0$ such that \[\sup_{k\in\mathbb{N}}\sup_{x\in \mathbb{R}^2}\|I_k(x)\| \le C_1.\] 5. There exists a constant $C_{2}>0$ such that \[\|I_k(x)-I_k(y)\| \le C_2 \|x-y\|\] for all $x,y \in \mathbb{R}^2$ and for all $k\in\mathbb{N}$. Now, denote by $P$ the corresponding orthogonal projection $P: \mathbb{L}^2 (\Omega) \rightarrow H$ and set the operators \[A =-\nu P \Delta\] and \[\mathcal{B}(t)(u, v) = q(t)P( (u \cdot \nabla) v).\] It is well known that the Stokes operator $A$ is positive self-adjoint with domain $D(A)$ dense in $H$, $0 \in \rho(A)$ and $A^{-1}$ is compact. Also, there exists $\alpha>0$ such that \[\label{JER1} \langle Au, u \rangle \geq \alpha \|u\|^2_H,\] for all $u \in H$. We also have the following orthogonality property of the nonlinear term which is fundamental and expresses the conservation of energy by the inertial forces: \[\label{JER2} \langle \mathcal{B}(t)(u,v), v \rangle = 0\] for all $u, v \in H$ and for all $t\in\mathbb{R}_+$. For the above properties see, for instance,, and. We set the Hilbert spaces $X^\alpha$, $\alpha \in(0, 1]$, as the domain of the powers of $A$ and we have \[V = X^{\frac{1}{2}} \quad \text{and} \quad \|u\|_{V} = \|\nabla u\|.\] Applying $P$ in the equation \[\displaystyle\frac{\partial u}{\partial t} + q(t)(u \cdot \nabla)u-\nu\Delta u +\nabla p = \phi(t, u),\] we obtain the evolution equation \[\label{I2} u'+ Au + \mathcal{B}(t)(u, u) = \mathcal{F}(t, u),\] where $\mathcal{F}(t, u) = P\phi(t, u)$ for all $t>0$ and $u\in H$, $\langle Au, v \rangle = \nu((u, v))$ for all $u, v \in H$ and \[\langle \mathcal{B}(t)(u, u), w \rangle = q(t) \displaystyle\sum_{i, j = 1}^2 \displaystyle\int_{\Omega} u_i \displaystyle\frac{\partial u_j}{\partial x_i} w_j dx \quad \text{for all} \quad u, w \in H.\] We also assume that: 1. $\mathcal{F} \in PC^+(\mathbb{R}_+\times H, H)$; 2. $\mathcal{B} \in C(\mathbb{R}_+, L^2(H, F))$, where $F = D(A^{-\delta})$ for some $0 < \delta < 1$. Denote $Y$ by $C(\mathbb{R}_+, L^2(H, F))\times PC^+(\mathbb{R}_+\times H, H)$ and let $(Y, \mathbb{R}_+, \sigma)$ be the semidynamical system of translations, that is, $\sigma(t, g) = g_t$ for all $g \in Y$ and $t\geq 0$. Now, set \[\mathcal{M} := \mathcal{H}(\mathcal{B}, \mathcal{F}) = \overline{\{(\mathcal{B}_{\tau}, \mathcal{F}_{\tau}): \, \tau \in \mathbb{R}_+\}},\] where $\mathcal{B}_{\tau}(t) = \mathcal{B}(t + \tau)$ for all $t \in \mathbb{R}_+$ and $\mathcal{F}_{\tau}(t, u) = \mathcal{F}(t + \tau, u)$ for all $t \in \mathbb{R}_+$ and $u\in H$. If $(\widetilde{\mathcal{B}}, \widetilde{\mathcal{F}}) \in \mathcal{M}$ and $\tau\geq 0$ we consider $\widetilde{\mathcal{B}}_{\tau}(t) = \widetilde{\mathcal{B}}(t+\tau)$ and $\widetilde{\mathcal{F}}_{\tau}(t, u) = \widetilde{\mathcal{F}}(t+\tau, u)$ for all $(t, u) \in \mathbb{R}_+\times H$. According to, the equation \[\label{L2} u' + Au + \widetilde{\mathcal{B}}(t)(u,u) = \widetilde{\mathcal{F}}(t, u),\] where $(\widetilde{\mathcal{B}}, \widetilde{\mathcal{F}}) \in \mathcal{H}(\mathcal{B}, \mathcal{F})$, is called the $\mathcal{H}-$class along with the equation [\[I2\]](#I2){reference-type="eqref" reference="I2"}. Define the mapping $B: \mathcal{M} \rightarrow L^2(H, F)$ by \[B(\omega) = B(\widetilde{\mathcal{B}}, \widetilde{\mathcal{F}}) := \widetilde{\mathcal{B}}(0)\] and the mapping $f: \mathbb{R}_+\times\mathcal{M}\times H \rightarrow H$ by \[f(t, \omega, u) = f(t, (\widetilde{\mathcal{B}}, \widetilde{\mathcal{F}}), u) := \widetilde{\mathcal{F}}(0, u).\] Then equation [\[L2\]](#L2){reference-type="eqref" reference="L2"} can be rewritten in the form \[\label{NSNA} u' + Au + B(\sigma(t, \omega))(u,u) = f(t, \sigma(t, \omega), u).\] From [\[JER2\]](#JER2){reference-type="eqref" reference="JER2"}, we obtain \[\label{JER3} \langle B(\omega)(u,v), v \rangle = 0\] for all $u, v \in H$ and for all $\omega \in \mathcal{M}$. Since $\phi$ is bounded and we have Lemma [\[LEMMA1\]](#LEMMA1){reference-type="ref" reference="LEMMA1"}, we may consider in $\mathcal{M}$ the metric $d_{\mathcal{M}}$ given by \[d_{\mathcal{M}}(\omega_1, \omega_2) = d_{\mathcal{M}}((\widetilde{\mathcal{B}}_1, \widetilde{\mathcal{F}}_1), (\widetilde{\mathcal{B}}_2, \widetilde{\mathcal{F}}_2)) = \\|\widetilde{\mathcal{B}}_1-\widetilde{\mathcal{B}}_2\\|_{\infty} + \\|\widetilde{\mathcal{F}}_1-\widetilde{\mathcal{F}}_2\\|_{PC^+_{\infty}},\] where $\displaystyle \\|\widetilde{\mathcal{F}}\\|_{PC^+_{\infty}}= \sup_{(t, u) \in [0,+\infty)\times H} \|\widetilde{\mathcal{F}}(t, u)\|_{H}$ and $\\|\widetilde{\mathcal{B}}\\|_{\infty} = \displaystyle\sup_{\omega \in \mathcal{M}} \\|\widetilde{\mathcal{B}}(\omega)\\|_{\mathcal{L}^2 (H, F)}$. Let us consider the following impulsive system associated to [\[NSNA\]](#NSNA){reference-type="eqref" reference="NSNA"}: \[\label{WeakEq} \left\{ \begin{array}{lll} u' + Au + B(\sigma(t, \omega))(u,u) = f(t, \sigma(t, \omega), u), \ \ u \in H, \; t > 0, \; t \neq t_k, \vspace{1mm}\\ u(t_k)-u(t_k^-) = I_k(u(t_k^-)), \ \ k \in \mathbb N, \vspace{1mm}\\ u(0) = u_0, \end{array} \right.\] which is a weak formulation of [\[l1\]](#l1){reference-type="eqref" reference="l1"}. In the next lines, we show that $f$ satisfies the conditions (C1), (C2), (C3) and (C4) presented in Section 2. This will help us to show that system [\[WeakEq\]](#WeakEq){reference-type="eqref" reference="WeakEq"} admits a unique global mild solution, see Theorems [\[TeoNS\]](#TeoNS){reference-type="ref" reference="TeoNS"} and [\[TEO2.6\]](#TEO2.6){reference-type="ref" reference="TEO2.6"} in the sequel. By Lemma [\[LNS222\]](#LNS222){reference-type="ref" reference="LNS222"}, we may define $\\|f\\|_1 = \sup\{\|f(t, \omega, u)\|_H: \, t\geq0, \omega \in\mathcal{M}, u\in H\}$ From Theorem [\[T1.3\]](#T1.3){reference-type="ref" reference="T1.3"}, we have the following straightforward result of existence and uniqueness of mild solutions. The mild solution of system [\[WeakEq\]](#WeakEq){reference-type="eqref" reference="WeakEq"} may be prolonged on $\mathbb{R}_+$, see Theorem [\[TEO2.6\]](#TEO2.6){reference-type="ref" reference="TEO2.6"}. The system [\[WeakEq\]](#WeakEq){reference-type="eqref" reference="WeakEq"} is called bounded dissipative* if there is a nonempty bounded set $B_0 \subset H$ such that for each bounded set $B \subset H$ there exists $T = T(B) > 0$ such that $\varphi(t, u_0, \omega) \in B_0$ for all $t\geq T$, $u_0\in B$ and $\omega\in\mathcal{M}$. In this case, $B_0$ is called a bounded attractor for the system [\[WeakEq\]](#WeakEq){reference-type="eqref" reference="WeakEq"}. In the next result, we obtain dissipativity for the system [\[WeakEq\]](#WeakEq){reference-type="eqref" reference="WeakEq"}. In the last result, we present an estimative between two solutions with different initial data in the same fiber $\omega \in \mathcal{M}$.	{'timestamp': '2016-07-18T02:08:43', 'yymm': '1607', 'arxiv_id': '1607.04582', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04582'}
# Introduction Germanium is widely used as a detector material in experiments searching for a rare process like the interaction of weakly interacting massive particles (WIMPs) . It is possible to build detectors with very good energy resolution based on the measurement of the ionization produced in the particle interaction, or of the increase of temperature . In addition, the combination of the ionization and heat signals is a powerful tool to distinguish nuclear recoils from electron recoils. Moreover, the crystal-growing process used in the semiconductor industry purifies the material to a high level that matches well the stringent ultra-low radioactivity requirements of rare event searches. The potential of germanium detectors for achieving very low threshold below 1 keV is particularly attractive for searches of WIMPs with masses below 10 GeV/c$^{2}$. The background at energies below 20 keV in such a detector is thus of particular interest. Notably, the contribution from tritium beta decays may have a significant impact on the sensitivity of the next generation of these detectors. The crystallization process removes all cosmogenically-produced radioactive atoms, with the exception of unstable germanium isotopes like $^{68}$Ge (see below). Their populations grow back again when the crystal is kept above ground, and therefore exposed to cosmic rays and the associated hadronic showers. Short-lived isotopes decay rapidly as soon as the detectors are stored underground, deep enough to suppress the hadronic component of the cosmic rays . The isotopes that merit attention have lifetimes exceeding a few tens of days, since shorter-lived nuclei can be eliminated just by storing the detectors in the underground site for some reasonable time before starting data taking. The cosmogenic products that have the most noticeable effect on the low-energy spectrum recorded in germanium detectors are those that decay via electronic capture (EC). The capture is often followed by the emission of a $K$-shell X-ray with characteristic energy between 4 and 11 keV. $L$-and $M$-shell captures will produce weaker lines at approximately 1 and 0.1 keV, respectively. The sharp line shapes and known $K$:$L$:$M$ intensity ratios can be used to identify and subtract the associated events. However, it is preferable to reduce their initial intensities to the lowest possible level. Measurements of the production rates of EC-decaying isotopes is helpful in designing a detector-production procedure that limits these backgrounds to acceptable levels, and, more generally, to constrain models predicting the production rates of all isotopes, including those that may prove to elude direct measurements. Another type of background of particular interest is the beta decay of tritium ($^{3}$H) originated from nuclear reactions induced by the interaction of the hadronic component of cosmic rays with atoms in the material . The electron emitted in the beta decay of tritium has an end point $Q_{\beta}$ of only 18.6 keV, and thus contributes to the background of low-energy events over the entire energy range relevant for low-mass WIMP searches. The lifetime of $^{3}$H is particularly long ($\tau$ = 17.79 y), so the tritium activity can essentially be expected to remain almost the same throughout the life of the detector. The only way to reduce this background is to limit the exposure of the crystal between the time it is grown and its installation underground. There are large uncertainties in model predictions for the production rate of $^{3}$H, and available measurements can only provide crude upper limits . The EDELWEISS collaboration has operated an array of 24 germanium heat-and-ionization detectors with the objective to perform searches for WIMPs with a total exposure close to 3000 kg$\cdot$d, and more specific searches for low-mass WIMPs with a subset of its detectors with the best experimental energy thresholds . The experiment is located in the Laboratoire Souterrain de Modane (LSM) and protected by a mean rock overburden of 1800 m (4800 m.w.e.) that reduces the cosmic ray flux to about 5 $\mu$/m$^2$/day , i.e. 10$^6$ times less than at the surface. The detectors are covered by interleaved electrodes that provide an efficient tool to reject surface events (i.e. particle interactions taking place within $\sim$`<!-- -->`{=html}2 mm from the detector surface) down to energies of $\sim$`<!-- -->`{=html}1 keV . The resolutions achieved with these detectors, the reduction of the external gamma-ray background and the excellent surface-event rejection performance of the interleaved electrodes , have made possible a precise measurement of decay rates of different nuclei in the bulk volume of germanium detectors, and in particular, for the first time, to measure unambiguously the intensity of the tritium spectrum. Efforts were made to keep to a minimum the exposure of each crystal to cosmic rays throughout the detector production. A history of the key steps in the detector production process is available. Despite this, there are non-negligible systematic uncertainties in the recorded history of exposure times. However, these uncertainties can be tested, because unforeseen production delays[^1] led to a relatively large spread (up to a factor of 4) in the exposures of the different detectors to cosmic rays. It was therefore possible to check on isotopes with the largest statistics that the observed activation rates scaled with the expectations from the recorded history of exposure times. In the following, we detail the EDELWEISS-III setup relevant to this measurement (Section [2](#sec:setup){reference-type="ref" reference="sec:setup"}), as well as the data selection (Section [3](#sec:data){reference-type="ref" reference="sec:data"}). We present the expected properties of the activation of tritium and other isotopes and of the energy spectrum of the emitted electrons, and describe the analysis used to extract their intensities from the data (Section [\[sec:analysis\]](#sec:analysis){reference-type="ref" reference="sec:analysis"}). These results are then converted into production rates during the exposure above ground (Section [\[sec-history\]](#sec-history){reference-type="ref" reference="sec-history"}) and compared to a previous measurement and calculations (Section [\[sec:pr\]](#sec:pr){reference-type="ref" reference="sec:pr"}). # Experimental setup {#sec:setup} The active target of the EDELWEISS-III WIMP search experiment consists of twenty-four 800-g Fully-InterDigit (FID) germanium detectors, cooled down to an operating temperature of 18 mK. All materials surrounding the detectors have been tested for their radiopurity . The cryostat is surrounded by 20 cm of lead and 50 cm of polyethylene shielding. A constant flow of de-radonized air in the vicinity of the cryostat reduces the radon level down to 30 mBq/m$^3$. The shielding is surrounded by a 100 m$^2$ muon veto made of plastic scintillator modules with a geometrical coverage of more than 98% . The detectors are high-purity germanium cylindrical crystals of 70 mm in diameter and 40 mm in height. All surfaces are entirely covered with interleaved Al electrodes, biased at alternate values of potentials. The electrodes are 150 $\mu$m in width and separated by 2 mm. The potentials applied to the electrodes are chosen to determine an axial electric field in the bulk of the detector , while in the volume within about 2 mm from the surfaces the electric field is parallel to them. As a consequence, electron-hole pairs created in the bulk volume are collected in the axial field by the fiducial electrodes on both sides of the detector, while surface events will be collected by adjacent electrodes. This scheme extends also to the cylindrical surfaces. Fiducial (or bulk) events can thus be selected on the basis of the presence of signals of opposite signs on the fiducial electrodes on each side of the detector, and on the absence of signals on all other electrodes. The fiducial ionization $E_{fid}$ is defined as the average of the signals on the two fiducial electrodes. Calibrations with a $^{210}$Pb source in equilibrium with its Bi and Po daughters have shown that this technique can reject surface $\beta$'s, $\alpha$'s and $^{206}$Pb recoils with an efficiency of the order of 99.999% . Nuclear recoils are identified by comparing the ionization signal with the $\sim\mu K$ rise in temperature accompanying each interaction, measured with two 15 mm$^{3}$ germanium NTD (Neutron Transmutation Doped) thermistors glued on each side of the detector. The charge and temperature signals are calibrated using a $^{133}$Ba $\gamma$-ray source: the units of both signals are thus in keV-electron-equivalent (keV$_{ee}$). Since this work is devoted to the study of bulk electron recoil populations, the subscript $ee$ will be omitted in the following. The signals on both thermistors are combined into a single heat measurement $E_c$. The average dispersion in ionization signals $E_{fid}$ at 0 keV is characterized by $\sigma_{fid}$ = 230 eV (baseline resolution). For the heat signal $E_c$, the corresponding values of $\sigma_{c}$ range from 150 to 500 eV depending on detectors. The resolution increases with the energy of the signal, with a linear term that is dominated by charge trapping effects . At 356 keV, the resolutions $\sigma(E)$ are larger than the baseline $\sigma_{fid}$ and $\sigma_{c}$ values and are approximately 4 keV. The data acquisition triggers if the amplitude of one of the heat signals is above a threshold value[^2]. These values are automatically adjusted every minute for each heat sensor according to the event rates recorded minute by minute. This, together with the 0.5 kHz sampling of the heat channels, results in the possibility to set the thresholds to levels that are close to $\sim$`<!-- -->`{=html}4.5$\sigma_{c}$, while keeping the trigger rates per detector at approximately 50 mHz. # Data selection {#sec:data} ## Data set and detector selection The data was recorded over a period of 280 calendar days, from July 2014 to April 2015. During that period, 160 days were devoted to WIMP searches. All twenty-four detectors were used to define coincidences between detectors. The average heat trigger threshold and baseline resolutions were monitored hour by hour. For this analysis, only hours when this trigger threshold is below 2 keV were selected. The hourly fiducial baseline resolution on the fiducial ionization measurement was also required to be less than 400 eV. Two detectors with a failing ionization channel, preventing an efficient application of the fiducial cut, were excluded from this exposure. Three detectors had less than one day of running time with an online threshold of below 2 keV and were also excluded from further analysis. The dead-time corrected exposure of the remaining 19 detectors is 1853 detector$\cdot$day. This exposure is considered for different global fits used in order to evaluate systematic uncertainties. As each detector had a different history of exposure to cosmic rays, the final results will be based on the fits to individual detector data for the 13 detectors with individual exposure greater than 60 days, corresponding to 87.0% of the total exposure. Similarly, a sub-sample of events with an online threshold below 0.8 keV is also defined, for precision tests of the efficiency correction and of the sample purity at energies lower than used in the final analysis. This sample corresponds to 499 detector$\cdot$day, to which 10 detectors contribute. ## Event selection {#sect-neg} An event is included in the analysis if its $E_{fid}$ ($E_c$) value is larger than 3.5$\sigma_{fid}$ ($3\sigma_{c}$). The values of $\sigma_{fid}$ and $\sigma_{c}$ are measured hour by hour from Gaussian fits to the distribution of amplitudes observed in events that have not triggered the online threshold. Cuts are also performed on the $\chi^2$ of the fit of template pulse shapes to the ionization signals, resulting in an efficiency loss of less than 1%. Fig. [\[fig-datasel\]](#fig-datasel){reference-type="ref" reference="fig-datasel"}a) shows the distribution of the fiducial ionization $E_{fid}$ as a function of the heat signal $E_c$ for selected events in a sample where the cut on $E_{fid}$ has been relaxed from $3.5\sigma_{fid}$ to $3\sigma_{fid}$. On this figure, events in gray are those rejected by the fiducial cut. Namely, events are considered as having occurred in the fiducial volume if the signals on the two non-fiducial electrodes, as well as the difference between the two fiducial electrodes, are each consistent with noise within $\pm$`<!-- -->`{=html}2.574$\sigma$ (99%). The rejected events are mostly distributed along the two dashed lines on Fig. [\[fig-datasel\]](#fig-datasel){reference-type="ref" reference="fig-datasel"}a) representing the expected location for surface beta events (blue dotted line) and surface gamma events (red dotted line). Fig. [\[fig-datasel\]](#fig-datasel){reference-type="ref" reference="fig-datasel"}b) shows the same distribution after the above-described selection of fiducial events. The events in gray are those where the values of $E_{fid}$ and $E_c$ are not consistent with each other within 2.574$\sigma$ of the experimental resolutions. They are associated with the background of heat-only events discussed in Ref. . Their origin is illustrated in Figs. [\[fig-datasel\]](#fig-datasel){reference-type="ref" reference="fig-datasel"}c) and d), where the low-energy part is shown together with events selected with cuts calculated using $\|E_{fid}\|$ instead of $E_{fid}$, namely, $\|E_{fid}\|$ $>$ 3$\sigma_{fid}$ and the consistency cut is applied on ($E_c-\|E_{fid}\|$). For clarity, events with trigger threshold below 0.8 keV (Fig. [\[fig-datasel\]](#fig-datasel){reference-type="ref" reference="fig-datasel"}c) are shown separately from those with trigger thresholds between 0.8 and 1.5 keV (Fig. [\[fig-datasel\]](#fig-datasel){reference-type="ref" reference="fig-datasel"}d). The distribution of $E_{fid}$ for heat-only events is a Gaussian centered at $E_{fid}=0$. In Figs. [\[fig-datasel\]](#fig-datasel){reference-type="ref" reference="fig-datasel"}c) and d), what is therefore observed after the rejection of events with $\|E_{fid}\| <3 \sigma_{fid}$ are two populations of events, symmetric in $\pm~E_{fid}$. A leakage from the population of surface gamma and beta events would be centered around the red and blue dotted lines, respectively, on Fig. [\[fig-datasel\]](#fig-datasel){reference-type="ref" reference="fig-datasel"}c) and d), and not distributed symmetrically at $\pm E_{fid}$. Such a symmetry is clearly observed for the events in gray, corresponding to those rejected by the consistent cut on $E_{c}-\|E_{fid}\|$. Most of the potential backgrounds below 2 keV thus comes from heat-only events. The population of heat-only events with $E_{fid}>0$ that are not rejected by the consistency cut will be estimated by mirroring the equivalent population with $E_{fid}<0$. For instance, in Fig. [\[fig-datasel\]](#fig-datasel){reference-type="ref" reference="fig-datasel"}c) and for $E_c<$`<!-- -->`{=html}0.8 keV, the population of gray points with $E_{fid}<0$ is nearly equal to those with $E_{fid}>0$. This is also true for the events in black, indicating that most events with positive ionization energies and $E_c$ values below 0.8 keV remaining after applying the consistency cut are due to the tail of heat-only events. [^1]: These delays occurred to solve problems related to surface current leakage, as described in . [^2]: The online filtering on the heat signals is less efficient than the offline version, resulting in a slight degradation of resolution. Here, $E_c$ and $\sigma_{c}$ refer to the offline-filtered heat signals. As the trigger is applied to the signals calculated online, some $E_c$ values may be less than their corresponding online trigger threshold by as much as $\sigma_{c}$.	{'timestamp': '2016-07-18T02:08:01', 'yymm': '1607', 'arxiv_id': '1607.04560', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04560'}
null	null
# Introduction Many problems, particularly in combinatorics, reduce to asking whether some graph with a given property exists, or alternatively, asking how many such non-isomorphic graphs exist. Such graph search and graph enumeration problems are notoriously difficult, in no small part due to the extremely large number of symmetries in graphs. In practical problem solving, it is often advantageous to eliminate these symmetries which arise naturally due to graph isomorphism: typically, if a graph $G$ is a solution then so is any other graph $G'$ that is isomorphic to $G$. General approaches to graph search problems typically involve either: generate and test, explicitly enumerating all (non-isomorphic) graphs and checking each for the given property, or constrain and generate, encoding the problem for some general-purpose discrete satisfiability solver (i.e. SAT, integer programming, constraint programming), which does the enumeration implicitly.	{'timestamp': '2016-08-02T02:11:05', 'yymm': '1607', 'arxiv_id': '1607.04829', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04829'}
null	null
null	null
null	null
# Handling batch insertions In this section, we study the dynamic DFS tree problem in the batch insertion setting. The goal of this section is to prove Theorem [\[batch-ins\]](#batch-ins){reference-type="ref" reference="batch-ins"}. Our algorithm basically follows the same framework for fully dynamic DFS proposed in. Since we are only interested in the dynamic DFS tree problem in the batch insertion setting, the algorithms [BatchInsert]{.sans-serif} and [DFS]{.sans-serif} presented below is a moderate simplification of the original algorithm in, by directly pruning those details unrelated to insertions. In Algorithm [BatchInsert]{.sans-serif}, we first attach each inserted vertex to the super root $r$, and pretend it has been there since the very beginning. Then only edge insertions are to be considered. All inserted edges are added into the reduced adjacency lists of corresponding vertices. We then use [DFS]{.sans-serif} to traverse the graph starting from $r$ based on $T$, $L$, and build the new DFS tree while traversing the entire graph and updating the reduced adjacency lists. In Algorithm [DFS]{.sans-serif}, the new DFS tree is built in a recursive fashion. Every time we enter an untouched subtree, say $T(u)$, from vertex $v \in T(u)$, we change the root of $T(u)$ to $v$ and go through $\mathit{path}(v, u)$; i.e., we wish to reverse the order of $\mathit{path}(u, v)$ in $T^$. One crucial step behind this operation is that we need to find a new root for each subtree $T(w)$ originally hanging on $\mathit{path}(u, v)$. The following lemma tells us where the $T(w)$ should be rerooted on $\mathit{path}(u, v)$ in $T^$. Let $Q(T(w), u, v)$ be the edge between the highest vertex on $\mathit{path}(u, v)$ incident to a vertex in subtree $T(w)$, and the corresponding vertex in $T(w)$. $Q(T(w), u, v)$ is defined to be $\mathsf{Null}$ if such an edge does not exist. By Lemma [\[feasible_edge\]](#feasible_edge){reference-type="ref" reference="feasible_edge"}, it suffices to ignore all other edges but just keep the edge returned by $Q(T(w), u, v)$; this is because we have reversed the order of $\mathit{path}(u, v)$ in $T^$ and thus $Q(T(w), u, v)$ connects to the lowest possible position in $T^$. Hence $T(w)$ should be rerooted at $Q(T(w), u, v)$. Denote $(x, y)$ to be the edge returned by $Q(T(w), u, v)$ where $x \in \mathit{path}(u, v)$, and then we add $y$ into $L(x)$. After finding an appropriate entering edge for each hanging subtree, we process each vertex $v \in \mathit{path}(u, v)$ in ascending order of depth (with respect to tree $T$). For every unvisited $w \in L(v)$, we set $\mathit{par}^(w) = v$, and recursively call $\mathrm{\textsf{DFS}}(w)$. For now we have not discussed how to implement $Q(T(w), u, v)$ and the above algorithm only assumes blackbox queries to $Q(T(\cdot), \cdot, \cdot)$. The remaining problem is to devise a data structure $\mathcal{D}$ to answer all the queries demanded by Algorithm [DFS]{.sans-serif} in $O(n)$ total time. We will show in the next section that there exists a data structure $\mathcal{D}$ with the desired performance, which is stated as the following lemma. # Introduction Depth First Search (DFS) is one of the most renowned graph traversal techniques. After Tarjan's seminal work , it demonstrates its power by leading to efficient algorithms to many fundamental graph problems, e.g., biconnected components, strongly connected components, topological sorting, bipartite matching, dominators in directed graph and planarity testing. Real world applications often deal with graphs that keep changing with time. Therefore it is natural to study the dynamic version of graph problems, where there is an online sequence of updates on the graph, and the algorithm aims to maintain the solution of the studied graph problem efficiently after seeing each update. The last two decades have witnessed a surge of research in this area, like connectivity , reachability , shortest path , bipartite matching , and min-cut . We consider the dynamic maintenance of DFS trees in undirected graphs. As observed by Baswana et al. and Nakamura and Sadakane , the incremental* setting, where edges/vertices are added but never deleted from the graph, is arguably easier than the fully dynamic setting where both kinds of updates can happen---in fact, they provide algorithms for incremental DFS with $\tilde{O}(n)$ worst case update time, which is close to the trivial $\Omega(n)$ lower bound when it is required to explicitly report a DFS tree after each update. *So, is there an algorithm that requires nearly linear preprocessing time and space, and reports a DFS tree after each incremental update in $O(n)$ time?* In this paper, we study the problem of maintaining a DFS tree in the incremental setting, and give an affirmative answer to this question. ## Previous works on dynamic DFS Despite the significant role of DFS tree in static algorithms, there is limited progress on maintaining a DFS tree in the dynamic setting. Many previous works focus on the total time of the algorithm for any arbitrary updates. Franciosa et al. designed an incremental algorithm for maintaining a DFS tree in a DAG from a given source, with $O(mn)$ total time for an arbitrary sequence of edge insertions; Baswana and Choudhary designed a decremental algorithm for maintaining a DFS tree in a DAG with expected $O(mn\log n)$ total time. For undirected graphs, Baswana and Khan designed an incremental algorithm for maintaining a DFS tree with $O(n^2)$ total time. These algorithms used to be the only results known for the dynamic DFS tree problem. However, none of these existing algorithms, despite that they are designed for only a partially dynamic environment, achieves a worst case bound of $o(m)$ on the update time. That barrier is overcome in the recent breakthrough work of Baswana et al. , they provide, for undirected graphs, a fully dynamic algorithm with worst case $O(\sqrt{mn} \log^{2.5} n)$ update time, and an incremental algorithm with worst case $O(n \log^{3} n)$ update time. Due to the rich information in a DFS tree, their results directly imply faster worst case fully dynamic algorithms for subgraph connectivity, biconnectivity and 2-edge connectivity. The results of Baswana et al. suggest a promising way to further improve the worst case update time or space consumption for those fully dynamic algorithms by designing better dynamic algorithms for maintaining a DFS tree. In particular, based on the framework by Baswana et al. , Nakamura and Sadakane propose an algorithm which takes $O(\sqrt{mn} \log^{1.75} n / \sqrt{\log \log n})$ time per update in the fully dynamic setting and $O(n \log n)$ time in the incremental setting, and $O(m \log n)$ bits of space. ## Our results In this paper, following the approach of , we improve the update time for the incremental setting, also studied in , by combining a better data structure, a novel tree-partition lemma by Duan and Zhang and the fractional-cascading technique by Chazelle and Guibas . For any set $U$ of incremental updates (insertion of a vertex/an edge), we let $G + U$ denote the graph obtained by applying the updates in $U$ to the graph $G$. Our results build on the following main theorem. By the above theorem combined with a de-amortization trick in , we establish the following corollary for maintaining a DFS tree in an undirected graph with incremental updates. ## Organization of the Paper In Section 2 we introduce frequently used notations and review two building blocks of our algorithm---the tree partition structure and the fractional cascading technique. In Section 3, we consider a batched version of the incremental setting, where all incremental updates are given at once, after which a single DFS tree is to be reported. Given an efficient scheme to answer queries of form $Q(T(\cdot), \cdot, \cdot)$, we prove Theorem [\[batch-ins\]](#batch-ins){reference-type="ref" reference="batch-ins"}, which essentially says there is an efficient algorithm, which we call [BatchInsert]{.sans-serif}, for the batched incremental setting. In Section 4, we elaborate on the implementation of the central query subroutine $Q(T(\cdot), \cdot, \cdot)$ used in the batched incremental algorithm. We first review a standard de-amortization technique, applying which our algorithm for the batched setting directly implies the efficient algorithm for the incremental setting stated in Corollary [\[cor-incre-dfs\]](#cor-incre-dfs){reference-type="ref" reference="cor-incre-dfs"}. We then, in Sections [4.1](#sec:logn){reference-type="ref" reference="sec:logn"} and [4.2](#sec:nsquare){reference-type="ref" reference="sec:nsquare"} respectively, introduce (1) an optimized data structure that takes $O(m \log n)$ time for preprocessing and answers each query in $O(\log n)$ time, and (2) a relatively simple data structure that takes $O(n^2)$ time for preprocessing and answers each query in $O(1)$ time. One of these two structures, depending on whether $m \log n > n^2$ or not, is then used in Section [4.3](#sec:mlogn){reference-type="ref" reference="sec:mlogn"} to implement a scheme that answers each query in amortized $O(1)$ time. This is straightforward when the $(n^2, 1)$ structure is used. When instead the $(m \log n, \log n)$ structure is used, we apply a nontrivial combination of the tree partition structure and the fractional cascading technique to bundle queries together, and answer each bundle using a single call to the $(m \log n, \log n)$ structure. We show that the number of such bundles from queries made by [BatchInsert]{.sans-serif} cannot exceed $O(n / \log n)$, so the total time needed for queries is $O(n)$. This finishes the proof of Theorem [\[batch-ins\]](#batch-ins){reference-type="ref" reference="batch-ins"} and Corollary [\[cor-incre-dfs\]](#cor-incre-dfs){reference-type="ref" reference="cor-incre-dfs"} and concludes the paper. # Preliminaries Let $G = (V, E)$ denote the original graph, $T$ a corresponding DFS tree, and $U$ a set of inserted vertices and edges. We first introduce necessary notations. - $T(x)$: The subtree of $T$ rooted at $x$. - $\mathit{path}(x, y)$: The path from $x$ to $y$ in $T$. - $\mathit{par}(v)$: The parent of $v$ in $T$. - $N(x)$: The adjacency list of $x$ in $G$. - $L(x)$: The reduced adjacency list for vertex $x$, which is maintained during the algorithm. - $T^$: The newly generated DFS tree. - $\mathit{par}^(v)$: The parent of $v$ in $T^*$. Our result uses a tree partition lemma in and the famous fractional cascading structure in, which are summarized as the following two lemmas. # Dealing with queries in [BatchInsert]{.sans-serif} In this section we prove Lemma [\[query_time\]](#query_time){reference-type="ref" reference="query_time"}. Once this goal is achieved, the overall time complexity of batch insertion taken by Algorithm [BatchInsert]{.sans-serif} would be $O(n + \|U\|)$. In the following part of this section, we will first devise a data structure in Section [4.1](#sec:logn){reference-type="ref" reference="sec:logn"}, that answers any single query $Q(T(w), u, v)$ in $O(\log n)$ time, which would be useful in other parts of the algorithm. We will then present another simple data structure in Section [4.2](#sec:nsquare){reference-type="ref" reference="sec:nsquare"}, which requires $O(n^2)$ preprocessing time and $O(n^2)$ space and answers each query in $O(1)$ time. Finally, we propose a more sophisticated data structure in Section [4.3](#sec:mlogn){reference-type="ref" reference="sec:mlogn"}, which requires $O(m\log n)$ preprocessing time and $O(m \log n)$ space and answer all queries $Q(T(w), x, y)$ in a single run of [BatchInsert]{.sans-serif} in $O(n)$ time. Hence, we can always have an algorithm that handles a batch insertion $U$ in $O(n + \|U\|)$ time using $O(\min\{m\log n, n^2\})$ preprocessing time and $O(\min\{m\log n, n^2\})$ space, thus proving Theorem [\[batch-ins\]](#batch-ins){reference-type="ref" reference="batch-ins"}. We can then prove Corollary [\[cor-incre-dfs\]](#cor-incre-dfs){reference-type="ref" reference="cor-incre-dfs"} using the following standard de-amortization argument. ## Answering a single query in $O(\log n)$ time {#sec:logn} We show in this subsection that the query $Q(T(\cdot), \cdot, \cdot)$ can be reduced efficiently to the range successor query (see, e.g.,, for the definition of range successor query), and show how to answer the range successor query, and thus any individual query $Q(T(\cdot), \cdot, \cdot)$, in $O(\log n)$ time. To deal with a query $Q(T(w), x, y)$, first note that since $T$ is a DFS tree, all edges not in $T$ but in the original graph $G$ must be ancestor-descendant edges. Querying edges between $T(w)$ and $\mathit{path}(x, y)$ where $x$ is an ancestor of $y$ and $T(w)$ is hanging from $\mathit{path}(x, y)$ is therefore equivalent to querying edges between $T(w)$ and $\mathit{path}(x, \mathit{par}(w))$, i.e., $Q(T(w), x, y) = Q(T(w), x, \mathit{par}(w))$. From now on, we will consider queries of the latter form only. Consider the DFS sequence of $T$, where the $i$-th element is the $i$-th vertex reached during the DFS on $T$. Note that every subtree $T(w)$ corresponds to an interval in the DFS sequence. Denote the index of vertex $v$ in the DFS sequence by $\mathit{first}(v)$, and the index of the last vertex in $T(v)$ by $\mathit{last}(v)$. During the preprocessing, we build a 2D point set $S$. For each edge $(u, v) \in E$, we add a point $p = (\mathit{first}(u), \mathit{first}(v))$ into $S$. Notice that for each point $p \in S$, there exists exactly one edge $(u, v)$ associated with $p$. Finally we build a 2D range tree on point set $S$ with $O(m\log n)$ space and $O(m\log n)$ preprocessing time. To answer an arbitrary query $Q(T(w), x, \mathit{par}(w))$, we query the point with minimum $x$-coordinate lying in the rectangle $\Omega = [\mathit{first}(x), \mathit{first}(w)-1] \times [\mathit{first}(w), \mathit{last}(w)]$. If no such point exists, we return [Null]{.sans-serif} for $Q(T(w), x, \mathit{par}(w))$. Otherwise we return the edge corresponding to the point with minimum $x$-coordinate. Now we prove the correctness of our approach. - If our method returns [Null]{.sans-serif}, $Q(T(w), x, \mathit{par}(w))$ must equal [Null]{.sans-serif}. Otherwise, suppose $Q(T(w), x, \mathit{par}(w)) = (u, v)$. Noticing that $(\mathit{first}(u), \mathit{first}(v))$ is in $\Omega$, it means our method will not return [Null]{.sans-serif} in that case. - If our method does not return [Null]{.sans-serif}, denote $(u', v')$ to be the edge returned by our method. We can deduce from the query rectangle that $u' \in T(x) \backslash T(w)$ and $v' \in T(w)$. Thus, $Q(T(w), x, \mathit{par}(w)) \neq \textsf{Null}$. Suppose $Q(T(w), x, \mathit{par}(w)) = (u, v)$. Notice that $(\mathit{first}(u), \mathit{first}(v))$ is in $\Omega$, which means $\mathit{first}(u') \le \mathit{first}(u)$. If $u' = u$, then our method returns a feasible solution. Otherwise, from the fact that $\mathit{first}(u') < \mathit{first}(u)$, we know that $u'$ is an ancestor of $u$, which contradicts the definition of $Q(T(w), x, \mathit{par}(w))$. ## An $O(n^2)$-space data structure {#sec:nsquare} [\[pre_all\]]{#pre_all label="pre_all"} In this subsection we propose a data structure with quadratic preprocessing time and space complexity that answers any $Q(T(\cdot), \cdot, \cdot)$ in constant time. Since we allow quadratic space, it suffices to precompute and store answers to all possible queries $Q(T(w), u, \mathit{par}(w))$. For preprocessing, we enumerate each subtree $T(w)$, and fix the lower end of the path to be $v = \mathit{par}(w)$ while we let the upper end $u$ go upward from $v$ by one vertex at a time to calculate $Q(T(w), u, v)$ incrementally, in order to get of the form $Q(T(w), \cdot, \cdot)$ in $O(n)$ total time. As $u$ goes up, we check whether there is an edge from $T(w)$ to the new upper end $u$ in $O(1)$ time; for this task we build an array (based on the DFS sequence of $T$) for each vertex, and insert an 1 into the appropriate array for each edge, and apply the standard prefix summation trick to check whether there is an 1 in the range corresponding to $T(w)$. To be precise, let $A_u: [n] \rightarrow \{0, 1\}$ denote the array for vertex $u$. Recall that $\mathit{first}(v)$ denotes the index of vertex $v$ in the DFS sequence, and $\mathit{last}(v)$ the index of the last vertex in $T(v)$. For a vertex $u$, we set $A_u[\mathit{first}(v)]$ to be 1 if and only if there is an edge $(u, v)$ where $u$ is the higher end. Now say, we have the answer to $Q(T(w), u, v)$ already, and want to get $Q(T(w), u', v)$ in $O(1)$ time, where $u' = \mathit{par}(u)$. If there is an edge between $T(w)$ and $u'$, then it will be the answer. Or else the answer to $Q(T(w), u', v)$ will be the same as to $Q(T(w), u, v)$. In order to know whether there is an edge between $T(w)$ and $u'$, we check the range $[\mathit{first}(w), \mathit{last}(w)]$ in $A_{u'}$, and see if there is an $1$ in $O(1)$ time using the prefix summation trick. ## An $O(m\log n)$-space data structure {#sec:mlogn} Observe that in [BatchInsert]{.sans-serif} (and [DFS]{.sans-serif}), a bunch of queries $\{Q(T(w_i), x, y)\}$ are always made simultaneously, where $\{T(w_i)\}$ is the set of subtrees hanging from $\mathit{path}(x, y)$. We may therefore answer all queries for a path in one pass, instead of answering them one by one. By doing so we confront two types of hard queries. First consider an example where the original DFS tree $T$ is a chain $L$ where $a_1$ is the root of $L$ and for $1 \le i \le n-1$, $a_{i+1}$ is the unique child of $a_i$. When we invoke $\textsf{DFS}(a_1)$ on $L$, $path(u, v)$ is the single node $a_1$. Thus, we will call $Q(T(a_2), a_1, a_1)$ and add the returned edge into $L(a_1)$. Supposing there are no back-edges in this graph, the answer of $Q(T(a_2), a_1, a_1)$ will be the edge $(a_1, a_2)$. Therefore, we will recursively call the $\textsf{DFS}(a_2)$ on the chain $(a_2, a_n)$. Following further steps of [DFS]{.sans-serif}, we can see that we will call the query $Q(T(w), x, y)$ for $\Omega(n)$ times. For the rest of this subsection, we will show that we can deal with this example in linear time. The idea is to answer queries involving short paths in constant time. For instance, in the example shown above, $path(u, v)$ always has constant length. We show that when the length of $path(u, v)$ is smaller than $2\log n$, it is affordable to preprocess all the answers to queries of this kind in $O(m\log n)$ time and $O(n \log n)$ space The second example we considered is given as Figure [\[fig:heavy\]](#fig:heavy){reference-type="ref" reference="fig:heavy"}. In this tree, the original root is $r$. Suppose the distance between $r$ and $r'$ is $n / 2$. When we invoke $\textsf{DFS}(r')$, $path(u, v)$ the path from $r$ to $r'$. Thus, we will call $T(a_1, r, r')$, $T(a_2, r, r')$, $\ldots$, $T(a_{n-2}, r, r')$, which means we make $\Omega(n)$ queries. In order to deal with this example in linear time, the main idea is using fractional cascading to answer all queries $Q(T(w), x, y)$ with a fixed $path(u, v)$, for all subtrees $T(w)$ with small size. In the examples shown above, all subtrees cut off $path(u, v)$ have constant size and thus the total time complexity for this example is $O(n)$. We will finally show that, by combining the two techniques mentioned above, it is enough to answer all queries $Q(T(w), x, y)$ in linear time, thus proving Lemma [\[query_time\]](#query_time){reference-type="ref" reference="query_time"}. ## Data structure {#data-structure .unnumbered} The data structure consists of the following parts. 1. Build the 2D-range successor data structure that answers any $Q(T(\cdot), \cdot, \cdot)$ in $O(\log n)$ time. 2. For each ancestor-descendent pair $(u, v)$ such that $u$ is at most $2\log n$ hops above $v$, precompute and store the value of $Q(T(v), u, \mathit{par}(v))$. 3. Apply Lemma [\[tree_partition_lem\]](#tree_partition_lem){reference-type="ref" reference="tree_partition_lem"} with parameter $k = \log n$ and obtain a marked set of size $O(n / \log n)$. Let $M$ be the set of all marked vertices $x$ such that $\|T(x)\| \geq \log n$. For every $v\notin M$, let $\mathit{anc}_v\in M$ be the nearest ancestor of $v$ in set $M$. Next we build a fractional cascading data structure for each $u\in M$ in the following way. Let $M_u$ be the set of all vertices in $T(u)$ whose tree paths to $u$ do not intersect any other vertices $u^\prime \neq u$ from $M$, namely $M_u = \{v\mid \mathit{anc}_v = u \}$; see Figure [\[ds\]](#ds){reference-type="ref" reference="ds"} for an example. Then, apply Lemma [\[fractional_cascading\]](#fractional_cascading){reference-type="ref" reference="fractional_cascading"} on all $N(v), v\in M_u$ where $N(v)$ is treated as sorted array in an ascending order with respect to depth of the edge endpoint opposite to $v$; this would build a fractional cascading data structure that, for any query encoded as a $w\in V$, answers for every $v\in M_u$ its highest neighbour below vertex $w$ in total time $O(\|M_u\| + \log n)$. Here is a structural property of $M$ that will be used when answering queries. ## Preprocessing time {#preprocessing-time .unnumbered} First of all, for part (), as discussed in a previous subsection, 2D-range successor data structure takes time $O(m\log n)$ to initialize. Secondly, for part (), on the one hand by Lemma [\[tree_partition_lem\]](#tree_partition_lem){reference-type="ref" reference="tree_partition_lem"} computing a tree partition takes time $O(n\log n)$; on the other hand, by Lemma [\[fractional_cascading\]](#fractional_cascading){reference-type="ref" reference="fractional_cascading"}, initializing the fractional cascading with respect to $u\in M$ costs $O(\sum_{v\in M_u}\|N(v)\|)$ time. Since, by definition of $M_u$, each $v\in V$ is contained in at most one $M_u, u\in M$, the overall time induced by this part would be $O(\sum_{u\in M}\sum_{v\in M_u}\|N(v)\|) = O(m)$. Preprocessing part () requires a bit of cautions. The procedure consists of two steps. (1) For every ancestor-descendent pair $(u, v)$ such that $u$ is at most $2\log n$ hops above $v$, we mark $(u, v)$ if $u$ is incident to $T(v)$. Here goes the algorithm: for every edge $(u, w)\in E$ ($u$ being the ancestor), let $z\in \mathit{path}(u, w)$ be the vertex which is $2\log n$ hops below $u$ (if $\mathit{path}(u, w)$ has less than $2\log n$ hops, then simply let $z = w$); note that this $z$ can be found in constant time using the level-ancestor data structure which can be initialized in $O(n)$ time. Then, for every vertex $v\in \mathit{path}(u, z)$, we mark the pair $(u, v)$. The total running time of this procedure is $O(m\log n)$ since each edge $(u, w)$ takes up $O(\log n)$ time. (2) Next, for each $v\in V$, we compute all entries $Q(T(v), u, \mathit{par}(v))$ required by () in an incremental manner. Let $u_1, u_2, \cdots, u_{2\log n}$ be the nearest $2\log n$ ancestors of $v$ sorted in descending order with respect to depth, and then we directly solve the recursion $Q(T(v), u_{i+1}, \mathit{par}(v)) = \begin{cases} Q(T(v), u_{i}, \mathit{par}(v)) & (u_{i+1}, v)\text{ is not marked}\\ u_{i+1} & i=0\text{ or }(u_{i+1}, v)\text{ is marked}\\ \end{cases}$ for all $0\leq i < 2\log n$ in $O(\log n)$ time. The total running time would thus be $O(n\log n)$. Summing up ()()(), the preprocessing time is bounded by $O(m\log n)$. ## Query algorithm and total running time {#query-algorithm-and-total-running-time .unnumbered} We show how to utilize the above data structures ()()() to implement $Q(T(\cdot), \cdot, \cdot)$ on line 9-11 in Algorithm [DFS]{.sans-serif} such that the overall time complexity induced by this part throughout a single execution of Algorithm [BatchInsert]{.sans-serif} is bounded by $O(n)$. Let us say we are given $(w_1, w_2, \cdots, w_t) = \mathit{path}(u, v)$ and we need to compute $Q(T(x), u, v)$ for every subtree $T(x)$ that is hanging on $\mathit{path}(u, v)$. There are three cases to discuss. (1) If $\mathit{path}(u, v)\cap M = \emptyset$, by Lemma [\[struct-marked\]](#struct-marked){reference-type="ref" reference="struct-marked"} we claim $\mathit{path}(u, v)$ has at most $2\log n$ hops, and then we can directly retrieve the answer of $Q(T(x), u, v)$ from precomputed entries of (), each taking constant query time. (2) Second, consider the case where $\mathit{path}(u, v)\cap M \neq \emptyset$. Let $s_1, s_2, \cdots, s_l, l\geq 1$ be the consecutive sequence (in ascending order with respect to depth in tree $T$) of all vertices from $M$ that are on $\mathit{path}(u, v)$. For those subtrees $T(x)$ that are hanging on $\mathit{path}(u, \mathit{par}(s_1))$, we can directly retrieve the value of $Q(T(x), u, \mathit{par}(x))$ from () in constant time, as by Lemma [\[struct-marked\]](#struct-marked){reference-type="ref" reference="struct-marked"} $\mathit{path}(u, \mathit{par}(s_1))$ has at most $2\log n$ hops. (3) Third, we turn to study the value of $Q(T(x), u, \mathit{par}(x))$ when $\mathit{par}(x)$ belongs to a $\mathit{path}(s_i, \mathit{par}(s_{i+1})), i<l$ or $\mathit{path}(s_l, v)$. The algorithm is two-fold. (a) First, we make a query of $u$ to the fractional cascading data structure built at vertex $s_i$ ($1\leq i\leq l$), namely part (), which would give us, for every descendent $y\in M_{s_i}$, the highest neighbour of $y$ below $u$. Using this information we are able to derive the result of $Q(T(x), u, v)$ if $\|T(x)\| < \log n$, since in this case $T(x)\cap M = \emptyset$ and thus $T(x)\subseteq M_{s_i}$. By Lemma [\[fractional_cascading\]](#fractional_cascading){reference-type="ref" reference="fractional_cascading"} the total time of this procedure is $O(\|M_{s_i}\| + \log n)$. (b) We are left to deal with cases where $\|T(x)\| \geq \log n$. In this case, we directly compute $Q(T(x), u, v)$ using the 2D-range successor built in () which takes $O(\log n)$ time. Correctness of the query algorithm is self-evident by the algorithm. The total query time is analysed as following. Throughout an execution of Algorithm [BatchInsert]{.sans-serif}, (1) and (2) contribute at most $O(n)$ time since each $T(x)$ is involved in at most one such query $Q(T(x), u, v)$ which takes constant time. As for (3)(a), since each marked vertex $s\in M$ lies in at most one such path $(w_1, w_2, \cdots, w_t) = \mathit{path}(u, v)$, the fractional cascading data structure associated with $M_s$ is queried for at most once. Hence the total time of (3)(a) is $O(\sum_{s\in M}(\|M_s\| + \log n)) = O(n + \|M\|\log n) = O(n)$; the last equality holds by $\|M\|\leq O(n / \log n)$ due to Lemma [\[tree_partition_lem\]](#tree_partition_lem){reference-type="ref" reference="tree_partition_lem"}. Finally we analyse the total time taken by (3)(b). It suffices to upper-bound by $O(n / \log n)$ the total number of such $x$ with the property that $\|T(x)\| \geq \log n$ and $\mathit{path}(u, \mathit{par}(x))\cap M \neq \emptyset$. Let $X$ be the set of all such $x$'s. Now we prove $\|X\| = O(n / \log n)$. Build a tree $T[X]$ on vertices $X\cup \{r\}$ in the natural way: for each $x\in X$, let its parent in $T[X]$ be $x$'s nearest ancestor in $X\cup \{r\}$. Because of \[\|X\| < 2\#\text{leaves of }T[X] + \#\text{vertices with a unique child in }T[X]\] it suffices to bound the two terms on the right-hand side: on the one hand, the number of leaves of $T[X]$ is at most $n / \log n$ since for each leave $x$ it has $\|T(x)\|\geq \log n$; on the other hand, for each $x\in T[X]$ with a unique child $y\in T[X]$, by Lemma [\[struct-large-tree\]](#struct-large-tree){reference-type="ref" reference="struct-large-tree"} $\mathit{path}(x, y)\cap M \neq \emptyset$, and so we can charge this $x$ to an arbitrary vertex in $\mathit{path}(x, y)\cap M$, which immediately bounds the total number of such $x$'s by $\|M\| = O(n / \log n)$; see Figure [\[qry\]](#qry){reference-type="ref" reference="qry"} for an illustration. Overall, $\|X\| \leq O(n / \log n)$.	{'timestamp': '2018-02-21T02:08:14', 'yymm': '1607', 'arxiv_id': '1607.04913', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04913'}
null	null
# Introduction A concordance of observations support a universe whose energy budget is dominated by the unknown elements of dark energy (DE) and dark matter (DM), while baryonic matter occupies only around 5% of the total energy. Focusing on DM, all concrete evidence for its existence is solely based upon its gravitational interactions, prompting some to hypothesize alternative explanations to DM, like modified gravitational interactions (e.g., Ref ), but these alternatives are not viable in light of observations of colliding galaxy clusters where it is shown that the bulk of the clusters consists of non-luminous matter that does not interact (except gravitationally). Though there is no evidence that DM interacts through non-gravitational channels, theories beyond the standard model that incorporate DM candidates often contain a rich panoply of interactions between dark and standard model particles. In fact, DM will interact electromagnetically even if it is electrically neutral, provided that it couples to charged particles. Granted, these electromagnetic interactions are suppressed, but they are generally nonzero. Given this, a DM medium will have an optical index of refraction which is generally dispersive; i.e., the phase velocity of light in the medium is frequency dependent. In Ref. , we computed the refractive index for various particulate DM models. The forward Compton scattering amplitude links the medium's optical properties with the particle-level interaction between the photon and DM. Rather general considerations (namely, Lorentz covariance and invariance under charge conjugation, parity, and time-reversal symmetries) restrict the structure of this forward Compton amplitude at low photon energies, $\omega$. As a result, the leading order contributions to the forward scattering amplitude are model independent, attributable to the charge, mass, and magnetic dipole moment of the scatterer, and the higher order, model dependent, terms follow a known form such that the index is $n(\omega) = 1-A\omega^{-2} +B +C \omega^2 +\dots$ with each coefficient non-negative assuming small $\omega$. For electrically neutral DM candidates, the $\mathcal{O}(\omega^{-2})$ term in the index of refraction vanishes because $A = 2 q^2$ where $q$ is the electric charge of the DM. The resulting index of refraction simplifies to $n(\omega) \approx 1 + B + C \omega^2$. In principle, one can experimentally assess the coefficient $C$ through astrophysical observation. Given the normally dispersive nature of the DM in the cosmos, high energy photons will travel more slowly than ones with lower energy. If a broadband pulse of photons travels over a sufficient baseline through the DM medium, then, statistically, the arrival time of photons from that pulse will be energy dependent. For this study, a near ideal source of photons is a gamma ray burst (GRB), observable out to redshifts of $z>9$. So, if the arrival time of photons from a large sample of GRBs shows energy and baseline dependence characteristic of matter dispersion, then one can observationally assess the coefficient $C$. The brightness of bursts is a boon for measuring dispersive matter effects, but their varied spectra and lower frequency afterglows are a significant confounding factor. Because dispersion measurements rely on temporal knowledge of the emission spectra, a single GRB event cannot yet be used to constrain DM properties. But with a large number of GRB observations located at a variety of redshifts, it is expected that, statistically, random variations amongst the sources should wash out, and the expected redshift and energy dependence that indicate dispersion should survive. The Fermi Gamma-Ray Space Telescope is dedicated to the task of gamma-ray observations, so that in the future, a sufficient number of GRBs may be used to provide meaningful constraints. Regardless, from our computations of the refractive index for several neutral pointlike DM models, we expect the DM dispersion, i.e., the coefficient $C$, to be extremely small; as a result, any time lags would be immeasurable. But, the theoretical landscape is rife with DM candidates, and the work in Ref. only considered a small subset. Here, we expound upon our previous work by computing the optical dispersion of composite DM comprised of millicharge constituents. Assessing the size of DM dispersive effects is crucial to evaluating potential claims of beyond-SM physics because matter dispersion is not the only mechanism by which one can achieve such energy-dependent photon time lags. In theories with Lorentz invariance violation (LIV), the photon's dispersion relation is modified. Following Ref. , LIV effects can modify the dispersion relation for photons at an energy scale $E_\text{LIV}$ \[E^2-p^2 = \pm p^2 \left( \frac{p}{E_\text{LIV}} \right)^n \label{disp_rel}\] for some integer $n\ge 1$. If the modification in the dispersion relation in Eq. ([\[disp_rel\]](#disp_rel){reference-type="ref" reference="disp_rel"}) comes with the minus sign (rather than the plus sign), then high energy photons will lag simultaneously emitted lower energy photons. Photon time lags due to matter dispersion have a different baseline dependence relative to LIV effects (i.e., they share a different dependence upon the source's redshift $z$), but if $n=2$ in Eq. ([\[disp_rel\]](#disp_rel){reference-type="ref" reference="disp_rel"}), the two effects have a common photon energy dependence. For this reason, it is useful to know the relative magnitudes of the LIV and matter effects at a given baseline so that matter effects cannot confound potential claims of LIV gleaned from GRB photon arrival times. Because DM comprises the bulk of the matter in the universe, we will consider its impact on dispersion in detail. In Ref. , we found that the potential LIV effects would dwarf any dispersion due to various models of pointlike DM, assuming $E_\text{LIV}$ is around the Planck scale. In fact, we found that dispersion due to matter effects was irrelevant until energies around $10^{29}$ GeV. But composite DM models [@Faraggi:2000pv; @Gudnason:2006ug; @Gudnason:2006yj; @feng_hiddendm; @kaplan_da1; @Hamaguchi:2009db; @Barbieri:2010mn; @Lisanti:2009am; @cidm_alves; @quirky; @formfactor_dm; @kaplan_da2; @Hur:2007uz; @Blennow:2010qp; @DelNobile:2011je; @Frigerio:2012uc; @Cline:2012bz; @cline_da; @Kumar:2011iy; @kouvaris; @wallemacq; @Holthausen:2013ota; @Buckley:2012ky; @Higaki:2013vuv; @cline_strong; @Cline:2014eaa; @Bhattacharya:2013kma; @Boddy:2014yra; @bro; @Carmona:2015haa; @Choquette:2015mca; @Wallemacq:2014sta; @Antipin:2015xia] might prove to be more reactive, particularly if the DM is comprised of charged constituents. The motivation for composite dark matter models is varied. Some models are constructed so as to explain possible photon signals of indirect dark matter detection. Others introduce composite systems designed to smooth out the cuspiness of simulated DM galactic halo profiles. Furthermore, several models attempt to rectify seemingly contradictory results in the experimental search for DM. The DAMA/LIBRA experiment reports a statistically significant annual modulation in its detector which could be attributed to the relative motion of the detector through the galaxy's dark matter halo. Furthermore, the CoGeNT experiment reports signals above background in its detector which, if due to dark matter, would be consistent with the apparent signal from DAMA/LIBRA. If these results are due to DM interactions, they occupy a region of parameter space that has been seemingly ruled out by the CDMS-II and XENON100 experiments. To reconcile results from DAMA with null results from other experiments, the notion that DM could interact through inelastic channels has been proposed. A natural way to incorporate inelastic interactions into a model is to allow DM to be composite, rather than point-like, and a host of models take this tack as means to accommodate the DAMA or CoGeNT results in light of other DM constraints. In what follows, we will consider composite dark matter particles which are electrically neutral, but comprised of millicharged constituents. Taken en masse, we are interested in the dispersive refractive index of such a medium. As the DM is electrically neutral, it is essentially invisible to low energy photons. However, for photons which are near the threshold energy needed to transition the composite DM to an excited state, the photon interaction is substantive. Considering dark matter as a bulk medium, the interaction between dark matter and light can be characterized in terms of an electric susceptibility and index of refraction. For photon energies below the transition energy $\omega \ll \omega_0$, the medium will rather generically exhibit dispersion quadratic in the photon energy $n(\omega) \approx 1 + B + C\omega^2$ for constants $B$ and $C$ as with the pointlike DM, because the low-energy theorems of Compton scattering can be generalized to composite structures. # Index of refraction Classically, a linear dielectric medium, such as a dilute gas of dark matter, will acquire a polarization when subjected to an external electric field. The degree of polarization, or dipole moment per unit volume, can be characterized through the electric susceptibility $\mathbf{P} = \chi_e \mathbf{E}$. From the susceptibility, we can compute the medium's index of refraction $n =\sqrt{1+\chi_e}$. To compute the susceptibility of a dark-matter medium, we take a semiclassical approach in which the quantum mechanical DM system interacts with a classical electromagnetic wave via electric dipole transitions. The constituents which comprise the DM will be assumed to be effectively non-relativistic so that they can be described via the Schrödinger equation. We assume the system consists of two constituents of masses $m_{a}, m_{b}$ with electric millicharge $\pm \epsilon e$ bound via a potential $V(r)$ with $r$ the relative separation between the particles. Though we assume a pair of constituent particles, the analysis can be extended to bound states consisting of more particles if need be. Defining the reduced mass $m := m_a m_b/(m_a + m_b)$ and relative momentum $\mathbf{p} := m\dot{\mathbf{r}}=(m_b \mathbf{p}_a-m_a \mathbf{p}_b)/(m_a +m_b)$, we construct the unperturbed Hamiltonian $H_0 :=-\nabla^2/(2m) + V(r)$ for the system. To determine the interaction between the DM and light, a classical electromagnetic wave of frequency $\omega$ interacts with the quantum mechanical electric dipole moment of the DM, $\pmb{\mathpzc{p}} :=-\epsilon e \mathbf{r}$, which introduces to the Hamiltonian a perturbation, $H' =-\pmb{\mathpzc{p}} \cdot \mathbf{E}$. In the long wavelength limit, the spatial variation of the electric field is irrelevant leaving the time-dependent perturbation $H' =-\pmb{\mathpzc{p}} \cdot \mathbf{E}_0 \cos \omega t$. ## Millicharged atomic dark matter To create composite particles in a dark sector, modelers introduce a new dark gauge group which results in a binding force among the composite's constituents. The simplest gauge group is $U(1)$. If the symmetry is unbroken, then the dark photon is massless and can kinetically mix with the Standard Model (SM) photon. In such models, the particles can effectively couple to the SM photon thereby acquiring a fractional electric charge . This permits the existence of dark atoms which are overall electrically neutral, but made of constituents with electric millicharge. A neutral dark atom comprised of millicharged particles consists of two fermions $\psi_\mathbf{p}$ and $\psi_\mathbf{e}$, charged under the unbroken gauge group $U(1)'$, coupling to the dark photon with opposite charges $\pm \mathbf{e}$. \[NB: The boldface type is meant to refer to the particles and couplings in the dark sector.\] The dark "proton\" and "electron\" can form bound states under the dark Coulombic potential $V(r) = -\boldsymbol{\alpha}/r$, where we define the dark fine structure constant $\boldsymbol{\alpha} := \mathbf{e}^2/ (4\pi)$, and the relative separation between the particles is $r$. We can use non-relativistic quantum mechanics to describe this dark "hydrogen\", $\mathbf{H}$. Without loss of generality, we assume $m_\mathbf{e}\le m_\mathbf{p}$ and define the reduced mass and relative momentum as above with $m_a= m_\mathbf{e}$ and $m_b= m_\mathbf{p}$. The unperturbed Hamiltonian $H_0$ yields the usual hydrogenic eigenstates $\psi_{n\ell m}$ and energy spectrum $E_n =-\boldsymbol{\alpha}^2 m/(2n^2)$ indexed by principal quantum number $n$, a positive integer. The dark and SM sectors are coupled through photon kinetic mixing which gives rise to electric millicharges $\pm \epsilon e$ of the dark particles. Electromagnetic waves can induce transitions between dark atomic energy states, but we argue that the bulk of the dark atoms exist in the ground state. There are three main mechanisms by which the atoms can be excited beyond the ground state: dark atom self-interactions, absorption of dark photons, or absorption of SM photons. The existence of elliptical DM halos severely constrains the DM self-interaction cross section; from the limits in Ref. , the ratio of the DM self interaction to the dark atom's mass must be $\sigma/m_\mathbf{H} < 0.02$ cm$^2$/g though more recent studies have relaxed this bound to 0.1 cm$^2$/g . These limits can be satisfied either through tuning the model parameters or dilution of the dark atom component of DM. For models which satisfy this constraint, we can estimate the mean free time between collisions for dark atoms in the Milky Way's galactic halo. The mean free path can be estimated as $\lambda \sim (\sigma N)^{-1}$ where $N$ the number density of dark atoms; the number density is related to the DM mass density via $N = \rho/m_\mathbf{H}$. Then the time between collisions is $t_\text{fp} \sim \lambda/v = m_\mathbf{H} /(\sigma \rho v)$. Taking as typical parameters the local dark matter density $\rho \sim 0.3$ GeV/cm$^3$ and $v\sim 200$ km/s, we find $t_\text{fp} \gtrsim 10^{18}$ s; i.e., they are non-interacting. Dark-atom absorption of dark photons also produces excited states. The greatest energy density of dark radiation is found in the dark analog of the cosmic microwave background (CMB). Viable models require the dark radiation to be slightly cooler than the CMB , so these dark photons will not have sufficient energy to excite the dark atoms. All that remains is the interaction with SM electromagnetic waves, which we discuss below. We consider a dark atom in its ground state which can be excited by SM photons. We restrict our study to a two-state system, limiting the electromagnetic wave frequency to $\omega \lesssim \omega_0 := E_2-E_1$. In the presence of the electromagnetic wave, the Hamiltonian for the dark atom is $H= H_0 + H'$, and a general state is $\Psi(t) = c_1(t) e^{-i E_1t} \psi_1+c_2(t) e^{-i E_2t} \psi_2$, with stationary eigenstates $\psi_{1,2}$. Using the Schrödinger equation, we can develop coupled differential equations for the coefficients $c_1$ and $c_2$. These equations must be amended to account for spontaneous emission of a dark photon from the excited state. With this extra term proportional to the decay constant $\Gamma$, we have the equation for $c_2$ \[\frac{\mathrm{d}}{\mathrm{d}t} c_2(t) =-i \, (\pmb{\mathpzc{p}}_{21}\cdot \mathbf{E}_0) e^{i\omega_0t} \cos(\omega t) c_1 (t) -\Gamma c_2(t)\] where $\pmb{\mathpzc{p}}_{21} := \langle \psi_2 \| \pmb{\mathpzc{p}} \| \psi_1 \rangle$ and $\Gamma= \frac{2^7}{3^8}\boldsymbol{\alpha}^5 m$, adapted from atomic hydrogen. We solve perturbatively for the coefficients $c_{j}(t) = c_{j}^{(0)}(t)+ c_{j}^{(1)}(t)+\cdots$ with initial conditions $c_1(0)=1$ and $c_2(0)=0$. To determine the linear response of the atom to the field $E_0$, we only need the zeroth order approximation for $c_1 \approx 1$ and the first order approximation for $c_2$ \[c_2(t) \approx-\frac{1}{2} (\pmb{\mathpzc{p}}_{21}\cdot \mathbf{E}_0) \left[ \frac{e^{i(\omega_0 + \omega)t}}{\omega_0 + \omega-i \Gamma}+\frac{e^{i(\omega_0-\omega)t}}{\omega_0-\omega-i \Gamma }\right].\] The induced dipole moment for $\Psi$ is thus \[\pmb{\mathpzc{p}}(t) =-\epsilon e \langle \Psi(t) \| \mathbf{r} \| \Psi(t) \rangle =-2\epsilon e\, \mathrm{Re} [ \mathbf{r}_{12} c_1^c_2 e^{-i \omega_0 t}].\] We note that the polarization $P(t)$ will be the product of the average induced dipole moment in the direction of $\mathbf{E}_0$ and the number density $N$. Averaging over the relative orientation between $\mathbf{r}_{12}$ and $\mathbf{E}_0$, the susceptibility is \[\chi_e = N\pi\frac{2^{18} }{3^{11}} \frac{\epsilon^2 \alpha }{\boldsymbol{\alpha}^2 m^2}\frac{\omega_0}{\omega_0^2-(\omega+i\Gamma)^2}.\] Given that the DM medium is weakly interacting and dilute, $\chi_e$ is nearly zero. As such we can approximate the index of refraction as $n \approx 1 +\frac{1}{2} \chi_e$. For frequencies below the transition energy $\omega < \omega_0$, the index of refraction exhibits normal dispersion; that is, $n(\omega)$ increases with $\omega$. In fact, for $\omega \ll \omega_0$, the dispersion is quadratic in frequency \[\mathrm{Re}(n) \approx 1 + N\pi\frac{2^{20} }{3^{12}} \frac{\epsilon^2 \alpha }{\boldsymbol{\alpha}^4 m^3}\left( 1 + \frac{\omega^2}{\omega_0^2}\right), \label{adm_index}\] neglecting the small term proportional to $\Gamma^2$. ## Other millicharged composite particles [\[sho_sect\]]{#sho_sect label="sho_sect"} Millicharged atomic DM represents only a fraction of proposed composite DM models. Many models posit composite states strongly bound by a nonabelian gauge force, some of which contain electrically charged constituents. The details of these strongly composite systems with charged constituents are model dependent, so we will only sketch an approach as to how one would estimate the susceptibility of such a DM medium. As with the atomic DM system, we will assume a composite state consisting of two particles with effective (dressed) masses of $m_a$ and $m_b$ with electric millicharges $\pm \epsilon e$. To model a tightly bound system, we approximate their interaction potential as that of a harmonic oscillator, $V(r) = \frac{1}{2}m\omega_0^2 r^2$ with $m$ the reduced mass. As above, we will only consider a two state system consisting of the ground state and first excited state; the energy difference between these states is $\omega_0$. Our naïve assumptions result in a simplistic model, yet for a system as complex as a nucleon, the SHO potential adapted to three constituent quarks yields an order of magnitude estimate of the nucleon polarizability. With this new potential, we merely need to compute the transition dipole moment $\mathpzc{p}_{21}$ and the decay constant $\Gamma$. We find $\mathpzc{p}_{21} = \epsilon e/\sqrt{2m \omega_0}$ and $\Gamma =\epsilon^2 \alpha \omega_0^2/(3m)$. Given our assumption of nonrelativistic QM, $\Gamma/\omega_0$ is small so that we can approximate the susceptibility of this DM medium as \[\chi_e = \frac{4}{3} \pi N \frac{\epsilon^2 \alpha }{ m \omega_0^2}\left( 1 + \frac{\omega^2}{\omega_0^2}\right).\] Finally, because the susceptibility is small, the refractive index can be approximated by $n \approx 1 + \frac{1}{2} \chi_e$ for $\omega \ll \omega_0$. # Observational consequences for GRB photons As a broadband pulse of electromagnetic radiation travels through a dispersive medium, the pulse shape spatially broadens because the phase speed of each component wave is frequency dependent. The DM medium we consider is normally dispersive, $n \approx 1 + B + C \omega^2$, so higher frequency components of the pulse will lag the lower frequency components. Over large distances, a time lag can accrue between these two components. Gamma-ray bursts are apt photon sources for dispersive studies because they are explosive events that occur over short time scales and their brightness permits observation over cosmological distances. Because cosmological distances are involved, we must account for the universe's expansion as a light pulse travels from source to observer. There are three factors that need to be considered. First, the light-travel time between source and detector is dependent upon the redshift of the source and the local expansion rate. Second, at redshift $z$ the number density of dark matter increases by a factor of $(1+z)^3$; a factor of present day DM number density $N$ is contained in the coefficient $C$. Finally, as we look into the past, the wavelength of light blue shifts relative to its value $\omega$ at the detector (at $z=0$). Incorporating these three factors, the time lag accrued between (detected) frequencies $\omega_\text{hi}$ and $\omega_\text{lo}$ simultaneously emitted from a source at redshift $z$ becomes \[\tau \approx C\left( \omega_\text{hi}^2-\omega_\text{lo}^2 \right) \int_0^z \frac{(1+z')^5 \mathrm{d}z'}{H(z')}, \label{matter_lag1}\] where the Hubble expansion rate at redshift $z'$ is $H(z') = H_0 \sqrt{(1+z')^3\Omega_M + \Omega_\Lambda}$, assuming a simple $\Lambda$CDM cosmology. If the two photon energies are well separated, we can neglect the low energy term in the difference in Eq. ([\[matter_lag1\]](#matter_lag1){reference-type="ref" reference="matter_lag1"}) and set $\omega = \omega_\text{hi} \gg \omega_\text{lo}$ so that \[\tau \approx C\omega^2 K_5 (z), \label{matter_lag2}\] where we define the integral over the baseline $K_j (z) := \int_0^z \frac{(1+z')^j \mathrm{d}z'}{H(z')}$. We employ the cosmological parameters in the 2015 data release from the Planck satellite. The Hubble constant today is $H_0 = 67.8 \pm 0.9 \,\hbox{km/s/Mpc}$, whereas the fraction of the energy density in matter relative to the critical density today is $\Omega_M = 0.308\pm 0.012$. For the simple $\Lambda$CDM model, the universe is flat so that the corresponding fraction of the energy density in the cosmological constant $\Lambda$ is $\Omega_\Lambda = 1-\Omega_M$. The high energy photons from distant GRBs have already been used to place lower bounds on the scale at which Lorentz-invariance violating effects could modify the photon's dispersion relation [@AmelinoCamelia:1997gz; @ellis; @Boggs:2003kxa; @Ellis:2005wr; @jacobpiran; @ellis2; @Ellis:2009yx; @AmelinoCamelia:2009pg; @Vasileiou:2013vra]. For some LIV models, the modified dispersion relation acquires terms quadratic in the photon energy, i.e., $n=2$ in Eq. ([\[disp_rel\]](#disp_rel){reference-type="ref" reference="disp_rel"}). For these models, the following time lag can accrue between the low and high energy photons emitted by a GRB \[\Delta t_\text{LIV} \approx \frac{3}{2} \left(\frac{\omega}{E_\text{LIV}}\right)^2 K_2(z). \label{t_liv}\] where the (present day) photon energy $\omega = \omega_\text{hi} \gg \omega_\text{lo}$ was emitted at redshift $z$. The energy $E_\text{LIV}$ characterizes the scale at which LIV effects become appreciable. Data from GRBs and AGNs have placed limits on this scale (for quadratic dependence) to be $E_\text{LIV} > 1.3 \times 10^{11}$ GeV. We wish to assess the size of dispersive matter effects relative to those attributable to LIV. From Eqs. ([\[matter_lag2\]](#matter_lag2){reference-type="ref" reference="matter_lag2"}) and ([\[t_liv\]](#t_liv){reference-type="ref" reference="t_liv"}), we will consider the ratio $\tau/\Delta t_\text{LIV}\sim \frac{2}{3} C E_\text{LIV}^2 K_5(z)/K_2(z)$ for millicharged composite DM candidates. The factor $C$ is dependent on the DM model, whereas the ratio of the $K_j$ integrals depends on the redshift of the source with $K_5(z)/K_2(z)\sim \mathcal{O}((1+z)^3)$. The GRB constraint in Ref. is derived from GRB 090510 which is located at a reshift of $z=0.903\pm 0.003$ with a high energy photon $\omega_\text{hi}=30.53^{+5.79}_{-2.56}$ GeV detected 0.829 s after the trigger of the GRB monitor . In this case, the ratio of integrals $K_j$ is $\mathcal{O}(10)$. On the other hand, there have been two GRB observations with confirmed redshift $z>8$: GRB 090423 with $z=8.3\,$ and GRB 090429B with $z=9.4$. For these, we find the maximum value of $K_5/K_2 \sim \mathcal{O}(10^3)$. Thus, matter effects are enhanced dramatically relative to LIV effects for higher redshift sources. As stated previously, we work within the context of a simple $\Lambda$CDM model; dark energy is attributed to a cosmological constant with an equation of state $w=-1$, where $w$ is the ratio of DE's pressure to energy density. This is consistent with the 2015 Planck data which, along with external astrophysical data, determines $w=-1.006 \pm 0.045$, if a constant $w$ is assumed . However, when the Planck data are combined with weak lensing data, cosmologies with a time-dependent equation of state are at least marginally preferred. Extensions of the simple $\Lambda$CDM model will affect the expansion rate of the universe, $H(z')$, and thus the integrals $K_j(z)$. The impact of cosmology upon GRB photon and neutrino time lags due to LIV was previously considered in Refs. , where the authors compare results from a simple $\Lambda$CDM model to quintessence, Chaplygin gas, and braneworld cosmologies. Amongst the models, differences in time lags exist which could affect the measurability of the effect or the interpretation of a measurement if the cosmology is unknown. Using the models considered in Refs. , we determine the impact of cosmology upon the relative* size of dispersive matter effects with those from LIV. We find the ratio of integrals $K_5(z)/K_2(z)$ varies little (less than $5\%$) out to $z=10$ for the $\Lambda$CDM, quintessence, Chaplygin gas, and braneworld models. On the other hand, for the variable quintessence model, we find that $K_5(z)/K_2(z)$ is commensurate in size with this ratio in the other models up to $z=2$, but beyond this redshift, the ratio of integrals approaches an asymptotic value around 18 (while the ratio scales as $\mathcal{O}((1+z)^3)$ for the other models). For GRBs at a distant redshift of $z=10$, the difference between the variable quintessence model and the others will be a factor of 25. On the face of it, this difference is substantial and could perhaps be even larger for different models of quintessence; however, when we examine the size of DM dispersive effects below, we will see that the choice of cosmology is of subleading significance. Given this, we opt to use the simple $\Lambda$CDM model, though acknowledge that other cosmologies are likely and will quantitatively affect our results to a degree. ## Millicharged atomic dark matter We now examine the size of the dispersive coefficient $C$ in the refractive index relative to $E_\text{LIV}$. This coefficient sets the energy scale, $C^{-\frac{1}{2}}$, at which DM dispersion becomes appreciable. For atomic dark matter, we find from Eq. ([\[adm_index\]](#adm_index){reference-type="ref" reference="adm_index"}) this coefficient to be $C= N\pi\frac{2^{20} }{3^{12}} \frac{\epsilon^2 \alpha }{\boldsymbol{\alpha}^4 m^3 \omega_0^2}$. The DM number density is $N = \rho/m_\mathbf{H}$ where the mass of the dark atom is $m_\mathbf{H} = m_\mathbf{p} + m_\mathbf{e}-\frac{1}{2}\boldsymbol{\alpha}^2 m$. For propagation of light across cosmological distances, the average DM density is $\rho \simeq 1.25\times 10^{-6}$ GeV/cm$^3$. As a benchmark, it is useful to estimate the value of the coefficient for various atomic dark matter models. In Ref. , the best-fit model employs a heavy dark proton $m_\mathbf{p} = 650$ GeV and much lighter dark electron $m_\mathbf{e}= 0.426$ MeV with an electric millicharge $\epsilon = 6.7\times 10^{-5}$ and a dark fine structure constant which takes the SM value, $\boldsymbol{\alpha} = \alpha$. For this model, the dispersive term is $C = 1.8 \times 10^{-25}$ GeV$^{-2}$. As a consequence, the energy scale at which the DM dispersion is appreciable is $C^{-\frac{1}{2}} = 2.4 \times 10^{12}$ GeV, which is commensurate with the limit on $E_\text{LIV}$ from Ref. . On the face of it, it would seem that for large $z$ the $K_5$ integral could make dispersion due to millicharge atomic DM competitive in magnitude with potential LIV dispersive effects, but we must recall one crucial point--the binding energy of the dark atom. Our estimate of the DM refractive index is only valid in the limit in which observed photon energies are much smaller than the energy difference between the ground and excited state of the dark atom, $\omega \ll \omega_0 = \frac{3}{8} m \boldsymbol{\alpha}^2$. The dark electron in Ref. has a mass similar to the SM electron, so the ionization energy of the dark hydrogen is on the scale of electronvolts with $\omega_0 = 8.5$ eV. As such, the limits derived from high energy GRB photons are not relevant for this model of atomic dark matter because the dark electrons are so weakly bound. Let us now consider another model of atomic dark matter which employs very different masses and electric millicharge. Reference has an atomic DM model which can account for the CoGeNT experimental results. The authors' dark atom consists of massive particles $m_\mathbf{p} = m_\mathbf{e} = 3$ GeV with a slightly larger dark fine structure constant $\boldsymbol{\alpha}= 0.062$ and much larger millicharge $\epsilon = 10^{-2}$. In this case, the relevant energy scale for matter dispersive effects is $C^{-\frac{1}{2}} = 6 \times10^{21}$ GeV with transition threshold of $\omega_0 = 2.2$ MeV. Due to the larger masses and dark fine structure constant, the threshold energy $\omega_0$ rises, allowing the dark matter to be probed with more energetic photons, but for such massive models, the energy scale at which matter effects become appreciable is well beyond the current bounds on $E_\text{LIV}$. These two specific examples are exemplars of a general trend. Namely, for light DM masses, the energy scale at which matter dispersion is operative is commensurate with limits on the LIV energy scale, but the binding energy is too small to actually probe such dispersion. For more massive models, the threshold energy $\omega_0$ increases, but $C^{-\frac{1}{2}}$ increases at a greater rate. To confirm these generalities, we explore more fully the allowed parameter space of exotic millicharged dark atoms without regard to their feasibility as a dark matter candidate. In particular, we compute the dispersive energy scale $C^{-\frac{1}{2}}$ for a range of dark electron and proton masses subject only to the provision that the existence of the constituent millicharged particles have not been excluded through other considerations. There are, in fact, stringent constraints on electric millicharge. Figure 1 of Ref. shows a current summary of these constraints for particle masses ranging from 100 eV to 100 TeV. Stellar evolution severely constrains electric millicharge $\epsilon < 2\times 10^{-14}$ for masses below 10 keV. Around 100 keV to 1 GeV, BBN and CMB constraints upon the light degrees of freedom limit electric millicharge $\epsilon \lesssim 10^{-9}$. For masses between 1 MeV and 100 GeV, collider constraints limit $\epsilon < 0.2$. Dispersive effects scale as $\epsilon^2$, so they will be maximized for the largest allowed $\epsilon$. To explore the largest possible matter dispersion, we choose, for a given particle mass, the millicharge which saturates the bounds in Fig. 1 of Ref. . Setting the dark fine structure constant equal to the SM value $\boldsymbol{\alpha} = \alpha$, we plot in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"} contours representing the energy scale $C^{-\frac{1}{2}}$ for various dark electron and proton masses. These contours are superimposed upon filled contours which represent the threshold energy $\omega_0$. For masses ranging from 100 eV to 1 GeV, we find the energy scale charactering matter dispersion range from $10^9$ GeV to $10^{21}$ GeV, spanning twelve decades of energy, while $\omega_0$ ranges from meV to keV, spanning six decades. The broad scan of parameter space depicted in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"} confirms our preliminary conjecture. We see rather generally that if the energy scale at which atomic DM dispersive effects are near current limits on $E_\text{LIV}$ then the binding energy of the dark atom is sub-eV. The photon energy needed to probe the DM dispersion would ionize the dark atoms. For higher threshold energies, matter dispersive effects quickly become irrelevant. Supposing a LIV scale around the Planck mass, $\sim 10^{19}$ GeV, the ionization energy of the dark atom is on the order of keV. Thus, in a search for LIV dispersive effects, dispersive effects due to atomic dark matter are not a confounding source of background. ## Other millicharged composite particles We now turn to our millicharged composite system bound under the harmonic oscillator potential. There is wide variation amongst the strongly bound composite DM models, but we will explore the dispersive matters effects for a limited range of particle mass and oscillator energy $\omega_0$. Again, to determine the maximal effect, we assume electric millicharge values which saturate the bounds in Fig. 1 of Ref. . From above, we find the dispersive coefficent to be $C = \frac{2}{3} \pi N \frac{\epsilon^2 \alpha }{ m \omega_0^4}$. We will assume that the constituent millicharged particles have the same effective mass $m_a = m_b$ so that the reduced mass is one-half this mass $m = \frac{1}{2} m_a$ and the total mass of the system is roughly $2m_a$. With these simplifying assumptions, the dispersive energy scale depends rather simply upon the composite parameters, $C^{-\frac{1}{2}} \sim m_a \omega_0^2/\epsilon$. As a result, this energy is lowest for masses $m_a$ in the GeV to 100 GeV region where $\epsilon \lesssim 0.2$. In Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}, we plot the dispersive energy scale as a function of the mass $m_a$, choosing $\omega_0 =1$ keV. The threshold energy could take a host of values, but we choose 1 keV because inelastic DM models with an excitation scale of order $\mathcal{O}(1-100$ keV) can accommodate the DAMA anomalous experimental results or potential indirect detection signals. Given that $C^{-\frac{1}{2}}$ is proportional to $\omega_0^2$, it is not difficult to extrapolate these results to different threshold values. Referring to Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}, we find that the lowest dispersive energy scale, $C^{-\frac{1}{2}}= 1.7 \times 10^{13}$ GeV, occurs for a millicharged particle mass of 1.3 GeV. This energy is a few decades above the current limit on the LIV scale, but over long baselines the size of matter dispersive effects could rival those due to LIV. Of course, unlike LIV, the quadratic dispersive terms is only relevant for photon energies below the threshold of 1 keV. This difference between the threshold and dispersive energy scales makes the matter dispersive effects immeasurable. Indeed, for a nearby GRB, $z\sim 1$, the time lag between keV and lower energy photons is on the order of $10^{-20}$ s. Given the same millicharged particle mass of 1.3 GeV, a dispersive scale near the Planck mass, $C^{-\frac{1}{2}} \sim 10^{19}$ GeV, involves a threshold energy near 1 MeV, but probing the DM with near threshold photons results in the same time lag for the $\omega_0 = 1$ keV case. As with atomic dark matter, whenever the energy scale of the dispersive coefficient is commensurate with the LIV scale, the threshold energy of the composite system is too small to probe this scale. Though one has some freedom to tune the threshold of the SHO model independent of the millicharged particle properties, this freedom is not sufficient to construct a composite DM candidate that can have measurable dispersive effects competitive with potential LIV effects. ## Baryonic matter One potential source of background to the dispersive matter effects from composite millicharge DM is ordinary baryonic matter which, for the most part, consists of the hydrogen and helium produced in big-bang nucleosynthesis. Dispersion due to baryonic matter occurs on varying scales, the first being the atomic scale, which is relevant for photon energies in the eV range. As we discussed above, the probes of LIV are high-energy (MeV or greater) photons that would ionize the hydrogen and helium of baryonic matter, so these cannot be confounded with LIV effects. Beyond the atomic energy scale, a medium of hydrogen and helium will effectively appear as a plasma to high energy photons, and the plasma's optical properties will be predominantly determined by its electron component. Such a medium possesses a dispersive index of refraction of the form $n \approx = 1-A\omega^{-2}$ with $A= e^2 N/(2m_e)$ where $N$ is the number density of the electrons. This does not have the $\mathcal{O}(\omega^2)$ dispersive behavior typical for a neutral scatterer, so it cannot be confused with dispersion due to DM or LIV. Furthermore, for a low density cloud of atoms, dispersion is negligible for high energy photons because $n-1 \sim 1/\omega^2$. For photon energies approaching the nuclear scale, there are $\mathcal{O}(\omega^2)$ dispersive effects to consider arising from the polarizabilities of the nucleons, but this physics is reasonably well understood both from an experimental and theoretical standpoint. As a result, the impact of baryonic dispersion can be studied in detail, but here we opt to estimate the effect by combining measured values of a nucleon's electric polarizability with our SHO model for strongly bound systems in Sec. [\[sho_sect\]](#sho_sect){reference-type="ref" reference="sho_sect"}. From Ref. , we find electric polarizabilities for the proton and neutron, $\alpha_\text{E}^p=11 \times 10^{-4}$ fm$^3$ and $\alpha_\text{E}^n=12 \times 10^{-4}$ fm$^3$, so for a nucleon we take $\alpha_\text{E}^{N}=12 \times 10^{-4}$ fm$^3$. The small size of the polarizability indicates a tightly bound system, and we expect dispersive effects to be small. With our SHO model for dispersion, we find $n \approx 1 + 2\pi N \alpha_\text{E}^N(1 + \omega^2/\omega_0^2)$ where $N$ is the number density of nucleons; we estimate the excitation energy to be $\omega_0 \approx 300$ MeV. From the Planck 2015 data, the baryonic component of the universes's energy budget is $\Omega_b =0.048$ so that number density of nucleons is $N = 2.5 \times 10^{-7}$ cm$^{-3}$. This renders a dispersive term of $n-1 \approx (2 \times 10^{-48})\, \omega^2/\omega_0^2$. For photon energies well below $\omega_0$, the dispersive energy scale is $C^{-\frac{1}{2}} \sim 2\times 10^{26}$ MeV. In short, this is not competitive with LIV probes and likely immeasurable. But, even if a suspected baryonic signal arose from a photon time lag, one can assess the presence of hydrogen and helium along the line of sight through the absorption spectrum that will arise in the optical afterglow of the burst. This absorption spectrum can aid in determining the presence of atomic scale physics, e.g., a Lyman-alpha system, differentiating the matter from millicharged DM. # Conclusion Models of composite dark matter represent an attractive alternative to simple pointlike DM candidates. Through their inelastic interactions, one can account for potential DM direct and indirect detection signals, produce more realistic DM galactic halo profiles, and provide a natural explanation for the relative abundance of dark to baryonic matter. In addition to these physical motivations, composite systems are aesthetically pleasing because the dark sector mirrors some of the complexity of the SM sector. In this work, we focused upon composite systems which consist of electrically charged constituents. Such particles naturally couple to the electromagnetic field via an electric dipole moment, rendering the cosmos with a dispersive optical index of refraction. We computed this index of refraction for atomic DM and more strongly bound composite systems, modeled through a harmonic oscillator potential. Given the dispersive nature of the refractive index, higher energy photons will lag those with lower energy as they travel through the universe from a distant GRB. A time lag accrued over long baselines could be confused with similar dispersive effects which result from theories of LIV. However, we found that, for both atomic and more strongly bound composite DM, whenever the energy scale of matter dispersion is commensurate with that of LIV dispersion, the threshold energy $\omega_0$ is too small to actually probe the matter dispersive effects. As a result, potential claims of LIV achieved through time lags cannot be confused for matter dispersion due to composite DM comprised of charged constituents.	{'timestamp': '2016-07-18T02:08:35', 'yymm': '1607', 'arxiv_id': '1607.04577', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04577'}
null	null
null	null
	{'timestamp': '2018-01-15T02:09:57', 'yymm': '1607', 'arxiv_id': '1607.04532', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04532'}
null	null
# Introduction {#sec:introduction} Inference of the underlying point sources in photon count maps is a recurring problem in astronomy. Potential challenges include poorly known backgrounds, detector noise, shot noise from faint or modeling degeneracies from overlapping point sources. The resulting symptom is that of flux incompleteness, where the faintest sources are not resolved, but instead absorbed into the diffuse background prediction. Hence the flux distribution of an incomplete catalog exhibits an unphysical roll off at the faint end, even when there are yet fainter true point sources in the image. A commonly applied approach to point source inference is the frequentist method of asking whether an additional point source increases the maximum likelihood compared to the null model, i.e., without the point source. Iteratively performing this exercise over the image while potentially floating other parameters such as the background normalization, one can produce a map of delta log-likelihood, which can be used to identify features in the count data that are unlikely to come from a diffuse emission component. A standard in gamma-ray astronomy is to reject model point sources that yield a test statistic below 25, which, for a $\chi^2$ distribution with four degrees of freedom (spatial coordinates, flux and spectral index), corresponds to just above 4$\sigma$. This approach is computationally cheap and has been the standard algorithm to reduce full-sky maps to catalogs such as the 3FGL catalog of point sources in the gamma-ray sky. In the limit where an image is populated by well-separated point sources, this method can capture the spatial and spectral uncertainties of individual sources in the form of ellipsoids assuming Gaussian covariance. However as sources start to overlap, covariances between positions and spectra of model point sources cannot be captured by whether it is favorable to reject the null hypothesis for a single source. Instead, there can be multiple sources consistent with the observed data with a potentially complicated spatial and spectral covariance. Generalization of the frequentist approach to pairs and even multiplets of point sources can in principle probe such covariances between point source positions and fluxes. However the computational complexity quickly increases before a significant fraction of the parameter space volume can be explored. Also, the best fit delta log-likelihood comparison of models with different numbers of point sources becomes ill-defined since a point source model can fit the data at least as well as another with fewer sources. Therefore one needs to balance the goodness of fit of a point source model with its predictivity. If this type of across-model comparison is not properly handled, the maximum likelihood catalog will either blend point sources or over-fit the image by introducing spurious point sources. Traditional (deterministic) catalogs can reduce large amounts of observations to relatively concise lists of point sources, precluding false positives with the use of a hard significance cut that discards subthreshold information. However it is important to keep in mind that catalogs are still models that describe our state of knowledge consistent with the given data up to statistical and systematic errors. Especially in the crowded field limit, the uncertainties on the number, localizations and spectra of candidate point sources are more complicated than the usually adopted Gaussian form. Therefore keeping a fair ensemble of realizations of the underlying catalog space properly propagates the uncertainties to subsequent analyses that rely on the catalog. This is in contrast with the frequentist approach of representing our state of knowledge about the point sources with an estimator of the most likely catalog. These concerns make a case for adopting a Bayesian approach to point source inference. In this paper we construct a Bayesian framework, Probabilistic Cataloger (`PCAT`), to perform probabilistic point source inference. In this setting, the hypothesis space is the union of emission models that have a number of point sources between $N_{min}$ and $N_{max}$, along with parameters characterizing the diffuse emission and the PSF. Therefore the number of point sources in a member model, $N$, itself becomes a discrete parameter of the metamodel, i.e., top-level model. We then sample from the posterior of the metamodel given the observed data by implementing the necessary transdimensional proposals to jump across models. This precludes the necessity to run separate MCMC simulations for each model in order to estimate the Bayesian evidence, which is subject to large uncertainties. When marginalized over all other parameters, the posterior distribution of the model indicator, i.e., the number of point sources in the model, can be used to calculate the relative evidence (the Bayes factor) between models. Given that detailed balance is ensured at each across-model proposal, models with too many point sources not justified by the data are less frequently visited, since most of the added parameter space is wasted, i.e., inconsistent with the data. Therefore the resulting Bayes factor penalizes models for the loss of predictivity as well as goodness of fit. Our work inherits elements from and builds on probabilistic cataloging discussed in. The resulting statistical model is hierarchical in the sense that the hierarchical priors we place on the point source parameters, e.g., positions, fluxes and colors, are parametrized by a small set of hyperparameters, which, in turn, admit hyperpriors. During the evolution of the MCMC state, the sampler can also propose to change the hyperparameters by respecting the imposed prior. Because they parametrize the prior distribution of the point source parameters, the posterior distribution of the hyperparameters encode our state of knowledge about the population characteristics. This allows one to statistically probe the source count function below the detection threshold of traditional catalogs. In general, there are two distinct questions that can be posed about the underlying distribution of point sources in a given image: - What is the flux distribution of the most significant $N$ point sources? - What is the number of point sources above a given minimum flux, $f_{min}$? For small values of $f_{min}$, traditional catalogs, by construction, can only address the first question, whereas a probabilistic catalog can provide an answer to both questions after proper marginalization of the posterior samples. In principle, it is also possible to float the hyperparameter, $f_{min}$, during the generation of a probabilistic catalog. However, in the limit of arbitrarily small $f_{min}$, diffuse origin hypothesis is nearly degenerate with that of a population of unresolved point sources. Therefore, our hierarchical model becomes insensitive to changes in $f_{min}$, when $f_{min}$ becomes much smaller than the typical fluctuations in the image. We pay particular attention to how we choose $f_{min}$, which will be discussed in Appendix [\[sect:info\]](#sect:info){reference-type="ref" reference="sect:info"}. Another approach to probing the source count function at the faint end is the fluctuation analysis where the 1-point probability distribution function of the emission is used to estimate the contribution of unresolved point sources to the total emission. By modeling the tail of the distribution of deviations due to unresolved point sources just below the detection threshold, fluctuation analysis can distinguish truly isotropic emission from unresolved point sources. The method has been considered across the whole electromagnetic spectrum, e.g., in radio, Far Infrared (FIR), x-rays and gamma-rays. After introducing the Bayesian framework, we use our formalism to construct a probabilistic catalog of the gamma-ray sky in the North Galactic Polar Cap (NGPC). We then compare our results with the traditional 3FGL catalog published by the Fermi-LAT Collaboration and show that our median catalog agrees with the 3FGL as well as revealing low-significance point sources. Nevertheless, the real benefit of constructing a probabilistic catalog becomes apparent in crowded regions such as the inner regions of the Milky Way. Given the intriguing possibility that the inner galaxy GeV excess could be due to unresolved point sources, we leave the application of probabilistic cataloging to the inner galaxy data, to a dedicated future work. The rest of the paper is structured as follows. We begin in Section [2](#sect:modl){reference-type="ref" reference="sect:modl"} by introducing our emission model, leading into a discussion of our hierarchical inference framework. Then, in Section [3](#sect:samp){reference-type="ref" reference="sect:samp"} we present the sampling method we use for probabilistic cataloging. We present our results on mock data and the NGPC region in Sections [4](#sect:mock){reference-type="ref" reference="sect:mock"} and [5](#sect:ngal){reference-type="ref" reference="sect:ngal"}, respectively. We then provide a discussion of our results in Section [6](#sect:disc){reference-type="ref" reference="sect:disc"} and conclude in Section [7](#sect:conc){reference-type="ref" reference="sect:conc"}. # The hierarchical Bayesian model {#sect:modl} ## Modeling photon emission The emission from a point source can be modeled as a delta function in position space with some parametrized energy spectrum, convolved with the spatial and spectral instrument response, i.e., the point and line spread functions, of the measuring instrument, respectively. Given that the Fermi-LAT energy resolution of $\sim 10\%$ is smaller than our energy bin width of $\sim 100\%$, we assume infinite energy resolution. The delta function at the position of each point source gets convolved with the PSF, $\dd{\mathcal{F}}/\dd{\Omega}$, in units of the fraction of total flux per solid angle. This yields the model point source flux map, $\mathcal{P}_{im}$, in energy bin $i$ and PSF class $m$, with units of 1/cm$^2$/s/sr/GeV, when summed over all point sources. \[\begin{aligned} \mathcal{P}_{im} (l, b) = \sum_{n=1}^N & \dv{\mathcal{F}_{im}}{\Omega} (l_n, b_n) f_{ni} \end{aligned}\] Here, $f_{ni}$ is the flux of the $n^{th}$ point source in the $i^{th}$ energy bin. We assume a power law for the spectral energy distribution so that \[\begin{aligned} f_{ni} = f_n \Bigg(\frac{E_i}{E_0}\Bigg)^{-s_n}, \end{aligned}\] where $E_0=1.7$ GeV is the pivot energy, $f_n$ and $s_n$ are the normalization and power law index of the spectral flux of each point source. At $\sim$ GeV energies, simply modeling the spectra of the point sources using a power law is not an accurate description of the data. Galactic gamma-ray emitters such as pulsars are known to exhibit an exponential cutoff in their spectra at $\sim 1-10$ GeV. Nevertheless, since our Region of Interest (ROI) in this work is restricted to the NGPC, most of the sources are expected to be extragalactic. Moreover, extragalactic gamma-ray sources such as blazars and radio galaxies have curved spectra, i.e., a power law with a running index, due to their broadband inverse Compton emission on the external radiation field of the Active Galactic Nucleus (AGN). To address this concern we restrict the energy span to 1.5 decades, i.e., between 0.3 GeV and 10 GeV, where curved spectra can be approximated using a single power law. We do not attempt the full probabilistic regeneration of the 3FGL, which uses data between 100 MeV and 100 GeV by relying on a spectral model with a larger number of free parameters. For a given ROI, there may also be emission from diffuse or extended sources. We therefore include into our model the diffuse emission prediction provided by the Fermi-LAT Collaboration, which accounts for Inverse Compton Scattering (ICS) emission due to upscattering of star light by relativistic electrons as well as gamma-rays from pion decay and bremsstrahlung due to the interaction of cosmic rays with gas and dust. At high galactic latitudes such as in the NGPC, where the latter process dominates, the model is roughly proportional to the ISM column density as traced by dust. In addition to the spatially varying diffuse emission model, we also add an isotropic template into our background prediction. This component serves two purposes. First, it models isotropic emission, whether of cosmic or instrumental origin. Second, it absorbs potential emission from point sources below the inclusion flux limit. With the addition of the background emission, the total model becomes \[\begin{aligned} \mathcal{M}_{im} = \mathcal{P}_{im} + \mathcal{D}_i + \mathcal{I}_i. \end{aligned}\] We sample $\mathcal{M}_{im}$ over a `HealPix` grid of resolution $n_{side} = 256$. The normalization of both templates are allowed to float in all energy and PSF classes and admit log-uniform priors. Ideally, the diffuse model prediction, $\mathcal{D}_i$, should be smoothed to match the PSF kernel of the data in each PSF class, $m$. We did not need to perform this operation, however, given the fact that the diffuse model does not have bright features in the NGPC. In order to be able to marginalize over uncertainties in the background prediction, we allow the normalizations of the background templates to float in each energy bin. We will denote the normalization parameters with the parameter vector $\vec{A}$ in what follows. Finally, in a photon counting experiment, the likelihood function is the Poisson probability of observing $k_{ijm}$ counts given a mean of $\mu_{ijm}$ counts in the $i^{th}$ energy bin, $j^{th}$ pixel and $m^{th}$ PSF class. We then sum the log-likelihood over all pixels, energy bins and PSF classes to construct a joint log-likelihood across different energy bands and PSF classes. ## Modeling the population characteristics of point sources In general, the set of prior beliefs about the statistical behavior of a model parameter, $\theta$, can be encapsulated in its prior probability distribution, $P(\theta)$. However, in this work we express the prior on the model parameters using a hierarchical structure, which requires a distinction between different levels of prior belief. By using the word prior, we refer to the first level in the hierarchy. This includes the prior belief that all model point source fluxes are drawn from a power-law with index $\alpha$ and that the Poisson mean of the number of point sources, $N$, is $\mu$. For this work on the NGPC, we assume that the point source positions are drawn from the uniform distribution and that the colors have a Gaussian distribution. We further assume that the model point sources have vanishing n-point spatial or spectral correlations and are independent and identically distributed realizations of an underlying population. The prior probability distribution of the number of point sources, $N$, is given by \[P(N\|\mu) = \frac{\mu^N}{N!}e^{-\mu}. \label{equa:numbpnts}\] The hyperparameter $\mu$ is taken to be log-uniform distributed such that \[P(\mu) = \dfrac{1}{\ln \mu_{max}-\ln \mu_{min}} \dfrac{1}{\mu}\] for $N_{min} < \mu < N_{max}$ and vanishes otherwise. This choice yields a scale-free prior on the number of point sources. Note that we use the same notation to refer to probability densities of continuous variables, e.g., $P(\mu)$, and probabilities of discrete variables, e.g., $P(N)$. When the parameter is discrete, the notation implies probability, whereas it refers to the probability per differential interval in the parameter, if the parameter is continuous. Furthermore, we assume that the flux of the $n^{th}$ point source in the pivot energy bin 1-3 GeV, $f_n$, is distributed as a power law between some $f_{min}$ and $f_{max}$ with the index $\alpha$ at the central flux bin of $\sim 10^{-9}$/cm$^2$/s/GeV. \[\begin{aligned} P(f_n\|\alpha) = \begin{cases} \dfrac{1-\alpha}{f_{max}^{1-\alpha}-f_{min}^{1-\alpha}} f_n^{-\alpha} \text{ for } f_{min} < f_n < f_{max} \\ 0 \text{ otherwise} \end{cases}. \label{equa:flux} \end{aligned}\] We place a uniform prior on the angle described by this slope, which yields \[P(\alpha) = \dfrac{1}{\tan^{-1}(\alpha_{max})-\tan^{-1}(\alpha_{min})}\dfrac{1}{1+\alpha^2}\] for $\alpha_{min} < \alpha < \alpha_{max}$ and 0 otherwise. ## PSF modeling {#sect:psfn} A particle detector inevitably introduces errors when the arrival direction of a photon event is reconstructed. At small angular deviations, the random scatter around the true arrival direction can be modeled as a Gaussian \[\begin{aligned} \mathcal{G}(\theta_0; \sigma) &= \frac{1}{\sqrt{2\pi \sigma^2}} \exp\Bigg(-\frac{\theta^2_0}{2\sigma^2}\Bigg), \end{aligned}\] where $\theta_0$ is the angular distance from the true direction. The variance of the Gaussian is the sum of variances due to the resolution of the silicon strips and multiple scattering, which itself scales with energy. At large angular deviations, however, the Fermi-LAT PSF instead follows a power-law. A convenient function that approximates a Gaussian at small deviations while converging to a power law at large deviations is the King function \[\begin{aligned} \mathcal{K}(\theta_0; \sigma, \gamma) &= \frac{1}{\sqrt{2\pi \sigma^2}} \Bigg(1-\frac{1}{\gamma}\Bigg) \Bigg(1 + \frac{\theta_0^2}{2\sigma^2\gamma}\Bigg)^{-\gamma}, \end{aligned}\] where the Gaussian scale is given by $\sigma$ and the power law slope is set by $\gamma$. The PSF modeling of the Fermi-LAT Collaboration uses a weighted sum of two such functions. Therefore, there are $5N_eN_{psf}$ independent parameters in the model, where $N_e$ and $N_{psf}$ are the number of energy bins and PSF classes. We use this model when we do not float the PSF parameters. When none of the point sources in the ROI are bright enough, such as Geminga or Vela, to constrain the PSF tails, floating them causes large uncertainties in the inferred background and point source fluxes. This is because the tails of the PSF are nearly degenerate with the background normalization and allowing them to float without strong priors causes a significant bias in the flux predictions. We generalize the algorithm to sample from the PSF parameters, when needed, and float the PSF during our mock runs. However, we fix the PSF during the nominal data run and discuss the results of letting the PSF float in Appendix [\[sect:psfnvari\]](#sect:psfnvari){reference-type="ref" reference="sect:psfnvari"}. The ability to float the PSF is especially relevant for ground-based optical astronomy, where the PSF is different in each exposure. In the other case, where we do float the PSF, a typical ROI without exceptionally bright point sources cannot constrain a double-King function with a floating scale factor. In particular, a bright source is needed to break the near-degeneracy between the power law slopes of the core and tail components of the radial profile. Therefore, in that case we fix the scale factor to the best-fit value provided by the Fermi-LAT Collaboration and use a linear combination of a King function and a Gaussian with only $4N_eN_{psf}$ free parameters. We place uniform priors on the logarithm of the angular scale, $\sigma$, the relative fraction of the Gaussian and King components, $f$, and the inverse tangent of the slope, $\gamma$. \[\mathcal{P} = a_G \mathcal{G}(\theta_0; \sigma_G) + a_K \mathcal{K}(\theta_0; \sigma_K, \gamma)\] Inference of the point source catalog sensitively depends on the PSF modeling. Due to the power-law tails of the PSF, bright members of the point source population can outshine the faintest point sources even $\sim$ a few degrees away as shown in Figure [\[figr:eval\]](#figr:eval){reference-type="ref" reference="figr:eval"}. In the rest of the paper, we will collectively denote the set of parameters characterizing the PSF, with the parameter vector $\vec{\eta}$. ## Prior structure A probabilistic graphical model of our inference framework is presented in Figure [\[figr:grap\]](#figr:grap){reference-type="ref" reference="figr:grap"}. In this representation, nodes denote random variables, while edges directed into a node denote the set of nodes that hierarchically parameterize the probability distribution of the given node, such as in Equations [\[equa:numbpnts\]](#equa:numbpnts){reference-type="ref" reference="equa:numbpnts"} and [\[equa:flux\]](#equa:flux){reference-type="ref" reference="equa:flux"}. The red, blue, green and yellow nodes represent our model parameters, which are assigned prior probability distributions. In particular, the red nodes are the hyperparameters $\mu$ and $\alpha$, which set the normalization and slope of the point source flux distribution, respectively. The blue node indicates the number of point sources in a model, i.e., is the multiplicity of each green node. Likewise, the green nodes are the point source parameters, i.e., longitude, latitude, flux and spectral index from left to right. The yellow nodes are the background normalizations, $\vec{A}$, and PSF parameters, $\vec{\eta}$. $\mathcal{M}$ node is a deterministic function of the above model parameters representing the set of forward modeled photon count maps. Finally, $\mathcal{D}$ represents the observed photon count maps, whose consistency with the former drives the evolution of the MCMC state through the Poisson likelihood. We also color code the directed edges such that black edges denote a probabilistic relation, whereas olive lines show a deterministic relation. Finally, magenta lines imply that the multiplicity of the destination node is set by the origin. Note that we do not use plate notation, since the multiplicity itself is a discrete parameter in our model, which admits a hierarchical prior. Given the parameter and hyperparameters introduced above, the joint prior probability distribution of a model with $N$ point sources becomes \[\begin{gathered} P(\mu, \alpha, N, \vec{A}, \vec{\eta}, \vec{l}, \vec{b}, \vec{f}, \vec{s}) = \\ P(\mu) P(\alpha) P(N\|\mu) P(\vec{A}) P(\vec{\eta}) \prod_{n=1}^N P(l_n) P(b_n) P(f_n\|\alpha) P(s_n) \end{gathered}\] where we use the vector notation to refer to the union parameter set of $N$ point sources, e.g., $\vec{l} \equiv (l_1, l_2,..., l_N)$. ## ROI margins The observed count data in the ROI can potentially be affected by point sources outside the ROI. In order to model such emission from point sources just outside the ROI, we make the spatial prior region slightly larger (1 degree larger on all sides) then the ROI window over which the likelihood is calculated. Therefore the model point sources can move slightly out of the image and probe whether a feature close to the boundary can be fit better by a model point source outside the ROI. As a result the offset provides a smooth transition from a data-informed region well inside the ROI to a prior dominated region outside the ROI, where the posterior asymptotes to the prior. This can be clearly seen in the artificial accumulation of sampled point sources along the boundary of the stacked posterior, as will be shown in Section [5](#sect:ngal){reference-type="ref" reference="sect:ngal"}. # Sampling Method {#sect:samp} In this section, we describe the method used to sample from the probability distribution of catalogs consistent with the given photon count map. The starting point is the Bayesian assumption that there exists an underlying probability distribution that characterizes our knowledge of the model parameters. These parameters are the longitude, $l$, latitude, $b$, flux, $f$, and color, $s$, of each point source. Refer to Section [2](#sect:modl){reference-type="ref" reference="sect:modl"} for details on how we define and place priors on these parameters. The parameter space of the point source metamodel is, then, the union of the parameter spaces of the point source models that contain from $N_{min}$ up to $N_{max}$ point sources. Denoting this space as the catalog space, $\mathcal{C}$, we therefore sample from \[\mathcal{C} = \bigcup_{N=N_{min}}^{N_{max}} \mathcal{C}_N = \bigcup_{N=N_{min}}^{N_{max}} \mathcal{L}_N \times \mathcal{B}_N \times \mathcal{F}_N \times \mathcal{S}_N,\] where $\mathcal{C}_N$ is the catalog subspace with $N$ point sources. $\mathcal{C}_N$ can further be written as the product space of $\mathcal{L}_N$, $\mathcal{B}_N$, $\mathcal{F}_N$ and $\mathcal{S}_N$ that denote the longitude, latitude, flux and color spaces of point sources in the model with $N$ point sources. Although sampling from a given $\mathcal{C}_N$ can be performed by constructing an MCMC chain using ordinary Metropolis updates over $\mathcal{L}_N$, $\mathcal{B}_N$, $\mathcal{F}_N$ and $\mathcal{S}_N$, the fact that the catalog space is transdimensional means that it cannot be explored by simple within-model proposals. We assume that the number of point sources in the image is not known a priori. Therefore, the dimensionality of the point source model, $N$, becomes a discrete parameter subject to inference. In addition to the catalog space, the complete hypothesis space includes normalizations of the isotropic and spatially varying background models in each energy bin and PSF class, as well as parameters that characterize the PSF. We will defer the discussion of these degrees of freedom to Section [2](#sect:modl){reference-type="ref" reference="sect:modl"}, since they have fixed dimensionality and can be explored using ordinary within-model updates. ## Trans-dimensional sampling In the rest of the paper we will generically refer to a parameter in the model by $\theta$. The objective is to sample from the posterior distribution of $\theta$ given the observed count map $D$, $P(\theta\|D)$, after updating our prior beliefs about $\theta$, $P(\theta)$. The Bayesian update is accomplished through the likelihood of observing the data given our model, $P(D\|\theta)$. For this purpose we construct a Markov chain of states, $\{\theta\}$, whose asymptotic stationary distribution is the desired posterior distribution. Therefore we require that the chain is reversible with respect to the posterior, i.e., satisfies detailed balance condition. If the sampling space was fixed dimensional, this would imply that \[\begin{gathered} \int P(\theta\|D) Q(\theta^\prime\|\theta) \alpha(\theta^\prime\|\theta) \dd{\theta} = \\ \int P(\theta^\prime\|D) Q(\theta\|\theta^\prime) \alpha(\theta\|\theta^\prime) \dd{\theta^\prime}, \end{gathered}\] where $\theta$ and $\theta^\prime$ are the current and the proposed states, $Q(\theta^\prime\|\theta)$ is the transition kernel from $\theta$ to $\theta^\prime$ and $\alpha(\theta^\prime\|\theta)$ is the appropriate proposal acceptance probability, \[\alpha(\theta^\prime\|\theta) = \min \Bigg( 1, \frac{P(\theta^\prime\|D)}{P(\theta\|D)} \times \frac{Q(\theta\|\theta^\prime)}{Q(\theta^\prime\|\theta)}\Bigg). \label{equa:acppmths}\] Note that when sampling from the catalog space, $\theta$ and $\theta^\prime$ can have different dimensions. This brings up an interesting issue that the chain can no longer be reversible in the transdimensional case, since the transition is not one to one. The remedy is to draw random auxiliary variables, $u$ and $u^\prime$, using the densities $g(u)$ and $g^\prime(u^\prime)$, to match the dimensions of the initial and final states of the transition such that the mapping $(\theta, u) \leftrightarrows (\theta^\prime, u^\prime)$ is a diffeomorphism, where \[\begin{aligned} H(\theta, u) &= (\theta^\prime, u^\prime) \\ H^{-1}(\theta^\prime, u^\prime) &= (\theta, u). \end{aligned}\] The dimension matching, \[\begin{aligned} \mathcal{D}(\theta) + \mathcal{D}(u) = \mathcal{D}(\theta^\prime) + \mathcal{D}(u^\prime), \end{aligned}\] where $\mathcal{D}$ denotes the dimension operator, conceals the transdimensional nature of the proposal and ensures reversibility. In this case we require \[\begin{gathered} \int P(\theta\|D)g(u)\alpha(\theta^\prime\|\theta) \dd{\theta}\dd{u} = \\ \int P(\theta^\prime\|D)g^\prime(u^\prime)\alpha(\theta\|\theta^\prime) \dd{\theta^\prime}\dd{u^\prime}. \label{equa:rjmcbala} \end{gathered}\] The transition is accomplished through the transition kernel, $H(\theta, u) = (\theta^\prime, u^\prime)$, which replaces the probabilistic transition kernel, $Q(\theta^\prime\|\theta)$, in the fixed-dimension case. For example, in the case where the proposal is to add a new point source, the auxiliary parameters, $u$, are simply the parameters describing the new point source. This method is known as the Reversible Jump MCMC (RJMCMC), which is a variant of MCMC that allows across-model moves in a pool of models indexed by their dimensionality. We inherit the reversible jump formalism in implementing transdimensional proposals. Given the freedom to jump across models, there are an infinite number of ways to propose such transitions using the current state. However, only some are useful schedules to explore the catalog space. Denoting the probability of the $m^{th}$ type of proposal in state $\theta$ by $j_m(\theta)$, the ratio, \[\frac{j_m(\theta^\prime)}{j_m(\theta)},\] should be also included in the resulting acceptance rate in order to compensate for any bias in the proposal frequencies. Using the new detailed balance condition, Equation [\[equa:rjmcbala\]](#equa:rjmcbala){reference-type="ref" reference="equa:rjmcbala"}, one finally obtains, \[\begin{gathered} \alpha(\theta^\prime\|\theta) = \min(1, \alpha_0), \\ \alpha_0 = \frac{P(D\|\theta^\prime)}{P(D\|\theta)} \frac{P(\theta^\prime)}{P(\theta)} \frac{j_m(\theta^\prime)}{j_m(\theta)} \frac{g(u^\prime)}{g(u)} \Bigg\lvert \frac{\partial(\theta^\prime, u^\prime)}{\partial(\theta, u)} \Bigg\rvert, \label{equa:accp} \end{gathered}\] since the coordinate transformation requires that probability is conserved, i.e., \[\dd{\theta^\prime}\dd{u^\prime} = \Bigg\lvert \frac{\partial(\theta^\prime, u^\prime)}{\partial(\theta, u)} \Bigg\rvert \dd{\theta} \dd{u},\] for all $\theta$ and $u$. The usual within-model proposals are recovered when both $u$ and $u^\prime$ have the same dimension. Given that the parameter space of the point sources has large covariances in crowded fields, it is likely that the sampler can get stuck in a likelihood island without being able to efficiently visit all high-likelihood regions. In order to prevent this, we take heavy-tailed Gaussian steps when proposing within-model transitions. However, large changes in parameters can also suffer from the prior ratio in Equation [\[equa:accp\]](#equa:accp){reference-type="ref" reference="equa:accp"}, especially when the prior is a power law such as for the fluxes of point sources. We therefore set the prior ratio to unity by transforming the parameters such that their prior distributions become uniform. This is accomplished through the use of the Cumulative Distribution Function (CDF) and its inverse, which map the parameters to the unit space, $\mathcal{U} = [0, 1]$, and back. Therefore, the actual sampling is performed in the unit space with the prior ratio set to unity by definition at the expense of inverse transforming the parameters back to their genuine values when the likelihood function is to be evaluated or a hyperparameter is updated. This excludes the number of point sources in the model, $N$, which is an integer parameter. Conventional updates to point source positions, fluxes or colors, i.e., within-model proposals, only allow the sampler to explore one catalog subspace, $\mathcal{C}_N$. To explore the full space, $\mathcal{C}$, the sampler must propose updates that change $N$, by adding or removing sources. We use two pairs of such updates, where $u$ and $u^\prime$ have different dimensions. The first type is the elementary operation of adding or deleting a point source to or from the current list of point sources. We denote this pair of proposals birth and death. They are the reverse proposals of each other unlike within-model updates that are manifestly reversible. Therefore both birth and death have to be present in the set of possible proposal types in order for detailed balance to be respected. When a point source is to be added, an auxiliary vector, $u$, is drawn whose elements are distributed uniformly between 0 and 1. We then use the inverse CDF of the position, flux and color distribution functions to transform the uniformly distributed elements to random draws of the position, flux and color. These parameters determine the parameters of the point source to be added. In contrast, when a point source is to be killed, the CDF transform of its parameters define $u$. The second pair of transdimensional proposals are splits and merges, where a point source is split into two, and two point sources are merged into one, respectively. This pair is especially important in the crowded field, where splits and merges make the exploration of the nearly degenerate regions of the parameter space more efficient. We refer the reader to Appendix [\[sect:accpprob\]](#sect:accpprob){reference-type="ref" reference="sect:accpprob"} for the implementation details of the move types. ## Time performance of the sampler Sampling from the posterior probability distribution of catalogs is a computationally demanding task. The primary reason is the large and variable number of parameters in the sample vector. For a typical MCMC run on a mock image with 200 point sources, the sample vector can have as many as $\sim$ 1000 parameters. Since position and flux changes are proposed one at a time in order to keep the acceptance ratio high, the typical autocorrelation time becomes $\sim$ 10000 MCMC steps. Furthermore the least significant features in the image require the highest number of MCMC steps for convergence, because they are sampled only slightly more frequently compared to a uniform Poisson background. Increasing the performance requirements of the sampler is the fact that hyperparameters parametrize the hierarchical priors on the point source parameters. This means that the convergence of parameters precedes that of the hyperparameters. A typical inference thus requires $\sim 10^8$ MCMC steps. For each proposal in an MCMC run, the current parameter vector is used to compute the model count map, which is then compared to the data count map through the Poisson likelihood. This process typically dominates the time budget and requires careful optimization for the sampler to be scaled up to large ROIs, larger number of point sources or energy bins. We break the optimization into four steps. - Although the sampling is actually performed in the CDF-transformed parameter space, the inverse CDF transforms of the parameters are stored along with those of the CDF-transformed parameters, precluding redundant CDF transformations. - Flux maps are calculated perturbatively. During each proposal the current flux map is modified by the updated, added or killed point sources. This decreases the time complexity of processing a single sample from $\mathcal{O}(N_{pix}N)$ to $\mathcal{O}(N_{pix})$, where $N$ is the number of point sources and $N_{pix}$ is the number of pixels in the ROI. - For each point source, the PSF is evaluated over a subset of pixels that lie inside a circle. The list of nearby pixels for each pixel is precomputed and stored as a look-up table. The radius of the circle depends on the flux of the associated point source and is determined such that the largest flux allowed by the prior contaminates the lowest flux at most by one percent. This is illustrated by Figure [\[figr:eval\]](#figr:eval){reference-type="ref" reference="figr:eval"}, where the horizontal line indicates 0.01$f_{min}$, where $f_{min}$ is the minimum flux allowed by the model. Hence, point sources contribute to the total flux map up to the radius, after which the bias introduced by their neglect is significantly below the faintest possible point source. We monitor the error introduced by this approximation and ensure that the bias is negligible in all pixels, energy bins and PSF classes. As a result of this approximation the time complexity of a single proposal is further reduced to $\mathcal{O}(1)$. This implies that the time complexity of a single proposal (except those that update the PSF or hyperparameters) does not depend on how many sources or pixels there are in the image. - The leading contribution to the time complexity of the average sample comes from the likelihood evaluation, which requires the computation of the proposed change to the flux over a set of data cubes, e.g., pixels, energy bins and PSF classes. In this operation, the PSF is evaluated by computing the angular distance from the point sources to the pixel centers. In order to accelerate this computation, we precompute unit vectors to the `HealPix` pixel centers, $\hat{u}_{pix}$. This allows us to compute the angular distance, $\theta_0$, using the dot product \[\theta_0 = \arccos(\hat{u}_{PS} \cdot \hat{u}_{pix}),\] where $\hat{u}_{PS}$ is the unit vector along the point source of interest. Furthermore, we store the radial profile of the PSF in the memory and interpolate it when calculating the flux updates. In overall, the employed acceleration scheme reduces the mean computation time per sample to $\sim 6$ ms. Nevertheless, the evaluation of the PSF takes $\sim 5$ ms per sample on the average and dominates the time budget of a typical sample. In order to scale the algorithm to larger ROIs, multiple energy bins or epochs, the sampler has been parallelized to run over multiple cores. The current time performance is adequate for analyzing full-sky datasets such as that of Fermi-LAT in $\sim 5000$ CPU hours. Although the execution time per sample increases linearly with the typical number of point sources in the sampler, the time it takes to get an independent sample from the catalog space scales roughly as the square of the number of point sources. This is due to the fact that as the number of parameters increases, one either has to take smaller steps in the model space, or make changes in fewer number of parameters in a given proposal, both of which increase the autocorrelation of the resulting chain. Therefore, a slight increase in the ROI size may significantly increase the convergence time. Because this scaling is not desirable for large fields or full sky analyses, it is more feasible to separately sample from the catalog space of patches that are $\sim 10$ PSFs wide. This neglects covariances between sources and background emission in different patches, but allows scaling to much larger fields. However, further speed improvements would be needed for analyzing optical photometric data, where the typical pixel size is less than an arcsecond as opposed to being tens of arcminutes as in Fermi-LAT. In addition to the performance improvement gained by working on a Cartesian grid, this could be achieved using GPUs (Graphics Processing Unit) given the parallelizable nature of the sampling problem. Note that the time complexity of the overall convergence still depends on the minimum flux allowed by the model. This is because it sets the typical number of model point sources and, hence, the size of the parameter space. In a typical run with 3 energy bins, 2 PSF classes, a $40^\circ\times40^\circ$ ROI, $\sim 10^5$ pixels and $\sim 250$ model point sources, the execution time is 250 CPU hours. ## Convergence diagnostics MCMC formalism allows the exploration of complex posterior distributions with the caveat that convergence to the stationary distribution can require a long simulation time. Given that a typical catalog has thousands of parameters, the finite running time of the Markov chain might raise concerns over whether the sampled distribution is representative of the desired target distribution. One method for evaluating the chain convergence is to inspect the variance of the sampled chain. However the variance of a single MCMC chain can underestimate the true variance, since the realized chain may not have converged to the target distribution despite having a small variance. We therefore run multiple, usually around 20, chains and compare the mean of the chain variances to the variance of the means of the chains, ensuring that the initial states of the chains are over-dispersed relative to the target distribution. Having $N_c$ chains and $N_s$ samples in each chain, the resulting test statistic, \[\hat{R} = \sqrt{1 + \frac{B}{W}-\frac{1}{N_s}},\] is known as the Potential Scale Reduction Factor (PSRF) and can be used to assess whether the chains have converged. Here, $W$ is the within-chain variance, i.e., the mean of the chain variances, \[W = \frac{1}{N_c} \sum_c^{N_c} \Bigg( \frac{1}{N_s-1}\sum_s^{N_s}(y_s-\bar{y}_c)^2 \Bigg),\] $B$ is the between-chain variance, i.e., the variance of the chain means, \[B = \frac{1}{N_c-1} \sum_c^{N_c}\Bigg(\bar{y}_c-\frac{1}{N_c} \sum_c^{N_c} \bar{y}_c \Bigg)^2,\] and the mean of the $c^{th}$ chain is \[\bar{y}_c = \frac{1}{N_s} \sum_s^{N_s} y_s.\] Our definition of $B$ differs from that commonly found in the literature by a factor of $N_s$, which is absorbed into the definition of $\hat{R}$. In a well-mixed chain the PSRF should be close to unity. One caveat of using the PSRF as an estimator of the convergence in this framework is that the sampled chains are transdimensional. Note that the problem of point source inference has a labeling degeneracy. That is to say that there is an N!-fold degeneracy in the likelihood function of a point source model with $N$ point sources, since permuting the parameter labels of these $N$ point sources leaves the likelihood invariant. For any reasonably large $N$, $N!$ is larger than the number of samples that can be drawn from the posterior. Therefore, formal convergence is not possible, but also unnecessary, since well sampling only one of the degenerate likelihood peaks reveals the unique likelihood topology of the problem. In order to probe convergence, we instead monitor the variance of the resulting model emission map. We draw 1000 random voxels (triplets of pixel, energy bin and PSF class) and show the distribution of the PSRF in our data run (Section [5](#sect:ngal){reference-type="ref" reference="sect:ngal"}) in Figure [\[figr:gmrb\]](#figr:gmrb){reference-type="ref" reference="figr:gmrb"}, which confirms that the between and within chain variances are similar for most pixels. ## Autocorrelation of the chain Given that the sampler updates, adds or kills one point source at a time, nearby samples in the chain are expected to be similar. In order to retain the Markovian property, we therefore thin the chain by a factor equal to the maximum number of parameters allowed by the metamodel. Typically this implies that the chain is thinned by a factor of 1000-10000. In order to ensure that the resulting chain has the Markovian property, we compute the autocorrelation of the chain, where we follow a similar method to that of diagnosing convergence using the predicted emission map. After calculating the autocorrelation for the randomly drawn 1000 voxels, we take the average over chains and parameters. We plot the resulting autocorrelation of the chain in Figure [\[figr:atcr\]](#figr:atcr){reference-type="ref" reference="figr:atcr"}, which shows that samples in the diluted chain is memoryless. ## `PCAT` We make the resulting software, `PCAT`, available to the astronomy community. `PCAT` is a pure `Python 2.7` implementation of the described sampling algorithm along with extensive routines to customize the sampler for the problem at hand, further process the output, diagnose convergence and visualize probabilistic catalogs. It is designed to sample from the catalog space for a given photon count map, choice of data binning and prior structure, where the level of background and the PSF are potentially unknown. As of version 0.2, it is only intended to process binned count maps from a Poisson process. # Mock runs {#sect:mock} In this section we will first present an ensemble of catalogs sampled using a mock, i.e., simulated, dataset. Hence we will compare the true (input) parameter values with those obtained from the ensemble. The Poisson mean of the mock map is generated by sampling all parameters randomly from the prior and then calculating the resulting emission map due to the background emission and the point sources. Finally mock data is generated as a Poisson realization of the mean emission. In total 300 point sources are generated from a flux distribution with a power law slope set to-1.8. We use the `Pass 7`, `source` class exposure map of the Fermi-LAT instrument between weeks 9 and 217, when calculating the mock number of counts from the synthetic emission map. This ensures that the mock point sources are subject to incompleteness at the same flux as the real point sources. Figure [\[figr:sampcnts\]](#figr:sampcnts){reference-type="ref" reference="figr:sampcnts"} summarizes a fair sample from the probabilistic catalog. The panels show the number of counts in the generated mock map, the sampled model map and the residual, respectively from left to right. Here, the shown sample is only one of many realizations that constitute a fair draw from the underlying probability distribution. The number of counts are given in the 1 GeV-3 GeV energy bin. As can be seen in the right panel, Poisson fluctuations near bright point sources can be large even if the data is a realization of a generative model, i.e., the model is a good description of the data. In practice, further mismodeling of the PSF can result in large, and possibly, even coherent residuals around bright point sources. Depending on how low the model point source fluxes are allowed to go, this can result in the sampling of spurious point sources around bright ones. Therefore care must be taken to employ a PSF modeling that does not bias the flux distribution of the point source population. Figure [\[figr:arrydatacnts\]](#figr:arrydatacnts){reference-type="ref" reference="figr:arrydatacnts"} tiles together fair samples from the posterior showing the number of data counts in each pixel, i.e., similar to the left panel in Figure [\[figr:sampcnts\]](#figr:sampcnts){reference-type="ref" reference="figr:sampcnts"}, but showing real NGPC data instead. The grid illustrates the typical evolution of the MCMC state, where bright true point sources have an associated model point source with precise spatial and spectral localization, whereas faint true point sources are only sometimes associated with model sources. Next, Figure [\[figr:mockhistflux\]](#figr:mockhistflux){reference-type="ref" reference="figr:mockhistflux"} shows the flux distribution of the mock (green) and median sample (black) with the 68% credible interval. The 3FGL is also shown with red for reference. The 1$\sigma$ uncertainties cover the truth for most of the flux bins. Note that a flux-incomplete traditional catalog rolls off at the faint end since such sources are fainter than the typical fluctuations of the background. In a probabilistic catalog the flux distribution can be probed even below where a traditional catalog would roll off. This is because even though none of the point sources in the catalog samples is to be taken as true, repeated sampling of the flux distribution can constrain the population characteristics. However, this is only true for point sources more significant than $\sim 1 \sigma$. At yet lower fluxes, the sampled flux distribution function is informed by the prior more than the likelihood, and tends to follow whatever characteristic is imposed by the hierarchical prior, i.e., a power law with a range of indices allowed by the hyperprior. This transition from a likelihood dominated region to a prior dominated one is controlled by the minimum allowed flux of the point sources, $f_{min}$, and indicates where probabilistic cataloging becomes ineffective in constraining the population characteristics. We discuss how we determine the transition region in Appendix [\[sect:info\]](#sect:info){reference-type="ref" reference="sect:info"}. The former comparison confirms the statistical agreement between the true and median catalogs. However it is also desirable to perform an element-wise comparison between the two catalogs. Towards this purpose we associate a traditional (in this case, the true) catalog and a probabilistic catalog as follows. For a given sample from the catalog space, we initialize the association algorithm by setting the fluxes of all associations to the true catalog, to zero. Then, for each point source in the sample catalog, we ask if there are any point sources in the true catalog within 0.5 degrees. If so, we take the closest such true point source and add the model point source to the true point source's list of possible associations. We then repeat this for all model point sources in the sample catalog. Along the way it happens that multiple model point sources get matched to the same true point source. In that case, we use the gathered list of possible associations to select the closest model point source as the association. After all samples in the probabilistic catalog has been so processed, we take the median flux associated with each point source in the traditional catalog. Figure [\[figr:mockscatspec\]](#figr:mockscatspec){reference-type="ref" reference="figr:mockscatspec"} illustrates the resulting correlation. The horizontal axis shows the true flux of the source, while the vertical axis marks the median of the fluxes associated to it, including potential zero fluxes due to missed associations. This procedure is generic to associations between any traditional and probabilistic catalog and not just to the association of the probabilistic catalog to the underlying true catalog. Furthermore, in Figure [\[figr:mockscatspec\]](#figr:mockscatspec){reference-type="ref" reference="figr:mockscatspec"} the vertical error bars denote the statistical uncertainty due to the stacking of the ensemble of catalogs. Horizontal error bars are not provided, since the true catalog is a generated catalog without any instrumental uncertainties. At the bright end, model point sources are statistically significant and well localized, yielding small vertical error bars. Moving towards the faint end, the correlation first broadens due to associated features on the image becoming comparable to Poisson fluctuations of the background. In the extreme faint limit, one expects a given true point source to be associated with a random model point source, completely suppressing the correlation. We then plot the posterior distribution of the flux distribution normalization and power law slope (Figures [\[figr:mockfdfnnorm\]](#figr:mockfdfnnorm){reference-type="ref" reference="figr:mockfdfnnorm"} and [\[figr:mockfdfnslop\]](#figr:mockfdfnslop){reference-type="ref" reference="figr:mockfdfnslop"}). Similarly, the posterior distribution of the hyperparameters cover the true values. For the mean number point sources, the relevant true value is the imposed number of mock point sources, which is 300. Lastly, Figure [\[figr:mocknumbpnts\]](#figr:mocknumbpnts){reference-type="ref" reference="figr:mocknumbpnts"} shows the posterior distribution of the model indicator, i.e., the number of point sources in the sample catalogs. This distribution is formally proportional to the posterior probability of the models assuming that the metamodel is true. Therefore it can be used to calculate the relative evidence (Bayes factor) for any two models. # Application to Fermi-LAT Data {#sect:ngal} Next, we show results using the gamma-ray data as measured by the Fermi-LAT instrument. In order to be able to compare with the 3FGL traditional catalog published by the Fermi-LAT Collaboration, we used the reprocessed `Pass 7`, `source` class data and generated `HealPix` sky maps with base resolution 256, in three energy bins between 0.3-1 GeV, 1-3 GeV and 3-10 GeV. Then we constructed a joint-likelihood by fitting the front and back converted sky maps separately and taking the total likelihood. This allows a more precise modeling of the PSF, since photons that convert at the top of the instrument, i.e., front-type events, have better angular reconstruction. The choice of energy binning coincides with that of the 3FGL catalog, in order to facilitate the comparison. Furthermore, since most of the sources at high galactic latitudes are time-variable blazars, we use data in the same time interval as that used to construct the 3FGL catalog, i.e., weeks 9 through 217. A probabilistic catalog, by construction, cannot be reduced to a single list of point sources. Nevertheless its statistical summary can still be compared to a traditional catalog. This provides a means of assessing the performance of our ensemble of catalogs against the well established 3FGL. As in the mock data case, Figures [\[figr:inpthistflux\]](#figr:inpthistflux){reference-type="ref" reference="figr:inpthistflux"} and [\[figr:inptscatspec\]](#figr:inptscatspec){reference-type="ref" reference="figr:inptscatspec"} show the flux distribution function and associations between the 3FGL and probabilistic catalog. The association between the 3FGL and the probabilistic catalog indicates that there is an agreement between the two. The correlation is stronger at larger fluxes, where covariances with the background normalization and PSF affect flux determinations negligibly. The level of agreement decreases towards lower fluxes since covariances with the background level and the radial profile of the PSF as well as the shape and normalization of the flux distribution widens the prediction of the probabilistic catalog. It is also worth noting that the scatter in Figure [\[figr:inptscatspec\]](#figr:inptscatspec){reference-type="ref" reference="figr:inptscatspec"} is partially due to the different spectral modeling used by the two catalogs. In other words, even if the flux predictions of the two catalogs were perfectly correlated in the pivot energy bin, 1 GeV-3 GeV, fluxes in the other energy bins would show some dispersion. Determining the nature of a point source population, i.e., whether it is a pulsar or AGN and what subclass it belongs to, requires the reconstruction of its light curve and color distribution. We show the latter in Figure [\[figr:inpthistsind\]](#figr:inpthistsind){reference-type="ref" reference="figr:inpthistsind"}. The distribution around 2.2 implies that most of the point sources are blazars, where the upper tail is dominated by Flat-spectrum Radio Quasars (FSRQs) whereas BL Lacs make up the harder sub-population. By collecting the point source fluxes in the ensemble of catalogs, we also estimate the contribution of the point sources to the total emission in the NGPC. The median spectra of the point sources, isotropic component, and diffuse model are given in Figure [\[figr:inptcompfracspec\]](#figr:inptcompfracspec){reference-type="ref" reference="figr:inptcompfracspec"}. The diffuse model is observed to be the dominant component, accounting for $57\%\substack{+6 \\-5}$ of the total emission. The isotropic component and the point sources account for the rest in roughly equal amounts, i.e., $25\%\substack{+4 \\-3}$ and $18\%\substack{+2 \\-2}$, respectively. However we note that the partitioning between the point sources and the isotropic component is set by our choice of the minimum allowed point source flux. If we lower the minimum flux allowed for the point sources, then the relative contribution of the point sources would account for some of the isotropic component. The reason for this near-degeneracy is that allowing the sampler to populate the image with point sources much fainter than the level of Poisson fluctuations of the background is equivalent to decreasing the isotropic background emission. In this sense, the question of how much of the emission is accounted for by the point sources should be addressed with reference to a particular minimum point source flux. Samples from the catalog space can be stacked together by binning catalog samples in space and flux. This yields a map of the probability of finding a point source at a certain direction and flux, given our model. In general, this should not be interpreted as an unconditional probability, since the model used may not contain the set of all backgrounds and point sources that could be consistent with the observed data. In this work, however, the background is relatively featureless and the allowed flux distribution covers nearly four decades with a floating power law slope. This makes the stacked posterior approximately equal to the probability of finding a point source at a certain direction and flux. We show in Figure [\[figr:pntsbind\]](#figr:pntsbind){reference-type="ref" reference="figr:pntsbind"}, the catalog samples binned spatially and stacked spectrally. The color scale gives the number of catalog samples, where a model point source lands in the associated pixel. All green stars, which show the locations of the true point sources are associated with model point sources. The hot pixels away from the true point sources show regions that have $\sim1-4\sigma$ count features. The posterior distribution of the hyperparameters provides another handle on the population characteristics. For example Figures [\[figr:inptfdfnnorm\]](#figr:inptfdfnnorm){reference-type="ref" reference="figr:inptfdfnnorm"} and [\[figr:inptfdfnslop\]](#figr:inptfdfnslop){reference-type="ref" reference="figr:inptfdfnslop"} show the posterior of the normalization and power-law slope of the flux distribution function, respectively. Note that hyperparameters do not directly affect the likelihood. However their posterior distribution is still informed by the data. This is because the hyperparameter updates respect detailed balance with respect to the Poisson probability of observing the sampled model flux distribution and, in turn, model point source updates respect the Poisson probability of observing the data. We find the power law slope to be $-1.92\substack{+0.07 \\-0.05}$. This is smaller than the expectation from a uniform distribution of equally bright blazars, i.e., \[\dv{N}{f} = \dv{N}{r}\Bigg(\dv{f}{r}\Bigg)^{-1} \propto r^2 \times \Bigg(\frac{1}{r^3}\Bigg)^{-1} = f^{-5/2}.\] Previously, found that the source count function has a slope of $-2.6 \pm 0.2$ at the bright end, which hardens to $-1.6 \pm 0.1$ at the faint end. Because we use a single power law, the resulting posterior converges to an intermediate value. Moreover, given that we perform sampling only over a $1600$ degree$^2$ patch around the NGP, this analysis is subject to more shot noise. We leave a large ROI, high latitude sampling to future work. As for the mock data, we show in Figures [\[figr:inptnormisot\]](#figr:inptnormisot){reference-type="ref" reference="figr:inptnormisot"} and [\[figr:inptfdfnnorm\]](#figr:inptfdfnnorm){reference-type="ref" reference="figr:inptfdfnnorm"}, the normalization of the isotropic template and the Fermi diffuse model for each energy bin, respectively. We find the median of the isotropic and diffuse backgrounds to be larger than unity by a factor of $\sim 1.3$ and $\sim 1.1$. Moreover, because the diffuse model is relatively featureless in the NGPC, the two normalizations have a large covariance. Lastly, the Figure [\[figr:inptnumbpnts\]](#figr:inptnumbpnts){reference-type="ref" reference="figr:inptnumbpnts"} shows the posterior distribution of the number of point sources. We infer that there are $270\substack{+30 \\-10}$ point sources in the ROI above $3 \times 10^{-11}$/cm$^2$/s/sr/GeV at the $1-3$ GeV energy bin. We make our probabilistic catalog available at the Harvard Dataverse <https://dataverse.harvard.edu> and refer the reader to Appendix [\[sect:fits\]](#sect:fits){reference-type="ref" reference="sect:fits"} for details on the data format. # Discussion {#sect:disc} The Bayesian approach to point source inference allows a more robust treatment of the parameter covariances compared to finding the most likely catalog. Furthermore, information encoded in the subthreshold features in the image is not discarded. An important motivational distinction between conventional and probabilistic cataloging is that the latter aims to find a set of point sources that is free of false positives, whereas member point sources in the former are not guaranteed to exist. It is the repeated sampling of the raw data along with the false positives that make probabilistic cataloging a very useful tool in probing population characteristics. The median flux of spatial associations of our catalog with the 3FGL agrees remarkably well when associated with the 3FGL fluxes. This demonstrates the feasibility and reliability of our probabilistic approach. We further verify the performance of probabilistic sampling in the unresolved regime by using mock data and demonstrate that the truth information is covered by our posterior. With the current implementation several thousand realizations of even full sky gamma-ray catalogs such as the 3FGL or 3LAC, which is an AGN catalog based on the 3FGL, can be generated. In fact, given the potential increase of computational resources available to catalog generation in astrophysics, it may even be feasible to make the next generation standard catalog reduction pipelines, probabilistic. As for the time domain analysis, FAVA, does not use a likelihood approach and therefore cannot be generalized using probabilistic cataloging. Given the time complexity of catalog sampling, it is unlikely that full sky catalogs can be sampled by binning data in time domain. Furthermore, flares or periodic flux variations of individual point sources will likely be washed out in the flux distribution. Since probabilistic sampling is most informative when studying the population characteristics, it is not particularly useful for time domain analyses. It may also be possible to generalize our method to other wavelengths such as optical datasets. However typical PSF size of less than an arcsecond in optical photometry restricts the application of probabilistic cataloging to $\sim$ deg$^2$ sized ROIs due to the large number of pixels. See for an example of this technique applied to optical photometry in a crowded stellar field. When multiple samples are drawn from the catalog space, point sources in one sample catalog cannot simply be matched to those in other samples. This is due to the fact that the underlying likelihood function is invariant to permuting the labels of model point sources. This is known as the labeling degeneracy problem and is common to all mixture models, where model components are not individually labeled. Therefore taking the ergodic average of a given model parameter without breaking this degeneracy becomes meaningless, since parameters change identity during the evolution of the chain. For instance in the limit of an infinitely long chain, the ergodic average of all flux parameters would be identical and, hence, not useful for inferring the flux posteriors of individual point sources. Nevertheless, inference of population characteristics does not necessitate breaking of the labeling degeneracy, since population characteristics are also invariant under permutations of point source labels. The time we need to explicitly break the labeling degeneracy is when comparing the chain of catalog samples to a traditional catalog. This requires a prescription to associate point sources. Associations at low fluxes inevitably causes loss of information and is the major reason for the increasing spread in Figures [\[figr:mockscatspec\]](#figr:mockscatspec){reference-type="ref" reference="figr:mockscatspec"} and [\[figr:inptscatspec\]](#figr:inptscatspec){reference-type="ref" reference="figr:inptscatspec"} towards low fluxes. Model choice in the Bayesian framework tries to establish a balance between the goodness of fit of a model and its predictive power. A point source model will fit a given count map at least as well as another with fewer point sources. In fact, in the limit of very large number of point sources, the point source model essentially becomes indistinguishable from a diffuse model with arbitrary number of spatial degrees of freedom. It is thus desirable that the chosen model covers the smallest possible volume when projected onto the data space while still fitting the observed data reasonably well. Known as the Occam's razor, this principle is encoded in the marginal likelihood. This quantity penalizes point source models for wasted parameter space, i.e., for only marginally increasing the goodness of fit at the expense of predicting extraordinary data, which significantly reduces the likelihood for most of the added parameter space. The difficulty in probing faint point sources is the inability of the Poisson-noise count data in constraining point sources fainter than the typical Poisson fluctuations of the background. This sets a fundamental count scale, $\sqrt{C_B}$, where $C_B$ is the mean number of counts expected from the background inside a FWHM for the given exposure. This scale necessitates a careful choice of the minimum point source flux allowed by the model, $f_{min}$. Let us denote the number counts that corresponds to $f_{min}$ by $C_{min}$ for the given exposure. In the limit $C_{min} \gtrsim \sqrt{C_{B}}$, the goodness of fit of the point source model gets reduced due to lack of favorable parameter space. On the other hand, if $C_{min} \lesssim \sqrt{C_{B}}$, then the data cannot constrain the parameter space. In order to assess the regime of a given prior choice, we evaluate the relative information gain in going from prior to the posterior and choose $f_{min}$ such that the posterior is sufficiently informed by the data. See Appendix [\[sect:info\]](#sect:info){reference-type="ref" reference="sect:info"} for a more detailed discussion. A well known problem of Poisson regression of photon count data is that of mismodeling. It is important to note that when a test statistic of maximum log-likelihood difference between the alternative and null models is interpreted as a detection significance, it is implicitly assumed that the underlying model is a good description of the data. Otherwise the unquoted systematic uncertainty can be much larger than the statistical uncertainty. For example, when a flux-incomplete point source model is used, this assumption will not hold true. Therefore predictions for the diffuse templates will be biased so as to minimize the residuals of the flux-incomplete point source model. Probabilistic cataloging addresses this problem by shunting uncertainties due to flux-incompleteness to a pure statistical form. This is accomplished through sampling in the catalog space above a given cutoff flux, which contains all point source configurations that the data is consistent with. By doing so, it allows many false-positives in the model. But these are sampled less frequently compared to the true-positives as long as the sampling is performed in the likelihood-dominated region. Note that probabilistic cataloging can still suffer from mismodeling due to imperfect background models. In this work, we follow Bayesian statistics to perform inference with the motivation that models can be penalized for introducing unnecessary point sources. Strictly speaking a frequentist approach that adds a degree of freedom penalizing term to the test statistic in the form of Akaike Information Criterion (AIC) or Bayesian Information Criterion (BIC) could achieve the same goal by iterating the fit over a range of number of point sources. However in the limit of faint point sources, the likelihood topology becomes nearly degenerate and the maximum likelihood becomes a poor indicator of the goodness of fit of a model. In a similar sampler was implemented and shown to work on mock data using a single energy bin. The sampling was performed directly on the prior with a hard likelihood threshold that depends on a series of levels whose prior mass decreases successively. Known as the Diffusive Nested Sampling (DNS), this method can sample multimodal distributions with high likelihood contrast. Furthermore, as a by-product, it provides an estimator for the Bayesian evidence. In this work, however, we sample from the posterior, since we do not need the Bayesian evidence for the point source metamodel. Instead, the model choice is based on the relative frequency of visiting different models via transdimensional jumps. In addition, the likelihood topology of the problem, although being highly multi-modal, has shallow islands since most point sources are faint. This argument excludes the well-localized bright point sources. Efficient exploration of these bright members would be impossible by Metropolis-Hastings updates alone. However we employ splits and merges for this reason in order to facilitate chain mixing in the crowded field limit. Similarly, `HELP: XID+` is another Bayesian point source extraction framework that can fold in prior information on the sources. Although both algorithms are tailored for the crowded limit with a rich covariance structure, `HELP: XID+` differs from `PCAT` in that it relies on the prior knowledge of a certain number of known source positions in order to infer flux and flux uncertainties of sources consistent with the observed data. # Conclusion {#sect:conc} In this paper, we implemented a transdimensional sampler to infer point sources in a given photon count map by sampling from the probability distribution of the underlying catalog space. Our approach allows a consistent Bayesian exploration of the background normalization and the PSF, where they are co-sampled along with the model point source positions, fluxes and spectral parameters. The output is an ensemble of catalogs consistent with the count data, which represents the probability distribution of the catalog given the image. Compared to the traditional maximum likelihood solution to point source inference, probabilistic cataloging is computationally expensive. However this is a price paid for superior control over covariances in the catalog space, which becomes critical when a subsequent analysis using the inferred catalog as an input tries to reach a conclusion. Instead of using the maximum likelihood catalog with Gaussian errors around this solution, probabilistic cataloging provides a robust treatment of uncertainties. As a case study we apply our technique first to mock data and then to gamma-ray data in the NGPC. We associate the resulting probabilistic catalog with the 3FGL, and measure the flux distribution function down to $\sim$ 1-sigma sources. We thank David W. Hogg, Brendon Brewer, Benjamin Lee, Zachary Slepian, Greg Green, Albert Lee, Can Gökler and Emre Ergeçen for useful discussions during the course of the project.	{'timestamp': '2017-03-13T01:01:46', 'yymm': '1607', 'arxiv_id': '1607.04637', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04637'}
null	null
# Introduction It is well known that a gas composed of bosonic atoms with repulsive interparticle interaction at appropriate values of density and temperature undergoes Bose-Einstein condensation (BEC), a phase transition which shares similarities to transitions to superfluid and superconductive states. Since the first experimental demonstration of BEC, efforts have directed toward in investigating the thermodynamic properties of such a macroscopic quantum system and finding suitable theoretical descriptions of the phase transitions. Recently there is a revival of experimental interest devoted to the study of the thermodynamics of quantum gases. On one hand, distinguished works have explored the thermodynamics: a Fermi gas with repulsive interactions, a Fermi gas in the limit of very strong interactions, i.e., near the unitary regime, a Fermi gas in a three-dimensional optical lattice showing fermionic Mott-insulator transition, and the Boson gas in a two-dimensional optical lattice showing a bosonic Mott-Insulator transition. On the other hand, works on weakly interacting bosonic gases have demonstrated that, even in this simpler system, the understanding and characterization of the thermodynamic behavior, especially across the phase transition, are not yet complete and that more experimental works is needed to validate the theoretical predictions. New approaches to investigating these systems and new experimental results can therefore contribute, in general, to advance the understanding of the thermodynamics of quantum gases and, in particular, of their phase transitions. In this work, we experimentally determine a global susceptibility from a global thermodynamical variables approach for a harmonically trapped Bose gas. We investigate and characterize the behavior of the susceptibility when the gas undergoes a BEC. In standard thermodynamics, the equivalent quantity of the global susceptibility that we define in this work is the isothermal compressibility. This parameter describes the relative variation of the volume $V$ of a system due to a change in the pressure $P$ at constant temperature $T$: $k_T=-\frac{1}{V}\left(\frac{\partial V}{\partial P}\right)_{N,T}$. It is a property associated with density fluctuations and it can also be expressed in terms of a second derivative of the free energy with respect to the pressure. At a second-order phase transition it is therefore expected to show a singularity. Here we provide experimental evidence of such a singular behavior by taking advantage of the global thermodynamic approach. # Thermodynamics based on global variables Global variables have already been successfully employed to obtain the phase diagram and measure the heat capacity of a gas in a harmonic potential. The need to review standard thermodynamics when dealing with quantum gases comes naturally from the fact that they are usually trapped in nonhomogeneous (normally harmonic) potentials. In this situation standard definitions of pressure and volume do not apply. In fact, $P$ and $V$ are conjugate variables of thermodynamical systems defined for homogeneous densities. In particular, $P$ is an intensive variable having the same value in every position inside the volume occupied by the gas. The local density approximation (LDA) is often used in non-homogeneous situations to define local variables. A different approach, involving a set of thermodynamic variables with single values for the entire gas, allows a global description of the thermodynamics of an inhomogeneous gas and of its phase transitions. This global approach is particularly suited, compared to the LDA, for the case in which the gas is characterized by abrupt spatial variations of the density, as in the occurrence of a phase transition or in a more exotic situation such as in the presence of vortices or local potential impurities. The use of global variables to describe the thermodynamics of an inhomogeneous system has been extensively described elsewhere. In brief, within the basis of thermodynamic and statistical mechanics one can infer a volume parameter and a pressure parameter respectively: \[{\cal V}=\frac{1}{\omega_x\omega_y\omega_z}, \label{eq:V}\] \[\Pi=\frac{2}{3{\cal V}}\langle U({\bf r})\rangle=\frac{m}{3{\cal V}}\int d^3r~n({\bf r})(\omega^2_xx^2+\omega^2_yy^2+\omega^2_zz^2),\label{eq:Pi}\] where $\omega_i$ with $\left(i=x,y,z\right)$ are the harmonic trap frequencies, $\langle U({\bf r})\rangle$ is the spatial mean of the external potential, and $n({\bf r})$ is the density of the sample. ${\cal V}$ is a natural extensive "volume" for the trapped gas and the thermodynamic limit can be achieved by making the density parameter $n_{\cal V}=N/{\cal V}$ constant as $N$ and ${\cal V}$ grow indefinitely. $\Pi$ is its intensive conjugated variable $(\Pi=-\left(\frac{\partial F}{\partial{\cal V}}\right)_{N,T})$, where $F=F(N,{\cal V} ,T)$ is the Helmholtz free energy. A nice proof that $\Pi = \Pi\left( N,{\cal V},T \right)$ and ${\cal V}$ are a good set of variables to describe the system is obtained through the determination of the heat capacity, $C_{\cal V}$, whose behavior is close that one expected from treatment of a harmonic trapped Bose gas. In this framework, the isothermal compressibility parameter can be obtained from the following relation: \[\kappa_T=-\frac{1}{{\cal V}}\left(\frac{\partial{\cal V}}{\partial\Pi}\right)_{N,T}.\] $\kappa_T$ is a quantity with the same properties of the standard compressibility $k_T$ and indicates the thermodynamic stability defined by the second derivative of Gibbs free energy. The convexity property of the free energy is maintained with the condition, $0 \leq \kappa_T < \infty$. Therefore, with this susceptibility we characterize a system in thermodynamic equilibrium. # Experimental system and measurement We performed the measurements to determine $\kappa_T$ across the transition from a thermal cloud to a BEC of $^{87}{\rm Rb}$ atoms with a new experimental setup in which the volume parameter can be easily varied. The system is built in a standard double magneto-optical trap (MOT) configuration. In the first vacuum cell we load a MOT of $10^8$ atoms from a dispenser and then we transfer the atoms to the second cell using an on-resonance beam. Here, we recapture the atoms in a second MOT and, after performing a sub-Doppler cooling, we spin polarize the atomic sample in the hyperfine state $F=2,\, m_F=2$. Afterwards, we transfer the atoms at temperatures of about $40~\mu{\rm K}$, in a pure quadrupole magnetic trap where a first radio-frequency evaporation is performed. Simultaneously, we ramp on a far-detuned beam (with wavelength, $\lambda=1064~{\rm nm}$) focused on a waist $w_0=85~\mu{\rm m}$, dislocated by $z_0=300~\mu{\rm m}$ along the gravity direction below the center of the quadrupole trap. When the temperature of the atomic cloud decreases to approximately $10~\mu {\rm K}$, atoms migrate from the quadrupole trap to the center of the beam, which serves as an optical dipole trap (ODT). At that point we reduce the vertical magnetic-field gradient to a value that no longer compensates for the gravity. The atoms are thus confined in a hybrid trap given by the combination of the optical and magnetic confinements. Here we further decrease the temperature of the cloud by a second stage of radio-frequency evaporation followed by optical evaporation obtained by exponentially ramping down the power of the laser beam. We can eventually achieve a pure BEC of $\sim10^5$ atoms at typical temperatures of $100--200~{\rm nK}$. The hybrid potential, including gravity can be described by the following expression: \[\begin{aligned} U({\bf r})=\mu B'_x\sqrt{x^2+\frac{y^2}{2}+\frac{z^2}{2}}-\frac{U_0}{(1+y^2/y^2_R)}\nonumber\\ \exp\left[-\frac{2x^2+2(z-z_0)^2}{w_0^2(1+y^2/y^2_R)}\right]+mg(z-z_0)+E_0 \end{aligned}\] $\mu$ is the atomic magnetic moment, $B'_x$ is the gradient of the quadrupole trap along the $x$ direction, $y_R = w^2_0\pi/\lambda$ is the Rayleigh range of the beam which propagates along direction $y$ and $U_0$ is the optical trap depth. $g$ is the gravitational acceleration, $m$ is the atomic mass and $E_0$ is the energy difference between the zero-field point absent the dipole trap and the total trap minimum, giving the trap minimum $U({\bf r}_{min})=0$. At low temperatures the effective potential of the HT can be safely approximated by a three dimensional harmonic potential, whose frequencies are \[\omega_x\simeq\omega_z=\sqrt{\frac{4U_0}{m w_0^2}},~~\omega_y=\sqrt{\frac{\mu B'_x}{2 m \left\| z_0 \right\|}}.\label{eq:freq}\] The trap has a cylindrical symmetry where the radial frequency confinement is due the ODT and the axial weaker confinement is due to magnetic-field gradient. We characterize the atomic cloud by using absorption imaging after a free expansion from the trapping potential with a time of flight of $30~{\rm ms}$. Each image is fitted to a two-dimensional bimodal distribution composed of a Gaussian function and a Thomas-Fermi function, which are known to properly describe the thermal and the condensed component of the gas, respectively. The number of particles and temperature are obtained from the fitted images following conventional procedure. The volume parameter can be easily changed by varying the radial frequencies of the hybrid confinement, which directly depend on the final laser power of the ODT. We consider measurements for seven different sets of frequencies, i.e., for seven different volume parameters. Different temperatures have been obtained by changing the radio-frequency evaporation ramp; in this way the initial conditions for the optical evaporation change, allow us to achieve different final temperatures with the same trapping frequencies since the final power the ODT is the same. For each volume parameter we have performed many experimental runs for temperatures within the range $40--400~{\rm nK}$ and postselected atomic clouds containing $(1\pm 0.1)\times10^5$ atoms to be taken in consideration. In order to calculate the pressure parameter $\Pi$ by performing the integral in Eq. ([\[eq:Pi\]](#eq:Pi){reference-type="ref" reference="eq:Pi"}), it is necessary to reconstruct the density profile of the atoms in the trap, $n(\mathbf{r})$, from the measured profiles in the time of flight and the trap frequencies. Toward this aim, for the thermal component we can safely assume a free expansion, whereas for the interacting condensed component we apply the Castin-Dum procedure. In Fig. [\[PixT\]](#PixT){reference-type="ref" reference="PixT"} we plot the calculated $\Pi(T)$ for each volume parameter. With a decrease in the temperature the atomic gas undergoes BEC: at high temperatures we observe a linear dependence of the pressure parameter on $T$ until an abrupt change takes place at a critical temperature $T_c$, and the decrease become faster than linear. Above $T_c$, experimental data are well reproduced by the the ideal gas law $\Pi {\cal V} = N k_B T$, plotted in Fig. [\[PixT\]](#PixT){reference-type="ref" reference="PixT"} for the known number of particles and the different volumes. Below $T_c$ we perform a proper empirical exponential fit which follows the behavior of the experimental points. These fitting functions, in principle, are not related to any theoretical model. For each volume parameter we can extract the critical pressure for condensation: lower volumes demand higher pressure to condense. The transition line from a thermal atomic cloud to a BEC in the $\Pi{\cal V}$-plane is shown in Fig. [\[logVxPi\]](#logVxPi){reference-type="ref" reference="logVxPi"} marking the separation between the white (thermal) and the gray (BEC) zone. # Isotherms and determination of compressibility parameter From the measurements shown in Fig. [\[PixT\]](#PixT){reference-type="ref" reference="PixT"}, we extract different isotherms relating the volume and pressure parameters, ${\cal V}={\cal V}_T(\Pi)$, which we plot in Fig. [\[logVxPi\]](#logVxPi){reference-type="ref" reference="logVxPi"}. As the temperature decreases, the overall isothermal lines shift towards a lower pressure. We can clearly identify two different behaviors in the two different regions of the thermal and condensed regimes. In the thermal region, experimental points are well reproduced by the ideal gas law for the known number of atoms and temperatures (plotted as lines on the log-log scale of the figure). When an isotherm crosses the critical line for condensation an abrupt change occurs and it departs from the ideal gas behavior. We can now extract the isothermal compressibility $\kappa_T$ from derivation of the isotherms in Fig. [\[logVxPi\]](#logVxPi){reference-type="ref" reference="logVxPi"}. Derivation is performed point by point in correspondence with the experimental data in order not to rely on the arbitrary fitting curves, which do not correspond to any theoretical model. The obtained $\kappa_T$ values for three isothermal curves are shown in Fig. [\[kTxPi\]](#kTxPi){reference-type="ref" reference="kTxPi"}. We have chosen the curves for $T=150~{\rm nK}$ \[Fig. [\[kTxPi\]](#kTxPi){reference-type="ref" reference="kTxPi"}(a))\], $T=80~{\rm nK}$ \[Fig. [\[kTxPi\]](#kTxPi){reference-type="ref" reference="kTxPi"}(b)\] and $T=40~{\rm nK}$ \[Fig. [\[kTxPi\]](#kTxPi){reference-type="ref" reference="kTxPi"}(c)\] because they demonstrate the three classes of behavior: pure thermal gas, gas undergoing BEC transition, and gas in the single BEC region, respectively. The isothermal curve at $150~{\rm nK}$ shows the decrease in $\kappa_T$ with $1/\Pi$, as expected for an ideal gas. Let us now consider the isotherm at $80~{\rm nK}$: at low pressures the gas is thermal and the compressibility $\kappa_T$ decreases with increasing $\Pi$; when the pressure reaches the region between $20$ and $30~(\times 10^{-19}{\rm J}\cdot{\rm s}^{-3})$, the sudden increase in $\kappa_T$ indicates the transition. The compressibility reaches a maximum value before returning close to the base-line after $40 \times 10^{-19}{\rm J}\cdot{\rm s}^{-3}$. In this pressure range the compressibility acquires values 4 to 8 times higher than the base-line. The behavior of $\kappa_T$ in Fig. [\[kTxPi\]](#kTxPi){reference-type="ref" reference="kTxPi"} is typical for a second-order phase transition. An investigation of $\kappa_T$ vs $\Pi$ for different isothermal curves, where the transition takes place, reveals that at higher temperatures the transition occurs at a higher pressure and the peak of compressibility is broader for higher temperatures. Contrary to the expectation that quantities involving integration of density over the potential would be weakly sensitive to the phase transition, our data shows a sudden large variation in the compressibility at the thermal-BEC transition. # Discussion We performed the data analysis using Castin-Dum procedure to reconstruct the in situ density distribution starting with a Thomas-Fermi fit of the condensed component in the time-of-flight images. In order to probe that the general results we found do not depend on the specific model for the analysis, we also tested an alternative, less constrained, model. We fitted our images with two Gaussians for the thermal and condensed components and we reconstructed the in situ profiles by applying a variational method which has already proved to be valid to study the ballistic expansion dynamics of a condensate. We checked that the $\Pi(T)$ curves, and therefore all the derived thermodynamic quantities, extracted with the two reconstructing methods are quantitatively comparable. A complete theory predicting the exact behavior of the compressibility parameter across the transition does not exist. Nevertheless, the need to make a prediction about the behavior and the shape of the compressibility around the critical point arises naturally. We have therefore attempted a comparison between our findings and the results of a toy model. We calculate $\Pi$ for synthetic density profiles consisting in a Gaussian thermal component and a Thomas-Fermi condensed one with a relative atom number given by the ideal BEC result. This model qualitatively catches the general experimental findings. In particular, the position and the shape of the compressibility peak are reproduced by the model as presented in the inset in Fig. [\[kTxPi\]](#kTxPi){reference-type="ref" reference="kTxPi"}. Nevertheless this simple model cannot give quantitative predictions, for example, of the absolute value of the compressibility, because it is over-simplified. A fair quantitative comparison would demand a more elaborate model, beyond the scope of this experimental report. The introduction of the global variable approach has proven to be a valid complementary approach to the LDA. Generally speaking, the LDA approach in fact in fact has strong intrinsic limitations in the case where sudden variation of the densities occurs, as at the thermal-condensed interface in a Bose gas. In this situation the LDA would in fact require a very high imaging resolution, which is experimentally challenging. With the global approach we overcome this limitation by describing the system undergoing phase transition as a whole and we can provide evidence of the compressibility peak at the transition. On the other hand, the global variables approach needs many measurements for different volumes with the same atom number to trace a single isothermal curve, and this can be experimentally nontrivial. In this sense the LDA has the advantage of leading to a complete isothermal curve from the analysis of a single image. Due to the lack of experimental points, we cannot precisely measure the compressibility in the close vicinity of the phase transition. Nevertheless, the expected sharp peak in $\kappa_T$ near the critical point is quite clear and shares remarkable similarities to the behavior of the isothermal compressibility for liquid helium as observed across the $\lambda$-point. # Conclusions In this article, we have used the concept of global thermodynamic variables to measure, the most appropriate susceptibility to understand the phase transition, the isothermal compressibility parameter of a harmonically confined Bose gas. Once the sample had undergone BEC we characterized this phase transition, from the classical to the quantum regime, indicating a second-order transition likely related to a spontaneous symmetry breaking. The concept of using global variables to determine the global compressibility is quite useful in situations where the LDA cannot be applied. In another more complex physical systems in which there are abrupt changes in the density are of interest for superfluid physics, such as vortices, vortex lattices, solitons, inter alia, and, especially, superfluid turbulence, recently demonstrated by our group. In this case the local variables do not make sense and the global behavior in the compressibility may indicate new characteristics of the turbulent regime. Such an investigation is currently in progress. We acknowledge financial support from FAPESP (Brazil), CNPq (Brazil), CAPES (Brazil), and LENS (Italy).	{'timestamp': '2016-07-19T02:08:09', 'yymm': '1607', 'arxiv_id': '1607.04897', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04897'}
null	null
# Introduction {#introduction .unnumbered} Carbon nanotubes (CNTs) are promising nanomaterials, which have been extensively studied by many researchers. Due to different combinations of structural variation, CNTs can exhibit a wide range of electronic and optical properties, which can be of great use in the design of novel techniques. CNTs are also polyfunctional macromolecules, where specific reactions can occur at various sites with different efficiencies. There are three major types of CNTs: armchair CNTs, chiral CNTs, and zigzag CNTs, which are distinguished by the geometrical vector ($n$,$m$), with $n$ and $m$ being integers. CNTs can behave as either metals or semiconductors depending on their chiral angles, diameters, and lengths. Therefore, a further investigation of how these factors affect the properties of CNTs is essential for the comprehensive understanding of these materials. In particular, it is useful to study the basic repeating units of CNTs, which still need further fundamental research exploration. The targeting units of the present study, a series of $n$-cyclacenes, consisting of $n$ fused benzene rings forming a closed loop (see ), are the shortest ($n$,0) zigzag CNTs with hydrogen passivation, which have attracted considerable interest in the research community due to their fascinating electronic properties. As $n$-cyclacenes belong to the category of cata-condensed aromatics (i.e., molecules that have no carbon atoms belonging to more than two rings), each carbon atom is on the periphery of the conjugated system. Before $n$-cyclacenes are intensively connected to zigzag CNTs, they have been studied mainly due to the research curiosity in highly conjugated cyclic systems. The studies of $n$-cyclacenes can also be important for atomic-level structural control in the synthesis of CNTs. In addition, bottom-up approaches to the synthesis of CNTs not only provide a fundamental understanding of the relationship between the design of CNTs and their electronic properties, but also greatly lower the synthetic temperatures. While zigzag CNTs may be synthesized from cycloarylenes by devising the cutout positions of CNTs, it remains important to systematically investigate the properties of $n$-cyclacenes, which can be useful for exploring the possible utility of their cylindrical cavities in host-guest chemistry. The structure of $n$-cyclacene has two types of components: an arenoid belt (composed of fused benzene rings) and two peripheral circuits (the top and bottom peripheral circuits). The peripheral circuits are of two types: $4k$ and $4k+2$ (where $k$ is an integer), depending on the number of benzene rings in $n$-cyclacene. In previous studies, it has been shown that $n$-cyclacene with even-number benzene rings ($4k$ type) is more stable than that with odd-number benzene rings ($4k+2$ type). Therefore, the nature of peripheral circuits (i.e., the cryptoannulenic effect) is expected to be responsible for the properties of $n$-cyclacene. Besides, the structure of $n$-cyclacene can also be regarded as two fused trannulenes (i.e., circular, all-trans cyclic polyene ribbons). From the bond length analysis of $n$-cyclacene, there is bond length alternation in the benzene ring, and the aromaticity is reduced due to the structural strain, which can hence be responsible for the properties of $n$-cyclacene. Even though there has been a keen interest in $n$-cyclacenes, the studies of their electronic properties are scarce. While $n$-cyclacene may be synthesized via an intramolecular cyclization of $n$-acene (a chain-like molecule with $n$ linearly fused benzene rings, e.g., see Figure 1 of Ref. ), the synthetic procedure has been very challenging, and has not succeeded in producing pure $n$-cyclacene, possibly due to its highly strained structure and highly reactive nature. As the stabilities of annulated polycyclic saturated hydrocarbons decrease rapidly with the number of fused benzene rings, the synthesis of larger $n$-cyclacenes should be even more difficult. To date, the reported properties of $n$-cyclacenes are based on theoretical calculations. Nevertheless, accurate prediction of the electronic properties of larger $n$-cyclacenes has been very challenging for traditional electronic structure methods, due to the presence of strong static correlation effects. Kohn-Sham density functional theory (KS-DFT) with conventional (i.e., semilocal, hybrid, and double-hybrid ) exchange-correlation (XC) density functionals can yield unreliable results for systems with strong static correlation effects. High-level ab initio multi-reference methods are typically required to accurately predict the properties of larger $n$-cyclacenes. However, as the number of electrons in $n$-cyclacene quickly increases with increasing $n$, there have been very few studies on the properties of larger $n$-cyclacenes using multi-reference methods, due to their prohibitively high cost. To circumvent the formidable computational expense of high-level ab initio multi-reference methods, we have recently developed thermally-assisted-occupation density functional theory (TAO-DFT), a very efficient electronic structure method for studying the properties of large ground-state systems (e.g., containing up to a few thousand electrons) with strong static correlation effects. In contrast to KS-DFT, TAO-DFT is a density functional theory with fractional orbital occupations, wherein strong static correlation is explicitly described by the entropy contribution (see Eq. (26) of Ref. ), a function of the fictitious temperature and orbital occupation numbers. Note that the entropy contribution is completely missing in KS-DFT. Recently, we have studied the electronic properties of zigzag graphene nanoribbons (ZGNRs) using TAO-DFT. The ground states of ZGNRs are found to be singlets for all the widths and lengths studied. The longer ZGNRs should possess increasing polyradical character in their ground states, with the active orbitals being mainly localized at the zigzag edges. Our results are in good agreement with the available experimental and highly accurate ab initio data. Besides, on the basis of our TAO-DFT calculations, the active orbital occupation numbers for the ground states of ZGNRs should exhibit a curve crossing behavior in the approach to unity (singly occupied) with increasing ribbon length. Very recently, the curve crossing behavior has been confirmed by highly accurate ab initio multi-reference methods ! TAO-DFT has similar computational cost as KS-DFT for single-point energy and analytical nuclear gradient calculations, and reduces to KS-DFT in the absence of strong static correlation effects. Besides, existing XC density functionals in KS-DFT may also be adopted in TAO-DFT. Relative to high-level ab initio multi-reference methods, TAO-DFT is computationally efficient, and hence very powerful for the study of large polyradical systems. In addition, the orbital occupation numbers from TAO-DFT, which are intended to simulate the natural orbital occupation numbers (NOONs) \[i.e., the eigenvalues of one-electron reduced density matrix\], can be very useful for assessing the possible polyradical character of systems. Recent studies have demonstrated that the orbital occupation numbers from TAO-DFT are qualitatively similar to the NOONs from high-level ab initio multi-reference methods, giving promise for applying TAO-DFT to large polyradical systems. Due to its computational efficiency and reasonable accuracy for large systems with strong static correlation effects, in this work, TAO-DFT is adopted to study the electronic properties of $n$-cyclacenes ($n$ = 4--100). As $n$-cyclacenes have not been successfully synthesized, no experimental data are currently available for comparison. Therefore, our results are compared with the available high-level ab initio data as well as those obtained from various XC density functionals in KS-DFT. In addition, as $n$-cyclacene can be considered as an interconnection of $n$-acene, the electronic properties of $n$-cyclacene are also compared with those of $n$-acene to assess the role of cyclic topology. # Computational Details {#computational-details .unnumbered} All calculations are performed with a development version of [Q-Chem 4.0]{.sans-serif}, using the 6-31G(d) basis set with the fine grid EML(75,302), consisting of 75 Euler-Maclaurin radial grid points and 302 Lebedev angular grid points. Results are calculated using KS-LDA (i.e., KS-DFT with the LDA XC density functional ) and TAO-LDA (i.e., TAO-DFT with the LDA XC density functional and the LDA $\theta$-dependent density functional $E_{\theta}^{\text {LDA}}$ (see Eq. (41) of Ref. ) with the fictitious temperature $\theta$ = 7 mhartree (as defined in Ref. ). Note that KS-LDA is simply TAO-LDA with $\theta$ = 0, and hence it is important to assess the performance of KS-LDA here to assess the significance of TAO-LDA. The ground state of $n$-cyclacene/$n$-acene ($n$ = 4--100) is obtained by performing spin-unrestricted KS-LDA and TAO-LDA calculations for the lowest singlet and triplet energies of $n$-cyclacene/$n$-acene on the respective geometries that were fully optimized at the same level of theory. The singlet-triplet energy (ST) gap of $n$-cyclacene/$n$-acene is calculated as $(E_{\text{T}}-E_{\text{S}})$, the energy difference between the lowest triplet (T) and singlet (S) states of $n$-cyclacene/$n$-acene. # Results and Discussion {#results-and-discussion .unnumbered} ## Singlet-Triplet Energy Gap {#singlet-triplet-energy-gap .unnumbered} shows the ST gap of $n$-cyclacene as a function of the number of benzene rings, calculated using spin-unrestricted KS-LDA and TAO-LDA. The results are compared with the available data, calculated using the complete-active-space second-order perturbation theory (CASPT2) (a high-level ab initio multi-reference method) as well as the M06L functional (a popular semilocal XC density functional) and the B3LYP functional (a popular hybrid XC density functional) in KS-DFT. As can be seen, the anticipated even-odd oscillations in the ST gaps may be attributed to the cryptoannulenic effects of $n$-cyclacenes. However, the amplitudes of the even-odd oscillations are considerably larger for KS-DFT with the XC functionals, which are closely related to the degree of spin contamination (as discussed in Ref. ). In general, the larger fraction of Hartree-Fock (HF) exchange adopted in the XC functional in KS-DFT, the higher the degree of spin contamination for systems with strong static correlation effects. For example, the ST gap obtained with KS-B3LYP is unexpectedly large at $n$ = 10, due to the high degree of spin contamination. On the other hand, as commented in Ref. the ST gaps obtained with CASPT2 are rather sensitive to the choice of active space. Since the complete $\pi$-valence space was not selected as the active space (due to the prohibitively high cost), the CASPT2 results here should be taken with caution. Recent studies have shown that a sufficiently large active space should be adopted in high-level ab initio multi-reference calculations for accurate prediction of the electronic properties of systems with strong static correlation effects, which can, however, be prohibitively expensive for large systems. Note that the ST gap obtained with CASPT2 unexpectedly increases at $n$ = 12, possibly due to the insufficiently large active space adopted in the calculations. To assess the role of cyclic topology, show the ST gap of $n$-cyclacene/$n$-acene as a function of the number of benzene rings, calculated with spin-unrestricted TAO-LDA. Similar to $n$-acenes, the ground states of $n$-cyclacenes remain singlets for all the cases investigated. In contrast to $n$-acene, the ST gap of $n$-cyclacene, however, displays oscillatory behavior for small $n$, and the oscillation vanishes gradually with increasing $n$. For small $n$, $n$-cyclacene with even-number benzene rings exhibits a larger ST gap (i.e., greater stability) than that with odd-number benzene rings. For sufficiently large $n$ ($n > 30$), the ST gap of $n$-cyclacene converges monotonically from below to the ST gap of $n$-acene (which monotonically decreases with increasing $n$). At the level of TAO-LDA, the ST gaps of the largest $n$-cyclacene and $n$-acene studied (i.e., $n$ = 100) are essentially the same (0.49 kcal/mol). On the basis of the ST gaps obtained with TAO-LDA, the cryptoannulenic effect and structural strain of $n$-cyclacene are more important for the smaller $n$, and less important for the larger $n$. Due to the symmetry constraint, the spin-restricted and spin-unrestricted energies for the lowest singlet state of $n$-cyclacene/$n$-acene, calculated using the exact theory, should be identical. Recent studies have shown that KS-DFT with conventional XC density functionals cannot satisfy this condition for the larger $n$-cyclacene/$n$-acene, due to the aforementioned spin contamination. To assess the possible symmetry-breaking effects, spin-restricted TAO-LDA calculations are also performed for the lowest singlet energies on the respective optimized geometries. Within the numerical accuracy of our calculations, the spin-restricted and spin-unrestricted TAO-LDA energies for the lowest singlet state of $n$-cyclacene/$n$-acene are essentially the same (i.e., essentially no unphysical symmetry-breaking effects occur in our spin-unrestricted TAO-LDA calculations). ## Vertical Ionization Potential, Vertical Electron Affinity, and Fundamental Gap {#vertical-ionization-potential-vertical-electron-affinity-and-fundamental-gap .unnumbered} At the lowest singlet state (i.e., the ground-state) geometry of $n$-cyclacene/$n$-acene (containing $N$ electrons), TAO-LDA is adopted to calculate the vertical ionization potential $\text{IP}_{v}={E}_{N-1}-{E}_{N}$, vertical electron affinity $\text{EA}_{v}={E}_{N}-{E}_{N+1}$, and fundamental gap $E_{g}=\text{IP}_{v}-\text{EA}_{v}={E}_{N+1}+{E}_{N-1}-2{E}_{N}$ using multiple energy-difference methods, with ${E}_{N}$ being the total energy of the $N$-electron system. With increasing number of benzene rings in $n$-cyclacene, $\text{IP}_{v}$ oscillatorily decreases (see ), $\text{EA}_{v}$ oscillatorily increases (see ), and hence $E_{g}$ oscillatorily decreases (see ). However, these oscillations are damped and eventually disappear with increasing $n$. For sufficiently large $n$ ($n > 30$), the $\text{IP}_{v}$ and $E_{g}$ values of $n$-cyclacene converge monotonically from above to those of $n$-acene (which monotonically decrease with increasing $n$), while the $\text{EA}_{v}$ value of $n$-cyclacene converges monotonically from below to that of $n$-acene (which monotonically increases with increasing $n$). Note also that the $E_{g}$ value of $n$-cyclacene ($n$ = 13--54) is within the most interesting range (1 to 3 eV), giving promise for applications of $n$-cyclacenes in nanophotonics. ## Symmetrized von Neumann Entropy {#symmetrized-von-neumann-entropy .unnumbered} To investigate the possible polyradical character of $n$-cyclacene/$n$-acene, we calculate the symmetrized von Neumann entropy (e.g., see Eq. (9) of Ref. ) \[\label{eq1} S_{\text{vN}} =-\frac{1}{2} \sum_{i=1}^{\infty} \bigg\lbrace f_{i}\ \text{ln}(f_{i}) + (1-f_{i})\ \text{ln}(1-f_{i}) \bigg\rbrace,\] for the lowest singlet state of $n$-cyclacene/$n$-acene as a function of the number of benzene rings, using TAO-LDA. Here, $f_{i}$ the occupation number of the $i^{\text{th}}$ orbital obtained with TAO-LDA, which ranges from 0 to 1, is approximately the same as the occupation number of the $i^{\text{th}}$ natural orbital. For a system without strong static correlation ($\{f_{i}\}$ are close to either 0 or 1), $S_{\text{vN}}$ provides insignificant contributions, while for a system with strong static correlation ($\{f_{i}\}$ are fractional for active orbitals and are close to either 0 or 1 for others), $S_{\text{vN}}$ increases with the number of active orbitals. As shown in , the $S_{\text{vN}}$ value of $n$-cyclacene oscillatorily increases with increasing number of benzene rings. Nonetheless, the oscillation is damped and eventually disappears with the increase of $n$. For sufficiently large $n$ ($n > 30$), the $S_{\text{vN}}$ value of $n$-cyclacene converges monotonically from above to that of $n$-acene (which monotonically increases with increasing $n$). Therefore, similar to $n$-acenes, the larger $n$-cyclacenes should possess increasing polyradical character. ## Active Orbital Occupation Numbers {#active-orbital-occupation-numbers .unnumbered} To illustrate the causes of the increase of $S_{\text{vN}}$ with $n$, we plot the active orbital occupation numbers for the lowest singlet state of $n$-cyclacene as a function of the number of benzene rings, calculated using TAO-LDA. Here, the highest occupied molecular orbital (HOMO) is the ${(N/2)}^{\text{th}}$ orbital, and the lowest unoccupied molecular orbital (LUMO) is the ${(N/2 + 1)}^{\text{th}}$ orbital, where $N$ is the number of electrons in $n$-cyclacene. For brevity, HOMO, HOMO$-$`<!-- -->`{=html}1,\..., and HOMO$-$`<!-- -->`{=html}15, are denoted as H, H$-$`<!-- -->`{=html}1,\..., and H$-$`<!-- -->`{=html}15, respectively, while LUMO, LUMO+1,\..., and LUMO+15, are denoted as L, L+1,\..., and L+15, respectively. As presented in , the number of fractionally occupied orbitals increases with increasing cyclacene size, clearly indicating that the polyradical character of $n$-cyclacene indeed increases with the cyclacene size. Similar to the previously discussed properties, the active orbital occupation numbers of $n$-cyclacene also exhibit oscillatory behavior, showing wave-packet oscillations. ## Real-Space Representation of Active Orbitals {#real-space-representation-of-active-orbitals .unnumbered} For the lowest singlet states of some representative $n$-cyclacenes ($n$ = 4--7), we explore the real-space representation of active orbitals (e.g., HOMOs and LUMOs), obtained with TAO-LDA. Similar to previous findings for $n$-acenes, the HOMOs and LUMOs of $n$-cyclacenes are mainly localized at the peripheral carbon atoms (see ). # Conclusions {#conclusions .unnumbered} In conclusion, we have studied the electronic properties of $n$-cyclacenes ($n$ = 4--100), including the ST gaps, vertical ionization potentials, vertical electron affinities, fundamental gaps, symmetrized von Neumann entropy, active orbital occupation numbers, and real-space representation of active orbitals, using our newly developed TAO-DFT, a very efficient electronic structure method for the study of large systems with strong static correlation effects. To assess the effects of cyclic nature, the electronic properties of $n$-cyclacenes have also been compared with those of $n$-acenes. Similar to $n$-acenes, the ground states of $n$-cyclacenes are singlets for all the cases investigated. In contrast to $n$-acenes, the electronic properties of $n$-cyclacenes, however, display oscillatory behavior for small $n$ ($n \le 30$) in the approach to the corresponding properties of $n$-acenes with increasing number of benzene rings, which to the best of our knowledge have never been addressed in the literature. The oscillatory behavior may be related to the cryptoannulenic effect and structural strain of $n$-cyclacene, which have been shown to be important for small $n$, and unimportant for sufficiently large $n$. On the basis of several measures (e.g., the smaller ST gap, the smaller $E_{g}$, and the larger $S_{\text{vN}}$), for small $n$, $n$-cyclacene with odd-number benzene rings should possess stronger radical character than that with even-number benzene rings. In addition, based on the calculated orbitals and their occupation numbers, the larger $n$-cyclacenes are expected to possess increasing polyradical character in their ground states, where the active orbitals are mainly localized at the peripheral carbon atoms. Since TAO-DFT is computationally efficient, it appears to be a promising method for studying the electronic properties of large systems with strong static correlation effects. Nevertheless, as with all approximate electronic structure methods, a few limitations remain. Relative to the exact full configuration interaction (FCI) method, TAO-LDA (with $\theta$ = 7 mhartree) is not variationally correct (i.e., overcorrelation can occur), and hence, the orbital occupation numbers from TAO-LDA may not be the same as the NOONs from the FCI method. To assess the accuracy of our TAO-LDA results, as the computational cost of the FCI method is prohibitive, the electronic properties of $n$-cyclacenes from relatively affordable ab initio multi-reference methods are called for. # Author Contributions {#author-contributions .unnumbered} C.-S.W. and P.-Y.L. contributed equally to this work. J.-D.C. conceived and designed the project. C.-S.W. and J.-D.C. performed the calculations. P.-Y.L. and J.-D.C. wrote the paper. All authors performed the data analysis.	{'timestamp': '2016-10-27T02:06:29', 'yymm': '1607', 'arxiv_id': '1607.04900', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04900'}
null	null
null	null
# Introduction Artificial neural networks are currently considered the state of the art in applications ranging from image classification, to speech recognition and even machine translation. However, little is understood about the process by which they are trained for supervised learning tasks. The problem of optimizing their parameters is an active area both practical and theoretical research. Despite considerable sensitivity to initialization and choice of hyperparameters, neural networks often achieve compelling results after optimization by gradient descent methods. Due to the nonconvexity and massive parameter space of these functions, it is poorly understood how these sub-optimal methods have proven so successful. Indeed, training a certain kind of neural network is known to be NP-Complete, making it difficult to provide any worst-case training guarantees. Much recent work has attempted to reconcile these differences between theory and practice. This article attempts a modest step towards understanding the dynamics of the training procedure. We establish three main convexity results for a certain class of neural network, which is the current the state of the art. First, that the objective is piecewise convex as a function of the input data, with parameters fixed, which corresponds to the behavior at test time. Second, that the objective is again piecewise convex as a function of the parameters of a single layer, with the input data and all other parameters held constant. Third, that the training objective function, for which all parameters are variable but the input data is fixed, is piecewise multi-convex. That is, it is a continuous function which can be represented by a finite number of multi-convex functions, each active on a multi-convex parameter set. This generalizes the notion of biconvexity found in the optimization literature to piecewise functions and arbitrary index sets. To prove these results, we require two main restrictions on the definition of a neural network: that its layers are piecewise affine functions, and that its objective function is convex and continuously differentiable. Our definition includes many contemporary use cases, such as least squares or logistic regression on a convolutional neural network with rectified linear unit (ReLU) activation functions and either max- or mean-pooling. In recent years these networks have mostly supplanted the classic sigmoid type, except in the case of recurrent networks. We make no assumptions about the training data, so our results apply to the current state of the art in many practical scenarios. Piecewise multi-convexity allows us to characterize the extrema of the training objective. As in the case of biconvex functions, stationary points and local minima are guaranteed optimality on larger sets than we would have for general smooth functions. Specifically, these points are partial minima when restricted to the relevant piece. That is, they are points for which no decrease can be made in the training objective without simultaneously varying the parameters across multiple layers, or crossing the boundary into a different piece of the function. Unlike global minima, we show that partial minima are reliably found by the optimization algorithms used in current practice. Finally, we provide some guarantees for solving general multi-convex optimization problems by various algorithms. First we analyze gradient descent, proving necessary convergence conditions. We show that every point to which gradient descent converges is a piecewise partial minimum, excepting some boundary conditions. To prove stronger results, we define a different optimization procedure breaking each parameter update into a number of convex sub-problems. For this procedure, we show both necessary and sufficient conditions for convergence to a piecewise partial minimum. Interestingly, adding regularization to the training objective is all that is needed to prove necessary conditions. Similar results have been independently established for many kinds of optimization problems, including bilinear and biconvex optimization, and in machine learning the special case of linear autoencoders. Our analysis extends existing results on alternating convex optimization to the case of arbitrary index sets, and general multi-convex point sets, which is needed for neural networks. We admit biconvex problems, and therefore linear autoencoders, as a special case. Despite these results, we find that it is difficult to pass from partial to global optimality results. Unlike the encouraging case of linear autoencoders, we show that a single rectifier neuron, under the squared error objective, admits arbitrarily poor local minima. This suggests that much work remains to be done in understanding how sub-optimal methods can succeed with neural networks. Still, piecewise multi-convex functions are in some senses easier to minimize than the general class of smooth functions, for which none of our previous guarantees can be made. We hope that our characterization of neural networks could contribute to a better understanding of these important machine learning systems. # Preliminary material We begin with some preliminary definitions and basic results concerning continuous piecewise functions. Note that this definition of piecewise convexity differs from that found in the convex optimization literature, which focuses on convex piecewise convex functions, i.e. maxima of convex functions. Note also that we do not claim a unique representation in terms of active functions $g_{k}$ and pieces $S_{k}$, only that there exists at least one such representation. Before proceeding, we shall extend definition [\[def:piecewise_affine\]](#def:piecewise_affine){reference-type="ref" reference="def:piecewise_affine"} to functions of multidimensional codomain for the affine case. First, we prove an intuitive statement about the geometry of the pieces of continuous piecewise affine functions. See figure [\[fig:bad_local_minimum\]](#fig:bad_local_minimum){reference-type="ref" reference="fig:bad_local_minimum"} in section [\[sec:Local-minima\]](#sec:Local-minima){reference-type="ref" reference="sec:Local-minima"} for an example of this result on a specific neural network. Our next result concerns the composition of piecewise functions, which is essential for the later sections. We now turn to continuous piecewise convex functions, of which continuous piecewise affine functions are a subset. Our final theorem concerns the arithmetic mean of continuous piecewise convex functions, which is essential for the analysis of neural networks. The proof takes the form of two lemmas. Having established that continuous piecewise convexity is closed under addition and positive scalar multiplication, we can see that it is closed under the arithmetic mean, which is just the composition of these two operations. # Neural networks[\[sec:Neural-networks\]]{#sec:Neural-networks label="sec:Neural-networks"} In this work, we define a neural network to be a composition of functions of two kinds: a convex continuously differentiable objective (or loss) function $h$, and continuous piecewise affine functions $g_{1},g_{2}...,g_{N}$, constituting the $N$ layers. Furthermore, the outermost function must be $h$, so that we have \[f=h\circ g_{N}\circ g_{N-1}\circ...\circ g_{1}\] where $f$ denotes the entire network. This definition is not as restrictive as it may seem upon first glance. For example, it is easily verified that the rectified linear unit (ReLU) neuron is continuous piecewise affine, as we have \[g(\boldsymbol{x})=\max(\boldsymbol{0},A\boldsymbol{x}+\boldsymbol{b}),\] where the maximum is taken pointwise. It can be shown that maxima and minima of affine functions are piecewise affine. This includes the convolutional variant, in which $A$ is a Toeplitz matrix. Similarly, max pooling is continuous piecewise linear, while mean pooling is simply linear. Furthermore, many of the objective functions commonly seen in machine learning are convex and continuously differentiable, as in least squares and logistic regression. Thus this seemingly restrictive class of neural networks actually encompasses the current state of the art. By theorem [\[theorem:composition_piecewise_affine\]](#theorem:composition_piecewise_affine){reference-type="ref" reference="theorem:composition_piecewise_affine"}, the composition of all layers $g=g_{N}\circ g_{N-1}\circ...\circ g_{1}$ is continuous piecewise affine. Therefore, a neural network is ultimately the composition of a continuous convex function with a single continuous piecewise affine function. Thus by theorem [\[theorem:covex_piecewise\]](#theorem:covex_piecewise){reference-type="ref" reference="theorem:covex_piecewise"} the network is continuous piecewise convex. Figure [\[fig:network\]](#fig:network){reference-type="ref" reference="fig:network"} provides a visualization of this result for the example network \[f(x,y)=\left(2-\left[\left[x-y\right]_{+}-\left[x+y\right]_{+}+1\right]_{+}\right)^{2},\label{eq:example_network}\] where $\left[x\right]_{+}=\max(x,0)$. For clarity, this is just the two-layer ReLU network \[f(x,y,z)=\left(z-\left[a_{5}\left[a_{1}x+a_{2}y\right]_{+}+a_{6}\left[a_{3}x+a_{4}y\right]_{+}+b_{1}\right]_{+}\right)^{2}\] with the squared error objective and a single data point $((x,y),z)$, setting $z=2$ and $a_{2}=a_{6}=-1$, with all other parameters set to $1$. Before proceeding further, we must define a special kind of differentiability for piecewise continuous functions, and show that this holds for neural networks. To see that neural networks are piecewise continuously differentiable, note that the objective $h$ is continuously differentiable, as are the affine active functions of the layers. Thus their composition is continuously differentiable. It follows that non-differentiable points are found only on the boundaries between pieces. # Network parameters of a single layer[\[sec:Network-parameters-of-single-layer\]]{#sec:Network-parameters-of-single-layer label="sec:Network-parameters-of-single-layer"} In the previous section we have defined neural networks as functions of labeled data. These are the functions relevant during testing, where parameters are constant and data is variable. In this section, we extend these results to the case where data is constant and parameters are variable, which is the function to optimized during training. For example, consider the familiar equation \[f=(ax+b-y)^{2}\] with parameters $(a,b)$ and data $(x,y$). During testing, we hold $(a,b)$ constant, and consider $f$ as a function of the data $(x,y)$. During training, we hold $(x,y)$ constant and consider $f$ as a function of the parameters $(a,b)$. This is what we mean when we say that a network is being "considered as a function of its parameters[^1]." This leads us to an additional stipulation on our definition of a neural network. That is, each layer must be piecewise affine as a function of its parameters as well. This is easily verified for all of the layer types previously mentioned. For example, with the ReLU neuron we have \[f(A,\boldsymbol{b})=\left[A\boldsymbol{x}+\boldsymbol{b}\right]_{+}\label{eq:ReLU}\] so for $\left(A\boldsymbol{x}+\boldsymbol{b}\right)_{k}\ge0$ we have that the $k^{th}$ component of $f$ is linear in $(A,\boldsymbol{b})$, while for $\left(A\boldsymbol{x}+\boldsymbol{b}\right)_{k}<0$ it is constant. To see this, we can re-arrange the elements of $A$ into a column vector $\boldsymbol{a}$, in row-major order, so that we have \[\begin{aligned} A\boldsymbol{x}+\boldsymbol{b} & =\begin{pmatrix}\boldsymbol{x^{T}} & \boldsymbol{0^{T}} &... &... & \boldsymbol{0^{T}} & \boldsymbol{1}^{T}\\ \boldsymbol{0^{T}} & \boldsymbol{x^{T}} & \boldsymbol{0}^{T} &... & \boldsymbol{0}^{T} & \boldsymbol{1}^{T}\\... &... &... &... &... &...\\ \boldsymbol{0}^{T} &... &... & \boldsymbol{0}^{T} & \boldsymbol{x}^{T} & \boldsymbol{1}^{T} \end{pmatrix}\begin{pmatrix}\boldsymbol{a}\\ \boldsymbol{b} \end{pmatrix}.\label{eq:ReLU_expanded} \end{aligned}\] In section [\[sec:Neural-networks\]](#sec:Neural-networks){reference-type="ref" reference="sec:Neural-networks"} we have said that a neural network, considered as a function of its input data, is convex and continuously differentiable on each piece. Now, a neural network need not be piecewise convex as a function of the entirety of its parameters[^2]. However, we can regain piecewise convexity by considering it only as a function of the parameters in a single layer, all others held constant. We conclude this section with a simple remark which will be useful in later sections. Let $f_{m}$ be a neural network, considered as a function of the parameters of the $m^{th}$ layer, and let $S$ be a piece of $f_{m}$. Then the optimization problem \[\begin{aligned} \mbox{minimize } & f_{m}(\boldsymbol{x})\nonumber \\ \mbox{subject to } & \boldsymbol{x}\in S\label{eq:convex_single_layer} \end{aligned}\] is convex. # Network parameters of multiple layers[\[sec:Network-parameters-of-multiple-layers\]]{#sec:Network-parameters-of-multiple-layers label="sec:Network-parameters-of-multiple-layers"} In the previous section we analyzed the convexity properties of neural networks when optimizing the parameters of a single layer, all others held constant. Now we are ready to extend these results to the ultimate goal of simultaneously optimizing all network parameters. Although not convex, the problem has a special convex substructure that we can exploit in proving future results. We begin by defining this substructure for point sets and functions. In other words, $S_{I}(\boldsymbol{x})$ is the subset of $S$ for which every point is equal to $\boldsymbol{x}$ in the components not indexed by $I$. Note that this differs from the typical definition, which is the intersection of a set with a hyperplane. For example, $\mathbb{R}_{\{1\}}^{3}(\boldsymbol{0})$ is the $x$-axis, whereas $\mathbb{R}_{\{1,2\}}^{3}(\boldsymbol{0})$ is the $xy$-plane. Note also that cross-sections are not unique, for example $\mathbb{R}_{\{1,2\}}^{3}(0,0,0)=\mathbb{R}_{\{1,2\}}^{3}(1,2,0)$. In this case the first two components of the cross section are irrelevant, but we will maintain them for notational convenience. We can now apply this concept to functions on $\mathbb{R}^{n}$. This formalizes the notion of restricting a non-convex function to a variable subset on which it is convex, as in section [\[sec:Network-parameters-of-single-layer\]](#sec:Network-parameters-of-single-layer){reference-type="ref" reference="sec:Network-parameters-of-single-layer"} when a neural network was restricted to the parameters of a single layer. For example, let $f(x,y,z)=xy+z$, and let $I_{1}=\{1,3\}$, and $I_{2}=\{2,3\}$. Then $f_{1}(x,y_{0},z)$ is a convex function of $(x,z)$ with $y$ fixed at $y_{0}$. Similarly, $f_{2}(x_{0},y,z)$ is a convex function of $(y,z)$ with $x$ fixed at $x_{0}$. Thus $f$ is multi-convex with respect to $\mathcal{I}=\{I_{1},I_{2}\}$. To fully define a multi-convex optimization problem, we introduce a similar concept for point sets. This generalizes the notion of biconvexity found in the optimization literature. From here, we can extend definition [\[def:piecewise_affine\]](#def:piecewise_affine){reference-type="ref" reference="def:piecewise_affine"} to multi-convex functions. However, we will drop the topological restrictions on the pieces of our function, since multi-convex sets need not be connected. From this definition, it is easily verified that a continuous piecewise multi-convex function $f:\mathbb{R}^{m}\rightarrow\mathbb{R}^{n}$ admits a representation where all pieces are multi-convex, as in the proof of theorem [\[theorem:convex_polytope\]](#theorem:convex_polytope){reference-type="ref" reference="theorem:convex_polytope"}. Before we can extend the results of section [\[sec:Network-parameters-of-single-layer\]](#sec:Network-parameters-of-single-layer){reference-type="ref" reference="sec:Network-parameters-of-single-layer"} to multiple layers, we must add one final constraint on the definition of a neural network. That is, each of the layers must be continuous piecewise multi-convex, considered as functions of both the parameters and the input. Again, this is easily verified for the all of the layer types previously mentioned. We have already shown they are piecewise convex on each cross-section, taking our index sets to separate the parameters from the input data. It only remains to show that the number of pieces is finite. The only layer which merits consideration is the ReLU, which we can see from equation [\[eq:ReLU\]](#eq:ReLU){reference-type="ref" reference="eq:ReLU"} consists of two pieces for each component: the "dead" or constant region, with $(A\boldsymbol{x})_{j}+b_{j}<0$, and its compliment. With $n$ components we have at most $2^{n}$ pieces, corresponding to binary assignments of "dead" or "alive" for each component. Having said that each layer is continuous piecewise multi-convex, we can extend these results to the whole network. We begin the proof with a lemma for more general multi-convex functions. Our next lemma extends theorem [\[theorem:mean_piecewise_convex\]](#theorem:mean_piecewise_convex){reference-type="ref" reference="theorem:mean_piecewise_convex"} to multi-convex functions. We can now prove the theorem. In the coming sections, we shall see that multi-convexity allows us to give certain guarantees about the convergence of various optimization algorithms. But first, we shall prove some basic results independent of the optimization procedure. These results were summarized by Gorksi et al. for the case of biconvex differentiable functions. Here we extend them to piecewise functions and arbitrary index sets. First we define a special type of minimum relevant for multi-convex functions. In other words, $\boldsymbol{x}_{0}$ is a partial minimum of $f$ with respect to $\mathcal{I}$ if it minimizes $f$ on every cross-section of $S$ intersecting $\boldsymbol{x}_{0}$, as shown in figure [\[fig:cross_section\]](#fig:cross_section){reference-type="ref" reference="fig:cross_section"}. By convexity, these points are intimately related to the stationary points of $f$. It is clear that multi-convexity provides a wealth of results concerning partial minima, while piecewise multi-convexity restricts those results to a subset of the domain. Less obvious is that partial minima of smooth multi-convex functions need not be local minima. An example was pointed out by a reviewer of this work, that the biconvex function $f(x,y)=xy$ has a partial minimum at the origin which is not a local minimum. However, the converse is easily verified, even in the absence of differentiability. We have seen that for multi-convex functions there is a close relationship between stationary points, local minima and partial minima. For these functions, infinitesimal results concerning derivatives and local minima can be extended to larger sets. However, we make no guarantees about global minima. The good news is that, unlike global minima, we shall see that we can easily solve for partial minima. # Gradient descent[\[sec:Optimization-and-convergence\]]{#sec:Optimization-and-convergence label="sec:Optimization-and-convergence"} In the realm of non-convex optimization, also called global optimization, methods can be divided into two groups: those which can certifiably find a global minimum, and those which cannot. In the former group we sacrifice speed, in the latter correctness. This work focuses on algorithms of the latter kind, called local or sub-optimal methods, as only this type is used in practice for deep neural networks. In particular, the most common methods are variants of gradient descent, where the gradient of the network with respect its parameters is computed by a procedure called backpropagation. Since its explanation is often obscured by jargon, we shall provide a simple summary here. Backpropagation is nothing but the chain rule applied to the layers of a network. Splitting the network into two functions $f=u\circ v$, where $u:\mathbb{R}^{n}\rightarrow\mathbb{R}$, and $v:\mathbb{R}^{m}\rightarrow\mathbb{R}^{n}$, we have \[\nabla f=\nabla u\mathcal{D}v\] where $\mathcal{D}$ denotes the Jacobian operator. Note that here the parameters of $u$ are considered fixed, whereas the parameters of $v$ are variable and the input data is fixed. Thus $\nabla f$ is the gradient of $f$ with respect to the parameters of $v$, if it exists. The special observation is that we can proceed from the top layer of the neural network $g_{N}$ to the bottom $g_{1}$, with $u=g_{N}\circ g_{N-1}\circ...\circ g_{m+1}$, and $v=g_{m}$, each time computing the gradient of $f$ with respect to the parameters of $g_{m}$. In this way, we need only store the vector $\nabla u$ and the matrix $\mathcal{D}v$ can be forgotten at each step. This is known as the "backward pass," which allows for efficient computation of the gradient of a neural network with respect to its parameters. A similar algorithm computes the value of $g_{m-1}\circ g_{m-2}\circ...\circ g_{1}$ as a function of the input data, which is often needed to evaluate $\mathcal{D}v$. First we compute and store $g_{1}$ as a function of the input data, then $g_{2}\circ g_{1}$, and so on until we have $f$. This is known as the "forward pass." After one forward and one backward pass, we have computed $\nabla f$ with respect to all the network parameters. Having computed $\nabla f$, we can update the parameters by gradient descent, defined as follows. Variants of this basic procedure are preferred in practice because their computational cost scales well with the number of network parameters. There are many different ways to choose the step size, but our assumption that $\sum_{k=0}^{\infty}a_{k}=\infty$ covers what is usually done with deep neural networks. Note that we have not defined what happens if $\boldsymbol{x}_{k}\notin S$. Since we are ultimately interested in neural networks on $\mathbb{R}^{n}$, we can ignore this case and say that the sequence diverges. Gradient descent is not guaranteed to converge to a global minimum for all differentiable functions. However, it is natural to ask to which points it can converge. This brings us to a basic but important result. In the convex optimization literature, this simple result is sometimes stated in connection with Zangwill's much more general convergence theorem. Note, however, that unlike Zangwill we state necessary, rather than sufficient conditions for convergence. While many similar results are known, it is difficult to strictly weaken the conditions of theorem [\[theorem:gradient_descent\]](#theorem:gradient_descent){reference-type="ref" reference="theorem:gradient_descent"}. For example, if we relax the condition that $\alpha_{k}$ is not summable, and take $f(x)=x$, then $x_{k}$ will always converge to a non-stationary point. Similarly, if we relax the constraint that $f$ is continuously differentiable, taking $f(x)=\|x\|$ and $a_{k}$ decreasing monotonically to zero, we will always converge to the origin, which is not differentiable. Furthermore, if we have $f(x)=\|x\|$ with $\alpha_{k}$ constant, then $x_{k}$ will not converge for almost all $x_{0}$. It is possible to prove much stronger necessary and sufficient conditions for gradient descent, but these results require additional assumptions about the step size policy as well as the function to be minimized, and possibly even the initialization $\boldsymbol{x}_{0}$. It is worth discussing $f(x)=\|x\|$ in greater detail, since this is a piecewise affine function and thus of interest in our investigation of neural networks. While we have said its only convergence point is not differentiable, it remains subdifferentiable, and convergence results are known for subgradient descent. In this work we shall not make use of subgradients, instead considering descent on a piecewise continuously differentiable function, where the pieces are $x\le0$ and $x\ge0$. Although theorem [\[theorem:gradient_descent\]](#theorem:gradient_descent){reference-type="ref" reference="theorem:gradient_descent"} does not apply to this function, the relevant results hold anyways. That is, $x=0$ is minimal on some piece of $f$, a result which extends to any continuous piecewise convex function, as any saddle point is guaranteed to minimize some piece. Here we should note one way in which this analysis fails in practice. So far we have assumed the gradient $\nabla f$ is precisely known. In practice, it is often prohibitively expensive to compute the average gradient over large datasets. Instead we take random subsamples, in a procedure known as stochastic gradient descent. We will not analyze its properties here, as current results on the topic impose additional restrictions on the objective function and step size, or require different definitions of convergence. Restricting ourselves to the true gradient $\nabla f$ allows us to provide simple proofs applying to an extensive class of neural networks. We are now ready to generalize these results to neural networks. There is a slight ambiguity in that the boundary points between pieces need not be differentiable, nor even sub-differentiable. Since we are interested only in necessary conditions, we will say that gradient descent diverges when $\nabla f(\boldsymbol{x}_{k})$ does not exist. However, our next theorem can at least handle non-differentiable limit points. The first condition of theorem [\[theorem:multi_convex_gradient_descent\]](#theorem:multi_convex_gradient_descent){reference-type="ref" reference="theorem:multi_convex_gradient_descent"} holds for every point in the interior of a piece, and some boundary points. The second condition extends these results to non-differentiable boundary points so long as gradient descent is eventually confined to a single piece of the function. For example, consider the continuous piecewise convex function $f(x)=\min(x,x^{4})$ as shown in figure [\[fig:convergence\]](#fig:convergence){reference-type="ref" reference="fig:convergence"}. When we converge to $x=0$ from the piece $[0,1]$, it is as if we were converging on the smooth function $g(x)=x^{4}$. This example also illustrates an important caveat regarding boundary points: although $x=0$ is an extremum of $f$ on $[0,1]$, it is not an extremum on $\mathbb{R}$. # Iterated convex optimization Although the previous section contained some powerful results, theorem [\[theorem:multi_convex_gradient_descent\]](#theorem:multi_convex_gradient_descent){reference-type="ref" reference="theorem:multi_convex_gradient_descent"} suffers from two main weaknesses, that it is a necessary condition and that it requires extra care at non-differentiable points. It is difficult to overcome these limitations with gradient descent. Instead, we shall define a different optimization technique, from which necessary and sufficient convergence results follow, regardless of differentiability. Iterated convex optimization splits a non-convex optimization problem into a number of convex sub-problems, solving the sub-problems in each iteration. For a neural network, we have shown that the problem of optimizing the parameters of a single layer, all others held constant, is piecewise convex. Thus, restricting ourselves to a given piece yields a convex optimization problem. In this section, we show that these convex sub-problems can be solved repeatedly, converging to a piecewise partial optimum. We call this iterated convex optimization because problem [\[eq:iterated_convex\]](#eq:iterated_convex){reference-type="ref" reference="eq:iterated_convex"} can be divided into convex sub-problems \[\begin{aligned} \mbox{minimize } & f(\boldsymbol{y})\label{eq:convex_subproblem}\\ \mbox{subject to } & \boldsymbol{y}\in S_{I}(\boldsymbol{x}_{k-1}).\nonumber \end{aligned}\] for each $I\in\mathcal{I}$. In this work, we assume the convex sub-problems are solvable, without delving into specific solution techniques. Methods for alternating between solvable sub-problems have been studied by many authors, for many different types of sub-problems. In the context of machine learning, the same results have been developed for the special case of linear autoencoders. Still, extra care must be taken in extending these results to arbitrary index sets. The key is that $\boldsymbol{x}_{k}$ is not updated until all sub-problems have been solved, so that each iteration consists of solving $m$ convex sub-problems. This is equivalent to the usual alternating convex optimization for biconvex functions, where $\mathcal{I}$ consists of two sets, but not for general multi-convex functions. Some basic convergence results follow immediately from the solvability of problem [\[eq:iterated_convex\]](#eq:iterated_convex){reference-type="ref" reference="eq:iterated_convex"}. First, note that $\boldsymbol{x}_{k-1}$ is a feasible point, so we have $f(\boldsymbol{x}_{k})\le f(\boldsymbol{x}_{k-1})$. This implies that $\lim_{k\rightarrow\infty}f(\boldsymbol{x}_{k})$ exists, so long as $f$ is bounded below. However, this does not imply the existence of $\lim_{k\rightarrow\infty}\boldsymbol{x}_{k}$. See Gorski et al. for an example of a biconvex function on which $\boldsymbol{x}_{k}$ diverges. To prove stronger convergence results, we introduce regularization to the objective. In theorem [\[theorem:regularization\]](#theorem:regularization){reference-type="ref" reference="theorem:regularization"}, the function $g$ is called the regularized version of $f$. In practice, regularization often makes a non-convex optimization problem easier to solve, and can reduce over-fitting. The theorem shows that iterated convex optimization on a regularized function always has at least one convergent subsequence. Next, we shall establish some rather strong properties of the limits of these subsequences. The previous theorem is an extension of results reviewed in Gorski et al. to arbitrary index sets. While Gorski et al. explicitly constrain the domain to a compact biconvex set, we show that regularization guarantees $\boldsymbol{x}_{k}$ cannot escape a certain compact set, establishing the necessary condition for convergence. Furthermore, our results hold for general multi-convex sets, while the earlier result is restricted to Cartesian products of compact sets. These results for iterated convex optimization are considerably stronger than what we have shown for gradient descent. While any bounded sequence in $\mathbb{R}^{n}$ has a convergent subsequence, and we can guarantee boundedness for some variants of gradient descent, we cannot normally say much about the limits of subsequences. For iterated convex optimization, we have shown that the limit of any subsequence is a partial minimum, and all limits of subsequences are equal in objective value. For all practical purposes, this is just as good as saying that the original sequence converges to partial minimum. # Global optimization[\[sec:Local-minima\]]{#sec:Local-minima label="sec:Local-minima"} Although we have provided necessary and sufficient conditions for convergence of various optimization algorithms on neural networks, the points of convergence need only minimize cross-sections of pieces of the domain. Of course we would prefer results relating the points of convergence to global minima of the training objective. In this section we illustrate the difficulty of establishing such results, even for the simplest of neural networks. In recent years much work has been devoted to providing theoretical explanations for the empirical success of deep neural networks, a full accounting of which is beyond the scope of this article. In order to simplify the problem, many authors have studied linear neural networks, in which the layers have the form $g(\boldsymbol{x})=A\boldsymbol{x}$, where $A$ is the parameter matrix. With multiple layers this is clearly a linear function of the output, but not of the parameters. As a special case of piecewise affine functions, our previous results suffice to show that these networks are multi-convex as functions of their parameters. This was proven for the special case of linear autoencoders by Baldi and Lu. Many authors have claimed that linear neural networks contain no "bad" local minima, i.e. every local minimum is a global minimum. This is especially evident in the study of linear autoencoders, which were shown to admit many points of inflection, but only a single strict minimum. While powerful, this claim does not apply to the networks seen in practice. To see this, consider the dataset $D=\{(0,1/2),(-1,\alpha),(1,2\alpha)\}$ consisting of three $(x,y)$ pairs, parameterized by $\alpha>1$. Note that the dataset has zero mean and unit variance in the $x$ variable, which is common practice in machine learning. However, we do not take zero mean in the $y$ variable, as the model we shall adopt is non-negative. Next, consider the simple neural network \[\begin{aligned} f(a,b) & =\sum_{(x,y)\in D}\left(y-\left[ax+b\right]_{+}\right)^{2}\label{eq:single_layer_objective}\\ & =\left(\frac{1}{2}-\left[b\right]_{+}\right)^{2}+\left(\alpha-\left[b-a\right]_{+}\right)^{2}+\left(2\alpha-\left[b+a\right]_{+}\right)^{2}.\nonumber \end{aligned}\] This is the squared error of a single ReLU neuron, parameterized by $(a,b)\in\mathbb{R}^{2}$. We have chosen this simplest of all networks because we can solve for the local minima in closed form, and show they are indeed very bad. First, note that $f$ is a continuous piecewise convex function of six pieces, realized by dividing the plane along the line $ax+b=0$ for each $x\in D$, as shown in figure [\[fig:bad_local_minimum\]](#fig:bad_local_minimum){reference-type="ref" reference="fig:bad_local_minimum"}. Now, for all but one of the pieces, the ReLU is "dead" for at least one of the three data points, i.e. $ax+b<0$. On these pieces, at least one of the three terms of equation [\[eq:single_layer_objective\]](#eq:single_layer_objective){reference-type="ref" reference="eq:single_layer_objective"} is constant. The remaining terms are minimized when $y=ax+b$, represented by the three dashed lines in figure [\[fig:bad_local_minimum\]](#fig:bad_local_minimum){reference-type="ref" reference="fig:bad_local_minimum"}. There are exactly three points where two of these lines intersect, and we can easily show that two of them are strict local minima. Specifically, the point $(a_{1},b_{1})=(1/2-\alpha,1/2)$ minimizes the first two terms of equation [\[eq:single_layer_objective\]](#eq:single_layer_objective){reference-type="ref" reference="eq:single_layer_objective"}, while $(a_{2},b_{2})=(2\alpha-1/2,1/2)$ minimizes the first and last term. In each case, the remaining term is constant over the piece containing the point of intersection. Thus these points are strict global minima on their respective pieces, and strict local minima on $\mathbb{R}^{2}$. Furthermore, we can compute $f(a_{1},b_{1})=4\alpha^{2}$ and $f(a_{2},b_{2})=\alpha^{2}$. This gives \[\begin{aligned} \lim_{\alpha\rightarrow\infty}a_{1} & =-\infty,\\ \lim_{\alpha\rightarrow\infty}a_{2} & =+\infty, \end{aligned}\] and \[\lim_{\alpha\rightarrow\infty}\left(f(a_{1},b_{1})-f(a_{2},b_{2})\right)=\infty.\] Now, it might be objected that we are not permitted to take $\alpha\rightarrow\infty$ if we require that the $y$ variable has unit variance. However, these same limits can be achieved with variance tending to unity by adding $\left\lfloor \alpha\right\rfloor$ instances of the point $(1,2\alpha)$ to our dataset. Thus even under fairly stringent requirements we can construct a dataset yielding arbitrarily bad local minima, both in the parameter space and the objective value. This provides some weak justification for the empirical observation that success in deep learning depends greatly on the data at hand. We have shown that the results concerning local minima in linear networks do not extend to the nonlinear case. Ultimately this should not be a surprise, as with linear networks the problem can be relaxed to linear regression on a convex objective. That is, the composition of all linear layers $g(\boldsymbol{x})=A_{1}A_{2}...A_{n}\boldsymbol{x}$ is equivalent to the function $f(\boldsymbol{x})=A\boldsymbol{x}$ for some matrix $A$, and under our previous assumptions the problem of finding the optimal $A$ is convex. Furthermore, it is easily shown that the number of parameters in the relaxed problem is polynomial in the number of original parameters. Since the relaxed problem fits the data at least as well as the original, it is not surprising that the original problem is computationally tractable. This simple example was merely meant to illustrate the difficulty of establishing results for every local minimum of every neural network. Since training a certain kind of network is known to be NP-Complete, it is difficult to give any guarantees about worst-case global behavior. We have made no claims, however, about probabilistic behavior on the average practical dataset, nor have we ruled out the effects of more specialized networks, such as very deep ones. # Conclusion We showed that a common class of neural networks is piecewise convex in each layer, with all other parameters fixed. We extended this to a theory of a piecewise multi-convex functions, showing that the training objective function can be represented by a finite number of multi-convex functions, each active on a multi-convex parameter set. From here we derived various results concerning the extrema and stationary points of piecewise multi-convex functions. We established convergence conditions for both gradient descent and iterated convex optimization on this class of functions, showing they converge to piecewise partial minima. Similar results are likely to hold for a variety of other optimization algorithms, especially those guaranteed to converge at stationary points or local minima. We have witnessed the utility of multi-convexity in proving convergence results for various optimization algorithms. However, this property may be of practical use as well. Better understanding of the training objective could lead to the development of faster or more reliable optimization methods, heuristic or otherwise. These results may provide some insight into the practical success of sub-optimal algorithms on neural networks. However, we have also seen that local optimality results do not extend to global optimality as they do for linear autoencoders. Clearly there is much left to discover about how, or even if we can optimize deep, nonlinear neural networks.	{'timestamp': '2016-12-30T02:01:59', 'yymm': '1607', 'arxiv_id': '1607.04917', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04917'}
# Introduction ## Background and Importance The minor planets orbiting beyond Neptune provide valuable insight on our solar system's formation and evolution, but they have only been studied since 1992 when the first Trans-Neptunian Object (TNO) after Pluto was discovered. Almost a quarter century later, $\sim2000$ TNOs and Centaurs are known (see Figure [\[real\]](#real){reference-type="ref" reference="real"}) and they are revealing their properties slowly because of the difficulties involved with detecting the faint, slow-moving members of this distant population. Multiple dedicated TNO surveys have been conducted over the years [e.g., @Larsen2001; @Gladman2001; @Bernstein2004; @Elliot2005; @Petit2006; @Petit2008; @Sheppard2011; @Gladman2012; @Alexandersen2014; @Brown2015] including stellar occultation surveys which focussed on the discovery of sub-km objects below the sensitivity limit of optical telescopes. Thanks to these studies, the large $100-1000$ km objects have a well characterized size-frequency distribution [SFD; @Petit2008; @Fuentes2008] while TNOs smaller than $100$ km have only more recently been studied. However, there is a need for more observational data to confirm the apparent transition from a steep to shallow SFD slope among the Neptune Trojans and SDOs around $D\sim100$ km (corresponding to absolute magnitude $H\sim8.5$). If the transition is present within all TNO sub-populations it would suggest the formation scenario in which "asteroids were born big" and imply that objects smaller than $100$ km are dominantly the result of collisional evolution. Most of the known TNOs were discovered in 'deep and narrow' observing campaigns using large telescopes with small fields of view. Current Near Earth Object (NEO) surveys [@Larson1998; @Kaiser2004] have the advantage of continuously monitoring large portions of the sky over several years, but are disadvantaged because they use smaller telescopes with cadences designed to identify NEOs that move more than $10\times$ faster than TNOs. searched archival data from the Catalina Sky Survey and Siding Spring Survey and independently identified the eight brightest known TNOs. Even though they did not discover any new objects they predicted a 32% chance that an object having magnitude $V<19.1$ remains undiscovered in the unsurveyed region of the sky. Evidence has been mounting in the past few years that there is a large planetary-sized distant object in our solar system whose gravitational perturbations influence the orbits of Scattered-Disc Objects (SDOs), particularly those on orbits similar to the dwarf planet (90377) Sedna. These works suggest that all currently known extreme TNOs with semi-major axis greater than $150$ AU (including the only other Sedna-like object: 2012 VP$_{113}$) show a pronounced clustering in their arguments of perihelia ($\omega$) not present in the closer TNO population. suggest this clustering is centred at $\omega\sim0^{\circ}$ and that it is due to the Lidov---Kozai effect, a three-body interaction capable of constraining $\omega$. They propose that a super-Earth mass body located at $\sim250$ AU would be capable of restricting $\omega$ for these objects and be stable for billions of years. However, made a similar calculation, but excluded orbits which do not demonstrate long term stability because of Neptune, and found that the distant TNOs cluster around $\omega\sim318^{\circ} \pm 8^{\circ}$ which is inconsistent with the Kozai mechanism. They suggest instead that the clustering can be maintained by a distant ten Earth-mass planet on an eccentric orbit with semi-major axis $700$ AU, nearly co-planar with the distant TNOs, but with $\omega$ shifted by $180^{\circ}$. In addition, such a planet might explain the presence of highly inclined TNOs whose existence has not yet been explained. also state that another plausible explanation for such a peculiar asymmetric $\omega$ configuration would be a strong stellar encounter with the Oort cloud in the past. Increasing the number of known retrograde TNOs and Sedna-like SDOs is needed to further test these hypotheses and to constrain the orbital elements and mass of any potentially undiscovered planet. Towards that end, in this work we report on the discovery and detection of the largest number of TNOs by a single asteroid survey, which due to its long-duration and wide-field coverage, provides an excellent complement to targeted deep-and-narrow surveys, resulting in a relatively unbiased TNO sample. ## Pan-STARRS The prototype telescope for the Panoramic Survey Telescope and Rapid Response System (Pan-STARRS1, hereafter referred to as PS1) located in the United States on Haleakala, Maui, Hawaii, has been surveying the sky since 2010. Many of the observations by PS1 were taken as a sequence of four exposures, each separated by a Transient Time Interval (TTI) of $\sim20$ minutes. This cadence was selected to optimise detection of Near Earth Objects (NEOs)---objects which have perihelia $q<1.3$ AU. Observations from each night are rapidly processed by the Image Processing Pipeline, and all detected moving objects identified by the Moving Object Processing System are reported to the Minor Planet Center. PS1 has become the leading discovery telescope for NEOs, discovering almost half of the new Near Earth Asteroids in 2015, and discovering more than half of the new comets in 2015. The detection of NEOs is done using subtraction of image pairs which have a TTI of $\sim20$ minutes and are well matched in image quality and telescope pointing. This TTI spacing produces a lower limit on the rate of motion for detection of moving objects, below which moving objects are self-subtracted in their image pairs. The lower limit is typically $\sim0.04^{\circ}$ per day ($=2\mathrm{"}$ in $20$ minutes), and is seeing dependent. A substantial number of Centaurs (which we define as having perihelia between Jupiter and Neptune) have been discovered from the pair-subtracted images, but only a few more-distant objects have been reported from PS1 (see Figure [\[real\]](#real){reference-type="ref" reference="real"}), some of which were discovered via the International Astronomical Search Collaboration (IASC[^1]), an educational outreach program in which images were blinked manually. Now that PS1 has thoroughly surveyed the sky north of $-30^{\circ}$ declination, other methods become viable for object detection that are potentially more sensitive to both fainter and slower moving objects. One method uses subtraction of a high-quality static sky image, derived from the cumulative survey data. The other method uses the historical survey to establish a catalogue of stationary objects, and compares catalogues of new detections in new images to the static sky, to reveal moving objects. Over the course of the PS1 survey, image quality has improved, but the grid structure in the PS1 CCDs requires many dithered images to produce a clean static sky image. And although good images in the gri passbands are now available for much of the sky north of $-30^{\circ}$ declination, the coverage in the more sensitive w passband is more sparse, because surveying in that band has been more focused on the ecliptic for the purpose of NEO discovery. The PS1 survey has also only recently been extended south to $-49^{\circ}$ declination. For these reasons, we have focussed our initial exploration of methods to extract fainter moving objects on the catalogue based approach. # Methodology To locate moving sources in the PS1 data, a new search method was developed and run on source catalogues previously generated by the IPP from PS1 images taken between 2010 Feb 24 and 2015 July 31. These catalogues are generated by a source extraction program which identifies and measures the point spread function for objects in the images. A detection is the information recorded about an object in a single exposure, and the catalogues contain detections of moving objects as well as stationary sources which must be removed. The method, as described in the following sections, links sets of detections (corresponding to the same object) from a single night into a 'tracklet'. Tracklets from multiple nights which correspond to the same object are then searched for. Figure [\[path\]](#path){reference-type="ref" reference="path"} shows the apparent path across the celestial sphere over four years for a typical TNO. Our method first searches for two related tracklets which are used to generate an initial orbit from which ephemerides are calculated to identify additional tracklets. All detections identified for an object have their image stamps extracted, which are visually inspected to ensure they are real and do not correspond to image artefacts or stationary sources that were not removed. While the PS1 MOPS pipeline was designed to identify TNOs, it requires a survey pattern and cadence that was not implemented. The method presented here is able to link TNOs using a survey pattern intended for NEO detection. ## Stationary Source Identification and Removal Objects in the outer solar system appear to move slowly across the celestial sphere, with their apparent paths dominated by the motion of Earth. TNOs in a 3:2 mean motion resonance with Neptune, such as Pluto, will have a maximum angular speed at opposition of $0.022^{\circ}$ per day, equal to $3.2\mathrm{"}$ per hour for orbits near zero inclination and eccentricity. An object much farther out at $550$ AU will move $0.25\mathrm{"}$ (the PS1 pixel scale) in one hour. However, even the most distant objects which move less than the astrometric uncertainty in one night will move a noticeable amount over multiple nights, allowing visual verification that they are indeed moving targets. To identify stationary sources, consider a specific telescope pointing (defined by the boresight direction and telescope rotation). All source detections for all exposures overlapping the field-of-view at this boresight direction are loaded. For each detection, the number of neighbouring detections within two magnitudes and $0.6\mathrm{"}$ are counted. Because each PS1 telescope pointing is typically observed four times in a given night (the average TTI is $19$ minutes, giving an average $57$ minute arc), we consider a detection stationary if it is present more than four times, i.e., if there are detections at the same location on the celestial sphere within $0.6\mathrm{"}$ over multiple nights. These criteria were determined empirically and relate to the astrometric and photometric uncertainty in the PS1 images. Once all stationary sources are identified, they are removed from the catalogues. The detections which remain are then used for tracklet creation. Note that while valid detections may be rejected, $100\%$ efficiency is not required as candidate objects are likely to have been observed multiple times at different positions along their orbits. ## Tracklet Creation Tracklets are formed from detections following a procedure similar to the stationary source removal. The detections from all exposures of a given pointing for a single night are loaded and iterated over. All detections within $0.4$ magnitudes and $16\mathrm{"}$ from each other are assigned a tracklet number. The choice of $16\mathrm{"}$ means many tracklets for inner Solar System objects will be present in the tracklet dataset, as this corresponds to $0.43^ {\circ}$ per day for exposures with a TTI of $15$ minutes. While these objects can be linked with our algorithm, we exclude any identified object having semi-major axis $< 4.8$ AU, as we are only interested in objects at heliocentric distances corresponding to Jupiter's orbit and beyond. While we could use a search distance $<16\mathrm{"}$, this choice does allow us to identify tracklets where the second or third detection occurred within a CCD cell gap. Once formed, each tracklet has its motion along a great circle fit with a constant angular speed model. Only tracklets with RMS residual $< 0.3\mathrm{"}$ and having $\ge3$ detections are kept, and their fitted angular speed and position angle are recorded. We do not create tracklets from detection pairs because we cannot apply an RMS test to judge their astrometric quality. Some fraction of the formed tracklets may be faint stars or image artefacts which exhibit linear motion, and their presence will increase the required computational time since there are more tracklet comparisons to be made. We validate all orbit fits using residual checks. This linking process is different from the kd-tree based algorithm used by MOPS described in detail by. The method presented here is computationally faster, but is limited to tracklets moving at much slower speed, and it cannot handle the intersection of tracklets which correspond to detections of different objects with similar apparent magnitude. ## Tracklet Pairing To test if two tracklets correspond to the same object, a brute-force approach is used. All tracklet pairs occurring within up to $6^{\circ}$ (equal to the maximum angular speed of $0.1^{\circ}$ per day with a $60$ day window) have a test orbit fit using FindOrb [^2], and we require its reported mean residual $<0.3\mathrm{"}$. The choice of a $60$ day window is a trade off between an improved initial arc length, and the computation time needed to test all tracklet pairs, and is reasonable since the PS1 survey pattern has observed much of the celestial sphere on at least three separate nights within this timespan (see Figure [\[look\]](#look){reference-type="ref" reference="look"}). However, this method is affected by missed tracklets which may have fallen into CCD cell gaps, been obscured by other image artefacts, or are too close to bright stars. We do not require every tracklet pair for an object to be identified, as many TNOs should be detected in more than two tracklets for the timespan considered here (see Figure [\[look\]](#look){reference-type="ref" reference="look"}). It is important to note however, that some orbital geometries will not have any observations: for example, there are no PS1 observations of the Centaur (10199) Chariklo which is located near the galactic centre. To increase computational efficiency, tracklet pairs are identified using a kd-tree indexed approach, and the position angle of both tracklets must be within $24^{\circ}$ of a line connecting the two tracklets. This value was empirically chosen and considerably reduces the computation time required as it excludes certain unphysical geometries. We consider only tracklets having angular speed between $0.1^{\circ}$ and $0.001^{\circ}$ per day, as our priority in this work is to find TNOs. An extended search for more distant objects will be considered in future work. ## Identifying additional tracklets The search for additional tracklets to extend an object's arc length requires a starting orbit, for which we use the tracklet pairing results described in the previous section. The metadata for each PS1 exposure is loaded, sorted by the observation time relative to the starting orbit's epoch (where its mean anomaly is defined), and iterated over. For each exposure, an ephemeris is generated, and if located within the field-of-view, the corresponding tracklet database is searched. For each tracklet which contains detections within $20\arcmin$ of the test ephemeris, a new orbit is fit using FindOrb. If the reported mean residual is $<0.2\mathrm{"}$, the tracklet is considered linked, its astrometry appended, and the updated orbit fit kept. The search then continues for additional tracklets. The choice of $0.2\mathrm{"}$ is more strict than the limit used during tracklet pairing, but has identified up to $27$ additional tracklets for the candidate objects. [\[morb\]]{#morb label="morb"} Table [1](#morb){reference-type="ref" reference="morb"} shows how the initial test orbit of a sample object converges as additional tracklets were found. The difference in the ephemeris for the initial and final orbit is illustrated in Figure [\[path\]](#path){reference-type="ref" reference="path"}. The choice of the $20\arcmin$ search distance was chosen based on identified TNOs: Figure [\[slip\]](#slip){reference-type="ref" reference="slip"} shows the angular distance between ephemerides from the initial tracklet-pair based orbit and the final fitted orbit for all identified objects. The majority of events show $<20\arcmin$ difference over $\pm 120$ days, where (based on Figure [\[look\]](#look){reference-type="ref" reference="look"}) most of the celestial sphere has been observed on at least three nights during the PS1 survey. While we could integrate the test orbits using Mercury6 to account for their orbital evolution, this is not required for TNOs as the generated ephemerides will change by much less than the $20\arcmin$ search distance. ## Visual Inspection and Classification The image stamps for each detection from all identified objects were extracted and visually inspected to confirm they were real and not affected by image artefacts. The orbital elements of the identified objects were used to classify them as potential Trojans, classical or resonant TNOs, Scattered Disc Objects (SDOs), or Centaurs. The orbital elements were also used to identify known objects by comparing the location and position angle of all ephemerides at their orbit epoch to those computed using the Minor Planet Center (MPC) orbit catalogue after integration using Mercury6. This method assumes the orbits from the MPC are more accurate than those identified in this study, which may not be true if our identified objects were independently previously observed by other telescopes over shorter orbital arcs. # Results and Discussion {#results} The source catalogues used in this study span the time from 2010 Feb 24 to 2015 July 31, and are generated from $529\,609$ exposures in $61\,065$ pointings, with $93\,799\,429\,652$ total detections. The stationary source removal left $7\,655\,731\,998$ detections of which $232\,447\,038$ were linked into $65\,524\,472$ tracklets. Figure [\[figA\]](#figA){reference-type="ref" reference="figA"} shows eccentricity vs semi-major axis for all identified objects, and we present a breakdown of these in Table [2](#count){reference-type="ref" reference="count"}, including the number of unknown and known objects, as well as the number expected (both total and to a limiting $V=22.5$) based on the MPC catalogue. We further discuss our identified objects in the following sections. [\[count\]]{#count label="count"} ## Bright Objects estimate that there is a 32% chance that a TNO having magnitude $V<19.1$ remains undiscovered after their archival search of the Catalina Sky Survey database. We computed ephemerides for the five year period of data used in this study, for all identified objects, and list the peak brightness of the top eight in Table [3](#bright){reference-type="ref" reference="bright"}. There is one object which reaches $V=18.5$, which is the third brightest TNO over the timespan of data used in this study. [\[bright\]]{#bright label="bright"} ## Large Objects Figure [\[figC\]](#figC){reference-type="ref" reference="figC"} plots semi-major axis vs absolute magnitude for all identified objects. The largest identified objects are all known, and are further listed in Table [4](#biggies){reference-type="ref" reference="biggies"}. However, (136472) Makemake and (134340) Pluto are not identified. For the first case, its tracklets do not occur within a $60$ day window, and for the latter, the only tracklet pair occurred with a two day arc, and additional tracklets were not linked. There are no unknown distant planetary-sized objects identified in our study. However, our minimum tracklet angular speed cut-off of $0.001^{\circ}$ per day could exclude them if they are present in the source catalogues. [\[biggies\]]{#biggies label="biggies"} ## Distant Objects Figure [\[figE\]](#figE){reference-type="ref" reference="figE"} plots semi-major axis vs distance from the Sun (on 2015 July 31). The farthest of these objects are listed in Table [5](#distant){reference-type="ref" reference="distant"}, with the six most distant being known (including the recent 2015 RR$_{245}$ [^3]). In addition to these, we have identified two previously unknown objects which have perihelia well beyond the orbit of Neptune. [\[distant\]]{#distant label="distant"} ## Orbital Clustering suggest the orbital configuration of TNOs decoupled from Neptune's influence support the presence of a planetary-sized perturber in the outer solar system. Due to the sensitivity limit of PS1 ($V\sim22.5$), we can only identify two of these objects, namely (90377) Sedna and 2007 TG$_{422}$, and both were identified in this study. We present the arguments of perihelia of our identified objects in Figure [\[figB\]](#figB){reference-type="ref" reference="figB"}. Using a Kolmogorov-Smirnov test on all identified objects having $q > 32$ AU and $q > 36$ AU, we do not see clustering in $\omega$. To our sensitivity limit, this is significant as our wide field survey might be expected to have less observational bias than targeted narrow searches. Survey bias might also be expected to induce apparent clustering in an observed population, rather than remove it. ## High Inclination Objects From Figure [\[figD\]](#figD){reference-type="ref" reference="figD"}, there are several identified objects with high inclination, and we list those having the largest inclination in Table [6](#retro){reference-type="ref" reference="retro"}. We identified in our study the recent 2011 KT$_{19}$ [^4], and also have one unknown object with $i=84^{\circ}$. The orbital evolution of highly-inclined TNOs is not understood, but suggest they could have been perturbed into their current orbits by a planetary-sized perturber in the outer solar system. Another plausible explanation is that they are captured objects. [\[retro\]]{#retro label="retro"} ## Neptune Trojans We list in Table [7](#trojs){reference-type="ref" reference="trojs"} two possible Neptune Trojans present in our identified objects, based solely on their orbital elements. One of these objects was identified in a parallel study using PS1 data, however we did not identify their other candidates. We present a discussion as to why this might be in Section [3.8](#whyyyy){reference-type="ref" reference="whyyyy"}. [\[trojs\]]{#trojs label="trojs"} ## Centaurs and Comets We define Centaurs to have perihelia between Jupiter and Neptune, but which are not in a mean motion resonance with Neptune. We identified $136$ Centaurs in our study, as well as $10$ comets which were all known objects. Our identification method does not link objects having $e\ge1$. Although call for a better characterised survey of Centaur objects, we consider them a by-product of the current study and leave this to a future publication---our primary interest here is identifying TNOs. ## Identification Efficiency {#whyyyy} Our discussion would not be complete without mentioning the efficiency of creating and pairing tracklets, as well as identifying additional tracklets. As suggested by Table [2](#count){reference-type="ref" reference="count"}, our method identifies about half of the expected (i.e., having $V<22.5$) known population of classical and resonant TNOs from the MPC catalogue, and two-thirds of the known SDOs. We do not count Jupiter Trojans or Centaur objects here because our detection parameters were empirically chosen to optimise detection of TNOs. A complete end-to-end measure of the identification efficiency could be made by injecting synthetic detections into our software routines, but ideally this must be done directly into the original PS1 image data which would then be used to produce new source catalogues. This reprocessing by the IPP is necessary for a realistic efficiency determination as the fill-factor of the PS1 camera is limited to $\sim70\%$, and our tracklet linking routines have no knowledge of which regions on the CCD are masked. Also, our limiting sensitivity is dependent on the photometric passband as well as local weather conditions. Because of these reasons, determining an accurate efficiency is a significant undertaking, which we will address in subsequent work to obtain an unbiased estimate of the true TNO population. From the current MPC catalogue for all known objects, the average distance between each asteroids within $30^{\circ}$ of the ecliptic and its closest neighbour having apparent magnitude within $0.4$ is $\sim820\mathrm{"}$. This is much greater than the $16\mathrm{"}$ search radius used during the tracklet creation process, suggesting that the majority of tracklets in the source catalogue (after removing stationary sources) will not be contaminated by detections associated with different objects. Our tracklets may be limited in terms of their RMS residual, but even for a distant object at $100$ AU, detections will be spaced $0.45\mathrm{"}$ for a $20$ minute TTI, which is twice the PS1 pixel scale. Because Figure [\[slip\]](#slip){reference-type="ref" reference="slip"} suggests additional tracklets should be readily found, this implies the efficiency of our search implementation is limited by either the stationary source removal process, or the tracklet pairing stage. We also note that while the PS1 camera fill-factor may lead to missed detections for NEOs, it can result in completely missed tracklets for TNOs due to their much slower speed across the celestial sphere. # Conclusions A search for distant objects was made using the archival PS1 data, with $1420$ objects identified, consisting of $255$ classical TNOs, $121$ resonant TNOs, $89$ SDOs, $154$ Centaurs, and $789$ Jupiter Trojans. Excluding the trojans, $371$ of these are new objects which we could not link to known objects. While our identified objects do not show a clustering in their arguments of perihelia, increasing the number of known retrograde TNOs and Sedna-like SDOs is important in better understanding the distant population in our solar system, especially to constrain the orbital elements and mass of any potentially undiscovered large planetary-sized objects. Future work will focus on validating the detection efficiency, as well as optimising our detection parameters to work well beyond the classical TNO regime.	{'timestamp': '2016-07-19T02:08:05', 'yymm': '1607', 'arxiv_id': '1607.04895', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04895'}
null	null
# Introduction Observations have established the presence of magnetic field of various magnitudes and on various spatial scales in our universe. Galaxies such as Milky Wave contain regular magnetic fields of the order of $\mu$G, while coherent fields of the order of $100~\mu$G are detected in distant galaxies. There is a strong evidence for the presence of magnetic field in intergalactic medium, including voids, with strengths exceeding $\sim 10^{-15}$ G. This supports the idea of cosmological origin of magnetic fields, which are subsequently amplified in galaxies, probably by the dynamo mechanism (see reviews ). The origin of cosmological magnetic field is a problem yet to be solved, with several possible mechanisms under discussion. These can broadly be classified into inflationary and post-inflationary scenarios. Both types still face problems to overcome: inflationary magnetic fields are constrained to be rather weak, while those produced after inflation typically have too small coherence lengths (see [@Widrow:2002ud; @Kandus:2010nw; @Durrer:2013pga; @Subramanian:2015lua] for a review of these mechanisms and assessment of these difficulties). It should also be noted that generation of helical hypermagnetic field prior to the electroweak phase transition may explain the observed baryon asymmetry of the universe. One of the mechanisms of generation of cosmological magnetic fields which is currently under scrutiny is based on the Abelian anomaly of the electroweak interactions. If the difference between the number densities of right-handed and left-handed charged fermions in the early hot universe happens to be non-zero (as in the leptogenesis scenario involving physics beyond the standard model; see for reviews), then a specific instability arises with respect to generation of helical (hypercharge) magnetic field. The generated helical magnetic field, in turn, is capable of supporting the fermion chiral asymmetry, thus prolonging its own existence to cosmological temperatures as low as tens of MeV. In this process, magnetic-field power is permanently transferred from small to large spatial scales (the phenomenon known as 'inverse cascade'). Further investigation of the general properties of the regime of inverse cascade revealed certain universal scaling laws in its late-time asymptotics. In this paper, we study analytically the process of generation of helical magnetic field in the early hot universe by an unspecified external source of lepton chiral asymmetry. Helical magnetic field is produced due to the presence of thermal background, which we extrapolate to all spatial scales, including the super-horizon scales.[^1] We consider a simple model of generation of magnetic field which assumes that the source of chiral anomaly maintains a constant value of the (conformal) chiral chemical potential of charged leptons. After generation of magnetic field of near maximal helicity, its evolution is traced in the absence of the external source of lepton chiral asymmetry. In this case, the helical magnetic field and the lepton chiral asymmetry are mutually sustained (decaying slowly) by quantum anomaly until temperatures of the order of tens of MeV, with magnetic-field power being permanently transferred from small to large spatial scales in the regime of inverse cascade. We obtain analytic expressions describing the evolution of the lepton chiral chemical potential and magnetic-field energy density. The evolution of both these quantities exhibits certain scaling behavior, effectively depending on a single combined variable. In this case, the late-time asymptotics of the chiral chemical potential reproduces the universal scaling law previously found in the literature for the system under investigation. As the temperature drops down because of the cosmological expansion, the processes of lepton scattering with the change of chirality (the so-called chirality-flipping processes) start playing important role, eventually leading to a rapid decay of the lepton chiral asymmetry. We give an analytic expression for the temperature at which this happens, depending on the initially generated values of the magnetic-field energy density and lepton chiral asymmetry. # Helical magnetic fields {#sec:helical} A spatially flat expanding universe filled by relativistic matter is conveniently described in the comoving conformal coordinate system $(\eta, \bx)$ with the conformal time $\eta$ and scale factor $a (\eta)$ entering the metric line element as $ds^2 = a^2 (\eta) \left( d \eta^2-d \bx^2 \right)$. By rescaling the conformal coordinates $(\eta, \bx)$, one can suitably normalize the scale factor $a (\eta)$. A divergence-free statistically homogeneous and isotropic cosmological magnetic field has the following general Fourier representation of the two-point correlation function[^2] : [\[correl\]]{#correl label="correl"} = ( 2 )\^3 (-' ) , where $\hat k_i = k_i / k$, $P_{ij} = \delta_{ij}-\hat{k}_i \hat{k}_j$ is the symmetric projector to the plane orthogonal to $\bk$, and $\epsilon_{ijk}$ is the normalized totally antisymmetric tensor. It is useful to introduce the helicity components $B_\pm (\bk)$ of the magnetic field via B_i () = B\_+ () \^+\_i () + B\_-() \^-\_i () , where the complex basis $\re^\pm_i = \frac{1}{\sqrt{2}}\left( \re^1_i \pm i \re^2_i \right)$ is formed from a right-handed (with respect to the orientation $\epsilon_{ijk}$) and orthonormal (with respect to the metric $\delta_{ij}$) basis ${\bf e}^{1} (\bk)$, ${\bf e}^{2} (\bk)$, ${\bf e}^{3} (\bk) = \bk /k$. The coefficients of the symmetric and antisymmetric parts of the correlation function are then expressed through these components as follows: [\[S\]]{#S label="S"} &=& 2 ( 2 )\^3 (-' ) S (k) ,\ &=& 2 ( 2 )\^3 (-' ) A (k) . [\[A\]]{#A label="A"} We note an obvious constraint $\|A (k)\| \leq S (k)$. The spectrum $A (k)$ of the magnetic-field correlation function characterizes the difference in the power between the left-handed and right-handed magnetic field, i.e., its helicity. The spectrum $S (k)$ characterizes the magnetic field energy density. In the case of so-called maximally helical magnetic field, one has $\|A(k)\| = S(k)$, and magnetic field is dominated by its left-handed or right-handed part, depending on the sign of $A (k)$. In this paper, we consider the effects of Abelian anomaly in the presence of spatially homogeneous chiral asymmetry.[^3] In this case, the evolution of the comoving magnetic field in conformal coordinates in cosmic plasma with high conductivity $\sigma$ takes the form [\[magev\]]{#magev label="magev"} = \^2 - , where $\Delta \mu \equiv a \left(\mu_\rL-\mu_\rR \right)$ is the spatially homogeneous difference between the (conformal) chemical potentials of the left-handed and right-handed charged leptons, $\sigma_c \equiv a \sigma \approx {\rm const}$ characterizes the plasma conductivity, and $\alpha \approx 1/137$ is the fine structure constant. The last term in equation ([\[magev\]](#magev){reference-type="ref" reference="magev"}) is connected with the anomalous current in Maxwell's equations [@Joyce:1997uy; @Frohlich:2000en; @Frohlich:2002fg; @Vilenkin:1980fu; @Redlich:1984md; @Fukushima:2008xe; @Tashiro:2012mf]. Using equations ([\[correl\]](#correl){reference-type="ref" reference="correl"}) and ([\[magev\]](#magev){reference-type="ref" reference="magev"}), one can obtain the following system of equations for the spectra $S (k, \eta)$ and $A (k, \eta)$ (see ): &=&- (S-S\_eq) + A , [\[dotS\]]{#dotS label="dotS"}\ &=&- A + S . [\[dotA\]]{#dotA label="dotA"} In equation ([\[dotS\]](#dotS){reference-type="ref" reference="dotS"}), we have added a term with the thermal equilibrium distribution [\[Seq\]]{#Seq label="Seq"} S\_eq (k, ) = , whose role is to ensure relaxation of the spectral energy distribution $S$ to its equilibrium value $S_{\rm eq}$ rather than to zero. This mechanism will not work in the long-wavelength domain $k \lesssim 1/\eta$, which is not causally connected in the expanding hot unverse. This, however, will be of no practical importance, since the anomalous term in equation ([\[dotS\]](#dotS){reference-type="ref" reference="dotS"}) will dominate in this spectral region. The initial spectra in the domain of small values of $k$ will also depend on their cosmological origin. We do not consider this issue in the present paper, assuming the initial spectrum to be given by ([\[Seq\]](#Seq){reference-type="ref" reference="Seq"}) on all scales. In an early radiation-dominated universe expanding adiabatically with the bulk matter in local thermal equilibrium, the entropy density $(a T)^3 g_$ remains constant. Here, $g_$ is the number of relativistic degrees of freedom $g_$ in thermal equilibrium. In the range of temperatures $80~{\rm GeV} < T < 150~{\rm MeV}$, the value of $g_$ changes insignificantly from about 86 to 72, and at the quantum-chromodynamical crossover, at $T \simeq 150$ MeV, drops to about 17. The quantity $g_^{1/3}$ thus drops from about 4.4 to 2.6, and we can see that the product $a T$ remains constant to a great extent. It is then convenient to normalize the scale factor as the inverse of the temperature, $a = 1/ T$. With this choice, we have $a = \eta / M_$, where $M_* = \left( 45 / 4 \pi^3 g_* \right)^{1/2} M_{\rm P} \simeq 10^{18}$ GeV is the effective Planck mass, and $\sigma_c = \sigma/ T \approx 70$ is constant in time [@Boyarsky:2011uy; @Turner:1987bw; @Baym:1997gq]. The equilibrium spectrum ([\[Seq\]](#Seq){reference-type="ref" reference="Seq"}) is independent of time and, with this normalization, reads [\[S0\]]{#S0 label="S0"} S\_eq (k, ) S_0 (k) = . The excess $\rho_B (\eta)$ of the energy density of magnetic field over the thermal energy density is then determined by [\[rhoB\]]{#rhoB label="rhoB"} \_B () = \_0\^P (k, ) k\^2 d k , where $P (k, \eta) \equiv S (k, \eta)-S_0 (k)$ is the excess over the thermal power spectrum. The system of equations ([\[dotS\]](#dotS){reference-type="ref" reference="dotS"}) and ([\[dotA\]](#dotA){reference-type="ref" reference="dotA"}) is supplemented by the evolution equation for the chiral chemical potential : [\[muevo\]]{#muevo label="muevo"} =- \_0\^ k d k-\_f () () + () . Here, $c_\Delta$ is a numerical constant of order unity (it would be equal to $3/4$ in pure quantum electrodynamics) that reflects the particle content of the primordial plasma, $\beta (\eta)$ is an external source for the evolution of chiral chemical potential which, for definiteness, we assume to be positive, and $\Gamma_{\rm f} (\eta)$ is the coefficient of the so-called 'flipping' term which describes decay of lepton chiral asymmetry caused by chirality flips in the electroweak interactions. # Development of helical magnetic fields Since the conformal conductivity $\sigma_c$ is constant during the period of evolution under investigation, it is more convenient to work in terms of a rescaled conformal time $\tau = \eta / \sigma_c$. With regard of ([\[S0\]](#S0){reference-type="ref" reference="S0"}), system ([\[dotS\]](#dotS){reference-type="ref" reference="dotS"}), ([\[dotA\]](#dotA){reference-type="ref" reference="dotA"}) can then be written in the form P &=&-2 k\^2 P + A , [\[dotP1\]]{#dotP1 label="dotP1"}\ A &=&-2 k\^2 A + ( P + S_0 ) . [\[dotA1\]]{#dotA1 label="dotA1"} Here and in what follows, the overdot denotes the derivative with respect to $\tau$. We then set the initial conditions for system ([\[muevo\]](#muevo){reference-type="ref" reference="muevo"}), ([\[dotP1\]](#dotP1){reference-type="ref" reference="dotP1"}), ([\[dotA1\]](#dotA1){reference-type="ref" reference="dotA1"}) at the formal cosmological singularity $\tau = 0$ in the form [\[inicon\]]{#inicon label="inicon"} P (k, 0) , A (k, 0) , (0) = 0 . Hence, we also have $\Delta \mu (0) = 0$. Thus, in the presence of lepton chiral asymmetry ($\Delta \mu \ne 0$), generation of the helicity spectrum $A (k)$ commences, according to ([\[dotA1\]](#dotA1){reference-type="ref" reference="dotA1"}), due to the presence of thermal distribution $S_0 (k)$. Solution of system ([\[dotP1\]](#dotP1){reference-type="ref" reference="dotP1"}) and ([\[dotA1\]](#dotA1){reference-type="ref" reference="dotA1"}) with respect to the spectral functions $P (k, \tau)$ and $A (k, \tau)$ with the initial conditions ([\[inicon\]](#inicon){reference-type="ref" reference="inicon"}) is given by P (k, ) &=& S_0 (k) e\^-2 k\^2 \_0\^e\^2 k\^2 ' ( 2 k ) (') d ' , [\[solP\]]{#solP label="solP"}\ A (k, ) &=& S_0 (k) e\^-2 k\^2 \_0\^e\^2 k\^2 ' ( 2 k ) (') d ' , [\[solA\]]{#solA label="solA"} where [\[AB\]]{#AB label="AB"} () = \^\_0 (') d ' . One can see that $A (k, \tau)$ has the same sign as $\Delta \mu (\tau)$, while $P (k, \tau)$ is always positive. To estimate the behavior of the spectral functions $P (k, \tau)$ and $A (k, \tau)$, let us evaluate them under the condition $\Delta \mu \equiv {\rm const} > 0$ \[which can be ensured by an appropriate behavior of the source $\beta (\tau)$ in ([\[muevo\]](#muevo){reference-type="ref" reference="muevo"})\]. Using ([\[AB\]](#AB){reference-type="ref" reference="AB"}) and taking the elementary integrals in ([\[solP\]](#solP){reference-type="ref" reference="solP"}) and ([\[solA\]](#solA){reference-type="ref" reference="solA"}), we have [\[spec-P\]]{#spec-P label="spec-P"} P (k, ) &=& (1-e\^-2 k\^2 ) ,\ A (k, ) &=& (-e\^-2 k\^2 ) , [\[spec-A\]]{#spec-A label="spec-A"} where [\[kmu\]]{#kmu label="kmu"} k\_= . Spectra ([\[spec-P\]](#spec-P){reference-type="ref" reference="spec-P"}) and ([\[spec-A\]](#spec-A){reference-type="ref" reference="spec-A"}) for $\Delta \mu = 5 \times 10^{-6}$ ($k_\mu \simeq 10^{-8}$) are plotted in figure [\[fig:spec\]](#fig:spec){reference-type="ref" reference="fig:spec"} for temperatures $T = 10$ GeV and 1 GeV, corresponding to $\tau \approx 1.2 \times 10^{15}$ and $1.2 \times 10^{16}$, respectively. As can be seen from expressions ([\[spec-P\]](#spec-P){reference-type="ref" reference="spec-P"}) and ([\[spec-A\]](#spec-A){reference-type="ref" reference="spec-A"}) and from figure [\[fig:spec\]](#fig:spec){reference-type="ref" reference="fig:spec"}, there arise two characteristic regions of wavenumbers: the region of relatively small $k$ (of order $k_\mu$), where the spectra keep growing and approach the property $A (k) \simeq P (k) \gtrsim S_0 (k)$ of maximal helicity, and the region of 'tails' of these spectra, where they quickly reach the threshold values $A (k) \simeq S_0 (k) k_\mu /k$ and $P (k) \simeq S_0 (k) \left(k_\mu /k\right)^2$. Indeed, in the region of large wavenumbers $k \gg k_\mu$, equation ([\[dotA1\]](#dotA1){reference-type="ref" reference="dotA1"}) is approximated as A k\^2 A + S_0 =-2 k\^2 A + 2 k_k S_0 , with the solution A (k) = S_0 (k) ( 1-e\^-2 k\^2 ) , that exponentially with time approaches the equilibrium $A_{\rm eq} = S_0 k_\mu / k$. Solution of ([\[dotP1\]](#dotP1){reference-type="ref" reference="dotP1"}) is then given by P (k) = S_0 (k) ( )\^2 , with a rapid exponential convergence to the equilibrium $P_{\rm eq} = S_0 \left(k_\mu /k\right)^2$. For sufficiently slow evolution of $\Delta \mu (\tau)$, these expressions for the spectral 'tails' will retain their forms, with $k_\mu (\tau)$ expressed through $\Delta \mu (\tau)$ by ([\[kmu\]](#kmu){reference-type="ref" reference="kmu"}). It should be noted that the magnetohydrodynamical description of cosmic plasma cannot be trusted in the domain of large physical wavenumbers $k / a \gtrsim \alpha T$, which, in our system of conformal units, corresponds to values of $k \gtrsim \alpha \sim 10^{-2}$. This limitation is insignificant for the evolution of magnetic instability developing on much larger spatial scales (as is usually the case in the scenarios under consideration). In the physically relevant domain $k \lesssim \alpha$, the thermal spectrum $S_0 (k) \approx 1$. Therefore, we can replace the factor $S_0 (k)$ by unity in all our equations; however, for the sake of better physical clarity, we will retain it. In the region of relatively small wavenumbers, the spectra $P (k, \tau)$ and $A (k, \tau)$ rapidly become equal to each other, as can be seen from figure [\[fig:spec\]](#fig:spec){reference-type="ref" reference="fig:spec"}. In the regime $P \gg S_0$, in which the quantity $S_0$ on the right-hand side of equation ([\[dotA1\]](#dotA1){reference-type="ref" reference="dotA1"}) can be neglected, one can multiply equation ([\[dotP1\]](#dotP1){reference-type="ref" reference="dotP1"}) by $A (k, \tau)$, equation ([\[dotA1\]](#dotA1){reference-type="ref" reference="dotA1"}) by $P (k, \tau)$ and subtract them to obtain an equation relating the spectral functions: ( P\^2 )\^+ 4 k\^2 P\^2 = ( A\^2 )\^+ 4 k\^2 A\^2 . This implies the relation [\[hellim\]]{#hellim label="hellim"} P\^2 (k, ) = A\^2 (k, ) + f_0\^2 (k) e\^-4 k\^2 , where $f_0^2(k)$ is an integration constant. We see that, if the quantity $A^2 (k, \tau)$ is not decaying or decaying slower than the last exponent in ([\[hellim\]](#hellim){reference-type="ref" reference="hellim"}), then a maximally helical state develops, with $P \simeq A$. In this regime, system ([\[dotP1\]](#dotP1){reference-type="ref" reference="dotP1"}), ([\[dotA1\]](#dotA1){reference-type="ref" reference="dotA1"}) reduces to a single equation for $A (k, \tau) \simeq P (k, \tau)$: [\[helA\]]{#helA label="helA"} A = (-2 k\^2 + ) A . # Evolution of chiral asymmetry and magnetic field Assuming that a maximally helical configuration quickly develops at some initial time $\tau_\ini$, we are going to establish how it will evolve together with $\Delta \mu (\tau)$ after the source $\beta (\tau)$ in ([\[muevo\]](#muevo){reference-type="ref" reference="muevo"}) is switched off. Let us make the notation $\Delta \mu_\ini = \Delta \mu (\tau_\ini)$, $P_\ini (k) = P (k, \tau_\ini)$, $A_\ini (k) = A (k, \tau_\ini)$, and introduce the momentum $k_\ini$ similarly to ([\[kmu\]](#kmu){reference-type="ref" reference="kmu"}): [\[kini\]]{#kini label="kini"} k\_= . The initial spectra can be presented as [\[APini\]]{#APini label="APini"} P\_(k) = P_0 Z ( ) , A\_(k) P\_(k) ( 1 + ) , where $Z (x)$ describes the shape of the spectrum, and the factor $(1 + k / k_\ini)$ is introduced to reflect the relation in the 'tails' of the spectra. The normalization constant $P_0$ is chosen so that [\[normal\]]{#normal label="normal"} \_0\^Z (x) x\^2 d x = 1 , and the initial excess ([\[rhoB\]](#rhoB){reference-type="ref" reference="rhoB"}) of the energy density of magnetic field over the thermal energy density is then equal to [\[helrhoB\]]{#helrhoB label="helrhoB"} \_B (\_) = \_0\^P (k, \_) k\^2 d k = . It is convenient to relate this quantity to the total radiation energy density by introducing the dimensionless parameter [\[rB\]]{#rB label="rB"} r_B\^ = = P_0 k\_\^3 . Asymptotically, as $\tau \to \infty$, the maximum of spectrum ([\[spec-P\]](#spec-P){reference-type="ref" reference="spec-P"}) is reached at $k = k_\mu / 2$, and one can derive an approximate asymptotic estimate for $r_B^\ini$: r_B\^ e\^k\_\^2 \_/ 2 , where $k_\ini$ is given by ([\[kini\]](#kini){reference-type="ref" reference="kini"}). One can see the exponential dependence of $r_B^\ini$ on the (rescaled) conformal time $\tau_\ini$ (or temperature $T_\ini = M_* / \sigma_c \tau_\ini$) at which the spectrum ([\[spec-P\]](#spec-P){reference-type="ref" reference="spec-P"}) is finally developed by the external source of chiral asymmetry. In what follows, we take $r_B^\ini$ and $\Delta \mu_\ini$ \[related to $k_\ini$ by ([\[kini\]](#kini){reference-type="ref" reference="kini"})\] to be our independent parameters. The subsequent evolution of the spectrum in the domain where $P (k, \tau) \approx A (k, \tau) \gg S_0 (k)$ is described by equation ([\[helA\]](#helA){reference-type="ref" reference="helA"}). Its solution with the initial condition $A (k, \tau_\ini) = A_\ini (k)$ is given by A (k, ) = g\^2 (k, ) A\_(k) , where [\[g\]]{#g label="g"} g (k, ) = e\^-k\^2 + k () , = -\_ , () = ()-(\_) , is the growth factor. Solution of ([\[muevo\]](#muevo){reference-type="ref" reference="muevo"}) with the zero source $\beta$ and with the flipping term neglected can be written as [\[differ\]]{#differ label="differ"} \_-() = \_0\^k d k . The contribution of the 'tail' in the distributions $A (k, \tau)$ and $A_\ini (k)$ to the value of the difference $\Delta \mu_\ini-\Delta \mu (\tau)$ in this expression is negligibly small. Indeed, the integral over the 'tail' region is estimated as \_k\_tail\^k d k && \_k\_tail\^S_0 (k) d k\ && , which is much smaller by absolute value than the left-hand side of ([\[differ\]](#differ){reference-type="ref" reference="differ"}) because of the smallness of $\alpha \approx 1/137$. Thus, we can ignore the presence of power-law tails in the spectra in ([\[differ\]](#differ){reference-type="ref" reference="differ"}), and write, using ([\[APini\]](#APini){reference-type="ref" reference="APini"}), \_-() = \_0\^Z (x) x d x . Dividing this by $\Delta \mu_\ini$ and using ([\[kini\]](#kini){reference-type="ref" reference="kini"}) and ([\[rB\]](#rB){reference-type="ref" reference="rB"}), we obtain the estimate [\[solmu\]]{#solmu label="solmu"} 1- = \_0\^Z (x) x d x . For values in a broad typical range of parameters in different cosmological scenarios, the factor in front of the integral in ([\[solmu\]](#solmu){reference-type="ref" reference="solmu"}) is much larger than unity. For instance, for $g_* = 75$ and $\Delta \mu_\ini = 5 \times 10^{-6}$, this factor is estimated to be $\sim 10^{13} r_B^\ini$, and will be very large for $r_B^\ini \gg 10^{-13}$. Since the left-hand side of ([\[solmu\]](#solmu){reference-type="ref" reference="solmu"}) is always bounded by unity, this implies that the integral on the right-hand side should be extremely small. The relation [\[fincon\]]{#fincon label="fincon"} \_0\^Z (x) x d x can then be regarded as an integral equation implicitly expressing the quantity to be found $\Delta \Psi$ through the known quantity $\Delta \Phi$ \[both enter the function $g^2 (k_\ini x, \tau)$ under this integral; see ([\[g\]](#g){reference-type="ref" reference="g"})\]. It is convenient to introduce the variables [\[phi\]]{#phi label="phi"} = 2 k\_\^2 , = 2 k\_ . In terms of these variables, we have [\[g2\]]{#g2 label="g2"} g\^2 (k_x, ) = e\^-x\^2 + x , and equation ([\[fincon\]](#fincon){reference-type="ref" reference="fincon"}) establishes the dependence $\psi (\phi)$, which is determined only by the form $Z (x)$ of the initial distribution. To find the dependence $\psi (\phi)$, we differentiate ([\[fincon\]](#fincon){reference-type="ref" reference="fincon"}) with respect to $\phi$. We obtain the Cauchy problem [\[dBA\]]{#dBA label="dBA"} = , (0) = 0 . The evolution of the chiral chemical potential can then be calculated by using ([\[AB\]](#AB){reference-type="ref" reference="AB"}): [\[muev\]]{#muev label="muev"} = = '() . Remarkably, the evolution of the chiral chemical potential $\Delta \mu$ depends on the initial conditions and on time through a single scaling parameter $\phi$, defined in ([\[phi\]](#phi){reference-type="ref" reference="phi"}). As an example, let us approximate the initial spectrum $P_\ini (k)$ with subtracted high-frequency 'tail' by a corresponding normalized spectral function with exponential cut-off:[^4] [\[zx\]]{#zx label="zx"} Z (x) = 2 x e\^-x\^2 . Then, introducing the variable [\[zeta\]]{#zeta label="zeta"} = , we can present problem ([\[dBA\]](#dBA){reference-type="ref" reference="dBA"}) in the form [\[dzeta\]]{#dzeta label="dzeta"} = , (0) = 0 , where [\[F\]]{#F label="F"} F () = . Equation ([\[dzeta\]](#dzeta){reference-type="ref" reference="dzeta"}) can, in principle, be integrated, and the function $\zeta (\phi)$ can be found. The evolution ([\[muev\]](#muev){reference-type="ref" reference="muev"}) of the chiral chemical potential is then given by [\[muevol\]]{#muevol label="muevol"} = '() = , where $F (\zeta)$ is given by ([\[F\]](#F){reference-type="ref" reference="F"}). Solution ([\[muevol\]](#muevol){reference-type="ref" reference="muevol"}) is plotted in logarithmic scale in figure [\[fig:muevol\]](#fig:muevol){reference-type="ref" reference="fig:muevol"}. Let us establish the late-time asymptotics of the solution to ([\[dzeta\]](#dzeta){reference-type="ref" reference="dzeta"}), ([\[F\]](#F){reference-type="ref" reference="F"}). In the regime $\zeta \gg 1$, we have $F (\zeta) \approx \zeta/2 + 3/\zeta + {\cal O} \left( \zeta^{-3} \right)$. Solution of ([\[dzeta\]](#dzeta){reference-type="ref" reference="dzeta"}) in this case behaves as [\[large-a\]]{#large-a label="large-a"} () \^1/2 , . This qualitative behavior does not depend on the specific shape ([\[zx\]](#zx){reference-type="ref" reference="zx"}) of the initial spectrum and is caused by the inverse cascade that transfers the spectral power to small-frequency region. Indeed, for large enough values of $\tau$, function ([\[g2\]](#g2){reference-type="ref" reference="g2"}) develops a strong Gaussian peak at small values of $x$, where $Z (x)$ behaves rather smoothly (typically, as a power of $x$). Expression ([\[dBA\]](#dBA){reference-type="ref" reference="dBA"}) does not then depend on the concrete form of $Z (x)$ in this limit. With the account of asymptotics ([\[large-a\]](#large-a){reference-type="ref" reference="large-a"}), the solution $\zeta (\phi)$ of the differential equation ([\[dzeta\]](#dzeta){reference-type="ref" reference="dzeta"}), ([\[F\]](#F){reference-type="ref" reference="F"}) can be approximated by the expression [\[zeta1\]]{#zeta1 label="zeta1"} () = . Numerical integration confirms this approximation within about $1\%$ precision (see figure [\[fig:interpol\]](#fig:interpol){reference-type="ref" reference="fig:interpol"}). Then [\[interpol\]]{#interpol label="interpol"} () = () = . Using ([\[muevol\]](#muevol){reference-type="ref" reference="muevol"}), we then obtain the universal late-time asymptotics [\[muas\]]{#muas label="muas"} - (1 + ) + (1 + ) . This describes very well the almost ideal power-law behavior observed in figure [\[fig:muevol\]](#fig:muevol){reference-type="ref" reference="fig:muevol"} for large values of $\phi$, with the asymptotic power index equal to $-1/2$. The late-time asymptotics $\Delta \mu \propto \eta^{-1/2}$ in the system under consideration has been previously established in (and in with a leading logarithmic correction, $\Delta \mu \propto \eta^{-1/2} \log^{1/2} \eta$), and also noted in in the context of chiral magnetohydrodynamics. Let us also determine the behavior of the magnetic-field energy density, described by the parameter $r_B (\tau)$. For our developed chiral distribution, we have $P (k, \tau) \approx g^2 (k, \tau) P_\ini (k)$. Hence, [\[NB\]]{#NB label="NB"} r_B () = N_B\^ r_B\^ , N_B () = \_0\^Z (x) e\^-x\^2 + () x x\^2 d x . Thus, the energy density of magnetic field also depends on time through a single scaling parameter $\phi$, defined in ([\[phi\]](#phi){reference-type="ref" reference="phi"}). Its behavior at large values of $\phi$ will depend only on the behavior of the function $Z (x)$ at small $x$. For the initial shape ([\[zx\]](#zx){reference-type="ref" reference="zx"}) of the magnetic-field spectrum, the scaling function $N_B (\phi)$ is determined by approximations ([\[zeta1\]](#zeta1){reference-type="ref" reference="zeta1"}) and ([\[interpol\]](#interpol){reference-type="ref" reference="interpol"}). Analytic estimate of integral ([\[NB\]](#NB){reference-type="ref" reference="NB"}) gives a rather complicated asymptotics at $\phi \gg 1$ : N_B () e\^- . However, in a very wide range of the values of argument, $10^2 \lesssim \phi \lesssim 10^{50}$, function ([\[NB\]](#NB){reference-type="ref" reference="NB"}) is excellently interpolated by a simple power law (see figure [\[fig:interpol-1\]](#fig:interpol-1){reference-type="ref" reference="fig:interpol-1"}) [\[interpol-1\]]{#interpol-1 label="interpol-1"} N_B () \^-5/9 . With $\phi$ being asymptotically given by \[see ([\[phi\]](#phi){reference-type="ref" reference="phi"}) and ([\[kini\]](#kini){reference-type="ref" reference="kini"})\] [\[phi-as\]]{#phi-as label="phi-as"} = , equations ([\[muas\]](#muas){reference-type="ref" reference="muas"}) and ([\[NB\]](#NB){reference-type="ref" reference="NB"}), ([\[interpol-1\]](#interpol-1){reference-type="ref" reference="interpol-1"}) give the asymptotic behavior of the quantities $\Delta \mu$ and $r_B$ as functions of temperature and of their initial values: = \^1/2 , = 9 ( )\^5/9 # Decay of chiral asymmetry caused by chirality flipping In the preceding analysis, we totally neglected the flipping term with coefficient $\Gamma_{\rm f} (\eta)$ in ([\[muevo\]](#muevo){reference-type="ref" reference="muevo"}), which is justified at high temperatures. However, as the temperature drops down because of cosmological expansion, at some point this term starts dominating over the other terms on the right-hand side of ([\[muevo\]](#muevo){reference-type="ref" reference="muevo"}), after which the chiral chemical potential decays exponentially as [\[expodec\]]{#expodec label="expodec"} (T) (- ) . In this section, we estimate the temperature $T_{\rm f}$ at which this decay commences. The contribution to the coefficient $\Gamma_{\rm f} (\tau)$ comes from weak and electromagnetic processes, so that we have $\Gamma_{\rm f} = \Gamma_w + \Gamma_e$, where the weak and electromagnetic contributions are estimated, respectively, as (see ) [\[gammas\]]{#gammas label="gammas"} \_w \~G\_F\^2 T\^4 ( )\^2 , \_e \~\^2 ( )\^2 . Here, $G_{\rm F}$ is the Fermi constant, $\alpha$ is the fine structure constant, and $m_e$ is the electron mass. The factors in the brackets in ([\[gammas\]](#gammas){reference-type="ref" reference="gammas"}) with electron mass $m_e$ describe suppression of chirality-flipping scattering rates with respect to 'chirality-preserving' ones. The weak contribution $\Gamma_w$ dominates at temperatures $T > T_{\rm eq} \simeq \sqrt{\alpha/ G_{\rm F}} \approx 25$ GeV, while, at $T < T_{\rm eq}$, chirality flipping is dominated by the electromagnetic processes and is characterized by $\Gamma_e$ in ([\[gammas\]](#gammas){reference-type="ref" reference="gammas"}). In view of equation ([\[dotA1\]](#dotA1){reference-type="ref" reference="dotA1"}), the first term on the right-hand side of ([\[muevo\]](#muevo){reference-type="ref" reference="muevo"}) is itself a sum of two terms with opposite signs: [\[two-term\]]{#two-term label="two-term"}- \_0\^A (k, ) k d k = \_0\^A (k, ) k\^3 d k- \_0\^k\^2 d k . In this expression, the thermal 'tail' in the spectral function $A (k, \tau)$ compensates the thermal contribution from $S_0 (k)$. Indeed, at the 'tail,' we have $A = A_{\rm eq} = S_0 \left(k_\mu/ k \right)$, and taking into account ([\[kmu\]](#kmu){reference-type="ref" reference="kmu"}), we observe cancellation of the corresponding integrals in ([\[two-term\]](#two-term){reference-type="ref" reference="two-term"}). Therefore, it is the negative term with the integral of the spectral function $P (k, \tau)$ in ([\[two-term\]](#two-term){reference-type="ref" reference="two-term"}) that is to be considered. To determine whether the neglect of the flipping term in ([\[muevo\]](#muevo){reference-type="ref" reference="muevo"}) is legitimate, we should compare the absolute value of this term with the absolute value $\Gamma_{\rm f} (\tau) \Delta \mu (\tau)$ of the flipping term. It is convenient to divide both quantities by $\Delta \mu (\tau)$. For the first expression, we have [\[regdec\]]{#regdec label="regdec"} \_0\^P (k, ) k\^2 d k = g\_\* r_B () \_0\^S_0 (k) k\^2 d k = g\_\* r_B () , Using ([\[NB\]](#NB){reference-type="ref" reference="NB"}) and ([\[interpol-1\]](#interpol-1){reference-type="ref" reference="interpol-1"}) to express $r_B (\tau)$ through $r_B^\ini$, we estimate ([\[regdec\]](#regdec){reference-type="ref" reference="regdec"}) as [\[regdec1\]]{#regdec1 label="regdec1"} g\_\* r_B () g\_\* r_B\^ . This expression should be compared to each of the quantities in ([\[gammas\]](#gammas){reference-type="ref" reference="gammas"}). Assuming that chirality flipping comes into play at temperatures $T < T_{\rm eq} \approx 25$ GeV (this will be confirmed by the final estimate), we only need to take into account the electromagnetic part $\Gamma_e$. We then have an equation for the estimate of the temperature of decay caused by chirality flipping: [\[finest\]]{#finest label="finest"} g\_\* r_B\^( )\^2 . The asymptotic value of $\phi \gg 1$ is given by ([\[phi-as\]](#phi-as){reference-type="ref" reference="phi-as"}). Substituting it into ([\[finest\]](#finest){reference-type="ref" reference="finest"}) and solving the resulting equation with respect to $T$, we obtain [\[Tf\]]{#Tf label="Tf"} T\_f = ( )\^1/23 ( )\^1/23 MeV , where we have put the numerical values for physical constants. For $g_* = 30$, $\Delta \mu_\ini = 3 \times 10^{-5}$, and $r_B^\ini = 5 \times 10^{-5}$, this equation gives $T_{\rm f} \simeq 150$ MeV (at this time, $\phi \approx 1200$). For $r_B^\ini = 5 \times 10^{-4}$, we obtain $T_{\rm f} \simeq 60$ MeV (with $\phi \approx 3000$). This is in good qualitative agreement with the numerical results of. Note that the resulting temperature ([\[Tf\]](#Tf){reference-type="ref" reference="Tf"}) does not depend on the temperature at which the initial values $\Delta \mu_\ini$ and $r_B^\ini$ are set (and which is assumed to be much higher than $T_{\rm f}$). This is due to the asymptotic scaling $\Delta \mu \propto \phi^{-1/2}$ \[see ([\[muas\]](#muas){reference-type="ref" reference="muas"})\], and $r_B \propto \phi^{-5/9}$ \[see ([\[NB\]](#NB){reference-type="ref" reference="NB"}) and ([\[interpol-1\]](#interpol-1){reference-type="ref" reference="interpol-1"})\], ensuring that the ratio $\Delta \mu^{10}/ r_B^9$ remains to be roughly constant in the regime of inverse cascade. # Summary We provided an analytic treatment of the process of generation of helical magnetic field in an early hot universe in the presence of externally induced lepton chiral asymmetry, and of the subsequent mutual evolution of the chiral asymmetry and magnetic field. Helical magnetic field is generated from the thermal initial spectrum (extrapolated to all scales including the super-horizon ones) owing to the effects of quantum chiral anomaly. The thermal bath also serves as a medium of relaxation of magnetic field to its thermal state. The generated helical magnetic field and the lepton chiral asymmetry are capable of supporting each other, thus prolonging their existence to cosmological temperatures as low as tens of MeV, with spectral power being permanently transformed from small to large spatial scales (the so-called 'inverse cascade'). Our main results are summarized as follows. We obtained analytic expressions describing the evolution of the lepton chiral chemical potential and magnetic-field energy density. For a developed maximally helical magnetic field, both the chiral chemical potential $\Delta \mu$ and the relative fraction of magnetic-field energy density $r_B$ depend on their initial values and on time through a single variable $\phi$ introduced in ([\[phi\]](#phi){reference-type="ref" reference="phi"}). This scaling property is encoded in equations ([\[dzeta\]](#dzeta){reference-type="ref" reference="dzeta"})--([\[muevol\]](#muevol){reference-type="ref" reference="muevol"}) and ([\[NB\]](#NB){reference-type="ref" reference="NB"})--([\[interpol-1\]](#interpol-1){reference-type="ref" reference="interpol-1"}), and depicted in figures [\[fig:muevol\]](#fig:muevol){reference-type="ref" reference="fig:muevol"} and [\[fig:interpol-1\]](#fig:interpol-1){reference-type="ref" reference="fig:interpol-1"}. The late-time asymptotics for $\Delta \mu$ reproduces the scaling law $\Delta \mu \propto \eta^{-1/2} \log^{1/2} \eta$ \[see equation ([\[muas\]](#muas){reference-type="ref" reference="muas"})\] previously found in this system in. By numerical interpolation, we find that the relative fraction $r_B$ of the magnetic-field energy density in this regime decays as $r_B \propto \eta^{-5/9}$ all through the relevant part of the cosmological history. Since the conformal time $\eta$ in our units is related to the temperature $T$ as $\eta = M_* / T$, this also describes the evolution of these quantities with temperature. As the temperature drops to sufficiently low values due to the cosmological expansion, the chirality-flipping lepton scattering processes take control over the evolution of chiral asymmetry, leading to its rapid decay ([\[expodec\]](#expodec){reference-type="ref" reference="expodec"}). We derived a simple expression ([\[Tf\]](#Tf){reference-type="ref" reference="Tf"}) for the temperature at which this happens, depending on the initially generated values of the energy density of magnetic field and of the lepton chiral asymmetry. The analytic expressions obtained in this paper are sufficiently general and may be used for primary evaluation of scenarios of cosmological magnetogenesis by lepton chiral asymmetry. We are grateful to Alexey Boyarsky and Oleg Ruchayskiy for valuable comments. M. S. and O. T. acknowledge support from the Scientific and Educational Center of the Bogolyubov Institute for Theoretical Physics. The work of O. T. was supported by the WFS National Scholarship Programme and by the Deutsche Forschungsgemeinschaft (DFG) in part through the Collaborative Research Center "The Low-Energy Frontier of the Standard Model" (SFB 1044), in part through the Graduate School "Symmetry Breaking in Fundamental Interactions" (DFG/GRK 1581), and in part through the Cluster of Excellence "Precision Physics, Fundamental Interactions and Structure of Matter" (PRISMA). The work of Y. S. was supported by the Swiss National Science Foundation grant SCOPE IZ7370-152581. [^1]: The spectral properties of magnetic fields on superhorizon spatial scales depend on a concrete model of generation of primordial magnetic fields (see [@Kandus:2010nw; @Durrer:2013pga; @Subramanian:2015lua] for recent reviews). [^2]: The quantities $B_i$ are the components of the so-called comoving magnetic field, which is related to the observable magnetic field strength $\bB_{\rm obs}$ by the equation $\bB = a^2 \bB_{\rm obs}$. The spatial vector indices are treated by using the Kronecker delta-symbol, and their position does not matter. [^3]: Effects of spatial inhomogeneity in chiral relativistic plasma were under investigation in. [^4]: Equation ([\[zx\]](#zx){reference-type="ref" reference="zx"}) gives a correct linear growth at small $x$, observed in figure [\[fig:spec\]](#fig:spec){reference-type="ref" reference="fig:spec"}.	{'timestamp': '2016-10-18T02:03:51', 'yymm': '1607', 'arxiv_id': '1607.04845', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04845'}
null	null
# Introduction {#intro} Suitable models for the theory of computation and approximation are certain (quasi-)ordered sets, whose elements represent states of computation, knowledge or information, while the order abstractly describes refinement, improvement or temporal sequence. Let us briefly record the relevant order-theoretical terminology. A quasi-ordered set or qoset is a pair $Q = (X,\leq)$ with a reflexive and transitive relation $\leq$ on $X$. The dual order is denoted by $\geq$, and the dual qoset $(X,\geq)$ by $\widetilde{Q}$.\ If $\,\leq\,$ is antisymmetric, we speak of a (partial) order and an ordered set or a poset.\ A lower set, downset or decreasing set is a subset $Y$ that coincides with its down-closure ${\downarrow\! Y}$, consisting of all $x \!\in\! X$ with $x \!\leq\! y$ for at least one $y \!\in\! Y$. The up-closure ${\uparrow\! Y}$ and upper sets (upsets, increasing sets) are defined dually. The upper sets form the upper Alexandroff topology $\alpha Q$, and the lower sets the lower Alexandroff topology $\alpha \widetilde{Q}$. A set $D\subseteq Q$ is (up-)directed, resp. filtered or down-directed, if every finite subset of $D$ has an upper, resp. lower bound in $D$; in particular, $D$ cannot be empty. An ideal of $D$ is a directed lower set, and a filter is a filtered upper set; for $x \!\in\! X$, the set ${\downarrow\!x} \!=\! {\downarrow\!\{ x\}}$ is the principal ideal, and ${\uparrow\!x} \!=\! {\uparrow\!\{ x\}}$ is the principal filter generated by $x$. A poset $P$ is called up-complete, directed complete, a dcpo, or a cpo if each directed subset $D$ or, equivalently, each ideal has a join, that is, a least upper bound (supremum), denoted by $\bigvee\! D$. The (ideal) up-completion of a qoset $Q$ is ${\cal I} Q$, the set of all ideals, ordered by inclusion. The Scott topology $\sigma P$ of a poset $P$ consists of all upper sets $U$ that meet any directed set having a join in $U$. Both in the mathematical and in the computer-theoretically oriented literature (see e.g. ,, , ), the word 'domain' represents quite diverse structures, and in order-theoretical contexts, its meaning ranges from rather general notions like dcpos to quite specific kinds of posets like $\omega$-algebraic dcpos, sometimes with additional properties. Here, we adopt the convention to call arbitrary up-complete posets domains and to speak of a continuous poset if for each element $x$ there is a least ideal having a join above $x$. Notice that our continuous domains are the continuous posets in and, but they are the domains in, whereas our continuous posets and those in need not be up-complete. Although continuous domains usually are defined in order-theoretical terms, there exist also topological descriptions of them, for example, as sober locally supercompact spaces (Erné, Hoffmann, Lawson ). It is one of our main purposes in the subsequent investigations to drop the completeness or sobriety hypotheses without loosing relevant results applicable to domain theory. The term space always means topological space, but extensions to arbitrary closure spaces are possible (see and ). Several classes of spaces may be characterized by certain infinite distribution laws for their lattices of open sets. Recall that a frame or locale is a complete lattice $L$ satisfying the identity \[\textstyle{ {\rm (d)} \ x\wedge \bigvee Y = \bigvee \{ x\wedge y: y\in Y \} \vspace{-.5ex}}\] for all $x\in L$ and $Y \subseteq L$; the dual of (d) characterizes coframes. The identity \[\textstyle{{\rm (D)} \ \bigwedge \,\{ \bigvee Y: Y \!\in {\cal Y}\} = \bigvee\bigcap {\cal Y}} \vspace{-.5ex}\] for all collections ${\cal Y}$ of lower sets, defining complete distributivity, is much stronger. However, frames may also be defined by the identity (D) for all finite collections ${\cal Y}$ of lower sets. An up-complete meet-semilattice satisfying (d) for all directed sets (or ideals) $Y$ is called meet-continuous. Similarly, the continuous lattices in the sense of Scott, are the complete lattices enjoying the identity (D) for ideals instead of lower sets. Therefore, completely distributive lattices are also called supercontinuous; alternative descriptions of complete distributivity by equations involving choice functions are equivalent to the Axiom of Choice (see ). A complete lattice satisfying (D) for all collections of finitely generated lower sets is called ${\cal F}$-distributive or a wide coframe, and its dual a quasitopology or a wide frame (cf. ). A lattice is spatial iff it is isomorphic to a topology. All spatial lattices are wide frames. For any space $(X,{\cal S})$, the frame of open sets is ${\cal S}$, and the coframe of closed sets is denoted by ${\cal S}^c$. The closure of a subset $Y$ is denoted by $cl_{{\cal S}}Y$ or $Y^-\!$, and the interior by $int_{{\cal S}}Y\!$ or $Y^{\circ}$. The specialization order is given by \[x\leq y \ \Leftrightarrow \ x\leq_{{\cal S}} y \ \Leftrightarrow \ x\in \{ y\}^-\ \Leftrightarrow \ \forall\,U\! \in {\cal S}\ (x\in U \,\Rightarrow \, y\in U).\] It is antisymmetric iff $(X,{\cal S})$ is T$_0$, but we speak of a specialization order also in the non--T$_0$ setting. The saturation of a subset $Y$ is the intersection of all its neighborhoods, and this is the up-closure of $Y$ relative to the specialization order. In the specialization qoset $\Sigma^{-\!} (X,{\cal S}) = (X,\leq_{{\cal S}})$, the principal ideals are the point closures, and the principal filters are the cores, where the core of a point $x$ is the saturation of the singleton $\{ x\}$; the lower sets are the unions of cores, or of arbitrary closed sets, and the upper sets are the saturated sets. A topology ${\cal S}$ on $X$ is compatible with a quasi-order $\leq$ if $Q = (X,\leq)$ is the specialization qoset of $(X,{\cal S})$ or, equivalently, $\upsilon Q \subseteq {\cal S} \subseteq \alpha Q$, where $\upsilon Q$ is the weak upper topology, generated by the complements of principal ideals; the weak lower topology of $Q$ is the weak upper topology $\upsilon \widetilde{Q}$ of the order-dual $\widetilde{Q}$. Of course, in other contexts, compatibility of a topology with an order relation may have a different meaning (cf. ). In, we have introduced three classes of spaces that might be useful for the mathematical foundation of communication and information theory (order-theoretical notions refer to the specialization order): - web spaces have neighborhood bases of webs at each point $x$, i.e. unions of filtered sets each of which contains $x$, - wide web spaces have neighborhood bases of filtered sets at each point, - worldwide web spaces or core spaces have neighborhood bases of principal filters (cores) at each point. As shown in, each of these three classes of spaces may be described by an infinite distribution law for their topologies: a space is a - web space iff its topology is a coframe, - wide web space iff its topology is a wide coframe, - worldwide web space iff its topology is completely distributive. In Section [2](#convex){reference-type="ref" reference="convex"}, we briefly review the construction of patch spaces and some applications to web spaces, as developed in. The patch spaces of a given space are obtained by joining its topology with a cotopology (in: complementary topology), that is, a topology having the dual specialization order. Useful for patch constructions are so-called coselections $\zeta$, which choose for any topology ${\cal S}$ a subbase $\zeta {\cal S}$ of a cotopology $\tau_{\zeta}{\cal S}$. The topology ${\cal S}^{\zeta}$ generated by ${\cal S} \cup \zeta {\cal S}$ is then a patch topology, and the corresponding (quasi-ordered!) $\zeta$-patch space is ${\rm P}_{\zeta}(X,{\cal S}) = (X, \leq_{{\cal S}} ,{\cal S}^{\zeta})$. As demonstrated in, web spaces may be characterized by the property that their open sets are exactly the up-closures of the open sets in any patch space. For us, compactness does not include the Hausdorff separation axiom T$_2$. Locally compact spaces undoubtedly form one of the most important classes of topological spaces. In the non-Hausdorff setting, one has to require whole bases of compact neighborhoods at each point, because one compact neighborhood for each point would not be enough for an efficient theory. In certain concrete cases, one observes that each point of the space under consideration has even a neighborhood base consisting of supercompact sets, i.e. sets each open cover of which has already one member that contains them. Such spaces occur, sometimes unexpectedly, in diverse fields of mathematics -- not only topological but also algebraic ones -- and in theoretical computer science. Section [3](#CDS){reference-type="ref" reference="CDS"} is devoted to a closer look at such locally supercompact spaces; they are nothing but the core spaces, because the supercompact saturated sets are just the cores. These spaces have been introduced in, where the name core spaces referred to the larger class of closure spaces, and discussed further in and; core spaces are also called C-spaces (in: c-spaces), but that term has a different meaning in other contexts (e.g. in and ). The core spaces are exactly the locally compact wide web spaces, but also the locally hypercompact web spaces, where a set is hypercompact if its saturation is finitely generated; while the interior operator of a web space preserves finite unions of saturated sets, the interior operator of a core space preserves arbitrary unions of saturated sets. Moreover, the category of core spaces has a strong order-theoretical feature, being concretely isomorphic to a category of generalized quasi-ordered sets (see and Section [3](#CDS){reference-type="ref" reference="CDS"} for precise definitions and results). Core spaces share useful properties with the more restricted basic spaces or B-spaces (having a least base, which then necessarily consists of all open cores) and with the still more limited Alexandroff-discrete spaces or A-spaces (in which all cores are open), ,; but, in contrast to A-and B-spaces, core spaces are general enough to cover important examples of classical analysis. For instance, the Euclidean topology on ${\mathbb R}^n$ (which, ordered componentwise, is a continuous poset but not a domain!) is the weak(est) patch topology of the Scott topology, which makes ${\mathbb R}^n$ a core space. In Section [4](#sectorspaces){reference-type="ref" reference="sectorspaces"}, we characterize the patch spaces of core spaces as sector spaces. These are ${\uparrow}$-stable semi-qospaces (meaning that up-closures of open sets are open, and principal ideals and principal filters are closed) with neighborhood bases of so-called sectors, a special kind of webs having least elements. The restriction of the patch functor ${\rm P}_{\zeta}$ to the category of core spaces yields a concrete isomorphism to the category of $\zeta$-sector spaces, which fulfil strong convexity and separation axioms. In particular, the weak patch functor ${\rm P}_{\upsilon}$ induces a concrete categorical isomorphism between core spaces and fan spaces, i.e. ${\uparrow}$-stable semi-qospaces in which each point has a neighborhood base of fans ${\uparrow\!u}\setminus{\uparrow\!F}$ with finite sets $F$. In Section [5](#fanspaces){reference-type="ref" reference="fanspaces"}, such fan spaces are investigated and characterized by diverse order-topological properties. Our considerations have useful consequences for topological aspects of domain theory, as the continuous domains, equipped with the Scott topology, are nothing but the sober core spaces, and these correspond to fan spaces that carry the Lawson topology, the weak patch topology of the Scott topology. We find alternative descriptions of such ordered spaces, including convexity properties, separation axioms and conditions on the interior operator. This enables us to generalize the characterization of continuous lattices as meet-continuous lattices whose Lawson topology is Hausdorff and the Fundamental Theorem of Compact Semilattices to non-complete situations. Crucial is the fact that a semilattice with a compatible topology is semitopological iff it is a web space, and (locally compact) topological with small semilattices iff it is a (world) wide webspace. The category of T$_0$ core spaces and that of ordered fan spaces are not only equivalent to the category C-ordered sets, but also to the category of based domains, i.e. pairs consisting of a continuous domain and a basis of it (in the sense of ). In the last section, we study weight and density of the spaces under consideration, using the order-theoretical description of core spaces. For example, the weight of a core space is equal to the density of any of its patch spaces, but also to the weight of the lattice of closed sets. This leads to the conclusion that the weight of a completely distributive lattice is always equal to the weight of the dual lattice. If not otherwise stated, all results are derived in a choice-free set-theoretical environment; i.e., we work in ZF or NBG (Zermelo--Fraenkel or Neumann--Bernays--Gödel set theory) but not in ZFC (i.e. ZF plus Axiom of Choice). For basic categorical concepts, in particular, concrete categories, functors and isomorphisms, see Adámek, Herrlich and Strecker. For relevant order-theoretical and topological definitions and facts, refer to the monograph Continuous Lattices and Domains by G. Gierz, K. H. Hofmann, K. Keimel, J. D. Lawson, M. Mislove, and D. S. Scott ̇ # Patch spaces and web spaces {#convex} A (quasi-)ordered space is a (quasi-)ordered set equipped with a topology. In this elementary definition, no separation properties and no relationship between order and topology are required. However, some classical separation axioms extend to the ordered case as follows. A quasi-ordered space is a lower semi-qospace if all principal ideals are closed, an upper qospace if all principal filters are closed, and a semi-qospace if both conditions hold (these conditions mean that the quasiorder is lower semiclosed, upper semiclosed or semiclosed, respectively, in the sense of ). An ordered semi-qospace is a semi-pospace or T$_1$-ordered . A space equipped with a closed quasi-order $\leq$ (regarded as a subset of the square of the space) is called a qospace, and a pospace in case $\leq$ is a (partial) order. Alternatively, qospaces may be characterized by the condition that for $x\not\leq y$, there are open $U$ and $V$ with $x \!\in\! U$, $y \!\in\! V$, and ${\uparrow\!U} \mathop{\cap} {\downarrow\! V} = \emptyset$. Similarly, we define T$_2$-ordered spaces to be ordered spaces in which for $x \not\leq y$ there are an open upper set containing $x$ and a disjoint open lower set containing $y$; some authors call such spaces strongly $T_2$-ordered and mean by a T$_2$-ordered space a pospace (cf. Künzi , McCartan ). A quasi-ordered space is said to be upper regular if for each open upper set $O$ containing a point $x$, there is an open upper set $U$ and an closed upper set $B$ such that $x\in U \subseteq B \subseteq O$, or equivalently, for each closed lower set $A$ and each $x$ not in $A$, there is an open upper set $U$ and a disjoint open lower set $V$ with $x\in U$ and $A\subseteq V$. An upper regular T$_1$-ordered space is said to be upper T$_3$-ordered. Lower regular spaces are defined dually. Note the following irreversible implications: compact qospace $\Rightarrow$ upper regular semi-qospace $\Rightarrow$ qospace $\Rightarrow$ semi-qospace,\ For any quasi-ordered space $T = (Q,{\cal T}) = (X,\leq,{\cal T})$, ${\cal T}^{\,\leq} = {\cal T} \cap \alpha Q\ $ is the topology of all open upper sets (also denoted by ${\cal T}^{\,\sharp}$),\ ${\cal T}^{\,\geq} = {\cal T} \cap \alpha \widetilde{Q}\ $ is the topology of all open lower sets (also denoted by ${\cal T}^{\,\flat}$). We call ${\rm U} T = (X,{\cal T}^{\leq})$ the upper space and ${\rm L} T = (X, {\cal T}^{\geq})$ the lower space of $T$. A basic observation is that for lower semi-qospaces, the specialization order of ${\cal T}^{\leq}$ is $\leq$, while for upper semi-qospaces, the specialization order of ${\cal T}^{\geq}$ is $\geq$. Recall that a subset $Y$ of a qoset $Q$ is (order) convex iff it is the intersection of an upper and a lower set. A quasi-ordered space is locally convex if the convex open subsets form a base, strongly convex if its topology ${\cal T}$ is generated by ${\cal T}^{\leq} \cup {\cal T}^{\geq}$, and $\zeta$-convex if ${\cal T}$ is generated by ${\cal T}^{\leq} \mathop{\cup} \zeta ({\cal T}^{\leq})$, where $\zeta$ is a coselection (see the introduction). Specifically, $\upsilon$-convex quasi-ordered spaces are called hyperconvex. Thus, hyperconvexity means that the sets $U\setminus {\uparrow\!F}$ with $U\in {\cal T}^{\leq}$ and $F$ finite form a base. Observe that $\zeta$-convexity implies strong convexity, which in turn implies local convexity, but not conversely (counterexamples are given in ). Let $\zeta$ be any coselection. A space $(X,{\cal S})$ is said to be $\zeta$-determined if $\,{\cal S}^{\,\zeta\leq} = {\cal S}$. A map between spaces is called $\zeta$-proper if it is continuous and preimages of closed sets relative to the $\zeta$-cotopology are $\zeta$-patch closed (whence such a map is $\zeta$-patch continuous); and a map between quasi-ordered spaces is lower semicontinuous if preimages of closed lower sets are closed. In, many examples and counterexamples concerning these notions are discussed, and the following facts are established: Any upset selection $\zeta$, assigning to each qoset $Q$ a collection $\zeta Q$ of upper sets such that ${\uparrow\!x} = \bigcap\,\{ V\in \zeta Q: x\in V\}$ for all $x$ in $Q$, gives rise to a coselection by putting $\zeta {\cal S} = \zeta \widetilde {Q}$ for any space $(X,{\cal S} )$ with specialization qoset $Q$. If each $\zeta Q$ is a topology, we call $\zeta$ a topological (upset) selection; the largest one is $\alpha$, while the smallest one is $\upsilon$. By Lemma [\[pat\]](#pat){reference-type="ref" reference="pat"}, the weak patch functor ${\rm P}_{\upsilon}$ induces an isomorphism between the category of $\upsilon$-determined spaces and that of hyperconvex semi-qospaces. An important intermediate topological selection is $\sigma$, where $\sigma Q$ is the Scott topology, consisting of all upper sets $U$ that meet every directed subset having a least upper bound that belongs to $U$ (in arbitrary qosets, $y$ is a least upper bound of $D$ iff $D\subseteq {\downarrow\! z} \Leftrightarrow y\leq z$). The weak patch topology of $\sigma Q$ is the Lawson topology $\lambda Q = \sigma Q ^{\,\upsilon}$. We denote by $\Sigma Q$ the Scott space $(X,\sigma Q )$ and by $\Lambda Q$ the (quasi-ordered) Lawson space $(Q,\lambda Q )$, whose upper space in turn is $\Sigma Q$. Thus, all Scott spaces are $\upsilon$-determined, and all Lawson spaces are hyperconvex semi-qospaces. A quasi-ordered space $(Q,{\cal T})$ is said to be upwards stable or ${\uparrow}$-stable if it satisfies the following equivalent conditions: - $O\in {\cal T}$ implies ${\uparrow\!O}\in {\cal T}$. - ${\cal T}^{\leq} = \{ {\uparrow\! O}: O\in {\cal T}\}$. - The interior of each upper set is an upper set: $int_{{\cal T}} Y\! = int_{{\cal T}^{\leq}}Y$ if $Y \!= {\uparrow\!Y}$. - The closure of each lower set is a lower set: $cl_{{\cal T}} Y\! = cl_{{\cal T}^{\leq}}Y$ if $Y \!= {\downarrow\!Y}$. A web around a point $x$ in a qoset is a subset containing $x$ and with each point $y$ a common lower bound of $x$ and $y$; if $\leq$ is a specialization order, that condition means that the closures of $x$ and $y$ have a common point in the web. By a web-(quasi-)ordered space we mean an ${\uparrow}$-stable (quasi-)ordered space in which every point has a neighborhood base of webs around it. In the case of a space equipped with its specialization order, this is simply the definition of a web space. Many characteristic properties of web spaces and of web-quasi-ordered spaces are given in and. Note that the meet-continuous dcpos in the sense of are just those domains whose Scott space is a web space. The following result from underscores the relevance of web spaces for patch constructions: We now are going to derive an analogous result for wide web spaces; the case of worldwide web spaces (core spaces) is deferred to the next section. The notion of wide web spaces is a bit subtle: while in a web space every point has a neighborhood base consisting of open webs around it, in a wide web space it need not be the case that any point has a neighborhood base consisting of open filtered sets; the spaces with the latter property are those which have a 'dual' (Hoffmann ), that is, whose topology is dually isomorphic to another topology; see for an investigation of such spaces and a separating counterexample. Note the implications completely distributive $\,\Rightarrow\,\ $ dually spatial $\,\ \ \Rightarrow\ $ wide coframe $\hspace{1.8ex}\Rightarrow\,$ coframe and the corresponding (irreversible) implications for spaces: worldwide web space $\ \ \ \Rightarrow\ $ space with dual $\,\Rightarrow\,$ wide web space $\,\Rightarrow\,$ web space. It is now obvious to introduce an ordered version of wide web spaces by calling a quasi-ordered space locally filtered if each point has a base of filtered neighborhoods. Let $S \!=\! (X,{\cal S} )$ be a wide web space and $T \!=\! (X,\leq, {\cal T} )$ a patch space of $S$. By Proposition [\[webpatch\]](#webpatch){reference-type="ref" reference="webpatch"}, $T$ is ${\uparrow}$-stable, and $S$ is the upper space of $T$, i.e., ${\cal S} = {\cal T}^{\leq}$. Given $x \!\in\! O \!\in\! {\cal T}$, find $U \!\in\! {\cal S}$ and $V \!\in\! {\cal T}^{\geq}$ with $x \in U \cap V \subseteq O$, and a filtered set $D$ with $x \in W = int_{{\cal S}} D \subseteq D \subseteq U$. Then, $D\cap V$ is filtered (since $V = {\downarrow\!V}$) with $x\in W\cap V \subseteq D\cap V \subseteq U \cap V \subseteq O$, and $W\cap V \in {\cal T}$. Thus, $T$ is locally filtered. Conversely, let $T = (X, \leq, {\cal T} )$ be a strongly convex, locally filtered and ${\uparrow}$-stable semi-qospace. Then $T$ is web-quasi-ordered and, by Proposition [\[webpatch\]](#webpatch){reference-type="ref" reference="webpatch"}, a patch space of the web space $(X,{\cal T}^{\leq})$. For $x \in O \in {\cal T}^{\leq}$, there is a filtered $D\subseteq O$ with $x \in int_{{\cal T}}D$. Then ${\uparrow\!D}$ is filtered, $x \in W = {\uparrow\! int_{{\cal T}} D} \subseteq {\uparrow\!D} \!\subseteq\! O$, and by ${\uparrow}$-stability, $W \in {\cal T}^{\leq}$; thus, $(X,{\cal T}^{\leq})$ is a wide web space. The rest follows from Proposition [\[pat\]](#pat){reference-type="ref" reference="pat"}. # Core spaces and C-quasi-ordered sets {#CDS} Strengthening the notion of compactness, we call a subset $C$ of a space supercompact if every open cover of $C$ has a member that contains $C$; and local supercompactness means the existence of supercompact neighborhood bases at each point. As mentioned in the introduction, an equivalent condition is that each point has a neighborhood base of cores (possibly of different points!) -- in other words, that we have a core space. Under the assumption of the Ultrafilter Theorem (a consequence of the Axiom of Choice), many properties of locally supercompact spaces are shared by the more general locally hypercompact spaces, where a subset is called hypercompact if its saturation is generated by a finite subset. In view of the next proposition, proven choice-freely in and, core spaces may be viewed as an infinitary analogue of web spaces (whence the name 'worldwide web spaces' ). Core spaces have pleasant properties. For example, on account of (6) resp. (7), the interior resp. closure operator of a core space induces a complete homomorphism from the completely distributive lattice of upper resp. lower sets onto the lattice of open resp. closed sets. In, it is shown that a nonempty product of spaces is a core space iff all factors are core spaces and all but a finite number of them are supercompact. Similarly, a nonempty product of spaces is a (wide) web space iff all factors are (wide) web spaces and all but a finite number are filtered. Assuming the Principle of Dependent Choices (another consequence of the Axiom of Choice), one can show that all core spaces have a dual (a base of filtered open sets; see ). Computationally convenient is the fact that core spaces are in bijective correspondence to so-called idempotent ideal relations or C-quasi-orders. These are not really quasi-orders but idempotent relations $R$ (satisfying $x\,R\,z \Leftrightarrow \exists\, y \,(x\,R\,y\,R\,z)$) on a set $X$ so that the sets \[R y = \{ x \in X: x \,R\, y\} \ \ (y\in X)\] are ideals with respect to the lower quasi-order $\leq_{R}$ defined by \[x\leq_{R} y \ \Leftrightarrow R x \subseteq R y.\] And $R$ is called a C-order if, moreover, $\leq_{R}$ is an order, i.e., $R y \!=\! R z$ implies $y \!=\! z$. The pair $(X,R )$ is referred to as a C-(quasi-)ordered set. Any C-order $R$ is an approximating auxiliary relation for the poset $(X,\leq_{R})$ in the sense of. For an arbitrary relation $R$ on a set $X$ and any subset $Y$ of $X$, put \[R\, Y = \{ x\in X: \exists\,y\in Y\, (x\,R\,y)\}, \ YR = \{ x\in X: \exists\,y\in Y\, (y\,R\,x)\},\vspace{-.5ex}\] \[\ _{R}{\cal O} = \{ \,R\,Y: Y\subseteq X\}, \ {\cal O}_{R} = \{ YR: Y\subseteq X\}. \ \] $_{R}{\cal O}$ and ${\cal O}_{R}$ are closed under arbitrary unions, and in case $R$ is an ideal relation, ${\cal O}_{R}$ is even a topology. On the other hand, if $R$ is idempotent then $_{R}{\cal O}$ consists of all rounded subsets, where a subset $Y$ is said to be round(ed) if $Y = R Y$ ,. Typical examples of C-orders are the way-below relations $\ll$ of continuous posets $P$, in which for any element $y$ the set \[\textstyle{\ll\! y = \{ x\in P: x \ll y\} = \bigcap \,\{ D\in {\cal I} P: \bigvee\! D \mbox{ exists and } y\leq \bigvee\! D\}}\] is an ideal (called the way-below ideal of $y$) with join $y$ (recall that ${\cal I} P$ is the set of all ideals). The next result, borrowed from, leans on the interpolation property of continuous posets, saying that their way-below relation is idempotent. (240,0) The aforementioned one-to-one correspondence between core spaces and C-quasiorders is based on the following remark and definition: every space $(X,{\cal S})$ carries a transitive (but only for A-spaces reflexive) interior relation $R_{{\cal S}}$, given by \[x \,R_{{\cal S}}\,y \, \Leftrightarrow \, y \in ({\uparrow\!x})^{\circ} = int_{{\cal S}} ({\uparrow\!x}), \vspace{-.5ex}\] where ${\uparrow\!x} = \bigcap\, \{ U \in {\cal S}: x \in U\}$ is the core of $x$. Note that for any subset $A$ of $X$,\ $R_{{\cal S}}A$ is a lower set in $(X, \leq_{{\cal S}})$, as $\,x \!\leq\! y\,R_{{\cal S}} \, z$ entails $z\in ({\uparrow\!y})^{\circ} \!\subseteq\! ({\uparrow\!x})^{\circ}$, hence $x \, R_{{\cal S}} \,z$.\ A map $f$ between C-quasi-ordered sets $(X,R)$ and $(X',R')$ interpolates if for all $y \!\in\! X$ and $x' \!\in\! X'$ with $x' R\, f(y)$ there is an $x \in X$ with $x' R' f(x)$ and $x \,R\, y$, and $f$ is isotone (preserve the lower quasi-orders) iff $x\leq_{R} y$ implies $f(x) \leq_{R '}\! f(y)$. $1$ and (2) have been established in. There are several reasonable alternative choices for the morphisms. For example, the quasiopen maps (for which the saturations of images of open sets are open) between core spaces are the relation preserving isotone maps between the associated C-quasi-ordered sets (cf. ). (3) Let $R$ be the interior relation and $\leq$ the specialization order $\leq_{R}$. We prove the identity $(R A)^-= A^-$, using idempotency of $R$ in the equivalence \* below: $y\in (R A)^-\ \Leftrightarrow \ \forall\, U\in {\cal O}_{R}\ (y\in U \Rightarrow U\cap R A \neq \emptyset)$ $\Leftrightarrow \ \forall \, x\in X\ (x \,R \,y \,\Rightarrow\, xR \cap R A \neq \emptyset) \ \stackrel{}{\Leftrightarrow} \ \forall \, x\in X\ (x \,R \,y \Rightarrow xR \cap A \neq \emptyset)$ $\Leftrightarrow \ \forall\, U\in {\cal O}_{R}\ (y\in U \Rightarrow U\cap A \neq \emptyset) \ \Leftrightarrow \ y\in A^-.$ In particular, $(R A)^-= A$ in case $A$ is closed. On the other hand, any rounded set $Y$ satisfies the identity $Y = R Y = R (Y^-)$: since each $xR$ is an open set, we have $x\in R Y \ \Leftrightarrow \ xR \cap Y \neq \emptyset \ \Leftrightarrow \ xR \cap Y^-\neq \emptyset \ \Leftrightarrow \ x\in R (Y^-).$ And if $Y$ is any lower set with $Y^-\! = A$ then $R A = R (Y^-) = R Y \subseteq {\downarrow\!Y} = Y.$ (4) Recall that a subset $A$ is irreducible iff it is nonempty and $A \subseteq B\cup C$ implies $A \subseteq B$ or $A \subseteq C$ for any closed sets $B,C$. Directed subsets and their closures are irreducible. The rounded ideals of a C-quasi-ordered set (or a core space) form a domain ${\cal I}_{R} = \{ I\in {\cal I} (X,\leq_{R}): I = R I\}$, being closed under directed unions. It is easy to see that $I \ll J$ holds in ${\cal I}_{R}$ iff there is an $x\in J$ with $I\subseteq {\downarrow\!x}$, and that the join of the ideal $\ll \!J = {\downarrow_{{\cal I}_{R}}} \{ R x: x\in J\}$ is $J$. Thus, ${\cal I}_{R}$ is continuous (cf. , ). For the isomorphism claim, observe that the coprime rounded sets are the rounded ideals, the coprime closed sets are the irreducible ones, and a lattice isomorphism preserves coprimes (an element $q$ is coprime* iff the complement of ${\uparrow\!q}$ is an ideal). (5) By (4), $I = R_{{\cal S}} y$ is an ideal with ${\downarrow\! y} = I^-$, so $y$ is a least upper bound of $I$; now, $y \in U \in \sigma (X,\leq)$ implies $U\cap R_{{\cal S}} y \not = \emptyset$, i.e. $y\in UR_{{\cal S}}$. Thus, $U = UR_{{\cal S}}\in {\cal S}$. (6) Let $C$ be a compact saturated set (i.e. $C \!\in\! \pi {\cal S}$) and $y \!\in\! X\setminus C$. There is an open neighborhood $U$ of $C$ not containing $y$. By local supercompactness of $(X,{\cal S})$, we have $U = \bigcup\, \{ int_{{\cal S}}({\uparrow\! x}): x\in U\}$, and by compactness of $C$, we find a finite $F \subseteq U$ with $C \!\subseteq \bigcup\, \{ int_{{\cal S}} ({\uparrow\! x}): x \!\in\! F\} \!\subseteq\! {\uparrow\! F} \!\subseteq\! U$. For $Q = (X,\leq)$, ${\uparrow\! F}$ is $\upsilon\widetilde{Q}$-closed and does not contain $y$. Thus, $C$ is $\upsilon\widetilde{Q}$-closed, and $\pi {\cal S}$ is contained in $\upsilon \widetilde{Q}$; the reverse inclusion is obvious, since cores (principal filters) are compact and saturated. Recall that a sober space is a T$_0$-space whose only irreducible closed sets are the point closures; and a monotone convergence space (mc-space) or d-space is a T$_0$ space in which the closure of any directed subset is the closure of a singleton (see, , ). Now, from Theorem [\[Cspace\]](#Cspace){reference-type="ref" reference="Cspace"}, one deduces (cf. ,, ): For the case of continuous posets that are not necessarily domains, see ,. Afficionados of domain theory might remark that continuous frames are automatically spatial (see, ). But that 'automatism' requires choice principles, which we wanted to avoid in the present discussion. However, it seems to be open whether the spatiality of supercontinuous frames may be proved in a choice-free manner. Since a T$_0$-space and its sobrification have isomorphic open set frames, it follows from Corollary [\[contdom\]](#contdom){reference-type="ref" reference="contdom"} that a T$_0$-space is a core space iff its sobrification is the Scott space of a continuous domain. This completion process is reflected, via Theorem [\[Cspace\]](#Cspace){reference-type="ref" reference="Cspace"}, by the fact that the rounded ideal completion of a C-ordered set is a continuous domain, and the C-order is extended to its completion, meaning that $x\,R\,y$ is equivalent to $R x \ll R y$ in the completion (cf., ,, ). # Core stable spaces and sector spaces {#sectorspaces} As every core space is a web space, it is equal to the upper spaces of its patch spaces and therefore $\zeta$-determined for any coselection $\zeta$ (see Proposition [\[webpatch\]](#webpatch){reference-type="ref" reference="webpatch"}). We now are going to determine explicitly these patch spaces, which turn out to have very good separation properties, whereas the only T$_1$ core spaces (in fact, the only T$_1$ web spaces) are the discrete ones. For alternative characterizations of such patch spaces, we need further properties of the interior operators. Call a quasi-ordered space $(Q, {\cal T} )$ core stable or c-stable if \[\textstyle{{\uparrow\!O} = \bigcup \,\{ \,int_{{\cal T}}({\uparrow\!u}): u\in O\} \ \mbox{ for each }O \in {\cal T}, }\] and d-stable if for any filtered (i.e. down-directed) subset $D$ of $Q$, \[\textstyle{int_{{\cal T}}D \subseteq \bigcup\,\{ int_{{\cal T}} \, ({\uparrow\!u}): u\in cl_{{\cal T}^{\geq}} D\}.}\] While core stability is a rather strong property, d-stability is a rather weak one (trivially fulfilled if all filters are principal). The terminology is justified by $1$ If the lower space $\,{\rm L}T = (X,{\cal T}^{\geq})$ is a d-space then for any filtered set $D$ in $(X,\leq)$ there is a $u$ with ${\uparrow\!u} = cl_{{\cal T}^{\geq}} D$ and $\,int_{{\cal T}} D \subseteq int_{{\cal T}} ({\uparrow\!u})$. Without proof, we note that ${\rm L}T$ is a d-space whenever $T$ is hyperconvex and $(X,\geq)$ is a domain. (2) Clearly, a core stable semi-qospace $(X,\leq ,{\cal T})$ is ${\uparrow}$-stable, whence ${\mathcal T}^{\leq} = \{ {\uparrow\!O}: O \in {\mathcal T}\}$. For $U\in {\cal T}^{\leq}$ and $x\in U$ there exists a $u\in U$ with $x\in int_{{\cal T}} ({\uparrow\!u})$. Then $W = {\uparrow int_{{\cal T}}({\uparrow\!u})}\in {\cal T}^{\leq}$ (by ${\uparrow}$-stability) and $x\in W\subseteq {\uparrow\!u} \subseteq {\uparrow\!U} = U$. Since $\leq$ is the specialization order of ${\cal T}^{\leq}$, this ensures that $(X,{\cal T}^{\leq})$ is a core space. Conversely, suppose $(X,{\cal T}^{\leq}) = (X, \{ {\uparrow\!O}: O \in {\mathcal T}\})$ is a core space. For $O\in {\mathcal T}$ and $y \in {\uparrow\!O}$, there is an $x\in {\uparrow\!O}$ and a $U\in {\cal T}^{\leq} \subseteq {\cal T}$ with $y\in U\subseteq {\uparrow\!x}$. Now, pick a $u\in O$ with $x\in {\uparrow\!u}$; then $y\in U \subseteq {\uparrow\!x} \subseteq {\uparrow\!u}$, whence ${\uparrow\!O} \subseteq \bigcup \,\{ \,int_{{\cal T}}({\uparrow\!u}): u\in O\}$. (3) Let $T = (X,\leq, {\cal T} )$ be a core stable semi-qospace. For $x\in O\in {\cal T}^{\leq}$, there is a $u\in O$ with $x\in U = int_{{\cal T}} ({\uparrow\!u}) \subseteq {\uparrow\!u} \subseteq O$. By ${\uparrow}$-stability, we have $U\in {\cal T}^{\leq}$, and since $T$ is a semi-qospace, ${\uparrow\!u}$ is a closed upper set. Hence, $T$ is upper regular. In order to check local filteredness, pick for $x\in O\in {\cal T}$ an element $u\in O$ with $x\in int_{{\cal T}} ({\uparrow\!u}) \mathop{\cap} O \subseteq {\uparrow\!u} \mathop{\cap} O\,$; this is a filtered set, possessing the least element $u$. Moreover, $T$ is not only ${\uparrow}$-stable (see (2)) but also d-stable, since for any subset $D$ and any $x\in O = int_{{\cal T}} D$, there is a $u \in O \subseteq D \subseteq cl_{{\cal T}^{\geq}} D$ with $x\in int_{{\cal T}}({\uparrow\!u})$. Conversely, suppose that $T$ is an upper regular, locally filtered, ${\uparrow}$-stable and d-stable semi-qospace. Then, for $O\in {\cal T}$ and $x\in {\uparrow\!O}$, we have: $x\in {\uparrow\!O} \in {\cal T}^{\leq}$ (by ${\uparrow}$-stability), there are $U\in {\cal T}^{\leq}$ and $B\in {\cal T}^{\geq c}$ with $x\in U \subseteq B \subseteq {\uparrow\!O}$ (by upper regularity), a filtered $D$ with $x\in int_{{\cal T}} D \subseteq D \subseteq U$ (by local filteredness), and an element $y\in cl_{{\cal T}^{\geq}} D$ with $x \in int_{{\cal T}} ({\uparrow\!y})$ (by d-stability). It follows that $y \in cl_{{\cal T}^{\geq}} D \subseteq cl_{{\cal T}^{\geq}} U \subseteq B \subseteq {\uparrow\!O}$. Now, choose a $u\in O$ with $u\leq y$, hence ${\uparrow\!y} \subseteq {\uparrow\!u}$, to obtain $x\in int_{{\cal T}} ({\uparrow\!u})$. Thus, ${\uparrow\!O} \subseteq \bigcup\,\{ int_{{\cal T}} ({\uparrow\!u}): u\in O\}$, showing that $T$ is core stable. By Lemma [\[cstable\]](#cstable){reference-type="ref" reference="cstable"}, every core stable semi-qospace is a qospace (being upper regular); in fact, core stability of a semi-qospace splits into the four properties (c1) upper regular (c2) locally filtered (c3) ${\uparrow}$-stable (c4) d-stable. These properties are independent: none of them follows from the other three. is a complete but not meet-continuous lattice and a compact pospace when equipped with the Lawson topology. It satisfies (c1), (c2) and (c4) (because all filters are principal). However, (c3) is violated, since $\{ a \}$ is open, while ${\uparrow\!a} = \{ a, \top\}$ is not. Given a quasi-ordered space $(X,\leq,{\cal T})$, we call any nonempty set of the form ${\uparrow\!u}\mathop{\cap} V$ with $V \!\in \!{\cal T}^{\,\geq}\!$ a sector, and a $\zeta$-sector if $V\in \zeta ({\cal T}^{\leq})$ for a coselection $\zeta$. Hence, $u$ is the least element of the sector, and every sector is obviously a web around any point it contains. By a ($\zeta$-)sector space we mean an ${\uparrow}$-stable semi-qospace in which each point $x$ has a base of ($\zeta$-)sector neighborhoods ${\uparrow\!u}\cap V$ (but the point $x$ need not be the minimum $u$ of such a sector). $1$ and (2): Let $T = (X,\leq,{\cal T})$ be a sector space. For $x\in O\in {\cal T}$, there are $u \!\in\! X$ and $V \!\in\! {\cal T}^{\,\geq}$ such that $x\in int_{{\cal T}} ({\uparrow\!u} \mathop{\cap} V) = int_{{\cal T}} ({\uparrow\!u})\cap V \subseteq {\uparrow\!u} \mathop{\cap} V \subseteq O$. Then $u\in {\uparrow\!u}\cap V \subseteq O$ and $x \in int_{{\cal T}}({\uparrow\!u})$, whence $O \subseteq \bigcup \,\{ int_{{\cal T}}({\uparrow\!u}): u \!\in\! O\}$. Using ${\uparrow}$-stability and applying the previous argument to ${\uparrow\!O}$ instead of $O$, one concludes that $T$ is core stable. Again by ${\uparrow}$-stability, $U = {\uparrow int_{{\cal T}}}({\uparrow\!u}\,\cap V)$ belongs to ${\cal T}^{\,\leq}$, and the above reasoning yields $x\in U\cap V \subseteq {\uparrow\!u}\cap V\subseteq O$, proving strong convexity. Now, let $(X,\leq,{\cal T})$ be any strongly convex core stable semi-qospace. Then, by Lemma [\[cstable\]](#cstable){reference-type="ref" reference="cstable"}, $(X,{\cal T}^{\leq})$ is a core space with specialization order $\leq$, while ${\cal T}^{\geq}$ has the dual specialization order. By strong convexity, ${\cal T}$ is the patch topology ${\cal T}^{\leq} \vee {\cal T}^{\geq}$. On the other hand, by Proposition [\[pat\]](#pat){reference-type="ref" reference="pat"}, any patch space $(X,\leq, {\cal T} )$ of a core (or web) space $(X,{\cal S} )$ is a web-quasi-ordered space, in particular ${\uparrow}$-stable. For $x\in U \cap V$ with $U\in {\cal S}$ and $V\in {\cal T}^{\geq}$, there are $u\in U$ and $W\in {\cal S}$ with $x\in W \subseteq {\uparrow\!u}$; it follows that $x\in W \mathop{\cap} V \subseteq {\uparrow\!u} \mathop{\cap} V \subseteq U \mathop{\cap} V$. Thus, $(X,\leq, {\cal T} )$ is a sector space. (3) follows from (2) and Lemma [\[cstable\]](#cstable){reference-type="ref" reference="cstable"}. Any upper regular semi-qospace is a qospace. The modified claims involving coselections $\zeta$ are now easily derived from Proposition [\[pat\]](#pat){reference-type="ref" reference="pat"} and Lemma [\[webpatch\]](#webpatch){reference-type="ref" reference="webpatch"}, using the fact that core spaces are web spaces. We are ready to establish a categorical equivalence between core spaces and $\zeta$-sector spaces. As explained in, the right choice of morphisms is a bit delicate. Continuous maps would be the obvious morphisms between core spaces. On the other hand, one would like to have as morphisms between quasi-ordered spaces the isotone continuous maps. But, as simple examples in show, a continuous map between two core spaces need not be continuous as a map between the associated $\zeta$-sector spaces, and a continuous isotone map between $\zeta$-sector spaces need not be continuous for the weak lower topologies (consider the map $f$ on the lattice $L$ in Example [\[Exnotcon\]](#Exnotcon){reference-type="ref" reference="Exnotcon"} with $f(a_n)=f(b_n) =a_0$, $f(c_n)=c_0$). Therefore, we take the $\zeta$-proper maps (see Section [2](#convex){reference-type="ref" reference="convex"}) as morphisms between core spaces in order to save the isomorphism theorem. Let us summarize the previous results. A related class of morphisms is formed by the core continuous maps, i.e. continuous maps for which preimages of cores are cores. In terms of the specialization orders, the latter condition means that these maps are residual (preimages of principal filters are principal filters), or equivalently, that they have lower adjoints (cf. and ). Core continuous maps are $\alpha$-, $\sigma$-and $\upsilon$-proper, but not conversely. The following facts are established in (see also ): In contrast to locally supercompact spaces (core spaces), locally hypercompact spaces need not be $\zeta$-determined, and the categorical equivalence between core spaces (i.e. locally hypercompact web spaces) and $\zeta$-sector spaces does not extend to locally hypercompact spaces without additional restrictions. The lattice $\widetilde{L}$ in Example [\[Exnotcon\]](#Exnotcon){reference-type="ref" reference="Exnotcon"} is locally hypercompact but not locally supercompact in the weak upper topology; in fact, it fails to be a web space: the points $b_n$ have no small web neighborhoods. Since $\widetilde{L}$ is not $\zeta$-determined unless $\zeta \widetilde{L} \!=\! \alpha \widetilde{L}$, Lemma [\[pat\]](#pat){reference-type="ref" reference="pat"} does not apply to that case. For the complete lattice $M = (X,\leq)$ in Example [\[Ex33\]](#Ex33){reference-type="ref" reference="Ex33"}, the Scott space $\Sigma M$ is locally hypercompact and $\upsilon$-determined, but its weak patch space $\Lambda M = {\rm P}_{\upsilon}\Sigma M$ is not ${\uparrow}$-stable: while $\{ a\}$ is open, ${\uparrow\!a}$ is not. # Fan spaces {#fanspaces} We have seen that, by virtue of the weak patch functor ${\rm P}_{\upsilon}$, the core spaces bijectively correspond to the $\upsilon$-sector spaces. We shall now give some effective descriptions of these specific qospaces. By a fan we mean a nonempty set of the form ${\uparrow\!u}\setminus {\uparrow\!F}$ for some finite $F$. In case $(X,\leq,{\cal T})$ is an upper semi-qospace, each fan is a sector, in fact, a $\upsilon$-sector. An ${\uparrow}$-stable semi-qospace space with a base of fans will be called a fan space. A nonempty product of quasi-ordered spaces is a fan space iff all factors are fan spaces and all but a finite number have a least element. (130,0) (255,0) In order to see that hyperconvex upper regular semi-qospaces are regular, it suffices to guarantee the separation of points from subbasic closed sets. For closed lower sets, this is possible by upper regularity. For subbasic closed sets ${\uparrow\!x}$ and points $y\in X\setminus {\uparrow\!x}$, one finds an open upper set $U$ and a closed upper set $B$ with ${\uparrow\!x} \subseteq U \subseteq B \subseteq X\setminus{\downarrow\!y}$; then $X\setminus B$ is an open set disjoint from $U$ and containing $y$. Now, from Theorem [\[sectors\]](#sectors){reference-type="ref" reference="sectors"} and Proposition [\[sectorcat\]](#sectorcat){reference-type="ref" reference="sectorcat"}, we infer for the case $\zeta = \upsilon$: Let us make the last statement in Theorem [\[fans\]](#fans){reference-type="ref" reference="fans"} more precise. For the definition of quasi-uniformities and a study of their relationships to (quasi-)ordered spaces, refer to Fletcher and Lindgren or Nachbin. For a quasi-uniformity ${\cal Q}$, the dual ${\cal Q}^{-1}$ is obtained by exchanging first and second coordinate, and ${\cal Q}^$ is the uniformity generated by ${\cal Q} \cup {\cal Q}^{-1}$. The topology $\tau ({\cal Q})$ generated by ${\cal Q}$ consists of all subsets $O$ such that for $x \!\in\! O$ there is a $\,U\!\in\! {\cal Q}$ with $xU = \{ y: (x,y) \in U\} \!\subseteq\! O$. The following results are due to Brümmer and Künzi and Lawson. For core spaces, these conclusions may be strengthened as follows. We apply Lemma [\[localunif\]](#localunif){reference-type="ref" reference="localunif"} to the base ${\cal B} = \{ xR: x\!\in\! X\}$ of ${\cal O}_{R} = {\cal S}$ (Theorem [\[Cspace\]](#Cspace){reference-type="ref" reference="Cspace"}). By idempotency, $x\,R\, y$ implies $x\,R\, x'R\, y$ for some $x'$, hence $x'R \ll xR$. This yields the equation $xR = \bigcup \,\{ x'\!R \in {\cal B}: x'\!R \ll xR\}$, and then (1) follows from Lemma [\[localunif\]](#localunif){reference-type="ref" reference="localunif"}. Now, Theorem [\[Cspace\]](#Cspace){reference-type="ref" reference="Cspace"} (6) and Lemma [\[localunif\]](#localunif){reference-type="ref" reference="localunif"} (2) give $\tau ({\cal Q}^{-1}) = \tau_{\pi}{\cal S} = \upsilon \widetilde{Q}$. Finally, $\bigcap {\cal Q}$ is the specialization order of $\tau ({\cal Q})$, and ${\cal S}^{\upsilon} = {\cal S}^{\pi} = \tau ({\cal Q}^)$, by Lemma [\[localunif\]](#localunif){reference-type="ref" reference="localunif"} (3). It is also easy to check the equations ${\cal S} = \tau ({\cal Q})$ and $\upsilon \widetilde{Q} = \tau ({\cal Q}^{-1})$ directly: For $U = x'\!R \!\rightarrow\! y'R$, the set $xU$ is equal to $y'R$ if $x'R\, x$, and to $X$ otherwise, while the set $Uy$ is equal to $X\setminus x'R$ if $y'\!\!\not\!R\, y$, and to $X$ otherwise. In case $y'R\,x'$ but not $y'R\, y$, the inclusion $y\in X\setminus y'R \subseteq X\setminus {\uparrow\!y'} \subseteq x'R = Uy$ shows that each $Uy$ is a neighborhood of $y$ in the weak lower topology. In (see also, , ), it is shown that in ZFC a pospace $T = (X,\leq,{\cal T})$ is determined by a quasi-uniformity ${\cal Q}$, i.e. $\leq \ = \bigcap {\cal Q}$ and ${\cal T} = \tau ({\cal Q}^)$, iff it has a (greatest) order compactification, which in turn is equivalent to saying that $T$ is completely regular(ly) ordered, i.e., a pospace such that for $x \!\in\! U \!\in\! {\cal T}$, there are continuous $f,g: X \rightarrow [0,1]$ with $f$ isotone and $g$ antitone, $f(x) = g(x) = 1$, and $f(y) = 0$ or $g(y) = 0$ for $y\in X\setminus U$. Dropping the antisymmetry, we arrive at # Continuous domains as pospaces {#domain} We now are going to establish a representation of continuous domains by certain pospaces, generalizing the famous characterization of continuous lattices as meet-continuous lattices whose Lawson topology is Hausdorff (see ). In contrast to the situation of complete lattices, we require at most up-completeness. For our purposes, we need a further definition. Given an upset selection $\zeta$ (see Section [2](#convex){reference-type="ref" reference="convex"}), a quasi-ordered space $(Q,{\cal T}) = (X,\leq,{\cal T} )$ or its topology is said to be $\zeta$-stable* if the interior operator $^{\circ}$ of the upper topology ${\cal T}^{\leq}$ satisfies the equation \[\textstyle{ Y^{\circ} = \bigcup\,\{ ({\uparrow\!y})^{\circ}: y\in Y\}} \ \mbox{ for all } Y \!\in \! \zeta Q.\] If $(Q,{\cal T})$ is ${\uparrow}$-stable, the operator $^\circ$ may be replaced by the interior operator of the original topology ${\cal T}$. For instance, core stable means $\alpha$-stable plus ${\uparrow}$-stable.\ We denote by $\curlyvee Q$ and $\curlywedge Q$ the set of all finite unions (incl. $\emptyset \!= \bigcup \emptyset$) resp. intersections (incl. $X \!= \bigcap \emptyset$) of principal filters, i.e., the unital join-resp. meet-subsemilattice they generate in the lattice $\alpha Q$ of all upper sets. Further, we denote by $\diamondsuit Q$ the sublattice of $\alpha Q$ generated by the principal filters and containing $\emptyset$ and $X$. $1$ Suppose $T$ is $\curlywedge$-stable. For finite $F\!\subseteq\! R y$, the set $F^{\uparrow\!} = \bigcap \, \{{\uparrow\! x}: x\in F\}$ is a member of $\curlywedge Q$, whence $y\in \bigcap \, \{ ({\uparrow\! x})^{\circ}: x\in F\} = (F^{\uparrow})^{\circ} = \bigcup\,\{ ({\uparrow\! u})^{\circ}: u\!\in\! F^{\uparrow\!}\}$, i.e., $F^{\uparrow}\cap R y \neq \emptyset$. In other words, $R y$ is directed (and always a lower set). Conversely, if that is the case then, for each finite subset $F$ of $Q$, one deduces from $y\in (F^{\uparrow})^{\circ}$ that $F$ is contained in $R y$, whence there is an upper bound $u$ of $F$ in $R y$; thus, $y\in \bigcup\,\{ ({\uparrow\!u})^{\circ}\!: u \!\in\! F^{\uparrow}\}$. The reverse inclusion $\bigcup\,\{ ({\uparrow\!u})^{\circ}\!: u \!\in\! F^{\uparrow}\} \!\subseteq\! (F^{\uparrow})^{\circ}$ is clear. (2) Obviously, $\diamondsuit$-stable quasi-ordered spaces are $\curlyvee$-and $\curlywedge$-stable. Conversely, suppose $T$ is $\curlyvee$-and $\curlywedge$-stable. Any $Y\in \diamondsuit Q$ is of the form $Y = \bigcap \,\{ {\uparrow\!F_i}: i \! < \! n\}$ where each of the finitely many sets $F_i$ is finite. Hence, $Y^{\circ} = \bigcap \,\{ ({\uparrow\!F_i})^{\circ}: i < n \} =$ $\bigcap \,\{ \,\bigcup \,\{ ({\uparrow\!x})^{\circ}: x\in F_i \}: i < n \} =$ (set distributivity) $\bigcup \,\{ \,\bigcap\,\{ ({\uparrow\!\chi_i})^{\circ}\!: i \!<\! n \}: \chi \in \prod_{i<n}\! F_i \} =$ $\bigcup\,\{ ({\uparrow\!y})^{\circ}\!: y\in \bigcap\,\{{\uparrow\!\chi_i}: i<n\}, \,\chi \in \prod_{i<n}\! F_i \} =$ (set distributivity) $\bigcup\,\{ ({\uparrow\!y})^{\circ}: y\in \bigcap\,\{ {\uparrow\!F_i}: i<n\} = Y \}.$ (3) If ${\rm U}T$ is a web space then its interior operator preserves finite unions; in particular, for finite $F$, one has $({\uparrow\!F})^{\circ}\! = \bigcup \,\{ ({\uparrow\!y})^{\circ}\!: y\in F\} = \bigcup \,\{ ({\uparrow\!z})^{\circ}\!: z\in {\uparrow\!F}\}$ (as $y\leq z$ implies ${\uparrow\!z}\subseteq {\uparrow\!y}$). (4) In a join-semilattice $Q$ with least element, $\curlywedge Q$ is the set of all principal filters. (5) If $Q$ is a lattice with least element and ${\cal T}$-continuous unary meet-operations $\wedge_x: y \mapsto x\wedge y$, then ${\rm U}T$ is a web space; hence, $T$ is $\curlyvee$-stable by (3). Furthermore, $T$ is $\curlywedge$-stable by (4), and finally $\diamondsuit$-stable by (2). (6) Let $T =(X,\leq,{\cal T} )$ be a lower semi-qospace. Then $\leq$ is the specialization order of ${\cal T}^{\leq}$. The upper space ${\rm U} T = (X,{\cal T}^{\leq})$ is a core space iff $U \!= \bigcup\,\{ ({\uparrow\!y})^{\circ}\!: y \!\in\! U\}$ for all $U \!\in\! {\cal T}^{\leq}$, which is equivalent to requiring that $T$ is $\curlyvee$-stable and satisfies the equation $U = \bigcup\,\{ ({\uparrow\!F})^{\circ}: F\subseteq U, \, F \mbox{ finite}\}$ for all $U\in {\cal T}^{\leq}$. But the latter condition means that the upper space is locally hypercompact. By an mc-ordered space we mean an ordered space such that every monotone net in the space or, equivalently, every directed subset of the space, regarded as a net, has a supremum to which it converges. It is shown in that the strongly convex mc-ordered semi-pospaces are exactly the patch spaces of monotone convergence spaces (d-spaces), and that the Lawson spaces of domains are just the hyperconvex, mc-ordered and upper m-determined semi-pospaces, where an ordered space is upper m-determined iff every upper set $U$ intersecting all directed sets whose closure meets $U$ is open. We are prepared for the main result in this section, characterizing the Lawson spaces of continuous domains in various ways. (1)$\,\Rightarrow\,$(2): Let $P$ be a continuous domain with $T = \Lambda P$. By Corollary [\[Cspace\]](#Cspace){reference-type="ref" reference="Cspace"}, its upper space $\Sigma P$ is a sober core space, and by Theorem [\[fans\]](#fans){reference-type="ref" reference="fans"}, $\Lambda P$ is a fan space. (2)$\,\Rightarrow\,$(3): By Theorem [\[fans\]](#fans){reference-type="ref" reference="fans"}, $T =(P,{\cal T} )$ is regular, upper regular and T$_1$, i.e., T$_3$ and upper T$_3$-ordered, a fortiori T$_2$-ordered; the upper space ${\rm U} T$ is a core space and a d-space. By Corollary [\[Cspace\]](#Cspace){reference-type="ref" reference="Cspace"}, $P$ is a continuous, hence meet-continuous domain, ${\cal T}^{\leq}$ is its Scott topology $\sigma P$, and ${\cal T} \!=\! {\cal T}^{\leq \upsilon}$ is the Lawson topology $\lambda P$. Since ${\rm U} T$ is a core space, $T$ is $\curlywedge$-stable, by Theorem [\[Cspace\]](#Cspace){reference-type="ref" reference="Cspace"} (1) and Lemma [\[stable\]](#stable){reference-type="ref" reference="stable"} (1). (2)$\,\Rightarrow\,$(4): By Theorem [\[fans\]](#fans){reference-type="ref" reference="fans"}, ordered fan spaces are hyperconvex core stable pospaces, and such spaces are mc-ordered if their upper space is a d-space. (3)$\,\Rightarrow\,$(5): If $P$ is a meet-continuous domain then $\Sigma P$ is a web space, so its weak patch space, the Lawson space $\Lambda P$, is web-ordered, hence ${\uparrow}$-stable (Proposition [\[webpatch\]](#webpatch){reference-type="ref" reference="webpatch"}); it is a hyperconvex semi-pospace, and mc-ordered since $\Sigma P$ is a d-space. By Lemma [\[stable\]](#stable){reference-type="ref" reference="stable"} (3), $T$ is $\curlyvee$-stable, and if it is $\curlywedge$-stable, it is $\diamondsuit$-stable by Lemma [\[stable\]](#stable){reference-type="ref" reference="stable"} (2). (4)$\,\Rightarrow\,$(5): Core stable semi-pospaces are ${\uparrow}$-stable, $\diamondsuit$-stable and T$_2$ (Lemma [\[cstable\]](#cstable){reference-type="ref" reference="cstable"}). (5)$\,\Rightarrow\,$(1): By Lemma [\[stable\]](#stable){reference-type="ref" reference="stable"} (1), the sets $R y = \{ x: y\!\in\! ({\uparrow\! x})^{\circ}\}$ are directed. As $T$ is mc-ordered, $R y$ has a join $x \!=\! \bigvee R y$. Assume $x \!<\! y$ and choose disjoint sets $V,W\in {\cal T}$ with $x\in V$ and $y\in W$. By hyperconvexity, there is a $U \in {\cal T}^{\leq}$ and a finite set $F$ with $x\in U\setminus {\uparrow\! F}\subseteq V$. Then $y\in U\cap W\subseteq {\uparrow\! F}$ (as $x < y$ and $V\cap W = \emptyset$), and $U\cap W\in {\cal T}$ yields ${\uparrow\!(U\cap W)}\in {\cal T}^{\leq}$ by ${\uparrow}$-stability. Thus, $y\in ({\uparrow\! F})^{\circ} = \bigcup\,\{ ({\uparrow\! u})^{\circ}\!: u\!\in\! F\}$ by $\curlyvee$-stability, and so $u\in R y$ for some $u\in F$, which leads to the contradiction $u\leq \bigvee \! R y = x\in U\setminus {\uparrow\! F}$. Hence, $y$ is the directed join of $R y$. Since $T =(P,{\cal T} )$ is mc-ordered, ${\cal T}^{\leq}$ is coarser than $\sigma P$. It follows that $R y$ coincides with the way-below ideal $\ll\! y$ (indeed, $x \ll y$ implies $x\,R\,y$, since $R y$ is an ideal with join $y$; and $x\,R\, y$ implies $x \!\ll\! y$, since for directed $D$, $x \,R \, y = \bigvee\! D$ entails $y\in int_{{\cal T}^{\leq}} {\uparrow\!x} \subseteq int_{\sigma P} {\uparrow\!x} \subseteq x\!\!\ll$). Therefore, $P$ is continuous, ${\cal T}^{\leq} = \sigma P$ (as the base $\{ xR: x\in P\}$ of $\sigma P$ is contained in ${\cal T}^{\leq}$), and finally, ${\cal T} = {\cal T}^{\leq \upsilon} = \lambda P$. Since all topologies on join-semilattices with 0 are $\curlywedge$-stable, we obtain at once: In categorical terminology, parts of Corollary [\[contdom\]](#contdom){reference-type="ref" reference="contdom"} and Theorem [\[pospace\]](#pospace){reference-type="ref" reference="pospace"} read as follows: Suitable examples show that the properties in Theorem [\[pospace\]](#pospace){reference-type="ref" reference="pospace"} (5) are independent; that is, none of them follows from the combination of the other four properties. This is a compact pospace, hence mc-ordered and T$_2$. It is $\curlyvee$-stable, since $({\uparrow\!F})^{\circ} = {\uparrow (F\cap \{ 0,2\})} \cup (F\cap \{ 3\}) = \bigcup \{ ({\uparrow\!x})^{\circ}: x\in F\}$ for all finite subsets $F$. It is $\curlywedge$-stable (because $(X,\sqsubseteq )$ is a complete lattice), and therefore $\diamondsuit$-stable (Lemma [\[stable\]](#stable){reference-type="ref" reference="stable"} (2)). Furthermore, it is hyperconvex, in view of the equations $\{ 0\} = X\setminus {\uparrow\!\{ 2,3\}}$, $\{ 3 \} = (X\setminus \{ 0\} )\setminus {\uparrow\!2}$, and $U = {\uparrow\!U}\setminus {\uparrow\!1}$ for $U \subseteq \ ]1,2\,]$. But this pospace fails to be ${\uparrow}$-stable, since $\{ 3\}$ is open, while $\{ 1,3\} = {\uparrow\!3}$ is not. The basic results about core spaces and fan spaces are more than 30 years old; they have been reported by the author at the Annual Meeting of the DMV, Dortmund 1980 and at the Summer School on Ordered Sets and Universal Algebra, Donovaly 1985 but have not been published in a systematic treatise until now. A common theory of web spaces and core spaces is possible by introducing so-called $\kappa$-web spaces, where $\kappa$ is a cardinal parameter; that approach was initiated in. A comprehensive theory of so-called $\zeta$-domains and their topological manifestation, providing common generalizations of continuous, hypercontinuous, algebraic and many other variants of domains or posets, is presented in. # Semitopological and topological semilattices {#semi} Let us now apply some of the previous results to the situation of semilattices, by which we always mean meet-semilattices. A semilattice-ordered space is an ordered space $S$ whose underlying poset is a semilattice; by slight abuse of language, we call it compatible if the topology is compatible with the semilattice order, and we call a semilattice-and T$_1$-ordered space a T$_1$-semilattice. A semitopological semilattice is a semilattice-ordered space whose unary meet operations $\wedge_x \!: y \mapsto x\wedge y$ are continuous, while in a topological semilattice the binary meet operation is continuous. $1$ For $x,y\in X$ and a filtered neighborhood $D$ of $x\wedge y$ with $D\subseteq U \in {\cal T} = {\cal T}^{\leq}$, the up-closure $F = {\uparrow\!D}$ is a filter still contained in $U$. For $W = int_{{\cal T}} F$, we obtain $W\wedge W \subseteq F\wedge F \subseteq F \subseteq U$ and $x,y\in W$, as $x\wedge y \in W = {\uparrow\!W}$. (2) Let $V$ be a subbasic open set such that $F = X\setminus V$ is a filter. If $x,y\in X$ satisfy $x\wedge y\in V$ then $x\in V$ or $y\in V$ (otherwise $x\wedge y \in F$). Hence, $(x,y)$ lies in $W = (V \!\times\! X)\cup (X \!\times\! V)$, and that is an open set in $S^2$ with $\wedge\, [W] \subseteq V$ (indeed, $(u,v)\in W$ implies $u\in V$ or $v\in V$, and so $u\wedge v\in V$, since $V$ is a lower set). (3) follows from (2), because $\upsilon (X,\geq )$ has a closed subbase of principal filters. (4) The ordered space $(X,\leq,{\cal T}^{\leq})$ is compatible and locally filtered: for $x \!\in\! U \!\in\! {\cal T}^{\leq}$, find a filtered $D\subseteq U$ with $x\in W = int_{{\cal T}}D$; then ${\uparrow\!D}$ is a filter with $x \!\in\! W \!\subseteq\! {\uparrow\!D} \!\subseteq\! U$, and ${\uparrow\!W}$ is ${\cal T}^{\leq}$-open by ${\uparrow}$-stability. By (1), the operation $\wedge$ is ${\cal T}^{\leq}$-continuous; by (3), it is $\upsilon (X,\geq)$-continuous, and then, by hyperconvexity, it is ${\cal T}$-continuous. We find it convenient to call a topological semilattice s-topological (resp. sc-topological) if it has small semilattices (resp. small convex semilattices), that is, each point has a neighborhood base consisting of (convex) subsemilattices (cf. ). We are ready for the characterization of hyperconvex semitopological, resp. s-topological T$_1$-semilattices as certain specific web-ordered spaces. Notice that hyperconvexity ensures closedness of all principal filters, so it suffices to assume that $S$ is a hyperconvex semilattice-ordered lower semi-pospace. (w11)$\,\Leftrightarrow \,$(w12): Apply Proposition [\[webpatch\]](#webpatch){reference-type="ref" reference="webpatch"} to $\zeta = \upsilon$. (w12)$\,\Rightarrow \,$(w13): By Proposition [\[webpatch\]](#webpatch){reference-type="ref" reference="webpatch"}, the upper space $(X,{\cal T}^{\leq})$ of $S = (X,\leq, {\cal T})$ is a web space, and $S$ is its weak patch space. Therefore, as remarked in and, the unary meet operations $\wedge_x$ are ${\cal T}^{\leq}$-continuous; by Lemma [\[top\]](#top){reference-type="ref" reference="top"} (3), they are $\upsilon (X,\geq)$-continuous, and by hyperconvexity, they are also ${\cal T}$-continuous. (w13)$\,\Rightarrow \,$(w12): This was shown in. (w21)$\,\Leftrightarrow \,$(w22): Apply Theorem [\[wideweb\]](#wideweb){reference-type="ref" reference="wideweb"} to $\zeta = \upsilon$. (w22)$\,\Rightarrow \,$(w23): By Lemma [\[top\]](#top){reference-type="ref" reference="top"} (4), $S$ is a topological semilattice. Given $x \!\in\! O \!\in\! {\cal T}$, use hyperconvexity and local filteredness in order to find a $U\in {\cal T}^{\leq}$, a finite set $F$, and a filtered set $D$ such that $x\in int_{{\cal T}} D \subseteq D \subseteq U \setminus {\uparrow\!F} \subseteq O$. Then ${\uparrow\!D}\setminus {\uparrow\!F}$ is a convex subsemilattice with $x\in int_{{\cal T}} D \subseteq {\uparrow\!D}\setminus {\uparrow\!F} \subseteq U\setminus {\uparrow\!F} \subseteq O$, which shows that $S$ has small convex semilattices. (w23)$\,\Rightarrow \,$(w22): By continuity of the unary operations $\wedge_x$, $S$ is ${\uparrow}$-stable: indeed, $O\in {\cal T}$ implies ${\uparrow\!O} = \bigcup\,\{ \wedge_x^{-1}[O]: x\in O\} \in {\cal T}$. Clearly, subsemilattices are filtered. The equivalence of (w31), (w32) and (w33) is now easily verified with the help of (2), Theorem [\[fans\]](#fans){reference-type="ref" reference="fans"} and Proposition [\[supertop\]](#supertop){reference-type="ref" reference="supertop"}, which says, among other things, that the core spaces are exactly the locally compact web spaces (cf. ). Combining Theorem [\[pospace\]](#pospace){reference-type="ref" reference="pospace"} with Theorem [\[topsemi\]](#topsemi){reference-type="ref" reference="topsemi"}, we arrive at Of course, most elegant results are available for compact pospaces. From or, we learn the following facts: every compact pospace is - mc-ordered and dually mc-ordered, - monotone normal, in particular upper and lower regular, - strongly convex. For any closed subset $C$ of a compact pospace $T$, the sets ${\uparrow\!C}$ and ${\downarrow\!C}$ are closed, but $T$ need neither be ${\uparrow}$-stable nor locally filtered: see Examples [\[Ex33\]](#Ex33){reference-type="ref" reference="Ex33"} and [\[Ex53\]](#Ex53){reference-type="ref" reference="Ex53"}. A semilattice is said to be complete if all directed subsets have suprema and all nonempty subsets have infima (this together with the existence of a top element defines a complete lattice). Recall from the famous\ The Lawson spaces of complete continuous semilattices are exactly the compact T$_2$ s-topological semilattices. This is now a rather easy consequence of the previous facts, the Ultrafilter Theorem (giving compactness of $\Lambda L$ ) and the following "non-complete" version: If $P$ is a continuous domain then, by Theorems [\[fans\]](#fans){reference-type="ref" reference="fans"} and [\[pospace\]](#pospace){reference-type="ref" reference="pospace"}, the Lawson space $\Lambda P$ is a fan space, hence a locally filtered ${\uparrow}$-stable pospace. Conversely, if $T = (X,\leq,{\cal T})$ is any locally filtered, ${\uparrow}$-stable compact pospace then $T$ is upper regular, mc-ordered and dually mc-ordered; by Lemma [\[cstable\]](#cstable){reference-type="ref" reference="cstable"}, it is d-stable, and its upper space is a core space and a d-space (because $T$ is mc-ordered); hence, by Corollary [\[contdom\]](#contdom){reference-type="ref" reference="contdom"}, it is the Scott space of the continuous domain $P =(X,\leq)$. It follows that the Lawson topology $\lambda P = \sigma P \vee \upsilon \widetilde{P}$ is contained in ${\cal T} = {\cal T}^{\leq}\vee {\cal T}^{\geq}$ (as $T$ is strongly convex). Now, $\lambda P$ is T$_2$ and ${\cal T}$ is compact, so both topologies must coincide. In, one finds a whole collection of various equivalent characterizations of continuous domains that are compact in their Lawson topology. For a detailed study of joint and separate continuity of operations in posets and lattices, see. # Domain bases and core bases Following, we mean by a basis of a domain $P$ a subset $B$ such that for each $y\in P$, the set $\{ b \in\! B: b \ll y\}$ is directed with join $y$. The pair $(P,B)$ is then referred to as a based domain. Note that $P$ is continuous iff it has at least one basis. By a core basis for a space $(X,{\cal S})$, we mean a subset $B$ of $X$ such that for all $U\in {\cal S}$ and all $y\in U$, the set $R_{{\cal S}}y$ meets $B\cap U$; i.e., $y\in int_{{\cal S}}({\uparrow\!x}) \subseteq {\uparrow\!x} \subseteq U$ for some $x\in B$; in other words, all points have neighborhood bases formed by cores of elements of $B$. (To avoid ambiguities, we use the word basis for subsets of $X$ and the word base for subsets of the power set of $X$.) Thus, a space is a core space iff it has a core basis. By a core based space we mean a pair consisting of a (core) space and a core basis of it. The following extension of Corollary [\[contdom\]](#contdom){reference-type="ref" reference="contdom"} is straightforward: That, for any basis $B$ of a domain, the pair $(B,\ll\!\!\|_B)$ is a C-ordered set follows easily from the interpolation property of $\ll$ (cf. Lemma [\[conti\]](#conti){reference-type="ref" reference="conti"}). Conversely, any C-ordered set $(X,R )$ is isomorphic to $({\cal B}_{R}, \ll\!\!\|_{{\cal B}_{R}})$, where ${\cal B}_{R} = \{ R x: x\in X\}$ is a basis of the continuous domain ${\cal I}_{R}$ of rounded ideals (see Theorem [\[Cspace\]](#Cspace){reference-type="ref" reference="Cspace"} (4)). On the pointfree side, we define a based supercontinuous lattice to be a pair consisting of a supercontinuous (i.e. completely distributive) lattice $L$ and a (coprime) basis of $L$, that is, a join-dense subset of coprime elements. We are ready for six different descriptions of C-ordered sets: On the object level, these equivalences easily follow from Theorem [\[Cspace\]](#Cspace){reference-type="ref" reference="Cspace"}, Corollary [\[contdom\]](#contdom){reference-type="ref" reference="contdom"}, Theorem [\[fans\]](#fans){reference-type="ref" reference="fans"}, Lemma [\[Cbasis\]](#Cbasis){reference-type="ref" reference="Cbasis"}, Proposition [\[Cequiv\]](#Cequiv){reference-type="ref" reference="Cequiv"} and Corollary [\[Cor\]](#Cor){reference-type="ref" reference="Cor"}. The verification of the claimed correspondences between the morphisms is left as an exercise. # Density and weight {#Cbases} We now are looking for characterizations of core bases in terms of the interior relation. Recall that the lower quasi-order $\leq_{R}$ of an arbitrary relation $R$ on a set $X$ is given by $x \leq_{R} y \,\Leftrightarrow\, R x \subseteq R y$. Now, we say a subset $B$ of $X$ is - $R$-dense if $x\,R\,y$ implies $x\,R\,b$ and $b\,R\,y$ for some $b\in B$, - $R$-cofinal if $x\,R\,y$ implies $x\!\leq_{R}\!b$ and $b\,R\,y$ for some $b\in B$. It is straightforward to see that $B$ is $R$-dense and $R$ is transitive iff $B$ is $R$-cofinal and $R$ is idempotent. Note that the strong patch topology ${\cal S}^{\alpha} = {\cal S} \vee \alpha (X,\geq)$ of a space $(X,{\cal S} )$ concides with the Skula topology generated by ${\cal S} \cup {\cal S}^c$. Hence, the sets $C\setminus D$ with $C,D\in {\cal S}^c$ form a base for ${\cal S}^{\alpha}$. (1)$\,\Leftrightarrow\,$(2): The above remark applies, as $R$ is idempotent by Theorem [\[Cspace\]](#Cspace){reference-type="ref" reference="Cspace"} (1). (1)$\,\Rightarrow\,$(3): $y\in U\in {\cal S}$ means $U = UR$ and $x\,R\,y$ for some $x\in U$. Choose $b\in B$ with $x\,R\, b \, R \, y$. Then, $b\in R y$ and, by transitivity, $b\in xR \subseteq x\!\leq_{R} \ = {\uparrow\!x} \subseteq U$. (3)$\,\Rightarrow\,$(4): If $C,D$ are ${\cal S}$-closed sets with $C \!\not\subseteq\! D$, pick an $x\in C\setminus D$ and find a $b\in B$ with $x\in int_{{\cal S}}({\uparrow\!b}) \subseteq {\uparrow\!b} \subseteq X\setminus D$. It follows that $b\in C$, since $b\leq x \in C = {\downarrow\!C}$. Thus, $b\in B\cap (C\setminus D)$. This proves density of $B$ in ${\cal S}^{\alpha} = {\cal S} \vee {\cal S}^c$. (4)$\,\Leftrightarrow\,$(5) holds for arbitrary spaces. Recall that ${\downarrow\!x}$ is the closure of the singleton $\{ x\}$ in $(X,{\cal S})$. Now, $\{ {\downarrow\!b}: b\in B\}$ is join-dense in ${\cal S}^c$ iff for $C,D\in {\cal S}^c$ with $C\not\subseteq D$ there is a $b\in B$ with ${\downarrow\!b}\subseteq C$ but not ${\downarrow\!b}\subseteq D$, i.e. $b\in C\setminus D$ (as $C$ and $D$ are lower sets). But the latter means that $B$ meets every nonempty open set in ${\cal S}^{\alpha} = {\cal S} \vee {\cal S}^c$. (4)$\,\Rightarrow\,$(1): Suppose $x\,R\, y$ and choose a $z$ with $x\,R\, z \, R\, y$. Then $xR \in {\cal O}_{R} = {\cal S}$ and therefore $z \in xR \cap {\downarrow\!z} \in {\cal S}^{\alpha}$. Hence, there is a $b \in B \cap xR \cap {\downarrow\!z}$, and it follows that $x \, R \, b \, R \, y$, since $R y$ is a lower set containing $z$ and so $b$. From now on, we assume the validity of the Axiom of Choice. Hence, each set $X$ has a cardinality, represented by the smallest ordinal number equipotent to $X$. The weight $w({\cal S})$ resp. density $d({\cal S})$ of a space or its topology ${\cal S}$ is the least possible cardinality of bases resp. dense subsets. The weight of a core space is at most the cardinality of any core basis $B$, since $B$ gives rise to a base $\{ bR = int_{{\cal S}}({\uparrow\!b}): b \in B\}$. Let ${\cal B}$ be an arbitrary base and $B$ a core basis for $(X,{\cal S})$. The Axiom of Choice gives a function picking an element $b_{\,U,V}$ from each nonempty set of the form \[B_{\,U,V} = \{ b \in B: U \subseteq {\uparrow\!b} \subseteq V \} \ \ (U,V\in {\cal B} ).\] Then $B_0 = \{ b_{\,U,V}: U,V\in {\cal B}, \ B_{\,U,V} \not = \emptyset\}$ is a subset of $B$ and still a core basis; in fact, $x\in V \in {\cal B}$ implies $x\in U \subseteq {\uparrow\!b} \subseteq V$ for suitable $b\in B$ and $U\in {\cal B}$, and it follows that $x \in U \subseteq {\uparrow\!b_{\,U,V}} \subseteq V$. If ${\cal S}$ is infinite then so is ${\cal B}$, and consequently $\|B_0\| \leq \|{\cal B}\|^2 = \|{\cal B}\|$. Thus, we get $\|B_0\| \leq w({\cal S})$, and the remark before yields equality. If ${\cal S}$ is finite then the cores form the least base $\{ {\uparrow\!x}: x\in X\}$, and choosing a set of representatives from these cores, one obtains a core basis of cardinality $w({\cal S})$. Generalizing the weight of topologies, one defines the weight $w(L)$ of a lattice $L$ as the least possible cardinality of join-dense subsets of $L$. For any relation $R$ on a set $X$, we define the $R$-cofinality, denoted by $c(X,R)$, to be the minimal cardinality of $R$-cofinal subsets of $X$. If $R$ is idempotent then $c(X,R)$ is also the $R$-density, the least cardinality of $R$-dense subsets. From Proposition [\[corebase\]](#corebase){reference-type="ref" reference="corebase"} and Lemma [\[cardbase\]](#cardbase){reference-type="ref" reference="cardbase"} we infer: Indeed, if $B$ is a core basis for $(X,{\cal S})$ then ${\cal B} = \{ bR \setminus {\uparrow\!F}: b\in B,\, F \!\subseteq\! B, \, F \,\mbox{finite}\}$ is a base for ${\cal S}^{\upsilon}$, and if $B$ is infinite, ${\cal B}$ and $B$ have the same cardinality. If $B$ is finite, the base $\{ {\uparrow\!b}: b \in B\}$ of ${\cal S}$ is equipotent to the base $\{ {\uparrow\!b}\cap {\downarrow\!b}: b \in B\}$ of ${\cal S}^{\upsilon} = {\cal S}^{\alpha}$. Since in ZCF any supercontinuous lattice is isomorphic to the topology of a core space, Theorem [\[w\]](#w){reference-type="ref" reference="w"} entails a fact that was shown choice-freely in:	{'timestamp': '2016-07-19T02:03:30', 'yymm': '1607', 'arxiv_id': '1607.04721', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04721'}
null	null
# Introduction The combinatorial structure treated in this paper is a $2 \rightarrow 1$ directed hypergraph defined as follows. For simplicity from this point on we will always refer to $2 \rightarrow 1$ directed hypergraphs as just graphs or sometimes as $(2 \rightarrow 1)$-graphs when needed to avoid confusion. This structure comes up as a particular instance of the model used to represent definite Horn formulas in the study of propositional logic and knowledge representation. Some combinatorial properties of this model have been recently studied by Langlois, Mubayi, Sloan, and Gy. Turán in and. In particular, they looked at the extremal numbers for a couple of different small graphs. Before we can discuss their results we will need the following definitions. These are often called Turán-type extremal problems after Paul Turán due to his important early results and conjectures concerning forbidden complete $r$-graphs. Turán problems for uniform hypergraphs make up a large and well-known area of research in combinatorics, and the questions are often surprisingly difficult. Extremal problems like this have also been considered for directed graphs and multigraphs (with bounded multiplicity) in and and for the more general directed multi-hypergraphs in. In, Brown and Harary determined the extremal numbers for several types of specific directed graphs. In, Brown, Erdős, and Simonovits determined the general structure of extremal sequences for every forbidden family of digraphs analogous to the Turán graphs for simple graphs. The model of directed hypergraphs studied in have $r$-uniform edges such that the vertices of each edge is given a linear ordering. However, there are many other ways that one could conceivably define a uniform directed hypergraph. The graph theoretic properties of a more general definition of a nonuniform directed hypergraph were studied by Gallo, Longo, Pallottino, and Nguyen in. There a directed hyperedge was defined to be some subset of vertices with a partition into head vertices and tail vertices. Recently in, this author tried to capture many of these possible definitions for "directed hypergraph\" into one umbrella class of relational structures called generalized directed hypergraphs. The structures in this class include the uniform and simple versions of undirected hypergraphs, the totally directed hypergraphs studied in, the directed hypergraphs studied in, and the $2 \rightarrow 1$ model studied here and in. In, they study the extremal numbers for two different graphs with two edges each. They refer to these two graphs as the 4-resolvent and the 3-resolvent configurations after their relevance in propositional logic. Here, we will denote these graphs as $R_4$ and $R_3$ respectively and define them formally as \[V(R_4) = \{a,b,c,d,e\} \text{ and } E(R_4) = \{ab \rightarrow c, cd \rightarrow e\}\] and \[V(R_3) = \{a,b,c,d\} \text{ and } E(R_3) = \{ab \rightarrow c, bc \rightarrow d\}.\] In the authors determined $\text{ex}(n,R_4)$ for sufficiently large $n$, and in they determined a sequence of numbers asymptotically equivalent to the sequence of numbers $\text{ex}(n,R_3)$ as $n$ increases to infinity. In these papers, the authors discuss a third graph with two edges which they call an Escher configuration because it calls to mind the Escher piece where two hands draw each other. This graph is on four vertices, $\{a,b,c,d\}$ and has edge set $\{ab \rightarrow c,cd \rightarrow b\}$. We will denote it by $E$. These three graphs along with the graph made up of two completely overlapping edges on the same 3-set actually turn out to be the only four nondegenerate graphs with exactly two edges. Their standard and oriented extremal numbers are shown in. An immediate consequence of a result shown in is that the extremal numbers for a graph $H$ are cubic in $n$ if and only if $H$ is not degenerate. In our model of directed hypergraph, there are nine different graphs with exactly two edges. Of these, five are degenerate. One of these is the graph with two independent edges, $V=\{a,b,c,d,e,f\}$ and $E=\{ab \rightarrow c, de \rightarrow f\}$. In this case the extremal numbers come directly from the known extremal number for two independent edges for undirected $3$-graphs. Therefore, the oriented extremal number is ${n-1 \choose 2}$ and the standard extremal number is $3{n-1 \choose 2}$. We will call the other four degenerate graphs with two edges $I_0$, $I_1$, $H_1$, and $H_2$ and define them as - $V(I_0) = \{a,b,c,d,x\} \text{ and } E(I_0) = \{ab \rightarrow x, cd \rightarrow x\}$ - $V(I_1) = \{a,b,c,d\} \text{ and } E(I_1) = \{ab \rightarrow c, ad \rightarrow c\}$ - $V(H_1) = \{a,b,c,d,x\} \text{ and } E(H_1) = \{ax \rightarrow b, cx \rightarrow d\}$ - $V(H_2) = \{a,b,c,d\} \text{ and } E(H_2) = \{ab \rightarrow c, ab \rightarrow d\}$ Here, the subscripts indicate the number of tail vertices common to both edges. The $I$ graphs also share a head vertex while the $H$ graphs do not. Some of the proofs that follow rely heavily on the concept of a link graph. For undirected $r$-graphs, the link graph of a vertex is the $(r-1)$-graph induced on the remaining vertices such that each $(r-1)$-set is an $(r-1)$-edge if and only if that set together with the specified vertex makes an $r$-edge in the original $r$-graph. In the directed hypergraph model here, there are a few ways we could define the link graph of a vertex. We will need the following two. The following notation will also be used when we want to count edges by tail sets. # Forbidden $I_0$ In this section $I_0$ denotes the forbidden graph where two edges intersect in exactly one vertex such that this vertex is a head for both edges. That is $V(I_0) = \{a,b,c,d,x\}$ and $E(I_0) = \{ab \rightarrow x, cd \rightarrow x\}$ (see Figure [\[A\]](#A){reference-type="ref" reference="A"}). In this section we will prove the following result on the oriented extremal numbers of $I_0$. The proof for this is rather long. However, in the standard version of the problem where each triple of vertices may hold up to three different directed edges, the problem is much simpler so we will begin there. The oriented version of this problem is less straight forward, but determining $\text{ex}_o(n,I_0)$ also begins with the observation that every tail link graph of an $I_0$-free graph will either be a triangle, a star, or empty. Broadly speaking, as $n$ gets large, it would make more sense for most, if not all, tail link graphs to be stars in order to fit as many edges into an $I_0$-free graph. This motivates the following auxiliary structure. ## Gates Let $H$ be some $I_0$-free graph. For each $x \in V(H)$ for which $T_x$ is a star (with at least one edge), let $g(x)$ denote the common vertex for the edges of $T_x$. We will refer to this vertex as the gatekeeper of $x$ (in that it is the gatekeeper that any other vertex must pair with in order to "access\" $x$). In the case where $T_x$ contains only a single edge we may choose either of its vertices to serve as the gatekeeper. In this way, we have constructed a partial function, $g: V(H) \nrightarrow V(H)$. Next, construct a directed $2$-graph $G$ on the vertex set $V(H)$ based on this partial function: \[y \rightarrow x \in E(G) \iff y = g(x).\] We'll call this digraph the gate of $H$ (or more properly, $G$ is the gate of $H$ under $g$ since $g$ isn't necessarily unique). The edge structure of any gate $G$ is not difficult to determine. Since $g$ is a partial function, then each vertex has in-degree at most one in $G$. Therefore, the structure of any connected component of $G$ can be described as a directed cycle on $k$ vertices, $C_k$, for $1 \leq k$ (where $k=1$ implies a single vertex) unioned with $k$ disjoint directed trees, each with its root vertex on this cycle (see Figure [\[Ckplus\]](#Ckplus){reference-type="ref" reference="Ckplus"}). We will refer to this kind of general structure as a $k$-cycle with branches. Let \[\mathcal{C} = \bigcup_{k=1}^{n} \mathcal{C}_k\] be the set of maximal connected components of a gate of $H$ where, for each $k$, $\mathcal{C}_k$ is the set of maximal connected components that are $k$-cycles with branches. Note that \[\|E(H)\| = \sum_{x \in V(H)} \|T_x\| = \sum_{C \in \mathcal{C}} \left( \sum_{x \in V(C)} \|T_x\| \right) = \sum_{k=1}^n \left(\sum_{C \in \mathcal{C}_k} \left( \sum_{x \in V(C)} \|T_x\| \right)\right).\] The next section determines for each $k$ an upper bound on \[\sum_{x \in V(C)} \|T_x\|\] as a function of the number of vertices, $\|V(C)\|$, for any $C \in \mathcal{C}_k$. ## Bounding $\sum_{x \in V(C)} \|T_x\|$ for any connected component $C$ of the gate Loosely speaking, each gatekeeper edge of a connected component $C$ represents at most $n-2$ edges of $H$. We will arrive at an upper bound on the sum $\sum_{x \in V(C)} \|T_x\|$ by adding this maximum for each edge of $C$, and then subtracting the number of triples of vertices that such a count has included more than once. This will happen for any triple of vertices which contain two or three gatekeeper edges. We make this observation formal in the following definition and lemma. Now comes the main counting lemma. This shows that the best we can hope for in terms of the average number of edges per vertex over any connected component of the gate is $n-3 + \frac{1}{3}$, and this could be attained only in the case where the component is a directed triangle with no branches. Otherwise, the average number of edges of a component is at most $n-3$, and this is attainable only if the component is a directed triangle with a single directed path coming off of one of its vertices or a directed $k$-cycle with no branches for some $k \geq 4$. This is enough for us to establish the upper bound for $\text{ex}_o(n,I_0)$ and to characterize the necessary structure of the gate for any graph attaining this upper bound. ## Upper Bound on $\text{ex}_o(n,I_0)$ Let $H$ be an $I_0$-free graph on $n \geq 9$ vertices. Let $G$ be a gate of $H$. Let $\mathcal{C}$ be the set of maximal connected components of $G$ and break $\mathcal{C}$ into three disjoint subsets based on the maximum average number of edges attainable for the components in each. That is, let \[\mathcal{C} = \mathcal{D}_1 \cup \mathcal{D}_2 \cup \mathcal{D}_3\] where $\mathcal{D}_1$ are all components with maximum average number of edges per vertex strictly less than $n-3$: those components that are either acyclic, contain a $C_2$, contain a $C_3$ with nonempty branches that are more than just a single path, or contain a $C_k$ for $k \geq 4$ with some nonempty branch; $\mathcal{D}_2$ is the set of all components with maximum number of edges per vertex of $n-3$: those that contain a directed $C_3$ and exactly one directed path or those that are a directed $k$-cycles for any $k \geq 4$ and no branches; and $\mathcal{D}_3$ is the set of components with a maximum average greater than $n-3$: the directed triangles. For each $i$ let $d_i$ be the total number of vertices contained in the components of $\mathcal{D}_i$. Then \[\|E(H)\| \leq d_3 \left(n-3 + \frac{1}{3} \right) + (n-d_3)(n-3)\] with equality possible only if $d_1=0$. Then this is enough to prove the following. The next two lemmas give the maximum number when $n \equiv 1,2 \text{ mod } 3$. There is only slightly more to consider in these cases. ## Lower bound constructions The structure of the gates necessary to attain the maximum number of edges for a $I_0$-free graph determined in the previous section are also sufficient. Of these gates, none of them have acyclic components. Therefore, any graph that produces one of these gates has only vertices with stars for tail link graphs. This immediately implies that there is no $I_0$ in any graph that has such a gate. Moreover, if $H$ is a graph with a gate $G$ that is one of these configurations, then \[E(H) \subseteq \bigcup_{C \in \mathcal{C}} P(C)\] where $\mathcal{C}$ is the set of maximal connected components of $G$. All that is left to do in order to construct an extremal example is to pick which edges of each $P(C)$ to delete in order to eliminate triples of vertices with more than one edge. This establishes the main result of this section. # Forbidden $I_1$ In this section $I_1$ denotes the forbidden graph where two edges intersect in exactly two vertices such that one vertex is a head for both edges and the other is a tail for each edge. That is $V(I_1) = \{a,b,c,d\}$ and $E(I_1) = \{ab \rightarrow c, ad \rightarrow c\}$ (see Figure [\[F\]](#F){reference-type="ref" reference="F"}). # Forbidden $H_1$ In this section $H_1$ denotes the forbidden graph where two edges intersect in exactly one vertex such that it is in the tail for each edge. That is $V(H_1) = \{a,b,c,d,x\}$ and $E(H_1) = \{ax \rightarrow b, cx \rightarrow d\}$ (see Figure [\[B\]](#B){reference-type="ref" reference="B"}). First we will show the following result for the oriented version of the problem. We will use this result to solve the standard version of the problem and get the following. First, note that the proof of Theorem [\[exB\]](#exB){reference-type="ref" reference="exB"} is straightforward when $n$ is even. To get a lower bound construction we can take a maximum matching of the $n$ vertices and use each pair of this matching as the tail set to point at all $n-2$ other vertices. That is, let $H$ be the graph with vertex set, \[V(H) = \{x_1,\ldots,x_{\frac{n}{2}}, y_1, \ldots, y_{\frac{n}{2}}\}\] and edge set, \[E(H) = \bigcup_{i=1}^{\frac{n}{2}} \left\{x_iy_i \rightarrow z: z \in V(H) \setminus \{x_i,y_i\}\right\}.\] To show that this is also an upper bound, let $H$ be an $H_1$-free oriented graph on $n$ vertices. Then for any $x \in V(H)$, the directed link graph $D_x$ cannot have two independent edges (see Figure [\[Bdirlink\]](#Bdirlink){reference-type="ref" reference="Bdirlink"}). Therefore, $D_x$ is either empty, a triangle, or a star with at most $n-2$ edges. Since $n \geq 5$, then $\|D_x\| \leq n-2$ for each $x$. So \[\|E(H)\| = \frac{1}{2} \sum_{x \in V(H)} \|D_x\| \leq \frac{1}{2}n(n-2).\] So we are finished for even $n$. However, this proof falls apart when $n$ is odd. We will need a different strategy. ## Counting edges by possible tail pairs The basis of our strategy in getting an upper bound on $\text{ex}_o(n,H_1)$ is to count the edges of an $H_1$-free graph $H$ by its tail sets. That is, \[\|E(H)\| = \sum_{\{x,y\} \in {V(H) \choose 2}} t(x,y)\] It is simple but important to note that if $H$ is $H_1$-free, then any two pairs of vertices that each point to two or more other vertices must necessarily be disjoint. Therefore, if we assume that $H$ is $H_1$-free on $n$ vertices, then we can split its vertices up into $k$ disjoint pairs that each serve as tail sets to at least two edges of $H$ plus a set of $n-2k$ vertices that belong to no such pair. That is, \[V(H) = \{x_1,y_1,\ldots,x_k,y_k\} \cup R\] so that $t(x_i,y_i) \geq 2$ for $i=1,\ldots,k$ and $t(w,v) \leq 1$ for all other vertex pairs, $\{w,v\}$ (see Figure [\[twopluspairs\]](#twopluspairs){reference-type="ref" reference="twopluspairs"}). We now have two cases to consider. Either there are no such pairs ($k=0$) or there is at least one ($k \geq 1$). ## No pair points to more than one vertex ($k=0$) Assume that $k=0$. Then $t(x,y) \leq 1$ for every pair $\{x,y\} \in {V(H) \choose 2}$. If $\|D_x\| \leq n-3$ for all $x \in V(H)$, then \[\|E(H)\| = \frac{1}{2} \sum_{x \in V(H)} \|D_x\| \leq \frac{1}{2}n(n-3) < \frac{1}{2}(n-1)(n-2)\] and we are done. Otherwise, there exists some vertex $x$ that belongs to $n-2$ tail sets. Therefore, $D_x$ is a star of directed edges with some focus $y$. Either $t(x,y)=0$ or $t(x,y)=1$. If $t(x,y) = 0$, then all of the $n-2$ directed edges of $D_x$ must point to $y$ (see Figure [\[CaseApic\]](#CaseApic){reference-type="ref" reference="CaseApic"}). Such a configuration in $H$ limits the number of edges to ${n-1 \choose 2}$ as proven in Lemma [\[CaseA\]](#CaseA){reference-type="ref" reference="CaseA"}. On the other hand, if $t(x,y) = 1$, then $xy \rightarrow z \in E(H)$ for some vertex $z$, and $xv \rightarrow y$ for all other vertices $v \neq x,y,z$. Such a configuration in $H$ will limit the number of edges to ${n-1 \choose 2}$ as proven in Lemma [\[CaseB\]](#CaseB){reference-type="ref" reference="CaseB"}. Together these two lemmas take care of the cases where all pairs of vertices point to at most one vertex in $H$. ## At least one pair of vertices is the tail set to more than one edge of $H$ ($k > 0$) So let's return to our description of an $H_1$-free oriented graph as being made up of $k \geq 1$ vertex pairs that each serve as tails to strictly more than one edge plus a set $R$ of the remaining $n-2k$ vertices, \[V(H) = \{x_1,y_1,\ldots,x_k,y_k\} \cup R\] (see Figure [\[twopluspairs\]](#twopluspairs){reference-type="ref" reference="twopluspairs"}). For each pair $\{x_i,y_i\}$ we want to prove the following upper bound, \[t(x_i,y_i) + \sum_{v \neq x_i,y_i} \left(t(x_i,v)+t(y_i,v)\right) \leq n-2.\] That is, the total number of edges that include either $x_i$ or $y_i$ or both in the tail set is at most $n-2$. Now, \[\|E(H)\| = \sum_{\{x,y\} \in {V(H) \choose 2}} t(x,y) \leq \sum_{i=1}^k \left(t(x_i,y_i) + \sum_{v \neq x_i,y_i} \left(t(x_i,v)+t(y_i,v)\right) \right) + \sum_{\{x,y\} \in {R \choose 2}} t(x,y).\] Note that each pair of vertices in $R$ act as a tail set at most once so \[\sum_{\{x,y\} \in {R \choose 2}} t(x,y) \leq {n-2k \choose 2}.\] Therefore, proving the upper bound for each $\{x_i,y_i\}$ pair would imply that \[\|E(H)\| \leq k(n-2) + {n-2k \choose 2}.\] Since \[k(n-2) + {n-2k \choose 2} = 2k^2-(n+1)k + {n \choose 2}\] is a quadratic polynomial with positive leading coefficient in terms of $k$, then it is maximized at the endpoints. Here, that means at $k=1$ and at $k= \left\lfloor\frac{n}{2}\right\rfloor$. When $n$ is odd, both of these values for $k$ give the upper bound, \[\|E(H)\| \leq {n-1 \choose 2}.\] Only when $n$ is even can we do better and get \[\|E(H)\| \leq \frac{n(n-2)}{2}\] in the case where $k=\frac{n}{2}$. In either case this give an upper bound of \[\|E(H)\| \leq \left\lfloor \frac{n}{2} \right\rfloor (n-2).\] So we need only prove that, in general, \[t(x_i,y_i) + \sum_{v \neq x_i,y_i} \left(t(x_i,v)+t(y_i,v)\right) \leq n-2.\] This is straightforward to show if $t(x_i,y_i) \geq 3$. However, when $t(x_i,y_i)=2$ there is a case where it fails to hold. This is taken care of in the following lemma. ## First main result, $\text{ex}_o(n,H_1) = \left\lfloor \frac{n}{2} \right\rfloor (n-2)$. Now we can proceed with establishing the upper bound under the assumption that the configuration presented in Lemma [\[CaseC\]](#CaseC){reference-type="ref" reference="CaseC"} does not occur in our directed hypergraph. As we've seen, all that's necessary to show is that \[t(x_i,y_i) + \sum_{v \neq x_i,y_i} \left(t(x_i,v)+t(y_i,v)\right) \leq n-2\] for any pair of vertices $\{x_i,y_i\}$ that serves as the tail set to at least two edges. So let $\{x,y\}$ be such a pair, and divide the rest of the vertices of $H$ into two groups, those that are a head vertex to some edge with $xy$ as the tail and those that are not. That is, \[V(H) \setminus \{x,y\} = \{h_1,\ldots,h_m\} \cup \{n_1,\ldots,n_t\}\] where for each $i=1,\ldots,m$, there exists an edge, $xy \rightarrow h_i \in E(H)$ and for each $j=1,\ldots,t$, $xy \rightarrow n_j \not \in E(H)$ (note that $t(x,y)=m$ and that $m+t=n-2$). Now, consider an edge that contains either $x$ or $y$ in the tail but not both. Then the other tail vertex is either some $h_i$ or some $n_j$. In the case of $n_j$, this edge must either be of the form $xn_j \rightarrow y$ or $yn_j \rightarrow x$ to avoid a copy of $H_1$ with both $xy \rightarrow h_1$ and $xy \rightarrow h_2$. Moreover, since $H$ is oriented, there can be at most one. Hence, \[\sum_{j=1}^t \left(t(x,n_j)+t(y,n_j)\right) \leq t.\] Now consider a tail set that includes either $x$ or $y$ and some $h_i$. Without loss of generality, assume that $xh_1$ is the tail to some edge. Since $t(x,y) \geq 2$, then there is some other vertex $h_2$ such that $xy \rightarrow h_2 \in E(H)$. In order to avoid a copy of $H_1$ with this edge, $xh_1$ must either point to $y$ or to $h_2$. However, $xh_1 \rightarrow y \not \in E(H)$ since this would give the triple $\{x,y,h_1\}$ more than one edge. Therefore, $xh_1 \rightarrow h_2$ is the only option. However, if $t(x,y) \geq 3$, then this will create a copy of $H_1$ with $xy \rightarrow h_3$. So $xh_i$ and $yh_i$ cannot be tails to any edge. So \[\sum_{i=1}^m \left(t(x,h_i)+t(y,h_i)\right) = 0.\] Therefore, \[\begin{aligned} &t(x,y) + \sum_{v \neq x,y} \left(t(x,v)+t(y,v)\right)\\ &= m + \sum_{j=1}^t \left(t(x,n_j)+t(y,n_j)\right) + \sum_{i=1}^m \left(t(x,h_i)+t(y,h_i)\right)\\ &\leq m+t\\ &=n-2 \end{aligned}\] when $t(x,y) \geq 3$. The only other possibility is that $t(x,y)=2$. So suppose this is the case and that the head vertices to $xy$ are $a$ and $b$. Without loss of generality, assume that $yb \rightarrow a \in E(H)$. Note that this precludes any edges of the form $yn_j \rightarrow x$. Similarly, if we added the edge $xa \rightarrow b$ or the edge $xb \rightarrow a$, then we could not add any edges of the form $xn_j \rightarrow y$ and so \[\sum_{j=1}^t \left(t(x,n_j)+t(y,n_j)\right) = 0.\] Moreover, $ya \rightarrow b$ would lead to more than one edge on the triple $\{y,a,b\}$. So \[\sum_{i=1}^m \left(t(x,h_i)+t(y,h_i)\right) = 2\] and total we would have, \[t(x,y) + \sum_{v \neq x,y} \left(t(x,v)+t(y,v)\right) = 4 \leq n-2.\] On the other hand, if $xa$ and $xb$ are not tails to any edge, then the only way we could get a sum more than $n-2$ is if $xn_j \rightarrow y \in E(H)$ for all $j=1,\ldots, n-4$. But this is exactly the configuration described in Lemma [\[CaseC\]](#CaseC){reference-type="ref" reference="CaseC"} which we have excluded. Therefore, \[t(x,y) + \sum_{v \neq x,y} \left(t(x,v)+t(y,v)\right) \leq n-2\] for any such pair, and this is enough to establish that \[\text{ex}_o(n,H_1) \leq \left\lfloor \frac{n}{2} \right\rfloor (n-2).\] Conversely, we have already considered an extremal construction in the case where $n$ is even, and this same construction will work when $n$ is odd. That is, take a maximum matching of the vertices (leaving one out) and then use each matched pair as the tail set for all $n-2$ possible edges. Another construction that works for odd $n$ that is not extremal for even $n$ is to designate one vertex as the only head vertex and then make all ${n-1 \choose 2}$ pairs of the rest of the vertices tail sets. Therefore, \[\text{ex}_o(n,H_1) = \left\lfloor \frac{n}{2} \right\rfloor (n-2).\] Also, note that the only way that any construction could have more than ${n-1 \choose 2}$ edges is if $n$ is even and the vertices are partitioned into $\frac{n}{2}$ pairs such that each points to at least two other vertices. This fact comes directly from the requirement that $k=\frac{n}{2}$ in the optimization of \[k(n-2) + {n-2k \choose 2}\] in order for the expression to be more than ${n-1 \choose 2}$. ## Intersections of multiedge triples in the standard version Now, let $H$ be an $H_1$-free graph on $n$ vertices under the standard version of the problem so that any triple of vertices can now have up to all three possible directed edges. If we let $t_H$ be the number of triples of vertices of $H$ that hold at least one edge, and we let $m_H$ be the number of triples that hold at least two, then we have the following simple observation: \[\|E(H)\| \leq t_H + 2m_H.\] We start our path towards an upper bound on $\|E(H)\|$ by finding an upper bound on the number of multiedge triples, $m_H$. We will need to prove some facts about the multiedge triples of $H$. First, any triple which holds two edges of $H$ might as well hold three. Next, we want to show that no two multiedge triples can intersect in exactly one vertex. Therefore, we can use an upper bound on the number of undirected 3-uniform hyperedges such that no two intersect in exactly one vertex as an upper bound on the number of multiedge triples. Moreover, the extremal examples are easy to describe which will be important for finding the upper bound for $\text{ex}(n,H_1)$ as well as for establishing the uniqueness of the lower bound construction. In general, the only way to actually have an $H_1$-free graph with $n$ multiedge triples is if the multiedge triples form an undirected 3-uniform hypergraph of $\frac{n}{4}$ disjoint $K_4^{(3)}$ blocks when $n \equiv 0 \text{ mod } 4$. In this case there can be no additional directed edges in $H$ since such an edge would either intersect one of these $K_4^{(3)}$s in one tail vertex which would create a copy of $H_1$ since this means it intersects three of the multiedge triples in exactly one tail vertex (we may assume that each multiedge has all three edges per Lemma [\[more\]](#more){reference-type="ref" reference="more"}) or it would intersect one of the $K_4^{(3)}$s in two tail vertices which means that it intersects two of the multiedge triples in exactly one tail vertex (see Figure [\[blocks\]](#blocks){reference-type="ref" reference="blocks"}). So in this case, the number of total edges would be bound by \[3n < {n+1 \choose 2}-3\] for all $n \geq 7$. Next, the only ways to have $n-1$ multiedge triples is to either have $\frac{n-1}{4}$ disjoint $K_4^{(3)}$ blocks when $n \equiv 1 \text{ mod } 4$ or to have $\frac{n}{4}-1$ disjoint $K_4^{(3)}$ blocks with one $K_4^{-}$ when $n \equiv 0 \text{ mod } 4$. In the first case any additional edge must have at least one and perhaps two of its tail vertices in a single $K_4^{(3)}$ block of multiedge triples which we have already seen will create a copy of $H_1$. So there are at most \[3(n-1) < 3n < {n+1 \choose 2}-3\] total edges in this case. In the second case, any additional edge that has no tail vertices in a $K_4^{(3)}$ block must have both tail vertices in the $K_4^{-}$. If the head to such an edge were outside of the $K_4^{-}$, then the edge must intersect one of the three multiedge triples of the block in exactly one tail vertex since there are two triples that it intersects in one tail vertex each, one of which must be a multiedge triple. On the other hand, it could have its head vertex inside the $K_4^{-}$. In this case, the additional edge must lie on the triple without multiple edges. This is the only edge that can be added so there are at most \[3(n-1)+1 <3n <{n+1 \choose 2}-3\] total edges in this case. ## An $H_1$-free graph with $n-2$ multiedge triples Now, the only ways to have exactly $n-2$ multiedge triples is either to have $\frac{n}{4}-2$ of the $K_4^{(3)}$ blocks plus two $K_4^{-}$ blocks of multiedge triples when $n \equiv 0 \text{ mod } 4$ or to have $k$ of the $K_4^{(3)}$ blocks of multiedge triples plus a sunflower with $n-4k-2$ petals. The first case is suboptimal for the same reasons already considered. So let's consider the second case. First, assume that $k=0$ and that we have $n-2$ multiedge triples that make a sunflower (see Figure [\[mexconst\]](#mexconst){reference-type="ref" reference="mexconst"}). How many edges can we add? This structure already has all possible edges with 2 vertices in the core (or so we may assume by Lemma [\[more\]](#more){reference-type="ref" reference="more"}). On the other hand, if an additional edge has no vertices in the core, then it would intersect two multiedge triples in one tail vertex each which would create a copy of $H_1$. Therefore, any additional edge must include exactly one vertex from the core. If this vertex is in the tail set to the additional edge and the sunflower has at least three petals, then the additional edge intersects in exactly one tail vertex one of the multiedge triples of the sunflower, a contradiction. Since we assume that $n \geq 6$, then the sunflower has at least three petals. Hence, any additional edge must intersect the core in only its head vertex. If any two additional edges have different core vertices as the head, then either the tails sets are the same or completely disjoint to avoid a copy of $H_1$. Hence, pairs of petal vertices that point to both core vertices must be independent of all other tails sets. And all other petal vertices fall into disjoint sets as to whether they are in additional edges that point to the first core vertex or the second. The number of additional edges will be maximized if every pair of petal vertices point to the same core vertex. Moreover, this will give a total of \[3(n-2) + {n-2 \choose 2} = {n+1 \choose 2}-3\] edges. We will soon see that this is the best that we can do and that this construction, where the multiedge triples make a sunflower with $n-2$ petals with ${n-2 \choose 2}$ additional edges pointing from pairs of petal vertices to a single core vertex, is unique up to isomorphism. First we will need to see that $k=0$ is the number of $K_4^{(3)}$ multiedge triple blocks that optimizes the total number of edges. So suppose there are $k$ such blocks and that the other $n-4k$ vertices are in a sunflower. Then from prior considerations we know that any additional edge must have both tail vertices in this sunflower. If one of these tail vertices coincides with a petal vertex of the sunflower, then there will be a copy of $H_1$. Therefore, the tail vertices must coincide with the core and the only possibility for such an edge is to point out to a vertex in one of the $k$ blocks. Therefore, there are at most \[3(4k) + 3(n-4k-2) + {n-4k-2 \choose 2} + 4k\] edges in such a construction. Since this expression is quadratic in $k$ with positive leading coefficient, then it must maximize at the endpoints, $k=0$ or $k=\frac{n}{4}$, and we already know that $k=\frac{n}{4}$ is suboptimal. Therefore, if there are exactly $n-2$ multiedge triples, then they must form a sunflower with a two-vertex core and from there the only way to maximize the total number of edges is to add every possible edge with tail set among the petal vertices all pointing to the same head vertex in the core. ## Fewer than $n-2$ multiedge triples Now suppose that $H$ has fewer than $n-2$ multiedge triples. If $t_H \leq {n-1 \choose 2}$, then \[\|E(H)\| \leq t_H + 2m_H < {n-1 \choose 2} + 2(n-2) = {n+1 \choose 2}-3.\] So we must assume that $t_H > {n-1 \choose 2}$. Also, if $m_H=0$, then we know that \[\|E(H)\| \leq \text{ex}_o(n,H_1) = \left\lfloor \frac{n}{2} \right\rfloor (n-2) < {n+1 \choose 2}-3.\] So assume that there is at least one multiedge triple, $\{x,y,z\}$. This triple has at least two edges. Assume without loss of generality that they are $xy \rightarrow z$ and $xz \rightarrow y$. Let $H'$ be an oriented graph arrived at by deleting edges from multiedge triples of $H$ until each triple has at most one edge and every triple that had at least one edge in $H$ still has at least one in $H'$. In other words, $H'$ is any subgraph of $H$ such that $t_{H'} = t_{H}$ and $m_{H'}=0$. Without loss of generality, assume that \[xy \rightarrow z \in E(H').\] Since $t_{H'} > {n-1 \choose 2}$, then $n$ must be even, and moreover, there is a matching on the vertices so that every matched pair $\{a,b\}$ points to at least two other vertices. That is, $t(a,b) \geq 2$. Now consider the directed link graphs of the vertices. As stated before, these are either triangles or stars with a common vertex. However, if two or more of these link digraphs have three or fewer edges each (for instance, if they are triangles), then there are fewer edges than we are assuming since \[\|E(H')\| = \frac{1}{2} \sum_{x \in V(H')} \|D_x\| \leq \frac{1}{2} \left(6 + (n-3)(n-2) \right) < {n-1 \choose 2}\] for all $n \geq 8$. We will show that it must be the case that here at least two directed link graphs are restricted to at most three directed edges each, contradicting our assumptions about the number of edges in $H$. First, note that $x \rightarrow z \in D_y$ and $y \rightarrow z \in D_x$. To avoid a contradiction, at least one of these two directed link graphs must have four or more edges. Without loss of generality, assume that it is $D_y$. Therefore, $D_y$ is a star and not a triangle. So the additional three directed edges in $D_y$ must either all be incident to $z$ or to $x$. If these directed edges are all incident to $z$, then $y$ and $z$ must be partners under the matching which means that $x$ has another partner $x'$ distinct from $y$ and $z$. Since $t(x,x') \geq 2$ in $H'$, then $x'$ must point to two vertices in $D_x$. Since $D_x$ already has $y \rightarrow z$ and no two edges may be independent in any directed link graph, then $x'$ must point to $y$ and to $z$, forming a triangle. Next, consider $D_{x'}$. We know that \[x \rightarrow y, x \rightarrow z \in D_{x'}.\] If there is an additional edge in $D_{x'}$ that does not complete this triangle then it is either of the form $x \rightarrow t$ or $t \rightarrow x$. If $x \rightarrow t \in D_{x'}$ then $x' \rightarrow t, y \rightarrow z \in D_x$, a contradiction. If $t \rightarrow x \in D_{x'}$, then $x' \rightarrow x \in D_t$. But since $t$ has its own matched vertex, then there exists a distinct $t'$ such that \[t' \rightarrow x, t' \rightarrow x' \in D_{t'}.\] So either $\|D_{x'}\| \leq 3$ or $\|D_{t'}\| \leq 3$. Either way, this gives us two directed link graphs that have at most three edges each. So $t_H > {n-1 \choose 2}$. Therefore, we must assume that the three additional edges in $D_y$ are incident to $x$ and that $y$ and $x$ are partners under the matching. So $z$ has some other partner under the matching $z'$ distinct from $x$ and $y$. Now, delete the edge $xy \rightarrow z$ from $H'$ and add $xz \rightarrow y$ to get a new directed hypergraph $H''$. It follows that $H''$ has no multiedge triples and is $H_1$-free since we still have a subgraph of $H$. In adding $xz \rightarrow y$ we have added $x \rightarrow y$ to $D_z$. Since $z'$ must point to two vertices in $D_z$, then this addition means that $D_z$ is a triangle under $H''$. Hence, $\|D_z\| = 2$ under $H'$. Now, the same argument as above applies to $D_{z'}$. The only way for $\|D_{z'}\| > 3$ would mean either $z \rightarrow a \in D_{z'}$ or $a \rightarrow z \in D_{z'}$ for some $a$ distinct from $x$, $y$, $z$, and $z'$. The first case would mean that two independent directed edges, $z' \rightarrow a$ and $x \rightarrow y$ are in $D_z$, a contradiction. The second case would mean that $z' \rightarrow z \in D_a$. Since $a$ has its own partner under the matching that must point to two vertices in $D_a$, then in this case, $D_a$ is a triangle. Therefore, $t_H > {n-1 \choose 2}$ and $m_H \geq 1$ cannot both be true in any $H_1$-free graph. This is enough to complete the result, \[\text{ex}(n,H_1) = {n+1 \choose 2}-3.\] This also exhausts the remaining cases in order to demonstrate that the extremal construction is unique. # Forbidden $H_2$ In this section $H_2$ denotes the forbidden graph where two edges intersect in exactly two vertices such that the set of intersection is the tail set to each edge. That is $V(H_2) = \{a,b,c,d\}$ and $E(H_2) = \{ab \rightarrow c, ab \rightarrow d\}$ (see Figure [\[D\]](#D){reference-type="ref" reference="D"}). # Conclusion The $2 \rightarrow 1$ version of directed hypergraph originally came to the author's attention as a way to model definite Horn clauses in propositional logic. Definite Horn clauses are more generally modeled by $r \rightarrow 1$ edges for any $r$. Therefore, it seems natural to ask about the extremal numbers for graphs with two $(r \rightarrow 1)$-edges. If we look at every $(r \rightarrow 1)$-graph with exactly two edges we see that these fall into four main types of graph. Let $i$ be the number of vertices that are in the tail set of both edges. Then let $I_r(i)$ denote the graph where both edges point to the same head vertex, let $H_r(i)$ denote the graph where the edges point to different head vertices neither of which are in the tail set of the other, let $R_r(i)$ denote the graph where the first edge points to a head vertex in the tail set of the second edge and the second edge points to a head not in the tail set of the first edge, and let $E_r(i)$ denote the graph where both edges point to heads in the tail sets of each other. This extends the notation used in this paper. The degenerate cases here would be $I_r(i)$ and $H_r(i)$. It would be interesting to find the extremal numbers for these graphs in general. To what extent do the current proofs extend to these graphs? For example, in the standard version of the problem it can easily be seen that \[\text{ex}(n,I_r(0)) = n{n-2 \choose r-1}\] using the same idea as we did for $I_0$. Will the other ideas generalize as well?	{'timestamp': '2016-07-19T02:09:39', 'yymm': '1607', 'arxiv_id': '1607.04932', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04932'}
null	null
# Introduction The study of hereditary properties of combinatorial structures is an important topic within the field of extremal combinatorics. Out of the many results in this line of research has emerged an pattern for how to prove approximate asymptotic enumeration and structure results. The aim of this paper is to provide a general framework in which to view these results and to formalize this pattern of proof. ## Background A nonempty class of graphs $\mathcal{P}$ is called a hereditary graph property if it is closed under isomorphism and induced subgraphs. Given a hereditary graph property $\mathcal{P}$, let $\mathcal{P}_n$ denote the set of elements of $\mathcal{P}$ with vertex set $[n]$. There has been extensive investigation into the properties of $\mathcal{P}_n$, where $\mathcal{P}$ is a hereditary property of graphs and $n$ is large, see for instance. The main questions addressed in these papers concern enumeration (finding an asymptotic formula for $\|\mathcal{P}_n\|$) and structure (understanding what properties elements of $\mathcal{P}_n$ have with high probability). Given a graph $H$, $\operatorname{Forb}(H)$ (respectively $\operatorname{Forb}_{ind}(H)$) is the class of finite graphs omitting $H$ as a non-induced (respectively induced) subgraph. For any graph $H$, both $\operatorname{Forb}(H)$ and $\operatorname{Forb}_{ind}(H)$ are hereditary graph properties. Therefore, work on hereditary graph properties can be seen as generalizing the many structure and enumeration results about graph properties of the form $\operatorname{Forb}(H)$ and $\operatorname{Forb}_{ind}(H)$, for instance those appearing in. From this perspective, the study of hereditary graph properties has been a central area of research in extremal combinatorics. There are many results which extend the investigation of hereditary graph properties to other combinatorial structures. Examples of this include for tournaments, for oriented graphs and posets, for $k$-uniform hypergraphs, and for colored $k$-uniform hypergraphs. The results in investigate asymptotic enumeration and structure results for specific classes of $H$-free hypergraphs, which are examples of hereditary properties of hypergraphs. Similarly, the results in concern specific examples of hereditary properties of digraphs. The results in for metric spaces are similar in flavor, although they have not been studied explicitly as instances of hereditary properties. Thus, extending the investigation of hereditary graph properties to other combinatorial structures has been an active area of research for many years. From this investigation, patterns have emerged for how to prove these kinds of results, along with a set of standard tools, including extremal results, stability[^1] theorems, regularity lemmas, supersaturation results, and the hypergraph containers theorem. In various combinations with extremal results, stability theorems, and supersaturation results, Szemerédi's regularity lemma has played a key role in proving many results in this area, especially those extending results for graphs to other settings. A sampling of these are for graphs, for oriented graphs, for hypergraphs, and for metric spaces. The hypergraph containers theorem, independently developed in, has been used in many recent papers in place of the regularity lemma. Examples of this include for graphs, for digraphs, and for metric spaces. In these papers, the commonalities in the proofs are especially clear. Given an extremal result, there is clear outline for how to prove an approximate enumeration theorem. If on top of this, one can characterize the extremal structures and prove a corresponding stability theorem, then there is a clear outline for how to prove an approximate structure theorem. The goal of this paper is to make these proof outlines formal using generalizations of tools, definitions, and theorems from these papers to the setting of structures in finite relational languages. ## Summary of Results Given a first-order language $\mathcal{L}$, we say a class $\mathcal{H}$ of $\mathcal{L}$-structures has the hereditary property if for all $A\in \mathcal{H}$, if $B$ is a model theoretic substructure of $A$, then $B\in \mathcal{H}$. This is the natural generalization of existing notions of hereditary properties of various combinatorial structures. Indeed, for appropriately chosen $\mathcal{L}$, almost all of the results cited so far are for hereditary $\mathcal{L}$-properties, including all hereditary properties of graphs, $k$-uniform hypergraphs, colored $k$-uniform hypergraphs, directed graphs, and posets, as well as the the metric spaces from. Thus hereditary $\mathcal{L}$-properties are the appropriate objects to study in order to generalize many of the results we are interested in. We now give a description of our results. The precise statements require extensive preliminaries and appear in Section [4](#mainresults){reference-type="ref" reference="mainresults"}. Fix a finite relational language $\mathcal{L}$ with maximum arity $r\geq 2$. Given a hereditary $\mathcal{L}$-property $\mathcal{H}$, we will define an invariant associated to $\mathcal{H}$, called the asymptotic density of $\mathcal{H}$, denoted by $\pi(\mathcal{H})$ (see Definition [\[pidef\]](#pidef){reference-type="ref" reference="pidef"}). Our first main result, Theorem [\[enumeration\]](#enumeration){reference-type="ref" reference="enumeration"}, gives an asymptotic enumeration of $\mathcal{H}_n$ in terms of $\pi(\mathcal{H})$, where $\mathcal{H}_n$ denotes the set of elements in $\mathcal{H}$ with domain $[n]$. The tools we use to prove this theorem include a general supersaturation theorem for $\mathcal{L}$-structures (Theorem [\[GENSUPERSAT\]](#GENSUPERSAT){reference-type="ref" reference="GENSUPERSAT"}) and a new adaptation of the hypergraph containers theorem to the setting of $\mathcal{L}$-structures (Theorems [\[version1\]](#version1){reference-type="ref" reference="version1"}). The proof of Theorem [\[GENSUPERSAT\]](#GENSUPERSAT){reference-type="ref" reference="GENSUPERSAT"} uses our hypergraph containers theorem for $\mathcal{L}$-structures (Theorem [\[version1\]](#version1){reference-type="ref" reference="version1"}) and a powerful generalization by Aroskar and Cummings of the triangle removal lemma (Theorem [\[triangleremoval2\]](#triangleremoval2){reference-type="ref" reference="triangleremoval2"}). Our proof strategies for these theorems draw on a series of enumeration results for combinatorial structures which employ the hypergraph containers theorem, namely those in. We will also define generalizations of extremal graphs (Definition [\[genexdef\]](#genexdef){reference-type="ref" reference="genexdef"}) and graph stability theorems (Definition [\[stabdef\]](#stabdef){reference-type="ref" reference="stabdef"}). We will prove that the existence of a stability theorem, along with an understanding of extremal structure, always yield an approximate structure theorem. This result, Theorem [\[stab\]](#stab){reference-type="ref" reference="stab"}, generalizes arguments appearing in many papers, including. The main tool used to prove Theorem [\[stab\]](#stab){reference-type="ref" reference="stab"} is a second adaptation of the hypergraph containers theorem to the setting of $\mathcal{L}$-structures, namely Theorem [\[COROLLARY2\]](#COROLLARY2){reference-type="ref" reference="COROLLARY2"}. Our adaptations of the hypergraph containers theorem, Theorems [\[version1\]](#version1){reference-type="ref" reference="version1"} and [\[COROLLARY2\]](#COROLLARY2){reference-type="ref" reference="COROLLARY2"}, rely on the original hypergraph containers theorem of, the general triangle removal lemma in, and the model theoretic tools developed in this paper. In the last main section of the paper, we consider how our results relate to theorems about discrete metric spaces. Given integers $r\geq 3$ and $n\geq 1$, let $M_r(n)$ be the set of metric spaces with distances all in $[r]$ and underlying set $[n]$. We will reprove structure and enumeration theorems from using the machinery of this paper, along with combinatorial ingredients from. We include this example because it demonstrates interesting behavior with regards to the existence of a stability theorem. In particular, we will prove that when $r$ is even, the hereditary property associated to $\bigcup_{n\in \mathbb{N}}M_r(n)$ has a stability theorem in the sense of Definition [\[stabdef\]](#stabdef){reference-type="ref" reference="stabdef"}, but when $r$ is odd, this is not the case. This formalizes an intuitive difference between the even and odd cases observed in. It is important to note that our results apply to languages with relations of arity larger than two, and to structures with non-symmetric relations. To illustrate this we have included appendices explaining how our results apply to examples in the settings of colored hypergraphs (Appendix [\[coloredhg\]](#coloredhg){reference-type="ref" reference="coloredhg"}), directed graphs (Appendix [\[dgsec\]](#dgsec){reference-type="ref" reference="dgsec"}), and triangle-free hypergraphs (Appendix [\[trifreesec\]](#trifreesec){reference-type="ref" reference="trifreesec"}). We now clarify what the results in this paper do and what they do not do. Our main theorem, Theorem [\[enumeration\]](#enumeration){reference-type="ref" reference="enumeration"}, gives an enumeration theorem for a hereditary $\mathcal{L}$-property in terms of its asymptotic density. However, determining the asymptotic density of a specific hereditary $\mathcal{L}$-property is often a hard combinatorial problem which this paper does not address. Similarly, while Theorem [\[stab\]](#stab){reference-type="ref" reference="stab"} shows that a stability theorem and an understanding of extremal structure implies an approximate structure theorem, proving a specific family $\mathcal{H}$ has a stability theorem and understanding its extremal structures are often difficult problems in practice. These difficulties are not addressed by the results in this paper. The role of this paper is to generalize how extremal results and stability theorems give rise to approximate structure and enumeration theorems. While our proofs are inspired by and modeled on those appearing in, our results are more than just straightforward generalizations of existing combinatorial theorems. We use new tools called $\mathcal{L}_{\mathcal{H}}$-templates (see Section [3](#tildeLstructures){reference-type="ref" reference="tildeLstructures"}) and an application of the hyergraph containers theorem to a hypergraph whose vertices and edges correspond to certain atomic diagrams (see Theorem [\[VERSION2\]](#VERSION2){reference-type="ref" reference="VERSION2"}). These technical tools and their appearances in our results are non-obvious and of independent interest from a model theoretic perspective. We also provide a simplified version of the generalized triangle removal lemma appearing in, by noting that a simpler notion of the distance between $\mathcal{L}$-structures may be used. ## Conclusion The work in this paper is significant from the perspective of combinatorics for three main reasons. First, problems in finite combinatorics are most often approached one by one, and techniques developed for specific structures often do not translate well into other contexts. While this style of approach is necessary to solve problems, it creates the impression that generalization is not possible. This work serves as an example that searching to generalize results and proofs within finite combinatorics can be highly fruitful. Second, this work will save researchers time by allowing them to avoid repeating arguments which now appear here in a general context. Third, we believe this paper gives the correct general framework in which to view these questions, which may aid in finding answers to open problems in the area. This work is also of significance from the model theoretic perspective due to the following connection to logical $0$-$1$ laws. Suppose $\mathcal{L}$ is a finite language, and for each $n$, $F(n)$ is a set of $\mathcal{L}$-structures with domain $[n]$. We say $F:=\bigcup_{n\in \mathbb{N}}F(n)$ has a $0$-$1$ law if for every first-order $\mathcal{L}$-sentence $\phi$, the limit \[\mu(\phi):=\lim_{n\rightarrow \infty}\frac{\|\{G\in F(n): G\models \phi\}\|}{\|F(n)\|}\] exists and is equal to $0$ or $1$. If $F$ has a $0$-$1$ law, then $T_{as}(F):=\{\phi: \mu(\phi)=1\}$ is a complete consistent first-order theory. There are many interesting model theoretic questions related to $0$-$1$ laws and almost sure theories. For instance, it is not well understood in general why some classes of finite structure have $0$-$1$ laws and why others do not. One source of known $0$-$1$ laws are asymptotic structure results from extremal combinatorics. For instance, fix $s \geq 3$ and suppose for each $n$, $F(n)$ is the set of all graphs with vertex set $[n]$ omitting the complete graph $K_s$ on $s$ vertices. In Kolaitis, Prömel and Rothschild show $F:=\bigcup_{n\in \mathbb{N}}F(n)$ has a $0$-$1$ law. Their proof relies crucially on first proving asymptotic structure and enumeration theorems for $F$. In particular, they show that if $S(n)$ is the set of $(s-1)$-partite graphs on $[n]$, then $S(n)\subseteq F(n)$ for all $n$ and \[\begin{aligned} \label{KPRfact} \lim_{n\rightarrow \infty}\frac{\|S(n)\|}{\|F(n)\|}=1. \end{aligned}\] They then prove a $0$-$1$ law for $S:=\bigcup_{n\in \mathbb{N}}S(n)$ which combines with ([\[KPRfact\]](#KPRfact){reference-type="ref" reference="KPRfact"}) to imply $F$ has a $0$-$1$ law. Other asymptotic structure results which imply $0$-$1$ laws include (for details on how these structure results imply $0$-$1$ laws, see ). In these papers,, the precise structure results (which are needed to prove the $0$-$1$ laws) are proven using approximate structure and enumeration theorems as stepping stones. This trend suggests that a systematic understanding of precise structure and enumeration will use some version of this "approximate version\" stepping stone. Therefore, understanding approximate structure and enumeration results from a model theoretic perspective is a necessary step in gaining a general understanding of precise structure and enumeration results, and consequently of the logical $0$-$1$ laws which rely on them. # Preliminaries {#prelims} Our goal here is to include enough preliminaries so that anyone with a rudimentary knowledge of first-order logic will be able to read this paper. Definitions we expect the reader to understand include: first-order languages, constant and relation symbols, formulas, variables, structures, substructures, satisfaction, and consistency. We refer the reader to for these definitions. ## Notation and Setup {#notation} In this subsection we fix some notational conventions and definitions. We will use the word "collection" to denote either a set or a class. Suppose $\ell \geq 1$ is an integer and $X$ is a set. Let $Perm(\ell)$ be the set of permutations of $[\ell]$. We let $\mathcal{P}(X)$ or $2^X$ denote the power set of $X$. Given a finite tuple $\bar{x}=(x_1,\ldots, x_{\ell})$ and $\mu\in Perm(\ell)$, let $\cup \bar{x} =\{x_1,\ldots, x_{\ell}\}$, $\|\bar{x}\|=\ell$, and $\mu(\bar{x})=(x_{\mu(1)},\ldots, x_{\mu(\ell)})$. An enumeration of $X$ is a tuple $\bar{x}=(x_1,\ldots, x_{\|X\|})$ such that $\cup \bar{x} =X$. Given $x\neq y\in X$, we will write $xy$ as shorthand for the set $\{x,y\}$. Set \[X^{\underline{\ell}}=\{(x_1,\ldots, x_{\ell}) \in X^{\ell}: x_i\neq x_j\text{ for each }i\neq j\}\quad \hbox{ and }\quad {X\choose \ell}=\{Y\subseteq X: \|Y\|=\ell\}.\] Suppose $\mathcal{L}$ is a finite relational first-order language. Let $r_{\mathcal{L}}$ denote the maximum arity of any relation symbol in $\mathcal{L}$. Given a formula $\phi$ and a tuple of variables $\bar{x}$, we write $\phi(\bar{x})$ to denote that the free variables in $\phi$ are all in the set $\cup \bar{x}$. Similarly, if $p$ is a set of formulas, we will write $p(\bar{x})$ if every formula in $p$ has free variables in the set $\cup \bar{x}$. We will sometimes abuse notation and write $\bar{x}$ instead of $\cup \bar{x}$ when it is clear from context what is meant. Suppose $M$ is an $\mathcal{L}$-structure. Then $dom(M)$ denotes the underlying set of $M$, and the size of $M$ is $\|dom(M)\|$. If $\mathcal{L'}\subseteq \mathcal{L}$, $M\upharpoonright_{\mathcal{L}'}$ is the $\mathcal{L}'$-structure with underlying set $dom(M)$ such that for all $\ell\geq 1$, if $\bar{a} \in dom(M)^{\ell}$ and $R$ is an $\ell$-ary relation symbol from $\mathcal{L}'$, then $M\upharpoonright_{\mathcal{L}'}\models R(\bar{a})$ if and only if $M\models R(\bar{a})$. We call $M\upharpoonright_{\mathcal{L}'}$ the reduct of $M$ to $\mathcal{L}'$. Given $X\subseteq dom(M)$, $M[X]$ is the $\mathcal{L}$-structure with domain $X$ such that for all $\ell\geq 1$, if $\bar{a}\in X^{\ell}$ and $R$ is an $\ell$-ary relation symbol from $\mathcal{L}$, then $M[X]\models R(\bar{a})$ if and only if $M\models R(\bar{a})$. We call $M[X]$ the $\mathcal{L}$-structure induced by $M$ on $X$. Given a tuple $\bar{a}\in dom(M)^{\ell}$, the quantifier-free type of $\bar{a}$ is \[qftp^M(\bar{a})=\{\phi(x_1,\ldots, x_{\ell}): \text{$\phi(x_1,\ldots, x_{\ell})$ is a quantifier-free $\mathcal{L}$-formula and $M\models \phi(\bar{a})\}$}.\] If $\bar{x}=(x_1,\ldots, x_{\ell})$ and $p(\bar{x})$ is a set of quantifier-free $\mathcal{L}$-formulas, then $p$ is called a quantifier-free $\ell$-type if there is some $\mathcal{L}$-structure $N$ and a tuple $\bar{a}\in dom(N)^{\ell}$ such that $N\models \phi(\bar{a})$ for all $\phi(\bar{x})\in p$. In this case we say $\bar{a}$ realizes $p$ in $N$. If there is some $\bar{a} \in dom(N)^{\ell}$ realizing $p$ in $N$, we say $p$ is realized in $N$. A quantifier-free $\ell$-type $p(\bar{x})$ is complete if for every quantifier-free formula $\phi(\bar{x})$, either $\phi(\bar{x})$ or $\neg \phi(\bar{x})$ is in $p(\bar{x})$. Note that any type of the form $qftp^M(\bar{a})$ is complete. All types and formulas we consider will be quantifier-free, so for the rest of the paper, any use of the words type and formula means quantifier-free type and quantifier-free formula. If $X$ and $Y$ are both $\mathcal{L}$-structures, let $X\subseteq_{\mathcal{L}} Y$ denote that $X$ is a $\mathcal{L}$-substructure of $Y$. Given an $\mathcal{L}$-structure $H$, we say that $M$ is $H$-free if there is no $A\subseteq_{\mathcal{L}} M$ such that $A\cong_{\mathcal{L}}H$. Suppose $\mathcal{H}$ is a collection of $\mathcal{L}$-structures. We say $M$ is $\mathcal{H}$-free if $M$ is $H$-free for all $H\in \mathcal{H}$. For each positive integer $n$, let $\mathcal{H}(n)$ denote the collection of all elements in $\mathcal{H}$ of size $\ell$, and let $\mathcal{H}_n$ denote the set of elements in $\mathcal{H}$ with domain $[n]$. $\mathcal{H}$ is trivial if there is $N$ such that $\mathcal{H}(n)=\emptyset$ for all $n\geq N$. Otherwise $\mathcal{H}$ is non-trivial. We now define a modified version of the traditional type space, which is appropriate for working with families of finite structures instead of with complete first-order theories. Given $\bar{x}=(x_1,\ldots, x_{\ell})$, an $\ell$-type $p(\bar{x})$ is proper if it contains the formulas $x_i\neq x_j$ for each $i\neq j$. We would like to emphasize some important differences between this and the usual type space. First, the elements of these type spaces are proper and contain only quantifier-free formulas. Second, these type spaces are defined relative to families of finite structures instead of complete first-order theories. It will at times be convenient to expand our languages to contain constant symbols naming elements of the structures under consideration. If $V$ is a set, let $C_V$ denote the set of constant symbols $\{c_v: v\in V\}$. Given $\bar{v}=(v_1,\ldots, v_{\ell}) \in V^{\ell}$, let $c_{\bar{v}}=(c_{v_1},\ldots, c_{v_{\ell}})$. Suppose $M$ is an $\mathcal{L}$-structure. The diagram of $M$, denoted $Diag(M)$, is the following set of sentences in the language $\mathcal{L}\cup C_{dom(M)}$. \[Diag(M)=\{\phi(c_{\bar{a}}): \phi(\bar{x}) \text{ is a quantifier-free $\mathcal{L}$-formula, } \cup \bar{a} \subseteq dom(M)\text{, and }M\models \phi(\bar{a})\}.\] If $A\subseteq dom(M)$, the diagram of $A$ in $M$ is the following set of sentences in the language $\mathcal{L}\cup C_{A}$. \[Diag^M(A)=\{\phi(c_{\bar{a}}): \phi(\bar{x}) \text{ is a quantifier-free $\mathcal{L}$-formula, } \cup \bar{a} \subseteq A\text{, and }M\models \phi(\bar{a})\}.\] Observe that if $A=\{a_1,\ldots, a_r\}\subseteq dom(M)$ and $p(\bar{x})\in S_r(\mathcal{L})$ is such that $p(\bar{x})=qftp^M(a_1,\ldots, a_r)$, then $Diag^M(A)=p(c_{a_1},\ldots, c_{a_r})$. Given a set of constants $C$, a collection of $\mathcal{L}$-structures $\mathcal{F}$, and $\ell\geq 1$, set \[S_{\ell}(C)=\{p(\bar{c}): p(\bar{x})\in S_{\ell}(\mathcal{L})\text{ and } \bar{c}\in C^{\underline{\ell}}\}\quad \hbox{ and }\quad S_{\ell}(C, \mathcal{F})=\{p(\bar{c}): p(\bar{x})\in S_{\ell}(\mathcal{F})\text{ and } \bar{c}\in C^{\underline{\ell}}\}.\] We would like to emphasize that if $p(\bar{c})\in S_{\ell}(C)$, then $\bar{c}\in C^{\underline{\ell}}$ is a tuple of $\ell$ distinct constants. Note that by this definition, if $A\in {dom(M)\choose \ell}$, then $Diag^M(A)\in S_{\ell}(C_{dom(M)})$. ## Hypergraph Containers Theorem. {#containerssec} In this section we state a version of the hypergraph containers theorem, which was independently developed by Balogh-Morris-Samotij in and by Saxton-Thomason in. The particular statement we use, Theorem [\[containers\]](#containers){reference-type="ref" reference="containers"} below, is a simplified version of Corollary 3.6 in. We will use Theorem [\[containers\]](#containers){reference-type="ref" reference="containers"} directly in Section [9](#VERSION2pf){reference-type="ref" reference="VERSION2pf"}. We also think it will be useful for the reader to compare it to the versions for $\mathcal{L}$-structures stated in Section [4](#mainresults){reference-type="ref" reference="mainresults"} (Theorem [\[version1\]](#version1){reference-type="ref" reference="version1"} and Corollary [\[COROLLARY2\]](#COROLLARY2){reference-type="ref" reference="COROLLARY2"}). We begin with some definitions. Recall that if $s\geq 2$ is an integer, an $s$-uniform hypergraph is a pair $(V,E)$ where $V$ is a set of vertices and $E\subseteq {V\choose s}$ is a set of edges. Suppose $H$ is an $s$-uniform hypergraph. Then $V(H)$ and $E(H)$ denote the vertex and edge sets of $G$ respectively. We set $v(H)=\|V(H)\|$ and $e(H)=\|E(H)\|$. Given $X\subseteq V(H)$, $H[X]$ is the hypergraph $(X,E\cap{V(H)\choose s})$. If $v(H)$ is finite, then the average degree of $H$ is $d=e(H)s/v(H)$. Unless otherwise stated, $n$ is always a positive integer. ## Distance between first-order structures In this section we define a notion of distance between finite first-order structures. The following is based on definitions in. We now state our definition for the distance between two finite first-order structures. It is a simplified version of the distance notion appearing in. We will discuss the relationship between the two notions in Section [7](#rphrem){reference-type="ref" reference="rphrem"}. Observe that in the notation of Definition [\[deltaclosedef1\]](#deltaclosedef1){reference-type="ref" reference="deltaclosedef1"}, $\textnormal{diff}(M,N)=\{A\in {V\choose r}: Diag^M(A)\neq Diag^N(A)\}$. ## Facts about hereditary properties Suppose $\mathcal{L}$ is a finite relational language. In this subsection we state some well known facts about hereditary $\mathcal{L}$-properties. First we recall that hereditary $\mathcal{L}$-properties are the same as families of structures with forbidden configurations. This fact will be used throughout the chapter. It is easy to check that for any collection $\mathcal{F}$ of finite $\mathcal{L}$-structures, $\operatorname{Forb}(\mathcal{F})$ is a hereditary $\mathcal{L}$-property. The converse to this statement is also true in the sense of Observation [\[HP\]](#HP){reference-type="ref" reference="HP"} below. This fact is standard, but we include a proof for completeness. A sentence $\phi$ is universal if it is of the form $\forall \bar{x} \psi(\bar{x})$ where $\psi(\bar{x})$ is quantifier-free. The following well known fact is another reason hereditary $\mathcal{L}$-properties are natural objects of study. The proof of this is straightforward using Observation [\[HP\]](#HP){reference-type="ref" reference="HP"}, the fact that $\operatorname{Forb}(\mathcal{F})$ can be axiomatized using universal sentences for any class $\mathcal{F}$ of finite $\mathcal{L}$-structures, and the fact that universal sentences are preserved under taking substructures (see the Łos-Tarski Theorem in ). # $\mathcal{L}_{\mathcal{H}}$-structures {#tildeLstructures} From now on, $\mathcal{L}$ is a fixed finite relational language and $r:=r_{\mathcal{L}}\geq 2$. For this section, $\mathcal{H}$ is a nonempty collection of finite $\mathcal{L}$-structures. In this section we introduce a language $\mathcal{L}_{\mathcal{H}}$ associated to $\mathcal{L}$ and $\mathcal{H}$. Structures in this new language play key roles in our main theorems. The goal of this section is to formalize how an $\mathcal{L}_{\mathcal{H}}$-structure $M$ with the right properties can serve as a "template" for building $\mathcal{L}$-structures with the same underlying set as $M$. We now give an example of a hereditary property and its corresponding auxiliary language as in Definition [\[LHdef\]](#LHdef){reference-type="ref" reference="LHdef"}. Observe that in an arbitrary $\mathcal{L}_{\mathcal{H}}$-structure, the relation symbols in $\mathcal{L}_{\mathcal{H}}$ may have nothing to do with the properties of the type space $S_r(\mathcal{H})$. For instance, in the notation of Example [\[ex1\]](#ex1){reference-type="ref" reference="ex1"}, we can easily build an $\mathcal{L}_{\mathcal{P}}$-structure $M$ so that for some $a,b\in dom(M)$, $M\models R_{p_1}(a,b) \wedge \neg R_{p_1}(b,a)$, even though $p_1(x,y)=p_1(y,x)$ in $S_2(\mathcal{P})$. This kind of behavior will be undesirable for various technical reasons. We now define the class of $\mathcal{L}_{\mathcal{H}}$-structures which are most nicely behaved for our purposes, and where in particular, this bad behavior does not happen. The idea is that $\mathcal{L}_{\mathcal{H}}$-templates are the $\mathcal{L}_{\mathcal{H}}$-structures which most accurately reflect the properties of $S_r(\mathcal{H})$. While $\mathcal{L}_{\mathcal{H}}$-templates are important for the main results of this paper, many of the definitions and facts in the rest of this section will be presented for $\mathcal{L}_{\mathcal{H}}$-structures with weaker assumptions. ## Choice functions and subpatterns In this subsection, we give crucial definitions for how we can use $\mathcal{L}_{\mathcal{H}}$-structures to build $\mathcal{L}$-structures. In the notation of Definition [\[chdef\]](#chdef){reference-type="ref" reference="chdef"}, note $Ch(M)\neq \emptyset$ if and only if $Ch_M(A)\neq \emptyset$ for all $A\in {V\choose r}$. Observe that $Ch_M(A)\neq \emptyset$ for all $A\in {V\choose r}$ if and only if $M$ is complete. Therefore $Ch(M)\neq \emptyset$ if and only if $M$ is complete. The following observation is immediate from the definition of $\mathcal{L}_{\mathcal{H}}$-template. The following fact is one reason why $\mathcal{L}_{\mathcal{H}}$-templates are convenient. The next example shows Proposition [\[nice\]](#nice){reference-type="ref" reference="nice"} can fail when we are not dealing with $\mathcal{L}_{\mathcal{H}}$-templates. The next definition shows how choice functions give rise to $\mathcal{L}$-structures. This example demonstrates although $\mathcal{L}_{\mathcal{H}}$-templates are well behaved in certain ways, an $\mathcal{L}_{\mathcal{H}}$-template can have full subpatterns that are not in $\mathcal{H}$. We will give further definitions to address this in Section [3.3](#temsec){reference-type="ref" reference="temsec"}. ## Errors and counting subpatterns In this subsection we characterize when an $\mathcal{L}_{\mathcal{H}}$-structure has the property that every choice function gives rise to a subpattern. This will be important for counting subpatterns of $\mathcal{L}_{\mathcal{H}}$-structures. An $\mathcal{L}_{\mathcal{H}}$-structure $M$ is error-free if it is $\mathcal{E}$-free. Error-free $\mathcal{L}_{\mathcal{H}}$-structures will be important for the following reason. This definition and the following observation will be crucial to our enumeration theorem. Remark [\[specialob0\]](#specialob0){reference-type="ref" reference="specialob0"} applies to most examples we are interested in, including graphs, (colored) $k$-uniform hypergraphs for any $k\geq 2$, directed graphs, and discrete metric spaces. ## $\mathcal{H}$-random $\mathcal{L}_{\mathcal{H}}$-structures and $\mathcal{L}_{\mathcal{H}}$-templates {#temsec} In this subsection we consider $\mathcal{L}_{\mathcal{H}}$-structures with the property that all choice functions give rise to subpatterns in $\mathcal{H}$. Observe that by Proposition [\[Lrandom\]](#Lrandom){reference-type="ref" reference="Lrandom"}, any $\mathcal{H}$-random $\mathcal{L}_{\mathcal{H}}$-structure is error-free. The difference between being error-free and being $\mathcal{H}$-random is as follows. If an $\mathcal{L}_{\mathcal{H}}$-structure is error-free, then it must have at least one full subpattern, however some or all its subpatterns may not be in $\mathcal{H}$. On the other hand, if an $\mathcal{L}_{\mathcal{H}}$-structure is $\mathcal{H}$-random, then it must have at least one full subpattern, and further, all its full subpatterns must also be in $\mathcal{H}$. The most important $\mathcal{L}_{\mathcal{H}}$-structures for the rest of the paper are $\mathcal{H}$-random $\mathcal{L}_{\mathcal{H}}$-templates. We now fix notation for these special $\mathcal{L}_{\mathcal{H}}$-structures. In the above notation, $\mathcal{R}$ is for "random." Note that if $\mathcal{H}(n)=\emptyset$ for some $n$, then $\mathcal{R}(n,\mathcal{H})=\emptyset$. # Main Results {#mainresults} In this section we state the main results of this paper. Recall that $\mathcal{L}$ is a fixed finite relational language of maximum arity $r\geq 2$. We now define our generalization of extremal graphs. By convention, set $\max \emptyset=0$. The main idea is that when $\mathcal{H}$ is a hereditary $\mathcal{L}$-property, $\textnormal{ex}(n,\mathcal{H})$ is the correct generalization of the extremal number of a graph, and elements of $\mathcal{R}_{ex}(n,\mathcal{H})$ are the correct generalizations extremal graphs of size $n$. Using techniques similar to those in we will show the following. We now state our approximate enumeration theorem in terms of the asymptotic density. The notion $\pi(\mathcal{H})$ is related to many existing notions of asymptotic density for various combinatorial structures, and Theorem [\[enumeration\]](#enumeration){reference-type="ref" reference="enumeration"} can be seen as generalizing many existing enumeration theorems. Some of these connections will be discussed in Section [11](#end){reference-type="ref" reference="end"} and Appendices [\[coloredhg\]](#coloredhg){reference-type="ref" reference="coloredhg"}, [\[dgsec\]](#dgsec){reference-type="ref" reference="dgsec"}, and [\[trifreesec\]](#trifreesec){reference-type="ref" reference="trifreesec"}. We say a hereditary $\mathcal{L}$-property $\mathcal{H}$ is fast-growing if $\pi(\mathcal{H})>1$. In this case, we informally say $M\in \mathcal{R}(n,\mathcal{H})$ is almost extremal if $sub(M)\geq\textnormal{ex}(n,\mathcal{H})^{1-\epsilon}$ for some small $\epsilon$. Our next theorem shows that almost all elements in a fast-growing hereditary $\mathcal{L}$-property $\mathcal{H}$ are close to subpatterns of almost extremal elements of $\mathcal{R}(n,\mathcal{H})$. Given $\epsilon>0$, $n$, and a collection $\mathcal{H}$ of $\mathcal{L}$-structures, let \[\begin{aligned} E(n,\mathcal{H})&=\{G\in \mathcal{H}_n: G\unlhd_pM\text{ for some $M\in \mathcal{R}_{ex}(n,\mathcal{H})$}\}\text{ and }\\ E(\epsilon, n,\mathcal{H})&=\{G\in \mathcal{H}_n: G\unlhd_pM\text{ for some $M\in \mathcal{R}(n,\mathcal{H})$ with $sub(M)\geq\textnormal{ex}(n,\mathcal{H})^{1-\epsilon}$}\}. \end{aligned}\] Given $\delta>0$, let $E^{\delta}(n,\mathcal{H})$ and $E^{\delta}(\epsilon, n,\mathcal{H})$ denote the set of $G\in \mathcal{H}_n$ which are $\delta$-close to any element of $E(n,\mathcal{H})$ and $E(\epsilon, n,\mathcal{H})$, respectively. We now define our generalization of a graph stability theorem. Our next result, Theorem [\[stab\]](#stab){reference-type="ref" reference="stab"} below, shows that if a fast-growing hereditary $\mathcal{L}$-property $\mathcal{H}$ has a stability theorem, we can strengthen Theorem [\[b4stab\]](#b4stab){reference-type="ref" reference="b4stab"} to say that that almost all elements in $\mathcal{H}_n$ are approximately subpatterns of elements of $\mathcal{R}_{ex}(n,\mathcal{H})$. When one has a good understanding of the structure of elements in $\mathcal{R}_{ex}(n, \mathcal{H})$, Theorem [\[stab\]](#stab){reference-type="ref" reference="stab"} gives us a good description of the approximate structure of most elements in $\mathcal{H}_n$, when $n$ is large. The main new tool we will use to prove our main theorems is Theorem [\[version1\]](#version1){reference-type="ref" reference="version1"} below, which is an adaptation of the hypergraph containers theorem to the setting of $\mathcal{L}$-structures. We will combine Theorem [\[version1\]](#version1){reference-type="ref" reference="version1"} with a general version of the graph removal lemma proved by Aroskar and Cummings in to prove a supersaturation theorem for hereditary $\mathcal{L}$-properties (Theorem [\[GENSUPERSAT\]](#GENSUPERSAT){reference-type="ref" reference="GENSUPERSAT"} below), and a version of the hypergraph containers theorem for hereditary $\mathcal{L}$-properties (Theorem [\[COROLLARY2\]](#COROLLARY2){reference-type="ref" reference="COROLLARY2"} below). # Proofs of Main Theorems In this section we prove our main results using Theorems [\[version1\]](#version1){reference-type="ref" reference="version1"}, [\[GENSUPERSAT\]](#GENSUPERSAT){reference-type="ref" reference="GENSUPERSAT"}, and [\[COROLLARY2\]](#COROLLARY2){reference-type="ref" reference="COROLLARY2"}. For the rest of the section, $\mathcal{H}$ is a fixed hereditary $\mathcal{L}$-property. We now prove Theorem [\[enumeration\]](#enumeration){reference-type="ref" reference="enumeration"}. The proof is based on the method of proof in. Let $b_n=ex(n,\mathcal{H})^{1/{n\choose r}}$. If $\mathcal{H}$ is trivial, then for sufficiently large $n$, $\mathcal{R}(n,\mathcal{H})=\emptyset$ so by convention, $\textnormal{ex}(n,\mathcal{H})=0$. Thus, for sufficiently large $n$, $b_n=0$ and $\pi(\mathcal{H})$ exists and is equal to zero. Assume now $\mathcal{H}$ is nontrivial. We show that the sequence $b_n$ is bounded below and non-increasing. Since $\mathcal{H}$ is non-trivial and has the hereditary property, $\mathcal{H}_n\neq \emptyset$ for all $n$. Fix $n\geq 1$ and choose any $N\in \mathcal{H}_n$. Let $\tilde{N}$ be the $\mathcal{L}_{\mathcal{H}}$-structure defined as in Lemma [\[templatelem2\]](#templatelem2){reference-type="ref" reference="templatelem2"} for $N$. Then $\tilde{N}$ is an $\mathcal{L}_{\mathcal{H}}$-template, and its only full subpattern is $N$. Since $N\in \mathcal{H}$, this implies $\tilde{N}\in \mathcal{R}(n,\mathcal{H})$ and $sub(\tilde{N})=1$. So we have shown $b_n\geq 1$ for all $n\geq 1$. We now show the $b_n$ are non-increasing. Fix $n\geq 2$. Let $M\in \mathcal{R}(n,\mathcal{H})$ be such that $sub(M)\geq 1$ and let $V=dom(M)$. Fix $a\in V$ and set $V_a=V\setminus \{a\}$ and $M_a=M[V_a]$. We claim $M_a\in \mathcal{R}(n-1,\mathcal{H})$. Because $M$ is an $\mathcal{L}_{\mathcal{H}}$-template, the definition of $M_a$ implies $M_a$ is also an $\mathcal{L}_{\mathcal{H}}$-template. Suppose $\chi\in Ch(M_a)$. We want to show there exists $N_a\in \mathcal{H}$ with $N_a\unlhd_{\chi}M_a$. We define a function $\chi':{V\choose r}\rightarrow S_r(C_V,\mathcal{H})$ as follows. For $A\in {V_a\choose r}$, set $\chi'(A)=\chi(A)$, and for $A\in {V\choose r}\setminus A\in {V_a\choose r}$, choose $\chi'(A)$ to be any element of $Ch_{M_a}(A)=Ch_M(A)$ (this is possible since $M$ is complete). Note that for each $A\in {V_a\choose r}$, $\chi(A)\in Ch_M(A)$, so $\chi'\in Ch(M)$. Because $M$ is $\mathcal{H}$-random, there is $N\in \mathcal{H}$ such that $N\unlhd_{\chi'}M$. Let $N_a=N[V_a]$. Because $\mathcal{H}$ has the hereditary property and $N_a\subseteq_{\mathcal{L}}N$, $N_a\in \mathcal{H}$. For each $A\in {V_a\choose r}$, $Diag^{N_a}(A)=Diag^N(A)=\chi'(A)=\chi(A)$, so $N_a\unlhd_{\chi}M_a$. Thus we have verified that $M_a\in \mathcal{R}(n-1,\mathcal{H})$. By definition of $b_{n-1}$, this implies $sub(M_a)^{1/{n-1\choose r}}\leq b_{n-1}$. Because $M_a$ is $\mathcal{H}$-random, Lemma [\[Lrandom\]](#Lrandom){reference-type="ref" reference="Lrandom"} implies it is error-free, so Observation [\[ob0\]](#ob0){reference-type="ref" reference="ob0"} implies $sub(M_a)=\prod_{A\in {V_a\choose r}} \|Ch_{M_a}(A)\|$. Then observe that \[sub(M)=\Big(\prod_{a\in V} \prod_{A\in {V_a\choose r}} \|Ch_{M_a}(A)\|\Big)^{1/(n-r)}= \Big(\prod_{a\in V} sub(M_a)\Big)^{1/(n-r)}.\] Since $sub(M_a)\leq b_{n-1}^{n-1\choose r}$, this implies \[sub(M)\leq \Big(\prod_{a\in V} b_{n-1}^{n-1\choose r}\Big)^{1/(n-r)} = b_{n-1}^{n{n-1\choose r}/(n-r)}= b_{n-1}^{n\choose r}.\] Thus for all $M\in \mathcal{R}(n,\mathcal{H})$, $sub(M)^{1/{n\choose r}}\leq b_{n-1}$. So by definition, $b_n\leq b_{n-1}$. 0◻ The following observations follow from the proof of Theorem [\[densityexists\]](#densityexists){reference-type="ref" reference="densityexists"}. Assume $\mathcal{H}$ is a hereditary $\mathcal{L}$-property. Recall we want to show the following. 1. If $\pi(\mathcal{H})>1$, then $\|\mathcal{H}_n\|= \pi(\mathcal{H})^{{n\choose r}+o(n^r)}$. 2. If $\pi(\mathcal{H})\leq 1$, then $\|\mathcal{H}_n\|=2^{o(n^r)}$. Assume first that $\mathcal{H}$ is trivial. Then by Observation [\[ob5\]](#ob5){reference-type="ref" reference="ob5"}(b), $\pi(\mathcal{H})=0\leq 1$, so we are in case 2. Since $\|\mathcal{H}_n\|=0$ for all sufficiently large $n$, $\|\mathcal{H}_n\|=2^{o(n^2)}$ holds, as desired. Assume now $\mathcal{H}$ is non-trivial, so $\pi(\mathcal{H})\geq 1$ by Observation [\[ob5\]](#ob5){reference-type="ref" reference="ob5"}(b). We show that for all $0<\eta<1$, either $\pi(\mathcal{H})=1$ and $\|\mathcal{H}_n\|\leq 2^{\eta n^r}$ or $\pi(\mathcal{H})>1$ and $\pi(\mathcal{H})^{n\choose 2}\leq \|\mathcal{H}_n\|\leq \pi(\mathcal{H})^{{n\choose r}+\eta n^r}$. Fix $0<\eta<1$. Let $\mathcal{F}$ be as in Observation [\[HP\]](#HP){reference-type="ref" reference="HP"} for $\mathcal{H}$ so that $\mathcal{H}=\operatorname{Forb}(\mathcal{F})$. Choose $\epsilon>0$ and $K$ as in Theorem [\[GENSUPERSAT\]](#GENSUPERSAT){reference-type="ref" reference="GENSUPERSAT"} for $\delta=\eta/4$. Replacing $K$ if necessary, assume $K\geq r$. Apply Theorem [\[version1\]](#version1){reference-type="ref" reference="version1"} to $\epsilon$ and $\mathcal{F}(K)$ to obtain $m=m(K,r)>1$ and $c=c(K,r, \mathcal{L}, \epsilon)$. Assume $n$ is sufficiently large. Theorem [\[version1\]](#version1){reference-type="ref" reference="version1"} with $W=[n]$ and $\mathcal{B}:=\operatorname{Forb}(\mathcal{F}(K))$ implies there is a collection $\mathcal{C}$ of $\mathcal{L}_{\mathcal{B}}$-templates with domain $[n]$ such that the following hold. (i) For all $\mathcal{F}(K)$-free $\mathcal{L}$-structures $M$ with domain $[n]$, there is $C\in \mathcal{C}$ such that $M\unlhd_pC$, (ii) For all $C\in \mathcal{C}$, $\textnormal{prob}(\widetilde{\mathcal{F}(K)},C)\leq \epsilon$ and $\textnormal{prob}(\mathcal{E},C)\leq \epsilon$. (iii) $\log \|\mathcal{C}\|\leq cn^{r-\frac{1}{m}}\log n$. Note that because $K\geq r$, $\mathcal{H}=\operatorname{Forb}(\mathcal{F})$ and $\mathcal{B}=\operatorname{Forb}(\mathcal{F}(K))$ imply we must have $S_r(\mathcal{H})=S_r(\mathcal{B})$. Consequently all $\mathcal{L}_{\mathcal{B}}$-templates are also $\mathcal{L}_{\mathcal{H}}$-templates. In particular the elements in $\mathcal{C}$ are all $\mathcal{L}_{\mathcal{H}}$-templates. Therefore, (ii) and Theorem [\[GENSUPERSAT\]](#GENSUPERSAT){reference-type="ref" reference="GENSUPERSAT"} imply that for all $C\in \mathcal{C}$, either $sub(C)\leq \textnormal{ex}(n,\mathcal{H})^{1+\eta/4}$ (case $\pi(\mathcal{H})>1$) or $sub(C)\leq 2^{\eta{n\choose r}/4}$ (case $\pi(\mathcal{H})=1$). Note every element in $\mathcal{H}_n$ is $\mathcal{F}$-free, so is also $\mathcal{F}(K)$-free. This implies by (i) that every element of $\mathcal{H}_n$ is a full subpattern of some $C\in \mathcal{C}$. Therefore we can construct every element in $\mathcal{H}_n$ as follows. 1. Choose a $C\in \mathcal{C}$. There are at most $\|\mathcal{C}\|\leq 2^{cn^{r-\frac{1}{m}} \log n}$ choices. 2. Choose a full subpattern of $C$. There are at most $sub(C)\leq \textnormal{ex}(n,\mathcal{H})^{1+\eta/4}$ choices if $\pi(\mathcal{H})>1$ and at most $sub(C)\leq 2^{\eta{n\choose r}/4}$ choices if $\pi(\mathcal{H})=1$. This implies \[\begin{aligned} \label{AP} \|\mathcal{H}_n\|\leq \begin{cases} 2^{cn^{r-\frac{1}{m}}\log n}\textnormal{ex}(n,\mathcal{H})^{1+\eta/4} & \text{ if }\pi(\mathcal{H})>1\\ 2^{cn^{r-\frac{1}{m}}\log n}2^{\eta{n\choose r}/4} & \text{ if }\pi(\mathcal{H})=1. \end{cases}\end{aligned}\] If $\pi(\mathcal{H})>1$, then we may assume $n$ is sufficiently large so that $\textnormal{ex}(n,\mathcal{H})\leq \pi(\mathcal{H})^{(1+\eta/4){n\choose r}}$ (see Observation [\[ob5\]](#ob5){reference-type="ref" reference="ob5"}(a)). Combining this with ([\[AP\]](#AP){reference-type="ref" reference="AP"}), we have that when $\pi(\mathcal{H})>1$, \[\|\mathcal{H}_n\|\leq 2^{cn^{r-\frac{1}{m}}\log n}\pi(\mathcal{H})^{(1+\eta/4)^2{n\choose r}}\leq \pi(\mathcal{H})^{{n\choose r}+\eta {n\choose r}},\] where the last inequality is because $\pi(\mathcal{H})>1$, $(1+\eta/4)^2<1+\eta$, and $n$ is sufficiently large. If $\pi(\mathcal{H})=1$, then ([\[AP\]](#AP){reference-type="ref" reference="AP"}) implies \[\|\mathcal{H}_n\|\leq 2^{cn^{r-\frac{1}{m}}\log n}2^{\eta{n\choose r}/2}\leq 2^{\eta{n\choose r}},\] where the last inequality is because $n$ is sufficiently large. Thus, we have shown $\|\mathcal{H}_n\|\leq 2^{\eta n^r}$ when $\pi(\mathcal{H})=1$ and $\|\mathcal{H}_n\|\leq \pi(\mathcal{H})^{{n\choose r}+\eta n^r}$ when $\pi(\mathcal{H})>1$. We just have left to show that when $\pi(\mathcal{H})>1$, then $\|\mathcal{H}_n\|\geq \pi(\mathcal{H})^{n\choose r}$. This holds because for any $M\in \mathcal{R}_{ex}([n],\mathcal{H})$, all $\textnormal{ex}(n,\mathcal{H})$ many full subpatterns of $M$ are in $\mathcal{H}_n$. Thus $\|\mathcal{H}_n\|\geq \textnormal{ex}(n,\mathcal{H})\geq \pi(\mathcal{H})^{n\choose r}$, where the second inequality is by Observation [\[ob5\]](#ob5){reference-type="ref" reference="ob5"}(a). This finishes the proof. 0◻ We now prove a few lemmas needed for Theorems [\[b4stab\]](#b4stab){reference-type="ref" reference="b4stab"} and [\[stab\]](#stab){reference-type="ref" reference="stab"}. . Let $\mathcal{H}$ be a fast-growing hereditary $\mathcal{L}$-property. Fix $\epsilon$ and $\delta>0$. Given $n$, let $A(n, \epsilon, \delta)=\mathcal{H}_n\setminus E^{\delta}(\epsilon, n,\mathcal{H})$. Recall, we want to show there is $\beta>0$ such that for sufficiently large $n$, \[\begin{aligned} \label{m} \frac{\|A(n,\epsilon, \delta)\|}{\|\mathcal{H}_n\|}\leq 2^{-\beta {n\choose r}}. \end{aligned}\] Let $\gamma>0$ be as in Lemma [\[deltaclose1\]](#deltaclose1){reference-type="ref" reference="deltaclose1"} for $\mathcal{H}$. Choose $K>2r$ sufficiently large so that $1-\epsilon +\gamma \delta /K<1-\epsilon/2$. Apply Theorem [\[COROLLARY2\]](#COROLLARY2){reference-type="ref" reference="COROLLARY2"} to $\frac{\delta}{K}$ to obtain constants $c$ and $m>1$. Assume $n$ is sufficiently large. Then Theorem [\[COROLLARY2\]](#COROLLARY2){reference-type="ref" reference="COROLLARY2"} implies there is a collection $\mathcal{C}$ of $\mathcal{L}_{\mathcal{H}}$-templates with domain $[n]$ such that the following hold. (i) For every $H\in \mathcal{H}_n$, there is $C\in \mathcal{C}$ such that $H\unlhd_pC$. (ii) For every $C\in \mathcal{C}$, there is $C'\in \mathcal{R}([n],\mathcal{H})$ such that $dist(C,C')\leq \delta$. (iii) $\log \|\mathcal{C}\|\leq cn^{r-\frac{1}{m}}\log n$. Suppose $G\in A(n,\epsilon,\delta)$. By (i), there is $C\in \mathcal{C}_n$ such that $G\unlhd_pC$. By (ii), there is $M_C\in \mathcal{R}([n],\mathcal{H})$ such that $\textnormal{dist}(C,M_C)\leq \frac{\delta}{K}$. By Lemma [\[deltaclose2\]](#deltaclose2){reference-type="ref" reference="deltaclose2"}, there is $G'\unlhd_p M_C$ with $\textnormal{dist}(G,G')\leq \frac{\delta}{K}\leq \delta$. If $sub(M_C)\geq \textnormal{ex}(n,\mathcal{H})^{1-\epsilon}$, then by definition of $E^{\delta}(\epsilon, n,\mathcal{H})$, $\textnormal{dist}(G,G')\leq \delta$ and $G'\unlhd_pM_C$ would imply $G\in E^{\delta}(\epsilon, n,\mathcal{H})$, contradicting our assumption that $G\in A(n,\epsilon, \delta)=\mathcal{H}_n\setminus E^{\delta}(\epsilon, n,\mathcal{H})$. Therefore, we must have $sub(M_C)<\textnormal{ex}(n,\mathcal{H})^{1-\epsilon}$. Note $M_C\in \mathcal{R}([n],\mathcal{H})$ implies $M_C$ is error-free, so Lemma [\[deltaclose1\]](#deltaclose1){reference-type="ref" reference="deltaclose1"} and the fact that $\textnormal{dist}(C,M_C)\leq \delta/K$ imply $sub(C)\leq sub(M_C)\textnormal{ex}(n,\mathcal{H})^{\gamma\delta/K}$. Combining this with the fact that $sub(M_C)<\textnormal{ex}(n,\mathcal{H})^{1-\epsilon}$ we have that \[sub(C)<\textnormal{ex}(n,\mathcal{H})^{1-\epsilon}\textnormal{ex}(n,\mathcal{H})^{\gamma\delta/K}=\textnormal{ex}(n,\mathcal{H})^{1-\epsilon +\gamma \delta/K}\leq\textnormal{ex}(n,\mathcal{H})^{1-\epsilon/2},\] where the second inequality is by assumption on $K$. Therefore every $G\in A(n,\epsilon, \delta)$ can be constructed as follows. - Choose $C\in \mathcal{C}_n$ with $sub(C)<\textnormal{ex}(n,\mathcal{H})^{1-\epsilon/2}$. There are at most $\|\mathcal{C}_n\|\leq 2^{cn^{r-\frac{1}{m}}\log n}$ ways to do this, where the bound is by (iii). Since $n$ is large and $\pi(\mathcal{H})>1$, we may assume $2^{cn^{r-\frac{1}{m}}\log n}\leq \pi(\mathcal{H})^{\epsilon{n\choose r}/4}$. - Choose a full subpattern of $C$. There are at most $sub(C)<\textnormal{ex}(n,\mathcal{H})^{1-\epsilon/2}$ ways to do this. Combining these bounds yields $\|A(n,\epsilon,\delta)\|\leq \pi(\mathcal{H})^{\epsilon{n\choose r}/4}\textnormal{ex}(n,\mathcal{H})^{1-\epsilon/2}$. Recall that $\|\mathcal{H}_n\|\geq\textnormal{ex}(n,\mathcal{H})$ holds, since for any $M\in \mathcal{R}_{ex}([n],\mathcal{H})$, all $\textnormal{ex}(n,\mathcal{H})$-many full subpatterns of $M$ are all in $\mathcal{H}_n$. Therefore \[\begin{aligned} \label{l} \frac{\|A(n,\epsilon, \delta)\|}{\|\mathcal{H}_n\|}\leq \frac{\pi(\mathcal{H})^{\epsilon{n\choose r}/4}\textnormal{ex}(n,\mathcal{H})^{1-\epsilon/2}}{\textnormal{ex}(n,\mathcal{H})}=\pi(\mathcal{H})^{\epsilon{n\choose r}/4}\textnormal{ex}(n,\mathcal{H})^{-\epsilon/2}\leq \pi(\mathcal{H})^{-\epsilon {n\choose r}/4}, \end{aligned}\] where the last inequality is because $\pi(\mathcal{H})^{n\choose r}\leq\textnormal{ex}(n,\mathcal{H})$. Therefore we have $\frac{\|A(n,\epsilon, \delta)\|}{\|\mathcal{H}_n\|}\leq 2^{-\beta{n\choose r}}$, where $\beta = \frac{\epsilon \log \pi(\mathcal{H})}{4\log 2}$. Note $\beta>0$ since $\pi(\mathcal{H})>1$. 0◻ Suppose $\mathcal{H}$ is a fast growing hereditary $\mathcal{L}$-property with a stability theorem. Fix $\delta>0$. Recall we want to show there is $\beta>0$ such that for sufficiently large $n$, \[\begin{aligned} \frac{\|\mathcal{H}_n\setminus E^{\delta}(n,\mathcal{H})\|}{\|\mathcal{H}_n\|}\leq 2^{-\beta {n\choose r}} \end{aligned}\] By Theorem [\[b4stab\]](#b4stab){reference-type="ref" reference="b4stab"}, it suffices to show that there are $\epsilon_1, \delta_1>0$ such that for all sufficiently large $n$, $E^{\delta_1}(\epsilon_1, n,\mathcal{H})\subseteq E^{\delta}(n,\mathcal{H})$. Because $\mathcal{H}$ has a stability theorem, there is $\epsilon$ such that for all sufficiently large $n$, if $H\in \mathcal{R}([n],\mathcal{H})$ satisfies $sub(H)\geq\textnormal{ex}(n,\mathcal{H})^{1-\epsilon}$, then there is $H'\in \mathcal{R}_{ex}([n],\mathcal{H})$ with $\textnormal{dist}(H,H')\leq \frac{\delta}{2}$. Fix $n$ sufficiently large. We claim $E^{\delta/2}(\epsilon, n,\mathcal{H})\subseteq E^{\delta}(n,\mathcal{H})$. Suppose $G\in E^{\delta/2}(\epsilon, n,\mathcal{H})$. Then by definition, $G$ is $\delta/2$-close to some $G'$ such that $G'\unlhd_pH$, for some $H\in \mathcal{R}([n],\mathcal{H})$ satisfying $sub(H)\geq\textnormal{ex}(n,\mathcal{H})^{1-\epsilon}$. By choice of $\epsilon$ and because $n$ is sufficiently large, there is $H'\in \mathcal{R}_{ex}([n],\mathcal{H})$ such that $\textnormal{dist}(H,H')\leq \frac{\delta}{2}$. Lemma [\[deltaclose2\]](#deltaclose2){reference-type="ref" reference="deltaclose2"} implies there is some $G''\unlhd_pH'$ such that $\textnormal{dist}(G',G'')\leq \frac{\delta}{2}$. Observe that $G''\in E(n,\mathcal{H})$ and \[\textnormal{dist}(G, G'')\leq \textnormal{dist}(G,G')+\textnormal{dist}(G', G'') \leq \frac{\delta}{2}+\frac{\delta}{2}=\delta.\] This implies that $G\in E^{\delta}(n,\mathcal{H})$, as desired. 0◻ # Characterization of $\mathcal{H}$-random $\mathcal{L}_{\mathcal{H}}$-templates {#Hrandom} In this section we give an equivalent characterization for when an $\mathcal{L}_{\mathcal{H}}$-structure is an $\mathcal{H}$-random $\mathcal{L}_{\mathcal{H}}$-template, where $\mathcal{H}$ is a hereditary $\mathcal{L}$-property. The results in this section will be used in the proofs of our remaining results, Theorems [\[version1\]](#version1){reference-type="ref" reference="version1"}, [\[GENSUPERSAT\]](#GENSUPERSAT){reference-type="ref" reference="GENSUPERSAT"}, and [\[COROLLARY2\]](#COROLLARY2){reference-type="ref" reference="COROLLARY2"}. For the rest of this section, $\mathcal{H}$ is a fixed nonempty collection of finite $\mathcal{L}$-structures. We are now ready to prove the main result of this section. # Graph Removal and Proofs of Theorems [\[GENSUPERSAT\]](#GENSUPERSAT){reference-type="ref" reference="GENSUPERSAT"} and [\[COROLLARY2\]](#COROLLARY2){reference-type="ref" reference="COROLLARY2"}. {#rphrem} In this section we will use a version of the graph removal lemma from to prove Theorem [\[GENSUPERSAT\]](#GENSUPERSAT){reference-type="ref" reference="GENSUPERSAT"} and to prove Theorem [\[COROLLARY2\]](#COROLLARY2){reference-type="ref" reference="COROLLARY2"} from Theorem [\[version1\]](#version1){reference-type="ref" reference="version1"}. We now state definitions required to quote the graph removal lemma from. Throughout the rest of this section, $\mathcal{L}_0$ is a fixed finite relational language with $r_{\mathcal{L}_0}=r$. Note $\mathcal{L}_0$ is not necessarily the same as $\mathcal{L}$, although we are assuming $r_{\mathcal{L}_0}=r_{\mathcal{L}}=r$. Given a partition $p$ of a finite set $X$, let $\|\|p\|\|$ denote the number of parts in $p$. Now we can define the notion of distance between two $\mathcal{L}_0$-structures from. We will see below in Lemma [\[distlem\]](#distlem){reference-type="ref" reference="distlem"} that this notion of distance, $d(M,N)$, is related to our notion of distance, $\textnormal{dist}(M,N)$. We first state the graph removal lemma of Aroskar and Cummings as it appears in their paper (Theorem 2 from ). The following relationship between $d(M,N)$ and $\textnormal{dist}(M,N)$ will allow us to restate this graph removal lemma. Given a tuple $\bar{x}=(x_1,\ldots, x_{\ell})$, a subtuple of $\bar{x}$ is any tuple $\bar{x}'=(x_{i_1},\ldots, x_{i_{\ell'}})$ where $1\leq i_1<\ldots<i_{\ell'}\leq \ell$. If $\ell'<\ell$, we say $\bar{x}'$ is a proper subtuple of $\bar{x}$, denoted $\bar{x}'\subsetneq \bar{x}$. We will use the following version of Theorem [\[triangleremoval\]](#triangleremoval){reference-type="ref" reference="triangleremoval"}, now adapted to our notation. Let $\mathcal{H}$ be a nontrivial hereditary $\mathcal{L}$-property and let $\mathcal{F}$ be as in Observation [\[HP\]](#HP){reference-type="ref" reference="HP"} so that $\mathcal{H}=\operatorname{Forb}(F)$. Recall we want to show that for all $\delta>0$, there are $\epsilon>0$ and $K$ such that for sufficiently large $n$, for any $\mathcal{L}_{\mathcal{H}}$-template $M$ of size $n$, if $\textnormal{prob}(\tilde{\mathcal{F}}(K) \cup \mathcal{E}(K),M)\leq \epsilon$ then 1. If $\pi(\mathcal{H})>1$, then $sub(M)\leq \textnormal{ex}(n,\mathcal{H})^{1+\delta}$. 2. If $\pi(\mathcal{H})\leq 1$, then $sub(M)\leq 2^{\delta {n\choose r}}$. Fix $\delta>0$. Apply Lemma [\[deltaclose1\]](#deltaclose1){reference-type="ref" reference="deltaclose1"} to $\mathcal{H}$ to obtain $\gamma>0$. Let $\mathcal{A}=\tilde{\mathcal{F}}\cup \mathcal{E}\cup \textnormal{FLAW}$. Apply Theorem [\[triangleremoval2\]](#triangleremoval2){reference-type="ref" reference="triangleremoval2"} to obtain $K$ and $\epsilon$ for $\delta/2\gamma$ and $\mathcal{A}$. Suppose $n$ is sufficiently large and $M$ is an $\mathcal{L}_{\mathcal{H}}$-template of size $n$ satsifying $\textnormal{prob}(\tilde{\mathcal{F}}(K)\cup \mathcal{E}(K), M)<\epsilon$. Because $M$ is an $\mathcal{L}_{\mathcal{H}}$-template, Lemmas [\[flaw\]](#flaw){reference-type="ref" reference="flaw"} implies for all $B\in \textnormal{FLAW}$, $\textnormal{prob}(B,M)=0$. Therefore $\textnormal{prob}(\mathcal{A}(K), M)<\epsilon$, so by Theorem [\[triangleremoval2\]](#triangleremoval2){reference-type="ref" reference="triangleremoval2"}, there is an $\mathcal{L}_{\mathcal{H}}$-structure $M'$ with $dom(M)=dom(M')$ such that $\textnormal{prob}(\mathcal{A},M')=0$ and $\textnormal{dist}(M,M')\leq \delta/2\gamma$. Since $\textnormal{prob}(\mathcal{A},M')=0$, Corollary [\[charR\]](#charR){reference-type="ref" reference="charR"} implies $M'\in \mathcal{R}(n,\mathcal{H})$. Thus $sub(M')\leq \textnormal{ex}(n,\mathcal{H})$ holds by definition of $\textnormal{ex}(n,\mathcal{H})$. Combining this with Lemma [\[deltaclose1\]](#deltaclose1){reference-type="ref" reference="deltaclose1"} (note $M'\in \mathcal{R}(n,\mathcal{H})$ implies $M'$ is error-free), we have the following. 1. If $\pi(\mathcal{H})>1$, then $sub(M)\leq sub(M')\textnormal{ex}(n,\mathcal{H})^{\gamma (\delta/2\gamma)}=sub(M')\textnormal{ex}(n,\mathcal{H})^{\delta/2}\leq \textnormal{ex}(n,\mathcal{H})^{1+\delta/2}$. 2. If $\pi(\mathcal{H})=1$, then $sub(M)\leq sub(M')2^{\gamma (\delta/2\gamma) {n\choose r}}=sub(M')2^{\delta{n\choose r}/2}\leq \textnormal{ex}(n,\mathcal{H})2^{\delta/2 {n\choose r}}$. We are done in the case where $\pi(\mathcal{H})>1$. If $\pi(\mathcal{H})=1$, assume $n$ is sufficiently large so that $\textnormal{ex}(n,\mathcal{H})\leq 2^{\delta/2{n\choose r}}$. Then (2) implies $sub(M)\leq 2^{\delta{n\choose r}}$, as desired. 0◻ Suppose $\mathcal{H}$ is a hereditary $\mathcal{L}$-property. Let $\mathcal{F}$ be the class of finite $\mathcal{L}$-structures from Observation [\[HP\]](#HP){reference-type="ref" reference="HP"} so that $\mathcal{H}=\operatorname{Forb}(\mathcal{F})$. Then for each $n$, $\mathcal{H}_n$ is the set of all $\mathcal{F}$-free $\mathcal{L}$-structures with domain $[n]$. Let $\mathcal{A}=\tilde{\mathcal{F}}\cup \mathcal{E}\cup \textnormal{FLAW}$. Fix $\delta>0$ and choose $K$ and $\epsilon$ as in Theorem [\[triangleremoval2\]](#triangleremoval2){reference-type="ref" reference="triangleremoval2"} for $\delta$ and the family $\mathcal{A}$. By replacing $K$ if necessary, assume $K\geq r$. Apply Theorem [\[version1\]](#version1){reference-type="ref" reference="version1"} to $\mathcal{B}:=\mathcal{F}(K)$ to obtain $c=c(K,r,\mathcal{L},\epsilon)$, $m=m(K,r)$. Observe the choice of $K$ depended on $\mathcal{H}$ and $r=r_{\mathcal{L}}$, so $m=m(\mathcal{H}, r_{\mathcal{L}})$. Since $r_{\mathcal{L}}$ depends on $\mathcal{L}$, $c=c(\mathcal{H},\mathcal{L},\epsilon)$. Let $n$ be sufficiently large. Then Theorem [\[version1\]](#version1){reference-type="ref" reference="version1"} applied to $W=[n]$ implies there is a collection $\mathcal{C}$ of $\mathcal{L}_{\mathcal{B}}$-templates with domain $[n]$ such that the following hold. (i) For all $\mathcal{F}(K)$-free $\mathcal{L}$-structures $M$ with domain $[n]$, there is $C\in \mathcal{C}$ such that $M\unlhd_pC$. (ii) For all $C\in \mathcal{C}$, $\textnormal{prob}(\widetilde{\mathcal{F(K)}},C)\leq \epsilon$ and $\textnormal{prob}(\mathcal{E},C)\leq \epsilon$. (iii) $\log \|\mathcal{C}\|\leq cn^{r-\frac{1}{m}}\log n$. We show this $\mathcal{C}$ satisfies the conclusions of Theorem [\[COROLLARY2\]](#COROLLARY2){reference-type="ref" reference="COROLLARY2"} with $c$, $m$ and $\delta$. Note that because $K\geq r$, $S_r(\mathcal{H})=S_r(\mathcal{B})$, so all $\mathcal{L}_{\mathcal{B}}$-templates are also $\mathcal{L}_{\mathcal{H}}$-templates. In particular the elements in $\mathcal{C}$ are all $\mathcal{L}_{\mathcal{H}}$-templates. Clearly (iii) implies part (3) of Theorem [\[COROLLARY2\]](#COROLLARY2){reference-type="ref" reference="COROLLARY2"} holds. For part (1), since any $H\in \mathcal{H}_n$ is $\mathcal{F}$-free, it is also $\mathcal{F}(K)$-free, so (i) implies there is $C\in \mathcal{C}$ such that $H\unlhd_pC$. This shows part (1) of Theorem [\[COROLLARY2\]](#COROLLARY2){reference-type="ref" reference="COROLLARY2"} holds. For part (2), fix $C\in \mathcal{C}$. Since $C$ is an $\mathcal{L}_{\mathcal{H}}$-template, Lemma [\[flaw\]](#flaw){reference-type="ref" reference="flaw"} implies $\textnormal{prob}(G,C)=0$ for all $G\in \textnormal{FLAW}$. Then (ii) implies that for all $G\in \widetilde{\mathcal{F}(K)}\cup \mathcal{E}$, $\textnormal{prob}(G,C)\leq \epsilon$. Since $\widetilde{\mathcal{F}(K)}=\tilde{\mathcal{F}}(K)$, these facts imply that for all $G\in \mathcal{A}(K)$, $\textnormal{prob}(G,C)\leq \epsilon$. Thus Theorem [\[triangleremoval2\]](#triangleremoval2){reference-type="ref" reference="triangleremoval2"} implies there is an $\mathcal{L}_{\mathcal{H}}$-structure $C'$ with $dom(C)=dom(C')=[n]$ such that $\textnormal{dist}(C,C')\leq \delta$ and $\textnormal{prob}(\mathcal{A},C')=0$. Since $\textnormal{prob}(\mathcal{A},C')=0$, $C'$ is a $\textnormal{FLAW}$-free, $\tilde{\mathcal{F}}$-free, and error-free $\mathcal{L}_{\mathcal{H}}$-structure with domain $[n]$, so by Corollary [\[charR\]](#charR){reference-type="ref" reference="charR"}, $C'\in \mathcal{R}([n],\mathcal{H})$. This finishes the proof. 0◻ # A Reduction {#section2thm1} We have now proved all the results in this paper except Theorem [\[version1\]](#version1){reference-type="ref" reference="version1"}. In this section we prove Theorem [\[version1\]](#version1){reference-type="ref" reference="version1"} by reducing it to another result, Theorem [\[VERSION2\]](#VERSION2){reference-type="ref" reference="VERSION2"} (which is proved in Section [9](#VERSION2pf){reference-type="ref" reference="VERSION2pf"}). ## Preliminaries {#prelims} In this subsection we give preliminaries necessary for the statement of Theorem [\[VERSION2\]](#VERSION2){reference-type="ref" reference="VERSION2"}. Many of these notions are similar to definitions from Section [3](#tildeLstructures){reference-type="ref" reference="tildeLstructures"}. However, we will see that our proofs necessitate this more syntactic treatment. We will say $\sigma\subseteq S_r(C)$ is a syntactic type diagram if it is a syntactic $\|V(\sigma)\|$-diagram. Observe that if $\sigma$ is a syntactic $m$-diagram, then by definition, $\|V(\sigma)\|=m$ and $\|\sigma\|={m\choose r}$. Given a tuple of constants $\bar{c}=(c_1,\ldots, c_k)$, a first-order language $\mathcal{L}_0$ containing $\{c_1,\ldots, c_k\}$, and an $\mathcal{L}_0$-structure $M$, let $\bar{c}^M$ denote the tuple $(c_1^M,\ldots, c_k^M)\in dom(M)^k$. Suppose that $M$ is an $\mathcal{L}$-structure with $dom(M)=W$. The canonical type-diagram of $M$ is \[Diag^{tp}(M)=\{p(c_{\bar{a}})\in S_r(C_W): M\models p(\bar{a})\}.\] In other words, $Diag^{tp}(M)=Diag^{tp}(M,C_W)$ where $M$ is considered with its natural $\mathcal{L}\cup C_W$-structure. Observe that $Diatg^{tp}(M)$ is always a syntactic $\|dom(M)\|$-diagram. The difference between $Diag^{tp}(M)$ and $Diag(M)$ is that elements of $Diag^{tp}(M)$ are types (with constants plugged in for the variables) while the elements of $Diag(M)$ are formulas (with constants plugged in for the variables). Clearly $Diag(M)$ and $Diag^{tp}(M)$ contain the same information. We now make a few observations which will be used in the remainder of the chapter. For the rest of this subsection, $\mathcal{H}$ is a fixed collection of finite $\mathcal{L}$-structures. ## Proof of Theorem [\[version1\]](#version1){reference-type="ref" reference="version1"} {#section3thm1} In this section we state Theorem [\[VERSION2\]](#VERSION2){reference-type="ref" reference="VERSION2"} and use it to prove Theorem [\[version1\]](#version1){reference-type="ref" reference="version1"}. Given a collection $\mathcal{H}$ of finite $\mathcal{L}$-structures, we now define a way of building an $\mathcal{L}_{\mathcal{H}}$-template from a complete subset of $S_r(C_W,\mathcal{H})$. In the notation of Definition [\[D_Cdef\]](#D_Cdef){reference-type="ref" reference="D_Cdef"}, note that for all $A\in {W\choose r}$, $Ch_{D_{\sigma}}(A)=Ch_{\sigma}(A)$ (here $Ch_{D_{\sigma}}(A)$ is in the sense of Definition [\[chdef\]](#chdef){reference-type="ref" reference="chdef"} and $Ch_{\sigma}(A)$ is in the sense of Definition [\[chd2\]](#chd2){reference-type="ref" reference="chd2"}). We now prove two lemmas. Let $0<\epsilon<1$ and let $k\geq r$ be an integer. Choose the constants $c=c(k,r,\mathcal{L},\epsilon)$ and $m=m(k,r)$ to be the ones given by Theorem [\[VERSION2\]](#VERSION2){reference-type="ref" reference="VERSION2"}. Suppose $\mathcal{F}$ is a collection of finite $\mathcal{L}$-structures, each of size at most $k$, and $\mathcal{B}:=\operatorname{Forb}(\mathcal{F})\neq \emptyset$. Suppose $n$ is sufficiently large and $W$ is a set of size $n$. Theorem [\[VERSION2\]](#VERSION2){reference-type="ref" reference="VERSION2"} applied to $\mathcal{B}$ implies there exists a set $\Sigma\subseteq \mathcal{P}(S_r(C_W,\mathcal{B}))$ such that the following hold. (i) For all $\mathcal{F}$-free $\mathcal{L}$-structures $M$ with domain $W$, there is $\sigma\in \Sigma$ such that $Diag^{tp}(M)\subseteq \sigma$. (ii) For all $\sigma\in \Sigma$ the following hold. For each $1\leq \ell\leq k$, $\|Diag^{tp}(\mathcal{F}(\ell), C_W)\cap Span(\sigma)\|\leq \epsilon{n\choose \ell}$, and for each $r+1\leq \ell \leq 2r$, $\|Err_{\ell}(C_W)\cap Span(\sigma)\|\leq \epsilon{n\choose \ell}$. (iii) $\log \|\Sigma\|\leq cn^{r-\frac{1}{m}}\log n$. Set $\mathcal{D}=\{D_{\sigma}: \sigma \in \Sigma\}$, where for each $\sigma\in \Sigma$, $D_{\sigma}$ is the $\mathcal{L}_{\mathcal{B}}$-structure from Definition [\[D_Cdef\]](#D_Cdef){reference-type="ref" reference="D_Cdef"}. We claim this $\mathcal{D}$ satisfies conclusions of Theorem [\[version1\]](#version1){reference-type="ref" reference="version1"}. First note (i) and part 3 of Observation [\[ob1\]](#ob1){reference-type="ref" reference="ob1"} imply that every $\sigma\in \Sigma$ is complete in the sense of Definition [\[chd2\]](#chd2){reference-type="ref" reference="chd2"}. Therefore Lemma [\[templateclaim\]](#templateclaim){reference-type="ref" reference="templateclaim"} implies each $D_{\sigma}\in \mathcal{D}$ is an $\mathcal{L}_{\mathcal{B}}$-template. We now verify parts (1)-(3) of Theorem [\[version1\]](#version1){reference-type="ref" reference="version1"} hold for this $\mathcal{D}$. Clearly $\|\mathcal{D}\|\leq \|\Sigma\|$, so (iii) implies part (3) of Theorem [\[version1\]](#version1){reference-type="ref" reference="version1"} is satisfied. Suppose now $M$ is an $\mathcal{F}$-free $\mathcal{L}$-structure with $dom(M)=W$. By (i), there is $\sigma\in \Sigma$ such that $Diag^{tp}(M)\subseteq \sigma$. We claim that $M\unlhd_pD_{\sigma}$. Let $A \in {W\choose r}$ and suppose $p(\bar{x})\in S_r(\mathcal{H})$ is such that $M\models p(\bar{a})$ for some enumeration $\bar{a}$ of $A$. Then $Diag^M(A)=p(c_{\bar{a}})\in Diag^{tp}(M)\subseteq \sigma$ implies by definition of $D_{\sigma}$, $D_{\sigma}\models R_p(\bar{a})$, so $p(c_{\bar{a}})\in Ch_{D_{\sigma}}(A)$. This shows $M\leq_p D_{\sigma}$. Then $M\unlhd_pD_{\sigma}$ holds because by assumption $dom(M)=dom(D_{\sigma})=W$. Thus part (1) of Theorem [\[version1\]](#version1){reference-type="ref" reference="version1"} is satisfied. We now verify part (2) of Theorem [\[version1\]](#version1){reference-type="ref" reference="version1"}. Let $D_{\sigma}\in \mathcal{D}$. We need to show $\textnormal{prob}(\tilde{\mathcal{F}},D_{\sigma})\leq \epsilon$ and $\textnormal{prob}(\mathcal{E}, D_{\sigma})\leq \epsilon$. For each $1\leq \ell \leq k$, we have \[\|\textnormal{cop}(\tilde{\mathcal{F}}(\ell), D_{\sigma})\|\leq \|Diag^{tp}(\mathcal{F}(\ell),C_W)\cap Span(\sigma)\|\leq \epsilon{n\choose \ell},\] where the first inequality is because of Lemma [\[boundinglem\]](#boundinglem){reference-type="ref" reference="boundinglem"} and the second inequality is by (ii). This implies that for all $1\leq \ell \leq k$, $\|\textnormal{cop}(\tilde{\mathcal{F}}(\ell), D_{\sigma})\|\leq \epsilon {n\choose \ell}$, so $\textnormal{prob}(\tilde{\mathcal{F}}(\ell),D_{\sigma})\leq \epsilon$. Since every element in $\tilde{\mathcal{F}}$ has size at most $k$, we have $\textnormal{prob}(\tilde{\mathcal{F}}, D_{\sigma})\leq \epsilon$. Similarly, for each $r+1\leq \ell \leq 2r$, \[\|\textnormal{cop}(\mathcal{E}(\ell), D_{\sigma})\|\leq\|Err_{\ell}(C_W)\cap Span (\sigma)\|\leq \epsilon {n\choose \ell},\] where the first inequality is by Lemma [\[boundinglem\]](#boundinglem){reference-type="ref" reference="boundinglem"} and the second inequality is by (ii). This implies for all $r+1\leq \ell \leq 2r$, $\|\textnormal{cop}(\mathcal{E}(\ell), D_{\sigma})\|\leq \epsilon{n\choose \ell}$, so $\textnormal{prob}(\mathcal{E}(\ell), D_{\sigma})\leq \epsilon$. Since every element in $\mathcal{E}$ has size at least $r+1$ and at most $2r$, we have $\textnormal{prob}(\mathcal{E}, D_{\sigma})\leq \epsilon$. This finishes the proof. 0◻ # Applying Hypergraph Containers to Prove Theorem [\[VERSION2\]](#VERSION2){reference-type="ref" reference="VERSION2"} {#VERSION2pf} In this section we prove Theorem [\[VERSION2\]](#VERSION2){reference-type="ref" reference="VERSION2"}. We will use the hypergraph containers theorem. We begin with a definition. Observe that in the above notation, an $\mathcal{L}$-structure of size at least $K$ is $\mathcal{A}$-free if and only if it is $cl_K(\mathcal{A})$-free. We now state a key lemma. We now present a computational lemma which will be used in the proof of Theorem [\[VERSION2\]](#VERSION2){reference-type="ref" reference="VERSION2"}. We now give a procedure for defining a special hypergraph given a finite set and a collection of $\mathcal{L}$-structures satisfying certain properties. Assume we are given the following. 1. A nonempty collection, $\mathcal{F}$, of finite $\mathcal{L}$-structures, each of size at most $k$, where $k\geq r$ is an integer. 2. A set $W$ of size $n$, where $n\geq k$ is an integer. Let $\mathcal{H}$ be the class of all finite $cl_k(\mathcal{F})$-free $\mathcal{L}$-structures. Define the hypergraph $H=H(\mathcal{F}, W)$ as follows. \[\begin{aligned} V(H)&=S_r(C_W,\mathcal{H})\text{ and }\\ E(H)&=Diag^{tp}(cl_k(\mathcal{F}), C_W)\cup Err_k(C_W). \end{aligned}\] We now make a few observations about $H$. First, note that the edges of $H$ are syntactic $k$-diagrams, so $H$ is a ${k\choose r}$-uniform hypergraph. By definition $\|V(H)\|={n\choose r}\|S_r(\mathcal{H})\|$. If $X$ and $X'$ are both in ${C_W\choose k}$, then since relabeling constants does not change the satisfiability properties of a collection of $\mathcal{L}\cup C_W$-sentences, we must have $\|Diag^{tp}(Cl_k(\mathcal{F}), X)\cup Err_k(X)\|=\|Diag^{tp}(Cl_k(\mathcal{F}), X')\cup Err_k(X')\|$. Therefore, the following is well defined. We claim that $\|E(H)\|=\alpha {n\choose k}$. Indeed, any $\sigma\in E(H)$ can be constructed as follows. - Choose $X\in {C_W\choose k}$. There are ${n\choose k}$ choices. - Choose an element $\sigma \in Diag^{tp}(Cl_k(\mathcal{F}), C_W)\cup Err_k(C_W)$ such that $V(\sigma)=X$, i.e. choose an element $\sigma \in Diag^{tp}(Cl_k(\mathcal{F}), X)\cup Err_k(X)$. There are $\alpha$ choices. Each of these choices lead to distinct subsets $\sigma\in E(H)$, so this shows $\|E(H)\|=\alpha {n\choose k}$. Note that because $\mathcal{F}\neq \emptyset$, $\alpha\geq 1$. On the other hand, there are at most $\|S_r(\mathcal{H})\|^{k\choose r}$ syntactic $k$-diagrams $\sigma$ with $V(\sigma)=X$, so $\alpha\leq \|S_r(\mathcal{H})\|^{k\choose r}\leq \|S_r(\mathcal{L})\|^{k\choose r}$. We now make a key observation about this hypergraph. Observe that by definition of $H$, if $S\subseteq V(H)$, then \[\begin{aligned} \label{ob2} E(H[S])= \Big(Diag^{tp}(cl_k(\mathcal{F}), C_W)\cup Err_k(C_W)\Big)\cap Span(S). \end{aligned}\] We are now ready to prove Theorem [\[VERSION2\]](#VERSION2){reference-type="ref" reference="VERSION2"}. At this point the reader may wish to review Theorem [\[containers\]](#containers){reference-type="ref" reference="containers"} and its corresponding notation in Subsection [2.2](#containerssec){reference-type="ref" reference="containerssec"}, as Theorem [\[VERSION2\]](#VERSION2){reference-type="ref" reference="VERSION2"} is the key tool used in this proof. Clearly it suffices to show the theorem holds for all $0<\epsilon<1/2$. We claim that further, it suffices to show the theorem holds for any $k\geq 2r$. Indeed, suppose $k<2r$ and Theorem [\[VERSION2\]](#VERSION2){reference-type="ref" reference="VERSION2"} holds for all $k'\geq 2r$. Suppose $\mathcal{F}$ is a nonempty collection of $\mathcal{L}$-structures, each of size at most $k$ and $\mathcal{H}:=\operatorname{Forb}(\mathcal{F})\neq \emptyset$. Then $\mathcal{F}$ is also a collection of $\mathcal{L}$-structures, each of size at most $k'=2r$. Apply Theorem [\[VERSION2\]](#VERSION2){reference-type="ref" reference="VERSION2"} to $k'=2r$ to obtain constants $c=c(2r,r,\mathcal{L}, \epsilon)$ and $m=m(2r, r)$. Since $k<2r$, it is clear the conclusions of Theorem [\[VERSION2\]](#VERSION2){reference-type="ref" reference="VERSION2"} for $\mathcal{H}$ and $2r$ imply the conclusions of Theorem [\[VERSION2\]](#VERSION2){reference-type="ref" reference="VERSION2"} for $\mathcal{H}$ and $k$. Thus we may take $c(k,r,\epsilon)=c(2r,r,\epsilon)$ and $m(k,r)=m(2r,r)$. We now prove the theorem holds for all $0<\epsilon<1/2$ and $k\geq 2r$. Fix $0<\epsilon<1/2$ and $k\geq 2r$. Apply Theorem [\[containers\]](#containers){reference-type="ref" reference="containers"} to $s={k\choose r}$ to obtain the constant $c_0=c_0({k\choose r})$ and set \[m=m(k,r)=\max\Big\{ \frac{{\ell\choose r}-1}{\ell-r}: r< \ell \leq k \Big\}.\] By Lemma [\[m\]](#m){reference-type="ref" reference="m"}, since $2\leq r<k$, $m> 1$. Set $\epsilon'=\epsilon/\|S_r(\mathcal{L})\|^{k\choose r}$ and choose $0<\gamma<1$ sufficiently small so that \[\begin{aligned} \label{gamma} 2^{{{k\choose r}\choose 2}+1}\|S_r(\mathcal{L})\|r!(k-r)^{k-r} \gamma &\leq \frac{\epsilon'}{12{k\choose 2}!}. \end{aligned}\] Now set $c=c(k,r,\mathcal{L},\epsilon)=(c_0\log(\frac{1}{\epsilon'})\|S_r(\mathcal{L})\|)/\gamma m$. Observe that $c$ actually depends on $\mathcal{L}$, not just $r_{\mathcal{L}}$. Let $M\geq k$ be such that $n\geq M$ implies $(n-r)^{k-r}\geq n^{k-r}/2$, and $n^{-\frac{1}{m}}\gamma^{-1}<1/2$. We show Theorem [\[VERSION2\]](#VERSION2){reference-type="ref" reference="VERSION2"} holds for this $c$ and $m$ for all $n\geq M$. Fix $n\geq M$. Suppose $\mathcal{F}$ a nonempty collection of finite $\mathcal{L}$-structures, each of size at most $k$, such that $\mathcal{H}:=\operatorname{Forb}(\mathcal{F})\neq \emptyset$. Let $W$ be a set of size $n$ and let $H=H(\mathcal{F}, W)$ be the ${k\choose r}$-uniform hypergraph described above. Set $\tau=n^{\frac{-1}{m}}\gamma^{-1}$. By our assumptions we have that $0<\epsilon', \tau<\frac{1}{2}$. We show that $\delta(H,\tau)\leq \frac{\epsilon'}{12{k\choose r}!}$ so that we may apply Theorem [\[containers\]](#containers){reference-type="ref" reference="containers"} to $H$. Let $\alpha=\alpha(\mathcal{F})$ be as in Definition [\[alpha\]](#alpha){reference-type="ref" reference="alpha"} so that $E(H)=\alpha{n\choose k}$ and let $N=\|V(H)\|$. Given $2\leq j\leq {k\choose r}$, set \[\begin{aligned} \label{fdef} f(j)=\min\{ \ell: {\ell \choose r}\geq j\}. \end{aligned}\] Observe that for each $2\leq j\leq {k\choose r}$, $r<f(j)\leq k$. Indeed, $r<f(j)$ holds since ${f(j)\choose r}\geq j\geq 2$, and $f(j)\leq k$ holds since $k\in \{\ell: {\ell \choose r}\geq j\}$. Thus by definition of $m$, for each $2\leq j\leq {k\choose r}$, \[\begin{aligned} \label{minequality} m\geq \frac{{f(j)\choose r}-1}{f(j)-r} \geq \frac{j-1}{f(j)-r}, \end{aligned}\] where the inequality is because by ([\[fdef\]](#fdef){reference-type="ref" reference="fdef"}), ${f(j)\choose r}\geq j$. We now show that for all $\sigma \subseteq V(H)$ with $2\leq \|\sigma\|\leq {k\choose r}$, $d(\sigma)\leq \alpha n^{k-f(\|\sigma\|)}$. Fix $\sigma\subseteq V(H)$ so that $2\leq \|\sigma\|\leq {k\choose r}$. Observe that if $\|V(\sigma)\|>k$, then $\{e\in E(H): \sigma\subseteq e\}=\emptyset$ since every $e\in E(H)$ is a syntactic $k$-diagram, so must satisfy $\|V(e)\|=k$. So in this case $d(\sigma)=0\leq \alpha n^{k-f(\|\sigma\|)}$. Similarly, if there is $A\in {V(\sigma)\choose r}$ such that $\|Ch_{\sigma}(A)\|\geq 2$, then $\{e\in E(H): \sigma\subseteq e\}=\emptyset$, since every $e\in E(H)$ is a syntactic $k$-diagram, so must satisfy $\|Ch_e(A)\|=1$. So in this case, $d(\sigma)=0\leq \alpha n^{k-f(\|\sigma\|)}$. Suppose now $\|V(\sigma)\|\leq k$ and $\|Ch_{\sigma}(A)\|\leq 1$ for all $A\in {V(\sigma)\choose r}$. This implies $\|\sigma\|\leq {\|V(\sigma)\|\choose r}$, so by ([\[fdef\]](#fdef){reference-type="ref" reference="fdef"}), $f(\|\sigma\|)\leq \|V(\sigma)\|$. Every edge in $H$ containing $\sigma$ can be constructed as follows. - Choose a set $X\in {C_W\choose k}$ such that $V(\sigma)\subseteq X$ (this makes sense since $\|V(\sigma)\|\leq k$). There are ${n-\|V(\sigma)\|\choose k-\|V(\sigma)\|}$ ways to do this. - Choose an element of $Diag^{tp}(cl_k(\mathcal{F}), X) \cup Err_k(X)$ containing $\sigma$. By definition of $\alpha$, there are at most $\alpha$ choices for this. Therefore, $d(\sigma)=\|\{ e\in E(H): \sigma\subseteq e\}\| \leq \alpha {n-\|V(\sigma)\|\choose k-\|V(\sigma)\|}\leq \alpha n^{k-\|V(\sigma)\|}\leq \alpha n^{k-f(\|\sigma\|)}$, where the last inequality is because $f(\|\sigma\|)\leq \|V(\sigma)\|$. Thus we have shown that for any $2\leq j\leq {k\choose r}$ and $\sigma\subseteq V(H)$ with $\|\sigma\|=j$, $d(\sigma)\leq \alpha n^{k-f(j)}$. Thus given $2\leq j\leq {k\choose r}$ and a vertex $v\in V(H)$, \[d^{(j)}(v)=\max \{ d(\sigma): v\in \sigma, \|\sigma\|=j\}\leq \alpha n^{k-f(j)}.\] Note the average degree of $H$ is \[d={k\choose r}\|E(H)\|/\|V(H)\| = \frac{{k\choose r}\alpha {n\choose k}}{\|S_r(\mathcal{H})\|{n\choose r}}=\frac{\alpha}{\|S_r(\mathcal{H})\|}{n-r\choose k-r}\geq \frac{\alpha}{\|S_r(\mathcal{H})\|}\Big(\frac{n-r}{k-r}\Big)^{k-r}.\] Combining this with our assumption $n$, we obtain the following inequality. \[\begin{aligned} \label{dbound} d\geq \frac{\alpha}{\|S_r(\mathcal{H})\|}\Big(\frac{n-r}{k-r}\Big)^{k-r}=\frac{\alpha}{\|S_r(\mathcal{H})\|(k-r)^{k-r}} (n-r)^{k-r}\geq \frac{\alpha}{2\|S_r(\mathcal{H})\|(k-r)^{k-r}} n^{k-r}. \end{aligned}\] Combining all of these computations we have the following. \[\delta_j=\frac{\sum_{v\in V(H)} d^{(j)}(v)}{Nd\tau^{j-1}}\leq \frac{Nn^{k-f(j)}}{Nd\tau^{j-1}}= \frac{n^{k-f(j)+(j-1)\frac{1}{m}}\gamma^{j-1}}{d}.\] Using our lower bound for $d$ from ([\[dbound\]](#dbound){reference-type="ref" reference="dbound"}), this implies \[\begin{aligned} \delta_j \leq 2\|S_r(\mathcal{H})\|(k-r)^{k-r} \gamma^{j-1}\alpha^{-1} n^{k-f(j)+\frac{j-1}{m}-k+r}=2\|S_r(\mathcal{H})\|(k-r)^{k-r} \gamma^{j-1}\alpha^{-1} n^{r-f(j)+\frac{j-1}{m}}. \end{aligned}\] By ([\[minequality\]](#minequality){reference-type="ref" reference="minequality"}), $\frac{j-1}{m}\leq f(j)-r$, so this implies $\delta_j$ is at most \[2\|S_r(\mathcal{H})\|(k-r)^{k-r} \gamma^{j-1}\alpha^{-1} n^{r-f(j)+f(j)-r}= 2\|S_r(\mathcal{H})\|(k-r)^{k-r} \gamma^{j-1}\alpha^{-1}\leq 2\|S_r(\mathcal{H})\|(k-r)^{k-r} \gamma,\] where the last inequality is because $j\geq 2$ and $\gamma<1$ implies $\gamma^{j-1}<\gamma$ and $\mathcal{F}\neq \emptyset$ implies $\alpha^{-1}\leq 1$. Therefore \[\begin{aligned} \label{delta1} \delta(H,\tau)&=2^{{{k\choose r}\choose 2}-1}\sum_{j=2}^{{k\choose r}}2^{-{j-1\choose 2}} \delta_j\leq 2^{{{k\choose r}\choose 2}-1}2\|S_r(\mathcal{H})\|(k-r)^{k-r} \gamma \sum_{j=2}^{{k\choose r}}2^{-{j-1\choose 2}}. \end{aligned}\] If $t={{k\choose r}\choose 2}$, then $\sum_{j=2}^{{k\choose r}}2^{-{j-1\choose 2}}\leq\sum_{j=0}^t2^{-t}$. Using the formula for summing finite geometric series, $\sum_{j=0}^t2^{-t}=\frac{1-2^{-t-1}}{1-2^{-1}}=2(1-2^{-t-1})<2$. Plugging this into ([\[delta1\]](#delta1){reference-type="ref" reference="delta1"}) yields \[\begin{aligned} \delta(H,\tau)&\leq 2^{{{k\choose r}\choose 2}-1}2\|S_r(\mathcal{H})\|(k-r)^{k-r} \gamma 2=2^{{{k\choose r}\choose 2}+1}\|S_r(\mathcal{H})\|(k-r)^{k-r} \gamma\leq 2^{{{k\choose r}\choose 2}+1}\|S_r(\mathcal{L})\|(k-r)^{k-r} \gamma. \end{aligned}\] By ([\[gamma\]](#gamma){reference-type="ref" reference="gamma"}), the right hand side above is at most $\frac{\epsilon'}{12{k\choose r}!}$, so we have shown that $\delta(H,\tau) \leq \frac{\epsilon'}{12{k\choose r}!}$. Thus Theorem [\[containers\]](#containers){reference-type="ref" reference="containers"} implies there is $\Sigma_0\subseteq \mathcal{P}(V(H))$ with the following properties. (i) For every independent set $I\subseteq V(H)$, there is $\sigma \in \Sigma_0$ such that $I \subseteq \sigma$. (ii) For every $\sigma \in \Sigma_0$, $e(H[\sigma])\leq \epsilon' e(H)$. (iii) $\log \|\Sigma_0\| \leq c_0\log(1/\epsilon') N\tau \log (1/\tau)$. Define $\Sigma=\{\sigma \in \Sigma_0: \exists \text{ an $\mathcal{F}$-free $\mathcal{L}$-structure $M$ with domain $W$ such that $Diag^{tp}(M)\subseteq \sigma \}$}$. Observe that every $\sigma \in \Sigma$ is complete by part 3 of Observation [\[ob1\]](#ob1){reference-type="ref" reference="ob1"}. We show $\Sigma$ satisfies the conclusions (1)-(3) of Theorem [\[VERSION2\]](#VERSION2){reference-type="ref" reference="VERSION2"}. Suppose $M$ is an $\mathcal{F}$-free $\mathcal{L}$-structure with domain $W$. Proposition [\[part1prop\]](#part1prop){reference-type="ref" reference="part1prop"} implies $Diag^{tp}(M)$ is an independent subset of $V(H)$, so by (i) and definition of $\Sigma$, there is $\sigma\in \Sigma$ such that $Diag^{tp}(M)\subseteq \sigma$. Thus part (1) of Theorem [\[VERSION2\]](#VERSION2){reference-type="ref" reference="VERSION2"} holds. We now show part (2) holds. Fix $\sigma\in \Sigma$. By ([\[ob2\]](#ob2){reference-type="ref" reference="ob2"}), $(Diag^{tp}(cl_k(\mathcal{F}), C_W)\cup Err_k(C_W))\cap Span(\sigma)=E(H[\sigma])$. So (ii) implies \[e(H[\sigma])=\|(Diag^{tp}(cl_k(\mathcal{F}), C_W)\cup Err_k(C_W))\cap Span(\sigma)\|\leq \epsilon' e(H).\] By definition of $\epsilon'$ and because $\alpha \leq \|S_r(\mathcal{L})\|^{k\choose r}$, \[\epsilon'e(H)=\epsilon'\alpha{n\choose k}=\frac{\epsilon}{\|S_r(\mathcal{L})\|^{k\choose r}}\alpha {n\choose k}\leq \epsilon {n\choose k}.\] Thus $\|(Diag^{tp}(cl_k(\mathcal{F}), C_W)\cup Err_k(C_W))\cap Span(\sigma)\|\leq \epsilon{n\choose k}$. By Lemma [\[lemma\\\]](#lemma){reference-type="ref" reference="lemma"}, for all $1\leq \ell \leq k$, \[\|(Diag^{tp}(\mathcal{F}(\ell),C_W)\cup Err_{\ell}(C_W))\cap Span (\sigma)\|\leq \epsilon {n\choose \ell}.\] Since $k\geq 2r$, this immediately implies part (2) of Theorem [\[VERSION2\]](#VERSION2){reference-type="ref" reference="VERSION2"} holds. By (iii), definition of $c$, and because $\Sigma\subseteq \Sigma_0$ we have that \[\begin{aligned} \|\Sigma\|\leq \|\Sigma_0\|\leq c_0\log(1/\epsilon') N\tau \log (1/\tau)&=c_0\log(1/\epsilon') \|S_r(\mathcal{H})\|{n\choose r}\tau \log (1/\tau) \\ &\leq c_0\log(1/\epsilon') \|S_r(\mathcal{L})\|{n\choose r}\tau \log (1/\tau)=c\gamma m{n\choose r}\tau \log\Big(\frac{1}{\tau}\Big). \end{aligned}\] This shows $\|\Sigma\|\leq c\gamma mn^r\tau \log(\frac{1}{\tau})$. By definition of $\tau$, \[c\gamma mn^r\tau \log\Big(\frac{1}{\tau}\Big)=c mn^{r-\frac{1}{m}}\Big(\frac{\log n}{m}+\log\gamma\Big)= c n^{r-\frac{1}{m}}\Big(\log n+m\log\gamma\Big)\leq cn^{r-\frac{1}{m}}\log n,\] where the last inequality is because $\gamma \leq 1\leq m$ implies $m\log \gamma \leq 0$. This shows part (3) of Theorem [\[VERSION2\]](#VERSION2){reference-type="ref" reference="VERSION2"} holds, so we are done. 0◻ # Metric Spaces {#ms} Given integers $r\geq 3$ and $n\geq 1$, let $M_r(n)$ be the set of metric spaces with distances all in $[r]$ and underlying set $[n]$. In this section we will reprove certain structure and enumeration theorems about $M_r(n)$ from using the machinery of this paper, along with combinatorial ingredients from. We include this example because it demonstrates interesting behavior with regards to the existence of a stability theorem. In particular, we will prove that when $r$ is even, the hereditary property associated to $\bigcup_{n\in \mathbb{N}}M_r(n)$ has a stability theorem in the sense of Definition [\[stabdef\]](#stabdef){reference-type="ref" reference="stabdef"}, but when $r$ is odd, this is not the case. We would like to point out that in order to be consistent with, $r$ will always denote the largest distance appearing in our metric spaces. This $r$ has nothing to do with our use of the letter $r$ as shorthand for $r_{\mathcal{L}}$ throughout the rest of the paper. In this section all languages will binary (i.e. $r_{\mathcal{L}}=2$), so we do not think any confusion will arise. ## Results from. In this subsection we state results from. For the rest of this section, $r\geq 3$ is a fixed integer. We require some notation and definitions in order to state the relevant theorems from. An $[r]$-graph (respectively an $2^{[r]}$-graph) is a pair $(V,c)$ such that $V$ is a set of vertices and $c:{V\choose 2}\rightarrow [r]$ (respectively $c:{V\choose 2}\rightarrow 2^{[r]}$) is a function. A $2^{[r]}$-graph $(V,c)$ is complete if for all $xy\in {V\choose 2}$, $\|c(xy)\|\geq 1$. Two $[r]$-graphs (respectively $2^{[r]}$-graphs) $(V,d)$ and $(V',d')$ are isomorphic if there is a bijection $f:V\rightarrow V'$ such that for all $xy\in {V\choose 2}$, $d(xy)=d'(f(x)f(y))$. A violating triple is a tuple $(i,j,k)\in \mathbb{N}^3$ such that $\|i-j\|\leq k\leq i+j$ is false. An $[r]$-graph $G=(V,d)$ is a metric $[r]$-graph if for all $\{x,y,z\}\in {V\choose 3}$, $(d(xy), d(yz),d(xz))$ is not a violating triple. A $2^{[r]}$-graph $G=(V,c)$ is a metric $2^{[r]}$-graph for all $\{x,y,z\}\in {V\choose 3}$, no $(i,j,k)\in c(xy)\times c(yz)\times c(xz)$ is a violating triple. Observe that if $R$ is a complete $2^{[r]}$-graph, then $R$ is a metric $2^{[r]}$-graph if and only if all its full sub-$[r]$-graphs are metric $[r]$-graphs. Given integers $i<j$, set $[i,j]=\{i,i+1, \ldots, j\}$. Set \[m(r)=\Bigg\lceil \frac{r+1}{2}\Bigg\rceil.\] If $r$ is odd, let $L_r=[ \frac{r-1}{2}, r-1]$ and $U_r=[\frac{r+1}{2}, r]$. Observe that if $r$ is odd, then $m(r)=\|U_r\|=\|L_r\|$ and if $r$ is even, then $m(r)=\|[\frac{r}{2},r]\|$. We now define a special subfamily of $M_r(n)$. It is straightforward to check that in both even and odd cases, $C_r(n)\subseteq M_r(n)$. Set $C_r^{\delta}(n)=\{G\in M_r(n): \textnormal{there is some }G' \in C_r(n)\textnormal{ such that }\|\Delta(G,G')\|\leq \delta{\|V\|\choose 2} \}$. We can now state the approximate structure theorem from. Informally, it states that for all $\delta>0$, most members of $M_r(n)$ are in $C_r^{\delta}(n)$ when $n$ is sufficiently large depending on $\delta$. We now state the approximate enumeration theorem from. In fact, contains much more precise structural and enumerative results for $M_r(n)$ in the case when $r$ is even. Finding similar refinements of Theorem [\[DELTACLOSETHM\]](#DELTACLOSETHM){reference-type="ref" reference="DELTACLOSETHM"} and Corollary [\[oddcor\]](#oddcor){reference-type="ref" reference="oddcor"} in the case when $r$ is odd is still open. This suggests there is something "nicer" about the even case than the odd case. We will show in this section that one candidate for what makes the even case "nice" is that when $r$ is even, the hereditary property corresponding to $M_r(n)$ has a stability theorem in the sense of Definition [\[stabdef\]](#stabdef){reference-type="ref" reference="stabdef"}, while in the odd case it does not. ## Translation to hereditary $\mathcal{L}$-property. In this subsection we translate some of the combinatorial notions used in to the setup used in this paper. Recall $r\geq 3$ is a fixed integer. Let $\mathcal{L}_r=\{R_1(x,y),\ldots, R_r(x,y)\}$ consist of $r$ binary relation symbols. We consider elements $(V,d)$ of $M_r(n)$ as $\mathcal{L}_r$-structures by interpreting $R_i(x,y)$ if and only if $d(xy)=i$, for each $(x,y)\in V^2$. Let $\mathcal{M}_r$ denote the class of $\mathcal{L}_r$-structures obtained by closing $M_r=\bigcup_{n\in \mathbb{N}}M_r(n)$ under isomorphism. Clearly $\mathcal{M}_r$ is a hereditary $\mathcal{L}_r$-property, and $(\mathcal{M}_r)_n=M_r(n)$. For the rest of the section, $\mathcal{P}=\mathcal{M}_r$, and $\mathcal{L}=\mathcal{L}_r$. Observe that since $r_{\mathcal{L}}=2$, $\mathcal{L}_{\mathcal{P}}=\{R_p(x,y): p(x,y)\in S_2(\mathcal{P})\}$. For each $i\in [r]$, set \[q_i(x,y):=\{x\neq y\}\cup \{R_i(x,y), R_i(y,x)\}\cup \{\neg R_j(x,y), \neg R_j(y,x): j\in [r]\setminus \{i\}\},\] and let $p_i(x,y)$ be the unique quantifier-free $2$-type containing $q_i(x,y)$. In other words, $p_i(x,y)$ is the type saying the distance between $x$ and $y$ is $i$. Clearly $S_2(\mathcal{P})=\{p_i(x,y): i\in [r]\}$, so $\mathcal{L}_{\mathcal{P}}=\{R_{p_i}(x,y): i\in [r]\}$. The following observation will be useful and is obvious from the definition of a choice function. Given a $2^{[r]}$-graph $(V,c)$, define $\Psi^{-1}(V,c)$ to be the $\mathcal{L}_{\mathcal{P}}$-structure $G$ with domain $V$ such that for each $(x,y)\in V^2$ and $i\in [r]$, $G\models R_{p_i}(x,y)$ if and only if $x\neq y$ and $i\in c(xy)$. We leave the following observations to the reader. Suppose $G\in M_r(n)$ is the $[r]$-graph $(V,d)$ considered as an $\mathcal{L}$-structure. We will often abuse notation by conflating $G$ and $(V,d)$. For instance if $R$ is a $2^{[r]}$-graph, we will write $G\subseteqq_{[r]}R$ to mean $(V,d)\subseteqq_{[r]}R$. Similarly if $R$ is an $\mathcal{L}_{\mathcal{P}}$-template, we will write $(V,d)\unlhd_pR$ to mean $G\unlhd_pR$. Given $n$, let $\tilde{M}_r(n)$ be the set of complete metric $2^{[r]}$-graphs on $[n]$. We now recall a definition from. If $G=(V,c)$ is a $2^{[r]}$-graph, then $W(R)=\prod_{xy\in {V\choose 2}}\|c(xy)\|$. If $G\in \tilde{M}_r(n)$, say that $G$ is product-extremal if \[W(G)=\max\{W(G'): G'\in \tilde{M}_r(n)\}.\] ## Characterizing extremal structures and computing $\pi(\mathcal{P})$. In this subsection, we characterize product-extremal elements of $\tilde{M}_r(n)$. These results, Lemmas [\[extlemeven\]](#extlemeven){reference-type="ref" reference="extlemeven"} and [\[extlemodd\]](#extlemodd){reference-type="ref" reference="extlemodd"}, are new results. We will then use the correspondence between product-extremal elements of $\tilde{M}_r(n)$ and elements of $\mathcal{R}_{ex}([n],\mathcal{P})$ from the preceding section to compute $\pi(\mathcal{P})$. We begin by defining a special subfamily of $\tilde{M}_r(n)$ corresponding to the special subfamily $C_r(n)$ of $M_r(n)$. Note that for all $n$, $\tilde{C}_r(n)\subseteq \tilde{M}_r(n)$ and when $r$ is even, $\tilde{C}_r(n)$ consists of a single $2^{[r]}$-graph. We will need a few results about multigraphs. A multigraph is a pair $(V,w)$ where $V$ is a set of vertices and $w:{V\choose 2}\rightarrow \mathbb{N}$ is a function. Given integers $s\geq 2$ and $q\geq 0$, an $(s,q)$-graph is a multigraph $(V,w)$ such that for all $X\in {V\choose s}$, $\sum_{xy\in {X\choose 2}}w(xy)\leq q$. Given a multigraph $G=(V,w)$, set $P(G)=\prod_{xy\in {V\choose 2}}w(xy)$. Observe that if $G\in U_{1,a}(n)$, then $P(G)=a^{n\choose 2}$ and if $G\in U_{2,a}(n)$, then $P(G)=a^{n\choose 2}(\frac{a+1}{a})^{\lfloor n/2\rfloor }$. In the following lemma we use a result of Balogh and Wagner from. In particular, we use Lemma 3.2 of, which is a corollary of combinatorial results in. If $r\geq 3$ is odd and $n\geq 3$, define $\tilde{E}_r(n)$ to be the set of $2^{[r]}$-graphs $([n],c)$ such that there is a set $\{e_1,\ldots, e_{\lfloor n/2\rfloor}\}$ of pairwise disjoint elements in ${[n]\choose 2}$ such that $c(e_i)=U_r\cup L_r$ and if $xy\in {[n]\choose 2}\setminus \{e_1,\ldots, e_{\lfloor n/2\rfloor}\}$ then $c(xy)=U_r$. Observe that $\tilde{E}_r(n)\subseteq \tilde{C}_r(n)$ and \[\begin{aligned} \label{Efact} \text{ for any $G\in \tilde{E}_r(n)$, }W(G)=m(r)^{n\choose r}\Big(\frac{m(r)+1}{m(r)}\Big)^{\lfloor n/2 \rfloor}. \end{aligned}\] ## Proofs of Results In this subsection, we prove Corollary [\[oddcor\]](#oddcor){reference-type="ref" reference="oddcor"} and Theorem [\[DELTACLOSETHM\]](#DELTACLOSETHM){reference-type="ref" reference="DELTACLOSETHM"}. We also prove that if $r$ is even then $\mathcal{P}$ has a stability theorem, and if $r$ is odd, then $\mathcal{P}$ does not have a stability theorem. These results, Lemma [\[msstabeven\]](#msstabeven){reference-type="ref" reference="msstabeven"} and Corollary [\[msstabodd\]](#msstabodd){reference-type="ref" reference="msstabodd"}, are new. We begin by proving Corollary [\[oddcor\]](#oddcor){reference-type="ref" reference="oddcor"}. Theorem [\[enumeration\]](#enumeration){reference-type="ref" reference="enumeration"} implies $\|\mathcal{P}_n\|=\pi(\mathcal{P})^{(1+o(1)){n\choose 2}}$. By definition of $\mathcal{P}$, $M_r(n)=\mathcal{P}_n$, and Corollary [\[mspi\]](#mspi){reference-type="ref" reference="mspi"} implies $\pi(\mathcal{P})=m(r)$. Thus $\|\mathcal{P}_n\|=\|M_r(n)\|=m(r)^{(1+o(1)){n\choose 2}}$. 0◻ We now state a stability style result, Theorem 4.13 from. The following is straightforward from the definitions. We can now prove Theorem [\[DELTACLOSETHM\]](#DELTACLOSETHM){reference-type="ref" reference="DELTACLOSETHM"}. Fix $\delta>0$. Choose $\epsilon>0$ and $M_1$ from Theorem [\[stabthm\]](#stabthm){reference-type="ref" reference="stabthm"} such that if $n>M_1$ and $H\in \tilde{M}_r(n)$ satisfies $W(H)>m(r)^{(1-\epsilon){n\choose 2}}$ then $H$ is $\delta/2$-close in the sense of Definition [\[newdeltaclose\]](#newdeltaclose){reference-type="ref" reference="newdeltaclose"} to an element of $\tilde{C}_r(n)$. Now let $\beta>0$ and $M_2$ be as in Theorem [\[b4stab\]](#b4stab){reference-type="ref" reference="b4stab"} applied to $\delta/2$ and $\epsilon$. Let $N=\max\{M_1,M_2\}$. We show for all $n>N$, \[\frac{\|M_r(n)\setminus C^{\delta}_r(n)\|}{\|M_r(n)}\|\leq 2^{-\beta n^2}.\] By Theorem [\[b4stab\]](#b4stab){reference-type="ref" reference="b4stab"}, it suffices to show that for all $n\geq N$, $E^{\delta/2}(\epsilon, n, \mathcal{P})\subseteq C^{\delta}_r(n)$. Fix $n\geq N$ and suppose $H=(V,d)\in E^{\delta/2}(\epsilon, n, \mathcal{P})$. We want to show $H\in C^{\delta}_r(n)$. By definition of $E^{\delta/2}(\epsilon, n, \mathcal{P})$, there is some $H'=(V,d')\in E(\epsilon, n, \mathcal{P})$ such that $\textnormal{dist}(H,H')\leq \delta/2$. By Observation [\[ob8\]](#ob8){reference-type="ref" reference="ob8"}, this implies $\|\Delta(H,H')\|\leq \delta {n\choose 2}$. By definition of $E(\epsilon, n,\delta)$, there is $G'\in \mathcal{R}([n],\mathcal{P})$ such that $H'\unlhd_pG'$ and $sub(G')>ex(n,\mathcal{P})^{1-\epsilon}$. Recall that by Observation [\[ob5\]](#ob5){reference-type="ref" reference="ob5"}(a), for all $n$, $\textnormal{ex}(n,\mathcal{P})\geq \pi(\mathcal{P})^{{n\choose 2}}$, so this implies $sub(G')>\pi(\mathcal{P})^{(1-\epsilon){n\choose 2}}$. Corollary [\[cor1\]](#cor1){reference-type="ref" reference="cor1"} implies $\Psi(G'):=(V,c')\in \tilde{M}_r(n)$, and Proposition [\[0\]](#0){reference-type="ref" reference="0"} implies $W(\Psi(G'))=sub(G')>\pi(\mathcal{P})^{(1-\epsilon){n\choose 2}}$. Consequently, Theorem [\[stabthm\]](#stabthm){reference-type="ref" reference="stabthm"} implies $\Psi(G')$ is $\delta/2$-close in the sense of Definition [\[newdeltaclose\]](#newdeltaclose){reference-type="ref" reference="newdeltaclose"} to some $M=([n],c)\in \tilde{C}_r(n)$. Define $H''=([n],d'')$ as follows. 1. If $xy \notin \Delta(\Psi(G'), M) \cup \Delta (H,H')$, let $d''(xy)=d'(xy)=d(xy)$. 2. If $xy\in \Delta(\Psi(G'), M)\cup \Delta (H,H')$, let $d''(xy)$ be any element of $c(xy)$. We claim $H''\subseteqq_{[r]}M$. Fix $xy\in {[n]\choose 2}$. We want to show $d''(xy)\in c(xy)$. This holds by definition of $H''$ when $xy\in \Delta(\Psi(G'), G'')\cup \Delta (H,H')$. If $xy\not\in \Delta(\Psi(G'), M)\cup \Delta (H,H')$, then $d''(xy)=d'(xy)$ and $c(xy)=c'(xy)$. Since, $H'\unlhd_pG'$, Proposition [\[translation\]](#translation){reference-type="ref" reference="translation"} implies $H'\subseteqq_{[r]}\Psi(G')$, thus $d'(xy)\in c'(xy)$. Since $d''(xy)=d'(xy)$ and $c(xy)=c'(xy)$, this implies $d''(xy)\in c(xy)$, as desired. Therefore, $H''\subseteqq_{[r]}M$ and $M\in \tilde{C}_r(n)$ implies $H''\in C_r(n)$. We claim $\|\Delta(H,H'')\|\leq \delta {n\choose 2}$. By definition of $H''$, \[\Delta(H,H'')\subseteq \Delta(H, H')\cup \Delta(\Psi(G'), M),\] so $\|\Delta(H,H'')\|\leq \|\Delta(H,H')\|+\|\Delta(\Psi(G'),M)\|\leq \delta {n\choose 2}$, where the inequality is by our assumptions. This shows $H''\in C_r(n)$ and $\|\Delta(H,H'')\|\leq \delta{n\choose 2}$, i.e. $H\in C_r^{\delta}(n)$. 0◻ We leave the following lemma to the reader. We now show that when $r$ is even, $\mathcal{P}$ has a stability theorem in the sense of Definition [\[stabdef\]](#stabdef){reference-type="ref" reference="stabdef"}, but when $r$ is odd, this is not the case. # Concluding Remarks {#end} We end with some questions and conjectures. Returning to the metric spaces of the previous section, it was shown in the following is true. It was then conjectured in that this theorem is false in the case when $r$ is odd (to our knowledge, this is still open). In the previous section, we showed that in the case when $r$ is even, $\mathcal{M}_r$ has a stability theorem in the sense of Definition [\[stabdef\]](#stabdef){reference-type="ref" reference="stabdef"}, while when $r$ is odd, this is false. These facts lead us to make the following conjecture. The idea behind this conjecture is that if $\mathcal{H}$ has nice asymptotic structure, it should reflect the structure of $\mathcal{R}_{ex}([n],\mathcal{H})$. Another phenomenon which can be observed from known examples is that the structures in $\mathcal{R}_{ex}(n,\mathcal{H})$ are not very complicated. The following questions are various ways of asking if this is always the case. We direct the reader to for the definition of a formula having the $k$-order property. A weaker version of this question is the following. We direct the reader to for the definition of the $VC$-dimension of a formula. [^1]: This use of the word stability refers to a type of theorem from extremal combinatorics and is unrelated to the model theoretic notion of stability. [^2]: In, $\delta$-closeness is instead defines as $\|\Delta(G,G')\|\leq \delta n^2$. For large $n$, this is basically the same as Definition [\[newdeltaclose\]](#newdeltaclose){reference-type="ref" reference="newdeltaclose"} up to a factor of $2$, and therefore doesn't change the statements of the results from. We have chosen to use Definition [\[newdeltaclose\]](#newdeltaclose){reference-type="ref" reference="newdeltaclose"} for convenience.	{'timestamp': '2016-07-20T02:11:53', 'yymm': '1607', 'arxiv_id': '1607.04902', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04902'}
null	null
# Introduction {#sec:introduction} Let $G$ be a simple graph without isolated vertices. If the edge set of a simple graph $K$ can be partitioned into edge sets of graphs each isomorphic to $G$, we say that there exists a decomposition of $K$ into $G$. In the case where $K$ is the complete graph $K_n$ we refer to the decomposition as a $G$ design of order $n$. The spectrum of $G$ is the set of positive integers $n$ for which there exists a $G$ design of order $n$. We refer the reader to the survey article of Adams, Bryant and Buchanan, and, for more up to date results, the Web site maintained by Bryant and McCourt,. A snark is a connected, bridgeless 3-regular graph with chromatic index 4. However, a snark is usually regarded as trivial (or reducible) if it has girth less than 5 or if it has three edges the deletion of which results in a disconnected graph each of whose components is non-trivial (as a graph). The spectra for non-trivial snarks of up to 22 vertices have been successfully determined, as follows. 1. The smallest non-trivial snark is the 10-vertex Petersen graph for which designs of order $n$ exist if and only if $n \equiv 1$ or 10 (mod 15),. 2. There are no non-trivial snarks on 12, 14 or 16 vertices. 3. There are two non-trivial snarks on 18 vertices, namely the $(1,1)$-and $(1,2)$-Blanuša snarks. For each of them, designs of order $n$ exist if and only if $n \equiv 1$ (mod 27),. 4. For each of the six 20-vertex non-trivial snarks (including the flower snark J5), designs of order $n$ exist if and only if $n \equiv 1,$ 16, 25, 40 (mod 60), $n \neq 16$,. 5. For each of the twenty 22-vertex non-trivial snarks (including the two Loupekine snarks), designs of order $n$ exist if and only if $n \equiv 1$ or 22 (mod 33),. The purpose of this paper is to extend these results to 24-vertex non-trivial snarks of which there are precisely 38,. We prove the following. [\[thm:snark24\]]{#thm:snark24 label="thm:snark24"} Designs of order $n$ exist for each of the thirty-eight non-trivial snarks on $24$ vertices if and only if $n \equiv 1$ or $64~(\mathrm{mod}~72)$.\ The methods use to obtain results like Theorems [\[thm:snark24\]](#thm:snark24){reference-type="ref" reference="thm:snark24"} are explained in, to which the reader should regard this paper as a sequel. The main tool is , repeated here for convenience as Proposition [\[prop:d=3, v=24\]](#prop:d=3, v=24){reference-type="ref" reference="prop:d=3, v=24"}. [\[prop:d=3, v=24\]]{#prop:d=3, v=24 label="prop:d=3, v=24"} Let $G$ be a 3-regular graph on $24$ vertices. Suppose there exist $G$ designs of order $64$, $73$, $136$ and $145$. Suppose also that there exist decompositions into $G$ of the complete multipartite graphs $K_{12,12,12}$, $K_{24,24,15}$, $K_{72,72,63}$, $K_{24,24,24,24}$ and $K_{24,24,24,21}$. Then there exists a $G$ design of order $n$ if and only if $n \equiv 1$ or $64~(\textup{mod}~ 72)$.\ Proof. See.\ To prove Theorem [\[thm:snark24\]](#thm:snark24){reference-type="ref" reference="thm:snark24"} it is sufficient, therefore, to construct the four designs and five multipartite graph decompositions required by Proposition [\[prop:d=3, v=24\]](#prop:d=3, v=24){reference-type="ref" reference="prop:d=3, v=24"} for each of the 38 non-trivial snarks. To save space the details of the constructions are given only for designs of order 136, which, in the author's opinion, were by far the most difficult to obtain. The details of the remaining constructions are given in the Appendix, which is present only in the full version of this paper, available by request from the author. The decompositions were obtained and checked by computer programs in the same manner as those of. # The 38 non-trivial snarks on 24 vertices {#sec:Snark24} The thirty-eight non-trivial snarks on 24 vertices are represented by ordered 24-tuples of vertices: (1, 2, ..., 24)${}_{\mathrm{G1}}$, (1, 2, ..., 24)${}_{\mathrm{G2}}$, ..., (1, 2, ..., 24)${}_{\mathrm{G38}}$. Their edge sets, as appearing in Royale's list,, are respectively $\mathrm{G1}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,8\}$, $\{6,11\}$, $\{7,12\}$, $\{8,13\}$, $\{9,14\}$, $\{10,11\}$, $\{10,12\}$, $\{11,15\}$, $\{12,16\}$, $\{13,14\}$, $\{13,17\}$, $\{14,18\}$, $\{15,16\}$, $\{15,19\}$, $\{16,20\}$, $\{17,21\}$, $\{17,22\}$, $\{18,23\}$, $\{18,24\}$, $\{19,21\}$, $\{19,23\}$, $\{20,22\}$, $\{20,24\}$, $\{21,24\}$, $\{22,23\}$}, $\mathrm{G2}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,8\}$, $\{6,11\}$, $\{7,12\}$, $\{8,13\}$, $\{9,14\}$, $\{10,11\}$, $\{10,15\}$, $\{11,16\}$, $\{12,15\}$, $\{12,16\}$, $\{13,14\}$, $\{13,17\}$, $\{14,18\}$, $\{15,19\}$, $\{16,20\}$, $\{17,21\}$, $\{17,22\}$, $\{18,23\}$, $\{18,24\}$, $\{19,21\}$, $\{19,23\}$, $\{20,22\}$, $\{20,24\}$, $\{21,24\}$, $\{22,23\}$}, $\mathrm{G3}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,8\}$, $\{6,11\}$, $\{7,12\}$, $\{8,13\}$, $\{9,14\}$, $\{10,11\}$, $\{10,15\}$, $\{11,16\}$, $\{12,15\}$, $\{12,16\}$, $\{13,17\}$, $\{13,18\}$, $\{14,19\}$, $\{14,20\}$, $\{15,21\}$, $\{16,22\}$, $\{17,19\}$, $\{17,23\}$, $\{18,20\}$, $\{18,24\}$, $\{19,24\}$, $\{20,23\}$, $\{21,22\}$, $\{21,23\}$, $\{22,24\}$}, $\mathrm{G4}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,8\}$, $\{6,11\}$, $\{7,12\}$, $\{8,13\}$, $\{9,14\}$, $\{10,11\}$, $\{10,15\}$, $\{11,16\}$, $\{12,15\}$, $\{12,17\}$, $\{13,18\}$, $\{13,19\}$, $\{14,20\}$, $\{14,21\}$, $\{15,22\}$, $\{16,17\}$, $\{16,22\}$, $\{17,23\}$, $\{18,20\}$, $\{18,23\}$, $\{19,21\}$, $\{19,24\}$, $\{20,24\}$, $\{21,23\}$, $\{22,24\}$}, $\mathrm{G5}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,8\}$, $\{6,11\}$, $\{7,12\}$, $\{8,13\}$, $\{9,14\}$, $\{10,12\}$, $\{10,15\}$, $\{11,12\}$, $\{11,16\}$, $\{13,14\}$, $\{13,17\}$, $\{14,18\}$, $\{15,16\}$, $\{15,19\}$, $\{16,20\}$, $\{17,21\}$, $\{17,22\}$, $\{18,23\}$, $\{18,24\}$, $\{19,21\}$, $\{19,23\}$, $\{20,22\}$, $\{20,24\}$, $\{21,24\}$, $\{22,23\}$}, $\mathrm{G6}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,8\}$, $\{6,11\}$, $\{7,12\}$, $\{8,13\}$, $\{9,14\}$, $\{10,12\}$, $\{10,15\}$, $\{11,15\}$, $\{11,16\}$, $\{12,16\}$, $\{13,14\}$, $\{13,17\}$, $\{14,18\}$, $\{15,19\}$, $\{16,20\}$, $\{17,21\}$, $\{17,22\}$, $\{18,23\}$, $\{18,24\}$, $\{19,21\}$, $\{19,23\}$, $\{20,22\}$, $\{20,24\}$, $\{21,24\}$, $\{22,23\}$}, $\mathrm{G7}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,8\}$, $\{6,11\}$, $\{7,12\}$, $\{8,13\}$, $\{9,14\}$, $\{10,12\}$, $\{10,15\}$, $\{11,15\}$, $\{11,16\}$, $\{12,16\}$, $\{13,17\}$, $\{13,18\}$, $\{14,19\}$, $\{14,20\}$, $\{15,21\}$, $\{16,22\}$, $\{17,19\}$, $\{17,23\}$, $\{18,20\}$, $\{18,24\}$, $\{19,24\}$, $\{20,23\}$, $\{21,22\}$, $\{21,23\}$, $\{22,24\}$}, $\mathrm{G8}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,8\}$, $\{6,11\}$, $\{7,12\}$, $\{8,13\}$, $\{9,14\}$, $\{10,12\}$, $\{10,15\}$, $\{11,15\}$, $\{11,16\}$, $\{12,17\}$, $\{13,18\}$, $\{13,19\}$, $\{14,20\}$, $\{14,21\}$, $\{15,22\}$, $\{16,17\}$, $\{16,23\}$, $\{17,22\}$, $\{18,20\}$, $\{18,23\}$, $\{19,21\}$, $\{19,24\}$, $\{20,24\}$, $\{21,23\}$, $\{22,24\}$}, $\mathrm{G9}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,8\}$, $\{6,11\}$, $\{7,12\}$, $\{8,13\}$, $\{9,14\}$, $\{10,12\}$, $\{10,15\}$, $\{11,16\}$, $\{11,17\}$, $\{12,16\}$, $\{13,18\}$, $\{13,19\}$, $\{14,20\}$, $\{14,21\}$, $\{15,17\}$, $\{15,22\}$, $\{16,22\}$, $\{17,23\}$, $\{18,20\}$, $\{18,23\}$, $\{19,21\}$, $\{19,24\}$, $\{20,24\}$, $\{21,23\}$, $\{22,24\}$}, $\mathrm{G10}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,8\}$, $\{6,11\}$, $\{7,12\}$, $\{8,13\}$, $\{9,14\}$, $\{10,15\}$, $\{10,16\}$, $\{11,15\}$, $\{11,17\}$, $\{12,15\}$, $\{12,18\}$, $\{13,19\}$, $\{13,20\}$, $\{14,21\}$, $\{14,22\}$, $\{16,17\}$, $\{16,18\}$, $\{17,23\}$, $\{18,24\}$, $\{19,21\}$, $\{19,23\}$, $\{20,22\}$, $\{20,24\}$, $\{21,24\}$, $\{22,23\}$}, $\mathrm{G11}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,8\}$, $\{6,11\}$, $\{7,12\}$, $\{8,13\}$, $\{9,14\}$, $\{10,15\}$, $\{10,16\}$, $\{11,15\}$, $\{11,17\}$, $\{12,15\}$, $\{12,18\}$, $\{13,19\}$, $\{13,20\}$, $\{14,21\}$, $\{14,22\}$, $\{16,18\}$, $\{16,23\}$, $\{17,18\}$, $\{17,24\}$, $\{19,21\}$, $\{19,23\}$, $\{20,22\}$, $\{20,24\}$, $\{21,24\}$, $\{22,23\}$}, $\mathrm{G12}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,8\}$, $\{6,11\}$, $\{7,12\}$, $\{8,13\}$, $\{9,14\}$, $\{10,15\}$, $\{10,16\}$, $\{11,15\}$, $\{11,17\}$, $\{12,16\}$, $\{12,18\}$, $\{13,19\}$, $\{13,20\}$, $\{14,21\}$, $\{14,22\}$, $\{15,18\}$, $\{16,17\}$, $\{17,23\}$, $\{18,24\}$, $\{19,21\}$, $\{19,23\}$, $\{20,22\}$, $\{20,24\}$, $\{21,24\}$, $\{22,23\}$}, $\mathrm{G13}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,8\}$, $\{6,11\}$, $\{7,12\}$, $\{8,13\}$, $\{9,14\}$, $\{10,15\}$, $\{10,16\}$, $\{11,17\}$, $\{11,18\}$, $\{12,15\}$, $\{12,17\}$, $\{13,19\}$, $\{13,20\}$, $\{14,21\}$, $\{14,22\}$, $\{15,18\}$, $\{16,17\}$, $\{16,23\}$, $\{18,24\}$, $\{19,21\}$, $\{19,23\}$, $\{20,22\}$, $\{20,24\}$, $\{21,24\}$, $\{22,23\}$}, $\mathrm{G14}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,11\}$, $\{6,12\}$, $\{7,10\}$, $\{8,11\}$, $\{8,13\}$, $\{9,13\}$, $\{10,14\}$, $\{11,15\}$, $\{12,14\}$, $\{12,16\}$, $\{13,17\}$, $\{14,18\}$, $\{15,17\}$, $\{15,19\}$, $\{16,20\}$, $\{16,21\}$, $\{17,22\}$, $\{18,23\}$, $\{18,24\}$, $\{19,20\}$, $\{19,23\}$, $\{20,24\}$, $\{21,22\}$, $\{21,23\}$, $\{22,24\}$}, $\mathrm{G15}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,11\}$, $\{6,12\}$, $\{7,10\}$, $\{8,11\}$, $\{8,13\}$, $\{9,14\}$, $\{10,15\}$, $\{11,14\}$, $\{12,15\}$, $\{12,16\}$, $\{13,17\}$, $\{13,18\}$, $\{14,17\}$, $\{15,19\}$, $\{16,20\}$, $\{16,21\}$, $\{17,22\}$, $\{18,20\}$, $\{18,23\}$, $\{19,23\}$, $\{19,24\}$, $\{20,24\}$, $\{21,22\}$, $\{21,23\}$, $\{22,24\}$}, $\mathrm{G16}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,11\}$, $\{6,12\}$, $\{7,10\}$, $\{8,11\}$, $\{8,13\}$, $\{9,14\}$, $\{10,15\}$, $\{11,16\}$, $\{12,15\}$, $\{12,17\}$, $\{13,14\}$, $\{13,18\}$, $\{14,16\}$, $\{15,19\}$, $\{16,20\}$, $\{17,21\}$, $\{17,22\}$, $\{18,21\}$, $\{18,23\}$, $\{19,23\}$, $\{19,24\}$, $\{20,22\}$, $\{20,24\}$, $\{21,24\}$, $\{22,23\}$}, $\mathrm{G17}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,11\}$, $\{6,12\}$, $\{7,10\}$, $\{8,11\}$, $\{8,13\}$, $\{9,14\}$, $\{10,15\}$, $\{11,16\}$, $\{12,15\}$, $\{12,17\}$, $\{13,18\}$, $\{13,19\}$, $\{14,20\}$, $\{14,21\}$, $\{15,16\}$, $\{16,22\}$, $\{17,22\}$, $\{17,23\}$, $\{18,20\}$, $\{18,23\}$, $\{19,21\}$, $\{19,24\}$, $\{20,24\}$, $\{21,23\}$, $\{22,24\}$}, $\mathrm{G18}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,11\}$, $\{6,12\}$, $\{7,10\}$, $\{8,11\}$, $\{8,13\}$, $\{9,14\}$, $\{10,15\}$, $\{11,16\}$, $\{12,15\}$, $\{12,17\}$, $\{13,18\}$, $\{13,19\}$, $\{14,20\}$, $\{14,21\}$, $\{15,22\}$, $\{16,17\}$, $\{16,22\}$, $\{17,23\}$, $\{18,20\}$, $\{18,23\}$, $\{19,21\}$, $\{19,24\}$, $\{20,24\}$, $\{21,23\}$, $\{22,24\}$}, $\mathrm{G19}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,11\}$, $\{6,12\}$, $\{7,10\}$, $\{8,11\}$, $\{8,13\}$, $\{9,14\}$, $\{10,15\}$, $\{11,16\}$, $\{12,17\}$, $\{12,18\}$, $\{13,19\}$, $\{13,20\}$, $\{14,16\}$, $\{14,19\}$, $\{15,21\}$, $\{15,22\}$, $\{16,20\}$, $\{17,21\}$, $\{17,23\}$, $\{18,22\}$, $\{18,24\}$, $\{19,23\}$, $\{20,24\}$, $\{21,24\}$, $\{22,23\}$}, $\mathrm{G20}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,11\}$, $\{6,12\}$, $\{7,10\}$, $\{8,13\}$, $\{8,14\}$, $\{9,13\}$, $\{10,15\}$, $\{11,14\}$, $\{11,15\}$, $\{12,16\}$, $\{12,17\}$, $\{13,18\}$, $\{14,19\}$, $\{15,16\}$, $\{16,20\}$, $\{17,21\}$, $\{17,22\}$, $\{18,21\}$, $\{18,23\}$, $\{19,22\}$, $\{19,24\}$, $\{20,23\}$, $\{20,24\}$, $\{21,24\}$, $\{22,23\}$}, $\mathrm{G21}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,11\}$, $\{6,12\}$, $\{7,10\}$, $\{8,13\}$, $\{8,14\}$, $\{9,13\}$, $\{10,15\}$, $\{11,14\}$, $\{11,15\}$, $\{12,16\}$, $\{12,17\}$, $\{13,18\}$, $\{14,19\}$, $\{15,20\}$, $\{16,21\}$, $\{16,22\}$, $\{17,23\}$, $\{17,24\}$, $\{18,19\}$, $\{18,21\}$, $\{19,23\}$, $\{20,22\}$, $\{20,24\}$, $\{21,24\}$, $\{22,23\}$}, $\mathrm{G22}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,11\}$, $\{6,12\}$, $\{7,10\}$, $\{8,13\}$, $\{8,14\}$, $\{9,13\}$, $\{10,15\}$, $\{11,14\}$, $\{11,16\}$, $\{12,15\}$, $\{12,17\}$, $\{13,18\}$, $\{14,19\}$, $\{15,20\}$, $\{16,21\}$, $\{16,22\}$, $\{17,21\}$, $\{17,23\}$, $\{18,19\}$, $\{18,23\}$, $\{19,24\}$, $\{20,22\}$, $\{20,24\}$, $\{21,24\}$, $\{22,23\}$}, $\mathrm{G23}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,11\}$, $\{6,12\}$, $\{7,10\}$, $\{8,13\}$, $\{8,14\}$, $\{9,13\}$, $\{10,15\}$, $\{11,14\}$, $\{11,16\}$, $\{12,15\}$, $\{12,17\}$, $\{13,18\}$, $\{14,19\}$, $\{15,20\}$, $\{16,21\}$, $\{16,22\}$, $\{17,21\}$, $\{17,23\}$, $\{18,19\}$, $\{18,24\}$, $\{19,23\}$, $\{20,22\}$, $\{20,24\}$, $\{21,24\}$, $\{22,23\}$}, $\mathrm{G24}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,11\}$, $\{6,12\}$, $\{7,10\}$, $\{8,13\}$, $\{8,14\}$, $\{9,13\}$, $\{10,15\}$, $\{11,14\}$, $\{11,16\}$, $\{12,17\}$, $\{12,18\}$, $\{13,16\}$, $\{14,19\}$, $\{15,20\}$, $\{15,21\}$, $\{16,22\}$, $\{17,20\}$, $\{17,23\}$, $\{18,21\}$, $\{18,24\}$, $\{19,22\}$, $\{19,23\}$, $\{20,24\}$, $\{21,23\}$, $\{22,24\}$}, $\mathrm{G25}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,11\}$, $\{6,12\}$, $\{7,10\}$, $\{8,13\}$, $\{8,14\}$, $\{9,13\}$, $\{10,15\}$, $\{11,15\}$, $\{11,16\}$, $\{12,17\}$, $\{12,18\}$, $\{13,19\}$, $\{14,16\}$, $\{14,20\}$, $\{15,21\}$, $\{16,19\}$, $\{17,20\}$, $\{17,22\}$, $\{18,23\}$, $\{18,24\}$, $\{19,23\}$, $\{20,24\}$, $\{21,22\}$, $\{21,24\}$, $\{22,23\}$}, $\mathrm{G26}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,11\}$, $\{6,12\}$, $\{7,10\}$, $\{8,13\}$, $\{8,14\}$, $\{9,13\}$, $\{10,15\}$, $\{11,15\}$, $\{11,16\}$, $\{12,17\}$, $\{12,18\}$, $\{13,19\}$, $\{14,16\}$, $\{14,20\}$, $\{15,21\}$, $\{16,22\}$, $\{17,19\}$, $\{17,23\}$, $\{18,22\}$, $\{18,24\}$, $\{19,20\}$, $\{20,24\}$, $\{21,23\}$, $\{21,24\}$, $\{22,23\}$}, $\mathrm{G27}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,11\}$, $\{6,12\}$, $\{7,10\}$, $\{8,13\}$, $\{8,14\}$, $\{9,13\}$, $\{10,15\}$, $\{11,15\}$, $\{11,16\}$, $\{12,17\}$, $\{12,18\}$, $\{13,19\}$, $\{14,17\}$, $\{14,20\}$, $\{15,21\}$, $\{16,20\}$, $\{16,22\}$, $\{17,19\}$, $\{18,22\}$, $\{18,23\}$, $\{19,24\}$, $\{20,23\}$, $\{21,23\}$, $\{21,24\}$, $\{22,24\}$}, $\mathrm{G28}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,11\}$, $\{6,12\}$, $\{7,10\}$, $\{8,13\}$, $\{8,14\}$, $\{9,13\}$, $\{10,15\}$, $\{11,15\}$, $\{11,16\}$, $\{12,17\}$, $\{12,18\}$, $\{13,19\}$, $\{14,17\}$, $\{14,20\}$, $\{15,21\}$, $\{16,20\}$, $\{16,22\}$, $\{17,23\}$, $\{18,22\}$, $\{18,24\}$, $\{19,20\}$, $\{19,24\}$, $\{21,23\}$, $\{21,24\}$, $\{22,23\}$}, $\mathrm{G29}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,11\}$, $\{6,12\}$, $\{7,10\}$, $\{8,13\}$, $\{8,14\}$, $\{9,15\}$, $\{10,16\}$, $\{11,13\}$, $\{11,16\}$, $\{12,17\}$, $\{12,18\}$, $\{13,17\}$, $\{14,19\}$, $\{14,20\}$, $\{15,21\}$, $\{15,22\}$, $\{16,23\}$, $\{17,23\}$, $\{18,19\}$, $\{18,21\}$, $\{19,22\}$, $\{20,21\}$, $\{20,24\}$, $\{22,24\}$, $\{23,24\}$}, $\mathrm{G30}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,11\}$, $\{6,12\}$, $\{7,10\}$, $\{8,13\}$, $\{8,14\}$, $\{9,15\}$, $\{10,16\}$, $\{11,13\}$, $\{11,16\}$, $\{12,17\}$, $\{12,18\}$, $\{13,19\}$, $\{14,15\}$, $\{14,20\}$, $\{15,19\}$, $\{16,21\}$, $\{17,20\}$, $\{17,22\}$, $\{18,23\}$, $\{18,24\}$, $\{19,23\}$, $\{20,24\}$, $\{21,22\}$, $\{21,24\}$, $\{22,23\}$}, $\mathrm{G31}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,11\}$, $\{6,12\}$, $\{7,10\}$, $\{8,13\}$, $\{8,14\}$, $\{9,15\}$, $\{10,16\}$, $\{11,13\}$, $\{11,16\}$, $\{12,17\}$, $\{12,18\}$, $\{13,19\}$, $\{14,15\}$, $\{14,20\}$, $\{15,21\}$, $\{16,17\}$, $\{17,22\}$, $\{18,20\}$, $\{18,23\}$, $\{19,23\}$, $\{19,24\}$, $\{20,24\}$, $\{21,22\}$, $\{21,23\}$, $\{22,24\}$}, $\mathrm{G32}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,11\}$, $\{6,12\}$, $\{7,10\}$, $\{8,13\}$, $\{8,14\}$, $\{9,15\}$, $\{10,16\}$, $\{11,13\}$, $\{11,17\}$, $\{12,16\}$, $\{12,18\}$, $\{13,19\}$, $\{14,15\}$, $\{14,20\}$, $\{15,19\}$, $\{16,21\}$, $\{17,22\}$, $\{17,23\}$, $\{18,20\}$, $\{18,22\}$, $\{19,24\}$, $\{20,23\}$, $\{21,23\}$, $\{21,24\}$, $\{22,24\}$}, $\mathrm{G33}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,11\}$, $\{6,12\}$, $\{7,10\}$, $\{8,13\}$, $\{8,14\}$, $\{9,15\}$, $\{10,16\}$, $\{11,13\}$, $\{11,17\}$, $\{12,16\}$, $\{12,18\}$, $\{13,19\}$, $\{14,15\}$, $\{14,20\}$, $\{15,19\}$, $\{16,21\}$, $\{17,22\}$, $\{17,23\}$, $\{18,22\}$, $\{18,24\}$, $\{19,24\}$, $\{20,21\}$, $\{20,22\}$, $\{21,23\}$, $\{23,24\}$}, $\mathrm{G34}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,11\}$, $\{6,12\}$, $\{7,10\}$, $\{8,13\}$, $\{8,14\}$, $\{9,15\}$, $\{10,16\}$, $\{11,13\}$, $\{11,17\}$, $\{12,18\}$, $\{12,19\}$, $\{13,18\}$, $\{14,20\}$, $\{14,21\}$, $\{15,22\}$, $\{15,23\}$, $\{16,17\}$, $\{16,18\}$, $\{17,24\}$, $\{19,20\}$, $\{19,22\}$, $\{20,23\}$, $\{21,22\}$, $\{21,24\}$, $\{23,24\}$}, $\mathrm{G35}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,11\}$, $\{6,12\}$, $\{7,10\}$, $\{8,13\}$, $\{8,14\}$, $\{9,15\}$, $\{10,16\}$, $\{11,16\}$, $\{11,17\}$, $\{12,18\}$, $\{12,19\}$, $\{13,15\}$, $\{13,17\}$, $\{14,18\}$, $\{14,20\}$, $\{15,20\}$, $\{16,21\}$, $\{17,22\}$, $\{18,23\}$, $\{19,22\}$, $\{19,24\}$, $\{20,24\}$, $\{21,23\}$, $\{21,24\}$, $\{22,23\}$}, $\mathrm{G36}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,11\}$, $\{6,12\}$, $\{7,10\}$, $\{8,13\}$, $\{8,14\}$, $\{9,15\}$, $\{10,16\}$, $\{11,16\}$, $\{11,17\}$, $\{12,18\}$, $\{12,19\}$, $\{13,15\}$, $\{13,18\}$, $\{14,17\}$, $\{14,20\}$, $\{15,20\}$, $\{16,21\}$, $\{17,22\}$, $\{18,23\}$, $\{19,22\}$, $\{19,24\}$, $\{20,24\}$, $\{21,23\}$, $\{21,24\}$, $\{22,23\}$}, $\mathrm{G37}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,11\}$, $\{6,12\}$, $\{7,13\}$, $\{8,14\}$, $\{8,15\}$, $\{9,14\}$, $\{10,16\}$, $\{10,17\}$, $\{11,16\}$, $\{11,18\}$, $\{12,19\}$, $\{12,20\}$, $\{13,18\}$, $\{13,21\}$, $\{14,22\}$, $\{15,19\}$, $\{15,23\}$, $\{16,21\}$, $\{17,18\}$, $\{17,23\}$, $\{19,24\}$, $\{20,22\}$, $\{20,23\}$, $\{21,24\}$, $\{22,24\}$} and $\mathrm{G38}$: {$\{1,2\}$, $\{1,3\}$, $\{1,4\}$, $\{2,5\}$, $\{2,6\}$, $\{3,7\}$, $\{3,8\}$, $\{4,9\}$, $\{4,10\}$, $\{5,7\}$, $\{5,9\}$, $\{6,11\}$, $\{6,12\}$, $\{7,13\}$, $\{8,14\}$, $\{8,15\}$, $\{9,14\}$, $\{10,16\}$, $\{10,17\}$, $\{11,16\}$, $\{11,18\}$, $\{12,19\}$, $\{12,20\}$, $\{13,18\}$, $\{13,21\}$, $\{14,22\}$, $\{15,19\}$, $\{15,23\}$, $\{16,21\}$, $\{17,18\}$, $\{17,24\}$, $\{19,24\}$, $\{20,22\}$, $\{20,23\}$, $\{21,23\}$, $\{22,24\}$}. [\[lem:Snark24 136\]]{#lem:Snark24 136 label="lem:Snark24 136"} There exist a design of order $136$ for each of the thirty-eight non-trivial 24-vertex snarks.\ Proof. Let the vertex set of $K_{136}$ be $Z_{135} \cup \{\infty\}$. The decompositions consist of $(\infty,1,96,44,95,99,21,54,102,3,90,11,\adfsplit 130,94,5,133,120,31,2,25,77,116,98,9)_{\mathrm{G1}}$, $(33,35,88,18,101,9,132,34,58,126,42,114,\adfsplit 120,39,72,119,107,28,70,105,12,133,56,106)_{\mathrm{G1}}$, $(29,13,28,12,27,84,31,37,56,95,46,133,\adfsplit 122,10,114,121,130,60,3,4,85,14,41,119)_{\mathrm{G1}}$, $(120,47,86,95,25,14,80,29,104,58,26,7,\adfsplit 76,28,97,121,41,33,18,46,70,63,133,4)_{\mathrm{G1}}$, $(132,70,3,96,55,5,24,60,59,130,26,56,\adfsplit 37,76,98,16,111,102,109,22,79,47,63,74)_{\mathrm{G1}}$, $(97,84,54,95,83,35,119,77,41,9,12,65,\adfsplit 57,120,55,42,126,50,53,62,123,63,98,51)_{\mathrm{G1}}$, $(2,102,26,48,105,20,56,50,59,114,22,45,\adfsplit 80,24,9,98,116,77,29,11,74,62,113,122)_{\mathrm{G1}}$, $(\infty,71,112,54,34,92,74,91,27,13,99,46,\adfsplit 107,115,123,121,97,25,19,47,31,51,106,14)_{\mathrm{G2}}$, $(116,84,89,26,70,52,113,75,125,123,122,82,\adfsplit 132,81,95,43,78,32,97,13,62,20,83,115)_{\mathrm{G2}}$, $(30,93,125,4,59,85,79,15,78,23,0,81,\adfsplit 126,106,46,36,82,72,99,103,100,77,43,41)_{\mathrm{G2}}$, $(89,126,98,31,28,105,95,10,22,16,110,77,\adfsplit 61,86,107,106,133,60,14,96,101,3,68,79)_{\mathrm{G2}}$, $(98,49,63,78,25,76,67,60,117,30,19,70,\adfsplit 22,58,20,6,105,120,59,126,111,83,3,38)_{\mathrm{G2}}$, $(131,57,125,42,87,59,96,119,54,29,123,21,\adfsplit 51,84,12,62,120,38,45,128,56,53,36,50)_{\mathrm{G2}}$, $(44,90,56,122,26,77,41,27,111,62,60,117,\adfsplit 87,98,50,48,2,83,21,108,80,78,9,14)_{\mathrm{G2}}$, $(\infty,31,33,86,82,70,32,55,90,95,1,21,\adfsplit 115,68,16,2,13,91,19,73,39,111,45,134)_{\mathrm{G3}}$, $(107,50,31,72,78,116,33,48,42,12,91,30,\adfsplit 90,112,59,18,13,80,122,32,17,98,11,58)_{\mathrm{G3}}$, $(38,82,85,67,95,128,78,112,50,97,116,96,\adfsplit 132,13,101,55,36,16,93,83,104,12,76,66)_{\mathrm{G3}}$, $(92,68,119,3,76,104,105,60,44,85,54,32,\adfsplit 125,78,72,45,8,86,71,12,34,65,33,102)_{\mathrm{G3}}$, $(29,59,126,91,44,94,128,42,77,80,103,70,\adfsplit 119,24,9,114,111,33,10,0,92,130,32,58)_{\mathrm{G3}}$, $(35,88,118,1,21,24,97,73,69,100,112,93,\adfsplit 36,117,22,45,82,85,70,127,34,60,124,106)_{\mathrm{G3}}$, $(1,60,36,79,97,64,40,10,131,25,111,30,\adfsplit 31,49,15,76,67,90,101,75,115,22,112,19)_{\mathrm{G3}}$, $(\infty,75,70,68,60,122,6,33,35,109,71,48,\adfsplit 91,111,13,65,22,63,121,118,42,7,47,126)_{\mathrm{G4}}$, $(7,83,59,128,76,43,22,102,14,71,21,124,\adfsplit 8,129,72,98,5,132,52,125,39,110,95,97)_{\mathrm{G4}}$, $(79,81,129,16,52,40,29,109,69,98,118,56,\adfsplit 67,47,37,87,48,63,116,44,21,26,50,131)_{\mathrm{G4}}$, $(108,81,122,131,105,11,91,50,95,67,20,82,\adfsplit 117,26,17,75,128,73,89,107,106,66,96,45)_{\mathrm{G4}}$, $(62,91,38,64,95,83,20,73,126,65,117,3,\adfsplit 17,90,70,30,33,85,130,109,96,81,58,69)_{\mathrm{G4}}$, $(127,95,52,91,24,121,6,70,94,79,4,66,\adfsplit 38,51,115,19,1,102,64,84,18,67,0,21)_{\mathrm{G4}}$, $(7,4,121,39,124,64,109,85,127,99,13,25,\adfsplit 97,78,24,45,117,118,71,15,0,102,70,103)_{\mathrm{G4}}$, $(\infty,93,53,76,43,0,113,49,24,2,34,54,\adfsplit 64,77,6,82,28,21,125,117,96,122,87,71)_{\mathrm{G5}}$, $(2,105,34,106,112,78,114,68,10,24,79,63,\adfsplit 23,72,52,82,45,32,64,39,116,1,18,43)_{\mathrm{G5}}$, $(125,24,118,77,86,16,92,6,84,78,131,35,\adfsplit 76,23,18,4,117,52,95,60,130,80,29,128)_{\mathrm{G5}}$, $(46,112,67,127,26,106,114,131,128,110,61,101,\adfsplit 59,43,91,14,16,19,40,84,111,7,86,93)_{\mathrm{G5}}$, $(120,57,101,16,0,52,106,115,109,133,107,8,\adfsplit 55,87,76,104,82,75,39,80,93,38,119,72)_{\mathrm{G5}}$, $(76,98,51,38,71,83,45,68,50,65,104,93,\adfsplit 106,81,44,82,128,117,21,57,111,105,6,87)_{\mathrm{G5}}$, $(5,17,33,99,32,29,59,123,9,71,11,41,\adfsplit 129,20,53,42,18,125,36,90,51,120,21,84)_{\mathrm{G5}}$, $(\infty,64,33,83,116,34,85,45,111,82,28,109,\adfsplit 37,0,50,35,86,110,70,68,115,40,69,56)_{\mathrm{G6}}$, $(11,89,129,50,90,60,94,3,115,122,39,40,\adfsplit 88,92,102,133,33,101,66,32,46,100,112,99)_{\mathrm{G6}}$, $(55,11,50,92,35,24,53,101,132,118,64,47,\adfsplit 43,122,67,96,108,23,111,103,134,4,37,15)_{\mathrm{G6}}$, $(80,125,14,101,50,36,69,64,132,8,13,129,\adfsplit 130,115,15,68,112,106,74,58,34,71,44,96)_{\mathrm{G6}}$, $(8,25,72,60,97,39,1,115,63,76,74,27,\adfsplit 23,7,18,9,96,55,14,31,51,128,130,36)_{\mathrm{G6}}$, $(67,46,92,32,38,25,126,4,123,80,50,30,\adfsplit 131,84,65,0,39,78,21,66,69,72,105,117)_{\mathrm{G6}}$, $(5,52,92,129,17,44,127,0,60,95,33,80,\adfsplit 108,54,117,114,78,96,30,45,111,56,8,3)_{\mathrm{G6}}$, $(\infty,123,16,92,100,30,29,69,132,57,27,67,\adfsplit 121,18,95,3,77,60,90,23,61,2,66,12)_{\mathrm{G7}}$, $(131,5,95,72,13,10,56,49,57,3,119,48,\adfsplit 66,11,89,32,86,16,46,74,22,114,88,4)_{\mathrm{G7}}$, $(80,6,101,23,62,35,91,82,0,134,68,92,\adfsplit 48,26,119,123,61,76,4,53,50,130,75,20)_{\mathrm{G7}}$, $(34,95,67,83,91,51,85,88,42,89,65,39,\adfsplit 115,61,82,5,118,53,133,57,71,69,19,60)_{\mathrm{G7}}$, $(89,24,101,121,103,30,83,11,15,91,62,1,\adfsplit 65,98,31,25,93,37,28,21,22,102,33,132)_{\mathrm{G7}}$, $(98,101,95,4,84,104,48,10,3,91,87,132,\adfsplit 61,111,69,51,109,115,9,93,36,0,58,117)_{\mathrm{G7}}$, $(1,55,27,3,54,75,121,7,130,105,49,70,\adfsplit 66,80,103,69,25,117,33,63,16,42,110,109)_{\mathrm{G7}}$, $(\infty,40,32,105,117,59,76,93,77,68,104,37,\adfsplit 5,85,128,18,4,34,79,110,65,64,51,126)_{\mathrm{G8}}$, $(57,36,114,81,4,55,19,66,112,10,46,17,\adfsplit 40,31,113,84,125,134,86,117,129,95,9,7)_{\mathrm{G8}}$, $(20,37,120,112,122,129,56,3,74,0,6,28,\adfsplit 75,30,52,73,72,79,71,10,83,53,63,110)_{\mathrm{G8}}$, $(54,18,123,88,74,128,81,69,4,8,51,110,\adfsplit 17,40,62,5,28,103,120,85,50,57,27,119)_{\mathrm{G8}}$, $(20,31,82,26,105,7,103,87,52,111,0,3,\adfsplit 36,39,6,70,48,121,28,42,41,86,118,50)_{\mathrm{G8}}$, $(90,129,51,77,116,33,119,20,74,25,59,82,\adfsplit 83,89,19,22,94,67,88,73,104,124,62,31)_{\mathrm{G8}}$, $(2,85,95,11,28,89,100,98,32,7,110,130,\adfsplit 35,41,70,26,124,9,66,113,92,82,77,134)_{\mathrm{G8}}$, $(\infty,23,123,79,87,12,80,58,38,90,134,109,\adfsplit 42,15,50,122,29,37,132,121,82,104,99,1)_{\mathrm{G9}}$, $(85,50,125,73,47,78,16,96,133,120,25,48,\adfsplit 59,71,60,83,12,23,84,69,75,22,40,91)_{\mathrm{G9}}$, $(3,128,52,36,83,27,58,7,50,80,61,60,\adfsplit 53,18,6,59,120,133,11,35,69,126,12,107)_{\mathrm{G9}}$, $(84,48,90,134,87,45,16,21,127,46,53,7,\adfsplit 59,64,56,31,129,119,70,8,43,0,6,52)_{\mathrm{G9}}$, $(36,83,90,94,78,35,133,34,37,3,117,108,\adfsplit 24,40,45,109,86,107,104,116,110,22,43,38)_{\mathrm{G9}}$, $(86,67,60,58,119,71,34,44,22,38,19,53,\adfsplit 18,56,65,104,14,2,127,109,107,70,10,28)_{\mathrm{G9}}$, $(1,103,34,31,50,80,64,133,73,57,52,83,\adfsplit 125,107,41,106,56,40,130,121,122,68,70,88)_{\mathrm{G9}}$, $(\infty,17,37,72,38,19,104,108,74,117,122,48,\adfsplit 119,93,106,92,64,42,68,15,107,88,44,84)_{\mathrm{G10}}$, $(31,55,121,113,26,51,22,108,27,117,48,53,\adfsplit 95,10,88,41,103,40,46,47,24,3,36,101)_{\mathrm{G10}}$, $(132,71,12,108,9,85,106,120,98,134,93,45,\adfsplit 61,89,63,6,59,94,56,127,119,100,126,104)_{\mathrm{G10}}$, $(16,101,92,23,72,68,35,60,103,66,65,57,\adfsplit 32,69,117,78,30,95,58,122,106,4,111,76)_{\mathrm{G10}}$, $(133,21,48,16,0,93,9,49,36,112,53,25,\adfsplit 105,20,28,114,59,119,125,124,113,91,37,72)_{\mathrm{G10}}$, $(3,106,28,46,11,81,19,49,103,58,124,94,\adfsplit 112,109,115,67,61,29,97,52,37,14,131,89)_{\mathrm{G10}}$, $(1,43,16,44,35,31,73,68,62,4,83,52,\adfsplit 25,127,46,38,56,94,104,37,112,71,131,69)_{\mathrm{G10}}$, $(\infty,8,99,25,127,54,124,24,9,38,74,122,\adfsplit 2,62,53,66,101,47,79,44,5,109,51,4)_{\mathrm{G11}}$, $(20,54,93,44,43,48,125,32,21,124,100,33,\adfsplit 7,112,88,39,131,83,92,41,1,130,109,82)_{\mathrm{G11}}$, $(69,65,31,28,129,128,106,38,57,97,13,132,\adfsplit 29,113,35,107,55,119,58,34,110,4,60,102)_{\mathrm{G11}}$, $(45,91,72,62,19,105,64,13,76,48,66,8,\adfsplit 89,113,102,36,67,100,97,70,1,9,46,63)_{\mathrm{G11}}$, $(97,57,49,86,62,52,34,58,80,82,79,69,\adfsplit 78,105,48,63,132,87,26,127,131,114,66,95)_{\mathrm{G11}}$, $(130,27,95,62,29,60,105,28,97,33,11,116,\adfsplit 128,129,117,84,44,77,93,63,51,74,35,120)_{\mathrm{G11}}$, $(0,59,77,78,126,108,26,66,71,80,39,96,\adfsplit 131,81,29,12,6,41,47,92,48,15,14,99)_{\mathrm{G11}}$, $(\infty,6,19,47,82,37,55,86,93,98,40,88,\adfsplit 26,10,85,73,56,35,42,90,71,87,22,21)_{\mathrm{G12}}$, $(130,134,50,119,3,103,76,98,7,28,84,133,\adfsplit 71,104,60,58,116,99,36,111,23,124,17,57)_{\mathrm{G12}}$, $(58,29,80,131,59,88,47,72,87,107,24,120,\adfsplit 90,11,122,67,81,130,127,8,56,43,55,14)_{\mathrm{G12}}$, $(119,23,109,42,120,95,84,25,126,35,113,125,\adfsplit 64,38,96,127,0,116,102,16,15,39,22,105)_{\mathrm{G12}}$, $(52,98,87,59,101,32,44,115,22,106,42,96,\adfsplit 27,51,15,65,9,41,99,57,111,72,74,62)_{\mathrm{G12}}$, $(8,75,7,22,84,70,16,9,133,64,56,15,\adfsplit 18,19,55,82,123,76,61,86,21,85,67,100)_{\mathrm{G12}}$, $(1,6,129,96,132,13,130,79,46,4,34,49,\adfsplit 25,115,90,9,16,40,66,67,109,73,127,108)_{\mathrm{G12}}$, $(\infty,125,103,69,70,64,82,106,24,20,81,123,\adfsplit 47,66,77,33,15,49,115,40,45,16,61,75)_{\mathrm{G13}}$, $(109,91,70,124,77,55,118,117,71,20,129,49,\adfsplit 45,110,112,121,98,35,99,68,114,130,0,43)_{\mathrm{G13}}$, $(8,48,10,64,80,128,14,78,53,116,37,90,\adfsplit 85,38,51,106,66,39,79,84,60,41,1,32)_{\mathrm{G13}}$, $(14,31,13,51,116,54,69,122,42,113,68,103,\adfsplit 58,119,12,96,93,23,118,7,89,129,134,87)_{\mathrm{G13}}$, $(54,102,48,1,122,127,52,105,103,93,29,3,\adfsplit 104,114,34,73,38,7,60,51,76,35,86,84)_{\mathrm{G13}}$, $(58,21,104,49,72,96,123,125,132,11,12,25,\adfsplit 117,5,113,92,128,20,34,98,43,122,80,38)_{\mathrm{G13}}$, $(128,14,41,40,116,62,53,122,35,8,20,17,\adfsplit 59,83,119,46,134,95,23,91,11,54,129,100)_{\mathrm{G13}}$, $(\infty,124,62,102,110,112,34,70,15,93,86,103,\adfsplit 89,12,7,10,98,90,8,75,37,117,104,97)_{\mathrm{G14}}$, $(54,13,19,133,108,124,93,88,85,118,38,28,\adfsplit 8,128,34,5,33,94,4,77,114,16,134,42)_{\mathrm{G14}}$, $(54,120,1,72,56,18,76,22,45,78,85,77,\adfsplit 53,64,49,87,36,67,113,106,74,28,24,117)_{\mathrm{G14}}$, $(85,44,52,75,127,132,95,74,24,13,9,134,\adfsplit 96,64,62,1,59,26,33,91,48,89,69,116)_{\mathrm{G14}}$, $(56,67,0,24,13,96,132,4,120,101,10,127,\adfsplit 65,63,28,70,66,11,134,53,36,126,98,102)_{\mathrm{G14}}$, $(10,98,47,15,131,29,71,68,8,38,45,24,\adfsplit 125,20,3,66,110,134,26,32,59,22,41,27)_{\mathrm{G14}}$, $(5,29,26,52,41,104,83,77,57,89,20,108,\adfsplit 53,101,71,78,11,68,66,36,62,131,96,2)_{\mathrm{G14}}$, $(\infty,112,90,29,49,24,116,74,127,134,5,25,\adfsplit 34,124,1,77,2,119,56,125,89,13,78,132)_{\mathrm{G15}}$, $(11,19,4,118,5,131,119,69,60,21,38,1,\adfsplit 9,51,54,92,71,61,66,13,126,48,134,94)_{\mathrm{G15}}$, $(47,4,121,20,10,12,78,17,35,24,88,1,\adfsplit 8,114,98,18,108,96,28,118,90,80,40,77)_{\mathrm{G15}}$, $(28,134,24,132,88,102,63,34,127,12,79,120,\adfsplit 49,2,76,78,133,16,117,58,112,103,43,77)_{\mathrm{G15}}$, $(121,18,15,68,91,110,92,13,125,54,127,88,\adfsplit 130,108,53,51,83,9,48,96,81,11,12,75)_{\mathrm{G15}}$, $(97,61,48,42,121,1,72,88,66,125,81,76,\adfsplit 75,110,36,40,26,29,80,27,33,62,119,113)_{\mathrm{G15}}$, $(1,100,123,96,93,51,41,77,11,5,27,53,\adfsplit 23,110,3,107,56,47,32,131,83,116,122,71)_{\mathrm{G15}}$, $(\infty,77,30,61,108,51,14,37,58,60,27,102,\adfsplit 0,117,46,126,91,28,103,74,17,70,15,118)_{\mathrm{G16}}$, $(40,73,110,11,61,4,16,88,92,116,134,59,\adfsplit 69,21,106,71,120,36,35,103,14,28,52,83)_{\mathrm{G16}}$, $(68,11,16,17,53,8,98,12,112,2,9,63,\adfsplit 124,110,88,15,93,30,61,77,23,95,62,105)_{\mathrm{G16}}$, $(83,87,107,39,82,66,34,30,95,51,93,80,\adfsplit 17,27,118,44,74,8,126,28,16,124,106,111)_{\mathrm{G16}}$, $(105,60,62,116,0,41,54,24,57,15,29,20,\adfsplit 106,80,76,117,120,36,53,48,82,16,78,46)_{\mathrm{G16}}$, $(29,128,56,1,85,68,36,116,27,7,41,32,\adfsplit 73,127,100,21,13,92,109,55,67,126,25,4)_{\mathrm{G16}}$, $(5,97,112,115,48,43,106,121,22,134,21,73,\adfsplit 16,51,98,85,15,7,125,127,79,64,58,105)_{\mathrm{G16}}$, $(\infty,87,65,55,68,108,114,25,7,132,34,110,\adfsplit 120,100,128,75,31,89,117,78,2,99,4,48)_{\mathrm{G17}}$, $(12,1,111,20,129,57,31,79,3,65,13,61,\adfsplit 15,52,81,97,23,91,124,37,47,74,120,99)_{\mathrm{G17}}$, $(57,98,91,14,105,59,63,34,35,0,83,92,\adfsplit 87,93,120,97,18,37,32,39,1,8,17,72)_{\mathrm{G17}}$, $(117,17,90,50,16,97,76,15,110,86,116,96,\adfsplit 85,78,13,38,109,120,66,83,122,80,130,92)_{\mathrm{G17}}$, $(129,6,25,90,84,89,130,63,113,42,41,35,\adfsplit 18,46,48,107,120,13,65,1,109,103,112,71)_{\mathrm{G17}}$, $(23,16,40,106,59,127,122,58,130,95,110,68,\adfsplit 52,28,98,125,115,134,26,10,56,128,82,4)_{\mathrm{G17}}$, $(5,112,31,20,77,1,74,107,47,133,44,101,\adfsplit 7,110,94,61,37,35,124,22,28,125,50,118)_{\mathrm{G17}}$, $(\infty,41,28,45,96,58,76,32,134,65,21,13,\adfsplit 48,0,57,61,62,25,67,125,6,15,101,16)_{\mathrm{G18}}$, $(32,0,9,119,17,84,53,107,12,131,70,69,\adfsplit 40,79,56,104,42,35,47,50,112,59,124,126)_{\mathrm{G18}}$, $(29,100,87,20,115,24,52,128,76,62,93,121,\adfsplit 81,66,132,48,79,83,16,126,45,1,57,123)_{\mathrm{G18}}$, $(77,91,126,110,50,8,122,85,97,130,88,35,\adfsplit 117,114,13,24,58,7,108,106,52,96,118,74)_{\mathrm{G18}}$, $(74,8,20,0,32,112,2,73,57,109,82,27,\adfsplit 16,96,51,39,75,37,9,10,63,69,56,64)_{\mathrm{G18}}$, $(18,34,32,71,77,36,16,7,115,10,65,119,\adfsplit 47,53,35,96,94,49,109,55,59,112,110,17)_{\mathrm{G18}}$, $(1,41,3,17,118,20,74,56,102,68,6,26,\adfsplit 13,71,132,59,90,7,86,92,57,80,5,130)_{\mathrm{G18}}$, $(\infty,121,122,120,7,30,84,123,21,129,2,133,\adfsplit 124,44,11,67,82,116,113,58,48,91,4,63)_{\mathrm{G19}}$, $(79,18,124,51,86,47,36,89,8,43,7,91,\adfsplit 16,60,132,2,134,117,10,115,67,53,108,68)_{\mathrm{G19}}$, $(33,99,82,15,75,92,117,79,5,123,96,71,\adfsplit 89,7,98,101,104,34,30,27,63,62,46,47)_{\mathrm{G19}}$, $(99,27,107,112,6,8,121,101,53,102,37,46,\adfsplit 22,14,49,97,61,7,103,34,29,31,9,63)_{\mathrm{G19}}$, $(49,56,120,6,41,12,122,16,101,17,109,69,\adfsplit 124,11,89,8,47,73,20,50,71,77,28,23)_{\mathrm{G19}}$, $(17,68,30,63,81,119,0,126,33,32,87,132,\adfsplit 51,88,78,48,121,120,21,102,16,45,96,18)_{\mathrm{G19}}$, $(4,99,106,72,40,96,73,81,10,75,93,15,\adfsplit 129,78,123,47,48,55,19,98,121,0,124,66)_{\mathrm{G19}}$, $(\infty,2,52,0,66,95,87,17,63,118,80,48,\adfsplit 53,31,78,89,21,115,75,132,125,74,85,42)_{\mathrm{G20}}$, $(78,90,27,37,84,95,106,3,9,49,50,36,\adfsplit 17,29,0,10,61,30,75,121,94,113,8,2)_{\mathrm{G20}}$, $(0,56,124,32,110,90,44,65,93,102,81,25,\adfsplit 40,26,1,128,91,88,6,99,119,41,97,92)_{\mathrm{G20}}$, $(43,0,63,62,116,16,107,47,25,49,75,122,\adfsplit 88,52,67,13,58,65,15,64,103,132,4,45)_{\mathrm{G20}}$, $(54,121,0,123,24,36,131,87,81,91,75,21,\adfsplit 82,50,39,25,44,109,49,23,92,111,124,95)_{\mathrm{G20}}$, $(96,67,56,7,88,103,49,74,95,28,83,100,\adfsplit 62,23,8,32,106,102,19,20,13,112,73,16)_{\mathrm{G20}}$, $(5,45,134,62,91,16,52,116,55,119,86,115,\adfsplit 49,92,35,95,73,131,22,77,8,26,20,110)_{\mathrm{G20}}$, $(\infty,126,41,121,109,56,78,82,13,2,114,34,\adfsplit 127,124,113,54,49,33,3,9,91,55,42,95)_{\mathrm{G21}}$, $(43,56,119,122,55,19,11,85,117,79,52,101,\adfsplit 60,133,32,23,115,36,45,75,77,89,72,123)_{\mathrm{G21}}$, $(127,108,132,42,2,55,30,15,13,62,37,51,\adfsplit 85,69,75,112,74,20,9,14,35,76,24,56)_{\mathrm{G21}}$, $(48,127,131,64,76,44,87,56,9,90,14,73,\adfsplit 37,78,46,35,75,15,68,20,60,107,12,39)_{\mathrm{G21}}$, $(53,69,55,108,27,18,15,5,35,49,24,11,\adfsplit 44,93,73,14,45,13,70,43,88,114,47,83)_{\mathrm{G21}}$, $(90,86,55,116,61,22,67,65,71,133,29,57,\adfsplit 16,37,33,53,83,19,25,59,2,4,32,47)_{\mathrm{G21}}$, $(2,112,35,131,109,95,43,10,29,31,46,50,\adfsplit 17,110,37,62,86,88,116,103,97,113,80,94)_{\mathrm{G21}}$, $(\infty,61,119,105,104,129,46,118,24,10,1,60,\adfsplit 42,123,84,124,50,93,125,109,26,4,17,39)_{\mathrm{G22}}$, $(33,67,62,43,61,36,69,12,118,131,11,7,\adfsplit 89,80,90,37,55,6,52,101,59,87,88,98)_{\mathrm{G22}}$, $(54,48,34,109,126,26,30,52,29,9,67,74,\adfsplit 133,93,32,43,125,61,105,47,127,68,108,41)_{\mathrm{G22}}$, $(60,33,63,37,38,80,128,61,116,133,73,126,\adfsplit 72,70,88,52,0,121,42,24,20,69,99,50)_{\mathrm{G22}}$, $(37,17,79,51,21,28,5,63,20,104,128,60,\adfsplit 121,30,123,93,133,55,131,3,50,45,119,40)_{\mathrm{G22}}$, $(48,108,104,91,16,33,132,131,83,53,93,76,\adfsplit 56,112,116,65,49,106,89,134,119,74,11,7)_{\mathrm{G22}}$, $(2,14,119,125,32,22,100,1,44,50,38,104,\adfsplit 71,117,88,91,41,4,89,61,64,131,122,80)_{\mathrm{G22}}$, $(\infty,94,105,59,27,134,123,19,114,117,46,118,\adfsplit 47,8,121,64,22,131,14,65,36,29,6,1)_{\mathrm{G23}}$, $(25,35,88,54,2,94,107,56,91,120,48,86,\adfsplit 129,39,30,123,57,106,93,83,82,79,60,112)_{\mathrm{G23}}$, $(63,84,127,13,86,104,101,66,90,88,67,8,\adfsplit 73,118,112,69,68,72,98,22,111,103,129,64)_{\mathrm{G23}}$, $(112,95,25,121,29,71,105,116,78,1,39,22,\adfsplit 73,104,134,77,49,40,52,51,119,98,108,61)_{\mathrm{G23}}$, $(55,66,121,116,46,13,80,81,71,105,14,49,\adfsplit 62,86,127,82,56,112,134,113,29,114,84,32)_{\mathrm{G23}}$, $(127,57,54,65,114,120,3,66,30,111,84,58,\adfsplit 44,50,123,20,131,101,87,107,42,36,60,6)_{\mathrm{G23}}$, $(3,101,47,131,60,0,104,54,53,90,57,15,\adfsplit 18,17,27,81,99,96,111,97,59,35,126,24)_{\mathrm{G23}}$, $(\infty,95,16,90,68,19,79,22,85,66,10,106,\adfsplit 62,53,49,101,32,15,30,52,84,47,120,12)_{\mathrm{G24}}$, $(31,81,0,97,103,14,4,26,101,29,76,40,\adfsplit 59,78,48,46,67,125,134,112,91,83,105,3)_{\mathrm{G24}}$, $(38,44,24,101,16,37,91,74,100,115,25,78,\adfsplit 108,99,59,7,90,29,127,132,87,39,57,41)_{\mathrm{G24}}$, $(124,28,0,114,110,48,20,128,105,25,38,127,\adfsplit 106,98,6,81,70,2,130,5,43,97,123,26)_{\mathrm{G24}}$, $(113,110,129,94,13,64,67,65,42,18,43,35,\adfsplit 100,101,56,85,19,83,48,74,87,2,11,17)_{\mathrm{G24}}$, $(13,62,36,101,53,102,21,120,123,60,81,79,\adfsplit 131,33,114,122,128,32,111,96,54,57,119,0)_{\mathrm{G24}}$, $(0,6,11,108,80,57,72,51,102,126,9,30,\adfsplit 123,24,105,48,90,93,1,27,111,89,50,78)_{\mathrm{G24}}$, $(\infty,68,108,127,24,55,134,11,115,126,0,96,\adfsplit 45,33,114,118,121,28,73,2,72,21,97,6)_{\mathrm{G25}}$, $(103,106,68,54,93,47,117,108,95,35,19,24,\adfsplit 3,126,100,5,44,30,133,73,134,97,4,88)_{\mathrm{G25}}$, $(35,0,6,99,112,9,60,51,131,96,129,133,\adfsplit 12,115,83,81,72,53,11,5,120,44,50,62)_{\mathrm{G25}}$, $(91,20,19,64,3,27,124,94,59,113,54,58,\adfsplit 133,47,134,101,51,17,18,108,49,67,16,6)_{\mathrm{G25}}$, $(54,133,129,116,8,27,4,34,76,25,17,90,\adfsplit 30,86,134,2,43,109,22,81,26,74,31,131)_{\mathrm{G25}}$, $(96,34,64,133,131,66,28,97,71,101,4,103,\adfsplit 95,10,77,2,86,1,62,128,11,44,134,49)_{\mathrm{G25}}$, $(2,79,92,77,31,71,134,80,133,122,47,13,\adfsplit 97,44,59,110,45,49,74,82,95,118,41,26)_{\mathrm{G25}}$, $(\infty,58,111,29,43,22,106,26,54,80,66,61,\adfsplit 1,74,28,118,30,57,119,50,16,59,46,65)_{\mathrm{G26}}$, $(96,63,55,4,107,104,121,32,26,94,62,48,\adfsplit 67,87,17,36,13,38,69,60,28,39,23,75)_{\mathrm{G26}}$, $(55,132,45,63,115,8,93,58,40,122,100,96,\adfsplit 128,86,47,15,90,114,60,106,43,80,61,49)_{\mathrm{G26}}$, $(46,108,45,100,56,84,33,0,50,26,4,98,\adfsplit 49,63,130,88,36,3,70,103,101,95,122,23)_{\mathrm{G26}}$, $(124,3,53,82,118,78,115,13,50,6,10,7,\adfsplit 59,61,34,42,41,44,43,5,108,80,113,9)_{\mathrm{G26}}$, $(106,69,99,98,38,122,41,71,66,59,18,130,\adfsplit 83,101,72,131,93,123,80,62,114,57,0,45)_{\mathrm{G26}}$, $(3,84,99,42,26,63,131,38,23,119,44,24,\adfsplit 6,77,66,105,78,102,120,89,73,83,65,36)_{\mathrm{G26}}$, $(\infty,68,0,109,87,60,20,118,48,112,65,66,\adfsplit 127,25,37,9,92,98,126,21,26,114,14,32)_{\mathrm{G27}}$, $(95,53,38,92,19,83,90,129,54,22,25,118,\adfsplit 79,80,86,9,66,21,126,56,23,111,49,10)_{\mathrm{G27}}$, $(86,107,90,91,67,61,112,65,130,102,32,17,\adfsplit 13,28,45,64,76,40,29,123,21,50,69,70)_{\mathrm{G27}}$, $(48,66,113,16,39,83,111,122,70,69,56,34,\adfsplit 24,103,25,1,126,107,61,9,52,78,22,85)_{\mathrm{G27}}$, $(23,22,45,76,16,43,95,5,96,133,94,112,\adfsplit 105,129,116,130,124,97,126,128,14,122,24,83)_{\mathrm{G27}}$, $(115,119,6,129,101,60,21,57,53,83,2,122,\adfsplit 126,31,47,63,45,32,42,35,62,78,17,14)_{\mathrm{G27}}$, $(2,56,66,30,33,41,132,81,107,129,128,23,\adfsplit 84,35,68,53,99,77,0,134,123,113,54,109)_{\mathrm{G27}}$, $(\infty,54,7,8,112,66,103,72,134,131,96,53,\adfsplit 64,57,117,1,61,113,118,13,35,30,125,16)_{\mathrm{G28}}$, $(27,121,75,106,125,12,68,9,13,6,72,59,\adfsplit 74,126,26,8,84,112,25,10,117,88,5,20)_{\mathrm{G28}}$, $(109,66,21,10,69,120,19,56,94,26,97,82,\adfsplit 53,67,102,50,32,6,51,112,35,86,5,61)_{\mathrm{G28}}$, $(38,57,39,42,41,88,85,44,69,3,20,68,\adfsplit 33,81,35,52,10,58,27,49,74,79,96,40)_{\mathrm{G28}}$, $(124,14,87,64,108,99,88,19,75,101,115,9,\adfsplit 82,67,59,121,33,130,30,107,110,52,24,131)_{\mathrm{G28}}$, $(91,122,62,74,132,46,54,31,48,117,92,108,\adfsplit 77,93,86,61,30,125,8,32,20,107,81,51)_{\mathrm{G28}}$, $(2,71,129,61,77,38,51,98,44,123,62,56,\adfsplit 92,26,72,14,85,39,5,132,101,131,68,29)_{\mathrm{G28}}$, $(\infty,103,120,11,101,56,111,46,67,93,134,70,\adfsplit 57,115,35,118,28,77,50,7,131,36,119,126)_{\mathrm{G29}}$, $(2,76,77,78,31,47,93,80,50,104,64,95,\adfsplit 99,111,18,28,86,58,38,17,35,133,37,123)_{\mathrm{G29}}$, $(4,84,25,104,47,76,68,51,109,46,42,70,\adfsplit 134,118,38,98,121,44,132,36,129,127,14,69)_{\mathrm{G29}}$, $(117,77,73,20,118,57,88,5,42,90,105,11,\adfsplit 71,109,126,1,3,35,134,55,65,129,66,93)_{\mathrm{G29}}$, $(113,71,11,28,79,116,26,74,40,32,45,104,\adfsplit 51,72,2,57,108,18,132,3,82,102,76,4)_{\mathrm{G29}}$, $(25,48,28,12,1,7,123,45,132,27,114,57,\adfsplit 97,84,19,3,40,109,91,112,52,94,43,0)_{\mathrm{G29}}$, $(0,15,54,88,127,69,94,67,84,70,121,112,\adfsplit 34,45,3,10,27,130,73,49,42,25,51,1)_{\mathrm{G29}}$, $(\infty,116,40,57,7,90,43,53,59,73,42,130,\adfsplit 102,3,107,16,105,112,89,26,76,117,92,49)_{\mathrm{G30}}$, $(101,29,18,120,91,97,107,19,12,115,1,95,\adfsplit 69,126,2,98,34,76,8,103,131,92,61,47)_{\mathrm{G30}}$, $(64,50,113,107,78,84,83,112,119,73,102,21,\adfsplit 63,103,111,18,86,67,120,71,22,77,118,43)_{\mathrm{G30}}$, $(127,83,122,14,47,39,58,46,71,96,86,121,\adfsplit 133,75,54,93,51,67,18,45,38,88,99,114)_{\mathrm{G30}}$, $(34,59,93,22,27,32,49,112,12,16,25,28,\adfsplit 60,67,36,87,21,74,131,54,129,55,132,37)_{\mathrm{G30}}$, $(115,22,20,50,62,64,0,110,8,29,134,106,\adfsplit 96,56,98,66,120,71,23,116,44,30,92,87)_{\mathrm{G30}}$, $(1,15,101,41,11,35,33,31,86,78,0,72,\adfsplit 45,131,17,113,68,5,83,39,75,110,126,59)_{\mathrm{G30}}$, $(\infty,119,43,93,115,17,83,26,55,88,128,56,\adfsplit 132,63,120,74,33,129,82,98,113,36,20,84)_{\mathrm{G31}}$, $(27,117,61,49,83,50,128,45,93,82,132,40,\adfsplit 109,76,121,71,10,56,8,77,72,64,58,115)_{\mathrm{G31}}$, $(16,34,74,96,129,15,131,55,67,8,2,95,\adfsplit 124,46,70,105,114,51,99,120,23,108,72,45)_{\mathrm{G31}}$, $(122,100,28,66,85,126,81,60,26,105,80,108,\adfsplit 73,17,35,83,52,91,72,45,43,0,102,20)_{\mathrm{G31}}$, $(134,50,75,52,45,120,33,108,129,104,21,0,\adfsplit 115,118,61,5,11,37,23,31,96,76,49,131)_{\mathrm{G31}}$, $(45,53,44,62,125,5,88,67,110,2,123,68,\adfsplit 131,4,46,122,97,20,71,118,109,70,86,91)_{\mathrm{G31}}$, $(2,65,31,112,134,57,70,52,68,14,74,56,\adfsplit 94,124,88,44,100,131,121,101,46,71,82,22)_{\mathrm{G31}}$, $(\infty,70,51,113,45,67,106,81,101,82,97,85,\adfsplit 10,123,64,31,9,48,2,77,40,88,7,86)_{\mathrm{G32}}$, $(38,21,11,66,103,56,128,7,44,25,47,79,\adfsplit 115,96,33,107,52,92,41,87,85,95,68,121)_{\mathrm{G32}}$, $(28,67,70,38,129,102,27,1,63,133,81,12,\adfsplit 80,18,49,132,43,121,8,50,100,58,93,23)_{\mathrm{G32}}$, $(12,129,70,0,59,30,20,3,104,119,10,105,\adfsplit 44,101,52,48,16,132,117,13,40,35,2,123)_{\mathrm{G32}}$, $(31,121,43,88,72,9,23,103,67,75,37,19,\adfsplit 81,99,38,20,44,0,5,3,132,122,84,98)_{\mathrm{G32}}$, $(87,63,83,36,116,52,90,81,56,86,18,8,\adfsplit 95,14,35,38,62,2,89,68,82,69,48,93)_{\mathrm{G32}}$, $(2,89,83,77,87,41,113,93,121,8,51,116,\adfsplit 95,11,132,66,101,126,74,102,15,75,35,68)_{\mathrm{G32}}$, $(\infty,96,44,49,13,35,120,45,27,122,4,21,\adfsplit 98,3,127,88,89,28,129,85,11,22,23,67)_{\mathrm{G33}}$, $(126,120,92,15,53,54,8,99,111,55,11,75,\adfsplit 66,40,95,52,110,129,56,10,67,116,21,134)_{\mathrm{G33}}$, $(15,105,78,132,41,14,114,7,129,27,128,11,\adfsplit 16,26,3,64,72,37,130,34,84,80,110,61)_{\mathrm{G33}}$, $(22,123,3,21,31,43,64,7,112,53,75,27,\adfsplit 124,35,28,78,90,23,93,24,107,41,20,118)_{\mathrm{G33}}$, $(113,133,86,18,15,24,121,45,28,13,112,132,\adfsplit 33,120,92,54,76,32,91,5,35,64,128,104)_{\mathrm{G33}}$, $(117,79,32,80,37,101,41,17,97,58,22,134,\adfsplit 8,50,52,118,47,124,68,28,76,14,31,128)_{\mathrm{G33}}$, $(6,104,56,103,34,119,121,116,113,61,21,38,\adfsplit 118,37,32,47,71,55,70,127,62,17,67,65)_{\mathrm{G33}}$, $(\infty,103,69,134,98,133,46,114,16,104,38,83,\adfsplit 120,9,84,122,81,71,6,63,25,72,80,108)_{\mathrm{G34}}$, $(89,92,13,128,101,102,30,127,100,112,22,80,\adfsplit 83,52,6,54,133,76,94,46,97,123,48,35)_{\mathrm{G34}}$, $(66,126,68,86,112,94,8,53,115,117,109,28,\adfsplit 60,121,39,33,120,79,37,50,13,2,98,48)_{\mathrm{G34}}$, $(74,94,41,9,106,40,131,102,124,104,1,2,\adfsplit 134,10,46,16,35,65,56,111,120,82,92,13)_{\mathrm{G34}}$, $(109,68,7,44,76,26,119,129,34,8,60,106,\adfsplit 64,85,39,9,70,33,43,93,21,6,14,30)_{\mathrm{G34}}$, $(109,77,108,102,56,69,35,115,113,123,72,76,\adfsplit 83,114,119,23,111,75,44,0,128,90,96,134)_{\mathrm{G34}}$, $(35,87,102,21,92,14,3,6,24,65,71,126,\adfsplit 42,17,68,117,50,131,27,74,9,30,39,44)_{\mathrm{G34}}$, $(\infty,11,133,99,111,59,74,16,125,95,66,5,\adfsplit 94,76,51,79,89,30,82,93,13,128,12,113)_{\mathrm{G35}}$, $(57,134,10,133,132,126,118,4,31,131,100,28,\adfsplit 45,87,67,122,69,33,21,75,42,24,97,8)_{\mathrm{G35}}$, $(107,16,79,32,23,65,76,29,20,131,6,97,\adfsplit 47,52,115,88,93,77,41,54,72,55,87,112)_{\mathrm{G35}}$, $(10,78,55,98,58,13,97,17,93,35,22,63,\adfsplit 81,74,24,41,32,23,82,45,4,111,88,34)_{\mathrm{G35}}$, $(101,116,36,7,82,55,61,115,19,51,109,40,\adfsplit 4,66,8,83,110,23,92,35,59,48,111,125)_{\mathrm{G35}}$, $(5,110,47,129,80,118,69,6,30,120,90,114,\adfsplit 45,105,29,81,96,75,7,111,54,134,92,3)_{\mathrm{G35}}$, $(76,48,68,72,123,108,96,98,122,36,119,3,\adfsplit 114,45,105,69,12,44,9,23,99,50,66,116)_{\mathrm{G35}}$, $(\infty,72,73,113,28,115,61,130,3,117,129,66,\adfsplit 26,132,97,111,1,32,5,80,98,105,114,107)_{\mathrm{G36}}$, $(89,24,112,4,61,131,115,16,13,20,118,34,\adfsplit 26,31,80,3,54,14,33,32,85,73,63,71)_{\mathrm{G36}}$, $(84,60,19,74,87,126,25,62,48,12,50,5,\adfsplit 34,47,0,130,105,104,133,55,30,72,75,128)_{\mathrm{G36}}$, $(87,40,128,102,90,59,23,84,13,3,54,131,\adfsplit 112,63,71,126,25,110,106,117,2,1,76,34)_{\mathrm{G36}}$, $(40,124,22,133,41,108,110,44,98,106,49,100,\adfsplit 62,7,38,20,75,11,64,97,84,43,54,53)_{\mathrm{G36}}$, $(52,118,131,121,92,49,36,38,95,68,54,128,\adfsplit 91,123,96,12,101,35,125,45,62,93,88,102)_{\mathrm{G36}}$, $(41,129,60,126,78,107,18,92,125,134,51,105,\adfsplit 99,69,122,44,111,48,65,77,86,110,108,84)_{\mathrm{G36}}$, $(\infty,112,122,54,86,109,60,81,51,9,116,97,\adfsplit 13,90,26,19,36,103,78,98,47,37,87,80)_{\mathrm{G37}}$, $(24,38,68,119,33,18,132,73,49,51,75,111,\adfsplit 127,43,94,83,76,32,112,92,34,74,19,71)_{\mathrm{G37}}$, $(130,51,49,103,29,39,0,12,128,92,26,57,\adfsplit 63,78,133,40,20,5,64,11,127,13,90,72)_{\mathrm{G37}}$, $(27,40,133,6,128,64,49,34,53,30,60,44,\adfsplit 125,57,96,0,70,108,124,69,119,37,62,15)_{\mathrm{G37}}$, $(75,47,26,81,25,84,27,3,114,70,122,40,\adfsplit 112,97,64,80,18,77,55,17,131,34,99,119)_{\mathrm{G37}}$, $(24,82,39,9,29,48,97,54,134,67,38,79,\adfsplit 58,88,85,34,14,125,118,47,116,104,101,50)_{\mathrm{G37}}$, $(96,130,127,86,134,41,31,8,56,110,32,118,\adfsplit 2,47,35,64,14,71,89,16,68,74,122,98)_{\mathrm{G37}}$ and $(\infty,12,46,92,123,2,67,77,91,116,33,76,\adfsplit 3,58,113,70,125,7,100,114,63,20,131,22)_{\mathrm{G38}}$, $(108,111,17,69,5,82,80,62,46,124,53,13,\adfsplit 58,96,76,133,114,19,8,38,105,3,121,132)_{\mathrm{G38}}$, $(70,134,24,34,129,108,96,85,104,42,120,32,\adfsplit 83,74,12,109,23,76,88,113,55,133,90,118)_{\mathrm{G38}}$, $(131,47,54,89,113,108,81,13,105,96,22,23,\adfsplit 27,100,17,130,116,36,29,31,7,74,119,78)_{\mathrm{G38}}$, $(133,83,130,8,118,98,55,111,56,47,49,7,\adfsplit 72,19,65,31,124,102,117,1,73,123,3,40)_{\mathrm{G38}}$, $(57,93,56,128,5,23,83,134,126,36,21,80,\adfsplit 104,42,39,0,35,105,24,9,107,78,129,38)_{\mathrm{G38}}$, $(131,39,75,60,126,90,47,123,38,113,45,42,\adfsplit 84,108,95,5,0,98,132,89,2,87,116,66)_{\mathrm{G38}}$ under the action of the mapping $\infty \mapsto \infty$, $x \mapsto x + 3$ (mod 135) for the first five graphs in each design, and $x \mapsto x + 9$ (mod 135) for the last two. [\[lem:Snark24 designs\]]{#lem:Snark24 designs label="lem:Snark24 designs"} There exist designs of order $64$, $73$ and $145$ for each of the thirty-eight 24-vertex non-trivial snarks.\ Proof. The decompositions are given in the Appendix.\ Proof. The decompositions are given in the Appendix. Theorem [\[thm:snark24\]](#thm:snark24){reference-type="ref" reference="thm:snark24"} follows from Lemmas [\[lem:Snark24 136\]](#lem:Snark24 136){reference-type="ref" reference="lem:Snark24 136"}, [\[lem:Snark24 designs\]](#lem:Snark24 designs){reference-type="ref" reference="lem:Snark24 designs"} and [\[lem:Snark24 multipartite\]](#lem:Snark24 multipartite){reference-type="ref" reference="lem:Snark24 multipartite"}, and Proposition [\[prop:d=3, v=24\]](#prop:d=3, v=24){reference-type="ref" reference="prop:d=3, v=24"}.	{'timestamp': '2016-07-19T02:06:58', 'yymm': '1607', 'arxiv_id': '1607.04847', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04847'}
null	null
# Introduction {#sec:introduction} Bioinformaticians define the $k$th-order de Bruijn graph for a string or set of strings to be the directed graph whose nodes are the distinct $k$-tuples in those strings and in which there is an edge from $u$ to $v$ if there is a $(k + 1)$-tuple somewhere in those strings whose prefix of length $k$ is $u$ and whose suffix of length $k$ is $v$.[^1] These graphs have many uses in bioinformatics, including de novo assembly , read correction and pan-genomics . The datasets in these applications are massive and the graphs can be even larger, however, so pointer-based implementations are impractical. Researchers have suggested several approaches to representing de Bruijn graphs compactly, the two most popular of which are based on Bloom filters and the Burrows-Wheeler Transform , respectively. In this paper we describe a new approach, based on minimal perfect hash functions , that is similar to that using Bloom filters but has better theoretical bounds when the number of connected components in the graph is small, and is fully dynamic: i.e., we can both insert and delete nodes and edges efficiently, whereas implementations based on Bloom filters are usually semi-dynamic and support only insertions. We also show how to modify our implementation to support, e.g., jumbled pattern matching with fixed-length patterns. Our data structure is based on a combination of Karp-Rabin hashing and minimal perfect hashing, which we will describe in the full version of this paper and which we summarize for now with the following technical lemmas: Suppose $N$ is the node-set of a de Bruijn graph. In Section [2](#sec:static){reference-type="ref" reference="sec:static"} we show how we can store $\Oh{n \sigma}$ more bits than Lemma [\[lem:static\]](#lem:static){reference-type="ref" reference="lem:static"} such that, given a pair of $k$-tuples $u$ and $v$ of which at least one is in $N$, we can check whether the edge $(u, v)$ is in the graph. This means that, if we start with a $k$-tuple in $N$, then we can explore the entire connected component containing that $k$-tuple in the underlying undirected graph. On the other hand, if we start with a $k$-tuple not in $N$, then we will learn that fact as soon as we try to cross an edge to a $k$-tuple that is in $N$. To deal with the possibility that we never try to cross such an edge, however---i.e., # Static de Bruijn Graphs {#sec:static} Let $G$ be a de Bruijn graph of order $k$, let $N = \{v_0, \ldots, v_{n-1}\}$ be the set of its nodes, and let $E = \{a_0, \ldots, a_{e-1}\}$ be the set of its edges. We call each $v_i$ either a node or a $k$-tuple, using interchangeably the two terms since there is a one-to-one correspondence between nodes and labels. We maintain the structure of $G$ by storing two binary matrices, and , of size $n \times \sigma$. For each node, the former represents its incoming edges whereas the latter represents its outgoing edges. In particular, for each $k$-tuple $v_x = c_1 c_2 \ldots c_{k-1} a$, the former stores a row of length $\sigma$ such that, if there exists another $k$-tuple $v_y = b c_1 c_2 \ldots c_{k-1}$ and an edge from $v_y$ to $v_x$, then the position indexed by $b$ of such row is set to . Similarly, contains a row for $v_y$ and the position indexed by $a$ is set to . As previously stated, each $k$-tuple is uniquely mapped to a value between $0$ and $n-1$ by $f$, where $f$ is as defined in Lemma [\[lem:static\]](#lem:static){reference-type="ref" reference="lem:static"}, and therefore we can use these values as indices for the rows of the matrices and , i.e.,Suppose we want to check whether there is an edge from $b X$ to $X a$. Letting $f(b X) = i$ and $f(X a) = j$, we first assume $b X$ is in $G$ and check the values of $\OUT [i] [a]$ and $\IN [j] [b]$. If both values are , we report that the edge is present and we say that the edge is confirmed by and ; otherwise, if any of the two values is , we report that the edge is absent. Moreover, note that if $b X$ is in $G$ and $\OUT [i] [a] = \TRUE$, then $X a$ is in $G$ as well. Symmetrically, if $X a$ is in $G$ and $\IN [j] [b] = \TRUE$, then $b X$ is in $G$ as well. Therefore, if $\OUT [i] [a] = \IN [j] [b] = \TRUE$, then $b X$ is in $G$ if and only if $X a$ is. This means that, if we have a path $P$ and if all the edges in $P$ are confirmed by and , then either all the nodes touched by $P$ are in $G$ or none of them is. We now focus on detecting false positives in our data structure maintaining a reasonable memory usage. Our strategy is to sample a subset of nodes for which we store the plain-text $k$-tuple and connect all the unsampled nodes to the sampled ones. More precisely, we partition nodes in the undirected graph $G^\prime$ underlying $G$ into a forest of rooted trees of height at least $k \lg \sigma$ and at most $3 k \lg \sigma$. For each node we store a pointer to its parent in the tree, which takes $1 + \lg \sigma$ bits per node, and we sample the $k$-mer at the root of such tree. We allow a tree to have height smaller than $k \lg \sigma$ when necessary, e.g., if it covers a connected component. Figure [\[fig:trees\]](#fig:trees){reference-type="ref" reference="fig:trees"} shows an illustration of this idea. We can therefore check whether a given node $v_x$ is in $G$ by first computing $f(v_x)$ and then checking and ascending at most $3 k \lg \sigma$ edges, updating $v_x$ and $f(v_x)$ as we go. Once we reach the root of the tree we can compare the resulting $k$-tuple with the one sampled to check if $v_x$ is in the graph. This procedure requires time since computing the first value of $f(v_x)$ requires , ascending the tree requires constant time per edge, and comparing the $k$-tuples requires . We now describe a Las Vegas algorithm for the construction of this data structure that requires, with high probability, expected time. We recall that $N$ is the set of input nodes of size $n$. We first select a function $f$ and construct bitvector $B$ of size $n$ initialized with all its elements set to . For each elements $v_x$ of $N$ we compute $f(v_x) = i$ and check the value of $B[i]$. If this value is we set it to and proceed with the next element in $N$, if it is already set to , we reset $B$, select a different function $f$, and restart the procedure from the first element in $N$. Once we finish this procedure---i.e.,This procedure requires with high probability expected time for constructing $f$ and time for computing and . Notice that if $N$ is the set of $k$-tuples of a single text sorted by their starting position in the text, each $f(v_x)$ can be computed in constant time from $f(v_{x-1})$ except for $f(v_0)$ that still requires . More generally, if $N$ is the set of $k$-tuples of $t$ texts sorted by their initial position, we can compute $n-t$ values of the function $f(v_x)$ in constant time from $f(v_{x-1})$ and the remaining in . We will explain how to build the forest in the full version of this paper. In this case the construction requires, with high probability, $\Oh{kt + n + n\sigma} = \Oh{kt + n\sigma}$ expected time. Combining our forest with Lemma [\[lem:static\]](#lem:static){reference-type="ref" reference="lem:static"}, we can summarize our static data structure in the following theorem: In the full version we will show how to use monotone minimal perfect hashing to reduce the space to $(2+\epsilon)n\sigma$ bits of space (for any constant $\epsilon>0$). We will also show how to reduce the time to list the edges incident to a node of degree $d$ to $O(d)$, and the time to check whether a node is in $G$ to $\Oh{k}$. We note that the obtained space and query times are both optimal up to constant factors, which is unlike previous methods which have additional factor(s) depending on $k$ and/or $\sigma$ in space and/or time. # Dynamic de Bruijn Graphs {#sec:dynamic} In the previous section we presented a static representation of de Buijn graphs, we now present how we can make this data structure dynamic. In particular, we will show how we can insert and remove edges and nodes and that updating the graph reduces to managing the covering forest over $G$. In this section, when we refer to $f$ we mean the function defined in Lemma [\[lem:dynamic\]](#lem:dynamic){reference-type="ref" reference="lem:dynamic"}. We first show how to add or remove an edge in the graph and will later describe how to add or remove a node in it. The updates must maintain the following invariant: any tree must have size at least $k\log\sigma$ and height at most $3k\log\sigma$ except when the tree covers (all nodes in) a connected component of size at most $k\log\sigma$. Let $v_x$ and $v_y$ be two nodes in $G$, $e = (v_x, v_y)$ be an edge in $G$, and let $f(v_x) = i$ and $f(v_y) = j$. Suppose we want to add $e$ to $G$. First, we set to the values of $\OUT[i][a]$ and $\IN[j][b]$ in constant time. We then check whether $v_x$ or $v_y$ are in different components of size less than $k \lg \sigma$ in time for each node. If both components have size greater than $k \lg \sigma$ we do not have to proceed further since the trees will not change. If both connected components have size less than $k \lg \sigma$ we merge their trees in time by traversing both trees and switching the orientation of the edges in them, discarding the samples at the roots of the old trees and sampling the new root in time. If only one of the two connected components has size greater than $k \lg \sigma$ we select it and perform a tree traversal to check whether the depth of the node is less than $2 k \lg \sigma$. If it is, we connect the two trees as in the previous case. If it is not, we traverse the tree in the bigger components upwards for $k \lg \sigma$ steps, we delete the edge pointing to the parent of the node we reached creating a new tree, and merge it with the smaller one. This procedure requires time since deleting the edge pointing to the parent in the tree requires time, i.e.,Suppose now that we want to remove $e$ from $G$. First we set to the values of $\OUT[i][a]$ and $\IN[j][b]$ in constant time. Then, we check in time whether $e$ is an edge in some tree by computing $f(v_x)$ and $f(v_y)$ checking for each node if that edge is the one that points to their parent. If $e$ is not in any tree we do not have to proceed further whereas if it is we check the size of each tree in which $v_x$ and $v_y$ are. If any of the two trees is small (i.e.,If such an edge is not found we conclude that we are in a small connected component that is covered by the current tree and we sample a node in the tree as a root and switch directions of some edges if necessary. If such an edge is found, we merge the small tree with the bigger one by adding the edge and switch the direction of some edges originating from the small tree if necessary. Finally if the height of the new tree exceeds $3k\log\sigma$, we traverse the tree upwards from the deepest node in the tree (which was necessarily a node in the smaller tree before the merger) for $2k \lg \sigma$ steps, delete the edge pointing to the parent of the reached node, creating a new tree. This procedure requires $\Oh{k \lg \sigma}$ since the number of nodes traversed is at most $O(k \lg \sigma)$ and the number of changes to the data structures is also at most $O(k \lg \sigma)$ with each change taking expected constant time. It is clear that the insertion and deletion algorithms will maintain the invariant on the tree sizes. It is also clear that the invariant implies that the number of sampled nodes is $O(n/(k\log\sigma))$ plus the number of connected components. We now show how to add and remove a node from the graph. Adding a node is trivial since it will not have any edge connecting it to any other node. Therefore adding a node reduces to modify the function $f$ and requires amortized expected time. When we want to remove a node, we first remove all its edges one by one and, once the node is isolated from the graph, we remove it by updating the function $f$. Since a node will have at most $\sigma$ edges and updating $f$ requires amortized expected time, the amortized expected time complexity of this procedure is $\Oh{\sigma k\lg \sigma+ k}$. Combining these techniques for updating our forest with Lemma [\[lem:dynamic\]](#lem:dynamic){reference-type="ref" reference="lem:dynamic"}, we can summarize our dynamic data structure in the following theorem: # Jumbled Pattern Matching {#sec:jumbled} Karp-Rabin hash functions implicitly divide their domain into equivalence classes---i.e., subsets in which the elements hash to the same value. In this paper we have chosen Karp-Rabin hash functions such that each equivalence class contains only one $k$-tuple in the graph. Most of our efforts have gone into being able, given a $k$-tuple and a hash value, to determine whether that $k$-tuple is the unique element of its equivalence class in the graph. In some sense, therefore, we have treated the equivalence relation induced by our hash functions as a necessary evil, useful for space-efficiency but otherwise an obstacle to be overcome. For some applications, however---e.g., parameterized pattern matching, circular pattern matching or jumbled pattern matching---we are given an interesting equivalence relation on strings and asked to preprocess a text such that later, given a pattern, we can determine whether any substrings of the text are in the same equivalence class as the pattern. We can modify our data structure for some of these applications by replacing the Karp-Rabin hash function by other kinds of hash functions. For indexed jumbled pattern matching we are asked to pre-process a text such that later, given a pattern, we can determine quickly whether any substring of the text consists of exactly the same multiset of characters in the pattern. Consider fixed-length jumbled pattern matching, when the length of the patterns is fixed at pre-processing time. If we modify Lemmas [\[lem:static\]](#lem:static){reference-type="ref" reference="lem:static"} and [\[lem:dynamic\]](#lem:dynamic){reference-type="ref" reference="lem:dynamic"} so that, instead of using Karp-Rabin hashes in the definition of the function $f$, we use a hash function on the histograms of characters' frequencies in $k$-tuples, our function $f$ will map all permutations of a $k$-tuple to the same value. The rest of our implementation stays the same, but now the nodes of our graph are multisets of characters of size $k$ and there is an edge between two nodes $u$ and $v$ if it is possible to replace an element of $u$ and obtain $v$. If we build our graph for the multisets of characters in $k$-tuples in a string $S$, then our process for checking whether a node is in the graph tells us whether there is a jumbled match in $S$ for a pattern of length $k$. If we build a tree in which the root is a graph for all of $S$, the left and right children of the root are graphs for the first and second halves of $S$, etc., as described by Gagie et al. , then we increase the space by a logarithmic factor but we can return the locations of all matches quickly.	{'timestamp': '2016-07-21T02:01:17', 'yymm': '1607', 'arxiv_id': '1607.04909', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04909'}
null	null
# Preliminaries and Problem Formulation {#sec:prelims} [\[sec2\]]{#sec2 label="sec2"} #### Bipartite Graphs A bipartite graph $G=(\mathcal{M},\mathcal{B},E)$ is a graph such that the vertices $\mathcal{M}\cup \mathcal{B}$ can be divided into two disjoint subsets, $\mathcal{M}$ and $\mathcal{B}$, and there are no edges connecting vertices in the same subset, $E\subseteq \mathcal{M}\times \mathcal{B}$. Such a graph is balanced if $\|\mathcal{M}\| = \|\mathcal{B}\|$, i.e., if the two subsets have the same cardinality. A perfect matching in a balanced bipartite graph $G=(\mathcal{M},\mathcal{B},E)$ is a subset of edges $E_{pm} \subseteq E$ such that every vertex in $G$ is incident upon exactly one edge of the matching. We denote the neighbors of a set of vertices $S$ by $N(S)$, where $N(S)\triangleq \{j\in \mathcal{B}: \exists i \in S~\text{s.t.}~(i,j) \in E\}$ when $S \subseteq \mathcal{M}$ and $N(S) \triangleq \{j\in \mathcal{M}: \exists i \in S~\text{s.t.}~(j,i) \in E\}$ when $S \subseteq \mathcal{B}$. #### Matching Market We consider a matching market with a set $\mathcal{B}$ of buyers, and a set $\mathcal{M}$ heterogeneous merchandise with exactly one copy of each type of good. Each buyer $i\in \mathcal{B}$ has a non-negative valuation $v_{ij}\geq 0$ for good $j\in M$, and desires at most one good (e.g. they are unit-demand buyers). We denote the $\|\mathcal{B}\|\times \|\mathcal{M}\|$ valuation matrix by $\mathbf{V}$. Our assumption that $\|\mathcal{B}\|=\|\mathcal{M}\|=m$ is without loss of generality because we can always add dummy goods or dummy buyers for balance. Given a price vector $\mathbf{P}=[P_1 \, P_2 \,...\, P_m]$, we assume a quasi-linear utilities for the buyers, i.e., buyer $i$ receiving good $j$ has utility $U_{i,j}=v_{i,j}-P_j$. Since each buyer is unit-demand, we define $U^_i$ be the maximum (non-negative) payoff of buyer $i \in \mathcal{B}$, i.e., $U^_i=\max \big\{0,\underset{j\in \mathcal{M}}{\max}~v_{i,j}-P_j\big\}$. Since buyers can opt out of the market and obtain zero, we insist on the payoff being non-negative. Note that the preferred goods set of a buyer is empty if its payoff for all the goods is negative. To avoid any confusion, we always place goods on the left-hand side and buyers on the right-hand side of the preference graphs. Given a specific price vector, if the preference graph $G_{\mathrm{pref}}$ contains a perfect matching $E_{pm}\subseteq E_{\mathrm{pref}}$, then we can allocate to each buyer exactly one of the goods it prefers and also sell all the goods. A price vector that leads to a perfect matching in the realized preference graph is called a market-clearing price (MCP) (also called a Walrasian price). Given any valuation $V$, it is well known that the set of MCPs is non-empty and bounded. Boundedness is obvious from the finiteness of the valuations. Non-emptiness is established either using the characterization in, using a constructive ascending price algorithm that starts from all the prices being $0$, or by using the VCG mechanism price (see Chap 15 in ). Furthermore, the set of MCPs has a lattice structure, so that given any two different MCP vectors, the element-wise maximum of the vectors and the element-wise minimum of the vectors are also MCPs. This guarantees the existence of the maximum and the minimum MCPs. #### Complexity of Algorithms An algorithm runs in strongly polynomial time if the number of operations and the space used are bounded by a polynomial in the number of input parameters, i.e., $O(\text{polynomial of }\|M\|)$, but both do not depend on the size of the parameters (assuming unit time for basic mathematical operations). If this does not hold but the number of operations is still bounded by a polynomial in the number of input parameters where the coefficients depend on the size of the parameters, then we say that the algorithm runs in weakly polynomial time. # Related Work {#sec:related} While sponsored search auctions are a recent motivation to study matching markets, there is a vast history to the problem. The term "matching market\" can be traced back to the seminal paper "College Admissions and the Stability of Marriage\" work by Gale and Shapley. In matching markets, the necessary and sufficient condition for the existence of an efficient matching using Hall's marriage theorem has been proved and a widely used mathematical model of two-sided matching markets was introduced in "The Assignment Game I: The Core\" by Shapley and Shubik. In, the set of MCPs is further shown to be solutions of a linear programming (LP) problem and the lattice property is also established. Despite this the study of this problem goes back at least to the well-known strongly polynomial-time Hungarian algorithm for finding the maximum weight matching in a weighted bipartite graph, which in fact can also be used to find the minimum MCP. Furthermore, several auction algorithms enhancing the run-time efficiency in markets with specific properties have been presented in. Leonard considered mechanisms with sealed-bids and proved that charging the minimum competitive equilibrium price from bidders will result in an incentive compatible mechanism, and also that MCP coincides with the VCG price. Soon after, an ascending-price-based auction algorithm was presented by Demange, Gale, and Sotomayor (DGS) in, which starts at the zero-price vector and then increases the posted price for any of the minimal over-demanded sets of goods to obtain the minimum MCP. Thereafter, plenty of ascending-price-based auction mechanisms have been studied under different assumptions in. We pause here to remind the reader that the DGS ascending price algorithm is only known to be weakly polynomial-time. On the other hand, there has only been a limited study of descending-price auction algorithms to obtain the maximum MCP. Mishra and Parkes present a descending price auction called the Vickrey-Dutch auction to generate the VCG price in equilibrium. To aim for a higher revenue for sellers, Mishra and Garg generalized the Dutch auction to provide a descending-price-based approximation algorithm in. As mentioned in Section [5](#sec:intro){reference-type="ref" reference="sec:intro"}, Mishra and Garg's algorithm yields an approximation to the maximum MCP via a weak polynomial-time algorithm, and furthermore, there is no analysis of the strategic bidding in their work. We remark again that the sequential LP approach in can be used to obtain the maximum MCP via a weakly polynomial-time algorithm. Finally, there is a body of literature that attempts to raise the revenue of sellers in equilibrium in related problems, such as sponsored search auctions and combinatorial auctions. In sponsored search markets, Ghosh and Sayedi considered a two-dimensional bid on advertisers' valuations according to exclusive and nonexclusive display, then run a GSP-like auction to determine the allocation that maximizes the search engine's revenue. With this small variation, efficiency does not hold for GSP, and hence the revenue will be different from the VCG mechanism. Additionally, in combinatorial auctions, it is well-known that designing a revenue maximizing auction mechanism is still an open problem. To achieve a higher expected revenue of sellers, Likhodedov and Sandholm presented a class of auctions, called virtual valuations combinatorial auctions, to maximize the sum of a pre-determined weighted valuation and an evaluation function of allocation rather than maximizing the total valuations as in the VCG mechanism to get a higher revenue. # Design of Descending Price Algorithm {#sec3} The problem considered in our work, as mentioned earlier, is to find the generalization of the Dutch auction[^1] to matching markets. Specifically, we seek a descending price auction that always converges to the maximum MCP. Like the DGS mechanism, our mechanism will choose a particular constricted good set to ensure the convergence. Specifically, we will define a dual to the "minimal over-demanded set\" which we call the maximally skewed set. Unlike minimal over-demanded sets, the maximally skewed set is unique, and an example of failure to achieve the maximum MCP if this set is not chosen will be discussed in Section [12.1](#sec6.1){reference-type="ref" reference="sec6.1"}. ## Framework of Descending Price Algorithms [\[sec:framework\]]{#sec:framework label="sec:framework"} We design a descending auction, which is the analogue of the ascending auction, in a straightforward framework. We start from a high enough initial price, iteratively pick a constricted good set to decrement prices, and terminate the algorithm when there exists a perfect matching. Clearly, this framework does not guarantee the termination in finite time, let alone strongly polynomial time. In order to make the algorithm run in strongly polynomial time, we will exploit the combinatorial structure of the preference graph, and make the evolution of the preference graph in the run of the algorithm be such that any specific bipartite graph appears at most once. To achieve this goal, we will specify a particular initial configuration, and a particular price reduction to be carried out in each step of the algorithm. #### Initial Price Choice: A perfect matching requires every good be preferred by some buyers. Then a reasonable starting point should guarantee that the preferred-buyer set of every good is non-empty, otherwise it cannot be an MCP for any valuation matrix. Thus, the natural candidate for the initial price is $P_j =\max_{i \in \mathcal{M}}v_{i,j}$ for good $j$, which is (element-wise) greater than or equal to any MCP but ensures that every good is preferred by at least one buyer from the very outset. #### Price Reduction: In computing the price reduction for a given constricted set $S$, we need to reduce the price by a large enough amount to trigger a change in the preference graph (otherwise we still have a constricting set and the same set of goods can be chosen again), but we should also avoid reducing the price of any good below its price in the maximum MCP. In other words, we want to find the minimum value to compensate the buyers not in the $N(S)$ to make at least one buyer indifferent between one of the goods in $S$ and the good(s) she prefers initially; the buyer in question may have an empty preferred goods set, in which case it is sufficient to ensure that one of the goods in $S$ has a non-negative utility with the price reduction. Lemma [\[lemA\]](#lemA){reference-type="ref" reference="lemA"} formally states the price reduction to be used in the proposed family of descending price algorithms. ## Choice of Constricted Good Sets and The Skewness Function {#skewfunc} Since different choices of constricted good sets could generate different MCPs when the algorithm terminates, pinpointing the right constricted good sets iteratively has a pivotal role when designing the algorithm for finding the maximum MCP[^2]. Before detailing the selection criterion, we use Figure [\[fig:skewness\]](#fig:skewness){reference-type="ref" reference="fig:skewness"} to provide some quick intuition. On one hand, we prefer choosing the constricted good set in LL to LR because we want to choose the largest good set given the same set of neighbors (buyers). On the other hand, we prefer RL to RR because we do not include any subgraph (set of good-buyer pairs) that already has a perfect matching. With this intuition in mind, we present the following formal criteria for choosing constricted-good sets: 1. Pick the constricted goods sets $S$ with the largest difference $\|S\|-\|N(S)\|$. 2. If there are multiple sets with the same $\|S\|-\|N(S)\|$, choose the one with the smallest size. The first criterion ensures that at each step the algorithm (simultaneously) reduces the price of the most critical set of goods. The proof that our algorithm returns the maximum MCP will not hold without this property. As an added bonus it also positively impacts the speed of convergence. The second criterion excludes any subset of goods $S'\subset S$ which is already perfectly matched to a subset of buyers, i.e., $\|N(S')\setminus N(S)\|\geq \|S'\|$. Jointly the criteria imply that we are searching for the most "skewed\" constricted good set in the preference graph. To formulate this mathematically, we define a function to measure the skewness of a set. With this skewness function, the criteria described earlier are equivalent to choosing the constricted goods set with the maximal skewness. To formally make this statement we need to show two properties. The first one is the uniqueness of the maximally skewed set when the preference graph has no perfect matching; and the second one is that the maximally skewed set is a constricted goods set when the preference graph has no perfect matching. Lemma [\[lem2\]](#lem2){reference-type="ref" reference="lem2"} proves these. With Lemma [\[lem2\]](#lem2){reference-type="ref" reference="lem2"} in place, it easily follows that the two rules we imposed are equivalent to finding the maximally skewed set at every iteration (as we already know that a perfect matching doesn't exist). With the proper initial price vector choice, specified price reduction per round, and the unique choice of the maximally skewed set, the complete algorithm is described in Algorithm [\[alg:alg1\]](#alg:alg1){reference-type="ref" reference="alg:alg1"}. Note that the DGS algorithm, which searches for over-demanded sets to increase the price, has a dual structure to our algorithm. Thus, it is not surprising that the minimally over-demanded sets of items in the DGS algorithm, denoted as DGS sets below, have a relationship with the skewness function $f(\cdot)$. They are ones that obtain the minimum positive value of the function $\|N(S)\|-\|S\|-\frac{1}{\|N(S)\|}$ when the algorithm starts with initial price 0. We highlight the fact that the DGS sets may not be unique as there can be multiple sets of goods that yield the same minimum positive value for the function $\|N(S)\|-\|S\|-\frac{1}{\|N(S)\|}$. In contrast to our algorithm, the lack of uniqueness in the DGS algorithm is not as critical because different choices of DGS sets lead to the same minimum MCP. Understanding this contrast better is for future work. [\[alg:alg1\]]{#alg:alg1 label="alg:alg1"} For further clarification, any buyer who has zero surplus at the maximum MCP (by definition of the maximum there will exist at least one such buyer) will be indifferent between the matched good and the dummy good. Hence, for every buyer set $B$ containing a zero-surplus buyer, the dummy good $D$ will be in the neighbor set of this set, $D\in N^D(B)$. We also show a dual property to VCG prices of the maximum MCP in Section [\[sec:externality\]](#sec:externality){reference-type="ref" reference="sec:externality"}. With Theorem [\[lem4.1\]](#lem4.1){reference-type="ref" reference="lem4.1"} in hand, we will now establish the correctness of the algorithm, assuming that it halts (in finite-time). Since the skew-aided algorithm continually changes the preference graph, it is necessary to label the bipartite graph in each round of our algorithm before starting any analysis. Let $G_0$ be the initial bipartite graph, in the running of our algorithm, we obtain a bipartite graph $G_t$ at $t^{th}$ round. Then, we'll need to check whether the terminal condition holds. To avoid cumbersome notation, we will use $N_t^D(B)$ instead of $N_{G_t}^D(B)$. With Theorem [\[lem4.1\]](#lem4.1){reference-type="ref" reference="lem4.1"}, the proof of Theorem [\[thm2\]](#thm2){reference-type="ref" reference="thm2"} followings by checking that the preference graphs at termination coincides has the combinatorial characterization outlined above. ## Preference Graphs Converge Quadratically in the Number of Goods {#sec4.2} The Algorithm [\[alg:alg1\]](#alg:alg1){reference-type="ref" reference="alg:alg1"} changes the preference graph in each round to obtain the bipartite graph with combinatorial structure of MCP at termination. We will now show that the algorithm terminates in at most $m^2$ rounds. Given a specific preference graph $G$, we can define the skewness of the graph $W(G)$ to equal the skewness of the maximally skewed set. Therefore, by defining a sequence $W(G_t)=\max_{S \in \mathcal{M}, S \neq \emptyset} f_t(S)$, where $G_t$ is the preference graph obtained at the $t^{\mathrm{th}}$ iteration of Algorithm [\[alg:alg1\]](#alg:alg1){reference-type="ref" reference="alg:alg1"}, we show the convergence of the algorithm in finite rounds by proving that $W(G_t)$ strictly decreases with the decrease at least some positive constant. Thus, $W(\cdot)$ is a potential function that will be shown to strictly decrease in every iteration of the algorithm in the proof of Lemma [\[lem:conv\]](#lem:conv){reference-type="ref" reference="lem:conv"}. Given the minimum decrement in Lemma [\[lem:conv\]](#lem:conv){reference-type="ref" reference="lem:conv"}, it is straightforward that the preference graphs converge to the bipartite graph with combinatorial structure of MCP in time upper bounded by $m^3$ because $W(G)<m$. However, as there are only $m^2$ positive distinct feasible values of $W(G)$ [^4], we are ensured convergence in time at most $m^2$. ## Complexity of the Algorithm Based on the results thus far determining the complexity of our algorithm depends only on the run-time of finding the maximally skewed set. We now discuss two approaches for this. ### Algorithm design in search of the maximally skewed set Given the uniqueness, we can always perform a brute-force search to get the maximally skewed set. Since there are $2^m-1$ non-empty subsets of $\mathcal{M}$, the complexity is $O(2^m)$, which doesn't meet our goal. We will exploit the combinatorial structure of the preference graph to scale down the complexity of finding the maximally skewed set. For this we design a graph coloring algorithm to color the preference graphs in Algorithm [\[alg:alg3\]](#alg:alg3){reference-type="ref" reference="alg:alg3"}. In any colored preference graphs, we want red edges to represent edges connecting matched pairs of good-buyer in a maximum matching, and blue edges to represent the rest of the edges. Hence, each vertex has at most one red edge. Additionally, we want the set of red goods $\mathcal{M}_r$ to represent the set of goods not in the maximally skewed set, the set of blue goods $\mathcal{M}_b$ to be goods in the maximally skewed set but ones that do not have matched pairs to buyers in this maximum matching (because of the nature of constricted good set), and the set of green goods $\mathcal{M}_g$ are the rest of the goods. On the buyer side, the buyers that are neighbors of the maximally skewed set should be colored green, the buyers that are not the neighbors of the maximally skewed set but have a matched good should be colored red, and the rest of the buyers should be colored blue. Given the object we seek, we now present an algorithm to color vertices/edges properly in strongly polynomial-time complexity. The steps will include an initial coloring and followed by an update of the preference graph. Before detailing the initial coloring, we define various depth-first search and breadth-first search procedures relevant to the algorithm. #### Initial Coloring First, we find a maximum matching using the Hopcroft-Karp algorithm and color edges linked matched pairs red, and other edges blue. After that, we start from the set of good without matched buyer in this maximum matching, color them blue, run the br-BFS algorithm starting from the set of blue goods. When the br-BFS algorithm terminates, color the set of reachable goods $Rch(\mathcal{M}_b)\cap \mathcal{M}$ with matched buyers green, color the rest set of goods red. Then, color the matched buyers of red goods red, color the buyers in the $Rch(\mathcal{M}_b)\cap \mathcal{B}$ green (they are the neighbors of the most skewed set), and color the rest of buyer blue. Finally, the following lemma states that we get the maximally skewed set from the initial coloring. Given that the Hopcroft-Karp algorithm has complexity $O(m^{2.5})$ and the br-BFS has complexity upper-bounded by $O(m^2)$, we learn the initial coloring has the complexity $O(m^{2.5})$. Note that the initial coloring does not rely on the initial price, our first algorithm, say *initial coloring based decreasing price auction, will use this in every iteration to get the maximally skewed set. This algorithm has complexity $O(m^2\times m^{2.5})=O(m^{4.5})$, which is already strongly polynomial. #### Update the preference graph We further scale down the complexity of the initial coloring based decreasing price auction* algorithm by exploiting and updating the colored preference graph colored in previous round without completely coloring the whole graph. This is detailed in Algorithm [\[alg:alg3\]](#alg:alg3){reference-type="ref" reference="alg:alg3"}. Since the procedure in Algorithm [\[alg:alg3\]](#alg:alg3){reference-type="ref" reference="alg:alg3"} is elaborate, we will highlight some facts of perfect matching and give the sketch of how we use them in the Algorithm [\[alg:alg3\]](#alg:alg3){reference-type="ref" reference="alg:alg3"}. First, we know that if there is a perfect matching, no vertex should be colored blue otherwise we fail to get a maximum matching. Furthermore, the following Lemma [\[lem:color\]](#lem:color){reference-type="ref" reference="lem:color"} states that we will never need to change a vertex from red or green to blue. Given the property of a proper coloring stated in Lemma [\[lem:color\]](#lem:color){reference-type="ref" reference="lem:color"}, we know that the set of blue buyers in round $t$, denoted as $\mathcal{B}_b^t$, is a decreasing set. Therefore, the key idea of designing Algorithm [\[alg:alg3\]](#alg:alg3){reference-type="ref" reference="alg:alg3"} is to restrict operations irrelevant to reducing the set of blue buyers to some constant number of $O(m^2)$ operations, and to allow the complexity of operations reducing the set of blue buyers to be upper-bounded by $O(m^3)$. To achieve this, we observe that the set of buyers at round $t$ that are willing to get goods in previous maximally skewed set, denoted by $A_t$, only contains red and blue buyers, i.e., $A_t\subseteq \mathcal{B}_r^t \cup\mathcal{B}_b^t$. Then, we know that if $A_t \subseteq \mathcal{B}_r$ and we cannot reach any blue buyers from $A_t$ without passing green or blue goods, then the current matching is maximized, and all we need to do for updating colors is to run rb-BFS from the set of $\mathcal{M}_b$ with complexity $O(m^2)$ as we did in initial coloring. In other cases, we need to find the new maximum matching with number of matched pairs increased by at least 1, which can be achieved by at least $O(m^{2.5})$ using the Hopcroft-Karp algorithm, and at least one blue buyer will be recolored. Since the maximum number of matched pairs is $m$, the complexity of the whole algorithm attaining the maximum MCP will be upper-bounded by the {(complexity of updating process not recoloring blue buyers)$+$(complexity of computing price reduction)}$\times$ (convergence rate of the preference graph)$+$(complexity of updating process increasing maximum matched pairs)$\times$ (maximum number of blue buyers)$=O((m^2+m^2)\times m^2+m^{2.5}\times m)=O(m^4)$. # Introduction {#sec:intro} Online Advertising is an over \\(70 billion busines with double-digit growth in consecutive years over a period of many years. Since nearly all of the ads are sold via auction mechanisms, auction-based algorithm design, which focuses on the online advertising, has become an important class of mechanism design to study. Among all online advertising auctions, the sponsored search auction, also known as a keyword auction, is the one that most captures researchers' attention. In a typical sponsored search auction, the auctioneer has a set of web slots to sell and every advertiser has different valuations on different web slots. Problems in sponsored search auctions are usually modeled as problems in (cardinal-preference) matching markets, and prices are used to clear the market. The key assumption of sponsored search auctions is that every advertiser shares the identical ordinal preference on web slots. Under this assumption, the celebrated Vickrey-Clarke-Groves (VCG) mechanism, which makes truthful bidding by the advertisers as (weakly) dominant strategies but yields low revenue to the auctioneer, is adopted by some web giants such as Facebook[^5], and is also a robust option in scenarios where the revenue equivalence theorem holds. However, as the revenue equivalence theorem does not hold in multi-good auctions , auctioneers can look for greater expected revenue than the value obtained by VCG mechanism by using even different efficient and market-clearing auction mechanisms. The most popular auction among these mechanisms is the Generalized Second Price (GSP) auction employed by Google. Since the GSP is not incentive compatible, its equilibrium behavior needs to be analyzed , and there are some Bayesian Nash equilibria (BNE) that have greater expected revenue than the expected revenue of the VCG mechanism. It should also be noted that designing (revenue) optimal mechanisms is intractable even in the context of matching markets when there is more than one good. Thus, the possibility of higher expected revenue coupled with the ease of implementing the GSP auction and the intractability of optimal mechanisms has lead to the popularity of the GSP mechanism. Unlike a decade ago where there were only statically-listed ads, websites now serve a variety of ads simultaneously, including sidebar images, pop-ups, embedded animations, product recommendations, etc. With this in mind, and the growing heterogeneity in both advertisers and consumers, it is clear that the "shared ordinal preference" assumption is untenable in the context of market design. Search engines and ad-serving platforms will be faced with a growing need to implement general unit-demand matching markets , and such market settings are the focus of our work. We refer to the prices that efficiently allocate the set of goods to the bidders according to their private valuations as a vector of market-clearing prices (MCP). An ascending price auction algorithm that generalizes the English auction was presented by Demange, Gale and Sotomayor. This ascending price algorithm (DGS algorithm) obtains the element-wise minimum MCP, that coincides with the VCG price. DGS is thus incentive compatible yet obtains low expected revenue for the mechanism. Of course, simultaneously maximizing revenue and maintaining incentive compatibility is computationally intractable once we have more than one good for sale, but we should still hope to obtain better than the minimum MCP within efficient mechanisms. In the present paper we design a family of mechanisms that seek to elicit the maximum MCP from the participants without sacrificing computational efficiency. Here we focus explicitly on how we can efficiently compute the maximum MCP given some representation of the bidder utilities, and defer the general[^6] analysis of strategic behavior to future work. Critical to our paper would be answering whether there exists a strongly polynomial-time algorithm to obtain the maximum MCP exactly. Before discussing the literature for the maximum MCP, we discuss the state of the art for the minimum MCP. The intuitively appealing DGS ascending price algorithm that attains the minimum MCP is only weakly polynomial time: the potential function used to show convergence depends on the valuations; and its value decrements by at least a constant independent of the valuations in each step. In fact, it is the well-known strongly polynomial-time Hungarian algorithm for finding the maximum weight matching in a weighted bipartite graph, that yields a strongly polynomial time algorithm for finding the minimum MCP, \)O(m^4)$ in the original implementation that can then be reduced to $O(m^3)$ . This will be the aspirational goal of this work. Using the method outlined in, where one computes the solution of two linear programs, it is possible to determine the maximum MCP. Note, however, that this is at best a weakly polynomial-time algorithm, and is neither a combinatorial nor an auction algorithm. Given that the DGS ascending price mechanism returns the minimum MCP, it is also intuitive to study descending price mechanisms to obtain the maximum MCP, i.e., generalize the Dutch auction to multiple goods. The first attempt to obtain the maximum MCP through descending price auction is in the work by Mishra and Garg, where they provide a descending-price-based auction algorithm that yields an approximation algorithm. The algorithm doesn't require agents to bid their whole valuation but still yields a price-vector in weakly polynomial-time[^7] that is within $\epsilon$ in $l_\infty$ norm of the maximum MCP[^8]. Therefore, in this work, one of main goals is to develop a strongly polynomial-time combinatorial/auction algorithm using descending prices for the exact computation of the maximum MCP. Note that based on the analysis in choosing the maximum MCP in sponsored search markets has exactly the same complexity as the VCG, GSP and Generalized first-price (GFP) mechanisms: The web-slots are sold from best to the worst and in decreasing order of the bids of the agents, with the only difference being the price that's ascribed to each good. Once the computational problem is solved, setting the maximum MCP is a viable option for general unit-demand markets, and is an alternate efficient mechanism. ## Our contribution By judiciously exploiting the combinatorial structure in matching markets, we propose a strongly polynomial-time[^9] descending price auction algorithm that obtains the maximum MCPs in time $O(m^4)$ with $m$ goods (and bidders). Critical to the algorithm is determining the set of under-demanded goods (to be defined precisely later on) for which we reduce the prices simultaneously in each step of the algorithm. This we accomplish by choosing the subset of goods that maximize a skewness function, which is obtained by proposed graph coloring algorithm a simple combinatorial algorithm to keep updating the bipartite graph and the collection under-demanded goods set. We start by discussing an intuitively appealing algorithm to solve this problem that uses the Hopcraft-Karp algorithm and Breadth-First-Search (BFS). This procedure will only yield a complexity of $O(m^{4.5})$. We will then present a refinement that cleverly exploits past computations and the structure of the problem to reduce the complexity to $O(m^4)$. # Appendix-Details of Algorithms [\[alg:alg3\]]{#alg:alg3 label="alg:alg3"} [\[alg:alg2\]]{#alg:alg2 label="alg:alg2"} However, if a law of choosing the constricted good set pick the maximally skewed on in the last round before termination, the answer of whether it returns the maximum MCP will be case by case. Since Fig. [\[fig4\]](#fig4){reference-type="ref" reference="fig4"} provide an example of not achieving the maximum MCP, we now provide an example which still gets the maximum MCP in Fig. [\[fig5\]](#fig5){reference-type="ref" reference="fig5"}. In Fig. [\[fig5\]](#fig5){reference-type="ref" reference="fig5"}, the algorithm in upper flow runs the algorithm we proposed and choose maximally skewed set in each round. We pick $\{A,B\}$ in the first iteration and then pick $\{A,B,C\}$ in the second iteration; and the algorithm terminates at $P=[1,1,2]$. The algorithm in lower flow does not choose the maximally skewed set in the first round, but the algorithm is forced to choose the maximally skewed set in the second round because it is the only constricted good set. It picked $\{A,B,C\}$ and lower the price to $[2,2,2]$; and pick $\{A,B\}$ to get a MCP at $[1,1,2]$. As shown in the figure, these two algorithm both return the maximal MCP. Therefore, the space of sets we can choose which guaranteeing to get the maximum MCP is still an open problem. ## Interpretation of the Maximum MCP Using Buyers' Externalities, Since the minimum posted price derived from DGS algorithm corresponds to the personalized VCG price given by the Clarke pivot rule, it is not surprising that there is an analogous structure between the posted price and the personalized price also for the maximum MCP. The Clarke pivot rule determines the VCG payment of buyer $i$ using the externality that a buyer imposes on the others by her/his presence, i.e., payment of buyer $i=$ (social welfare[^10] of others if buyer $i$ were absent)-(social welfare of others when buyer $i$ is present). Using the combinatorial characterization of the maximum MCP, we obtain an exact analogue of the Clarke pivot rule when viewing the maximum MCP as the personalized price in Theorem [\[thm:externality\]](#thm:externality){reference-type="ref" reference="thm:externality"}. In the theorem below we can use any perfect matching to determine the good matched to buyer $i$. # Appendix-Preliminary Analysis of Strategic Buyers and Bayesian-Nash Equilibrium {#sec5} As mentioned in the introduction, the proposed descending price algorithm is not incentive compatible so that we need to explore Bayesian Nash equilibrium (BNE) to predict buyers' strategic behaviors. From the messaging viewpoint we will assume that we have a direct mechanism where the buyers bid their valuation: a scalar in sponsored search markets, and a vector in the general matching market. Algorithm [\[alg:alg1\]](#alg:alg1){reference-type="ref" reference="alg:alg1"} is then applied as a black-box to the inputs to produce a price on each good and a perfect matching. Although the strategic behavior of buyers is not the main subject of this work, with expected revenue being an important concern, we provide an instance achieving higher expected revenue than VCG mechanism in a BNE with asymmetric distributions of buyers' valuations. Furthermore, we analyze the BNE in two simple cases with symmetric buyers and the associated bidding strategy of the buyers. ## An Instance Achieving higher expected revenue than VCG {#sec:greatRev} Given the proposed algorithm, it is important to check if there exists a scenario that our algorithm can achieve a higher revenue than the VCG mechanism. It is well-known that the revenue equivalence theorem holds with an assumption that valuation of every player is drawn from a path-connected space in Chap. 9.6.3 of [^11]. Furthermore, there's a large body of literature that has discussed the failure of getting the VCG revenue under asymmetric distribution of buyers' valuations, e.g., see. With this knowledge, we demonstrate a $3\times 3$ matching market where our mechanism achieves higher than VCG revenue, where the buyers have asymmetric distributions of their valuations. Consider three advertisers named Alice, Bob, and Carol, and three different types of ads called listing ads, sidebar ads, and pop-ups. The realized valuation is only known to the advertiser (equivalently buyer), but the distribution of an advertisers' valuation is known to other advertisers but not the auctioneer. In other words, the auctioneer can only calculate the price according to the bids submitted by the advertisers. The minimum increment of the submitted bids is $\epsilon$, which is a positive infinitesimal, and the valuation matrix of advertisers is displayed in Table [4](#tbl:1){reference-type="ref" reference="tbl:1"}. \ Now, we give an asymmetric BNE[^12] in this matching market and provide a detailed verification in the Appendix [9.11](#asymBNE){reference-type="ref" reference="asymBNE"}. First, consider Alice always bids 0 on sidebar ads and pop-ups, and bids $\max \{1-\epsilon, \frac{w}{2}\}$ on listing ads. The best response of Bob is to bid 0 on both sidebar ads and pop-ups, and to bid $\max \{1, \frac{x-0.5}{2}\}$ on listing ads for any realized $x$. Now, given Bob's bidding function as above, one of Alice's best responses is to follow her original bidding function. Last, consider the bidding function of Alice and Bob mentioned above, a best response of Carol is to bid 0 on listing ads and pop-ups, and to bid $\epsilon$ on sidebar ads regardless of the outcomes of $y$ and $z$. [^13], even if $y>z$. This is because Carol will never win listing ads for any bids less than $1$ as the probability $\text{Pr}\{y-z \geq 1\}=0$. Now, given Carol's bidding strategy, Alice and Bob will not change their bidding functions. Therefore, the strategy $\{\beta_{Alice}(w,0,0), \beta_{Bob}(x,0,0.5),\beta_{Carol}(y,z,2)\}=\{(\max \{1-\epsilon, \frac{w}{2}\},0,0), (\max \{1, \frac{2x-1}{4}\},0,0),(0,\epsilon,0)\}$ can be verified as an asymmetric BNE. With the asymmetric BNE in hand, we want to calculate the expected revenue of the auctioneers and compare it with the expected revenue of the VCG mechanism. Since Carol always wins the sidebar ads and both Alice and Bob bid 0 on that, Carol will pay $\epsilon$. Additionally, in the asymmetric BNE, Alice and Bob compete on the listing ads and both bid 0 on sidebar ads and pop-ups, resulting in the payment of listing ads to be the same as the payment in the first price auction. Next, let's calculate the expected revenue of auctioneers, which is given by \[\begin{aligned} &&\epsilon+\frac{4}{9}\int_0^1\int_{1.5}^{2.5} 1 dxdw+\frac{2}{9}\int_2^3\int_{1.5}^{2.5} \frac{w}{2} dxdw+\frac{2}{9}\int_1^2\int_{2.5}^{3.5} \frac{2x-1}{4} dxdw \nonumber \\&+&\frac{1}{9}\int_2^3\int_{2.5}^{w+0.5} \frac{w}{2} dxdw+\frac{1}{9}\int_{2.5}^{3.5}\int_2^{x-0.5} \frac{2x-1}{4} dwdx =\dfrac{31}{27}+\epsilon \end{aligned}\] The last step is to calculate the expected revenue of the VCG mechanism, which is given by \[\begin{aligned} &&\frac{2}{3}\int_0^1\int_{1.5}^{2.5} w dxdw+\frac{2}{9}\int_2^3\int_{1.5}^{2.5} (x-0.5) dxdw+\frac{1}{9}\int_{2.5}^{3.5}\int_2^{x-0.5} w dwdx \nonumber \\&+&\frac{1}{9}\int_2^3\int_{2.5}^{w+0.5} (x-0.5) dxdw =\dfrac{1}{3}+\dfrac{2}{9}\times \dfrac{3}{2}+\dfrac{1}{9}\times \dfrac{7}{6}++\dfrac{1}{9}\times \dfrac{7}{6}=\dfrac{25}{27} \end{aligned}\] Even if we set $\epsilon$ to 0, it is obvious that the expected revenue derived under our descending price auction algorithm is strictly greater than the expected revenue of the VCG mechanism. This shows that in some instances the proposed descending price algorithm is preferred to the (DGS) ascending price algorithm, even taking the strategic behavior of buyers into account. ## Symmetric Bayesian Nash Equilibrium in $2\times 2$ Sponsored Search Market Now we focus on the analysis of symmetric distributions of buyers. Next we will detail the analysis of two cases for a market with two goods and two buyers, one in the following paragraphs and another one in the next subsection. Since the primary application of our algorithm is in online advertising auctions, it is useful for us to analyze the strategic behavior under the conventional assumptions made in the sponsored search market setting. The sponsored search market assumes every buyer's (advertiser's) value on goods (web slots) can be determined by a product of the buyer's private weight and a common click-through rate. Now, we consider a $2 \times 2$ case in sponsored search market with settings detailed below. There are two web slots with click-through rates $c_1$ and $c_2$, with $c_1\geq c_2$, and two advertisers with private weights $w_1, w_2$. We assume that $w_j$ is an i.i.d. non-negative random variable with PDF $f_{w_j}(\cdot)$, and the private value for getting web slots $j$ is $w_ic_j$. Each advertiser knows his/her true weight but only knows the distribution of another advertiser's value, and both know that it is a sponsored search market. Under our descending price auction algorithm, they each have to effectively place a one-dimensional bid, denoted by $b_i$, according to their bidding function $\beta_j(w_j,f_{w_{-j}}(\cdot))$, where ${-j}$ denotes the other advertiser(s). To simplify the analysis, we assume weights $w_1, w_2$ are uniformly distributed in $[0,1]$. Extensions to other symmetric distributions, asymmetric distribution/knowledge space and more web slots are all for future work. ### Payment of advertisers If $c_2=0$, the descending price algorithm terminates at the initial point. At this point, the advertiser $j$ wins the first web slots pays $c_1w_j$ and another advertiser pays $0$. This is, equivalent, to the single good case, which has been carefully studied before. When $c_2 \neq 0$, the descending price algorithm makes the advertiser indifferent between two slots if he/she bid truthfully. Therefore, if advertiser $i$ wins the first slot, the payment will be $(c_1-c_2)b_i+c_2b_{-i}$, if he/she gets the second slot, the payment will be $c_2b_i$. ### Analysis of strategic behavior Before deriving the bidding function, we show that the bidding function is a monotonic increasing function of the weight. Since the outcome of this auction is exactly the same as what the VCG mechanism provides, this monotonicity leads us to expect revenue equivalence to hold in this case. Exploring the generality of this result is for future work. This monotonicity property of bidding functions can be generalized to all sponsored search markets when every advertiser's private weight follows the same uniform distributions. Without loss of generality, we now derive the bidding function of advertiser 1. Since the optimal bidding function is monotonic in the private value, the surplus function of advertiser 1 can be written using an integral form. Advertiser 1 wants to maximize the following surplus function[^14]: \[\begin{aligned} \int_0^{\beta_2^{-1}(b_1)}(c_1w_1-[(c_1-c_2)b_1 +c_2\beta_2(x)])f_{w_2}(x)dx +c_2(w_1-b_1)[1-\int_0^{\beta_2^{-1}(b_1)}f_{w_2}(x)dx] \end{aligned}\] With detailed analysis presented in Appendix [\[apdx:analysis\]](#apdx:analysis){reference-type="ref" reference="apdx:analysis"}, the bidding function $\beta_1(w_1)$ is \[\begin{aligned} \begin{cases} e^{-w_1}+w_1-1 \qquad \qquad \qquad \qquad \qquad \qquad ~~~c_1=2c_2 \\ w_1-(2-w_1)\ln(1-0.5w_1) \qquad \qquad \qquad ~~~~c_1=1.5c_2 \\ \dfrac{c_2}{2c_1-3c_2}\Big[(c_1-c_2)w_1+\big(\dfrac{c_2}{c_2+(c_1-2c_2)w_1}\big)^{\tfrac{c_1-c_2}{c_1-2c_2}}-1\Big] \\ \nonumber \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \text{otherwise} \nonumber \end{cases} \end{aligned}\] Without loss of generality, we can let $c_1=1$ and $c_2 \in [0,1]$, Figure [\[fig5\]](#fig5){reference-type="ref" reference="fig5"} displays the optimal bidding bid corresponds to the buyer's private weight and the ratio of click-through rates. Then, the expected revenue of the search engine is \[\begin{aligned} &&2\bigg\{\int_0^1\int_0^{w_1}([(c_1-c_2) \beta_1(w_1)+c_2 \beta_1(w_2)])dw_2dw_1 +\int_0^1\int_0^{w_2}c_2\beta_1(w_2)dw_1dw_2 \bigg\} \nonumber\\ &=& 2\bigg\{(c_1-c_2)\int\limits_0^1 w_1\beta_1(w_1)dw_1+2c_2\int\limits_0^1\int\limits_0^{w_1}\beta_1(x)dxdw_1 \bigg\} =\begin{cases} \frac{c_2}{3}&~c_1=2c_2 \\ \frac{c_2}{6}&~c_1=1.5c_2 \\ \frac{c_1-c_2}{3} &~\text{otherwise} \nonumber \end{cases} \end{aligned}\] It is well known that in the VCG mechanism, the expected VCG revenue is $\dfrac{c_1-c_2}{3}$. Therefore, in this $2\times 2$ case, the proposed algorithm has the same expected revenue as the VCG mechanism after taking the strategic behavior into consideration. ## Symmetric Bayesian Nash Equilibrium in $2\times 2$ General Unit-demand Matching Markets Theoretically, the same method we used to studied the $2\times 2$ case of a sponsored search market can be used to solve the $2\times 2$ case for a general matching market when valuations are assumed to be i.i.d.. However, that method requires us to consider two variables simultaneously, and we cannot avoid solving the resulting partial differential equations. Hence, before directly solving the general $2\times 2$ cases, we present the following lemmas to simplify the analysis afterward. With the above two lemmas, the following corollary is straightforward. Now, we can get rid of all weakly-dominated strategies to analyze the buyer's strategic behavior to find a symmetric Bayesian Nash Equilibrium. It is worth to note that weakly-dominated strategies can still be rationalizable strategies, our approach (analyzing Bayesian Nash equilibrium(BNE) without taking weakly-dominated strategies into consideration) may miss some BNEs. Our current goal is not to find all BNEs but just one. Therefore, we can continue to find a BNE with restricted strategy space. First, assume $v_{i1}\geq v_{i2}$ without loss of generality, a rational buyer $i$ will bid 0 on the good 2, and bid no more than $v_{11}-v_{12}$ on good 1. Then, we can compute the equilibrium in symmetric bidding strategies. Assume the symmetric bidding function is $\beta(\cdot,\cdot)$ for the good with higher value and 0 for the other one, buyer $i$ bids $b$ on good 1 and the CDF of buyer $-i$'s valuation on good $j$ is $F_{-ij}(\cdot)$, the objective function is \[\begin{aligned} \max_b E_{(v_{-i1},v_{-i2})} [u_i(b,0,\beta\mathbf{1}_{\{v_{-i1}>v_{-i2}\}},\beta\mathbf{1}_{\{v_{-i1}\leq v_{-i2}\}})], \end{aligned}\] where $\beta$ denotes $\beta(v_{-i1},v_{-i2})$. Now, we need Lemma [\[pro1\]](#pro1){reference-type="ref" reference="pro1"} to simplify our analysis. With Lemma [\[pro1\]](#pro1){reference-type="ref" reference="pro1"}, we can revise and simplify the bidding function $\beta(v_{ih},v_{il})$ to be $\beta_{r}(v_{ih}-v_{il})$. Define the joint CDF of $v_{-i1}-v_{-i2}$ to be $F_{-i}(\cdot)$. Then the expected surplus of buyer $i$ for a specific bid $b$ of good 1 is: \[\begin{aligned} (v_{i1}-b)F_{-i}(\beta_r^{-1}(b))+v_{i2}(1-F_{-i}(\beta_r^{-1}(b))) =v_{i2}+(v_{i1}-v_{i2}-b_{i1})F_{-i}(\beta_r^{-1}(b)) \end{aligned}\] Therefore, determining the objective function is equivalent to solving the following optimization problem: \[\begin{aligned} \max_b ~(v_{i1}-v_{i2}-b)F_{-i}(\beta_r^{-1}(b)). \end{aligned}\] To simplify our problem, let's assume the distributions of private value on goods of two buyers are i.i.d. uniformly distributed on $[0,1]$. Then assume that the bidding function is differentiable and denote $x=v_{i1}-v_{i2}$. Using the same technique we used to solve the symmetric BNE in sponsored search market. With the analysis in Appendix [\[apdx:analysis2\]](#apdx:analysis2){reference-type="ref" reference="apdx:analysis2"}, the bidding function $\beta_r(x)$ is \[\begin{aligned} \beta_r(x)= \dfrac{1}{2-(1-x)^2}\int_0^{x} 2\tau(1-\tau)d\tau =\dfrac{x^2(1-\tfrac{2x}{3})}{2-(1-x)^2} \end{aligned}\] Then, the expected revenue is \[\begin{aligned} 4\int_0^1 \beta_r(x)(1-x)(\frac{1}{2}+x-\frac{x^2}{2}) dx =2\int_0^1 x^2(1-\tfrac{2x}{3})(1-x) dx = 2(\dfrac{1}{3}-\dfrac{5}{12} + \dfrac{2}{15})= \dfrac{1}{10} \end{aligned}\] Finally, computing the expected revenue with VCG price yields: \[\begin{aligned} 4\int_0^1 (1-x)\int_0^x y(1-y) dydx = 4\int_0^1 (1-x)(\dfrac{x^2}{2}-\dfrac{x^3}{3}) dx = 4 ( \dfrac{1}{6}-\dfrac{5}{24}+\dfrac{1}{15}) = \dfrac{1}{10} \end{aligned}\] It is shown here that the revenue equivalence theorem continues to hold in this case. # Appendix-Proofs ## Proof of Lemma [\[lemA\]](#lemA){reference-type="ref" reference="lemA"} {#pf:lem1} First, we will show that given a constricted good set $S$, reducing price of every good in $S$ with the amount specified in the statement will increase the size of its neighbor, i.e., $N(S)$. Given a constricted good set $S$ under a specific price vector $\mathbf{P}$. Consider another price vector $\mathbf{P'}$, $P'_j=\begin{cases} P_j, & j \notin S \\ P_j-c, & j \in S \end{cases}$, where $c:=\min_{i \in \mathcal{B} \setminus N(S),l\in S} \{\max_{k \in \mathcal{M} \setminus S}(v_{i,k}-P_k)-(v_{i,l}-P_l) \}$. For some $j \in S$, there must exists an $i\in \mathcal{B} \setminus N(S)$ satisfying $v_{i,j}-P_j'= \max_{k \in \mathcal{M} \setminus S} v_{i,k}-P_k$. Now, by an abuse of notation to denote $N'(S)$ as the neighbor of $S$ under $P'$, $i \in N'(S)$. For those $l \in N(S)$, $\max_{j \in S}v_{l,j}-P'_j>\max_{j \in S}v_{l,j}-P_j\geq \max_{j \in \mathcal{M}\setminus S}v_{l,j}-P_j$ implies $l \in N'(S)$. Therefore, $\|N(S)\|<\|N'(S)\|$. Then, we need to prove that $c$ is the minimum decrement. Consider another price vector Consider another price vector $\mathbf{P''}$, $P'_j=\begin{cases} P_j, & j \notin S \\ P_j-d, & j \in S \end{cases}$, where $d<c$. It is straightforward that for all $i \in \mathcal{B} \setminus N(S)$, $v_{i,j}-P''_j< v_{i,j}-P_j+c \le \max_k (v_{i,k}-P_k)$. Hence, no buyers will be added to the $N(S)$. Now, it is clear that $\min_{i \in \mathcal{B} \setminus N(S),l\in S} \{\max_{k \in \mathcal{M} \setminus S}(v_{i,k}-P_k)-(v_{i,l}-P_l) \}$ is the minimum price reduction which guarantees to add at least a new buyer to the $N(S)$. ## Proof of Lemma [\[lem2\]](#lem2){reference-type="ref" reference="lem2"} {#pf:lem2} In order to prove the statement, we start from showing the most skewed set is always a constricted good set when there is no perfect matching. Then, we prove the uniqueness of the most skewed set by contradiction. When there is no perfect matching, there must exists a constricted good set. By definition, the constricted good set $S$ has the property $\|S\|>\|N(S)\|$. Since $\|S\|$, $\|N(S)\|$ are integers, the skewness of a constricted good set \[\label{eq1} f(S)=\|S\|-\|N(S)\|+\frac{1}{\|S\|} \geq 1+\frac{1}{\|S\|} >1.\] Then, for any non-constricted good set $S'$, $\|S'\| \leq \|N(S')\|$. The skewness of $S'$ is \[\label{eq2} f(S')=\|S'\|-\|N(S')\|+\frac{1}{\|S'\|} \leq 0+\frac{1}{\|S'\|} \leq 1.\] With equation ([\[eq1\]](#eq1){reference-type="ref" reference="eq1"}), ([\[eq2\]](#eq2){reference-type="ref" reference="eq2"}), the skewness of a constricted good set is always greater than any non-constricted good sets. Therefore, if a preference graph exists a constricted good set, the most skewed set is always a constricted good set. Then, we start to prove the uniqueness of the most skewed set. Suppose there exists two disjoint sets $S_1$, $S_2$ and both sets are the most skewed sets, i.e., $f(S_1)=f(S_2)= \max_{S \subset \mathcal{M}, S \neq \emptyset} f(S)$. Consider the skewness of the union of $S_1$ and $S_2$. \[\begin{aligned} &&f(S_1 \cup S_2) \\ &=& \|S_1 \cup S_2\|-\|N(S_1 \cup S_2)\|+ \frac{1}{\|S_1 \cup S_2\|} \\ &=& \|S_1\|+ \|S_2\|-\|N(S_1) \cup N(S_2)\|+ \frac{1}{\|S_1 \cup S_2\|} \\ &\geq & \|S_1\|+ \|S_2\|-\|N(S_1)\|-\|N(S_2)\|+ \frac{1}{\|S_1 \cup S_2\|} \\ &=& f(S_1)+ \|S_2\|-\|N(S_2)\|-\frac{1}{\|S_1\|}+\frac{1}{\|S_1 \cup S_2\|} \label{eq3}\\ &\geq & f(S_1) +1-\frac{1}{\|S_1\|}+\frac{1}{\|S_1 \cup S_2\|} > f(S_1) \label{eq4} \end{aligned}\] From ([\[eq3\]](#eq3){reference-type="ref" reference="eq3"}) to ([\[eq4\]](#eq4){reference-type="ref" reference="eq4"}) is true because $S_2$ is a constricted good set. ([\[eq4\]](#eq4){reference-type="ref" reference="eq4"}) contradicts our assumption that $S_1$, $S_2$ are the most skewed set. Therefore, we know that if there exists multiple sets share the same highest skewness value, these sets are not disjoint. Now, suppose there exists two sets $S_1$, $S_2$ satisfying that $S_1 \cup S_2 \neq \emptyset$ and both sets are the most skewed sets, the following two inequalities must hold: \[\begin{aligned} f(S_1)-f(S_1 \cup S_2) \geq 0 \\ f(S_2)-f(S_1 \cap S_2) \geq 0 \end{aligned}\] Let's sum up the two inequalities and represent the formula in twelve terms. \[\begin{aligned} &&f(S_1)+f(S_2)-f(S_1 \cup S_2)-f(S_1 \cap S_2) \\ &=& \|S_1\|+\|S_2\|-\|S_1 \cup S_2\|-\|S_1 \cap S_2\| \nonumber \\ && +\|N(S_1 \cup S_2)\|+ \|N(S_1 \cap S_2)\|-\|N(S_1)\|-\|N(S_2)\| \nonumber \\ && +\dfrac{1}{\|S_1\|}+\dfrac{1}{\|S_2\|}-\dfrac{1}{\|S_1 \cup S_2\|}-\dfrac{1}{\|S_1 \cap S_2\|} \end{aligned}\] The first four terms $\|S_1\|+\|S_2\|-\|S_1 \cup S_2\|-\|S_1 \cap S_2\|=0$. Using the similar argument, $\|N(S_1)\|+\|N(S_2)\|=\|N(S_1) \cap N(S_2)\|+\|N(S_1) \cup N(S_2)\|$. \[\begin{aligned} \|N(S_1 \cup S_2)\|=\|N(S_1) \cup N(S_2)\| \\ \|N(S_1 \cap S_2)\| \leq \|N(S_1) \cap N(S_2)\| \label{eq5} \end{aligned}\] Equation ([\[eq5\]](#eq5){reference-type="ref" reference="eq5"}) is true because there may exist some elements in $S_1 \setminus S_2$ and $S_2 \setminus S_1$ but have common neighbors. Thus, the second four terms are smaller than or equal to $0$. To check the last four terms, let $\|S_1\|=a$, $\|S_2\|=b$, and $\|S_1 \cap S_2\|=c$, where $c<\min\{a,b\}$ because $S_1,S_2$ are not disjoint. The last four terms are \[\begin{aligned} &&\dfrac{1}{\|S_1\|}+\dfrac{1}{\|S_2\|}-\dfrac{1}{\|S_1 \cup S_2\|}-\dfrac{1}{\|S_1 \cap S_2\|} \\ &=& \dfrac{1}{a}+\dfrac{1}{b}-\dfrac{1}{a+b-c}-\dfrac{1}{c} \\ &=& \dfrac{a+b}{ab}-\dfrac{a+b}{(a+b-c)c} \\ &=& \dfrac{a+b}{abc(a+b-c)}(ac+bc-c^2-ab) \\ &=&-\dfrac{(a+b)(a-c)(b-c)}{abc(a+b-c)}<0 \end{aligned}\] To conclude, the first four terms are $0$, the second four terms are smaller than or equal to $0$, and the last four terms are strictly negative make \[\begin{aligned} f(S_1)+f(S_2)-f(S_1 \cup S_2)-f(S_1 \cap S_2)<0. \end{aligned}\] Therefore, at least one of set $S_1 \cup S_2$, $S_1 \cap S_2$ has the skewness value greater than $f(S_1)=f(S_2)$, which leads to a contradiction that $S_1$ and $S_2$ are the most skewed set. Finally, we can claim that the most skewed set is unique when there is no perfect matching. ## Proof of Lemma [\[lem:conv\]](#lem:conv){reference-type="ref" reference="lem:conv"} {#pf:lem:conv} First, we prove that the algorithm terminates in finite (at most $\|\mathcal{M}\|^3$) rounds by investigating the relationship of the most skewed sets in the consecutive rounds, $S^_t$, $S^_{t+1}$. The relationship of $S^_t$, $S^_{t+1}$ has four cases. 1. $S^_t=S^_{t+1}$ 2. $S^_t \subset S^_{t+1}$ 3. $S^_t \supset S^_{t+1}$ 4. $S^_t \nsubseteq S^_{t+1}$ and $S^_t \nsupseteq S^_{t+1}$ Recall that $W(G)$ is the skewness of the graph $G$ and $N_t(S)$ is the neighbor of $S$ based on the preference graph at round $k$.\ In case 1, $W(G_t)-W(G_{t+1})=f_t(S^_t)-f_{t+1}(S^_{t+1})=f_t(S^_t)-f_{t+1}(S^_t)\geq 1$.\ In case 2, define $S'=S^_{t+1}\setminus S^_t$, and it is trivial that $S^_t \subset S^_{t+1}$ implies $1>\frac{1}{\|S^_t\|}-\frac{1}{\|S^_{t+1}\|}>0$. \[\begin{aligned} &&W(G_t)-W(G_{t+1})=f_t(S^_t)-f_{t+1}(S^_{t+1})\\ &=&\|S^_t\|-\|N_t(S^_t)\|-\|S^_{t+1}\|+\|N_{t+1}(S^_{t+1})\| +\dfrac{1}{\|S^_t\|}-\dfrac{1}{\|S^_{t+1}\|} \label{eq6}\\ &\geq &\|S^_{t+1}\|-\|N_t(S^_{t+1})\|-\|S^_{t+1}\|+\|N_{t+1}(S^_{t+1})\|+\dfrac{1}{\|S^_t\|}-\dfrac{1}{\|S^_{t+1}\|} \label{eq7}\\ &=&\|N_{t+1}(S^_t \cup S')\|-\|N_t(S^_{t+1})\|+\frac{\|S^_{t+1}\|-\|S^_t\|}{\|S^_t\|\|S^_{t+1}\|} \nonumber \\ &\geq &\|N_{t+1}(S^_t \cup S')\|-\|N_t(S^_{t+1})\| + \dfrac{1}{\|\mathcal{M}\|(\|\mathcal{M}\|-1)} \label{eq8}\\ &=&\|N_{t+1}(S^_t)\|+\|N_{t+1}(S')\cap N_{t+1}(S^_t)^c\|-\|N_t(S^_{t+1})\|+ \dfrac{1}{\|\mathcal{M}\|^2-\|\mathcal{M}\|} \label{eq9}\\ &=&\|N_t(S^_t)\|+\|N_{t+1}(S_t)\setminus N_t(S^_t)\|-\|N_t(S^_{t+1})\| +\|N_{t+1}(S')\cap N_{t+1}(S^_t)^c\|+ \tfrac{1}{\|\mathcal{M}\|^2-\|\mathcal{M}\|} \label{eq10} \end{aligned}\] Before going to further steps, we have to briefly explain the logic behind the above equations. ([\[eq6\]](#eq6){reference-type="ref" reference="eq6"}) to ([\[eq7\]](#eq7){reference-type="ref" reference="eq7"}) is true because $N_t(S^_t) \subseteq N_t(S^_{t+1})$. ([\[eq8\]](#eq8){reference-type="ref" reference="eq8"}) to ([\[eq9\]](#eq9){reference-type="ref" reference="eq9"}) is to expand $\|N_{t+1}(S^_t \cup S')\|$ to $\|N_{t+1}(S^_t)\|+\|N_{t+1}(S')\cap N_{t+1}(S^_t)^c\|$. ([\[eq9\]](#eq9){reference-type="ref" reference="eq9"}) to ([\[eq10\]](#eq10){reference-type="ref" reference="eq10"}) is to expand $\|N_{t+1}(S^_t)\|$ to $\|N_t(S^_t)\|+\|N_{t+1}(S_t)\setminus N_t(S^_t)\|$. Since $\|N_{t+1}(S')\cap N_{t+1}(S^_t)^c\|=\|N_{t+1}(S')\cap N_{k}(S^_t)^c\|-\|N_{t+1}(S')\cap (N_{t+1}(S_t)\setminus N_t(S^_t))\|$, and $\|N_{t+1}(S_t)\setminus N_t(S^_t)\|\geq \|N_{t+1}(S')\cap (N_{t+1}(S_t)\setminus N_t(S^_t))\|$, we can further summarize the first four terms in ([\[eq10\]](#eq10){reference-type="ref" reference="eq10"}). \[\begin{aligned} &&\|N_t(S^_t)\|+\|N_{t+1}(S_t)\setminus N_t(S^_t)\| +\|N_{t+1}(S')\cap N_{t+1}(S^_t)^c\|-\|N_t(S^_{t+1})\| \\ &\geq & \|N_t(S^_t)\|+\|N_{t+1}(S')\cap N_{k}(S^_t)^c\|-\|N_t(S^_{t+1})\|\\ &=& \|N_t(S^_t)\|+\|N_{t+1}(S')\cap N_t(S^_t)^c\|-\|N_t(S^_t)\|-\|N_t(S')\cap N_t(S^_t)^c\| \\ &=& \|N_{t+1}(S')\cap N_t(S^_t)^c\|-\|N_t(S')\cap N_t(S^_t)^c\| \label{eq32} \\ &=& \|N_t(S')\cap N_t(S^_t)^c\|-\|N_t(S')\cap N_t(S^_t)^c\|=0 \label{eq33} \end{aligned}\] ([\[eq32\]](#eq32){reference-type="ref" reference="eq32"}) to ([\[eq33\]](#eq33){reference-type="ref" reference="eq33"}) is true because $S'$ is not in $S^_t$, the neighbor of $S'$ not in the neighbor of $S^_t$ remains the same from $k$ to $t+1$ round. With equation ([\[eq10\]](#eq10){reference-type="ref" reference="eq10"}) and ([\[eq33\]](#eq33){reference-type="ref" reference="eq33"}), we can conclude that $W(G_t)-W(G_{t+1})\geq \frac{1}{\|\mathcal{M}\|^2-\|\mathcal{M}\|}$ in case 2.\ \ In case 3, since every elements in $S^_t$ belongs to $S^_t$, $f_t(S^_{t+1}) \geq f_{t+1}(S^_{t+1})$. Given that $\dfrac{1}{\|S^_t\|}<\dfrac{1}{\|S^_{t+1}\|}$ and $f_t(S^_t)-f_t(S^_{t+1})>0$ by definition, $\|S^_t\|-\|S^_{t+1}\|+\|N_t(S^_t)\|-\|N_t(S^_{t+1})\|\geq 1$. With the knowledge that $S^_t$ and $S^_t$ are non-empty set, it is obvious that $f_t(S^_t)-f_t(S^_{t+1})$ is lower-bounded by $\tfrac{1}{2}$. Therefore, we can conclude that $W(G)_t-W(G_{t+1})=f_t(S^_t)-f_{t+1}(S^_{t+1}) \geq f_t(S^_t)-f_t(S^_{t+1}) \geq \frac{1}{2}$ in case 3.\ \ In case 4, define $S'=S^_{t+1} \setminus S^_t$, $S''=S^_t \setminus S^_{t+1}$, and $T=S^_t \cap S^_{t+1}$. \[\begin{aligned} && W(G_t)-W(G_{t+1}) \\ &=&f_t(S^_t)-f_{t+1}(S^_{t+1}) \\ &=& \|T\|+\|S''\|-\|N_t(T)\|-\|N_t(S'')\setminus N_t(T)\|+\frac{1}{\|S^_t\|}-\|T\|-\|S'\|+\|N_{t+1}(S^_{t+1})\|-\frac{1}{\|S^_{t+1}\|} \\ &=& \|S''\|-\|N_t(S'')\setminus N_t(T)\|+\frac{1}{\|S^_t\|}-\frac{1}{\|S^_{t+1}\|} +\|N_{t+1}(S^_{t+1})\|-\|N_t(T)\|-\|S'\| \\ &\geq & 1+\frac{1}{\|S^_t\|}-\frac{1}{\|S^_{t+1}\|}+\|N_{t+1}(S^_{t+1})\|-\|N_t(T)\|-\|S'\| \\ &\geq & 1+\frac{1}{\|\mathcal{M}\|}-1+\|N_{t+1}(T))\|-\|N_t(T)\|+\|N_{t+1}(S')\setminus N_{t+1}(T)\|-\|S'\| \\ &= & \frac{1}{\|\mathcal{M}\|}+\|N_{t+1}(T))\|-\|N_t(T)\| +\|N_{t+1}(S')\setminus N_{t+1}(T)\|-\|S'\| \label{eq11}\\ &\geq & \frac{1}{\|\mathcal{M}\|}+\|N_{t+1}(S')\setminus N_{k}(T)\|-\|S'\| \geq \frac{1}{\|\mathcal{M}\|} \label{eq12}\\ \end{aligned}\] Most of the equations in case 4 are straight-forward except from ([\[eq11\]](#eq11){reference-type="ref" reference="eq11"}) to ([\[eq12\]](#eq12){reference-type="ref" reference="eq12"}). ([\[eq11\]](#eq11){reference-type="ref" reference="eq11"}) to ([\[eq12\]](#eq12){reference-type="ref" reference="eq12"}) is true because of the following inequalities: \[\begin{aligned} &&\|N_{t+1}(S')\setminus N_{t+1}(T)\|+\|N_{t+1}(T))\|-\|N_t(T)\| \nonumber \\ &=&\|N_{t+1}(S')\setminus N_t(T)\|+\|N_{t+1}(T))\|-\|N_t(T)\|-\| \{N_{t+1}(S') \cap N_{t+1}(T) \}\setminus N_t(T)\| \\ &\geq & \|N_{t+1}(S')\setminus N_t(T)\|-\| N_{t+1}(T)\setminus N_t(T)\| +\|N_{t+1}(T))\|-\|N_t(T)\| \\ &=& \|N_{t+1}(S')\setminus N_t(T)\| \end{aligned}\] Combine these four cases, we know that \[\begin{aligned} W(G_t)-W(G_{t+1}) \geq \dfrac{1}{ \|\mathcal{M}\|^2-\|\mathcal{M}\|}. \end{aligned}\] Therefore, we can conclude that the proposed algorithm terminates in finite rounds. (At most $\|\mathcal{M}\|^3$ rounds because $W(G)<\|\mathcal{M}\|$ and minimum decrement is greater than $\tfrac{1}{\|\mathcal{M}\|^3}$). ## Proof of Lemma [\[lem:colorMSS\]](#lem:colorMSS){reference-type="ref" reference="lem:colorMSS"} First, we want to show that suppose $\mathcal{M}_g \cup\mathcal{M}_b$ is a constricted good set, we can not include any good colored red to increase the skewness of the set. For any set of good colored red $S_r$ being added to $\mathcal{M}_g \cup\mathcal{M}_b$, at least the same size of buyers being matched pairs of those red goods are join to $N(\mathcal{M}_g \cup\mathcal{M}_b \cup S_r)$. Since we know that $\|N(S_r)\setminus N(\mathcal{M}_g \cup\mathcal{M}_b)\|\geq \|S_r\|$, including any set of red goods will decrease the skewness of the set $\mathcal{M}_g \cup\mathcal{M}_b$. Then, we want to show remove any subset of goods $S\subseteq \mathcal{M}_g \cup\mathcal{M}_b$ will also reduce the skewness of the set. Clearly, removing any subset of blue good will not reduce $N(\mathcal{M}_g \cup\mathcal{M}_b)$, hence blue goods are definitely included in the maximally skewed set. Therefore, we only need to consider the impact of removing set of green goods $S \subseteq \mathcal{M}_g$. Since for any $S \subseteq \mathcal{M}_g$, there has has at least one red edge connecting $S$ and $N(\mathcal{M}_g \cup\mathcal{M}_b\setminus S)$ according to the algorithm, Hence, $\|N(\mathcal{M}_g \cup\mathcal{M}_b)\|-\|N(\mathcal{M}_g \cup\mathcal{M}_b\setminus S)\|<\|S\|$, which implies that removing any $S \subset \mathcal{M}_g$ will only increase the skewness of the set. With these facts and the uniqueness property of the maximally skewed set, we can conclude that $\mathcal{M}_g \cup\mathcal{M}_b$ is the maximally skewed set. ## Proof of Lemma [\[lem:alg2\]](#lem:alg2){reference-type="ref" reference="lem:alg2"} {#pf:lemalg2} We will prove Algorithm [\[alg:alg2\]](#alg:alg2){reference-type="ref" reference="alg:alg2"} always return the most skewed set by contradiction.\ Since every untraversed good upon termination can be matched with an untraversed buyer without repetition. Therefore, adding any set of runtraversed goods $S_U$ to the set $S$ will always reduce the skewness of $S$ (because the increase of cardinality in neighbor of $S$, $\|N(S_U)\|-\|N(S)\|$ is always greater than or equal to the increase of cardinality of $S$, $\|S_U\|-\|S\|$). Hence, the most skewed set will never contain any untraversed good. Suppose there exists a set $S'$ is the most skewed set, with a higher skewness than the set $S$ return by Algorithm [\[alg:alg2\]](#alg:alg2){reference-type="ref" reference="alg:alg2"}, $S'$ must be a subset of $S$ because there's no untraversed good in the most skewed set and $f(S')>f(S)$. Let $S^=S\setminus S'$, $f(S')>f(S)$ and $S' \subset S$ implies $\|N(S^)\setminus N(S')\|-\|S^\|>0$. If this happens, there must be a non-empty set of matching pairs match nodes from $S^$ to $N(S^)\setminus N(S')$ without repetition. Since $N(S^)\setminus N(S')$ is not in $N(S')$ under the directed graph, nodes in $S^$ will not be traversed in the algorithm contradicts that $S^ \subset S$. ## Proof of Theorem [\[lem4.1\]](#lem4.1){reference-type="ref" reference="lem4.1"} {#pf:lem4.1} First, let's begin with the proof of the first half statement of the theorem, which is a variational characterization of MCPs. ($\Rightarrow$) This direction is obvious, otherwise the MCP is not the maximum by definition. ($\Leftarrow$) Recall that $\mathcal{M}$ is the set of goods. Suppose there exists an MCP $P^1$ satisfying the conditions but it is not the maximum MCP. Then there must exist a set of goods $S_1$ such that for all $i \in S_1$, $P^1_i<P^_i$, and for all $i \in \mathcal{M}-S_1$, $P^1_i \geq P^_i$; $\mathcal{M}-S_1$ can be an empty set. Let $P^2_i= P^1_i$ for all $i \in \mathcal{M}-S_1$, and $P^2_i= P^_i$ for all $i \in S_1$. We will verify that $P^2$ is an MCP. WLOG, we can assume $P^1$ and $P^$ have the same allocation; this is true as every MCP supports all efficient matchings. Then, consider any buyer who is assigned a good in $S_1$ under $P^$. When the price vector changes from $P^$ to $P^2$, the buyer has no profitable deviation from his/her assigned good because $P^2_i \geq P^_i$ for all $i$ and $P^2_j= P^_j$ for all $j \in S_1$. Similarly, when the price vector changes from $P^1$ to $P^2$, the buyer who is assigned a good in $\mathcal{M} \setminus S_1$ under $P^1$ has no profitable deviation because $P^2_i=P^1_i$ for all $i \in \mathcal{M} \setminus S_1$ and $P^2_j>P^1_j$ for all $j \in S_1$. Finally, since $P^1$ and $P^$ have the same allocation, no buyer will deviate if we assign this allocation to buyers under $P^2$. Since all buyers have non-negative surpluses under $P^2$, it follows that $P^2$ is an MCP. Given $P^1$, there exists a set of goods $S_1$ whose price we can increase and still get market clearing because both $P^1$ and $P^2$ are MCPs. This contradicts the assumption that $P^1$ satisfies the stated conditions, and the proof of the variational characterization follows. For the second half part, which is a combinatorial characterization of MCPs, let's try to prove the statement using the result of the variational characterization. Given that we cannot increase the price for any subset of goods, it implies that for any subset of goods $S$, the set of corresponding matched buyer, either there exists at least one buyer has an edge connected to a good not in this subset or there exists at least one buyer with surplus zero. In the former case, it is obvious that $B<N(B)$ for such corresponding buyer set $B$. In the later case, $B\leq N(B)$ and $D\in N^D(B)$ guarantees $B < N^D(B)$. For the opposite direction, if every set of buyer with $B < N^D(B)$ under the current MCP, increasing the price for any set of good will make at least one corresponding buyer deviate from the matched good and cause no perfect matching. Therefore, the condition in variational characterization holds if and only if the condition in the combinatorial characterization holds. ## Proof of Theorem [\[thm2\]](#thm2){reference-type="ref" reference="thm2"} {#pf:thm2} As mentioned earlier, showing the algorithm returns the maximum MCP is equivalent to show that when the algorithm terminates, the bipartite structure guarantees that we can not increase the price of any subset of good with the result in Theorem [\[lem4.1\]](#lem4.1){reference-type="ref" reference="lem4.1"}. It implies that after adding a dummy good with zero price, the support of any subset of goods has a size less than the size of the subset of goods, i.e., $\|B\|< \|N^D_T(B)\|$. Therefore, the proof of the theorem can be transformed to prove that the algorithm satisfies $\|B\|< \|N^D_T(B)\|$ for any non-empty set of buyer $B$ on termination. Let's start the proof with several claims. Since our algorithm returns an MCP vector, if $\|B\|> \|N^D_T(B)\|$, there does not exist a perfect matching because of the existence of constricted buyer set. $\|B\|= \|N^D_T(B)\|$ and $D\in N^D_T(B)$, imply $\|B\|> \|N_T(B)\|$. The preference graph has constricted sets and has no perfect matching. Now, we can prove Claim [\[cl3\]](#cl3){reference-type="ref" reference="cl3"} by induction. At t=0, $\mathscr{B}^s_0=\emptyset$. At t=1, $\mathscr{B}^s_1=\mathcal{N_0}(S^_0)$. With Claim [\[cl4\]](#cl4){reference-type="ref" reference="cl4"}, $\|B\|<\|N_1(B)\|~~\forall B\subseteq \mathscr{B}^s_1$ and $B \neq \emptyset$. At a finite time $t$, suppose for all $B\subseteq \mathscr{B}^s_t$, $B \neq \emptyset$ satisfy $\|B\|<\|N_t(B)\|$, consider at time $t+1$: Since $\mathscr{B}^s_t \subseteq \mathscr{B}^s_{t+1}$, $\mathscr{B}^s_{t+1}$ contains three disjoint components: \[\mathscr{B}^s_{t+1}= \{\mathscr{B}^s_t \cap {N_t(S^_t)}^c \} \cup \{\mathscr{B}^s_t \cap N_t(S^_t) \} \cup \{{\mathscr{B}^s_t}^c \cap N_t(S^_t) \}\] Buyers in the first two parts are originally with positive utilities. Buyers in the last part have zero utilities at time $t$ but have positive utilities at time $t+1$. Consider the subset of buyers $B_{\alpha} \subseteq \{ \mathscr{B}^s_t \cap {N_t(S^_t)}^c \}$. Since every $b \in B_{\alpha}$ does not prefer any good in $S^_t$, the price reduction in $S^_t$ will never remove any edges between $\{ \mathscr{B}^s_t \cap {N_t(S^_t)}^c \}$ and $N_t(\{ \mathscr{B}^s_t \cap {N_t(S^_t)}^c \})$. Therefore, for any non-empty set of buyers $B_{\alpha}$, $\|B_{\alpha}\|< \|N_t(B_{\alpha})\|\leq \|N_{t+1}(B_{\alpha})\|$. Then, consider the second and the third parts. Since $\{\mathscr{B}^s_t \cap N_t(S^_t) \} \cup \{{\mathscr{B}^s_t}^c \cap N_t(S^_t) \}=N_t(S^_t)$, every non-empty set of buyers $B_{\beta} \subseteq N_t(S^_t)$ satisfies $\|B_{\beta}\|<\|N_{t+1}(B_{\beta})\|$ by Claim [\[cl4\]](#cl4){reference-type="ref" reference="cl4"}. At the last step, consider $B_{\gamma}=B_{\gamma_1} \cup B_{\gamma_2}$, where $B_{\gamma_1} \subseteq \{\mathscr{B}^s_t \cap {N_t(S^_t)}^c \}$, $B_{\gamma_2} \subseteq N_t(S^_t)$ and $B_{\gamma_1},B_{\gamma_2} \neq \emptyset$. \[\begin{aligned} \|N_{t+1}(B_{\gamma})\| &=&\|N_{t+1}(B_{\gamma_1}) \cup N_{t+1}(B_{\gamma_2})\| \\ &=& \|N_{t+1}(B_{\gamma_1})\|+\|N_{t+1} (B_{\gamma_2})\|-\|N_{t+1}(B_{\gamma_1}) \cap N_{t+1}(B_{\gamma_2})\| \\ &=& \|N_{t+1}(B_{\gamma_1})\|-\|N_{t+1}(B_{\gamma_1}) \cap N_{t+1}(B_{\gamma_2})\|+\|N_{t+1} (B_{\gamma_2})\| \\ &=& \|N_{t+1}(B_{\gamma_1}) \cap {N_{t+1}(B_{\gamma_2})}^c\|+\|N_{t+1} (B_{\gamma_2})\| \\ &\geq& \|N_{t+1}(B_{\gamma_1}) \cap {N_{t+1}(N_t(S^_t))}^c\|+\|N_{t+1} (B_{\gamma_2})\| \label{eqx1} \\ &=& \|N_{t+1}(B_{\gamma_1}) \cap {S^_t}^c\|+\|N_{t+1} (B_{\gamma_2})\| \label{eqx2}\\ &=& \|N_t(B_{\gamma_1})\|+\|N_{t+1} (B_{\gamma_2})\| \label{eqx3}\\ &>& \|B_{\gamma_1})\|+\|B_{\gamma_2}\|=\|B_{\gamma}\| \end{aligned}\] From ([\[eqx1\]](#eqx1){reference-type="ref" reference="eqx1"}) to ([\[eqx2\]](#eqx2){reference-type="ref" reference="eqx2"}) is true because after price reduction at time $t$, buyer belongs to the neighbor of constricted good set $S^_t$ will only prefer goods in $S^_t$, therefore $N_{t+1}(N_t(S^_t))=S^_t$. From ([\[eqx2\]](#eqx2){reference-type="ref" reference="eqx2"}) to ([\[eqx3\]](#eqx3){reference-type="ref" reference="eqx3"}) is true because $N_t(B_{\gamma_1}) \cap S^_t = \emptyset$ (by definition). Then, we have discussed before that the price reduction in $S^_t$ will never remove any edges between $\{ \mathscr{B}^s_t \cap {N_t(S^_t)}^c \}$ and $N_t(\{ \mathscr{B}^s_t \cap {N_t(S^_t)}^c \})$. Since goods not in the most skewed set at time $t$ will not add new buyers to their neighbor, the equivalence between $N_{t+1}(B_{\gamma_1}) \cap {S^_t}^c$ and $N_t(B_{\gamma_1})$ holds. Therefore, for every non empty set of buyers $B\subseteq \mathscr{B}^s_{t+1}$, $\|B\|<\|N_{t+1}(B)\|$. The mathematical induction works for all $k \in \mathbb{N}$. Since our algorithm terminates in finite round, $\|B\|< \|N_T(B)\|$, Q.E.D. ◻ Since our algorithm returns an MCP vector, if $\|B\|> \|N^D_T(B)\|$, there does not exist a perfect matching because of the existence of constricted buyer set. $\|B\|= \|N^D_T(B)\|$ and $D\in N^D_T(B)$, imply $\|B\|> \|N_T(B)\|$. The preference graph has constricted sets and has no perfect matching. Now, we can prove Claim [\[cl3\]](#cl3){reference-type="ref" reference="cl3"} by induction. At t=0, $\mathscr{B}^s_0=\emptyset$. At t=1, $\mathscr{B}^s_1=\mathcal{N_0}(S^_0)$. With Claim [\[cl4\]](#cl4){reference-type="ref" reference="cl4"}, $\|B\|<\|N_1(B)\|~~\forall B\subseteq \mathscr{B}^s_1$ and $B \neq \emptyset$. At a finite time $t$, suppose for all $B\subseteq \mathscr{B}^s_t$, $B \neq \emptyset$ satisfy $\|B\|<\|N_t(B)\|$, consider at time $t+1$: Since $\mathscr{B}^s_t \subseteq \mathscr{B}^s_{t+1}$, $\mathscr{B}^s_{t+1}$ contains three disjoint components: \[\mathscr{B}^s_{t+1}= \{\mathscr{B}^s_t \cap {N_t(S^_t)}^c \} \cup \{\mathscr{B}^s_t \cap N_t(S^_t) \} \cup \{{\mathscr{B}^s_t}^c \cap N_t(S^_t) \}\] Buyers in the first two parts are originally with positive utilities. Buyers in the last part have zero utilities at time $t$ but have positive utilities at time $t+1$. Consider the subset of buyers $B_{\alpha} \subseteq \{ \mathscr{B}^s_t \cap {N_t(S^_t)}^c \}$. Since every $b \in B_{\alpha}$ does not prefer any good in $S^_t$, the price reduction in $S^_t$ will never remove any edges between $\{ \mathscr{B}^s_t \cap {N_t(S^_t)}^c \}$ and $N_t(\{ \mathscr{B}^s_t \cap {N_t(S^_t)}^c \})$. Therefore, for any non-empty set of buyers $B_{\alpha}$, $\|B_{\alpha}\|< \|N_t(B_{\alpha})\|\leq \|N_{t+1}(B_{\alpha})\|$. Then, consider the second and the third parts. Since $\{\mathscr{B}^s_t \cap N_t(S^_t) \} \cup \{{\mathscr{B}^s_t}^c \cap N_t(S^_t) \}=N_t(S^_t)$, every non-empty set of buyers $B_{\beta} \subseteq N_t(S^_t)$ satisfies $\|B_{\beta}\|<\|N_{t+1}(B_{\beta})\|$ by Claim [\[cl4\]](#cl4){reference-type="ref" reference="cl4"}. At the last step, consider $B_{\gamma}=B_{\gamma_1} \cup B_{\gamma_2}$, where $B_{\gamma_1} \subseteq \{\mathscr{B}^s_t \cap {N_t(S^_t)}^c \}$, $B_{\gamma_2} \subseteq N_t(S^_t)$ and $B_{\gamma_1},B_{\gamma_2} \neq \emptyset$. \[\begin{aligned} \|N_{t+1}(B_{\gamma})\| &=&\|N_{t+1}(B_{\gamma_1}) \cup N_{t+1}(B_{\gamma_2})\| \\ &=& \|N_{t+1}(B_{\gamma_1})\|+\|N_{t+1} (B_{\gamma_2})\|-\|N_{t+1}(B_{\gamma_1}) \cap N_{t+1}(B_{\gamma_2})\| \\ &=& \|N_{t+1}(B_{\gamma_1})\|-\|N_{t+1}(B_{\gamma_1}) \cap N_{t+1}(B_{\gamma_2})\|+\|N_{t+1} (B_{\gamma_2})\| \\ &=& \|N_{t+1}(B_{\gamma_1}) \cap {N_{t+1}(B_{\gamma_2})}^c\|+\|N_{t+1} (B_{\gamma_2})\| \\ &\geq& \|N_{t+1}(B_{\gamma_1}) \cap {N_{t+1}(N_t(S^_t))}^c\|+\|N_{t+1} (B_{\gamma_2})\| \label{eqx1} \\ &=& \|N_{t+1}(B_{\gamma_1}) \cap {S^_t}^c\|+\|N_{t+1} (B_{\gamma_2})\| \label{eqx2}\\ &=& \|N_t(B_{\gamma_1})\|+\|N_{t+1} (B_{\gamma_2})\| \label{eqx3}\\ &>& \|B_{\gamma_1})\|+\|B_{\gamma_2}\|=\|B_{\gamma}\| \end{aligned}\] From ([\[eqx1\]](#eqx1){reference-type="ref" reference="eqx1"}) to ([\[eqx2\]](#eqx2){reference-type="ref" reference="eqx2"}) is true because after price reduction at time $t$, buyer belongs to the neighbor of constricted good set $S^_t$ will only prefer goods in $S^_t$, therefore $N_{t+1}(N_t(S^_t))=S^_t$. From ([\[eqx2\]](#eqx2){reference-type="ref" reference="eqx2"}) to ([\[eqx3\]](#eqx3){reference-type="ref" reference="eqx3"}) is true because $N_t(B_{\gamma_1}) \cap S^_t = \emptyset$ (by definition). Then, we have discussed before that the price reduction in $S^_t$ will never remove any edges between $\{ \mathscr{B}^s_t \cap {N_t(S^_t)}^c \}$ and $N_t(\{ \mathscr{B}^s_t \cap {N_t(S^_t)}^c \})$. Since goods not in the most skewed set at time $t$ will not add new buyers to their neighbor, the equivalence between $N_{t+1}(B_{\gamma_1}) \cap {S^*_t}^c$ and $N_t(B_{\gamma_1})$ holds. Therefore, for every non empty set of buyers $B\subseteq \mathscr{B}^s_{t+1}$, $\|B\|<\|N_{t+1}(B)\|$. The mathematical induction works for all $k \in \mathbb{N}$. Since our algorithm terminates in finite round, $\|B\|< \|N_T(B)\|$, Q.E.D. ◻ [\[alg:alg3\]]{#alg:alg3 label="alg:alg3"} In Lemma [\[lem8\]](#lem8){reference-type="ref" reference="lem8"}, we addressed the connectivity property in sponsored search markets during the execution of our algorithm. Unfortunately, the connectivity does not always hold in general matching markets. Here we demonstrate this by a $4$-by-$4$ counter example below. To make this case as simple as possible, we just consider a realized valuation matrix \[\begin{aligned} V= \begin{bmatrix} 5 & 4 & 1 & 1 \\ 3 & 3 & 2 & 2 \\ 2 & 2 & 3 & 3 \\ 1 & 1 & 4 & 5 \end{bmatrix}. \end{aligned}\] Initially, the first two goods are preferred by buyer 1, and the last two goods are preferred by buyer buyer 4. In our algorithm, the most skewed set contains all goods, and price is reduced by one in the first round. After that, the MCP vector $\mathbf{P}=[4,3,3,4]$ and the bipartite graphs of initial and final steps are in Fig. [\[fig:connectivity\]](#fig:connectivity){reference-type="ref" reference="fig:connectivity"}. Fig. [\[fig:connectivity\]](#fig:connectivity){reference-type="ref" reference="fig:connectivity"} demonstrates that both the initial and final bipartite graphs are disconnected in our algorithm as there are two components. Therefore, preserving the connectivity property in the running of our algorithm holds in sponsored search markets, and may not otherwise. ## Further Comments on Strategic Bidding Unlike most of the non-incentive compatible auction mechanisms that overbidding above buyers' actual value is always irrational, a buyer may benefit from bidding higher than her actual value to enjoy a lower price under the proposed algorithm in some scenarios. We can show this by a simple example below. First, we present two valuations matrices that will yield the same matching but different prices with the difference in the matrices only being the valuations of one buyer. These are \[\begin{aligned} V_1= \begin{bmatrix} 4 & 5 & 5 & 6 \\ 2 & 4 & 5 & 5 \\ 1 & 2 & 4 & 5 \\ 0 & 1 & 2 & 4 \end{bmatrix}, \quad V_2 = \begin{bmatrix} 4 & 5.9 & 5.9 & 6.9 \\ 2 & 4 & 5 & 5 \\ 1 & 2 & 4 & 5 \\ 0 & 1 & 2 & 4 \end{bmatrix}. \end{aligned}\] The maximum MCPs are $P_1=[1, 2, 3, 4]$ and $P_2=[0.1,2,3,4]$, respectively, and in both cases buyer $i$ gets matched to good $i$ for all $i\in \{1,\dotsc,4\}$. Now if we assume that the true valuations of all the buyers is given by matrix $V_1$, then by unilaterally deviating to a higher bid, buyer $1$ is able to dramatically increase her payoff without impacting any other buyers. While this is only an example, we believe that this will occur with non-zero probability and will necessitate an alternate means of determining the equilibrium bidding behavior. # Preliminary Analysis of Strategic Buyers and Bayesian-Nash Equilibrium {#sec5} As mentioned in the introduction, the proposed descending price algorithm is not incentive compatible, and so we need to explore Bayesian Nash equilibrium (BNE) to predict buyers' strategic behaviors. From the messaging viewpoint we will assume that we have a direct mechanism where the buyers bid their valuation: a scalar in sponsored search markets, and a vector in the general matching market. Algorithm [\[alg:alg1\]](#alg:alg1){reference-type="ref" reference="alg:alg1"} is then applied as a black-box to the inputs to produce a price on each good and a perfect matching. With expected revenue being an important concern, we provide an instance achieving higher expected revenue than VCG mechanism in an asymmetric BNE. Furthermore, a case provided to show the possibility of overbidding. Additional analyses of BNE in two simple cases with symmetric buyers and the associated bidding strategy of the buyers are provided in Appendices [\[apdx:analysis\]](#apdx:analysis){reference-type="ref" reference="apdx:analysis"} and [\[apdx:analysis2\]](#apdx:analysis2){reference-type="ref" reference="apdx:analysis2"}. ## An Instance Achieving higher expected revenue than VCG {#sec:greatRev} Given the proposed algorithm, it is important to check if there exists a scenario that our algorithm obtain a higher expected revenue than the VCG mechanism. There is a large body of literature discussing the failure of getting the VCG revenue under asymmetric distribution of buyers' valuations, e.g., see. Additionally, the well-known revenue equivalence theorem holds with an assumption that valuation of every player is drawn from a path-connected space in Chap. 9.6.3 of [^16]. With this knowledge, we demonstrate a $3\times 3$ matching market, where the buyers have asymmetric distributions of their valuations, for which our mechanism achieves higher than VCG revenue. Consider three advertisers named Alice, Bob, and Carol, and three different types of ads called listing ads, sidebar ads, and pop-ups. The realized valuation is only known to the advertiser (equivalently buyer), but the distribution of an advertisers' valuation is known to other advertisers but not the auctioneer. In other words, the auctioneer can only calculate the price according to the bids submitted by the advertisers. The minimum increment of the submitted bids is $\epsilon$, which is a positive infinitesimal, and the valuation matrix of advertisers is displayed in Table [4](#tbl:1){reference-type="ref" reference="tbl:1"}. \ Now, we give an asymmetric BNE[^17] in this matching market, with a detailed verification in the Appendix [9.11](#asymBNE){reference-type="ref" reference="asymBNE"}. First, assume that Alice always bids 0 on sidebar ads and pop-ups, and bids $\max \{1-\epsilon, \frac{w}{2}\}$ on listing ads. The best response of Bob is to bid 0 on both sidebar ads and pop-ups, and to bid $\max \{1, \frac{x-0.5}{2}\}$ on listing ads for any realized $x$. Now, given Bob's bidding function as above, one of Alice's best responses is to follow her original bidding function. Last, consider the bidding function of Alice and Bob mentioned above, a best response of Carol is to bid 0 on listing ads and pop-ups, and to bid $\epsilon$ on sidebar ads regardless of the outcomes of $y$ and $z$. [^18], even if $y>z$. This is because Carol will never win listing ads for any bids less than $1$ as the probability $\text{Pr}\{y-z \geq 1\}=0$. Now, given Carol's bidding strategy, Alice and Bob will not change their bidding functions. Therefore, the strategy $\{\beta_{\mathrm{Alice}}(w,0,0), \beta_{\mathrm{Bob}}(x,0,0.5),\beta_{\mathrm{Carol}}(y,z,2)\}=\{(\max \{1-\epsilon, \frac{w}{2}\},0,0), (\max \{1, \frac{2x-1}{4}\},0,0),(0,\epsilon,0)\}$ can be verified to be an asymmetric BNE. With the asymmetric BNE in hand, we want to calculate the expected revenue of the auctioneer and compare it with the expected revenue of the VCG mechanism. Since Carol always wins the sidebar ads and both Alice and Bob bid 0 on that, Carol will pay $\epsilon$. Additionally, in the asymmetric BNE, Alice and Bob compete on the listing ads and both bid 0 on sidebar ads and pop-ups, resulting in the payment of listing ads to be the same as the payment in the first price auction. Using this the expected revenue of the auctioneers is given by \[\begin{aligned} &&\epsilon+\frac{4}{9}\int_0^1\int_{1.5}^{2.5} 1 dxdw+\frac{2}{9}\int_2^3\int_{1.5}^{2.5} \frac{w}{2} dxdw+\frac{2}{9}\int_1^2\int_{2.5}^{3.5} \frac{2x-1}{4} dxdw \nonumber \\&+&\frac{1}{9}\int_2^3\int_{2.5}^{w+0.5} \frac{w}{2} dxdw+\frac{1}{9}\int_{2.5}^{3.5}\int_2^{x-0.5} \frac{2x-1}{4} dwdx =\dfrac{31}{27}+\epsilon \end{aligned}\] The last step is to calculate the expected revenue of the VCG mechanism, which is given by \[\begin{aligned} &&\frac{2}{3}\int_0^1\int_{1.5}^{2.5} w dxdw+\frac{2}{9}\int_2^3\int_{1.5}^{2.5} (x-0.5) dxdw+\frac{1}{9}\int_{2.5}^{3.5}\int_2^{x-0.5} w dwdx \nonumber \\&+&\frac{1}{9}\int_2^3\int_{2.5}^{w+0.5} (x-0.5) dxdw =\dfrac{1}{3}+\dfrac{2}{9}\times \dfrac{3}{2}+\dfrac{1}{9}\times \dfrac{7}{6}++\dfrac{1}{9}\times \dfrac{7}{6}=\dfrac{25}{27} \end{aligned}\] Even if we set $\epsilon$ to 0, it is obvious that the expected revenue derived under our descending price auction algorithm is strictly greater than the expected revenue of the VCG mechanism. This shows that in some instances the proposed descending price algorithm is preferred to the (DGS) ascending price algorithm, even taking the strategic behavior of buyers into account. ## Further Comments on Strategic Bidding Unlike most of the non-incentive compatible auction mechanisms where overbidding above buyers' actual value is always irrational, in some scenarios a buyer may benefit from bidding higher than her actual value to enjoy a lower price in the proposed algorithm. We show this by a simple example below. First, we present two valuations matrices that will yield the same matching but different prices with the difference in the matrices only being the valuations of one buyer. These are \[\begin{aligned} V_1= \begin{bmatrix} 4 & 5 & 5 & 6 \\ 2 & 4 & 5 & 5 \\ 1 & 2 & 4 & 5 \\ 0 & 1 & 2 & 4 \end{bmatrix}, \quad V_2 = \begin{bmatrix} 4 & 5.9 & 5.9 & 6.9 \\ 2 & 4 & 5 & 5 \\ 1 & 2 & 4 & 5 \\ 0 & 1 & 2 & 4 \end{bmatrix}. \end{aligned}\] The maximum MCPs are $P_1=[1, 2, 3, 4]$ and $P_2=[0.1,2,3,4]$, respectively, and in both cases buyer $i$ gets matched to good $i$ for all $i\in \{1,\dotsc,4\}$. Now if we assume that the true valuations of all the buyers is given by matrix $V_1$, then by unilaterally deviating to a higher bid on three goods, buyer $1$ is able to dramatically increase her payoff without impacting any other buyer. While this is only an example, we believe that this will occur with non-zero probability and will necessitate an alternate means of determining the equilibrium bidding behavior. [^1]: Despite a similar sounding name, what we seek to implement is completely different from the generalized first price (GFP) auction as in. [^2]: Appendix [12.1](#sec6.1){reference-type="ref" reference="sec6.1"} provides an example where a different choice fails to obtain the maximum MCP. [^3]: There is a history of such variational characterizations in the stable matching literature where agents are assumed to have ordinal preferences. [^4]: Since there are only $m$ possible values of $\|S\|-\|N(S)\|$ and $m$ possible values of $\frac{1}{\|S\|}$. [^5]: Facebook Ad Auction: See https://www.facebook.com/business/help/163066663757985. [^6]: Several illustrative instances of Bayesian equilibrium of strategic buyers are discussed in the full version. One particular instance explicitly demonstrates an example where our mechanism yields greater expected revenue when compared to the expected revenue of the VCG mechanism. [^7]: Again, as the number of iterations depends on both $\epsilon$ and the input valuation matrix. [^8]: Even though the final price may not be market clearing, decreasing it further by $\epsilon$ and then running the DGS algorithm, it is possible to obtain a market-clearing price that is within $2\epsilon$ of the maximum MCP. [^9]: See Section [1](#sec:prelims){reference-type="ref" reference="sec:prelims"} for a definition of strongly and weakly polynomial-time complexity. [^10]: Social welfare is defined as the sum of buyers' surplus plus the sum of goods' prices, which is the same as the sum of each buyer's value on her matched good. [^11]: Two other assumptions are (1) both mechanisms implement the same social welfare function; (2) for every player $i$, there exists a type that the expected payments are the same in both mechanisms [^12]: There need not be a unique equilibrium. [^13]: The $\epsilon$ is designed to avoid complex tie-breaking rules. [^14]: In this two advertisers case, for advertiser 1, $f_{w_{-1}}(\cdot)=f(w_2)(\cdot)$ and $\beta_{-1}(\cdot)=\beta_{2}(\cdot)$. [^15]: Every good is preferred by the buyer with the highest bid. [^16]: Two other assumptions are (1) both mechanisms implement the same social welfare function; (2) for every player $i$, there exists a type that the expected payments are the same in both mechanisms [^17]: Multiple equilibria may exist. [^18]: The $\epsilon$ is designed to avoid complex tie-breaking rules.	{'timestamp': '2017-11-07T02:05:04', 'yymm': '1607', 'arxiv_id': '1607.04710', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04710'}
null	null
null	null
# Introduction For future fusion devices such as ITER, operating at least in a partially detached state is important for reducing the heat flux incident on the divertor to below engineering limits (10 $\text{MW/m}^2$). Modelling for ITER demonstrates reduction of the peak heat flux near the separatrix by factors of up to 100 due to a number of atomic physics processes including line radiation, charge exchange and recombination. To address the need for further power removal before exhaust heat reaches the targets, which is needed for a DEMO fusion reactor and beyond, an enhanced understanding of the detachment process would be beneficial, which will enable better models for predicting ITER and DEMO performance and potentially provides insight in enhancing both detachment power and particle loss as well as control of detachment. There has been considerable work utilizing spectroscopic measurements for understanding detachment, where the characteristics of the recombining region are extracted from the Balmer series emission. Typically a high density recombination front forms at the target and moves rapidly towards the x-point as the core plasma density is increased. The aim of this study is to develop a detailed understanding of the detachment process at TCV (medium-sized tokamak ($R=0.89 \text{ m}$, $a = 0.25 \text{ m}$, $B_t = 1.4 \text{ T}$)) where low densities should give us insight into how the role of recombination changes as a function of plasma density in the divertor. In addition, this provides an understanding of how detachment on TCV relates to the general experience of detachment. This is required to interpret recent experiments on TCV, which have been performed to investigate how magnetic divertor geometry influences detachment. For this investigation a new spectroscopic diagnostic has been developed for the TCV divertor and improvements for extracting information on recombination and electron temperature from Balmer series spectra have been made. Using spectroscopic measurements we show that the observed high density recombination front at TCV during a density ramp likely stays near the target even after the target ion current drops. # Experimental setup {#ExpSetup} ## TCV's Divertor Spectroscopy System (DSS) {#DSS} The primary measurements of the recombination characteristics are made using a new spectrometer with views of the divertor, which we refer to as the DSS. The viewing optics provide a poloidal, line-integrated, view of the divertor, yielding 32 lines of sight (figure [\[fig:SpectraFigure\]](#fig:SpectraFigure){reference-type="ref" reference="fig:SpectraFigure"}B). The fibres of each system are coupled to a Princeton Instruments Isoplane SCT 320 spectrometer coupled to an Andor iXon Ultra 888 EMCCD camera with a 1024 x 1024 pixel sensor. A 1800 l/mm grating was used to allow $n_e$ measurements through Stark broadening of the n=7 Balmer series line with a measured FWHM resolution of 0.06 nm. The system has been absolutely calibrated in intensity ($\sim$`<!-- -->`{=html}15% inaccuracy) and wavelength ($<0.1$ nm), taking stray light contributions into account. A dark frame is acquired before and after the plasma discharge, which is subtracted from the measurements. Due to the long frame transfer time (1.2 ms) with respect to the acquisition time (5-10 ms), the measurements are susceptible to read-out smear of the CCD. At least 90% of the smearing is removed by post-processing using a numerical matrix-based algorithm. For the results analysed in this work, the measured spectra have been re-sampled by averaging frames and/or multiple chordal signals over the entire discharge, improving S/N ratio by up to a factor 40 which leads to improved determination of $n_e$ from line fitting (section [2.3](#NeStark){reference-type="ref" reference="NeStark"}). Figure [\[fig:SpectraFigure\]](#fig:SpectraFigure){reference-type="ref" reference="fig:SpectraFigure"}A shows that the observed intensity of medium-n Balmer lines ($n=6,7$) increases strongly during the density ramp. The observed spectra corresponds to the view line close to the target highlighted in figure [\[fig:SpectraFigure\]](#fig:SpectraFigure){reference-type="ref" reference="fig:SpectraFigure"}B (red), where the locations of the primary diagnostics used in this work are shown. ## Extracting information on recombination from Balmer lines using a collisional-radiative model {#CollRadMod} The brightness ($B_{n\rightarrow2}$ in $[\text{photons } \text{m}^{-2} \text{s}^{-1}]$) of a hydrogen Balmer line with quantum number $n$ can be modelled using the Photon Emissivity Coefficients ($PEC_{n\rightarrow2}^{rec,exc}$) \[$\text{photons } \text{m }^{3} \text{s }^{-1}$\] obtained from the ADAS collisional-radiative model for recombination and excitation, as indicated in equation [\[eq:EmissEQ\]](#eq:EmissEQ){reference-type="ref" reference="eq:EmissEQ"}. $B_{n\rightarrow2}$ consists of recombination and excitation parts: $B_{n\rightarrow2}^{rec,exc}$. It is assumed that all line emission comes from a plasma slab with spatially constant electron density $n_e$, electron temperature $T_e$, neutral density $n_o$ and width $\Delta L$. Additional assumptions are that hydrogen collisional radiative model results are valid for deuterium and that the contribution of charge exchange and molecular reactions (molecular reactions might be significant for detachment in low density plasmas ) to the emission of a certain Balmer line are negligible. For simplicity we have assumed all electrons come from hydrogen ($Z_{eff}=1$), which is discussed in section [2.5](#impeffect){reference-type="ref" reference="impeffect"}. For further discussion we define $F_{rec}$ as the fraction of total Balmer line radiation due to recombination ($F_{rec} (n) = B_{n \rightarrow 2}^{rec} / B_{n \rightarrow 2}$). We also define $F_{76}$ as the ratio of brightness of the $7\rightarrow2$ and $6\rightarrow2$ Balmer lines ($F_{76} = B_{7\rightarrow2}/B_{6\rightarrow2}$). We define $R_L$ $[\text{rec } / \text{ s} \text{ m}^2]$ as the volumetric recombination rate ($R$ $[\text{rec } / \text{ s} \text{ m}^3]$) line integrated along the line of sight through the plasma for a length $\Delta L$. Although the analysis in this section is mainly focused on the $n = 6,7$ Balmer lines, the analysis strategy is general and can be applied to other Balmer lines. \[B_{n \rightarrow 2} = \underbrace{\Delta L n_e^2 PEC_{n->2}^{rec} (n_e, T_e)}_{B_{n \rightarrow 2}^{rec}} + \underbrace{\Delta L n_o n_e PEC_{n->2}^{exc} (n_e, T_e)}_{B_{n \rightarrow 2}^{exc}} \label{eq:EmissEQ}\] ### Using Balmer line ratios to obtain the fraction of Balmer line emission due to recombination {#Frec} We have developed a method for inferring the recombination contribution to the Balmer line emission, which is important for determining several characteristics of the local plasma. For a fixed $n_e$ and $n_o$, both $F_{rec} (n)$ and $F_{76}$ only depend on $T_e$. In figure [\[fig:LineRat\]](#fig:LineRat){reference-type="ref" reference="fig:LineRat"} the relation between $F_{rec} (n)$ and $F_{76}$ is shown, where $T_e$ is varied between 0.2 and 1000 eV for each curve, while $n_o/n_e = [10^{-3}, 1]$ and $n_e = 10^{20} \text{ m}^{-3}$ are fixed. Figure [\[fig:LineRat\]](#fig:LineRat){reference-type="ref" reference="fig:LineRat"} indicates the ratio of two Balmer lines (e.g. $F_{76}$) changes as function of $F_{rec}$ and is thus useful to infer the dominance of recombination in the total emission of a particular Balmer line. The relation between $F_{rec} (n)$ and $F_{76}$ depends only weakly on $n_e$ and $n_o/n_e$. Divertor pressure measurements with an absolutely calibrated baratron gauge have been used to estimate $n_o$ and indicate $n_o/n_e$ rises from order $10^{-3}$ to order $10^{-1}$ as $T_e$ drops, which is supported by OSM-Eirene modelling and SOLPS-Eirene modelling of the TCV divertor. Based on the above $n_o/n_e$ estimates, we utilize a $n_o/n_e$ range between 0.01 and 0.25. In this $n_o/n_e$ range and the typical TCV divertor density range (between $10^{19} m^{-3}$ and $10^{20} m^{-3}$) $F_{rec}$ changes by $<0.1$. When inferring $T_e$ and $R_L$ from either $6,7\rightarrow2$ lines (using $F_{rec} (n=6,7)$) the result differs by $<3\%$. Line integration effects (section [2.4](#ProfEffect){reference-type="ref" reference="ProfEffect"}) are negligible to the determination of $F_{rec}$. Although $n_o/n_e$ only weakly influences the relation between $F_{rec} (n)$ and $F_{76}$, it strongly affects the temperature dependence of both $F_{rec} (n)$ and $F_{76}$, as shown in figure [\[fig:LineRat\]](#fig:LineRat){reference-type="ref" reference="fig:LineRat"}. Therefore, to determine $T_e$ from $F_{76}$ an accurate $n_o/n_e$ determination is needed, but is not currently possible. In section [2.2.2](#Recom){reference-type="ref" reference="Recom"}, we develop another method to derive $T_e$. Note that, even if $n_o/n_e$ would be accurately known, the $T_e$ obtained would be line averaged and weighted over both the excitation and recombination part of the Balmer line emission profile along the line of sight. This is in contrast to the $T_e$ determination described in section [2.2.2](#Recom){reference-type="ref" reference="Recom"}, where only the recombination part of the Balmer line emission is taken into account. ### Obtaining $R_L$ and $T_e^{avg}$ from absolute Balmer line intensities {#Recom} We have developed a method for calculating $R_L$, which has the advantage over previous work that no direct temperature estimate is required in the calculation. The first step is to determine the number of recombinations per photon as in for a particular Balmer line, which is (assuming the plasma is optically thin) the ratio of the ADAS effective recombination rate coefficient ($ACD (n_e, T_e)$), which takes into account both radiative and three body recombination, and the ADAS $PEC_{n \rightarrow 2}^{rec} (n_e, T_e)$. By multiplying the number of recombinations per emitted photon with $B_{n\rightarrow2}^{rec}$, we obtain $R_L (n_e, T_e, \Delta L)$ \[$\text{rec } / \text{m}^2 \text{s}$\]. Once $F_{rec}$ is determined from figure [\[fig:LineRat\]](#fig:LineRat){reference-type="ref" reference="fig:LineRat"}, we can obtain $B_{n\rightarrow2}^{rec} = F_{rec} (n) \times B_{n\rightarrow2}$, from which we can derive other important characteristics of the plasma along each chord. With fixed $n_e$ and $\Delta L$, both $R_L$ and $B_{n\rightarrow2}^{rec}$ only depend on $T_e$ and a one-to-one relationship between $R_L$ and $B_{n\rightarrow2}^{rec}$ is obtained (figure [\[fig:RecRate\]](#fig:RecRate){reference-type="ref" reference="fig:RecRate"}). In addition, as $T_e$ varies along each curve in figure [\[fig:RecRate\]](#fig:RecRate){reference-type="ref" reference="fig:RecRate"} from 0.2 to 1000 eV, $T_e$ is also obtained when determining $R_L$. We refer to this as $T_e^{avg}$ as it is line averaged and weighted by the recombination part of the Balmer line emission profile along the line of sight. Using $n_e$ (Stark broadening-section [2.3](#NeStark){reference-type="ref" reference="NeStark"}), $\Delta L$ and $B_{n\rightarrow2}^{rec}$ both $R_L$ and $T_e^{avg}$ can be determined. As shown in figure [\[fig:RecRate\]](#fig:RecRate){reference-type="ref" reference="fig:RecRate"}, determining $R_L$ through this method is only weakly affected by $n_e$ and $\Delta L$. The measurement inaccuracy of $R_L$ is generally $\sim$ 40 % when $F_{rec} \sim 1$ and is mostly due to the inaccuracy in $B_{n\rightarrow2}^{rec}$, which is affected by inaccuracies in both the absolute Balmer line intensity and the Balmer line ratio used to obtain $F_{rec}$. Line integration effects influence $R_L$ by $< 5 \%$, except for cases with a strongly hollow $n_e$ and peaked $T_e$ profile, where $R_L$ can be underestimated by up to 30 % (section [2.4](#ProfEffect){reference-type="ref" reference="ProfEffect"}). A similar approach as described here could be used to obtain excitation rates and track the excitation region, but with larger uncertainties. Obtaining $T_e^{avg}$ through the method above has the advantage that less spectral information is needed to obtain $T_e$ than for other methods. However, this method is sensitive to inaccuracies in $\Delta L$ and is strongly affected by line-integration effects. Assuming peaked $n_e, T_e$ profiles along the line of sight $T_e^{avg}$ is in between 50-100 % of the peak $T_e$ if $F_{rec} \sim 1$ (section [2.4](#ProfEffect){reference-type="ref" reference="ProfEffect"}). $T_e^{avg}$ should not be used as an absolute $T_e$ measurement, but as an indicator for trends in $T_e$ which shows the role $T_e$ plays in the increase of $R_L$ during a density ramp discharge. We define $\Delta L$ as the full-width $1/e$ fall-off length of the $n_e$ profile at the target measured by Langmuir probes, which is mapped along the flux surfaces to determine $\Delta L$ for each point where the DSS view line intersects with the separatrix at multiple time points. $\Delta L$ for TCV is generally between 2.5 and 10 cm, depending on the magnetic equilibrium used for that pulse and time. During a density ramp the density profile in the divertor broadens. Together with a constant magnetic equilibrium, $\Delta L$ can increase by up to 70 %. As a trend in the density profile at the target is not necessarily representable for trends in the density profile across the divertor leg above the target, both the Langmuir probe spatial resolution and experimental variations in $\Delta L$ during a single discharge with constant magnetic equilibrium are used to estimate the uncertainty of $\Delta L$, which makes up at most $25 \%$ of the measurement uncertainty in $R_L$. ## Obtaining $n_e$ from Stark broadening {#NeStark} The spectrally resolved line profile is affected by Stark broadening. Our chordal measurement provides a density weighted integral of contributions to the line shape and thus of the electron density $(n_e^{Stark})$. The Stark broadened line shape of a Balmer line can be expressed as a modified Lorentzian as function of $n_e$ and $T_e$, which is a parametrisation of the Microfield Model Method. The spectrometer induces additional instrumental broadening to the emitted spectral line, which is parametrized using a modified asymmetric Lorentzian whose parameters are obtained as function of wavelength and viewing chord. The experimentally observed Balmer line shape is fitted using a numerical algorithm based on the Gradient Expansion Algorithm. The fitting function used is the convolution of Stark broadening, Doppler broadening (depends on $T_i$) and the instrumental line shape. Magnetic effects are neglected. To lower the amount of fitting parameters it is assumed $T_e = T_i = 3 \text{ eV}$. For $T_e$ between 0.6 and 15 eV, the variations in $n_e^{Stark}$ are $<7\%$. For $T_i$ between 0.2 and 15 eV the variations in $n_e^{Stark}$ are $<10\%$. Assuming peaked $n_e$ profiles, $n_e^{\text{Stark}}$ is in between 65-100 % of the peak $n_e$ (section [2.4](#ProfEffect){reference-type="ref" reference="ProfEffect"}). The main parameter leading to measurement uncertainty in $n_e^{Stark}$ is the signal/noise level. By fitting synthetic spectra with a level of random noise, we have determined the measurement uncertainty of $n_e^{Stark}$ as function of $n_e$, S/N level and viewing chord. We utilize the $7\rightarrow2$ line for determining $n_e^{Stark}$ since, for the same $n_e$, higher-n Balmer lines lead to wider line shapes, which are more accurately analysed. ## Investigating line-integration effects on $n_e^{Stark}$, $T_e$ and recombination measurements {#ProfEffect} The sensitivity of the $n_e^{Stark}$, $T_e^{avg}$, $F_{rec}$ and $R_L$ inferences to line-integration effects have been discussed in sections [2.2.1](#Frec){reference-type="ref" reference="Frec"}, [2.2.2](#Recom){reference-type="ref" reference="Recom"} and [2.3](#NeStark){reference-type="ref" reference="NeStark"}. These sensitivities have been determined using the methods described in this section. Line integration effects have been studied by assuming various a priori peaked and hollow $n_e$, $T_e$ profiles along the integration chord. For peaked profiles Gaussian profile shapes have been assumed with widths varying from 0.5 to 7 cm using peak densities: $n_{e,0} = [3, 5, 10]. 10^{19} \text{ m}^{-3}$ and corresponding peak temperatures: $T_{e,0} = [15, 3, 1] \text{ eV}$. A flat neutral density profile using $n_o = [10^{18}, 10^{19}] \text{ m}^{-3}$ has been assumed. Using these profiles, the Balmer line emission is modelled at every point of the profile and the corresponding Stark line shape is calculated. The Stark line shapes, weighted by the Balmer line emission, are summed over all points of the profile to obtain a synthetic Balmer line spectrum. $n_e^{Stark}$, $R_L$, $F_{rec}$ and $T_e$ are inferred from the synthetic spectrum using the methods described in sections [2.2](#CollRadMod){reference-type="ref" reference="CollRadMod"} and [2.3](#NeStark){reference-type="ref" reference="NeStark"}. ## The role of impurity concentration on inferred results {#impeffect} For simplicity in section [2.2](#CollRadMod){reference-type="ref" reference="CollRadMod"} it has been assumed that the hydrogen ion density equals the electron density ($n_H^{+} = n_e$). However, a portion of the electrons can originate from plasma impurities. This can be taken into account by replacing the $n_e^2$ term in $B_{n\rightarrow2}^{rec}$ (equation [\[eq:EmissEQ\]](#eq:EmissEQ){reference-type="ref" reference="eq:EmissEQ"}) with $n_e n_H^{+}$, which can be written as $f n_e^2$, where $f=n_H^{+}/n_e$. By including $f$ in equation [\[eq:EmissEQ\]](#eq:EmissEQ){reference-type="ref" reference="eq:EmissEQ"} and propagating the effect of $f$ towards the inference of $R_L$, $F_{rec}$ and $T_e^{avg}$ the role of the impurity concentration has been investigated. Based on $Z_{eff}$ measurements and fractional abundance modelling through ADAS (using carbon and boron as the main plasma impurity species) we estimate that $f$ is in between 0.6 and 1.0. For this range $F_{rec}$ differs by 0.01, $R_L$ differs by 10 % and $T_e^{avg}$ differs by 20 %. Therefore, the impurity concentration is expected to have an effect on the inferred results which is small compared to the estimated uncertainty margins. # Experimental results {#ExpResults} In this section we will use the DSS data and analysis techniques described in section [2](#ExpSetup){reference-type="ref" reference="ExpSetup"} to illustrate how divertor conditions vary as detachment proceeds in TCV. Connections will be made to other diagnostic measurements to form a more complete picture of the detachment process. Observations of the Balmer line intensity ($B_{n\rightarrow2}$) and the inferred $F_{rec}$ from $F_{76}$ presented in this section correspond to the $n=7$ Balmer line. Our observation is that some of the characteristics of detachment on TCV are similar to that found at other, higher density, tokamaks. However, detachment in TCV does not lead to a large movement of the recombination region. ## Onset, evolution and dynamics of detachment {#DetachDyn} A reference plasma discharge is utilized for illustrating the process of detachment in TCV (\#52065). It has a single null magnetic divertor geometry with a plasma current of 340 kA and a reversed toroidal field direction ($\nabla B$ away from the x-point). The spectroscopic data has been acquired at 200 Hz and has been averaged over a number of time frames to improve S/N level, as indicated in the legends in figure [\[fig:Overview50648\]](#fig:Overview50648){reference-type="ref" reference="fig:Overview50648"}. The line colour and line style shown in figure [\[fig:Overview50648\]](#fig:Overview50648){reference-type="ref" reference="fig:Overview50648"}A-H correspond to the diagnostic locations shown in figure [\[fig:SpectraFigure\]](#fig:SpectraFigure){reference-type="ref" reference="fig:SpectraFigure"}B. Similar detachment characteristics as observed for \#52065 have been found for $\sim 20$ other density ramp discharges, with slight variation in timing of changes (e.g. drop in target density as determined by Langmuir probes) and magnitude (e.g. the total recombination). The vertical error bars shown in figures [\[fig:Overview50648\]](#fig:Overview50648){reference-type="ref" reference="fig:Overview50648"}A-H represent 95 % confidence intervals. Measurement uncertainties have been determined by propagating measurement uncertainties in the absolute calibration; in fit parameters (determined through the Gradient Expansion Method ); in $\Delta L$ and assuming $n_o/n_e$ is in between 0.01 and 0.25. Bolometry and spectral features consistently indicate an expansion of a cold plasma region from the target towards the x-point during a density ramp. During a considerable increase in Greenwald fraction from $\overline{n_e}/n_G = 0.3$ to $\overline{n_e}/n_G = 0.5$ (figure [\[fig:Overview50648\]](#fig:Overview50648){reference-type="ref" reference="fig:Overview50648"}A, $F_{76}$ increases resulting in an increase in $F_{rec}$ from $<0.35$ to $\sim 1$ (figure [\[fig:Overview50648\]](#fig:Overview50648){reference-type="ref" reference="fig:Overview50648"}C). Significant increases in $F_{rec}$ first occur near the target and later the region of enhanced $F_{rec}$ expands towards the x-point. The radiation front as measured by bolometry (figure [\[fig:Overview50648\]](#fig:Overview50648){reference-type="ref" reference="fig:Overview50648"}E), which is representative of higher temperatures than those at which recombination occurs, also moves from the target towards the x-point and is correlated with the increase in $F_{rec}$. The above spectral features are consistent with a strong recombining region near the target. Those features include a strong increase in $B_{n\rightarrow2}$ (figure [\[fig:Overview50648\]](#fig:Overview50648){reference-type="ref" reference="fig:Overview50648"}B) which, combined with a rising $F_{rec} (n=7)$, implies that $R_L$ (figure [\[fig:Overview50648\]](#fig:Overview50648){reference-type="ref" reference="fig:Overview50648"}G) is strongly increasing. Similar to trends in $F_{rec}$, the onset of this non-linear increase starts first close to the target and later increases closer to the x-point. The increase in $B_{n \rightarrow 2}$ during the density ramp is both due to the $n_e$ increase (figure [\[fig:Overview50648\]](#fig:Overview50648){reference-type="ref" reference="fig:Overview50648"}D) and $T_e^{avg}$ decrease (figure [\[fig:Overview50648\]](#fig:Overview50648){reference-type="ref" reference="fig:Overview50648"}H). Our results suggest that recombination is insufficient to effectively reduce the particle flux at the time of the particle flux roll-over. Furthermore, the (Stark) density close to the target does not decrease. After $F_{rec} \rightarrow 1$, $B_{n \rightarrow 2}$, $R_L$ and $n_e^{Stark}$ keep increasing until the end of the discharge while remaining highest at the lowest DSS chord 5 cm above the target. At first glance this and bolometric measurements (figure [\[fig:Overview50648\]](#fig:Overview50648){reference-type="ref" reference="fig:Overview50648"}E) would seem to indicate that while ionization and impurity radiation have detached from the target, the high density region has not. However, Langmuir probe data (taken from the probe closest to the separatrix) suggests the density has dropped at the target (figure [\[fig:Overview50648\]](#fig:Overview50648){reference-type="ref" reference="fig:Overview50648"}) as discussed in section [3.2](#RecLP){reference-type="ref" reference="RecLP"}. It is possible that the inferred $R_L$ is an underestimate, since the closest target DSS view line intersects the separatrix 5 cm above the strike point. If the $R_L$ spatial profile is extrapolated to the target, $R_L$ at the target is three times higher than at the DSS chord closest to the target. However, target probe measurements (figure [\[fig:Overview50648\]](#fig:Overview50648){reference-type="ref" reference="fig:Overview50648"}D) indicate $n_e$ drops in this non-observed region, which would lower $R_L$. Combining LP data and spectroscopic data (section [3.2](#RecLP){reference-type="ref" reference="RecLP"}) suggests that either the electron temperature at the target is very low ($< 0.06$ eV) or the high-density recombination front has moved off-target and is located in the region between the target and the lowest DSS chord. Detachment in TCV has so far never reached the level where the density and recombination region peak moves to points above the lowest DSS chord. The total recombination rate in the divertor $R_V$ \[rec./s\] is determined by integrating $R_L$ toroidally and poloidally across the chords. $R_V$ increases strongly during the last phase of the discharge (figure [\[fig:Overview50648\]](#fig:Overview50648){reference-type="ref" reference="fig:Overview50648"}F), and reaches values of up to $R_V = (6 \pm 2) \cdot 10^{21}$ rec/s, which is similar to the total particle flux measured by Langmuir probes at that time, indicating that $R_V$ contributes significantly to the particle flux drop at this time. However, the particle flux measured by the Langmuir probes drops at 1.0 s. $R_V$ at that time is relatively low, which indicates that recombination losses are not the main contributor to the initial particle flux drop. ## Recombination signatures compared with Langmuir probe data {#RecLP} Combining data from the DSS and divertor target Langmuir probe data is informative about the development of detachment. The peak target electron density determined from Langmuir probe (LP) I-V characteristics is in agreement with $n_e^{Stark}$ near the target until 0.9 s (figure [\[fig:Overview50648\]](#fig:Overview50648){reference-type="ref" reference="fig:Overview50648"}D), which is close to the time when $B_{n\rightarrow2}$ starts increasing strongly. Across many tokamaks it has been found that the temperature derived from Langmuir probes is overestimated for $T_e < 5 \text{ eV}$,. Assuming this is also true for TCV, we utilize $J_{sat}$ and $n_e^{Stark}$ (5 cm from the target) to calculate $T_e^{\text{mod}}$ (figure [\[fig:Overview50648\]](#fig:Overview50648){reference-type="ref" reference="fig:Overview50648"}H). $T_e^{\text{mod}}$ decreases during the density ramp in agreement with the Balmer line derived $T_e^{avg}$ up until 1.1 s when both the target particle flux and target density (LP) have started dropping. Near the end of the discharge, $T_e^{mod}$ reaches temperatures below 0.06 eV, much lower than $T_e^{avg}$ obtained from Balmer line analysis ($\sim 0.5$ eV). In addition it should be noted that the peak density in the density profile along the line of sight is likely higher (up to 35 %) than the density inferred from Stark broadening due to the weighted average along the chord, which would lead to an even lower $T_e^{mod}$. Therefore, if $T_e$ at the target would be higher than 0.06 eV, it would imply the target density would be lower than $n_e^{Stark}$. Hence, the density front would have moved between the target and the first DSS chord. # Discussion {#Disc} The onset of detachment observed spectroscopically at TCV is generally similar to the dynamics previously observed at higher density machines, but there are also significant differences. As the core density is increased in L-mode plasmas, the target density increases and the temperature decreases, which are general characteristics of a high-recycling divertor. However, the ion current to the target does not increase $\propto <n_e>^2$ as expected from the two point model (assuming $n_{e, up} \propto <n_e>$) (figure [\[fig:Overview50648\]](#fig:Overview50648){reference-type="ref" reference="fig:Overview50648"}F). This difference to other, high density machines and the two-point model may be due to the fact that the ionization mean free path in TCV ($\lambda_{ioniz} \sim 5-10 \text{ cm}$) is larger compared to the width of the divertor plasma ($d_{fan} \sim$ a few cm) near the target. Together with the open divertor geometry neutrals are not well-confined, which likely leads to less ionization and a slower rise in divertor density. That could reduce the amount of charge exchange and recombination, thus slowing down the detachment process. Once the detachment process starts with the drop in divertor target density and the rise in recombination signatures ($F_{rec}$ and $R_L$, figures [\[fig:Overview50648\]](#fig:Overview50648){reference-type="ref" reference="fig:Overview50648"}C and [\[fig:Overview50648\]](#fig:Overview50648){reference-type="ref" reference="fig:Overview50648"}G) the process of detachment proceeds slowly in TCV. Instead of a swift movement of the recombination and high-density regions observed at other higher density machines, the recombining region and peak density stays near the divertor target at TCV while recombination signatures extend towards the x-point. At the highest core and divertor densities in the TCV plasmas studied so far, the drop in target density (figure [\[fig:Overview50648\]](#fig:Overview50648){reference-type="ref" reference="fig:Overview50648"}D) concurrent with the continued increase in the DSS-inferred density is consistent with the detachment region (in the sense of both low density and low temperature) moving off the divertor target slightly, less than the 5 cm corresponding to the lowest DSS chord, as discussed in section [3.2](#RecLP){reference-type="ref" reference="RecLP"}. However, such a movement is very slow given that the Stark-derived density continues to rise throughout the remainder of the discharge. The inference of recombination rates through the DSS data analysis also provide some insight into the role of recombination in removing ions from the plasma and causing a density drop. As discussed earlier, recombination remains highest near the target throughout the discharge, with the total amount of recombination rising rapidly to levels at the end of the discharge comparable to the target ion flux. Given that the target density drops earlier in the pulse and that the total recombination rate is less than 1% of the particle flux at the point the particle flux starts dropping, the question is whether recombination is playing an important role at this time. The two possibilities are that the ion source upstream could start dropping at the same time as the target density falls. Or, that the recombination local to the flux tube of the peak ion flux is removing significant ion flux. We do not have enough spatial information at this time to comment further on the relative important of the two effects. # Summary The physics of the TCV divertor, including the detached regime, has been investigated at TCV, using a newly developed divertor spectroscopy system (DSS), together with advancements in techniques for extracting information from the Balmer spectrum. Analysis of the DSS data has been instrumental in characterizing the behaviour of detachment in TCV. We find that the detachment process develops slowly: the radiation first peaks near the divertor and then moves towards the x-point. The rise in the dominance of recombination signatures over excitation signatures follows the movement of the radiation peak, while the strongest level of density and recombination remains close to the target. Even as the plasma density above the target continues to increase the ion current to the target drops, which may imply that the detached (low pressure and density) region has moved off the target. But within the density range studied on TCV, the detachment front moves no further. The role of recombination in ion loss has been investigated. We find that there is no clear connection between the initial roll-over in the target ion current and the level of recombination. However, later in time, $R_V$ approaches the integral ion current and it may be that the recombination front moves further off the target if higher densities could be achieved. Further studies are needed.	{'timestamp': '2017-01-17T02:12:13', 'yymm': '1607', 'arxiv_id': '1607.04539', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04539'}
# Introduction: classical configuration theorems {#intro} Projective configuration theorems are among the oldest and best known mathematical results. The next figures depict the famous theorems of Pappus, Desargues, Pascal, Brianchon, and Poncelet. The literature on configuration theorems is vast; the reader interested in a panoramic view of the subject is recommended. Configuration theorems continue to be an area of active research. To a great extent, this is due to the advent of computer as a tool of experimental research in mathematics. In particular, interactive geometry software is a convenient tool for the study of geometric configurations. The illustration in this article are made using such a software, Cinderella 2. Another reason for the popularity of configuration theorems is that they play an important role in the emerging field of discrete differential geometry and the theory of completely integrable systems. The goal of this survey is to present some recent results motivated and inspired by the classical configuration theorems; these results make the old theorems fresh again. The selection of topics reflects this author's taste; no attempt was made to present a comprehensive description of the area. In the cases when proofs are discussed, they are only outlined; the reader interested in details is referred to the original papers. We assume that the reader is familiar with the basics of projective, Euclidean, spherical, and hyperbolic geometries. One of the standard references is, and is as indispensable as ever. Now let us specify what we mean by configuration theorems in this article. The point of view is dynamic, well adapted for using interactive geometry software. An initial data for a configuration theorem is a collection of labelled points $A_i$ and lines $b_j$ in the projective plane, such that, for some pairs of indices $(i,j)$, one has the incidence $A_i \in b_j$. If, in addition, a polarity is given, then one can associate the dual line to a point, and the dual point to a line. In presence of polarity, the initial data includes information that, for some pairs of indices $(k,l)$, the point $A_k$ is polar dual to the line $b_l$. One also has an ordered list of instructions consisting of two operations: draw a line through a pair of points, or intersect a pair of lines at a point. These new lines and points also receive labels. If polarity is involved, one also has the operation of taking the polar dual object, point to line, or line to point. The statement of a configuration theorem is that, among so constructed points and lines, certain incidence relations hold, that is, certain points lie on certain lines. It is assumed that the conclusion of a configuration theorem holds for almost every initial set of points and lines satisfying the initial conditions, that is, holds for a Zariski open set of such initial configurations. This is different from what is meant by a configuration of points and lines in chapter 3 of or in: the focus there is on whether a combinatorial incidence is realizable by points and lines in the projective plane. . I am grateful to R. Schwartz for numerous stimulating discussions and to P. Hooper for an explanation of his work. I was supported by NSF grant DMS-1510055. This article was written during my stay at ICERM; it is a pleasure to thank the Institute for its inspiring and friendly atmosphere. # Iterated Pappus theorem and the modular group {#itPapp} The Pappus theorem can be viewed as a construction in ${\mathbb {RP}}^2$ that inputs two ordered triples of collinear points $A_1,A_2,A_3$ and $B_1,B_2,B_3$, and outputs a new collinear triple of points $C_1,C_2,C_3$, see Figure [\[Pappusfig\]](#Pappusfig){reference-type="ref" reference="Pappusfig"}. One is tempted to iterate: say, take $A_1,A_2,A_3$ and $C_1,C_2,C_3$ as an input. Alas, this takes one back to the triple $B_1,B_2,B_3$. To remedy the situation, swap points $C_1$ and $C_3$. Then the input $A_1,A_2,A_3$ and $C_1,C_2,C_3$ yields a new collinear triple of points, and so does the input $C_1,C_2,C_3$ and $B_1,B_2,B_3$. And one can continue in the same way indefinitely, see Figure [\[Pappusiter\]](#Pappusiter){reference-type="ref" reference="Pappusiter"}. The study of these iterations was the topic of R. Schwartz's paper. Return to Figure [\[Pappusfig\]](#Pappusfig){reference-type="ref" reference="Pappusfig"}. The input of the Pappus construction is the marked box $(A_1,A_3, B_3, B_1; A_2, B_2)$, a quadrilateral $A_1 A_3 B_3 B_1$ with the top distinguished point $A_2$ and the bottom distinguished point $B_2$. The marked box is assumed to satisfy the convexity condition: the points $A_1$ and $A_3$ are separated by the points $A_2$ and $O$ on the projective line $a$, and likewise for the pairs of points $B_1, B_3$ and $B_2, O$ on the line $b$. Marked boxes that differ by the involution \[(A_1, A_3, B_3, B_1; A_2, B_2) \leftrightarrow (A_3, A_1, B_1, B_3; A_2, B_2)\] are considered to be the same. A convex set in ${\mathbb {RP}}^2$ is a set that is disjoint from some line and that is convex in the complement to this line, the affine plane. Two points in ${\mathbb {RP}}^2$ can be connected by a segment in two ways. The four points $A_1, A_3, B_3, B_1$, in this cyclic order, define 16 closed polygonal lines, but only one of them is the boundary of a convex quadrilateral, called the interior of the convex marked box. Recall that the points of the dual projective plane are the lines of the initial plane. Let $\Theta=(A_1,A_3, B_3, B_1; A_2, B_2)$ be a convex marked box in ${\mathbb {RP}}^2$. Its dual, $\Theta^$, is a marked box in the dual projective plane whose points are the lines \[(A_2B_1,A_2B_3,A_1B_2,A_3B_2; a,b).\] The dual marked box is also convex. The moduli space of projective equivalence classes of marked boxes in 2-dimensional. One can send the points $A_1, A_3, B_3, B_1$ to the vertices of a unit square; then the projective equivalence class of a convex marked box is determined by the positions of the points $A_2$ and $B_2$ on the horizontal sides of the square. Namely, let $x=\|A_1 A_2\|, y=\|B_1 B_2\|$. Then the equivalence class \[\label{invol} (x,y) \sim (1-x,1-y),\] where $0<x,y<1$, determines the the projective equivalence class of a convex marked box. We denote this equivalence class by $[x,y]$. The Pappus construction defines two operations on convex marked boxes, see Figure [\[complement\]](#complement){reference-type="ref" reference="complement"}: \[\begin{split} \tau_1: (A_1,A_3, B_3, B_1; A_2, B_2) \mapsto (A_1,A_3, C_3, C_1; A_2, C_2),\\ \tau_2: (A_1,A_3, B_3, B_1; A_2, B_2) \mapsto (C_1,C_3, B_3, B_1; C_2, B_2). \end{split}\] Add to it a third operation \[i: (A_1,A_3, B_3, B_1; A_2, B_2) \mapsto (B_1,B_3,A_1,A_3; B_2,A_2),\] also shown in Figure [\[complement\]](#complement){reference-type="ref" reference="complement"}. The three operations form a semigroup $G$. The operations satisfy the following identities, proved by inspection. As a consequence, $G$ is a group; for example, $\tau_1^{-1}= i \tau_2 i$. Recall that the modular group $M$ is the group of fractional-linear transformations with integral coefficients and determinant one, that is, the group $PSL(2,{\mathbb Z})$. Since $PGL(2,{\mathbb R})$ is the group of orientation preserving isometries of the hyperbolic plane, the modular group $M$ is realized as a group of isometries of $H^2$. Consider the tiling of $H^2$ by ideal triangles obtained from one such triangle by consecutive reflections in the sides, see Figure [\[Farey\]](#Farey){reference-type="ref" reference="Farey"} for the beginning of this construction. The modular group is generated by two symmetries of the tiling: the order three rotation about point $A$ and the order two rotation (central symmetry) about point $B$. Algebraically, $M$ is a free product of ${\mathbb Z}_3$ and ${\mathbb Z}_2$. Return to the group $G$. It is generated by the elements $\alpha=i\tau_1$ and $\beta = i$. Lemma [\[groupG\]](#groupG){reference-type="ref" reference="groupG"} implies that $\alpha^3=\beta^2=1$. One can prove that there are no other relations, and hence $G={\mathbb Z}_3 {\mathbb Z}_2$ is identified with the modular group. Given a convex marked box $\Theta$, consider its orbit $\Omega = G(\Theta)$ under the action of the group $G$. The orbit can be described by its oriented incidence graph $\Gamma$. The edges of $\Gamma$ correspond to the marked boxes of $\Omega$, oriented from top to bottom, and the vertices correspond to the tops and the bottoms of the boxes. One can embed $\Gamma$ in the hyperbolic plane as in Figure [\[Farey\]](#Farey){reference-type="ref" reference="Farey"} (the orientations of the edges are not shown). The group $G$ acts by permutations of the edges of $\Gamma$. The operation $i$ reverses the orientations of the edges. The operation $\tau_1$ rotates each edge counterclockwise one 'click' about its tail, and $\tau_2$ rotates the edges one 'click' clockwise about their heads. (This is a different action from the one generated by rotations about points $A$ and $B$ in Figure [\[Farey\]](#Farey){reference-type="ref" reference="Farey"}). Denote by $G'$ the index two subgroup of $G$ that consists of the transformations that preserve the orientations of the edges. The orbit $\Omega$ of a convex marked box $\Theta$ has a large group of projective symmetries, namely, an index two subgroup $M'$ of the modular group $M$. This is one of the main results of. Specifically, one has #### Proof. For the proof of the first statement, one can realize the box $\Theta$ in such a way that the three-fold rotational symmetry is manifestly present, see Figure [\[threefold\]](#threefold){reference-type="ref" reference="threefold"}. Namely, \[\begin{split} &\Theta = (B_3,B_1,A_1,A_3;B_2,A_2),\ i(\Theta) = (A_1,A_3,B_1,B_3;A_2,B_2),\\ &\tau_1(\Theta) = (B_1,B_3,C_1,C_3;B_2,C_2),\ \tau_2(\Theta) = (C_1,C_3,A_1,A_3;C_2,A_2). \end{split}\] In terms of the marked box coordinates $(x,y)$, described in ([\[invol\]](#invol){reference-type="ref" reference="invol"}), the three operations, $i, \tau_1$, and $\tau_2$, act in the same way: $[x,y] \mapsto [1-y,x].$ For the second statement, consider another realization depicted in Figure [\[inversion\]](#inversion){reference-type="ref" reference="inversion"}. The marked points $A_2$ and $B_2$ are at infinity, and $\|OA_1\| \|OA_3\| = \|OB_1\| \|OB_3\| = 1$. Then the polarity with respect to the unit circle centered at point $O$ acts as follows: \[A_1 \mapsto A_3B_2,\ A_3 \mapsto A_1 B_2,\ B_1 \mapsto B_3 A_2,\ B_3 \mapsto B_1 A_2,\] providing the desired projective equivalence. $\Box$If one identifies the projective plane with its dual by a polarity, then the above discussion describes a faithful representation of the modular group $M$ as the group of projective symmetries of the $G$-orbit $\Omega$ of a convex marked box. A marked box $\Theta$ determines a natural map $f$ of the set of vertices of the graph $\Gamma$ to the set of the marked points of the orbit $\Omega$. The map $f$ conjugates the actions of the group $G'$ on the graph $\Gamma$ and the group $M'$ of projective symmetries of the orbit $\Omega$. The set of vertices of $\Gamma$ is dense on the circle at infinity of the hyperbolic plane $S^1$, see Figure [\[Farey\]](#Farey){reference-type="ref" reference="Farey"}. Using the nested properties of the interiors of the boxes in $\Omega$ and estimates on their sizes (in the elliptic plane metric), Schwartz proves the following theorem. The image $\Lambda = f(S^1)$ is called the Pappus curve; see Figure [\[Pappusiter\]](#Pappusiter){reference-type="ref" reference="Pappusiter"} that provides an approximation of this curve. The above discussion shows that the Pappus curve is projectively self-similar. In the exceptional case of $x=y=1/2$, the curve $\Lambda$ is a straight line. Otherwise, it is not an algebraic curve, see. The tops and bottoms of the marked boxes form a countable collection of lines that also extends to a continuous family, a curve $L$ in the dual projective plane. Define a transverse line field along $\Lambda$ as a continuous family of lines such that each line from the family intersects the curve at exactly one point and every point of $\Lambda$ is contained in some line. This theorem, the fact that the Pappus curve is projectively self-similar, and computer experiments suggest that $\Lambda$ is a true fractal (unless it is a straight line). The thesis contains some preliminary numerical results on the box dimension of the Pappus curve and its dependence on the coordinates $[x,y]$ of the initial convex marked box. According to these experiments, the maximal possible box dimension of $\Lambda$ is about 1.25. Finding the fractal dimensions of the Pappus curves as a function of $[x,y]$ or, at least, proving that this dimension is greater than one in all non-exceptional cases $[x,y]\neq [1/2,1/2]$, is an outstanding open problem. # Steiner theorem and the twisted cubic {#itSteiner} This section is based on another recent ramification of the Pappus theorem, the work of J. F. Rigby and P. Hooper. Let us start with the dual Pappus theorem, see Figure [\[dualPap\]](#dualPap){reference-type="ref" reference="dualPap"} where the objects dual to the ones in Figure [\[Pappusfig\]](#Pappusfig){reference-type="ref" reference="Pappusfig"} are denoted by the same letters, with the upper and lower cases swapped (the Pappus theorem is equivalent to its dual). As an aside, let us mention that the dual Pappus theorem has an interpretation in the theory of webs: the 3-web, made of three pencils of lines, is flat, see, lecture 18. Now consider Pascal's theorem, Figure [\[Pascalfig\]](#Pascalfig){reference-type="ref" reference="Pascalfig"}. The six permutations of the points on the conic yield 60 Pascal lines. These lines and their intersection points, connected by further lines, form a intricate configuration of 95 points and 95 lines, the hexagrammum mysticum. There is a number of theorems describing this configuration, due to Steiner, Plücker, Kirkman, Cayley, and Salmon. See for a contemporary account of this subject. The Pappus theorem is a particular case of Pascal's theorem, and in this case, the number of lines that result from permuting the initial points (say, points $B_1,B_2,B_3$ in Figure [\[Pappusfig\]](#Pappusfig){reference-type="ref" reference="Pappusfig"}) reduces to six, shown in Figure [\[return2\]](#return2){reference-type="ref" reference="return2"}. Let us introduce notations. Consider Figure [\[Pappusfig\]](#Pappusfig){reference-type="ref" reference="Pappusfig"} and denote the triples of points: \[{\cal A} = (A_1,A_2,A_3),\ {\cal B} = (B_1,B_2,B_3).\] The Pappus theorem produces a new triple, ${\cal C} = (C_1,C_2,C_3)$. The lines containing these triples are denoted by $a,b,c$, respectively. We write: $c = \ell ({\cal A},{\cal B}).$ We use a similar notation for the dual Pappus theorem: if \[\alpha=(a_1,a_2,a_3),\ \beta=(b_1,b_2,b_3)\] are two triples of concurrent lines, then $\ell^* (\alpha,\beta)$ is the point of intersection of the triple of lines $(c_1,c_2,c_3)$, see Figure [\[dualPap\]](#dualPap){reference-type="ref" reference="dualPap"}. The permutation group $S_3$ acts on triples by the formula \[s({\cal B}) = (B_{s^{-1}(1)},B_{s^{-1}(2)},B_{s^{-1}(3)}).\] Let $\sigma\in S_3$ be a cyclic permutation, and $\tau\in S_3$ be a transposition of two elements. The following result, depicted in Figure [\[return2\]](#return2){reference-type="ref" reference="return2"}, is due to J. Steiner. Thus we obtain two triples of concurrent lines; denote them by \[\varphi = (\ell ({\cal A},{\cal B}), \ell ({\cal A},\sigma({\cal B}), \ell ({\cal A},\sigma^2({\cal B})), \psi = (\ell ({\cal A},\tau({\cal B})), \ell ({\cal A},\tau\sigma({\cal B}), \ell ({\cal A},\tau\sigma^2({\cal B})).\] Apply the dual Pappus theorem to the permutations of these triples of lines. By the dual Steiner theorem, the six points $\ell^(\varphi, s(\psi)),\ s\in S_3$, are collinear in threes. More is true. The next two theorems are due to Rigby. Let ${\cal B}'$ be another collinear triple of points such that the line $b'$ still passes through point $O = a\cap b$. Applying the above constructions to ${\cal A}, {\cal B}'$, we obtain new triples of lines $\varphi', \psi'$, and a new triple of points $\ell^(\varphi', s(\psi'))$ on line $a$ where $s$ is an even permutation. Theorems [\[Ri1\]](#Ri1){reference-type="ref" reference="Ri1"} and [\[Ri2\]](#Ri2){reference-type="ref" reference="Ri2"} are stated by Rigby without proof; to quote, > The theorems in this section have been verified in a long and tedious manner using coordinates. There seems little point in publishing the calculations; it is to be hoped that shorter and more elegant proofs will be found. Conceptual proofs are given in; the reader is referred to this paper and is encouraged to find an alternative approach to these results. The above theorems make it possible to define the Steiner map \[S_O: (A_1,A_2,A_3) \mapsto (\ell^(\varphi, \psi), \ell^(\varphi, \sigma^2(\psi)), \ell^(\varphi, \sigma(\psi))).\] This map depends on the point $O$, but not on the choice of the triple ${\cal B}$. The Steiner map commutes with permutations of the points involved, and hence it induces a map of the space of unordered triples of points of the projective line. Abusing notation, we denote this induced map by the same symbol. Hooper gives a complete description of the Steiner map. Assume that the ground field is complex. The space of unordered triples of points of ${\mathbb {CP}}^1$, that is, the symmetric cube $S^3({\mathbb {CP}}^1)$, is identified with ${\mathbb {CP}}^3$. This is a particular case of the Fundamental Theorem of Algebra, one of whose formulations is that $S^n({\mathbb {CP}}^1)={\mathbb {CP}}^n$ (given by projectivizing the Vieta formulas that relate the coefficients of a polynomial to its roots). Thus $S_O$ is a self-map of ${\mathbb {CP}}^3$. The set of cubic polynomials with a triple root corresponds to a curve $\Gamma \subset {\mathbb {CP}}^3$, the twisted cubic (the moment curve). The secant variety of the twisted cubic, that is, the union of its tangent and secant lines, covers ${\mathbb {CP}}^3$, and the lines are pairwise disjoint, except at the points of $\Gamma$. The set of cubic polynomials with a zero root corresponds to a plane in ${\mathbb {CP}}^3$. Denote this plane by $\Pi$. The Steiner map $S_O: {\mathbb {CP}}^3 \to {\mathbb {CP}}^3$ is described in the next theorem. In homogeneous coordinates of ${\mathbb {CP}}^3$, the map $S_O$ is polynomial of degree 6; see for an explicit formula for a particular choice $O=(0:1)$. In the real case, the secant lines are identified with the circle ${\mathbb R}/{\mathbb Z}$, and the Steiner map becomes the doubling map $t \mapsto 2t$ mod 1, a well known measure preserving ergodic transformation. # Pentagram-like maps on inscribed polygons {#pentalike} This section, based on, concerns eight configuration theorems of projective geometry that were discovered in the study of the pentagram map. The pentagram map, whose study was put forward by R. Schwartz , is a transformation of the moduli space of projective equivalence classes of polygons in the projective plane depicted in Figure [\[penta\]](#penta){reference-type="ref" reference="penta"}. The pentagram map has become a popular object of study: it is a discrete completely integrable system, closely related with the theory of cluster algebras. See for a sampler of the current literature on this subject. To formulate the results, let us introduce some notations. By a polygon in the projective plane we mean a cyclically ordered collection of its vertices (that also determines the cyclically ordered collection of lines, the sides of the polygon). Let ${\cal C}_n$ and ${\cal C}_n^$ be the spaces $n$-gons in the projective plane ${\mathbb {RP}}^2$ and its dual $({\mathbb {RP}}^2)^$. Define the $k$-diagonal map $T_k: {\cal C}_n \to {\cal C}_n^$: for $P=\{p_1...,p_n\}$, \[T_k(P)=\{(p_1p_{k+1}),(p_2p_{k+2}),\ldots, (p_np_{k+n})\}.\] Each map $T_k$ is an involution; the map $T_1$ is the projective duality that sends a polygon to the cyclically ordered collection of its sides. Extend the notation to muti-indices: $T_{ab}=T_a \circ T_b, T_{abc}=T_a \circ T_b\circ T_c$, etc. For example, the pentagram map is $T_{12}$. If $P$ is a polygon in ${\mathbb {RP}}^2$ and $Q$ a polygon in $({\mathbb {RP}}^2)^$, and there exists a projective transformation ${\mathbb {RP}}^2 \to ({\mathbb {RP}}^2)^$ that takes $P$ to $Q$, we write: $P \sim Q$. Now we are ready to formulate our results; they concern polygons inscribed into a conic or circumscribed about a conic. Surprisingly, this sequence doesn't continue! Theorem [\[thm12\]](#thm12){reference-type="ref" reference="thm12"} (iii) is depicted in Figure [\[octagon\]](#octagon){reference-type="ref" reference="octagon"}. See also Schwartz's applet for illustration of this and other results of this section. See Figure [\[nonagon\]](#nonagon){reference-type="ref" reference="nonagon"}. The next results have a somewhat different flavor: one does not claim anymore that the final polygon is projectively related to the initial one. Again, in spite of one's expectation, this sequence does not continue. Theorem [\[thm13\]](#thm13){reference-type="ref" reference="thm13"} (iii) is illustrated in Figure [\[dodecagon\]](#dodecagon){reference-type="ref" reference="dodecagon"}. Now about the discovery of these results and their proofs. Theorems [\[thm12\]](#thm12){reference-type="ref" reference="thm12"} (i) and (ii) were discovered in our study of the pentagram map. Then V. Zakharevich, a participant of the 2009 Penn State REU (Research Experience for Undergraduates) program, discovered Theorem [\[thm313\]](#thm313){reference-type="ref" reference="thm313"}. Inspired by this discovery, we did an extensive computer search for this kind of configuration theorems; the results are the above eight theorems. We think that the list above is exhaustive, but this remains a conjecture. Note that one may cyclically relabel the vertices of a polygon to deduce seemingly new theorems. Let us illustrate this by an example. Rephrase the statement of Theorem [\[thm13\]](#thm13){reference-type="ref" reference="thm13"} (iii) as follows: If $P$ is an inscribed dodecagon then $T_{131313}(P)$ is also inscribed. Now relabel the vertices by $\sigma(i)=5i$ mod 12. The map $T_3$ is conjugated by $\sigma$ as follows: \[i\mapsto 5i\mapsto 5i+3\mapsto 5(5i+3)=i+3\ \ {\rm mod}\ 12,\] that is, it is the map is $T_3$ again, and the map $T_1$ becomes \[i\mapsto 5i\mapsto 5i+1\mapsto 5(5i+1)=i+5\ \ {\rm mod}\ 12,\] that is, the map is $T_5$. One arrives at the statement: If $P$ is an inscribed dodecagon then $T_{535353}(P)$ is also inscribed. We proved all of the above theorems, except Theorem [\[thm13\]](#thm13){reference-type="ref" reference="thm13"} (iii), by uninspiring computer calculations (the symbolic manipulation required for a proof of Theorem [\[thm13\]](#thm13){reference-type="ref" reference="thm13"} (iii) was beyond what we could manage in Mathematica). Of course, one wishes for elegant geometric proofs. Stephen Wang found proofs of Theorems [\[thm13\]](#thm13){reference-type="ref" reference="thm13"} (i) and (ii) which are presented below, and Maria Nastasescu, a 2010 Penn State REU participant, found algebraic geometry proofs of the same two theorems. Fedor Nilov proved Theorem [\[thm13\]](#thm13){reference-type="ref" reference="thm13"} (iii) using a planar projection of hyperboloid of one sheet. Unfortunately, none of these proofs were published. Here is Wang's proof of Theorem [\[thm13\]](#thm13){reference-type="ref" reference="thm13"} (i). Consider Figure [\[Wang2\]](#Wang2){reference-type="ref" reference="Wang2"}. We need to prove that the points $B_1,\ldots, B_8$ lie on a conic. The hexagon $A_6 A_1 A_4 A_7 A_2 A_5$ is inscribed so, by Pascal's theorem, the points $B_1, B_6$ and $C$ are collinear. That is, the intersection points of the opposite sides of the hexagon $B_1 B_2 B_3 B_4 B_5 B_6$ are collinear. By the converse Pascal theorem, this hexagon is inscribed. A similar argument shows that the hexagon $B_2 B_3 B_4 B_5 B_6 B_7$ is inscribed. But the two hexagons share five vertices, hence they are inscribed in the same conic. Likewise, $B_8$ lies on this conic as well. Now, to the proof of Theorem [\[thm13\]](#thm13){reference-type="ref" reference="thm13"} (ii), see Figure [\[Wang3\]](#Wang3){reference-type="ref" reference="Wang3"}. Consider the inscribed hexagon $A_3A_6A_9A_{10}A_7A_4$. By Pascal's theorem, the points \[(A_3A_6)\cap (A_{10}A_7),\ (A_6A_9)\cap (A_7A_4),\ (A_9A_{10})\cap (A_4A_3)\] are collinear. Hence the triangles $A_3B_3A_4$ and $A_{10}A_9B_6$ are perspective. By the Desargues theorem, the points $A_4, A_9, (B_3B_6)\cap (A_3A_{10})$ are collinear. It follows that the triangles $B_9B_{10}Y$ and $B_3B_6X$ are perspective. By the Desargues theorem, the points $X,Y, (B_6B_9)\cap (B_3B_{10})$ are collinear. The same argument, with all indices shifted by five, implies that the points $X,Y, (B_1B_4)\cap (B_8B_5)$ are collinear as well. Hence the points \[(B_6B_9)\cap (B_3B_{10}), (B_1B_4)\cap (B_8B_5),\ {\rm and}\ X\] are collinear. Reinterpret this as the collinearity of \[(C_{10}B_{10})\cap (C_5B_9),\ (C_{10}B_4)\cap (C_5B_5),\ (B_3B_4)\cap (B_5B_6).\] It follows that the triangles $B_3B_4C_{10}$ and $B_5B_6C_5$ are perspective. By the Desargues theorem, the points $B_4,B_5$ and $(C_{10}C_5)\cap (C_2C_3)$ are collinear. That is, the points \[(C_{10}C_5)\cap (C_2C_3), (C_{10}C_1)\cap (C_3C_4),\ (C_1C_2)\cap (C_4C_5)\] are collinear, and by the converse Pascal theorem, the points \[C_{10},C_1,C_2,C_3,C_4,C_5\] lie on a conic. The rest is the same as in the previous proof. $\Box$One can add to Theorems [\[thm12\]](#thm12){reference-type="ref" reference="thm12"}--[\[thm13\]](#thm13){reference-type="ref" reference="thm13"} a statement about pentagons. Consider the following facts:\ (i) every pentagon is inscribed in a conic and circumscribed about a conic;\ (ii) every pentagon is projectively equivalent to its dual;\ (iii) the pentagram map sends every pentagon to a projectively-equivalent one.\ Therefore one may add the following theorem to our list: for a pentagon $P$, one has $P\sim T_2(P)$. The following result of R. Schwartz also has a similar flavor. One wonders whether there is a unifying theme here. A possibly relevant reference is. # Poncelet grid, string construction, and billiards in ellipses {#grid} A Poncelet polygon is a polygon that is inscribed into an ellipse $\Gamma$ and circumscribed about an ellipse $\gamma$. Let $L_1,\ldots, L_n$ be the lines containing the sides of a Poncelet $n$-gon, enumerated in such a way that their tangency points with $\gamma$ are in the cyclic order. The Poncelet grid is the collection of $n(n+1)/2$ points $L_i \cap L_j$, where $L_i \cap L_i$ is the tangency point of the line $L_i$ with $\gamma$. To simplify the exposition, assume that $n$ is odd (for even $n$, the formulations are slightly different). One can partition the Poncelet grid in two ways. Define the sets \[P_k = \cup_{i-j=k} \ell_i \cap \ell_j,\quad Q_k = \cup_{i+j=k} \ell_i \cap \ell_j,\] where the indices are understood mod $n$. There are $(n + 1)/2$ sets $P_k$, each containing $n$ points, and $n$ sets $Q_k$, each containing $(n + 1)/2$ points. The sets $P_k$ are called concentric, and the sets $Q_k$ are called radial, see Figure [\[Grid\]](#Grid){reference-type="ref" reference="Grid"}. The following theorem is proved in. In this section, following, we prove this projective theorem using Euclidean geometry, namely, the billiard properties of conics. As a by-product of this approach, we establish the Poncelet theorem and prove the theorem of Reye and Chasles on inscribed circles. See for general information about the Poncelet theorem, and for the theory of billiards. The reduction to billiards goes as follows. Any pair of nested ellipses $\gamma \subset \Gamma$ can be taken to a pair of confocal ellipses by a suitable projective transformation. This transformation takes a Poncelet polygon to a periodic billiard trajectory in $\Gamma$. The billiard inside a convex domain with smooth boundary is a transformation of the space of oriented lines (rays of light) that intersect the domain: an incoming billiard trajectory hits the boundary (a mirror) and optically reflects so that the angle of incidence equals the angle of reflection. The space of oriented lines has an area form, preserved by the optical reflections (independently of the shape of the mirror). Choose an origin, and introduce coordinates $(\alpha,p)$ on the space of rays: $\alpha$ is the direction of the ray, and $p$ is its signed distance to the origin. Then the invariant area form is $\omega=d\alpha\wedge d p$. A caustic of a billiard is a curve $\gamma$ with the property that if a segment of a billiard trajectory is tangent to $\gamma$, then so is each reflected segment. There is no general method of describing caustics of a given billiard curve,[^1] but the converse problem, to reconstruct a billiard table $\Gamma$ from its caustic $\gamma$, has a simple solution given by the following string construction: wrap a non-stretchable closed string around $\gamma$, pull it tight, and move the farthest point around $\gamma$; the trajectory of this point is the billiard curve $\Gamma$. This construction yields a 1-parameter family of billiard tables sharing the caustic $\gamma$: the parameter is the length of the string. The reason is as follows, see Figure [\[string\]](#string){reference-type="ref" reference="string"}. For a point $X$ outside of the oval $\gamma$, consider two functions: \[f(X)=\|XA\|+\stackrel{\smile}{\|AO\|},\ g(X)=\|XB\|+\stackrel{\smile}{\|BO\|}.\] The gradients of these functions are the unit vectors along the lines $AX$ and $BX$, respectively. It follows that these two lines make equal angles with the level curves of the functions $f+g$ and $f-g$, and that these level curves are orthogonal to each other. In particular, the level curves of $f+g$ are the billiard tables for which $\gamma$ is a caustic. Note that the function $f+g$ does not depend on the choice of the auxiliary point $O$, whereas the function $f-g$ is defined up to an additive constant, so its level curves are well defined. Here is a summary of the billiard properties of conics. The interior of an ellipse is foliated by confocal ellipses: these are the caustics of the billiard inside an ellipse. Thus one has Graves's theorem: wrapping a closed non-stretchable string around an ellipse yields a confocal ellipse. The space of rays $A$ that intersect an ellipse is topologically a cylinder, and the billiard system inside the ellipse is an area preserving transformation $T: A\to A$. The cylinder is foliated by the invariant curves of the map $T$ consisting of the rays tangent to confocal conics, see Figure [\[portrait\]](#portrait){reference-type="ref" reference="portrait"}. The curves that go around the cylinder correspond to the rays that are tangent to confocal ellipses, and the curves that form 'the eyes' to the rays that are tangent to confocal hyperbolas. A singular curve consists of the rays through the foci, and the two dots to the 2-periodic back and forth orbit along the minor axis of the ellipse. One can choose a cyclic parameter, say, $x$ modulo 1, on each invariant circle, such that the map $T$ becomes a shift $x \mapsto x+c$, where the constant $c$ depends on the invariant curve. Here is this construction (a particular case of the Arnold-Liouville theorem in the theory of integrable systems). Choose a function $H$ whose level curves are the invariant curves that foliate $A$, and consider its Hamiltonian vector field sgrad $H$ with respect to the area form $\omega$. This vector field is tangent to the invariant curves, and the desired coordinate $x$ on these curves is the one in which sgrad $H$ is a constant vector field $d/dx$. Changing $H$ scales the coordinate $x$ on each invariant curve and, normalizing the 'length' of the invariant curves to 1, fixes $x$ uniquely up to an additive constant. In other words, the 1-form $dx$ is well defined on each invariant curve. The billiard map $T$ preserves the area form and the invariant curves, therefore its restriction to each curve preserves the measure $dx$, hence, is a shift $x \mapsto x+c$. An immediate consequence is the Poncelet Porism: if a billiard trajectory in an ellipse closes up after $n$ reflections, then $nc \equiv 0$ mod 1, and hence all trajectories with the same caustic close up after $n$ reflections. Note that the invariant measure $dx$ on the invariant curves does not depend on the choice of the billiard ellipse from a confocal family: the confocal ellipses share their caustics. This implies that the billiard transformations with respect to two confocal ellipses commute: restricted to a common caustic, both are shifts in the same coordinate system. This statement can be considered as a configuration theorem; see Figure [\[commute\]](#commute){reference-type="ref" reference="commute"}. To summarize, an ellipse is a billiard caustic for the confocal family of ellipses. It carries a coordinate $x$, defined up to an additive constant, in which the billiard reflection in confocal ellipses is given by $x \mapsto x+c$. We refer to the coordinate $x$ as the canonical coordinate. Consider an ellipse $\gamma$, and let $x$ be the canonical coordinate on it. Define coordinates in the exterior of the ellipse: the coordinates of a point $X$ outside of $\gamma$ are the coordinates $x_1$ and $x_2$ of the tangency points of the tangent lines from $X$ to $\gamma$. Let us call $(x_1,x_2)$ the string coordinates of point $X$. The confocal ellipses are given by the equations $x_2-x_1=$ const. #### Proof. Consider Figure [\[circlequad\]](#circlequad){reference-type="ref" reference="circlequad"}. Let the canonical coordinates of the tangency points on the inner ellipse, from left to right, be $x_1,x_2,x_3,x_4$, so that the string coordinates are as follows: \[A(x_1,x_3),\ B(x_2,x_4),\ C(x_2,x_3),\ D(x_1,x_4).\] Since $A$ and $B$ are on a confocal ellipse, $x_4-x_2=x_3-x_1$, and hence $x_2+x_3=x_1+x_4$. By the billiard property, the arc of an ellipse $AB$ bisects the angles $CAD$ and $CBD$. Therefore, in the limit $A\to B$, the infinitesimal quadrilateral $ABCD$ becomes a kite: the diagonal $AB$ is its axis of symmetry. Hence $AB \perp CD$, and the locus of points given by the equation $x_1+x_4=$ const and containing points $C$ and $D$ is orthogonal to the ellipse through points $A$ and $B$. Therefore this locus is a confocal hyperbola. $\Box$The next result is due to Reye and Chasles. #### Proof. In the notation of the proof of the preceding lemma, $x_2+x_3=x_1+x_4$, hence points $C$ and $D$ lie on a confocal hyperbola. Furthermore, in terms of the string construction, \[f(A)+g(A)=f(B)+g(B),\ \ f(C)-g(C)=f(D)-g(D),\] hence \[f(D)-f(A)-g(A)+g(C)+f(B)-f(C)-g(D)+g(B)=0,\] or $\|AD\|-\|AC\|+\|BC\|-\|BD\|=0$. This is necessary and sufficient for the quadrilateral $ABCD$ to be circumscribed. $\Box$Now, consider a Poncelet $n$-gon, an $n$-periodic billiard trajectory in the ellipse $\Gamma$. Oner can choose the canonical coordinates of the tangency points of the sides of the polygon with the confocal ellipse $\gamma$ to be \[0,\ \frac{1}{n},\ \frac{2}{n},\ \dots,\ \frac{n-1}{n}.\] Then the string coordinates of the points of the concentric set $P_k$ are \[\left(0, \frac{k}{n}\right), \left(\frac{1}{n}, \frac{k+1}{n}\right), \left(\frac{2}{n}, \frac{k+2}{n}\right), \ldots,\] that is, their difference equals $k/n$, a constant. It follows that $P_k$ lies on a confocal ellipse. Likewise for the radial sets $Q_k$, proving the first claim of Theorem [\[Pgrid\]](#Pgrid){reference-type="ref" reference="Pgrid"}. Theorem [\[inscribed\]](#inscribed){reference-type="ref" reference="inscribed"} implies that each quadrilateral of the Poncelet grid is circumscribed, see Figure [\[circles\]](#circles){reference-type="ref" reference="circles"}. We refer to for circle patterns related to conics. Next, we prove the second claim of Theorem [\[Pgrid\]](#Pgrid){reference-type="ref" reference="Pgrid"}. The confocal family of conics is given by the equation \[\frac{x_1^2}{a_1^2+\lambda}+\frac{x_2^2}{a_2^2+\lambda}=1,\] where $\lambda$ is a parameter. Its dual family is the pencil \[(a_1^2+\lambda) x_1^2+(a_2^2+\lambda) x_2^2=1\] that consists of the conics that share four points, possibly, complex. Hence the confocal family consists of the conics that share four tangent lines, also possibly complex. To prove the last claim of Theorem [\[Pgrid\]](#Pgrid){reference-type="ref" reference="Pgrid"}, we need the following classical result. Let $\gamma$ and $\Gamma$ be confocal ellipses, centered at the origin and symmetric with respect to the coordinate axes, and let $A$ be the diagonal matrix with positive entries that takes $\gamma$ to $\Gamma$. Let us show that the linear map $A$ takes $P_k$ to $P_m$ or to its centrally symmetric set; the argument for the radial sets is similar. It is convenient to change the string coordinates $(x,y)$ to $u=(x+y)/2, v= (y-x)/2$. The $(u,v)$-coordinates of the points of the sets $P_k$ and $P_m$ are \[\left(\frac{k}{2n}+\frac{j}{n}, \frac{k}{2n}\right),\ \left(\frac{m}{2n}+\frac{j}{n}, \frac{m}{2n}\right)\ (j=0,1,\dots,n-1).\] We know that $P_k$ and $P_m$ lie on confocal ellipses $\gamma$ and $\Gamma$. According to Lemma [\[Ivory\]](#Ivory){reference-type="ref" reference="Ivory"}, the map $A$ preserves the $u$-coordinate. Therefore the coordinates of the points of the set $A (P_k)$ are \[\left(\frac{k}{2n}+\frac{j}{n}, \frac{m}{2n}\right)\ (j=0,1,\dots,n-1).\] If $m$ has the same parity as $k$, this coincides with the set $P_m$, and if the parity of $m$ is opposite to that of $k$, then this set is centrally symmetric to the set $P_m$. This completes the proof. # Identities in the Lie algebras of motions {#Lie} It is well known that the altitudes of a Euclidean triangle are concurrent. It is a lesser known fact that an analogous theorem holds in the spherical and hyperbolic geometries. In this section, we describe V. Arnold's observation that these results have interpretations as the Jacobi identity in the Lie algebras of motions of the respective geometries of constant positive or negative curvatures; see also. Following, we shall discuss the relation of othr classical configuration theorems with identities in these Lie algebras. In spherical geometry, one has the duality between points and lines that assigns the pole to an equator. There are two poles of a great circle; one can make the choice of the pole unique by considering oriented great circle, or by factorizing by the antipodal involution, that is, by replacing the sphere by the elliptic plane. This spherical duality can be expressed in terms of the cross-product: if $A$ and $B$ are two vectors in ${\mathbb R}^3$ representing points in the elliptic plane, then the vector $A \times B$ represents the point dual to the line $AB$. In the following argument, we do not distinguish between points and their dual lines. Given a spherical triangle $ABC$, the altitude dropped from $C$ to $AB$ is the great circle connecting the pole of the great circle $AB$ and point $C$. Using the identification of points and lines, and cross-product, this altitude is represented by the vector $(A\times B) \times C$, see Figure [\[sphere\]](#sphere){reference-type="ref" reference="sphere"}. Two other altitudes are given by similar cross-products, and the statement that the three great circles are concurrent is equivalent to linear dependence of the these three cross-products. But \[(A\times B) \times C + (B\times C)\times A + (C\times A)\times B =0,\] the Jacobi identity for cross-product, hence the three altitudes are concurrent. Note that the Lie algebra $({\mathbb R}^3,\times)$ is $so(3)$, the algebra of motions of the unit sphere. Thus the Jacobi identity in $so(3)$ implies the the existence of the spherical orthocenter. A similar, albeit somewhat more involved, argument works in the hyperbolic plane, with the Lie algebra of motions $sl(2,{\mathbb R})$ replacing $so(3)$. Note that these algebras are real forms of the complex Lie algebra $sl(2,{\mathbb C})$. Interestingly, the Euclidean theorem on concurrence of the three altitudes of a triangle does not seem to admit an interpretation as the Jacobi identity of the Lie algebra of motions of the plane. Developing these ideas, T. Tomihisa discovered the following identity. Note that the indices $1,3,5$ permute cyclically, while $2$ and $4$ are frozen. As above, the Tomihisa identity can be interpreted as a configuration theorem: the Lie bracket corresponds to one of the two basic operations: connecting a pair of points by a line or intersecting a pair of lines at a point. See Figure [\[Tomihisa\]](#Tomihisa){reference-type="ref" reference="Tomihisa"} for such an interpretation. # Skewers This section is based upon the recent paper. The main idea is that planar projective configuration theorems have space analogs where points and lines in the projective plane are replaced by lines in space, and the two operations, connecting two points by a line and intersecting two lines at a point, are replaced by taking the common perpendicular of two lines. The skewer of two lines in 3-dimensional space is their common perpendicular. We denote the skewer of lines $a$ and $b$ by $S(a,b)$. In Euclidean and hyperbolic spaces, a generic pair of lines has a unique skewer; in the spherical geometry, a generic pair of lines (great circles) has two skewers, similarly to a great circle on $S^2$ having two poles. We always assume that the lines involved in the formulations of the theorems are in general position. Here is the 'skewer translation' of the Pappus theorem, as depicted in Figure [\[Pappusfig\]](#Pappusfig){reference-type="ref" reference="Pappusfig"}: This theorem, as well as in the following ones, holds in ${\mathbb R}^3, S^3$ and $H^3$. And here is the skewer version of the Desargues theorem, as depicted in Figure [\[Desarguesfig\]](#Desarguesfig){reference-type="ref" reference="Desarguesfig"}: The 'rules of translation' should be clear from these examples. As a third example, consider a configuration theorem that involves polarity, namely, the theorem that the three altitudes of a triangle are concurrent that was discussed in Section [6](#Lie){reference-type="ref" reference="Lie"}. In its skewer version, one does not distinguish between polar dual objects, such as a great circle and its pole. This yields This is the Petersen-Morley, also known as Hjelmslev-Morley, theorem. An equivalent formulation: the common normals of the opposite sides of a rectangular hexagon have a common normal. See Figure [\[ten\]](#ten){reference-type="ref" reference="ten"}, borrowed from. Denote the 2-parameter family of lines that meet a given line $\ell$ at right angle by ${\cal N}_{\ell}$. The sets ${\cal N}_{\ell}$ plays the role of lines in the skewer versions of configuration theorems. Two-parameter families of lines in 3-space are called congruences. Next we describe line analogs of circles. Let $\ell$ be an oriented line in 3-space (elliptic, Euclidean, or hyperbolic). Let $G_\ell$ be the 2-dimensional subgroup of the group of orientation preserving isometries that preserve $\ell$. The orbit $G_\ell(m)$ of an oriented line $m$ is called the axial congruence with $\ell$ as axis (an analog of the center of a circle). In ${\mathbb R}^3$, the lines of an axial congruence with axis $\ell$ are at equal distances from $\ell$ and make equal angles with it. In the hyperbolic space, one can define the so-called complex distance between oriented lines, see. The complex distance between the lines of an axial congruence and its axis is constant. Axial congruences share the basic properties of circles: if two generic axial congruences share a line, then they share a unique other line; and three generic oriented lines belong to a unique axial congruence. The next result is a skewer analog of the Pascal theorem, see Figure [\[Pascalfig\]](#Pascalfig){reference-type="ref" reference="Pascalfig"}, in the particular case when the conic is a circle. As another, lesser known, example, consider the Clifford's Chain of Circles. This chain of theorems starts with a collection of concurrent circles labelled $1,2,3,\ldots, n$. The intersection point of the circles $i$ and $j$ is labelled $ij$. The circle through points $ij, jk$ and $ki$ is labelled $ijk$. The first statement of the theorem is that the circles $ijk, jkl, kli$ and $lij$ share a point; this point is labelled $ijkl$. The next statement is that the points $ijkl, jklm, klmi, lmij$ and $mijk$ are concyclic; this circle is labelled $ijklm$. And so on, with the alternating claims of being concurrent and concyclic; see, and Figure [\[Clifford\]](#Clifford){reference-type="ref" reference="Clifford"} where the initial circles are represented by lines (circles of infinite radius sharing a point at infinity). The next theorem, in the case of ${\mathbb R}^3$, is due to Richmond. Next one would like to define line analogs of conics. A first step in this direction is made in, but much more work is needed. In particular, one would like to have skewer analogs of various configuration theorems involving conics, including the Pascal theorem and the whole hexagrammum mysticum, the Poncelet Porism, and the theorems described in Section [4](#pentalike){reference-type="ref" reference="pentalike"}. As of now, this is an open problem. Now we outline two approaches to proofs of the above theorems and the skewer versions of other planar configuration theorems. The first approach is by way of the spherical geometry, and the second via the hyperbolic geometry. Either approach implies the results in all three classical geometries by 'analytic continuation'. This analytic continuation principle is well known in geometry; see, e.g., where it is discussed in detail. #### Elliptic approach. The space of oriented great circles in $S^3$, or lines in the elliptic space ${\mathbb {RP}}^3$, is he Grassmannian $G(2,4)$ of oriented 2-dimensional subspaces in ${\mathbb R}^4$. Below we collect pertinent facts concerning this Grassmannian. To every oriented line $\ell$ in ${\mathbb {RP}}^3$ there corresponds its dual oriented line $\ell^$: the respective oriented planes in ${\mathbb R}^4$ are the orthogonal complements of each other. The dual lines are equidistant and they have infinitely many skewers. The Grassmannian is a product of two spheres: $G(2,4)=S^2_-\times S^2_+$. This provides an identification of an oriented line in ${\mathbb {RP}}^3$ with a pair of points of the unit sphere $S^2$: $\ell\leftrightarrow (\ell_-,\ell_+)$. The antipodal involutions of the spheres $S^2_-$ and $S^2_+$ generate the action of the Klein group ${\mathbb Z}_2\times{\mathbb Z}_2$ on the space of oriented lines generated by reversing the orientation of a line and by taking the dual line. Two lines $\ell$ and $m$ intersect at right angle if and only if $d(\ell_-,m_-)=d(\ell_+,m_+)=\pi/2$, where $d$ denotes the spherical distance in $S^2$. A line $n$ is a skewer of lines $\ell$ and $m$ if and only if $n_-$ is a pole of the great circle $\ell_-m_-$, and $n_+$ is a pole of the great circle $\ell_+ m_+$. The set of lines that intersect $\ell$ at right angle coincides with the set of lines that intersect $\ell$ and $\ell^$. A generic pair of lines has exactly two skewers (four, if orientation is taken into account), and they are dual to each other. It follows that a configuration involving lines in elliptic space and their skewers can be identified with a pair of configurations on the spheres $S^2_-$ and $S^2_+$. Under this identification, the great circles of these spheres are not distinguished from their poles, just like in the proof described in Section [6](#Lie){reference-type="ref" reference="Lie"}. That is, the operation of taking the skewer of two lines is represented, on both spheres, by the cross-product. In this way, a configuration of lines in space becomes the direct product of the corresponding planar configurations. For example, the Petersen-Morley Theorem [\[skAlt\]](#skAlt){reference-type="ref" reference="skAlt"} splits into two statements that the altitudes of triangles, on the spheres $S^2_-$ and $S^2_+$, are concurrent. #### Hyperbolic approach. In a nutshell, a skewer configuration theorem in 3-dimensional hyperbolic space is a complexification of a configuration theorem in the hyperbolic plane. We follow the ideas of F. Morley, Coxeter, and V. Arnold. Consider the hyperbolic space in the upper halfspace model. The isometry group is $SL(2,{\mathbb C})$, and the sphere at infinity (the celestial sphere of ) is the Riemann sphere ${\mathbb {CP}}^1$. A line in $H^3$ intersects the sphere at infinity at two points, hence the space of (non-oriented) lines is the configuration space of unordered pairs of points. As we mentioned in Section [3](#itSteiner){reference-type="ref" reference="itSteiner"}, $S^2({\mathbb {CP}}^1)={\mathbb {CP}}^2$, namely, to a pair of points in the projective line one assigns the binary quadratic form having zeros at these points: \[(a_1:b_1,a_2:b_2) \longmapsto (a_1y-b_1x)(a_2y-b_2x).\] Thus a line in $H^3$ can be though of as a complex binary quadratic form, up to a factor. The space of binary quadratic forms $ax^2+2bxy+cy^2$ has the discriminant quadratic form $\Delta=ac-b^2$ and the respective bilinear form. The equation $\Delta=0$ defines the diagonal of $S^2({\mathbb {CP}}^1)$; this is a conic in ${\mathbb {CP}}^2$ that does not correspond to lines in $H^3$. The next result is contained in §52 of. If $(a_1:b_1:c_1)$ and $(a_2:b_2:c_2)$ are homogeneous coordinates in the projective plane and the dual projective plane, then ([\[ort\]](#ort){reference-type="ref" reference="ort"}) describes the incidence relation between points and lines. In particular, the set of lines in $H^3$ that meet a fixed line at right angle corresponds to a line in ${\mathbb {CP}}^2$. Suppose a configuration theorem involving polarity is given in ${\mathbb {RP}}^2$. The projective plane with a conic provides the projective model of the hyperbolic plane, so the configuration is realized in $H^2$. Consider the complexification, the respective configuration theorem in ${\mathbb {CP}}^2$ with the polarity induced by $\Delta$. According to Lemma [\[Jacobian\]](#Jacobian){reference-type="ref" reference="Jacobian"}, this yields a configuration of lines in $H^3$ such that the pairs of incident points and lines correspond to pairs of lines intersecting at right angle. We finish the section by discussing two results concerning lines in 3-space that do not follow the above described general pattern. The first of them is the skewer version of the Sylvester Problem. Given a finite set $S$ of points in the plane, assume that the line through every pair of points in $S$ contains at least one other point of $S$. J.J. Sylvester asked in 1893 whether $S$ necessarily consists of collinear points. See for the history of this problem and its generalizations. In ${\mathbb {RP}}^2$, the Sylvester Problem, along with its dual, has an affirmative answer (the Sylvester-Galai theorem), but in ${\mathbb {CP}}^2$ one has a counter-example: the 9 inflection points of a cubic curve (of which at most three can be real, according to a theorem of Klein), connected by 12 lines. The skewer version of the Sylvester Problem concerns a finite collection of pairwise skew lines in space such that the skewer of any pair intersects at least one other line at right angle. The question is whether a collection of lines with this skewer Sylvester property necessarily consists of the lines that intersect some line at right angle. #### Proof. In the elliptic case, we argue as in the above described elliptic proof. A collection of lines becomes two collections of points, in ${\mathbb {RP}}^2_-$ and in ${\mathbb {RP}}^2_+$, and the skewer Sylvester property implies that each of these sets has the property that the line through every pair of points contains another point, so one applies the Sylvester-Galai theorem on each sphere. In the hyperbolic case, we argue as in the hyperbolic proof. Let $a_1,\ldots,a_9$ be the nine inflection points of a cubic curve in ${\mathbb {CP}}^2$, and let $b_1,\ldots,b_{12}$ be the respective lines. Let $b_1^,\ldots,b_{12}^$ be the polar dual points. Then the points $a_i$ correspond to nine lines in $H^3$, and the points $b_j^*$ to their skewers. We obtain a collection of nine lines that has the skewer Sylvester property but does not possess a common skewer. In the intermediate case of ${\mathbb R}^3$, the argument is due to V. Timorin (private communication). Let us add to ${\mathbb R}^3$ the plane at infinity $H$; the points of $H$ are the directions of lines in space. One has a polarity in $H$ that assigns to a direction the set of the orthogonal directions, a line in $H$. Therefore, if three lines in ${\mathbb R}^3$ share a skewer, then their intersections with the plane $H$ are collinear. Let $L_1,\ldots, L_n$ be a collection of lines with the skewer Sylvester property. Then, by the Sylvester-Galai theorem in $H$, the points $L_1 \cap H,\ldots, L_n \cap H$ are collinear. This means that the lines $L_1,\ldots, L_n$ lie in parallel planes, say, the horizontal ones. Consider the vertical projection of these lines. We obtain a finite collection of non-parallel lines in the plane such that through the intersection point of any two there passes at least one other line. By the dual Sylvester-Galai theorem, all these lines are concurrent. Therefore the respective horizontal lines in ${\mathbb R}^3$ share a vertical skewer. $\Box$The second result is a different skewer version of the Pappus theorem. We proved this result, in the hyperbolic case, by a brute force calculation using the approach to hyperbolic geometry, developed in; see for details. It is not clear whether this theorem is a part of a general pattern. Let us close with inviting the reader to mull over the skewer versions of other constructions of planar projective geometry. For example, one can define the skewer pentagram map that acts on cyclically ordered tuples of lines in space: \[\{L_1, L_2,\ldots \} \mapsto \{S(S(L_1,L_3),S(L_2,L_4)), S(S(L_2,L_4),S(L_3,L_5)), \ldots \}\] Is this map completely integrable? [^1]: The existence of caustics for strictly convex and sufficiently smooth billiard curves is proved in the framework of the KAM theory.	{'timestamp': '2016-07-19T02:04:49', 'yymm': '1607', 'arxiv_id': '1607.04758', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04758'}
null	null
null	null
null	null
# Introduction Magneto-hydrodynamics (MHD) forms the basis of large-scale description of physics of dense plasma in compact stars. A key quantity in the dissipative formulations of MHD is the conductivity of matter. It determines, for example, the dissipation of currents and therefore the decay of magnetic fields, the dispersion of plasma waves, etc. In turn, magnetic field decay affects the rotational and thermal evolutions of neutron stars and consequently a broad array of their observational manifestations. Transport in compact star plasma was studied traditionally in the cold (essentially zero-temperature) and dense regime where the constituents form degenerate quantum liquids. This regime is relevant for mature isolated or accreting neutron stars as well as interiors of white dwarfs. The dilute and warm (non-zero temperature) regime is of interest in the context of transient, short-lived states of neutron stars, such as proto-neutron stars newly born in supernova explosions or hypermassive remnants formed in the aftermath of neutron star binary mergers. We start this article with an overview of the transport calculations of electrical conductivity of compact star matter in the density regime corresponding to their outer crusts ($\rho \le 10^{11}$ g cm$^{-3}$). Then we go on to describe our recent effort to calculate the electrical conductivity of non-zero temperature crustal plasma. We focus on sufficiently high temperatures where nuclei form a liquid coexisting with electronic background of arbitrary degeneracy. We close this review with a summary and outlook. Below we use the natural (Gaussian) units with $\hbar= c = k_B = k_e = 1$, $e=\sqrt{\alpha}$, $\alpha=1/137$ and the metric signature $(1,-1,-1,-1)$. # Overview At densities relevant to interiors of white dwarfs and neutron star crusts the electron-ion system is in a plasma state-the ions are fully ionized while free electrons are the most mobile carriers of charge. By charge conservation electron density is related to the ion charge $Z$ by $n_e=Zn_i$, where $n_i$ is the number density of nuclei. Electrons to a good accuracy form non-interacting gas which becomes degenerate below the Fermi temperature $T_F = \varepsilon_F-m = (p_F^2+m^2)^{1/2}-m$, where the electron Fermi momentum is given by $p_F = (3\pi^2n_e)^{1/3}$ and $m$ is the electron mass. The state of ions (mass number $A$ and charge $Z$) is controlled by the value of the Coulomb plasma parameter $\Gamma$ \[\begin{aligned} \label{eq:Gamma} \Gamma=\frac{e^2 Z^2}{Ta_i}\approx 22.73 \frac{Z^2}{T_6}\bigg(\frac{\rho_6}{A}\bigg)^{1/3}, \end{aligned}\] where $e$ is the elementary charge, $T$ is the temperature, $a_i=(4\pi n_i/3)^{-1/3}$ is the radius of the spherical volume per ion, $T_6$ is the temperature in units $10^6$ K and $\rho_6$ is the density in units of $10^6$ g cm$^{-3}$. If $\Gamma\ll 1$ or, equivalently $T\gg T_{\rm C}\equiv Z^2e^2/a_i$, ions form weakly coupled Boltzmann gas. In the regime $\Gamma\ge 1$ ions are strongly coupled and form a liquid for low values of $\Gamma\leq\Gamma_m\simeq 160$ and a lattice for $\Gamma>\Gamma_m$. The melting temperature of the lattice associated with $\Gamma_m$ is defined as $T_m=(Ze)^2/\Gamma_ma_i$. For temperatures below the ion plasma temperature \[\begin{aligned} T_p = \biggl(\frac{4\pi Z^2e^2n_i}{M }\biggl)^{1/2}, \end{aligned}\] where $M$ is the ion mass, the quantization of oscillations of the lattice becomes important. Figure [\[fig:PhaseDiagram\]](#fig:PhaseDiagram){reference-type="ref" reference="fig:PhaseDiagram"} shows the temperature-density phase diagram of the crustal material in the cases where it is composed of iron $\isotope[56]{Fe}$ (left panel) or carbon $\isotope[12]{C}$ (right panel). While the structure of the phase diagrams for $\isotope[56]{Fe}$ and $\isotope[12]{C}$ are similar there is an important difference as well: as the temperature is lowered the quantum effects become important for carbon prior to solidification, whereas iron solidifies close to the temperature where ionic quantum effects become important. Except of hydrogen and perhaps helium which may not solidify because of quantum zero point motions all heavier elements $Z> 2$ solidify at low enough temperature. The phase diagram in the case of density dependent composition is shown in Fig. [\[fig:PhaseDiagram2\]](#fig:PhaseDiagram2){reference-type="ref" reference="fig:PhaseDiagram2"}. The earliest studies of transport in dense matter go back to the work by Mestel, Hoyle and Lee in 1950s, who obtained the "conductive opacity", or equivalently the thermal conductivity of the electron-ion plasma in non-relativistic electron regime in the context of radiative and thermal transport in white dwarfs. Above densities of the order of $10^6$ g cm$^{-3}$ electrons are relativistic. Following the initial qualitative estimates of the conductivity of highly compressed matter by Abrikosov in 1963 more detailed calculations were carried out in the 1970s by many authors. In particular the transport in neutron star crusts in the relativistic electron regime was studied in much detail by Flowers and Itoh both in the solid and in the liquid regime using a variational method. They were able to cover a broad range of densities and low-temperature regime including multiple channels of scattering, relatively accurate description of collective modes (phonons) which contribute to the transport coefficients in the solid phase. Their discussion also extended to the neutron drip region where free neutrons contribute to the thermal conductivity and shear viscosity of matter. A critical analysis of the numerical values for the transport coefficients found by different authors was given by Yakovlev and Urpin , who also provided useful and simple approximations for the transport coefficients in the degenerate electron regime in terms of the Coulomb logarithm. Nandkumar and Pethick studied the temperature regime above the melting temperature, i.e., where ions form a liquid, showing that the screening of electron-ion interactions can lead to substantial corrections in this case. These calculations agree with those of Itoh et al. who also provide useful fitting formulae for the transport coefficient. Subsequent refinements of the results quoted above included, among other things, multi-phonon process and Debye-Waller factor in the solid phase and improved correlation functions in the liquid phase . The implementation of the transport coefficients of dense matter in the dissipative MHD equations in the case of cold neutron crust plasma in the presence of magnetic fields were discussed by a number of authors. We do not discuss here the physics in ultra-strong fields and confine our attention to non-quantizing fields, i.e., fields below the critical field $B\simeq 10^{14}$ G above which the Landau quantization of electron trajectories becomes important. # Formalism The Boltzmann equation for electron distribution function is given by \[\begin{aligned} \label{eq:Boltzmann} \frac{\partial f}{\partial t}+ {\bm V}\frac{\partial f}{\partial\bm r}-e(\bm E+{\bm V} \times \bm H)\frac{\partial f} {\partial{\bm p}}=I[f], \end{aligned}\] where ${\bm V}$ is the electron velocity, ${\bm E}$ and ${\bm H}$ are the electric and magnetic fields and $I$ is the collision intergal. We are interested in the regime where the collision integral describes electron-ion scattering and, therefore, has the form \[\begin{aligned} \label{eq:collision} I[f]=-(2\pi)^4\sum\limits_{234}\|M_{12\to 34}\|^2 \delta^{(4)}(p+p_2-p_3-p_4) [f(1-f_3)g_2-f_3(1-f)g_4], \end{aligned}\] where $p_i$ are four-momenta of particles, $g$ is the equilibrium distribution function of ions, which to a good accuracy can be described by the Maxwell-Boltzmann distribution with energy spectrum $\varepsilon=p^2/2M$, where $M$ is the ion mass. The sum Eq. [\[eq:collision\]](#eq:collision){reference-type="eqref" reference="eq:collision"} stands symbolically for the integrals over the phase-space of scattering particles, $M_{12\to 34}$ is the transition matrix element for scattering of relativistic electrons off correlated ions and is given by \[\begin{aligned} \label{eq:matrix_element} M_{12\to 34}=\frac{J_0J'_0}{q^2+\Pi_l(\omega,{\bm q})}-\frac{\bm J_t\bm J'_t}{q^2-\omega^2+\Pi_t(\omega,{\bm q})}, \end{aligned}\] where the electron and ion four-currents are given, respectively, \[\begin{aligned} \label{eq:currents} J^{\mu}=-e^\bar{u}^{s_3}(p_3)\gamma^\mu u^s(p),\quad J'^{\mu}=Ze^V'^{\mu}=Ze^(1,{\bm p}'/M), \end{aligned}\] $e^ = \sqrt{4\pi}e$, and $J_t, J'_t$ are the components of the currents transversal to the moment transfer ${\bm q}$, $\Pi_l(\omega,{\bm q})$ and $\Pi_t(\omega,{\bm q})$ are the longitudinal and transverse components of the polarization tensor, which describe respectively, the (irreducible) self-energies of longitudinal and transverse photons in the medium (plasma). The form of the matrix element [\[eq:matrix_element\]](#eq:matrix_element){reference-type="eqref" reference="eq:matrix_element"} includes thus the dynamical screening of the electron-ion interaction due to the exchange of transverse photons. Such separation has been employed in the treatment of transport in unpaired and superconducting quark matter and we adopt an analogous approach here. We linearize the Boltzmann equation [\[eq:Boltzmann\]](#eq:Boltzmann){reference-type="eqref" reference="eq:Boltzmann"} by writing \[\begin{aligned} \label{eq:perturbation} f= f^0+\delta f,\quad \delta f=-\Phi \frac{\partial f^0}{\partial\varepsilon}, \end{aligned}\] where $f^0$ is the equilibrium Fermi-Dirac distribution function, $\delta f \ll f^0,$ and $\Phi$ is the perturbation. The electric field appears in the drift term of linearized Boltzmann equation at $O(1)$ in perturbation, whereas the term involving magnetic field at order $O(\Phi)$, because $[{\bm V}\times {\bm H}]({\partial f^0}/{\partial{\bm p}})\propto [{\bm V}\times {\bm H}]{\bm V}=0.$ We next specify the form of the function $\Phi$ in the case of conduction as $\Phi={\bm p}\cdot {\bm \Xi}(\varepsilon),$ which after substitution in the linearized Boltzmann equation gives \[\begin{aligned} \label{eq:boltzmann2} e{\bm V}\cdot \left[{\bm E}+({\bm \Xi}\times{\bm H})\right]=-{\bm \Xi}\cdot {\bm p}~\tau^{-1}(\varepsilon), \end{aligned}\] where the relaxation time, which depends on electron energy $\varepsilon$, is defined by \[\begin{aligned} \label{eq:t_relax} \tau^{-1}(\varepsilon)=(2\pi)^{-5} \int d\omega d\bm q\int d\bm p_2\|{M}_{12\to 34}\|^2 \frac{\bm q\cdot \bm p}{p^2} \delta(\varepsilon-\varepsilon_3-\omega)\delta(\varepsilon_2-\varepsilon_4+\omega) g_2\frac{1-f^0_3}{1-f^0}. \end{aligned}\] (Here and below the indices 2 and 4 are reserved for ions, the index 3 corresponds to the outgoing electron). In transforming the linearized collision integral we introduced a dummy integration over energy and momentum transfers, i.e., $\omega = \varepsilon-\varepsilon_3$ and ${\bm q} = {\bm p}-{\bm p}_3$. It remains to express the vector ${\bm \Xi}$ describing the perturbation in terms of physical fields; its most general decomposition is given by \[\begin{aligned} \label{eq:Xi} {\bm \Xi}=\alpha{\bm e}+\beta{\bm h}+\gamma[{\bm e}\times{\bm h}], \end{aligned}\] where ${\bm h} \equiv {\bm H}/H$ and ${\bm e} \equiv {\bm E}/E$ and the coefficients $\alpha$, $\beta$, $\gamma$ are functions of the electron energy. Substituting Eq. [\[eq:Xi\]](#eq:Xi){reference-type="eqref" reference="eq:Xi"} in Eq. [\[eq:boltzmann2\]](#eq:boltzmann2){reference-type="eqref" reference="eq:boltzmann2"} one finds that $\alpha=-eE\tau /\varepsilon (1+\omega^2_c\tau^2)$, $\beta/\alpha=(\omega_c\tau)^2({\bm e}\cdot {\bm h})$ and $\gamma/\alpha=-\omega_c\tau$, where $\omega_c=eH\varepsilon^{-1}$ is the cyclotron frequency. As a result, the most general form of the perturbation is given by \[\begin{aligned} \label{eq:phi} \Phi=-\frac{e\tau}{1+(\omega_c\tau)^2} V_i\left[\delta_{ij}-\omega_c\tau\varepsilon_{ijk} h_k+(\omega_c\tau)^2h_ih_j\right]E_j. \end{aligned}\] Using the standard expression for the electrical current in terms of the perturbation $\Phi$ we arrive at the conductivity tensor $\sigma_{ij}=\delta_{ij}\sigma_0-\varepsilon_{ijm}h_m \sigma_1 +h_ih_j\sigma_2,$ where the components of the tensor are defined as \[\begin{aligned} \label{eq:sigma2} \sigma_n=\frac{e^2}{3\pi^2 T}\int_m^\infty\!\! d\varepsilon \frac{p^3}{\varepsilon}\frac{\tau(\omega_c\tau)^n} {1+(\omega_c\tau)^2}f^0(1-f^0),\quad n=0,1,2, \end{aligned}\] where $T$ is the temperature and the lower bound of the integral is given by the mass of the electron, which vanishes in the ultra-relativistic limit. The conductivity tensor has a particularly simple form if the magnetic field is along the $z$-direction \[\begin{aligned} \label{eq:sigma3} \hat{\sigma}= \begin{pmatrix} \sigma_0 &-\sigma_1 & 0 \\ \sigma_1 & \sigma_0 & 0 \\ 0 & 0 & \sigma \end{pmatrix}. \end{aligned}\] For zero magnetic field the current is along the electric field and we find the scalar conductivity \[\begin{aligned} \label{eq:sigma} &&\sigma=\frac{e^2}{3\pi^2 T}\int_m^\infty d\varepsilon \frac{p^3}{\varepsilon}\tau f^0(1-f^0)=\sigma_0+\sigma_2. \end{aligned}\] Thus, the components of the conductivity tensor are fully determined if the relaxation time $\tau$ is known. We evaluate the square of the scattering matrix using the standard QFT methods and then average over the positions of correlated ions, which effectively multiplies the transition probability by the structure function of ions. After some computations we find for the relaxation time \[\begin{aligned} \label{eq:relax_time2} \tau^{-1}(\varepsilon) &=&\frac{\pi Z^2e^4n_i}{\varepsilon p^3} \int_{-\infty}^{\varepsilon-m} d\omega e^{-\omega/2T} \frac{f^0(\varepsilon-\omega)}{f^0(\varepsilon)} \int_{q_-}^{q_+} dq(q^2-\omega^2+2\varepsilon\omega)S(q)F^2(q) \frac{1}{\sqrt{2\pi}\theta}\nonumber\\ &\times& e^{-\omega^2/2q^2\theta^2}e^{-q^2/8MT}\bigg\{ \frac{(2\varepsilon-\omega)^2-q^2}{\|q^2+\Pi_l\|^2}+\theta^2 \frac{(q^2-\omega^2)[(2\varepsilon-\omega)^2 +q^2]-4m^2q^2}{q^2\|q^2-\omega^2+\Pi_t\|^2}\bigg\},\nonumber\\ \end{aligned}\] where $S(q)$ is the ionic structure function, $\theta \equiv \sqrt{{T}/{M}}$, $q_{\pm} = \vert\pm p+ \sqrt{p^2-(2\omega\epsilon-\omega^2)}\vert$ and $\varepsilon = \sqrt{p^2+m^2}$ for non-interacting electrons. The contribution of longitudinal and transverse photons in [\[eq:relax_time2\]](#eq:relax_time2){reference-type="eqref" reference="eq:relax_time2"} separate. The dynamical screening effects contained in the transverse contribution are parametrically suppressed by the factor $T/M$ at low temperatures and for heavy nuclei. This contribution is clearly important in the cases where electron-electron ($e$-$e$) scattering contributes to the collision intergral. This is the case, for example, when ions form a solid lattice and, therefore, Umklapp $e$-$e$ processes are allowed, or in the case of thermal conduction and shear stresses when the $e$-$e$ collisions contribute to the dissipation. Finally, we note that in order to account for the finite size of the nuclei we have multiplied the transition probability in Eq. [\[eq:relax_time2\]](#eq:relax_time2){reference-type="eqref" reference="eq:relax_time2"} by the standard expression for the nuclear formfactor \[\begin{aligned} \label{formfactor} F(q)=-3\frac{qr_c\cos(qr_c)-\sin(qr_c)}{(qr_c)^3}, \end{aligned}\] where $r_c$ is the charge radius of the nucleus given by $r_c=1.15\, A^{1/3}$ fm. # Results For the numerical computations we need to specify the ion structure function $S(q)$. We assume that only one sort of ions exists at a given density, so that the structure functions of one-component plasma (OCP) can be used. These has been extensively computed using various numerical methods. We adopt the Monte-Carlo results of Galam and Hansen for Coulomb OCP provided in tabular form and set a two-dimension spline function in the space spanned by the magnitude of the momentum transfer $q$ and the plasma parameter $\Gamma$. In the low-$\Gamma$ regime ($\Gamma\le 2$) we used the analytical (leading order) expressions derived by Tamashiro et al. for Coulomb OCP derived using density functional methods. The resulting structure functions for various values of the plasma parameter $\Gamma$ are shown in Fig. [\[fig:Sq\]](#fig:Sq){reference-type="ref" reference="fig:Sq"} as a function of the dimensionless parameter $a_iq$, where $a_i$ is the ion-radius as defined after Eq. [\[eq:Gamma\]](#eq:Gamma){reference-type="eqref" reference="eq:Gamma"}. It is seen that the structure factor universally suppresses the contribution from small-$q$ scattering. The suppression sets in for larger $q$ at larger values of $\Gamma$. The large-$q$ asymptotics is independent of $\Gamma$ as $S(q)\to 1.$ The major difference arises for intermediate values of $q$ where the structure factor oscillates and the amplitude of oscillations increases with the value of $\Gamma$ parameter. The screening of longitudinal and transverse interactions is determined by the corresponding components of the polarization tensor. While expression [\[eq:relax_time2\]](#eq:relax_time2){reference-type="eqref" reference="eq:relax_time2"} is exact with respect to the form of the polarization tensor, in the numerical calculations we use the hard-thermal-loop approximation and next-to-leading expansion in $x = \omega/q$. For the real and imaginary parts of the polarization tensor we find \[\begin{aligned} \Pi_l (q,\omega) = q_D^2\chi_l, \qquad \Pi_t (q,\omega) = q_D^2\chi_t, \end{aligned}\] where $q_D$ is the Debye wave-length and the susceptibilities to order $O(x^2)$ are given by \[\begin{aligned} \label{eq:chi_l} &&{\rm Re}\chi_l (q,\omega) = 1-\frac{x^2}{\bar{v}^2}, \quad {\rm Im}\chi_l (q,\omega) =-\frac{\pi x}{2\bar{v}},\\ \label{eq:chi_t} &&{\rm Re}\chi_t (q,\omega) = x^2, \qquad {\rm Im}\chi_t (q,\omega) = \frac{\pi}{4}x\bar{v}, \end{aligned}\] where $\bar{v}$ is the electrons average velocity. Because the terms containing $\bar{v}$ are small as well as electrons are ultra-relativistic in the most of the regime of interest we approximate $\bar{v} = 1$ in our numerical calculations. For the longitudinal piece of the polarization tensor the screening is finite in the static case $x = 0$, while it vanishes for the transverse piece as $\Pi_t (q,\omega)\propto x^2$, hence the purely dynamical nature of the transverse screening. In the zero-temperature limit Eq. [\[eq:sigma\]](#eq:sigma){reference-type="eqref" reference="eq:sigma"} simplifies via the substitution $T\partial f^0/\partial\varepsilon=-f^0(1-f^0) \to -T\delta(\varepsilon-\varepsilon_F)$, i.e., \[\begin{aligned} \label{sigma_fermi} \sigma=\frac{e^2}{3\pi^2}\int_m^\infty d\varepsilon \frac{p^3}{\varepsilon}\tau(\varepsilon)\delta (\varepsilon-\varepsilon_F)= \frac{n_ee^2\tau_F}{\varepsilon_F}, \end{aligned}\] where $\tau_F$ is the relaxation time ([\[eq:relax_time2\]](#eq:relax_time2){reference-type="ref" reference="eq:relax_time2"}) taken on the Fermi surface in the $T=0$ limit \[\begin{aligned} \label{eq:relax_fermi} \tau^{-1}_F\equiv\tau^{-1}(\varepsilon_F) =\frac{4}{3\pi}Ze^4 \varepsilon_F\int_{0}^{2p_F}dq\frac{q^3}{\|q^2+\Pi_l\|^2}\bigg(1-\frac{q^2}{4\varepsilon^2_F}\bigg)S(q)F^2(q), \end{aligned}\] where employed the charge neutrality condition $n_e=Zn_i$. Neglecting the screening ($\Pi_l\to 0$) and the nuclear formfactor \[$F(q)\to 1$\] we obtain from [\[eq:relax_fermi\]](#eq:relax_fermi){reference-type="eqref" reference="eq:relax_fermi"} \[\begin{aligned} \label{eq:relax_Pethick} \tau^{-1}_F =\frac{4Ze^4 \varepsilon_F}{3\pi}\int_{0}^{2p_F}\frac{dq}{q}\bigg(1-\frac{q^2}{4\varepsilon^2_F}\bigg)S(q), \end{aligned}\] which coincides with Eqs. (9) and (11) of Ref. . With the input described above we have evaluated the relaxation time for electron scattering off the ions using Eq. [\[eq:relax_time2\]](#eq:relax_time2){reference-type="eqref" reference="eq:relax_time2"} and then the components of the conductivity tensor according to Eq. [\[eq:sigma2\]](#eq:sigma2){reference-type="eqref" reference="eq:sigma2"}. Here we demonstrate selected results, while our complete results are discussed elsewhere . The conductivity as a function of temperature is shown in Fig. [\[fig:sigma\]](#fig:sigma){reference-type="ref" reference="fig:sigma"} for carbon $\isotope[12]{C}$ and iron $\isotope[56]{Fe}$ nuclei. The magnitude of the magnetic field is fixed to $B_{12} = 1$, where $B_{12}$ is the magnetic field in units of $10^{12}$ G. The full results are compared to the case where the conductivity is evaluated from the Drude formula [\[sigma_fermi\]](#sigma_fermi){reference-type="eqref" reference="sigma_fermi"}, which is shown by dotted lines. The deviation from the zero temperature result are visible for temperatures in the range 0.1$-$`<!-- -->`{=html}1 MeV (MeV = $1.16\times 10^{10}$ K) when the density is varied from $10^6$ to $10^{10}$ g cm$^{-3}$. It is seen that the $\sigma$ component of conductivity has a minimum as a function of temperature: the low-temperature decrease is replace by a power-law increase with increasing temperature. This increase can be understood in terms of the smearing of the Fermi surface by temperature which makes more electrons available for conduction. The minimum of the conductivity is one of the key findings of our work. The same as in Fig. [\[fig:sigma\]](#fig:sigma){reference-type="ref" reference="fig:sigma"} but as a function of density for fixed temperature values is shown in Fig. [\[fig:sigma2\]](#fig:sigma2){reference-type="ref" reference="fig:sigma2"}. The scalar conductivity and $\sigma_0$ component are increasing functions of density and depend strongly on the temperature in the low-density limit, which is associated with lifting of the degeneracy as the temperature is increased. The behaviour of $\sigma_1$ is reversed: it is almost independent of temperature and has a maximum. This is the consequence of the different scaling of the components of the conductivity tensor with $\omega_c\tau$ parameter, which describes the effects of magnetic field. Our results in the cases of matter composed of $\isotope[12]{C}$ or matter composed of series of nuclei (when the composition varies with density) show the same general trends as for $\isotope[56]{Fe}$. The differences between these cases are quantitative and are discussed in detailed in Ref. . # Conclusions In this contribution we gave an overview of our current work on the conductivity of dense matter in the envelopes of neutron stars at non-zero temperature. One ingredient of our effort is the formulation of the transport in a manner which allows us to include the dynamical screening exactly, provided that the polarization tensor of electrons (or equivalently the self-energies of QED photons) in plasma can be computed to desired accuracy. Here we employed the results based on the hard-thermal-loop approximation and low-frequency expansion appropriate at not very high temperatures. We have shown that for electron-ion scattering the dynamical screening is suppressed parametrically by a factor $M/T$, but we anticipate that its effect would be substantial in the cases (a) of high temperatures and presence of light clusters, (b) of low temperatures, where the Umklapp processes with $e$-$e$ scattering are important, (c) of transport processes where $e$-$e$ scattering may be dominant from the outset, such as the thermal conductivity and shear viscosity. Our numerical results show that the scalar conductivity (no anisotropy due to the $B$-field) has a minimum as a function of temperature, with a power-law decrease at low-temperatures and a power-law increase at higher temperatures. The range of validity of the zero-temperature Drude formula extends from low temperatures up to 0.1$-$`<!-- -->`{=html}1 MeV ($10^9-10^{10}$ K), where the lower of these bounds corresponds to density $\rho \sim 10^{6}$ g cm$^{-3}$ and the upper one to $\rho \sim 10^{10}$ g cm$^{-3}$. The behaviour of the off-diagonal $\sigma_1$ component of the conductivity tensor is similar to the one described above except at low densities, where it remains almost constant. Finally the $\sigma_0$ component shows strongly density dependent behaviour: for large densities (high degeneracy) it behaves analogous to $\sigma$, but shows the inverse trend for low-densities which is associated with the transition from the regime $\omega_c\tau <1$ to $\omega_c\tau >1$. # Supplemental material {#supplemental-material .unnumbered} Below we present numerical tables for the conductivities log$_{10}\sigma$, log$_{10}\sigma_0$ and log$_{10}\sigma_1$ (in units of s$^{-1}$) for various values of magnetic field (in units of $10^{12}$ G) for sets of values of density \[g cm$^{-3}$\] and temperature \[MeV\]. The tables are provided for three types of composition of matter: $\isotope[12]{C}$ nuclei, $\isotope[56]{Fe}$ nuclei, and density-dependent composition as indicated in Fig. 2. Analytical fits to these results with relative error $\le 10\%$ can be found in Ref. . :::	{'timestamp': '2016-07-18T02:07:47', 'yymm': '1607', 'arxiv_id': '1607.04541', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04541'}
# Introduction A popular version of the third law of thermodynamics is that the entropy density of a physical system tends to zero in the $T \to 0$ limit. However, there is a class of theoretical models that violate this law: models in this class exhibit a ground-state degeneracy which grows exponentially with the system size, leading to a non-zero entropy density even at $T=0$. Nor can these be easily dismissed as theorists' abstractions, since one also sees ample evidence in experiment that there are systems in which the entropy plateaus at a non-zero value over a large range of temperature. In many such cases it is suspected that it eventually falls to zero at a much lower temperature scale, though recent theoretical work on skyrmion magnets suggests that this intuition may not always be reliable. Whatever the ultimate low-temperature fate of these materials, it is clear that over a broad range of temperatures they exhibit physics which is well captured by models with a non-zero residual entropy density. One important class of these are so-called ice models, in which the ground-state manifold consists of all configurations which satisfy a certain local 'ice rule' constraint. The first such model was Pauling's model for the residual configurational entropy of water ice. Here the local constraint is that two of the four hydrogens neighboring any given oxygen should be chemically bonded to it to form a water molecule. Similar models were subsequently discovered to apply to the orientations of spins along local Ising axes in certain rare-earth pyrochlores, which by analogy were dubbed 'spin ice' compounds. Such models develop power-law spin-spin correlations at low temperatures, with characteristic 'pinch points' in the momentum representation of the spin-spin correlation function, but they do not order. Their low-temperature state is often referred to as a 'co-operative paramagnet'. One interesting feature of such co-operative paramagnets is their response to an applied magnetic field. The configurations that make up the ice-rule manifold usually have different magnetizations; thus an applied field, depending on its direction, may either reduce or entirely eliminate the degeneracy. In the latter case, further interesting physics may arise when the system is heated, especially if the ice-rule constraints do not permit the thermal excitation of individual flipped spins. In such cases the lowest-free-energy excitation may be a string of flipped spins extending from one side of the system to the other. A demagnetization transition mediated by such excitations is known as a Kasteleyn transition. In spin ice research to date, insight has often been gained from the study of simplified models where the dimensionality is reduced or the geometry simplified while retaining the essential physics. In that spirit, we present in this paper a two-dimensional ice model which exhibits a Kasteleyn transition in an applied magnetic field. The model is especially interesting since, unlike its three-dimensional counterparts, it has the same Ising quantization axis for every spin. This raises the possibility that it could be extended to include a transverse magnetic field, thereby allowing the exploration of quantum Kasteleyn physics. The remainder of this paper is structured as follows. In section [2](#sec:model){reference-type="ref" reference="sec:model"}, we present our spin ice model, along with some analytical and numerical results on its thermodynamic properties in the absence of an applied magnetic field. In section [3](#sec:kasteleyn){reference-type="ref" reference="sec:kasteleyn"}, we analyse the model in the presence of a magnetic field: we show that it has a Kasteleyn transition, and we characterize it. In section [4](#sec:entropy){reference-type="ref" reference="sec:entropy"}, we use an alternative representation of the ice-rule states---the 'string representation'---to determine the model's entropy as a function of its magnetization. Finally, in section [5](#sec:summary){reference-type="ref" reference="sec:summary"}, we summarize our findings and discuss possible future lines of work. # The model {#sec:model} The model that we shall consider has the following Hamiltonian: H = \_ij J\_ij \_i \_j-h \_i \_i. [\[ham\]]{#ham label="ham"} Here $i$ and $j$ label the sites of a two-dimensional square lattice, $\sigma_i = \pm 1$ is an Ising variable on lattice site $i$, and $h$ is an externally applied (longitudinal) magnetic field. The exchange interaction $J_{ij}$ is given by: J\_ij = { . [\[exchanges\]]{#exchanges label="exchanges"} where ${\bf r}_i$ is the position vector of site $i$, ${\hat {\bf x}}$ and ${\hat {\bf y}}$ are the unit vectors of a Cartesian system in the two-dimensional plane, and $J$ is a positive constant. In this paper, we shall always work in the limit $J \gg \vert h \vert, k_B T$. Furthermore, where necessary we shall take the number of sites in the lattice to be $N$, always assuming $N$ to be large enough that edge effects can be neglected. When we refer to the density of something (e.g. the entropy density), we shall always mean that quantity divided by the number of spins---not, for example, by the number of plaquettes. The lattice described by ([\[exchanges\]](#exchanges){reference-type="ref" reference="exchanges"}) is shown in the upper-left inset of Fig. [\[defects\]](#defects){reference-type="ref" reference="defects"}, with ferromagnetic bonds represented by solid lines and antiferromagnetic bonds represented by dotted lines. One may view this lattice as made of corner-sharing plaquettes, one of which is shown in the lower-right inset of Fig. [\[defects\]](#defects){reference-type="ref" reference="defects"}. It is easy to see that the bonds on this plaquette cannnot all be satisfied at once: the model ([\[ham\]](#ham){reference-type="ref" reference="ham"}) is therefore magnetically frustrated. The sixteen spin configurations of the elementary plaquette, together with their energies, are shown in Table [3](#plaqconf){reference-type="ref" reference="plaqconf"}.	{'timestamp': '2016-07-19T02:00:32', 'yymm': '1607', 'arxiv_id': '1607.04657', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04657'}
null	null
# Introduction {#sec:intro} The precise manipulation of nano and sub-nanoscale physical systems lies at the heart of the ongoing quantum revolution, by which new communication and information technologies are expected to emerge. In this context, an amazing progress has been made in the study of non-equilibrium dynamics of many-body quantum systems, both theoretically and experimentally. A wide range of different phenomena has been closely studied in recent years, such as many-body localization, relaxation, thermalization, quantum phase transitions, among others.\ Understanding the dynamics of such complex quantum systems is the first step towards the ultimate goal: the ability to engineer its complete time-evolution using a small number of properly tailored control fields. To tackle this problem, optimal control theory (OCT) emerges as the natural tool. Routinely used in various branches of science, optimization techniques allows to derive the required shape for a control field $\epsilon(t)$ that optimizes a particular dynamical process for a quantum system described by a Hamiltonian $H(\epsilon)$. For example, a typical goal in quantum control is to connect a given initial $\Ket{\psi_0}$ and target states $\Ket{\psi_f}$ in some evolution time $T$. In recent years, optimal control has been applied with great success in systems of increasing complexity, with applications including state control of many-boson dynamics, the crossing of quantum-phase transitions, generation of many-body entangled states and optimization of quantum thermodynamic cycles. A lot of attention has also been devoted to investigate the fundamental limitations of OCT, most of all in connection with the study of the so-called quantum speed limit. In a recent work, OCT has even been used in a citizen-science scenario allowing to investigate the power of gamification techniques in solving quantum control problems.\ In this work, we investigate the connection between the complexity of a quantum system and its controllability. To this end, we study optimal control protocols on a spin-1/2 chain with short-range interactions, both in the few-and many-body regimes. By using this model, we are able to tune the physical complexity of the system in two different ways: (a) by adding excitations to the chain, we can increase the system space dimension; (b) by tuning the interparticle coupling, we can drive the system through a transition from a regular energy spectrum to a chaotic one. We perform an unconstrained optimization in order to obtain the control fields needed to drive various physical processes, and define two figures of merit based on the frequency spectrum of the fields: the spectral bandwidth, associated with the maximum frequency present in the field and the spectral inverse participation ratio (sIPR), related to the signal complexity. We find that the spectral bandwidth is strongly connected to the structure of the control Hamiltonian. In the common scenario where the control is applied locally on any site of the chain, we find that the bandwidth is independent of the state space dimension, for various processes. On the other hand, the complexity of the signal grows with the dimension, due to the increase of energy levels. Inspired by this, we asses the role of quantum chaos in the control of quantum many-body dynamics. We find that the transition between regular and chaotic energy spectrum does not affect the complexity of the control problem in a significant way.\ We point out that previous works have studied the relationship between optimal control and the integrable or non-integrable nature of the quantum system under analysis. There, a suitable measure for the control field complexity was defined, related with the number of frequencies in the field, as allowed by the optimization procedure. Then, it was shown that the complexity required to achieve control scaled exponentially with the dimension of the manifold supporting the dynamics. Here, we focus on studying the complexity of the control fields regardless of the details of the optimization method itself. We do this by deliberately allowing many frequency components in the control fields, and then analysing which of those components are required to effectively drive the system.\ This article is organized as follows. In Sec. [2](#sec:model){reference-type="ref" reference="sec:model"} we present the model of a spin-1/2 chain and discuss its symmetries. Also, we propose a few control protocols and put forward the main elements of QOC theory. In Sec. [3](#sec:result){reference-type="ref" reference="sec:result"} we define two measures of complexity for the optimal control fields, and present a systematic study on how they vary with both the system state dimension and chaos parameter. In. Sec. [4](#sec:discuss){reference-type="ref" reference="sec:discuss"} we analyse in detail the reasons why the chaotic regime does not affect the controllability of the system. Finally, Sec. [5](#sec:conclu){reference-type="ref" reference="sec:conclu"} contains some concluding remarks. # Model and methods {#sec:model} ## Spin chain model Let us consider a one-dimensional system of $L$ spin-1/2 particles that interact through nearest-neighbor (NN) and next-to-nearest-neighbor (NNN) homogenous couplings with open boundary conditions. The Hamiltonian for this models reads \[\begin{aligned} H_{01} &=& H_0 + \Gamma H_1, \label{ec:h01}\\ H_0 &=& \frac{J}{2}\sum_{i=1}^{L-1}\sigma_i^x\sigma_{i+1}^x+\sigma_i^y\sigma_{i+1}^y+\alpha_z\sigma_i^z\sigma_{i+1}^z,\\ H_1 &=& \frac{J}{2}\sum_{i=1}^{L-2}\sigma_i^x\sigma_{i+2}^x+\sigma_i^y\sigma_{i+2}^y+\alpha_z\sigma_i^z\sigma_{i+2}^z, \end{aligned}\] where $\sigma_i^{x,y,z}$ are the Pauli matrices for the i-th particle and we have taken $\hbar=1$. The Hamiltonian $H_0$, which has only NN couplings, is the usual $XXZ$ Heisenberg model, which can be exactly solved via the Bethe ansatz. The parameter $\Gamma$ measures the ratio between the NNN exchange and the NN couplings. This model has been extensively studied in the literature in many contexts, in particular when investigating quantum chaos, i.e., the study of the quantum mechanical properties of systems which classical analogs display a chaotic behavior. Albeit lacking a semiclassical counterpart, this spin model displays a transition in its level spacing distribution as $\Gamma$ changes. For $\Gamma\lesssim0.5$, the energy spectrum is regular, and its level spacings follow a Poisson distribution, while for $\Gamma\gtrsim0.5$, the distribution follows Wigner-Dyson statistics, and the spectrum is deemed \"chaotic\" (see the Appendix for more details).\ While the full Hilbert space $\mathcal{H}$ of this model has a dimension of $2^L$, we can identify two constant of motions such that $\mathcal{H}$ is decomposed into smaller subspaces. First, the total magnetization the in $z$ direction $\sigma_z=\sum_{i=1}^L \sigma_i^z$ is conserved, which defines $L+1$ subspaces with fixed $\sigma_z$. Each subspace can then be characterized by this quantum number, which can be interpreted as the number of sites with spins pointing "up" or "excitations" in the chain. The dimension of the subspace with $K$ excitations is given by \[D_K=\frac{L!}{K!(L-K)!}.\] We consider also conservation of parity. The parity operator $\Pi$ acts as permutation between mirrored sites of the chain, and commutes with the Hamiltonian $H_{01}$ for all values of $J$ and $\Gamma$. As a consequence, each of the above mentioned subspaces break up into two (positive and negative) parity subspaces, each of which of dimension $D_{K,\Pi}\simeq D_K/2$. Lastly, we avoid conservation of $\sigma_z^2$ by chosing $\alpha_z=0.5$, and choose an odd value of the chain length $L=15$. ## Control protocols We now describe an scenario to perform control operations on the spin chain. First, we define the control Hamiltonian by means of which we intend to steer the "free" chain Hamiltonian $H_{01}$. Several different proposals have been studied in the literature. For example, in Ref., the author proposed using a global parabolic magnetic potential to control the transfer of excitations from one end of the chain to the other. Later, the same configuration was used to study the optimal evolution time for such processes. Other control configurations have also been proposed in scenarios where only one or two sites are locally addressed by external fields. Here, we will consider time-dependent magnetic fields in the $z$ direction which are locally applied at each site of the chain. In order to comply with the system symmetries, we will consider the situation where the first and last spin of the chain are affected by the same field $\epsilon(t)$, whereas all the other spins do not interact with any external field. Consequently, the full Hamiltonian can be written as \[H(t) = H_{01} + \epsilon(t)H_c,\ \mathrm{where}\ H_c=\frac{J}{2}\left(\sigma_1^z+\sigma_L^z\right). \label{ec:hcontrol}\] We point out that $H(t)$ preserves the same symmetries as $H_{01}$ for any choice of $\epsilon(t)$. In addition to this, we checked that for any fixed value of the field, $H$ still shows a transition between a regular and a chaotic spectrum for $\Gamma\simeq0.5$. For more details about this issue, see the Appendix.\ The next step is to define the control processes we aim to perform. We will consider two different protocols (A and B) in order to obtain general results about the systems controllability. In both cases, we define initial and target states which we denote $\Ket{\psi_0^\alpha}$ and $\Ket{\psi_f^\alpha}$, where $\alpha=A,B$. These states are deliberately defined to allow the system evolve within a particular subspace with fixed (positive) parity and number of excitations $K$ of the complete Hilbert space, as discussed in the previous section. First, process "A" involves the system initially prepared in a state with all excitations in the middle sites of the chain (in this scheme, the central site has no excitations if K is even). We then intend to drive this configuration into a coherent superposition as defined by \[\begin{aligned} \Ket{\psi_0^A}&=&\Ket{\downarrow\ldots\downarrow\uparrow\uparrow\uparrow\downarrow\ldots\downarrow}\\ \Ket{\psi_f^A}&=&\frac{1}{\sqrt{2}}\left(\Ket{\uparrow\uparrow\uparrow\downarrow\ldots\downarrow}+\Ket{\downarrow\ldots\downarrow\uparrow\uparrow\uparrow}\right). \end{aligned}\] Process A then represents and orderered control process in which entanglement is generated between both ends of the chain. On the other hand, we define a disordered process B, where the system starts from the ground state of $H_0$ and its steered into a random superposition of excited states (with positive parity). \[\begin{aligned} \Ket{\psi_0^B}&=&\Ket{\mathrm{g.s.}_0}\\ \Ket{\psi_f^B}&=&\sum_{n=1}^{D_{K,+}-1}a_n\Ket{n_0} \end{aligned}\] where $\left\{\Ket{n_0}\right\}$, $n=0,\ldots,D_{K,+}$ are the positive eigenvectors of $H_0$ in the subspace of $K$ excitations and $\Ket{\mathrm{g.s.}_0}\equiv\Ket{0_0}$. The coefficientes $\left\{a_n\right\}$ are a set of random complex numbers so that $\sum_i \|a_n\|^2=1$.\ ## Optimal control methods In order to obtain the control fields $\epsilon(t)$ which drive the system for both processes, we use optimal control theory. Here we briefly sketch the Krotov optimization algorithm, as described in many previous works (see for example Refs. ). This procedure takes as an input a fixed evolution time $T$, an initial guess for the control field $\epsilon^{(0)}(t)$, and both the initial and final states $\Ket{\psi_0^\alpha}$ and $\Ket{\psi_f^\alpha}$, where $\alpha=A,B$. The procedure starts by evolving $\Ket{\psi_0^\alpha}$ according to the Hamiltonian $H(\epsilon^{(0)}(t))$, from $t=0$ to $t=T$. The final state $\Ket{\psi(T)}$ is then projected to $\Ket{\psi_f^\alpha}$ in order to obtain an auxiliary state $\Ket{\chi(T)}=\BraKet{\psi_f^\alpha}{\psi(T)}\Ket{\psi_f^\alpha}$. This state is finally evolved backwards with the same Hamiltonian, from $t=T$ to $t=0$. The process is then repeated, but the control field is updated following the recipe \[\epsilon^{(k+1)}(t)\rightarrow\epsilon^{(k)}(t) + \frac{1}{\lambda(t)}\mathrm{Im}\left\{\Bra{\chi(t)}\frac{\partial H}{\partial \epsilon}\Ket{\psi(t)}\right\},\] where $\lambda(t)$ is an weight function. Note that, for the model considered here, the operator $\partial H/\partial \epsilon$ is fully defined by equation ([\[ec:hcontrol\]](#ec:hcontrol){reference-type="ref" reference="ec:hcontrol"}) and equals simply to $H_c$. The iterative procedure stops when a certain target fidelity $\mathcal{F}=\|\BraKet{\psi(T)}{\psi_f}\|^2$ has been achieved.\ As we intend to compare the optimal control fields obtained by this optimization procedure, we fix the input parameters of the optimization as follows. For the total evolution time $T$ we set $T=15\times T_L$ where $T_L=(L-1)\frac{\pi}{J}$ can be regarded as the typical evolution time required for transferring a single excitation from one end of the chain to the other. We have checked that using this value we are operating well beyond the quantum speed limit, and so that fidelities up to 0.99 or greater can be achieved, for both control processes and every value of the number of excitations $K$ and the NNN coupling $\Gamma$ considered. Also, we used a constant initial guess $\epsilon^{(0)}(t)=0.1$ in all cases. We have checked that the results we present in the next section hold for other choices of this function. # Analysis of the optimal control fields {#sec:result} In Fig. [\[fig:fig1\]](#fig:fig1){reference-type="ref" reference="fig:fig1"} (a) we show a typical example of the control field $\epsilon(t)$ obtained by the optimization procedure outlined in the previous section, together with its Fourier spectrum. The time signal shown can be seen to be complex and to have many spectral components up to certain frequency threshold. In order to characterize quantitively this features, we define two measures of complexity for the control fields: the frequency bandwidth and the spectral inverse participation ratio (sIPR). In this section we investigate these quantites. ## Frequency bandwidth Given a time-varying signal $\epsilon(t)$ and its Fourier transform $\hat{\epsilon}(\omega)$, we first define its frequency bandwidth as the value $\omega_{bw}$ such that \[\int_0^{\omega_{bw}}d\omega\:\|\hat{\epsilon}(\omega)\|^2=1-\beta, \label{ec:bw}\] where $0<\beta<1$ and the frequency distribution is normalized such that $\int_0^{\infty}d\omega\:\|\hat{\epsilon}(\omega)\|^2=1$. By this definition, the frequency interval $[0,\omega_{bw}]$ concentrates the $[(1-\beta)\times100]\%$ of the power spectrum (here, we use $\beta=10^{-2}$). In other words, $\omega_{bw}$ is a measure of the maximum frequency present in $\epsilon(t)$.\ In Fig. [\[fig:fig1\]](#fig:fig1){reference-type="ref" reference="fig:fig1"} (b) we show the frequency bandwidth $\omega_{bw}$ as a function of the NNN or chaos parameter $\Gamma$, for different number $K$ of excitations in the chain. Results obtained for both processes A and B are shown in the same plot. Remarkably, we find that all data roughly coincides in the same curve. This result indicates that the bandwidth is independent not only of the control processes considered, but also of the state space dimension. Note that, in each case, $\omega_{bw}$ is approximately constant for $\Gamma<0.5$ and then increases steadly for $\Gamma>0.5$. Although this behavior correlates with the onset of chaos in the system (as discussed in Sec. [1](#sec:intro){reference-type="ref" reference="sec:intro"}), we must first consider that increasing the interparticle coupling $\Gamma$ necessarily increases the energy of the chain. As previously discussed in the context of QOC, we expect that the frequency distribution of the control fields presents peaks located at the resonances of the free Hamiltonian $H_{01}$. Following this criterion, the maximum frequency is bounded by the energy spread $\Delta E$ of $H_{01}$, defined as \[\Delta E = E_{max}-E_{0},\] where $E_{max}$ and $E_0$ are the maximum and minimum (ground state) energies of the Hamiltonian. Note that $\Delta E$ is a function of the interparticle interaction parameters $J$ and $\Gamma$ and of the number of excitations $K$. We show such functions as dashed lines in Fig. [\[fig:fig1\]](#fig:fig1){reference-type="ref" reference="fig:fig1"} (c). It is clear that the dependence of the bandwidth with $\Gamma$ closely resembles the energy spread with $K=1$. The same observation can be drawn by studying both quantities as a function the NN coupling $J$ (for fixed $\Gamma$). There, the behaviour is obviously linear, as shown in the inset of Fig. 1 (c).\ Note, however, that the striking independence of the bandwith with the state dimension cannot be explained by its relation with the energy spread of the free Hamiltonian, since $\omega_{bw}$ is in every case significantly smaller than $\Delta E$ for $K>1$. In order to gain a deeper insight about this result, we turn to investigate the role of the control Hamiltonian $H_c$, defined in eq. ([\[ec:hcontrol\]](#ec:hcontrol){reference-type="ref" reference="ec:hcontrol"}). We first study the structure of the matrix $H_c$ written in the basis of (positive) eigenvectors of the free chain Hamiltonian $H_{01}$. In the top row of Fig. [\[fig:fig2\]](#fig:fig2){reference-type="ref" reference="fig:fig2"} (a) we plot the absolute value of such matrix elements for fixed values of $\Gamma$ and $K$. From this plots we can see that $H_c$ does not connect eigenstates which distant energies: for example, the ground state is not connected with excited states beyond the middle of the spectrum. This explains the absence of such high transition frequencies in the spectrum of the control fields. In order to provide numerical proof about this feature, we studied the implementation of one the control processes with a different choice of control operator $H'_c$ which presents a higher connectivity between distant states in the spectrum. Such Hamiltonian matrix is shown in the bottom-right panel of Fig. [\[fig:fig2\]](#fig:fig2){reference-type="ref" reference="fig:fig2"} (a). Results for the new optimized fields are shown in Fig. [\[fig:fig2\]](#fig:fig2){reference-type="ref" reference="fig:fig2"} (b), where we show the frequency bandwidth as a function of $\Gamma$ for this case. It can be seen that $\omega_{bw}$ is greater for $K=2$ than for $K=1$, for all values of $\Gamma$ considered. We point out that, by looking at the representation of $H'_c$ in the computational basis (bottom left panel), we can see that this alternative control procedure would involve tuning a complex combination of multi-spin interactions, in clear contrast with the simple structure of $H_c$.\ The results shown so far allows us to assert that the control bandwidth, which measures the range of frequencies present in the fields is determined exclusively by the energy spread of the free Hamiltonian and the structure of the control Hamiltonian. This gives us a measure of the physical complexity of the control field which turns to be independent of the number of particles in the system. We point out here that we are not interested in analyzing the complexity of the optimization itself, as has been done in previous works which have obtained interesting results. We work our way around this issue by fixing the time step of our numerical implementation to very small values, $J\Delta t = 10^{-2}$. This determines that the maximum allowed frequency in the fields is en every case at least on order of magnitud higher than the actual physical frequencies found by Fourier analysis in the control fields. ## Spectral localization We now turn our attention to another measure of the control field complexity. In this case, we to study how the number of frequencies which appear in the signal spectrum varies as the systems complexity is increased. For this purpose, we define the following quantity \[\mathrm{sIPR} = \left(\int_0^\infty d\omega\:\|\hat{\epsilon}(\omega)\|^4\right)^{-1},\] which we call "spectral inverse participation ratio" (sIPR) as it is inspired in the commonly known IPR. The sIPR quantifies the localization in the Fourier transform of a time signal, and thus allows us to asses how complex the control field is inside its bandwidth. Note that localized frequency spectrums give sIPR$\rightarrow0$, and complex signals with delocalized spectrum tend to higher sIPR. As an example, take a completely random signal with frequency components up to $\omega_{bw}$. We expect such a signal to have a flat Fourier transform $\hat{\epsilon}(\omega)=1/\omega_{bw}$ for $0<\omega<\omega_{bw}$ and $\hat{\epsilon}(\omega)=0$ for $\omega>\omega_{bw}$. Calculating the sIPR in that case is straightforward and gives $\omega_{bw}$. We point out that here we intend to quantify the optimal control field complexity regardless of the frequency distribution width. For this purpose, we evaluate the normalized sIPR \[\mathrm{sIPRn} = \frac{\mathrm{sIPR}}{\omega_{bw}}.\] Following the discussion on the previous paragraph, we expect sIPRn to range between 0 and 1, and we can interpret it as a measure of resamblence between the signal under study and a completely random time field.\ In Fig. [\[fig:fig3\]](#fig:fig3){reference-type="ref" reference="fig:fig3"} we plot the normalized spectral IPR as a function of the NNN exchange $\Gamma$ for different values of the number $K$ of excitations in the spin chain, and for both control processes A and B. We show also some examples of the frequency spectrums we obtained, and it can be corroborated that sIPRn effectively measures how localized the spectrum is in Fourier space. More generally, it can be seen from the figure that sIPRn takes small values for $K=1$ and then grows with $K$, and thus with state space dimension of the system. This is in sharp contrast with the behaviour of the frequency bandwidth $\omega_{bw}$, which was found to be independent of $K$. We point out that this behaviour is common to both control processes. It is interesting to note that the high-dimensional cases ($K=3,4$) roughly converge to the same value of sIPRn, indicating that there maybe an upper bound for this measure which is below its maximal theoretical sIPRn$=1$, which is achieved when the frequency spectrum is flat. Physically, the existence of an upper bound $<1$ means that optimal control fields can always be distinguished from completely random, white noise-type fields. We leave this issue for future investigation.\ Despite the dependance of normalized spectral IPR with the space dimension, it can be seen also that this indicator does not exhibit any clear trend with the NNN parameter $\Gamma$. We observe that, for small values of $K$, this parameter shows large fluctuations which tend to attenuate when for larger space dimensions. We recall that, for $K\geq 3$, the system exhibits a clear transition from a regular energy spectrum to a more complex (chaotic) one at $\Gamma=0.5$. As can be seen from Fig. [\[fig:fig3\]](#fig:fig3){reference-type="ref" reference="fig:fig3"}, there is no evidence of such leap in complexity in our numerical study. In this way, we can assert that the optimal fields required to control the dynamics of regular or chaotic Hamiltonians display a similar spectral complexity.\ As a final remark, we point out that it would not be correct to claim that the spectral properties analized in this section are completely independent of the choice of initial and final state. This can be easily seen by considering a processes where we intend to connect the ground state of the free Hamiltonian $H_{01}$ and one of its excited states $\Ket{n(\Gamma)}$. If the control Hamiltonian $H_c$ connects both states, we expect that the bandwidth of the control field will be given by the energy difference between both levels $E_n(\Gamma)-E_0(\Gamma)$, which can be signficantly lower than the obtained $\omega_{bw}$ for processes A and B if $\Ket{n}$ lies in the low-energy region. Nevertheless, our results do apply to general linear combinations of energy eigenstates, which is the more common scenario. # Discussion {#sec:discuss} We will now look more closely at the connection between the spectral features of the optimal control fields and the structural properties of the system spectrum. We have already pointed out in Sec. [2](#sec:model){reference-type="ref" reference="sec:model"} that the free chain Hamiltonian $H_{01}$ [\[ec:h01\]](#ec:h01){reference-type="ref" reference="ec:h01"} shows a transition in its level spacing distribution $\{\delta E_n\}$ as the NNN coupling parameter $\Gamma$ changes, where \[\delta E_n = E_{n+1}-E_n, \label{ec:levelsp1}\] and $E_n$ is the $n$th ordered eigenvalue of $H_{01}$. If the space dimension is high enough ($K\geq 3$), the level spacings statistics show a Poisson distribution for $\Gamma\lesssim0.5$, and a Wigner-Dyson distribution $\Gamma\gtrsim0.5$ (see the Appendix for more details). We have also discussed in Sec. [3](#sec:result){reference-type="ref" reference="sec:result"} that we observed a connection between the frequency components present in the optimized control field and the resonances of the free Hamiltonian $H_{01}$. Thus, an interesting point arises: if the energy spectrum of $H_{01}$ changes its structure with $\Gamma$, why is there no evidence about those changes in the frequency distribution of the optimal control fields?\ The key point here is to note that the resonances of $H_{01}$, which feed the frequency distribution of the control field, are not only formed by the difference of two consecutive energies $\delta E_n$ ([\[ec:levelsp1\]](#ec:levelsp1){reference-type="ref" reference="ec:levelsp1"}). If connected by the control Hamiltonian, every energy difference present in the spectrum is also a suitable candidate for appearing the control field frequency spectrum. Following this discussion, we studied the distribution of the energy differences defined as \[\delta E_{n,m} = E_{n+m}-E_n\ \mathrm{with}\ 0<m\leq M, \label{ec:levelspm}\] such that $\delta E_{n,1}\equiv \delta E_n$. Note that, for every $n$, the value of $M$ indicates how many levels above $E_n$ are considered, and is thus bounded by the space dimension $D_{K,+}$. In Fig. [\[fig:fig4\]](#fig:fig4){reference-type="ref" reference="fig:fig4"} we show the distributions of normalized energy differences for different values of $M$, using $\Gamma=0$ and $\Gamma=1$. There, it can be seen that both distributions show the expected Poisson and Wigner-Dyson shapes when $M=1$ (as discussed in the previous paragraph), but start to converge to a common form when $M$ grows. As an example, for $K=4$, we have that dimension of the positive subspace is $D_{4,+}\simeq700$, and already taking $M\simeq D_{4,+}/10$ already gives near perfectly matching distributions for both values of the chaos parameter $\Gamma$. This analysis indicates that, while level spacing distributions are quite different for regular and chaotic spectra, the overall energy difference distributions converge to a same shape. This interesting behavior determines that the frequency spectrum of the optimal field which control both type of systems have the same complexity.\ # Final remarks {#sec:conclu} In this work we studied control processes in a chain of spin-1/2 particles and investigated how the complexity of the physical system relates to the complexity of the control field. We studied a Heisenberg chain model, which allowed us to consider separately different space dimensions (ranging from $\sim10$ to $\sim700$ states) by adding excitations to the system. By allowing next-to-nearest neighbour interactions, we were also able to parametrically tune the system from regular to chaotic. We find the time-dependent control fields required to drive different processes using optimal control theory and defined two measures of complexity based on the Fourier spectrum of those fields. By doing so, we could identify which aspects of the systems complexity affect the control fields. For instance, we found that the spectral bandwidth, which measures the maximimum frequency present in the field, is quite generally independent of the system space dimension. However, we showed that exceptions to this rule ocurr if we choose highly non-local control fields. Also, we investigated how many frequencies present inside the signal bandwidth, by defining a measure of localization: the spectral inverse participation ratio (sIPR). We found that this measure of field complexity does increase when excitations are added to system. Finally, we assesed the role of quantum chaos in the control of the system by studying the fields as a function of the chaos parameter $\Gamma$. We found no evidence of the regular-chaotic transition in the field spectral measures, allowing us to assert that the fields required to control chaotic and integrable systems display the same complexity. Concerning the role of quantum chaos in the dynamics of many-body systems, it is interesting to point out that a previous work studied relaxation processes in such systems. Although working in an opposite scenario to coherent control, the authors also found no trace of the chaoticity of the system in the relaxation dynamics. In our case, we present further evidence about the irrelevance of quantum chaos in the coherent dynamics of many-body system.\	{'timestamp': '2016-07-19T02:02:31', 'yymm': '1607', 'arxiv_id': '1607.04701', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04701'}
# Introduction The coupling of electronic and nuclear motions plays a significant role in many fascinating scientific phenomena, e.g., photoreactions, molecular electronics, and strong-field processes. In particular, the study of strong-field processes has been one of the most dynamic research areas in the past few decades with the advent of femtosecond and attosecond technology. Irradiation of atoms and molecules by intense laser pulses gives rise to highly nonlinear effects such as above-threshold ionization or dissociation, Coulomb explosion, or high-harmonic generation. To understand the mechanisms of any nonlinear molecular phenomenon and non-adiabatic reactions, and to move the technology further forward, it is essential to be able to correctly describe any type of coupled electron-nuclear dynamics. Developing such a theoretical tool has been one of the biggest issues in theoretical physics and chemistry. Numerous studies have been conducted and many sophisticated methods have been developed [@enpaper0; @enpaper01; @enpaper1; @enpaper2; @enpaper3; @enpaper4; @enpaper5; @enpaper6; @enpaper7; @enpaper8; @enpaper9; @enpaper10; @enpaper11], among them, multiple-spawning method, multiconfiguraton time-dependent Hartree method, and nonadiabatic Bohmian dynamics method are the methods that retain a quantum description of the nuclei and simulate nonadiabatic electron-nuclear dynamics very accurately. However, they incur huge computational cost when applied to systems with many atoms. Moreover, inclusion of a large number of electronic states is required when higher-intensity fields exist, and ionization processes are very difficult to treat within these approaches. Therefore, alternative approaches have also been developed extensively, which have significantly reduced the computational cost. One of the most widely used approaches is mixed quantum-classical (MQC) approximation, where the nuclei are treated as classical particles, while the electrons are treated quantum mechanically. Among these, the Ehrenfest and trajectory surface hopping (TSH) methods are the most widely used, and have been employed in many studies. However, both Ehrenfest and TSH have certain discrepancies that arise from the fact that, in both methods, the forces acting on classical nuclei and the potential acting on electrons are derived with approximations. There are ongoing intensive efforts to improve these approaches. Recently, a new approach to the coupled electron-nuclear motion, the so-called exact factorization of the electron-nuclear wavefunction, has been proposed. This method provides a new route to go beyond the Born-Oppenheimer (BO) approximation and to study the force acting on the classical nuclei, and then to develop a rigorous MQC method. In this framework, the full wavefunction is written as the product of a nuclear wavefunction and conditional electronic wavefunctions, which parametrically depend on the nuclear configuration. The coupled equations drive the dynamics of these two components, and the motion of the nuclear wavefunction is governed by a single time-dependent Schrödinger equation (TDSE), which contains a time-dependent potential-energy surface (TDPES) and a time-dependent vector potential. Since this nuclear wavefunction provides the exact nuclear and current densities, the TDSE that it satisfies has been identified as the exact nuclear TDSE. The presence of a single exact nuclear TDSE has been found to be very useful in developing the MQC approach systematically. In previous studies, the features of the TDPES in a one-dimensional nonadiabatic electron-transfer model system have been fully analyzed. Indeed, it has been shown that evolving an ensemble of classical nuclear trajectories using the force determined from the gradient of the TDPES reproduces the nuclear wavepacket dynamics very well. These led to the idea of developing the MQC method based on the TDPES and multiple trajectories. Recently, a novel MQC algorithm-the coupled-trajectory (CT) MQC algorithm-has been proposed and shown to be able to accurately simulate the coupled electron-nuclear dynamics in a one-dimensional field-free process. On the other hand, the features of the TDPES under external fields have also been studied. We proposed the reverse factorization, which allows us to define the exact electronic TDSE and the exact electronic TDPES. These are found to be very useful for exploring the mechanism of electron dynamics under an external field. Furthermore, we have recently studied the nuclear TDPES in laser-induced electron localization in the H$_2^+$ molecule, and showed that the propagation of an ensemble of classical trajectories using the gradient of the TDPES yields nuclear density dynamics that are very similar to the exact quantum nuclear ones. This result encourages the idea of developing the MQC dynamics method for the strong-field processes as well. This would be useful since none of the methods that presently exist can accurately simulate the coupled electron-nuclear dynamics of medium-and large-sized systems under a strong field. However, it is still not clear whether the gradient of the TDPES can reproduce the quantum nuclear dynamics in strong-field processes such as strong-field ionization and dissociation, where the quantum effects of the nuclei are significant. In fact, the gradient of the TDPES is not exactly the same as the force that appears in the quantum Hamilton-Jacobi equation derived by applying the Bohmian mechanics approach to the exact nuclear TDSE; it lacks the force from the so-called Bohmian quantum potential. In previous studies, it was shown that a single classical trajectory evolved by the gradient of the TDPES does not yield the correct time evolution of the mean nuclear distance in strong-field dissociation of the one-dimensional H$_2^+$ molecular model. The question then arises as to whether multiple classical trajectories evolved by the quantum Hamilton-Jacobi equation derived from the exact nuclear TDSE give the correct quantum nuclear dynamics in strong-field processes. In this paper, we show that multiple classical trajectories propagated by the gradient of the TDPES plus Bohmian quantum potential can reproduce quantum nuclear dynamics in the strong-field processes: it produces the correct dissociation dynamics and splitting of nuclear probability density in the one-dimensional H$_2^+$ molecular model. The Bohmian quantum potential is found to play a non-negligible role in giving the correct nuclear dynamics for the present strong-field processes, where ionization and/or splitting of the nuclear wavepacket occur/s. Our findings provide a useful basis toward the development of the MQC method for strong-field processes. The rest of the paper is organized as follows. In section II, we briefly review the concepts of the exact factorization of the full electron-nuclear wavefunction and the quantum Hamilton-Jacobi equation derived from the exact nuclear TDSE. There, we show the exact force acting on the classical nuclei and its relationship with the gradient of the TDPES. In section III, we first describe our model system of strong-field dissociation of H$_2^+$, and then show the quantum potential in this system together with the exact TDPES. We then propagate multiple classical trajectories with the force from the gradient of the TDPES plus quantum potential and demonstrate that it perfectly reproduces the quantum nuclear dynamics in strong-field processes. We also show the role of the quantum potential by showing the dynamics propagated only by the gradient of the TDPES. In section IV, we summarize the results and speculate on future directions. # THEORY In Ref., it was shown that the full electron-nuclear wavefunction $\Psi({\underline{\underline{\bf R}}},{\underline{\underline{\bf r}}}, t)$ that solves the TDSE $\hat{H}\Psi({\underline{\underline{\bf R}}},{\underline{\underline{\bf r}}},t)=i\partial_t\Psi({\underline{\underline{\bf R}}},{\underline{\underline{\bf r}}}, t)$ can be factorized exactly to the single product \[\Psi({\underline{\underline{\bf R}}},{\underline{\underline{\bf r}}}, t)=\chi({\underline{\underline{\bf R}}},t)\Phi_{\underline{\underline{\bf R}}}({\underline{\underline{\bf r}}},t) \label{eqn: factorization}\] of the nuclear wavefunction $\chi({\underline{\underline{\bf R}}},t)$ and the electronic wavefunction $\Phi_{\underline{\underline{\bf R}}}({\underline{\underline{\bf r}}},t)$ that parametrically depends on the nuclear positions ${\underline{\underline{\bf R}}}$ and satisfies the partial normalization condition \[\int d{\underline{\underline{\bf r}}} \|\Phi_{\underline{\underline{\bf R}}}({\underline{\underline{\bf r}}},t)\|^2=1\] for any ${\underline{\underline{\bf R}}}$ and $t$. Throughout this paper, ${\underline{\underline{\bf R}}}$ and ${\underline{\underline{\bf r}}}$ collectively represent the sets of nuclear and electronic coordinates, respectively (i.e., ${\underline{\underline{\bf R}}} \equiv \{ {\bf R}_1,{\bf R}_2,\cdots,{\bf R}_{N_n} \}$ and ${\underline{\underline{\bf r}}} \equiv \{ {\bf r}_1,{\bf r}_2,\cdots,{\bf r}_{N_e} \}$), and atomic units are used unless stated otherwise. The complete molecular Hamiltonian is \[\hat{H} = \hat{T}_{\rm n}({\underline{\underline{\bf R}}})+ \hat{V}^{\rm n}_{\rm ext}({\underline{\underline{\bf R}}},t) +\hat{H}_{\rm BO}({\underline{\underline{\bf r}}},{\underline{\underline{\bf R}}}) +\hat{v}^{\rm e}_{\rm ext}({\underline{\underline{\bf r}}},t),\] and $\hat{H}_{\rm BO}({\underline{\underline{\bf r}}},{\underline{\underline{\bf R}}})$ is the BO electronic Hamiltonian, \[\hat{H}_{\rm BO} = \hat{T}_{\rm e}({\underline{\underline{\bf r}}})+ \hat{W}_{\rm ee}({\underline{\underline{\bf r}}}) +\hat{W}_{\rm en}({\underline{\underline{\bf r}}},{\underline{\underline{\bf R}}})+\hat{W}_{\rm nn}({\underline{\underline{\bf R}}}),\] where $\hat{T}_{\rm n}=-\sum_{\alpha=1}^{N_{\rm n}}\frac{\nabla^2_\alpha}{2M_\alpha}$ and $\hat{T}_{\rm e}=-\sum_{j=1}^{N_{\rm e}}\frac{\nabla^2_j}{2}$ are the nuclear and electronic kinetic energy operators, $\hat{W}_{\rm ee}$, $\hat{W}_{\rm en}$ and $\hat{W}_{\rm nn}$ are the electron-electron, electron-nuclear and nuclear-nuclear interactions, and $\hat{V}^{\rm n}_{\rm ext}({\underline{\underline{\bf R}}},t)$ and $\hat{v}^{\rm e}_{\rm ext}({\underline{\underline{\bf r}}},t)$ are time-dependent (TD) external potentials acting on the nuclei and electrons, respectively. The stationary variations of the quantum mechanical action with respect to $\Phi_{\underline{\underline{\bf R}}}({\underline{\underline{\bf r}}},t)$ and $\chi({\underline{\underline{\bf R}}},t)$ under the normalization condition of $\Phi_{\underline{\underline{\bf R}}}({\underline{\underline{\bf r}}},t)$ lead to the following equations of motion for $\chi({\underline{\underline{\bf R}}},t)$ and $\Phi_{\underline{\underline{\bf R}}}({\underline{\underline{\bf r}}},t)$: \[\begin{split} \left(\hat{H}_{\rm BO}({\underline{\underline{\bf r}}},{\underline{\underline{\bf R}}})+\hat{v}^{\rm e}_{\rm ext}({\underline{\underline{\bf r}}},t) +\hat U_{\rm en}^{\rm coup}[\Phi_{\underline{\underline{\bf R}}},\chi]-\epsilon({\underline{\underline{\bf R}}},t)\right) \Phi_{{\underline{\underline{\bf R}}}}({\underline{\underline{\bf r}}},t)\\ =i\partial_t \Phi_{{\underline{\underline{\bf R}}}}({\underline{\underline{\bf r}}},t) \end{split}\label{eqn: exact electronic eqn}\] \[\begin{split} \left[\sum_{\alpha=1}^{N_{\rm n}} \frac{\left[-i\nabla_\alpha+{\bf A}_\alpha({\underline{\underline{\bf R}}},t)\right]^2}{2M_\alpha} +\hat{V}^{\rm n}_{\rm ext}({\underline{\underline{\bf R}}},t) + \epsilon({\underline{\underline{\bf R}}},t)\right]\chi({\underline{\underline{\bf R}}},t)\\ =i\partial_t \chi({\underline{\underline{\bf R}}},t) \label{eqn: exact nuclear eqn}. \end{split}\] Here, $\epsilon({\underline{\underline{\bf R}}},t)$ is the exact nuclear TDPES \[\label{eqn: tdpes} \epsilon({\underline{\underline{\bf R}}},t)=\left\langle\Phi_{\underline{\underline{\bf R}}}(t)\right\|\hat{H}_{\rm BO}+\hat{v}^{\rm e}_{\rm ext}({\underline{\underline{\bf r}}},t) +\hat U_{\rm en}^{\rm coup}-i\partial_t\left\| \Phi_{\underline{\underline{\bf R}}}(t)\right\rangle_{\underline{\underline{\bf r}}},\] $\hat U_{\rm en}^{\rm coup}[\Phi_{\underline{\underline{\bf R}}},\chi]$ is the electron-nuclear coupling operator, \[\begin{aligned} \hat U_{\rm en}^{\rm coup}&[\Phi_{\underline{\underline{\bf R}}},\chi]=\sum_{\alpha=1}^{N_{\rm n}}\frac{1}{M_\alpha}\left[ \frac{\left[-i\nabla_\alpha-{\bf A}_\alpha({\underline{\underline{\bf R}}},t)\right]^2}{2} \right.\label{eqn: enco} \\ & \left.+\left(\frac{-i\nabla_\alpha\chi}{\chi}+{\bf A}_\alpha({\underline{\underline{\bf R}}},t)\right) \left(-i\nabla_\alpha-{\bf A}_{\alpha}({\underline{\underline{\bf R}}},t)\right)\right],\nonumber \end{aligned}\] and ${\bf A}_{\alpha}\left({\underline{\underline{\bf R}}},t\right)$ is the TD vector potential, \[\label{eqn: vector potential} {\bf A}_{\alpha}\left({\underline{\underline{\bf R}}},t\right) = \left\langle\Phi_{\underline{\underline{\bf R}}}(t)\right\|-i\nabla_\alpha\left.\Phi_{\underline{\underline{\bf R}}}(t) \right\rangle_{\underline{\underline{\bf r}}}.\] The symbol $\left\langle\,\,\cdot\,\,\right\rangle_{\underline{\underline{\bf r}}}$ indicates an integration over electronic coordinates only. The partial normalization condition of $\Phi_{\underline{\underline{\bf R}}}({\underline{\underline{\bf r}}},t)$ makes the factorization ([\[eqn: factorization\]](#eqn: factorization){reference-type="ref" reference="eqn: factorization"}) unique up to within a $({\underline{\underline{\bf R}}},t)$-dependent gauge transformation, $\chi({\underline{\underline{\bf R}}},t)\rightarrow\tilde\chi({\underline{\underline{\bf R}}},t)=e^{-i\theta({\underline{\underline{\bf R}}},t)}\chi({\underline{\underline{\bf R}}},t)$ and $\Phi_{\underline{\underline{\bf R}}}({\underline{\underline{\bf r}}},t)\rightarrow\tilde\Phi_{\underline{\underline{\bf R}}}({\underline{\underline{\bf r}}},t)=e^{i\theta({\underline{\underline{\bf R}}},t)}\Phi_{\underline{\underline{\bf R}}}({\underline{\underline{\bf r}}},t)$. Eqs. ([\[eqn: exact electronic eqn\]](#eqn: exact electronic eqn){reference-type="ref" reference="eqn: exact electronic eqn"}) and ([\[eqn: exact nuclear eqn\]](#eqn: exact nuclear eqn){reference-type="ref" reference="eqn: exact nuclear eqn"}) are form invariant under this transformation while the scalar potential and the vector potential transform as $\tilde{\epsilon}({\underline{\underline{\bf R}}},t) = \epsilon({\underline{\underline{\bf R}}},t)+\partial_t\theta({\underline{\underline{\bf R}}},t)$ and $\tilde{\bf A}_{\alpha}({\underline{\underline{\bf R}}},t) = {\bf A}_{\alpha}({\underline{\underline{\bf R}}},t)+\nabla_\alpha\theta({\underline{\underline{\bf R}}},t)$. The equation for $\chi({\underline{\underline{\bf R}}},t)$, Eq. ([\[eqn: exact nuclear eqn\]](#eqn: exact nuclear eqn){reference-type="ref" reference="eqn: exact nuclear eqn"}), has the form of a Schrödinger equation. Note that $\chi({\underline{\underline{\bf R}}},t)$ can be interpreted as the exact nuclear wave-function since it leads to an $N$-body nuclear density, $\Gamma({\underline{\underline{\bf R}}},t)=\vert\chi({\underline{\underline{\bf R}}},t)\vert^2,$ and an $N$-body current density, ${\bf J}_\alpha({\underline{\underline{\bf R}}},t)=\frac{1}{M_\alpha}\Big[\mbox{Im}(\chi^({\underline{\underline{\bf R}}},t)\nabla_\alpha\chi({\underline{\underline{\bf R}}},t))+ \Gamma({\underline{\underline{\bf R}}},t){\bf A}_\alpha({\underline{\underline{\bf R}}},t)\Big],$ which yields the true nuclear $N$-body density and current density obtained from the full wavefunction $\Psi({\underline{\underline{\bf R}}},{\underline{\underline{\bf r}}},t)$. Therefore, the equation ([\[eqn: exact nuclear eqn\]](#eqn: exact nuclear eqn){reference-type="ref" reference="eqn: exact nuclear eqn"}) can be regarded the exact* nuclear TDSE. Having a single exact TDSE for the nuclear subsystem, it is possible to consider its hydrodynamic reformulation using the approach of Bohmian mechanics, to study how the exact force acting on the classical nuclei can be defined. To this end, the polar forms of the wavefunction $\chi({\underline{\underline{\bf R}}},t)=\vert\chi({\underline{\underline{\bf R}}},t)\vert e^{iS({\underline{\underline{\bf R}}},t)}$ ($\vert\chi({\underline{\underline{\bf R}}},t)\vert$ and $S({\underline{\underline{\bf R}}},t)$ are real-valued amplitude and action functions, respectively) are substituted into the exact nuclear TDSE ([\[eqn: exact nuclear eqn\]](#eqn: exact nuclear eqn){reference-type="ref" reference="eqn: exact nuclear eqn"}). Here, to easily find the exact force acting on the classical nuclei, we set the gauge of the wavefunction such that the vector potential ${\bf A}_\alpha({\underline{\underline{\bf R}}},t)$ is always zero. Note that whenever the vector potential is curl-free ($\nabla_\alpha\times{\bf A}_\alpha({\underline{\underline{\bf R}}},t)=0$), the gauge can be chosen such that ${\bf A}_\alpha({\underline{\underline{\bf R}}},t)$ is zero. Whether and under which conditions $\nabla_\alpha\times{\bf A}_\alpha({\underline{\underline{\bf R}}},t)=0$ is currently under investigation. Under this choice of the gauge, the exact nuclear TDSE ([\[eqn: exact nuclear eqn\]](#eqn: exact nuclear eqn){reference-type="ref" reference="eqn: exact nuclear eqn"}) can be written as \[\left[-\sum_{\alpha=1}^{N_n} \frac{\nabla_\alpha^2}{2M_\alpha} + \epsilon({\underline{\underline{\bf R}}},t)\right]\chi({\underline{\underline{\bf R}}},t) =i\partial_t \chi({\underline{\underline{\bf R}}},t) \label{eqn: exact nuclear eqn2}.\] Here we include $\hat{V}^{\rm n}_{\rm ext}({\underline{\underline{\bf R}}},t)$ in the exact TDPES $\epsilon({\underline{\underline{\bf R}}},t)$. Substituting $\chi({\underline{\underline{\bf R}}},t)=\vert\chi({\underline{\underline{\bf R}}},t)\vert e^{iS({\underline{\underline{\bf R}}},t)}$ into Eq. ([\[eqn: exact nuclear eqn2\]](#eqn: exact nuclear eqn2){reference-type="ref" reference="eqn: exact nuclear eqn2"}), the following two coupled equations are obtained: \[\begin{split} \sum_{\alpha=1}^{N_{\rm n}}\frac{(\nabla_\alpha S({\underline{\underline{\bf R}}},t))^2}{2M_\alpha}+\epsilon({\underline{\underline{\bf R}}},t)-\sum_{\alpha=1}^{N_{\rm n}}\frac{1}{2M_\alpha}\frac{\nabla_\alpha^2 \vert\chi({\underline{\underline{\bf R}}},t)\vert}{\vert\chi({\underline{\underline{\bf R}}},t)\vert}\\ =-\frac{\partial S({\underline{\underline{\bf R}}},t)}{\partial t} \label{eqn: QHJeqn} \end{split}\] and \[\begin{split}-\sum_{\alpha=1}^{N_{\rm n}}\frac{\nabla_\alpha S({\underline{\underline{\bf R}}},t)\cdot\nabla_\alpha \vert\chi({\underline{\underline{\bf R}}},t)\vert}{M_\alpha}-\sum_{\alpha=1}^{N_{\rm n}}\frac{\vert\chi({\underline{\underline{\bf R}}},t)\vert\nabla_\alpha^2 S({\underline{\underline{\bf R}}},t)}{2M_\alpha}\\ =\frac{\partial \vert\chi({\underline{\underline{\bf R}}},t)\vert}{\partial t}. \label{eqn: continuity} \end{split}\] Equation ([\[eqn: QHJeqn\]](#eqn: QHJeqn){reference-type="ref" reference="eqn: QHJeqn"}) and ([\[eqn: continuity\]](#eqn: continuity){reference-type="ref" reference="eqn: continuity"}) are the hydrodynamic formulation of the exact nuclear TDSE ([\[eqn: exact nuclear eqn2\]](#eqn: exact nuclear eqn2){reference-type="ref" reference="eqn: exact nuclear eqn2"}). Equation ([\[eqn: QHJeqn\]](#eqn: QHJeqn){reference-type="ref" reference="eqn: QHJeqn"}) can be regarded as the exact quantum Hamilton-Jacobi equation, while ([\[eqn: continuity\]](#eqn: continuity){reference-type="ref" reference="eqn: continuity"}) produces the continuity equation. Identifying $\nabla_\alpha S({\underline{\underline{\bf R}}},t)$ as a momentum ${\bf P}_\alpha$ of a classical trajectory, ([\[eqn: QHJeqn\]](#eqn: QHJeqn){reference-type="ref" reference="eqn: QHJeqn"}) can be solved by propagating an ensemble of classical trajectories that obey the following Newton's equations: \[\begin{split} \frac{d {\bf P}_\alpha}{dt}&=-\nabla_\alpha \left[ \epsilon({\underline{\underline{\bf R}}},t)-\sum_{\alpha=1}^{N_{\rm n}}\frac{1}{2M_\alpha}\frac{\nabla_\alpha^2 \vert\chi({\underline{\underline{\bf R}}},t)\vert}{\vert\chi({\underline{\underline{\bf R}}},t)\vert}\right]\\ &=-\nabla_\alpha \left[\epsilon({\underline{\underline{\bf R}}},t)+\epsilon^{\rm QP}({\underline{\underline{\bf R}}},t)\right]. \label{eqn: Newton} \end{split}\] The right-hand side of ([\[eqn: Newton\]](#eqn: Newton){reference-type="ref" reference="eqn: Newton"}) can now be considered as the exact force acting on the classical nuclei, since it is derived from the exact nuclear TDSE. It is a gradient of the sum of the exact TDPES $\epsilon({\underline{\underline{\bf R}}},t)$ and the additional time-dependent potential \[\epsilon^{\rm QP}({\underline{\underline{\bf R}}},t)=-\sum_{\alpha=1}^{N_{\rm n}}\frac{1}{2M_\alpha} \frac{\nabla_\alpha^2 \vert\chi({\underline{\underline{\bf R}}},t)\vert}{\vert\chi({\underline{\underline{\bf R}}},t)\vert}, \label{eqn: eQP}\] which is referred to as the quantum potential in Bohmian mechanics. In the previous studies, we propagated multiple classical trajectories according to ([\[eqn: Newton\]](#eqn: Newton){reference-type="ref" reference="eqn: Newton"}) without this Bohmian quantum potential $\epsilon^{\rm QP}({\underline{\underline{\bf R}}},t)$, i.e., \[\begin{split} \frac{d {\bf P}_\alpha}{dt}=-\nabla_\alpha \epsilon({\underline{\underline{\bf R}}},t) \label{eqn: Newton2} \end{split}\] for the field-free nonadiabatic charge-transfer process and the laser-induced electron localization processes in the H$_2^+$ molecule. We found that an ensemble of independent classical nuclear trajectories on $\epsilon({\underline{\underline{\bf R}}},t)$ provides dynamics that accurately reproduce the exact nuclear wavepacket dynamics. Here, we study whether the same multiple classical trajectory approach can also reproduce strong-field processes in which ionization and/or splitting of nuclear density occur/s. In such strong-field processes, nuclear quantum effects are significant; in fact, previous studies have shown that a single classical trajectory cannot yield the molecular dissociation via tunneling that occurs under the strong field, even though it is propagated by the force of the gradient of the exact TDPES, $\epsilon({\underline{\underline{\bf R}}},t)$. In the next section, we will propagate multiple classical trajectories by $\epsilon({\underline{\underline{\bf R}}},t)$ according to ([\[eqn: Newton2\]](#eqn: Newton2){reference-type="ref" reference="eqn: Newton2"}). We will also calculate the exact quantum potential $\epsilon^{\rm QP}({\underline{\underline{\bf R}}},t)$ and propagate multiple classical trajectories by $\epsilon({\underline{\underline{\bf R}}},t)+\epsilon^{\rm QP}({\underline{\underline{\bf R}}},t)$, i.e., ([\[eqn: Newton\]](#eqn: Newton){reference-type="ref" reference="eqn: Newton"}), to study the importance of the force from $\epsilon^{\rm QP}({\underline{\underline{\bf R}}},t)$. Note this study demonstrates the role of $\epsilon^{\rm QP}({\underline{\underline{\bf R}}},t)$ in multiple classical-trajectories dynamics for the first time. # Results and discussion ## Theoretical model To study whether the propagation of multiple classical trajectories can reproduce the exact quantum nuclear dynamics in strong laser fields, we employ a simplified model of the H$_2^+$ molecule, which is the same as that used in previous studies. In this model, the dimensionality of the problem is reduced by restricting the motion of the nuclei and the electron to the direction of the polarization axis of the laser field. In the center-of-mass system, the dynamics of this one-dimensional model of H$_2^+$ is governed by the full Hamiltonian $\hat{H}(R,r,t) =\hat{T}_{\rm n}(R)+\hat{T}_{\rm e}(r)+\hat{W}_{\rm nn}(R)+\hat{W}_{\rm en}(R,r)+\hat{v}_{\rm laser}(r,t)$, where $R$ is the internuclear distance and $r$ is the electronic coordinate as measured from the nuclear center of mass. The kinetic energy terms are $\hat{T}_{\rm n}(R) =-\frac{1}{2\mu_n}\frac{\partial^2}{\partial R^2}$ and, $\hat{T}_{\rm e}(r) =-\frac{1}{2\mu_{\rm e}}\frac{\partial^2}{\partial r^2}$, respectively, where the reduced mass of the nuclei is given by $\mu_{\rm n}=M_{\rm H}/2$, and reduced electronic mass is given by $\mu_{\rm e}=\frac{2M_{\rm H}}{2M_{\rm H}+1}$ ($M_{\rm H}$ is the proton mass). The interactions are soft-Coulomb: $\hat{W}_{\rm nn}(R) = \frac{1}{\sqrt{0.03+R^2}}$, and $\hat{W}_{\rm en}(R,r) =-\frac{1}{\sqrt{1.0+(r-\frac{R}{2})^2}}-\frac{1}{\sqrt{1.0+(r+\frac{R}{2})^2}}$ (and $\hat{W}_{\rm ee} = 0$). The field is described within the dipole approximation and length gauge, as $\hat{v}_{\rm laser}(r,t) = E(t)q_{\rm e}r$, where the reduced charge $q_{\rm e}=\frac{2M_{\rm H}+2}{2M_{\rm H}+1}$. This reduced-dimensional model has proven useful since it allows numerically exact solutions to the TDSE while capturing the essential physics in strong-field processes such as multiphoton ionization, above-threshold ionization and dissociation, enhanced ionization, non-sequential double ionization, and high-harmonic generation. In this study, we investigate the dynamics of the model H$_2^+$ system under a $\lambda=228$ nm ($5.4$ eV) UV-laser pulse, which is represented by $E(t)=E_0f(t)\sin(\omega t),$ with two peak intensities, $I_1 =\vert E_0\vert^2=10^{14}$ W/cm$^2$ and $I_2 =\vert E_0\vert^2=2.5\times10^{13}$ W/cm$^2$. This frequency provides an energy that is about twice as much as the dissociation energy of the model molecule ($2.88$ eV); thus, dissociation is expected. The envelope function $f(t)$ is chosen such that the field is linearly ramped from zero to its maximum strength at $t=7.6$ fs (over 10 optical cycles) and thereafter, held constant for an additional 15 laser cycles, corresponding to a total simulation time of about 19 fs. The same system and parameters were employed in previous studies, where the important role of the complex coupling between the electronic and nuclear motions in these strong-field systems was revealed. In, in particular, the exact TDPES ([\[eqn: tdpes\]](#eqn: tdpes){reference-type="ref" reference="eqn: tdpes"}) in these systems was calculated and shown to be a very useful tool for analyzing and interpreting the complicated quantum nuclear dynamics in the strong-field processes. Here, we will study the possibility of the Bohmian mechanics being established for these strong-field processes by using the concept of exact factorization, and whether multiple classical trajectories can give the correct quantum nuclear dynamics. We first calculated the full molecular wavefunction $\Psi(R,r,t)$ by propagating the full TDSE \[\hat{H}(R,r,t)\Psi(R,r,t)=i\partial_t\Psi(R,r,t) \label{eqn: fullTDSE}\] numerically exactly using the second-order split-operator method. As the initial state of the time propagation, $\Psi(R,r,t)$ was prepared in its ground state by imaginary-time propagation. In Fig. [\[fig:Fig1\]](#fig:Fig1){reference-type="ref" reference="fig:Fig1"}, the electron-nuclear density $\|\Psi(R,r,t)\|^2$ at $t=18.6$ fs (after the 24th optical cycle) is shown for both the higher-intensity case ($I_1 =10^{14}$ W/cm$^2$) (upper panel) and the lower-intensity case ($I_2 =2.5\times10^{13}$ W/cm$^2$) (lower panel). These indicate the probability of finding an electron at position $r$ and the nuclear separation at position $R$ at $t=18.6$ fs for each case. In the upper panel, it is observed that $\|\Psi(R,r,t)\|^2$ at $t=18.6$ fs exists at larger $R$ compared to the expectation value at the ground state $\langle R \rangle(t=0)=2.65$ a.u., indicating that dissociation occurred. We also observe large streaks of $\|\Psi(R,r,t)\|^2$ in both negative and positive $r$ directions, which shows that a considerable ionization occurred in this higher-intensity case. Therefore, dissociation occurs here via the Coulomb-explosion mechanism, as already discussed in previous studies. However, in the lower-intensity case (lower panel in Fig. [\[fig:Fig1\]](#fig:Fig1){reference-type="ref" reference="fig:Fig1"}), different dynamics occurred: a small amount of $\|\Psi(R,r,t)\|^2$ exists in the region larger than $\langle R \rangle(t=0)=2.65$ a.u., but a large part of it remains in the ground-state position around $\langle R \rangle(t=0)$. Therefore, a splitting of probability density occurred here and a small amount of $\|\Psi(R,r,t)\|^2$ went to dissociation. It is also seen that the probability of ionization is very low, and hence, dissociation occurred predominantly via the photodissociation channel (H$^+_2\rightarrow$H$+$H$^+$). Previous studies have shown that this lower-intensity case ($I_2 =2.5\times10^{13}$ W/cm$^2$) represents a particularly challenging system when we consider simulating it by using the approximated method. The Ehrenfest and time-dependent Hartree methods could not reproduce the probability of dissociation of this system. A more sophisticated correlated time-dependent variational approach succeeded in giving the dissociation probability to some degree, but still could not reproduce the nuclear density dynamics well. The exact TDPES in this system provided a clear picture that explains the difficulty in the simulation of these nuclear dynamics: it was shown that the quantum tunneling through the TDPES occurs, which is difficult to reproduce by the approximated methods. The question arises as to whether we get the force acting on classical nuclei that gives the correct dynamics in this system if we formulate Bohmian mechanics in the exact-factorization framework. ## Multiple classical trajectory dynamics on the exact TDPES + the exact quantum potential To answer the questions raised in the previous sections, we now present our results. We begin with the calculation of the exact TDPES $\epsilon(R,t)$ ([\[eqn: tdpes\]](#eqn: tdpes){reference-type="ref" reference="eqn: tdpes"}) and the quantum potential $\epsilon^{\rm QP}(R,t)$ ([\[eqn: eQP\]](#eqn: eQP){reference-type="ref" reference="eqn: eQP"}) that appear in the exact quantum Hamilton-Jacobi equation ([\[eqn: QHJeqn\]](#eqn: QHJeqn){reference-type="ref" reference="eqn: QHJeqn"}). Since we already have numerically exact $\Psi(R,r,t)$ at each time step obtained by propagating the full TDSE ([\[eqn: fullTDSE\]](#eqn: fullTDSE){reference-type="ref" reference="eqn: fullTDSE"}), we can easily calculate the TDPES $\epsilon(R,t)$ by fixing the gauge and $\epsilon^{\rm QP}(R,t)$ with $\|\chi(R,t)\|=\sqrt{\int{\|\Psi(R,r,t)\|^2 dr}}$. In Fig. [\[fig:Fig2\]](#fig:Fig2){reference-type="ref" reference="fig:Fig2"} (a) and (b), we show snapshots of the exact TDPES $\epsilon(R,t)$ (blue or dark gray solid lines) and the quantum potential $\epsilon^{\rm QP}(R,t)$ (pink or light gray solid lines) at times indicated for the system with peak intensities $I_1 =10^{14}$ W/cm$^2$ (Fig. [\[fig:Fig2\]](#fig:Fig2){reference-type="ref" reference="fig:Fig2"} (a)) and $I_2 =2.5\times10^{13}$ W/cm$^2$ (Fig. [\[fig:Fig2\]](#fig:Fig2){reference-type="ref" reference="fig:Fig2"} (b)), respectively. The nuclear density at each time is also shown (black solid lines). Note that the TDPES have already been reported in previous studies, and here we show the quantum potential for the first time. The TDPES in each system shows the characteristic feature of the time-dependent potentials that the nuclear wavepacket experiences. In the Coulomb explosion case (Fig. [\[fig:Fig2\]](#fig:Fig2){reference-type="ref" reference="fig:Fig2"}(a)), the wells in the TDPES $\epsilon(R,t)$ that confine the wavefunction in the ground state flatten out as the laser is switched on, causing the nuclear density to spill out to larger separations. However, in the photodissociation (without ionization) case (Fig. [\[fig:Fig2\]](#fig:Fig2){reference-type="ref" reference="fig:Fig2"} (b)), the well in the $\epsilon(R,t)$ exists at all times, and thus, the nuclear density can leak out from it only by tunneling. By comparing the upper and lower panels in both laser cases, we find that $\epsilon^{\rm QP}(R,t)$ has non-negligible structures, especially at earlier simulation times. It is found that $\epsilon^{\rm QP}(R,t)$ actually has an effect that flattens the well in $\epsilon(R,t)$, thus helping the nuclei delocalize against the force from the confining potential at the initial time and leak out from it. Especially at the initial time when the wavefunction is in its ground state, the shape of $\epsilon^{\rm QP}(R,t=0)$ is exactly opposite to that of TDPES $\epsilon(R,t=0)$ as seen in the left-hand panels of Figs. [\[fig:Fig2\]](#fig:Fig2){reference-type="ref" reference="fig:Fig2"} (a) and (b). This is understood by the fact that the ground state of a one-dimensional wavefunction is expressed by a real function multiplied by a complex constant, i.e., $\chi(R)=\|\chi(R)\|e^{ia}$ where $a$ is a real constant. Since the exact nuclear Schrödinger equation in the ground state is written as \[\left[-\frac{1}{2\mu_n}\frac{d^2}{dR^2} + \epsilon(R)\right]\chi(R) =E \chi(R) \label{eqn: static nuclear eqn}\] (where $E$ is the total energy of the system in the ground-state), the quantum potential at the initial time can be written as \[\epsilon^{\rm QP}(R)=\frac{-\frac{1}{2\mu_{\rm n}}\frac{d^2}{dR^2}\|\chi(R)\|}{\|\chi(R)\|}=E-\epsilon(R). \label{eqn: static eQP}\] Thus, $\epsilon^{\rm QP}(R)$ shows the opposite curvature to the exact TDPES $\epsilon(R)$. We now demonstrate that the effect of $\epsilon^{\rm QP}$ on the dynamics is significant, especially for the low-intensity case, as it causes the tunneling of the classical nuclei. Having the exact TDPES $\epsilon(R,t)$ and the quantum potential $\epsilon^{\rm QP}(r,t)$ at each time step, we now solve the exact quantum Hamilton-Jacobi equation ([\[eqn: QHJeqn\]](#eqn: QHJeqn){reference-type="ref" reference="eqn: QHJeqn"}) for these two systems by propagating multiple classical trajectories according to Newton's equation ([\[eqn: Newton\]](#eqn: Newton){reference-type="ref" reference="eqn: Newton"}). We first show the results for the higher-intensity ($I_1 =10^{14}$ W/cm$^2$) case. We propagate 1000 trajectories according to Eq. ([\[eqn: Newton\]](#eqn: Newton){reference-type="ref" reference="eqn: Newton"}), where the initial positions $R$ are sampled from the initial (ground-state) nuclear density $N(R,t=0)=\int{\|\Psi(R,r,t=0)\|^2 dr}$, which was obtained in the previous TDSE calculation, and the initial momentum is set to zero since $\frac{d}{dR}S(R,t)$ is zero at the initial time. In Fig. [\[fig:Fig3\]](#fig:Fig3){reference-type="ref" reference="fig:Fig3"} (a), the nuclear density reconstructed as a histogram from the distributions of classical positions evolving on $\epsilon(R,t)+\epsilon^{\rm QP}(R,t)$ for the system with the peak intensity $I_1 =10^{14}$ W/cm$^2$ is shown as red (or gray) linepoints. The exact nuclear density obtained from the full TDSE ([\[eqn: fullTDSE\]](#eqn: fullTDSE){reference-type="ref" reference="eqn: fullTDSE"}) is also shown as a black solid line. By comparing the red (or gray) linepoints and black line, we find that the nuclear density obtained from the multiple classical trajectories yields the exact quantum nuclear dynamics in this strong-field process: the characteristic Coulomb-explosion dynamics of the quantum nuclei are reproduced by the ensemble of classical trajectories. This result indicates that even strong-field processes in which ionization occurs can be simulated, in principle, by the MQC approximation method. We also show $\epsilon(R,t)+\epsilon^{\rm QP}(R,t)$ (green or light gray dashed line) and $\epsilon(R,t)$ (blue or dark gray solid line). By comparing these, the discussion above is confirmed: $\epsilon^{\rm QP}$ plays a role to flatten the well in $\epsilon(R,t)$, which initially confines the wavepacket to the equilibrium position, thus enhancing dissociation. Here, in the higher-intensity ($I_1 =10^{14}$ W/cm$^2$) case, together with the TDPES that gives a strong repulsive force that reflects ionization, almost the entire nuclear density leaks out from the well and moves to the dissociation. We observe that, when the nuclear density moves outside the well, the effect of $\epsilon^{\rm QP}$ becomes very small. Thus, in this higher-intensity case, the quantum potential affects the dynamics only during the earlier time of the propagation when the nuclear density is about to leak out from the well. Next, we turn to the lower-intensity case ($I_2 =2.5\times10^{13}$ W/cm$^2$), where tunneling of the nuclear wavepacket occurs and causes the difficulty to be simulated by the approximated method. We propagate 2,000 trajectories according to ([\[eqn: Newton\]](#eqn: Newton){reference-type="ref" reference="eqn: Newton"}) with $\epsilon(R,t)$ plus $\epsilon^{\rm QP}(R,t)$ calculated above. Snapshots of the nuclear density reconstructed as a histogram from the distributions of classical positions are plotted as red (or gray) linepoints in Fig. [\[fig:Fig3\]](#fig:Fig3){reference-type="ref" reference="fig:Fig3"} (b). Comparison with the exact nuclear density (black solid line in the same figure) again shows excellent agreement. Multiple classical trajectories evolving on $\epsilon(R,t)+\epsilon^{\rm QP}(R,t)$ perfectly reproduce the splitting of the exact nuclear density, i.e., long-reaching tails of the exact nuclear density are correctly reproduced. Thus, it is found that the strong-field photodissociation dynamics can also be simulated, in principle, by multiple classical trajectories when their motion is driven by the correct force, i.e., the gradient of $(\epsilon(R,t)+\epsilon^{\rm QP}(R,t))$. In the same figure, $\epsilon(R,t)+\epsilon^{\rm QP}(R,t)$ and $\epsilon(R,t)$ are also plotted. Similar to the higher-intensity case, $\epsilon^{\rm QP}(R,t)$ flattens the well in $\epsilon(R,t)$; thus, it allows the nuclear density to escape from the confining potential. This explains how quantum tunneling is reproduced by classical-trajectory dynamics: the quantum potential is responsible for it. Here, a large part of the nuclear density remains in the ground-state position and the exact TDPES $\epsilon(R,t)$ has a confining well at all times. However, the quantum potential $\epsilon^{\rm QP}(R,t)$ has the opposite curve to $\epsilon(R,t)$ around the equilibrium position. Thus, in total, $\epsilon(R,t)+\epsilon^{\rm QP}(R,t)$ always shows a somewhat flattened structure and some classical nuclei can experience the force to escape the well and go to the dissociation. Note that once the nuclei leak outside the well, the quantum potential $\epsilon^{\rm QP}(R,t)$ has an almost negligible effect on their dynamics. Finally, to confirm if the above discussion about the role of the quantum potential $\epsilon^{\rm QP}(R,t)$ is correct, we compute the multiple classical trajectory dynamics without taking account of the quantum potential, i.e., propagated according to ([\[eqn: Newton2\]](#eqn: Newton2){reference-type="ref" reference="eqn: Newton2"}). In Fig. [\[fig:Fig4\]](#fig:Fig4){reference-type="ref" reference="fig:Fig4"} (a) and (c), we show snapshots of the calculated dynamics: the blue (or dark gray) linepoints indicate the nuclear density reconstructed as a histogram from the distribution of the classical positions evolving only on $\epsilon(R,t)$. ((a) shows the results for the higher-intensity case and (c) is for the lower-intensity one). As a reference, the densities obtained from the multiple classical trajectories on $\epsilon(R,t)+\epsilon^{\rm QP}(R,t)$ (red or light gray linepoints) and the exact nuclear density (black solid line) are also shown. To see the difference between these results more clearly, in Figs. [\[fig:Fig4\]](#fig:Fig4){reference-type="ref" reference="fig:Fig4"} (b) and (d), we also plot the time evolution of the mean inter-nuclear distance $\langle R\rangle(t)$ obtained from each calculation ((b) is for the higher-intensity case and (d) is for the lower-intensity one). Exact results (black solid line) are obtained by $\langle R\rangle(t)=\langle\Psi(R,r,t)\|\hat{R}\|\Psi(R,r,t)\rangle$ while $\langle R\rangle(t)$ from classical trajectories (red or light gray solid line indicates the dynamics on $\epsilon(R,t)+\epsilon^{\rm QP}(R,t)$ and blue or dark gray solid line indicates the dynamics only on $\epsilon(R,t)$) are obtained as $\langle R\rangle(t)=\frac{1}{N_{\rm traj}}\sum^{N_{\rm traj}}_{I=1}R_I(t)$, where $N_{\rm traj}$ is the total number of trajectories and $R_I(t)$ is the distance at time $t$ of each trajectory. Comparison between the blue (or dark gray) and red (or light gray) lines clearly reveals the failure in the dynamics propagated only by the force from $\epsilon(R,t)$. In the higher-intensity case, the dynamics driven only by $\epsilon(R,t)$ succeeded in reproducing the dissociation ones, but their speed is lower than the exact result; the shape of the nuclear density (blue or dark gray) is more localized in the smaller $R$ region than the exact density. This is easily understood because the quantum potential has the effect of flattening the well of the ground-state potential, as shown above, and its absence makes the nuclear density likely to be trapped at the equilibrium position, leading to slower dissociation. In the lower-intensity case, failure in the dynamics propagated only by the force from $\epsilon(R,t)$ is more noticeable: the majority of the nuclear density shown as the blue (or dark gray) curve is trapped by the confining potential well at the equilibrium position, and only a very small fraction leaks out to the dissociation, even in the final snapshot ($t=18.6$ fs). The time evolution of $\langle R\rangle(t)$ in Fig. [\[fig:Fig4\]](#fig:Fig4){reference-type="ref" reference="fig:Fig4"} (d) clearly reflects this: the speed of the dissociation in the dynamics propagated only by $\epsilon(R,t)$ is much lower than that in the exact dynamics. This result is also supported by the above analysis, which shows that the quantum potential would have flattened the well and caused tunneling if it had been included in the force. From these results, we can conclude that the quantum potential plays a non-negligible role in reproducing the exact quantum nuclear dynamics in the strong-field processes studied here by propagating an ensemble of classical trajectories according to the Newton's equation ([\[eqn: Newton\]](#eqn: Newton){reference-type="ref" reference="eqn: Newton"}) derived from Bohmian mechanics in the exact nuclear TDSE ([\[eqn: exact nuclear eqn\]](#eqn: exact nuclear eqn){reference-type="ref" reference="eqn: exact nuclear eqn"}). This suggests that careful assessment of its effect is required when developing the MQC method based on the exact nuclear TDSE ([\[eqn: exact nuclear eqn\]](#eqn: exact nuclear eqn){reference-type="ref" reference="eqn: exact nuclear eqn"}), especially when we aim to develop the method for strong-field processes. # Conclusions In this paper, we have demonstrated that the propagation of multiple classical trajectories can reproduce quantum nuclear dynamics even in strong-field processes when they are propagated by Newton's equation with the force determined from the gradient of the exact TDPES plus the exact quantum potential, which is defined in Bohmian mechanics in the exact-factorization framework. We employed a one-dimensional H$_2^+$ model system subject to two different-intensity laser fields, which give rise to different types of dissociation dynamics, and showed that both processes can be simulated by multiple classical trajectory dynamics. We found that the exact quantum potential has a non-negligible effect on the classical dynamics: its force accelerates the classical nuclei to overcome the confining well of the ground-state potential-energy surface, and causes tunneling of the nuclear density in the lower-intensity case. The results here offer important knowledge for developing the MQC algorithm for coupled electron-nuclear dynamics based on the exact-factorization approach, especially when one wants to develop it for strong-field processes. Although it is a challenging task to develop suitable approximations to the quantum Hamilton-Jacobi equation ([\[eqn: QHJeqn\]](#eqn: QHJeqn){reference-type="ref" reference="eqn: QHJeqn"}) and make it solvable on-the-fly, our results are encouraging since they show that quantum nuclear dynamics can be simulated, in principle, by classical trajectories if the force is properly prepared, and the existence of such exact forces is proved. In this study, we explored the case where the gauge is fixed such that the vector potential in the exact nuclear TDSE vanishes. It is also desirable to formulate Bohmian mechanics in a more general gauge, i.e., where the vector potential exists and the gauge-dependent part of the TDPES vanishes. This would provide a more general basis for developing an MQC method that can yield the correct quantum nuclear-density dynamics in various non-adiabatic situations including strong-field processes. Acknowledgments: This study was supported by JSPS KAKENHI Grant No. 16K17768.	{'timestamp': '2016-09-13T02:03:59', 'yymm': '1607', 'arxiv_id': '1607.04854', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04854'}
# Introduction In robotics motion planning and control, one major challenge is to specify a high level task for a robot without specifying how to solve this task. For legged locomotion such a task include a goal position to reach or to manipulate an object without specifying gaits, contacts, balancing or other behaviors. Trajectory Optimization recently gained a lot of attention in robotics research since it promises to solve some of these problems. It could potentially solve complex motion planning tasks for robots with many degrees of freedom, leveraging the full dynamics of the system. Yet, there are two challenges of optimization optimization. Firstly, Trajectory Optimization problems are hard problems to solve, especially for robots with many degrees of freedom and for robots that make or break contact. Therefore, many approaches add heuristics or pre-specify contact points or sequences. However, this then defines again how the robot is supposed to solve the task, affecting optimality and generality. Secondly, Trajectory Optimization cannot be blindly applied to hardware but requires an accurate model as well as a good control and estimation framework. In this work, we are addressing both issues. In our Trajectory Optimization framework, we only specify high level tasks, allowing the solver to find the optimal solution to the problem, optimizing over the entire dynamics and automatically discovering the contact sequences and timings. Second, we also demonstrate how such a dynamic task can be generated online. This work does not present a general tracking controller suitable for all trajectories. However, we show the successful integration of our Trajectory Optimization with our estimation and control framework allowing for executing a selection of tasks even under disturbances. Trajectory Optimization tries to solve a general, time-varying, non-linear optimal control problem. There are various forms of defining and solving the control problem. An overview can be found in. With the increase of computational power, Trajectory Optimization can be applied to higher dimensional systems like legged robots. Thus, it has gained a lot of attention in recent years and impressive results have been demonstrated in simulation. One of the conceptually closest related work is. However, no gait discovery or very dynamic motions are shown and no hardware results are presented. Later work includes hardware results but the planning horizon is short, the motions are quasi-static and contact changes are slow. In Trajectory Optimization through contacts is demonstrated on hardware. While the results are promising, the approach is only tested on a low-dimensional, single leg platform and for very simple tasks. In general, work that demonstrates Trajectory Optimization through contacts applied to physical, legged systems is very rare. Automatically discovered gaits and dynamic motions demonstrated on quadruped hardware are presented in. The presented results are very convincing and they are also considering actuator dynamics. However, their approach differs in key areas to the presented work: Contact sequences are pre-specified and they optimize over control parameters for a fixed control structure rather than whole body trajectories and control inputs. Additionally, they are using black box optimization. However, the tracking controllers employed in both approaches are very similar. # Trajectory Optimization ## Optimal Control Problem In this work, we consider a general non-linear system of the following form \[\begin{aligned} \dot{\myvec{x}}(t) = f(\myvec{x}(t),\myvec{u}(t)) \label{eq:system_model} \end{aligned}\] where $x(t)$ and $u(t)$ denote state and input trajectories respectively. In Trajectory Optimization, we indirectly optimize these trajectories by altering a linear, time-varying feedback and feedforward controller of the form \[\begin{aligned} \myvec{u}(\myvec{x},t) = \myvec{u}_{ff}(t) + \mathcal{K}(t) \myvec{x}(t) \label{control_law} \end{aligned}\] where $\mathcal{K}(t)$ is a time-varying control gain matrix and $u_{ff}(t)$ a time-varying feedforward control action. Our Trajectory Optimization approach then tries to solve a finite-horizon optimal control problem which minimizes a given cost function of the following form \[\begin{aligned} J(\myvec{x},\myvec{u}) = h\left( \myvec{x}(t_f) \right) + \int_{t=0}^{t_f}l \left( \myvec{x}(t),\myvec{u}(t) \right) dt \label{eq:cost_function} \end{aligned}\] ## Sequential Linear Quadratic Control Sequential Linear Quadratic Control is an iterative optimal-control algorithm. SLQ first rolls out the system dynamics. Then, the non-linear system dynamics are linearized around the trajectory and a quadratic approximation of the cost function is computed. The resulting Linear-Quadratic Optimal Control problem is solved backwards. Since the solution to the Linear-Quadratic Optimal Control problem can be computed in closed form with Ricatti-like equations, SLQ is very efficient. Algorithm [\[alg:slq\]](#alg:slq){reference-type="ref" reference="alg:slq"} summarizes the algorithm. SLQ computes both, a feedforward control action as well as a time-varying linear-quadratic regulator that stabilizes the trajectory. To avoid discontinuities during contact changes, the models are non-linear towards zero ground penetration, i.e. for $p_n < \alpha$, where $\alpha_c$ is a smoothing coefficient. Typical values for the contact model are given in Table [1](#tbl:contact_parameters){reference-type="ref" reference="tbl:contact_parameters"}. The parameters are task dependent to trade-off physical correctness and convergence. Friction and friction limits are considered via friction cones. This allows the Trajectory Optimization algorithm to reason about possible slippage and contact force saturation. In this work, we assume that static and dynamic friction coefficients are the same. Therefore, the friction cone can be expressed as a contact force saturation \[\myvec{F}_t = max(\|\|\myvec{F}_t\|\|, \mu F_n) \myvec{n}_s\] where $F_t$ is the contact force parallel to the surface, $F_n$ is the force normal to the surface, $\mu$ is the friction coefficient and $\myvec{n}_s$ is the surface normal. # State Estimation and Tracking Control Our TO approach is fully integrated into our estimation and control framework, illustrated in Figure [\[fig:control_overview\]](#fig:control_overview){reference-type="ref" reference="fig:control_overview"}. Since SLQ assumes full state-feedback, joint position and velocity measurements are fused with IMU data to obtain a base estimate. Furthermore, since SLQ reasons about contacts with the environment, a ground estimator is added which estimates elevation and inclination of the ground. SLQ then optimizes a trajectory which is then fed to a task-space base controller and the joint controller. These control loops are closed at a rate of 200 Hz. The corrective output of both controllers is then added to the feedforward control signal as optimized by SLQ and sent to the robot. ## State Estimation SLQ assumes full state-feedback, i.e. all states are assumed to be known. While we can directly measure joint positions and velocities with the encoders on HyQ, the base state cannot be directly measured. Therefore, we use to estimate the base's pose and twist. The contact state is estimated by mapping joint torques to estimated ground reaction forces. If the normal component of the ground reaction force estimate is above a fixed threshold we assume the leg being in contact. In order to reason about the contact forces $\lambda$, we also have to know the elevation and contact surface normal at the contact points. In order for the algorithm to work online we cannot assume a perfectly leveled ground. Instead, we assume a ground plane which is fitted to the last contact point of each feet. This implementation allows for incorporating a ground map in the future. ## Tracking Controller SLQ does not only optimize feed-forward control action but also a feedback controller. While we have successfully applied theses feedback gains to robotic hardware, we are not using the optimized feedback gains in this work for two reasons. First, the optimized feedback gains correspond to a time-varying, linear-qaudratic regulator (TVLQR). Therefore, the gains significantly depend on the linearization point. Especially due to the non-linearity of the contacts, the robot's state will drift from the linearization point during trajectory execution. Secondly, due to the usage of a single cost function, any regularization in the cost function will also effect the feedback gains which is undesirable. One solution is to recompute the gains online. However, we have found that the following tracking controller shows great performance at reduced complexity. Assuming a known, fixed stance configuration with no slippage and at least three stance legs, the stance legs as well as the torso can be described by a 6 degrees of freedom state. This means, the base state, consisting of pose and twist, is coupled with joint positions and velocities of the stance legs. Therefore, controlling either set of states, the joint states of the stance legs or the base state, will subsequently also track the other set. A base controller allows us to directly track the base state and tune feedback gains on these states intuitively. Yet, for swing legs, we still require a joint controller. Hence, to get best of both worlds, we are using a combination of a task space base controller and a joint space controller. The joint space controller tracks the desired position and desired velocity of all joints independently. Additionally, the base controller regulates the base state with a PD controller. This task space control can be imagined as virtual springs and dampers attached to the robot's trunk on one side and the optimized body trajectory on the other. \[\myvec{F}_{cog} = \myvec{P}_x (\myvec{x}_{cog}^-\myvec{x}_{cog}) + \myvec{D}_x (\dot{\myvec{x}}_{cog}^-\dot{\myvec{x}}_{cog}) \label{eq:virtual_model}\] The desired body wrench $\mathbf{F}_{cog}$ is applied to the robot by converting it to forces in the contact points $\boldsymbol{\lambda}_{c}$ and then mapping these to the joint torques through $\boldsymbol{\tau}_{fb} = \mathbf{J}_{c}^T \boldsymbol{\lambda}_{c}$. These torques are then added to the feedforward control action obtained from SLQ. This feedforward control action already includes torques that counteract gravity and thus, no gravity compensation term is added in Equation [\[eq:virtual_model\]](#eq:virtual_model){reference-type="ref" reference="eq:virtual_model"}. # Experiments To evaluate the approach, we first show, how a galloping and a trotting gait with optimized contact switching sequences and timings can be obtained in simulation. Furthermore, we demonstrate a manipulation motion that includes contact switches and more complex base motions. To verify the applicability of our approach, we demonstrate hardware experiments for the last task. Finally, we show a highly under-actuated task where HyQ walks on its hind legs like a humanoid. For each task, the cost function is shown as a color code below the heading. The colors indicate the individual, relative weightings of the diagonal entries of each weighting matrix, ranging from lowest (green) to highest (red) on a logarithmic scale. No color means that the according value is zero and all off-diagonal elements are zero as well. The input cost matrix $\myvec{R}$ is set to identity in all experiments. All tasks are initialized with a simple PD feedback controller on all joints and no feedforward control. This leads to HyQ maintaining its initial state over the entire length of each trajectory. ## Simulation ### Galloping The first gait we demonstrate is galloping. The galloping gait is a direct outcome of the optimization with a very simple cost function. The cost function consists of final costs on the base position and leg positions. Additionally, we add some regularization on the base and leg motion to prevent excessive motions of the body or the limbs. However, the cost function does not include any terms related to contact sequences or timings and no priors on a galloping gait. Also, we are not using any intermediate cost terms here. HyQ starts in its default stance and is supposed to reach its final position 2 m in front. The chosen time horizon is 3 seconds. As for all tasks, the initial controller is a simple joint PD controller which results in HyQ staying in place and maintaining its intial configuration. As the results in Figure [\[fig:gallop_base\]](#fig:gallop_base){reference-type="ref" reference="fig:gallop_base"} illustrates, we obtain a gallop motion with 9 steps which includes acceleration and deceleration. Finally, the robot reaches the desired position at $x=2.0$ m. As expected, we see significant pitch motion of the upper body. From Figure [\[fig:gallop_torques\]](#fig:gallop_torques){reference-type="ref" reference="fig:gallop_torques"} we can tell that the hind legs, especially the HFE joints, are used for acceleration. While our findings are not directly comparable to biology, we can see similar effects: In fast running animals, like leopards and horses, the rear legs also bear a larger load during running than the front legs, which explains the different sizing of legs and muscles between rear and front. Snapshots of the gallop are shown in Figure [\[fig:strips\]](#fig:strips){reference-type="ref" reference="fig:strips"} and the full motion is shown in the video[^1]. ### Trotting One hypothesis for the previous task resulting in a galloping behavior is the short time horizon and no penalty on the orientation. Now if we increase the time horizon to 8 seconds and penalyze the base motion, i.e. we give HyQ more time and encourage smoother base motions, we see that the optimization starts to prefer a trotting motion instead of the galloping. Interestingly enough, the weightings allow to blend between galloping and trotting. As a result, HyQ first gallops and then smoothly transitions to a trotting gait. Again, neither the contact sequence, nor the gait or the transition is given. The parameters influencing the gait are the diagonal elements of $\myvec{Q}$ and $\myvec{R}$ as in Equation [\[eq:cost_function_experiments\]](#eq:cost_function_experiments){reference-type="ref" reference="eq:cost_function_experiments"}. By increasing the base motion penalty, the trajectory results in a pure trot. The trot consists of 4 steps per diagonal leg pair with almost constant stride length. By setting the desired position to the side instead of the front, the resulting trajectory is a sidestepping motion. If only a desired yaw angle is set, the robot turns on the spot. Both trajectories are trotting variants where the diagonal leg pairs are moved together. Both motions are illustrated in Figure [\[fig:strips\]](#fig:strips){reference-type="ref" reference="fig:strips"} and included in the video^[\[video\]](#video){reference-type="ref" reference="video"}^. ### Squat Jump Next, we test if our Trajectory Optimization approach can leverage and reason about the dynamics of our system. Thus, we define a task with intermediate waypoint states that are not statically reachable. First, we add an intermediate waypoint cost term for the base position and orientation. For this waypoint, we choose a strong weighting for the base height and set the desired value of the base height to $z=0.8 m$. Furthermore, the waypoint costs contain low weighting on the base orientation to keep the base level. By adding a cost on the deviation from default joint position, we ensure HyQ cannot just try to straighten its legs to try to reach the base pose, but that it has to jump. After running our optimization, we obtain a near symmetric squat jump. The overall optimization spans the entire motion, e.g. preparation for lift-off from default pose, the lift-off itself, going to default pose in the air, landing and returning to the default pose. The apex waypoint is localized in time but contact switches and timings are optimized. Also, we are not directly specifying a squat jump. Therefore, simultaneous lift-off and landing of the legs is an optimization outcome rather than pre-specified. Figure [\[fig:squat_base\]](#fig:squat_base){reference-type="ref" reference="fig:squat_base"} shows the optimized squat jump motion. We can see that unnecessary motions such as rotations or position changes in x-or y-direction are avoided. Also, the desired apex height of 0.8 m, despite being a soft constraint, is reached with sub-millimeter accuracy (measured 0.8004 m). The small pitch velocity results from an off-center mass and the optimization goal of low torques which promotes equal distribution of torques between legs as seen in Figure [\[fig:squat_torques\]](#fig:squat_torques){reference-type="ref" reference="fig:squat_torques"}. The same plot also shows that the largest torques appear in the KFE and HFE joints. Furthermore, the lower leg of HyQ weighs less than 1 kg compared to a total mass of about 80 kg. Therefore, the mass that the lower leg drives changes rapidly during contact changes occurring at around $t=0.55 s$ and $t=0.9 s$. Hence, the torques also change quickly. While the knee joint is still moved in the air to take the default angle and prepare for landing, the required torque is minimal. Again, both motions are shown as a snapshot series in Figure [\[fig:strips\]](#fig:strips){reference-type="ref" reference="fig:strips"} as well as in the video^[\[video\]](#video){reference-type="ref" reference="video"}^. ### Rearing For the next task, we are using a similar cost function as the squat jump. Instead of penalizing the deviation from a neutral base pose, we penalize deviations from a 30$^\circ$ pitched base orientation and we lower the desired apex height to 0.7 m. Yet again, we do not specify contact/stance configurations. Therefore, the optimization algorithm is free to optimize the trajectory. The final trajectory represents a rearing motion, where HyQ lifts off with the front legs, reaches the apex position and finally returns to full contact as well as its default pose. A screenshot of the apex position is shown in Figure [\[fig:strips\]](#fig:strips){reference-type="ref" reference="fig:strips"}. The full motion is shown in the accompanying video^[\[video\]](#video){reference-type="ref" reference="video"}^. ### Diagonal Balance In order to demonstrate that the SLQ can also find statically unstable trajectories, we are demonstrating a diagonal balance task. Here, we use a waypoint term in our cost function again. This term penalizes the orientation and height of the base. Furthermore, it encourages the robot to pull up its legs by bending HFE and KFE of the left front and right hind leg. The final trajectory shows the expected balancing behavior. Again a screenshot at apex is shown in Figure [\[fig:strips\]](#fig:strips){reference-type="ref" reference="fig:strips"} and the video^[\[video\]](#video){reference-type="ref" reference="video"}^ shows the full motion. Interestingly, while we are using a single intermediate term with a single time point and absolutely symmetric costs, the lift-off and touch-down of the swing legs is not synchronous but the front left leg lifts-off later and touches down earlier. This asymmetry most likely stems from the asymmetric location of the CoG and assymmetric inertia of the robot's main body. ### Humanoid Walk As a final test, we evaluate if our algorithm is capable of optimizing a behavior that allows HyQ to stand on its hind legs, similar to a humanoid robot, and then go to a target point in front. While standing on its hind legs exceeds the capabilities of the hardware, such a motion is physically possible. Compared to a humanoid with ankles, HyQ has point feet. This increases the level of difficulty since the absence of ankles makes HyQ highly under-actuated. This means that during a standing or walking task no torques can be applied to the ground but balance can only be maintained by moving the links. In this task, we again initialize HyQ in its default stance, requiring it to stand up in place first and then reach a goal point where it should stabilize for a second. The first waypoint has a very wide spread in time and penalizes deviations in base orientation and height. This cost ensures that HyQ stays upright during the entire task after getting up. We do not add this cost to the general intermediate cost, as it would encourage HyQ to stand up as fast as possible, rather than getting up in a controlled, efficient manner. The stand up motion is encouraged by the second waypoint cost penalizing base orientation, height and changes in forward position. We add a third waypoint one second before the end of the time horizon, defining the target pose and orientation. While the last waypoint and the final cost specify the same base pose, we separate them to demonstrate that HyQ can stay upright and stabilize in place for longer times without showing signs of falling. The resultant motion shows interesting, non-trivial behavior. Before getting up HyQ pulls its hind left leg in, moving the contact point more closely below the center of gravity. Also, it uses the front left leg ("left arm") to get up, resulting in a very natural, coordinated, asymmetric motion. After getting into a two-leg standing phase, a forward motion is initiated by a short symmetric hopping but quickly changes to a coordinated stepping/walking pattern. While we penalize joint velocity, the front limbs are moved to support balance during the motion until they are retracted to a target pose at the end of the trajectory. The entire motion underlines once more the capabilities of the approach. Classic controllers and motion planners might not find such a complex motion. Additionally, by allowing optimization over contact points and sequences, we allow the Trajectory Optimization to find the best solution for the task. While a rearing motion, as demonstrated above, is feasible and can be found using our problem formulation, the optimizer prefers a less dynamic, coordinated motion here. ## Hardware To verify that the optimized trajectories can be applied to hardware, we run an example task on hardware. This task is a rough manipulation task where HyQ pushes over a pallet with its front left leg. The trajectory requires a sequence of motions such as shifting its CoG in the support polygon of the three stance legs, lifting off of the front left leg, executing the push motion, putting the foot back down and shifting the CoG back. While in classic approaches these motions would possibly all handcoded, they result from a single cost function with only one intermediate cost function term for the front left leg in our Trajectory Optimization approach. While this trajectory works perfectly fine in simulation, we have to slightly alter it for the hardware. Since our TO approach is deterministic and we penalize control input, the algorithm tries to minimize the shift of the CoG, leading to a very risky trajectory. Thus, small model inaccuracies or disturbances make the robot loose its balance. This is a general problem of all deterministic TO approaches and can only be fundamentally solved by using risk aware or stochastic approaches. While there is a risk-aware SLQ variant, here we apply a work-around by simply encouraging a larger CoG shift in our cost function. For visual purposes only, we add additional intermediate waypoints for the front left leg only, leading to a slower, more appealing push motion. While we could easily add the push contact to our optimization, we leave it unmodelled on purpose such that it becomes a disturbance to our controller, underlining its robustness. The pallet is a 1200 by 800 mm Euro pallet that weighs approximately 24 kg and thus requires a force of about 70 N at the point of contact of the leg to be pushed over. To demonstrate that the algorithm is being run online we approach the pallet with a separate walking controller. Thus, the robot is faced with different initial conditions every time. Figure [\[fig:strips\]](#fig:strips){reference-type="ref" reference="fig:strips"} shows a sequence of images of the optimized motion while Figure [\[fig:hw_sequence\]](#fig:hw_sequence){reference-type="ref" reference="fig:hw_sequence"} shows a sequence of images taken during execution. The video^[\[video\]](#video){reference-type="ref" reference="video"}^ shows the entire motion as well as examples from our physics simulator, where initial conditions differ even more significantly. As can be seen in all examples, the motion is dynamic and lift-off and touch-down events of the front left leg partially overlap in time with the CoG shift. This underlines that static stability is not required in this task but the algorithm finds a dynamically stable trajectory. The base pose/twist tracking for this task during hardware experiments is shown in Figure [\[fig:hw_base_tracking\]](#fig:hw_base_tracking){reference-type="ref" reference="fig:hw_base_tracking"} and the joint position/velocity tracking in Figure [\[fig:hw_joint_tracking\]](#fig:hw_joint_tracking){reference-type="ref" reference="fig:hw_joint_tracking"}. We can see that the joint tracking is better than base tracking due to a more aggressive joint controller. However, the base pose and twist do not significantly deviate from their planned trajectories either. Figure [\[fig:hw_torques\]](#fig:hw_torques){reference-type="ref" reference="fig:hw_torques"} shows the combination of feedforward torques and task space control input. In the front left leg, we can nicely see the lift off and touch down events in between which joint torques are almost zero. This happens since this leg does not further contribute to sustaining the weight of the body. Subsequently, the other legs have to bear a higher load. Especially in the neighboring legs, i.e. the right front and left hind leg, we see an increase in torques. Yet, this increase is not very significant since the robot has already shifted its CoG towards the support polygon of the stance legs. ## Runtime and Convergence When running Trajectory Optimization online, runtime and convergence become a major concern since these measures define how long the robot "thinks" before executing a task. Especially in a dynamic environment or in presence of drifting estimates, we want to keep the optimization procedures short to be able to react to a given situation quickly. Even better than pre-optimizing a trajectory before execution, we reoptimize and adjust them during execution, forming a model-predictive control scheme. In this section, we will look at both the number of iterations for each task as well as the runtime of each iteration. This gives us an indicator of the complexity of a task and tells us how far we are from running our approach in an MPC fashion. As a first test, we take a look at convergence rates across tasks. To obtain comparable results between different tasks, we initialize all tasks with a standing controller in form of a pure joint position controller and normalize the costs with the initial cost. The results of this test are shown in Figure [\[fig:convergence\]](#fig:convergence){reference-type="ref" reference="fig:convergence"}. The curves suggest that there is a relation between the time horizon and complexity of a task and the corresponding convergence rate. The trotting behavior is one of the most complex behaviors and also has the longest time horizon. The rearing task is relatively simple and well guided by the waypoint costs which is a possible explanation of the fast convergence rate. Most tasks converge within 10 and 40 iterations, which given the runtime per iteration, shown in Figure [\[fig:runtime\]](#fig:runtime){reference-type="ref" reference="fig:runtime"}, usually means an overall optimization time of less than 1 minute. When it comes to runtime, SLQ has a major advantage over many other Trajectory Optimization approaches: It scales linear with discretization steps and thus with time horizon. Each SLQ step is fixed runtime which means the fewer steps we use, the faster the optimizer runs. Therefore, one would ideally increase step size but also shorten the trajectory length. Yet, there is a limit to both. Figure [\[fig:runtime\]](#fig:runtime){reference-type="ref" reference="fig:runtime"} underlines the linear relation between the number of trajectory points (or steps) and the runtime per iteration. The timings are measured on a standard quadcore laptop and averaged over multiple iterations. To achieve these runtimes we use a custom multi-threaded solver. Additionally, we use highly optimized, auto-generated code for the computation of forward dynamics, used for rollouts/line-search and linearization. This code is generated using RobCoGen, a code-generator framework for rigid body kinematics and dynamics. # Discussion, Conclusion and Outlook In this paper, we have presented a fully dynamic, whole-body Trajectory Optimization framework that is able to create motions which involve multiple contact changes. The approach does not require any priors or initial guesses on contact points, sequences or timings. We demonstrate the capabilities on various tasks including squat jumps, rearing, balancing and rough manipulation. Furthermore, our approach is able to discover gaits such as galloping, trotting and two legged walking. First hardware results show that the optimized trajectories can be transferred to hardware. While some motions might not be directly applicable to hardware, they can serve as an initialization for other algorithms and they could potentially provide insight into optimal gait parameters such as stepping frequencies or stride lengths. As with most optimal control approaches, a certain amount of cost function tuning is required. However, we use only one cost function per task and tried to stay away from complex terms that tune for numerical properties. Instead, we tuned for functionality and behavior and provide insight into the weightings used. While tuning cost functions is a manual procedure, most of the obtained motion would require a complex combination of different planners and controllers if implemented with state-of-the-art approaches, e.g.. In contrast to these methods, whole body Trajectory Optimization is versatile and the structure of the approach is not task dependent. Also, the optimized trajectories leverage the full dynamics of the system and do not rely on heuristics or simplified models. Using our formulation and solvers, the optimized trajectory is usually available in less than one minute, even for complex tasks and despite not pre-specifying contacts. This compares favorable to reported results in literature where solving time can be several minutes or even hours. However, there are also some short comings of the approach. The solution space of the approach is huge, e.g. we can apply our approach to a broad variety of tasks despite limiting ourselves to a single cost function structure. While this generality is a strength of the approach, it also requires the user to "choose" a solution by modifying cost function weights or adding additional term. While this is a more or less intuitive approach, one would wish to further reduce the number of open parameters by applying some form of meta algorithm. Another drawback is that the optimized motions might not be very robust to model inaccuracies, disturbances or noise. Thus, alterations of the cost functions might be required to successfully execute these motions on hardware. This is a known issue of deterministic Trajectory Optimization approaches and can be tackled only by considering the stochasticity of the problem. Lastly, SLQ cannot handle state constraints, which can be a limiting factor for complex Trajectory Optimization problems. We believe that SLQ can be especially useful when applied as a shorter horizon MPC controller where its fast run-times are leveraged and complex state constraints are taken care of by a higher level planner/optimizer. While the run-time of SLQ seems far off from an MPC scenario, warm starting is an efficient measure to make the algorithm converge in only a few iterations. Also, run-time can be further reduced by limiting the time horizon. Thus, as a next step, we are planning to apply our approach in MPC fashion to HyQ, extending our work in. Hopefully, SLQ-MPC will be able to act as a general stabilization and tracking controller as well as short horizon re-planner, allowing us to execute most tasks demonstrated in this work on hardware. One important step in this direction will be to include input constraints.	{'timestamp': '2016-07-18T02:07:44', 'yymm': '1607', 'arxiv_id': '1607.04537', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04537'}
# Introduction The combinatorial structure treated in this paper is a $2 \rightarrow 1$ directed hypergraph defined as follows. For simplicity from this point on we will always refer to $2 \rightarrow 1$ directed hypergraphs as just graphs or sometimes as $(2 \rightarrow 1)$-graphs when needed to avoid confusion. This structure comes up as a particular instance of the model used to represent definite Horn formulas in the study of propositional logic and knowledge representation. Some combinatorial properties of this model have been recently studied by Langlois, Mubayi, Sloan, and Gy. Turán in and. In particular, they looked at the extremal numbers for a couple of different small graphs. Before we can discuss their results we will need the following definitions. These are often called Turán-type extremal problems after Paul Turán due to his important early results and conjectures concerning forbidden complete $r$-graphs. Turán problems for uniform hypergraphs make up a large and well-known area of research in combinatorics, and the questions are often surprisingly difficult. Extremal problems like this have also been considered for directed graphs and multigraphs (with bounded multiplicity) in and and for the more general directed multi-hypergraphs in. In, Brown and Harary determined the extremal numbers for several types of specific directed graphs. In, Brown, Erdős, and Simonovits determined the general structure of extremal sequences for every forbidden family of digraphs analogous to the Turán graphs for simple graphs. The model of directed hypergraphs studied in have $r$-uniform edges such that the vertices of each edge is given a linear ordering. However, there are many other ways that one could conceivably define a uniform directed hypergraph. The graph theoretic properties of a more general definition of a nonuniform directed hypergraph were studied by Gallo, Longo, Pallottino, and Nguyen in. There a directed hyperedge was defined to be some subset of vertices with a partition into head vertices and tail vertices. Recently in, this author tried to capture many of these possible definitions for "directed hypergraph\" into one umbrella class of relational structures called generalized directed hypergraphs. The structures in this class include the uniform and simple versions of undirected hypergraphs, the totally directed hypergraphs studied in, the directed hypergraphs studied in, and the $2 \rightarrow 1$ model studied here and in. In, they study the extremal numbers for two different graphs with two edges each. They refer to these two graphs as the 4-resolvent and the 3-resolvent configurations after their relevance in propositional logic. Here, we will denote these graphs as $R_4$ and $R_3$ respectively and define them formally as \[V(R_4) = \{a,b,c,d,e\} \text{ and } E(R_4) = \{ab \rightarrow c, cd \rightarrow e\}\] and \[V(R_3) = \{a,b,c,d\} \text{ and } E(R_3) = \{ab \rightarrow c, bc \rightarrow d\}.\] In the authors determined $\text{ex}(n,R_4)$ for sufficiently large $n$, and in they determined a sequence of numbers asymptotically equivalent to the sequence of numbers $\text{ex}(n,R_3)$ as $n$ increases to infinity. In these papers, the authors discuss a third graph with two edges which they call an Escher configuration because it calls to mind the Escher piece where two hands draw each other. This graph is on four vertices, $\{a,b,c,d\}$ and has edge set $\{ab \rightarrow c,cd \rightarrow b\}$. We will denote it by $E$. These three graphs actually turn out to be the only three nondegenerate graphs with exactly two edges on more than three vertices. An immediate consequence of a result shown in is that the extremal numbers for a graph $H$ are cubic in $n$ if and only if $H$ is not degenerate. In our model of directed hypergraph, there are nine different graphs with exactly two edges. Of these, four are not degenerate. One of these is the graph on three vertices with exactly two edges, $V=\{a,b,c\}$ and $E=\{ab \rightarrow c, ac \rightarrow b\}$. It is trivial to see that both the standard and oriented extremal numbers for this graph are ${n \choose 3}$. The other three nondegenerate graphs are $R_4$, $R_3$, and $E$. We will determine both the standard and oriented extremal numbers for each of these graphs in Sections 2, 3, and 4 respectively. The extremal numbers for the five degenerate cases are determined in a companion paper. The proofs that follow rely heavily on the concept of a link graph. For undirected $r$-graphs, the link graph of a vertex is the $(r-1)$-graph induced on the remaining vertices such that each $(r-1)$-set is an $(r-1)$-edge if and only if that set together with the specified vertex makes an $r$-edge in the original $r$-graph. In the directed hypergraph model here, there are a few ways we could define the link graph of a vertex. We will need the following three. # The 4-resolvent graph $R_4$ In, the authors gave a simple construction for an $R_4$-free graph. Partition the vertices into sets $T$ and $K$ and take all possible edges with tail sets in $T$ and head vertex in $K$. When there are $n$ vertices, this construction gives ${t \choose 2}(n-t)$ edges where $t=\|T\|$. This is optimized when $t = \left\lceil \frac{2n}{3} \right\rceil$. In, they showed that this number of edges is maximum for $R_4$-free graphs for sufficiently large $n$ and that the construction is the unique extremal $R_4$-free graph. We now give an alternate shorter proof that $\left\lfloor \frac{n}{3} \right\rfloor{\left\lceil \frac{2n}{3} \right\rceil \choose 2}$ is an upper bound on the extremal number for $R_4$ for sufficiently large $n$ in both the standard and oriented versions of the problem. The proof also establishes the uniqueness of the construction. # The 3-resolvent graph $R_3$ In, the authors gave a simple construction for an $R_3$-free graph. Partition the vertices into sets $A$ and $B$ and take all possible edges with a tail set in $A$ and head vertex in $B$ plus all possible edges with a tail set in $B$ and a head in $A$. When there are $n$ vertices, this construction gives $(n-a){a \choose 2} + a{n-a \choose 2}$ edges where $a=\|A\|$. This is optimized when $a = \left\lceil \frac{n}{2} \right\rceil$. The authors showed that this number of edges is asymptotically optimal for $R_3$-free graphs. We show that in both the standard and the oriented versions of this problem that this construction is in fact the best that we can do. We will start with the oriented case since it is less technical. ## The oriented version ## The standard version # The Escher graph $E$ In this section, we will prove the following result on the maximum number of edges of a $E$-free. But first we will prove the easier oriented version of the problem. This result will be needed to prove Theorem [\[mexG\]](#mexG){reference-type="ref" reference="mexG"}. ## The oriented version We now shift our attention to the the standard version of the problem where a triple of vertices can have more than one edge. Here, both of the lower bound constructions are similar to the unique extremal construction in the oriented version. ## Two lower bound constructions for $\text{ex}(n,E)$ The first construction is the same as the extremal construction in the oriented case but with two additional edges placed on the "smallest\" triple. That is, let $H_1 = ([n],E_1)$ where \[E_1 = \left \{ ab \rightarrow c: a <b<c \right \} \cup \{13 \rightarrow 2, 23 \rightarrow 1\}.\] See Figure [\[mexG1\]](#mexG1){reference-type="ref" reference="mexG1"}. Moreover, it is important to note that if an $E$-free graph with ${n \choose 3} + 2$ edges has at least one edge on every vertex triple, then it must be isomorphic to $H_1$. This is because we can remove two edges to get an $E$-free subgraph where each triple has exactly one edge. Therefore, this must be the unique extremal construction established in Theorem [\[exG\]](#exG){reference-type="ref" reference="exG"}. The only way to add two edges to this construction and avoid creating an Escher graph is to add the additional edges to the smallest triple under the ordering. The second construction is also based on the oriented extremal construction. Let $H_2=([n],E_2)$ where \[E_2 = \left( E_1 \setminus \{23 \rightarrow 4, 23 \rightarrow 1\} \right) \cup \{14 \rightarrow 2, 14 \rightarrow 3\}.\] See Figure [\[mexG2\]](#mexG2){reference-type="ref" reference="mexG2"}. For the rest of this section we will show that any $E$-free graph is either isomorphic to one of these two constructions or has fewer than ${n \choose 3} + 2$ edges. Roughly speaking, the strategy is to take any $E$-free graph and show that we can add and remove edges to it so that we preserve $E$-freeness, remove most multiple edges from triples that had more than one, and never decrease the overall number of edges. ## Add and Remove Edges Let $H$ be an $E$-free graph and represent its vertices as the disjoint union of three sets: \[V(H) = D \cup R \cup T\] where $D$ (for 'Done') is the set of all vertices that have complete graphs on three or more vertices as tail link graphs, $R$ (for 'Ready to change') is the set of vertices not in $D$ that have at least three edges in their tail link graphs, and $T$ is the set of all other vertices (those with 'Two or fewer edges in their tail link graphs'). The plan is now to remove and add edges in order make a new graph $H'$ which is also $E$-free, has at least as many edges as $H$, and whose vertices make a disjoint union, \[V(H') = D' \cup T'\] where $D'$ and $T'$ are defined exactly the same as $D$ and $T$ except in terms of the vertices of $H'$. That is, for each vertex $x \in R$, we will add all possible edges to complete $T_x$. This moves $x$ from $R$ to $D$. The edges removed will be all those that pointed from $x$ to a vertex that points to $x$. This will destroy triples with more than one edge as we go. The following observation will ensure that this procedure only ever moves vertices from $R$ to $D$, from $R$ to $T$, from $R$ to $R$, and from $T$ to $T$. Since each step moves one vertex from $R$ to $D$ and ends when $R$ is empty, then the procedure is finite. Here is the observation: Now, let us make the procedure slightly more formal: While there exist vertices in $R$, pick one, $x \in R$, and for each pair $a,b \in V(T_x)$, add the edge $ab \rightarrow x$ to $E(H)$ if it is not already an edge. Then, for each $a \in V(T_x)$, remove all edges of $E(H)$ of the form $xs \rightarrow a$ for any third vertex $s$. Since there were at least three edges in $T_x$, then the added edges will move $x$ from $R$ to $D$. The removed edges, if any, will only affect vertices in $R$ or in $T$ since if $xs$ is removed from $T_a$, then this implies that $a \in T_x$ and that $x \in T_a$ and so both had degree one in the other's tail link graph. Hence, $a \not \in D$. Moreover, an affected vertex in $R$ will either stay in $R$ or move to $T$ while an affected vertex in $T$ will stay in $T$ since it is only losing edges from its tail link graph. Moreover, at the end of this process $D'$ will contain no triple of vertices with more than one edge. Therefore, the only such triples of vertices of $H'$ will be entirely in $T'$ or will consist of vertices from both $T'$ and $D'$. We will show later that there cannot be too many of these triples. First, we need to show that after each step of this procedure, no Escher graph is created and at least as many edges are added to the graph as removed. ## No copy of $E$ is created and the number of edges can only increase Fix a particular vertex $x \in R$ to move to $D$. Add and remove all of the designated edges. Suppose that we have created an Escher graph. Since the only edges added point to $x$, then the configuration must be of the form, $ab \rightarrow x, xc \rightarrow a$ for some distinct vertices, $a$, $b$, and $c$. Therefore, $a \in V(T_x)$ and so $xc \rightarrow a$ would have been removed in the process. Now we will show that at least as many edges have been added to $H$ as removed by induction on the number of independent edges in $T_x$. Start by assuming there are 0 independent edges in $T_x$ and assume that there are $k$ vertices in $T_x$ that have degree one. Then at most $k$ edges will be removed. If $k=0$, then no edges are removed and there is a strict increase in the number of edges. If $k=1$, then let $y_1$ be the degree one vertex and let $y_2$ be the vertex it is incident to. Since $d_x(y_2) \neq 1$ and $d_x(y_2) \geq 1$, then $d_x(y_2) \geq 2$. So there exists a third vertex, $y_3$, and similarly, $d_x(y_3) \geq 2$ but $y_2$ is not adjacent to $y_1$. Hence, there exists a fourth vertex, $y_4$. So at most one edge is removed and at least two edges are added, $y_1y_3 \rightarrow x$ and $y_1y_4 \rightarrow x$. Therefore, there is a strict increase in the number of edges. If $k = 2$, then the fact that $T_x$ has at least three edges means that there must be at least two additional vertices in $T_x$. Hence, at most two edges are removed but at least three are added. If $k \geq 3$, then at most $k$ are removed but ${k \choose 2}$ are added which nets \[{k \choose 2}-k = \frac{k(k-3)}{2} \geq 0\] edges added. Now, for the induction step, assume that $T_x$ has $m>0$ independent edges and that the process on a $T_x$ with $m-1$ independent edges adds just as many edges as it removes. Let $yz$ be an independent edge in $T_x$ and let $A$ be the set of vertices of $T_x$ that are not $y$ or $z$. Since $T_x$ has at least three edges, then $A$ contains at least three vertices. Therefore, the number of added edges is at least 6 between $A$ and $\{y,z\}$. The number of edges removed from $T_y$ and $T_z$ together is at most 2. By assumption, the number of edges removed from the other tail link graphs of vertices in $A$ is offset by the number of edges added inside $A$. Therefore, there is a strict increase in the number of edges. To summarize, we have shown that $H'$ is an $E$-free graph such that \[\|E(H)\| \leq \|E(H')\|\] and \[V(H') = D' \cup T'\] such that any triple of vertices of $H'$ with more than one edge must intersect the set $T'$. We will now consider what is happening in $T'$ by cases. ## Case 1: $\|T'\| \geq 5$ Let $T' = \{x_1,x_2,\ldots,x_t\}$ for $t \geq 5$. For each $x_i$ remove all edges of $H'$ that have $x_i$ as a head. By the definition of $T'$ this will remove at most $2t$ edges from $H'$. Next, add all edges to $T'$ that follow the index ordering. That is, for each triple $\{x_i,x_j,x_k\}$ add the edge that points to the largest index, $x_ix_j \rightarrow x_k$ where $i<j<k$. This will add ${t \choose 3}$ edges. The new graph has \[{t \choose 3}-2t \geq 0\] more edges than $H'$. Moreover, it is $E$-free and oriented. Therefore, $\|E(H)\| < {n \choose 3}$. ## Case 2: $\|T'\| \leq 4$ and there exists an $x \in T'$ such that $T_x$ is two independent edges Assume that some $x \in T'$ has a tail link graph $T_x$ such that $ab, cd \in E(T_x)$ for four distinct vertices, $\{a,b,c,d\}$. If \[d_a(x) = d_b(x) = d_c(x) = d_d(x) = 1,\] then $a,b,c,d,x \in T'$, a contradiction of the assumption that $\|T'\| \leq 4$. Therefore, we can add the edges \[ac \rightarrow x, ad \rightarrow x, bc \rightarrow x, bd \rightarrow x\] and remove any edges that point to a vertex from $\{a,b,c,d\}$ with $x$ in the tail set. Because $x$ has zero degree in at least one of those tail link graphs, then we have removed at most three edges and added four, a strict increase. We have also not created any triples of vertices with more than one edge or any Escher graphs. We may now assume that $\|T'\| \leq 4$ and that the tail link graphs of vertices in $T'$ are never two independent edges. ## Case 3: $\|T'\| =0,1,2$ First, note that if $H'$ has a triple with more than one edge $\{x,y,z\}$ then at least two of its vertices must be in $T'$ as a consequence of Lemma [\[observation\]](#observation){reference-type="ref" reference="observation"}. Therefore, if $\|T'\|=0,1$, then $H'$ is oriented and so \[\|E(H)\| \leq \|E(H')\| < {n \choose 3}.\] Moreover, if $T'=\{x,y\}$ and $H'$ is not oriented, then any vertex triple with more than one edge must have two edges of the form, \[zx \rightarrow y,zy \rightarrow x\] for some third vertex $z$. If there exist two such vertices $z_1 \neq z_2$ that satisfy this, then there would be an Escher graph. Hence, there is at most one vertex triple with more than one edge and it would have at most two edges. Therefore, \[\|E(H)\| \leq \|E(H')\| \leq {n \choose 3}+1.\] ## Case 4: $\|T'\| =3$ First, suppose that there exists a triple, $\{x,y,z\}$ with all three possible edges. Then $T' = \{x,y,z\}$. Since any triple with multiple edges must intersect $T'$ in at least two vertices, then any additional such triple would make an Escher graph with one of the edges in $T'$. Therefore, $H'$ has exactly one triple of vertices with all three edges on it and no others. So \[\|E(H) \leq \|E(H')\| \leq {n \choose 3} + 2.\] Moreover, to attain this number of edges, no triple of vertices can be empty of edges. In this case,$H'$ must be isomorphic to the first construction $H_1$. Next, assume that no triple of vertices has all three edges and let $T' = \{x,y,z\}$. Therefore, $H'$ needs at least two triples of vertices that each hold two edges or else \[\|E(H)\| \leq \|E(H')\| \leq {n \choose 3} + 1\] automatically. Suppose one of the multiedges is $\{x,y,z\}$ itself. Then without loss of generality let the edges be $xy \rightarrow z$ and $xz \rightarrow y$. The second triple with two edges must have its third vertex in $D'$. Call this vertex $v$.The vertex $x$ cannot be in this second triple of vertices without creating an Escher graph. So the edges must be $vy \rightarrow z$ and $vz \rightarrow y$. But this also creates an Escher graph. Therefore, neither of the two triples that hold two edges are contained entirely within $T'$. So without loss of generality they must be $vx \rightarrow y, vy \rightarrow x$ and $wy \rightarrow z, wz \rightarrow y$. If $v \neq w$, then $vx, wz \in T_y$, a contradiction to our assumption that $T'$ contains no vertices with tail link graphs that are two independent edges. Hence, $v=w$. Since $v \in D$, then $T_v$ has at least three vertices. Moreover, since $v$ is in the tail link graphs of each vertex of $T'$, then none of these vertices can be in $T_v$. Remove all edges pointing to the vertices of $T'$. This is at most 6 edges. Add all possible edges with $v$ as the head and a tail set among the set $V(T_v) \cup \{x,y,z\}$. This adds at least 12 new edges. The new graph is oriented and $E$-free. Therefore, $\|E(H)\| < {n \choose 3}$. ## Case 5: $\|T'\| =4$ First, assume that there is some triple $\{x,y,z\}$ that contains all three possible edges. As before, there are no additional triples with more than one edge. So \[\|E(H)\| \leq \|E(H')\| \leq {n \choose 3} + 2.\] The first construction $H_1$ is the unique extremal construction under this condition since all triples must be used at least once. So assume that all triples with more than one edge have two edges each. Then we must have at least two. Assume that one of them is contained within $T' = \{a,b,c,d\}$. Without loss of generality let it be $ab \rightarrow c, ac \rightarrow b$. Since the second such triple intersects $T'$ in at least two vertices, then it must intersect $\{a,b,c\}$ in at least one vertex. If it intersects $\{a,b,c\}$ in two vertices, then without loss of generality (to avoid a copy of $E$) the second triple must be of the form $ab \rightarrow x, ax \rightarrow b$. Hence, $x \in T'$ so $x=d$. But now there is no edge possible on $\{b,c,d\}$. Therefore, there must be a third such triple for $H'$ to have ${n \choose 3} +2$ edges. This triple must be $ac \rightarrow d, ad \rightarrow c$. And the only way to actually make it to the maximum number of edges now must be to have an edge on every other triple. Every triple of the form $\{b,c,s\}$ for $s \in D$ must have the edge $bc \rightarrow s$ since the other two options would create an Escher graph. Similarly, $bd \rightarrow s$ and $cd \rightarrow s$ are the only options for triples of the form $\{b,d,s\}$ and $\{c,d,s\}$ respectively. Next, any triple of the form $\{a,b,s\}$ must hold the edge $ab \rightarrow s$ since the other two edges create Escher graphs. Similarly, every triple of the forms $\{a,c,s\}$ and $\{a,d,s\}$ must hold the edges $ac \rightarrow s$ and $ad \rightarrow s$ respectively. Since each triple contained in $D$ holds exactly one edge, then the induced subgraph on $D$ must be isomorphic to the oriented extremal example of an $E$-free graph on $n-4$ vertices. Therefore, the entire graph $H'$ must be isomorphic to the second extremal construction $H_2$ in order to attain ${n \choose 3} + 2$ edges. So assume that the second triple with two edges intersects $\{a,b,c\}$ in only one vertex. Then these edges must be $xa \rightarrow d, xd \rightarrow a$. This can be the only additional triple with two edges. So to make it to ${n \choose 3} +2$ edges we need each triple to have an edge. However, the edge for $\{a,b,d\}$ is forced to be $ad \rightarrow b$ and the edge for $\{b,c,d\}$ is forced to be $bc \rightarrow d$. This makes an Escher graph. So \[\|E(H)\| \leq \|E(H')\| \leq {n \choose 3} + 1.\] Now assume that no vertex triple with multiple edges is contained entirely within $T'$, but assume that there are at least two such triples in $H'$. The only way that two triples could have distinct vertices in $D'$ is if they were of the forms (without loss of generality), $xa \rightarrow b, xb \rightarrow a$, and $yc \rightarrow d, yd \rightarrow c$. Otherwise, the pairs of the two triples that are in $T'$ would intersect resulting in either a copy of $E$ (if both triples use the same pair) or a vertex in $T'$ with two independent edges as a tail link graph. So there must be exactly two such triples. Therefore, all other triples of vertices must contain exactly one edge in order to reach ${n \choose 3}+2$ edges overall. To avoid the forbidden subgraph this edge must be $ab \rightarrow c$ for the triple $\{a,b,c\}$ and $cd \rightarrow a$ for the triple $\{a,c,d\}$. But this is an Escher graph. Hence, not all triples may be used and so \[\|E(H)\| \leq \|E(H')\| \leq {n \choose 3} + 1.\] Therefore, we may now assume for each multiedge triple that the vertex from $D'$ is always $x$. First, assume that there are only two such triples. As before, if we assume that the only two such triples are $xa \rightarrow b, xb \rightarrow a$ and $xc \rightarrow d, xd \rightarrow c$, then there can be not be an edge on both $\{a,b,c\}$ and $\{a,c,d\}$. Hence, there would be a suboptimal number of edges overall. On the other hand, if the only two such triples are adjacent in $T'$, then they are, without loss of generality, $xa \rightarrow b, xb \rightarrow a$ and $xb \rightarrow c, xc \rightarrow a$. In this case, no edge can go on the triple $\{a,b,c\}$ at all and so there are at most ${n \choose 3} + 1$ edges overall. Therefore, we must assume there are at least three such triples that meet at $x$. If these three triples make a triangle in $T'$, then they are $xa \rightarrow b, xb \rightarrow a$, $xb \rightarrow c, xc \rightarrow b$, and $xc \rightarrow a, xa \rightarrow c$. Again, there can be no edges on the triple $\{a,b,c\}$. Hence, every other triple must hold an edge to attain ${n \choose 3} + 2$ edges overall. On the triple $\{a,b,d\}$ this edge must be $ab \rightarrow d$ to avoid making a copy of $E$. Similarly, we must have the edges $ac \rightarrow d$ and $bc \rightarrow d$. But this means that $d \not \in T'$, a contradiction. On the other hand, if there are three triples of vertices with more than one edge on each that do not make a triangle in $T'$ or if there are four or more such triples, then $x$ is in the tail link graphs for each vertex in $T'$. Hence, none of these vertices may be in the tail link graph, $T_x$. However, $x \in D'$ so its tail link graph has at least three vertices. Remove all edges pointing to vertices of $T'$ (at most 8). Add all edges pointing to $x$ with tail sets in $T'$ (6 new edges) and between $T'$ and $V(T_x)$ (at least 12 new edges). So this adds at least ten edges to $H'$ to create $H''$. $H''$ is oriented so \[\|E(H)\| < \|E(H'')\| \leq {n \choose 3}.\] This exhausts all of the cases and establishes that \[\text{ex}(n,E) = {n \choose 3} +2\] with exactly two extremal examples up to isomorphism. # Conclusion In, Brown and Harary started studying extremal problems for directed $2$-graphs by determining the extremal numbers for many "small\" digraphs and for some specific types of digraphs such as tournaments-a digraph where every pair of vertices has exactly one directed edge. We could follow their plan of attack in studying this $2 \rightarrow 1$ model and also look for the extremal numbers of tournaments. Here, a tournament would be a graph with exactly one directed edge on every three vertices. In particular, a transitive tournament might be an interesting place to begin. A transitive tournament is a tournament where the direction of each edge is based on an underlying linear ordering of the vertices as in the oriented lower bound construction in Theorem [\[exG\]](#exG){reference-type="ref" reference="exG"}. Denote the $2 \rightarrow 1$ transitive tournament on $k$ vertices by $TT_k$. Since the "winning\" vertex of the tournament will have a complete $K_{k-1}$ as its tail link graph, then any $H$ on $n$ vertices for which each $T_{x}$ is $K_{k-1}$-free must be $TT_k$-free. Therefore, \[n \left(\frac{n-1}{k-2}\right)^2 {k-2 \choose 2} \leq \text{ex}(n,TT_k), \text{ex}_o(n,TT_k).\] This also immediately shows that the transitive tournament on four vertices with the "bottom\" edge removed has this extremal number exactly. Is it still true if we add an edge to $\{a,b,c\}$? Another way of generalizing the extremal questions asked in this paper is to ask about $r \rightarrow 1$ models of directed hypergraphs. If we look at every $(r \rightarrow 1)$-graph with exactly two edges we see that these fall into four main types of graph. Let $i$ be the number of vertices that are in the tail set of both edges. Then let $T_r(i)$ denote the graph where both edges point to the same head vertex, let $H_r(i)$ denote the graph where the edges point to different head vertices neither of which are in the tail set of the other, let $R_r(i)$ denote the graph where the first edge points to a head vertex in the tail set of the second edge and the second edge points to a head not in the tail set of the first edge, and let $E_r(i)$ denote the graph where both edges point to heads in the tail sets of each other. So in terms of the graphs discussed in this paper, the 3-resolvent would be a $R_2(1)$, the 4-resolvent would be a $R_2(0)$, the Escher graph would be a $E_2(0)$, and two edges on the same triple of vertices would be an $E_2(1)$. The nondegenerate cases here would be $R_r(i)$ and $E_r(i)$. It would be interesting to find the extremal numbers for these graphs in general. To what extent do the current proofs extend to these graphs? For example, any $r \rightarrow 1$ transitive tournament on $n$ vertices would be $E_r(i)$-free. This solves the oriented version and gives a lower bound for the standard version: \[\text{ex}_o(n,E_r(i)) = {n \choose r+1}.\] For the generalized resolvent configurations, the lower bound extremal constructions for $R_3$ and $R_4$ generalize easily to the $r \rightarrow 1$ setting, but are they ever tight? Can we generalize these constructions to get extremal numbers for all $R_r(i)$?	{'timestamp': '2016-07-19T02:09:37', 'yymm': '1607', 'arxiv_id': '1607.04930', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04930'}
# Introduction {#sec:Intro} Wireless communication, by its inherent broadcast nature, is vulnerable to eavesdropping by illegitimate receivers within communication range of the source. Wyner in, for the first time, information-theoretically addressed the problem of secure communication in the presence of an eavesdropper and showed that secure communication is possible if the eavesdropper channel is a degraded version of the destination channel. The rate at which information can be transferred from the source to the intended destination while ensuring complete equivocation at the eavesdropper is termed as secrecy rate and its maximum over all input probability distributions is defined as the secrecy capacity of channel. Later, extended Wyner's result to Gaussian channels. These results are further extended to various models such as multi-antenna systems, multiuser scenarios, fading channels. An interesting direction of work on secure communication in the presence of eavesdropper(s) is one in which the source communicates with the destination via relay nodes. Such work has considered various scenarios such as different relaying schemes (amplify-and-forward and decode-and-forward), constraints on total or individual relay power consumption, one or more eavesdroppers. However, except for a few specific scenarios, such work does not provide tight characterization of secrecy capacity or even optimal secrecy rate achievable with the given relaying scheme. Further, all previous work only considered secure communication scenarios where the source communicates with the legitimate destination(s) in two hops, over so called diamond network. We consider a multihop unicast communication over layered network of relays in the presence of a single eavesdropper. The relays nodes are arranged in layers where all relays in a particular layer can communicate only with the relays in the next layer. The relay nodes, operating under individual power constraints, amplify-and-forward the signals received at their input. In this scenario, multiple relay nodes in each layer can cooperate to enhance the end-to-end achievable rate. Also, the signals transmitted simultaneously by the relays add in the air, thus providing an opportunity for the relays in the second layer onward to perform Analog Network Coding (ANC) on the received noisy sum of these signals, where each relay merely amplifies and forwards this noisy sum. The eavesdropper can overhear the transmissions from the relay nodes of any of the layers depending on its location. The objective is to maximize the rate of secure transmission from the source to the destination by choosing the optimal set of scaling factors for the ANC-relays, irrespective of the relays that the eavesdropper listens to. However, so far, there exists no closed-form expression or polynomial time algorithm to exactly characterize the optimal AF secrecy rate even for general two-hop (diamond) relay networks, except for a few specific cases where eavesdropper's channel is a degraded or scaled version of destination channel and characterizing the optimal AF secrecy rate for general layered network is an even harder problem than general diamond network. Thus, to get some insights into the nature of the optimal solution for such networks, we consider symmetric layered networks, where all channel gains between the nodes in two adjacent layers are equal, thus the nomenclature of these networks as "Equal Channel Gains between Adjacent Layers (ECGAL)" networks. We provide closed-form solutions for the optimal secure AF rate for such networks. We envision that these results may help us gain insight into the nature of the optimal solution and develop techniques which may further help in construction of low-complexity optimal schemes for general relay networks. The eavesdropper being a passive entity, a realistic eavesdropper scenario is the one where nothing about the eavesdropper's channel is known, neither its existence, nor its channel state information (CSI). However, the existing work on secrecy rate characterization assumes one of the following: (1) the transmitter has prefect knowledge of the eavesdropper channel states, (2) compound channel: the transmitter knows that the eavesdropper channel can take values from a finite set, and (3) fading channel: the transmitter only knows distribution of the eavesdropper channel. In this paper, we assume that the CSI of the eavesdropper channel is known perfectly for the following two reasons. First, this provides an upper bound to the achievable secure ANC rate for the scenarios where we have imperfect knowledge of the eavesdropper channel. For example, the lower (upper) bound on the compound channel problem can be computed by solving the perfect CSI problem with the worst (best) channel gain from the corresponding finite set. Further, this also provides a benchmark to evaluate the performance of achievability schemes in such imperfect knowledge scenarios. Second, this assumption allows us to focus on the nature of the optimal solution and information flow, instead of on complexities arising out of imperfect channel models. The key contribution of this work is the computation of the globally optimal set of scaling factors for the nodes that maximizes the end-to-end secrecy rate for a class of layered networks. We also show that in the high-SNR regime, ANC achieves secrecy rates within an explicitly computed constant gap of the cutset upper bound on the secrecy capacity. To the best of our knowledge, this work offers the first characterization of the performance of secure ANC in multi-layered networks in the presence of an eavesdropper. Organization: In Section [2](#sec:sysMdl){reference-type="ref" reference="sec:sysMdl"} we introduce the system model and formulate the problem of maximum secure ANC rate achievable in the proposed system model. In section [3](#sec:OptBeta){reference-type="ref" reference="sec:OptBeta"} we compute the optimal vector of scaling factors of the nodes of an ECGAL network when eavesdropper snoops on the transmissions of the nodes in any one of the $L$ layers. Then, in Section [4](#sec:highSNRanalysis){reference-type="ref" reference="sec:highSNRanalysis"} we analyze the high-SNR behavior of the achievable secure ANC rate and show that it lies within a constant gap from the corresponding cutset upper bound on the secrecy capacity and Section [5](#sec:numSim){reference-type="ref" reference="sec:numSim"} we numerically validate these results. Finally, Section [6](#sec:cnclsn){reference-type="ref" reference="sec:cnclsn"} concludes the paper. # System Model {#sec:sysMdl} Consider a $(L+2)$-layer wireless network with directed links. The source $s$ is at layer '$0$', the destination $t$ is at layer '$L+1$' and the relays from the set $R$ are arranged in $L$ layers between them. The $l^{th}$ layer contains $n_l$ relay nodes, $\sum _{l-1}^{L} n_l = \|R\|$. The source $s$ transmits message signals to the destination $t$ via $L$ relay layers. However, the signals transmitted by the relays in a layer are also overheard by the eavesdropper $e$. An instance of such a network is given in Figure [\[fig:layrdNetExa\]](#fig:layrdNetExa){reference-type="ref" reference="fig:layrdNetExa"}. Each node is assumed to have a single antenna and operate in full-duplex mode, e.g. as in. At instant $n$, the channel output at node $i, i \in R \cup \{t, e\}$, is \[\label{eqn:channelOut} y_i[n] = \sum_{j \in {\mathcal N}(i)} h_{ji} x_j[n] + z_i[n], \quad-\infty < n < \infty,\] where $x_j[n]$ is the channel input of node $j$ in neighbor set ${\mathcal N}(i)$ of node $i$. In [\[eqn:channelOut\]](#eqn:channelOut){reference-type="eqref" reference="eqn:channelOut"}, $h_{ji}$ is a real number representing the channel gain along the link from the node $j$ to the node $i$ and constant over time (as in, for example) and known (even for the eavesdropper channels) throughout the network. All channel gains between the nodes in two adjacent layers are assumed to be equal, thus the nomenclature of these networks as "Equal Channel Gains between Adjacent Layers (ECGAL)" networks. The source symbols $x_s[n],-\infty < n < \infty$, are i.i.d. Gaussian random variables with zero mean and variance $P_s$ that satisfy an average source power constraint, $x_s[n] \sim {\cal N}(0, P_s)$. Further, $\{z_i[n]\}$ is a sequence (in $n$) of i.i.d. Gaussian random variables with $z_i[n] \sim {\cal N}(0, \sigma^2)$. We assume that $z_i$ are independent of the input signal and of each other. We also assume that each relay's transmit power is constrained as: \[\label{eqn:pwrConstraint} E[x_i^2[n]] \le P, \quad i \in R,-\infty < n < \infty\] In ANC, each relay node amplifies and forwards the noisy signal sum received at its input. More precisely, relay node $i, i \in R$, at instant $n+1$ transmits the scaled version of $y_i[n]$, its input at time instant $n$, as follows \[\label{eqn:AFdef} x_i[n+1] = \beta_i y_i[n], \quad 0 \le \beta_i^2 \le \beta_{i,max}^2 = P/P_{R,i},\] where $P_{R,i}$ is the received power at node $i$ and choice of scaling factor $\beta_i$ satisfies the power constraint [\[eqn:pwrConstraint\]](#eqn:pwrConstraint){reference-type="eqref" reference="eqn:pwrConstraint"}. Assuming equal delay along each path, for the network in Figure [\[fig:layrdNetExa\]](#fig:layrdNetExa){reference-type="ref" reference="fig:layrdNetExa"}, the copies of the source signal ($x_s[.]$) and noise signals ($z_i[.]$), respectively, arrive at the destination and the eavesdropper along multiple paths of the same delay. Therefore, the signals received at the destination and eavesdropper are free from intersymbol interference (ISI). Thus, we can omit the time indices and use equations [\[eqn:channelOut\]](#eqn:channelOut){reference-type="eqref" reference="eqn:channelOut"} and [\[eqn:AFdef\]](#eqn:AFdef){reference-type="eqref" reference="eqn:AFdef"} to write the input-output channel between the source $s$ and the destination $t$ as \[\label{eqn:destchnl} y_t = \left[\sum\limits_{(i_1...,i_L) \in K_{st}}\!\!\!\!\!\!\!\!\! h_{s,i_1}\beta_{i_1}h_{i_1, i_2}...h_{i_{L-1}, i_L}\beta_{i_L} h_{i_L, t}\right] x_s + \sum\limits_{l=1}^L \sum\limits_{j-1}^{n_l}\left[\sum\limits_{(i_1...,i_{L-l}) \in K_{lj,t}} \!\!\!\!\!\!\!\!\!\beta_{lj} h_{lj,i_1}...\beta_{i_{L-l}} h_{i_{L-l},t}\right] z_{lj} + z_t\] where $K_{st}$ is the set of $L$-tuples of node indices corresponding to all paths from the source $s$ to the destination $t$ with path delay $L$. Similarly, $K_{lj,t}$ is the set of $L-l$-tuples of node indices corresponding to all paths from the $j^{th}$ relay of the $l^{th}$ layer to the destination with path delay $L-l + 1$. We introduce modified channel gains as follows. For all the paths between the source and the destination: \[h_{st} = \sum\limits_{(i_1...,i_L) \in K_s} h_{s,i_1}\beta_{i_1}h_{i_1, i_2}...h_{i_{L-1}, i_L}\beta_{i_L} h_{i_L, t}\] For all the paths between the $j^{th}$ relay of the $l^{th}$ layer to the destination $t$ with path delay $L-l + 1$: \[h_{lj,t} = \sum\limits_{(i_1...,i_{L-l}) \in K_{lj}} \beta_{lj} h_{lj,i_1}...\beta_{i_{L-l}} h_{i_{L-l},t}\] In terms of these modified channel gains, the source-destination channel in [\[eqn:destchnl\]](#eqn:destchnl){reference-type="eqref" reference="eqn:destchnl"} can be written as: \[\label{eqn:destchnlmod} y_t = h_{st} x_s + \sum_{l=1}^{L} \sum_{j=1}^{n_l} h_{lj,t} z_{lj} + z_t,\] Similarly, assuming that the eavesdropper is overhearing the transmissions of the relays in the layer $E, 1\le E \le L$, the input-output channel between the source and the eavesdropper can be written as \[\label{eqn:evechnlmod} y_e = h_{se} x_s + \sum_{l=1}^{E} \sum_{j=1}^{n_l} h_{lj,e} z_{lj} + z_t,\] The secrecy rate at the destination for such a network model can be written as, $R_s(P_s)= [I(x_s;y_t)-I(x_s;y_e)]^+$, where $I(x_s;y)$ represents the mutual information between random variable $x_s$ and $y$ and $[u]^+=\max\{u,0\}$. The secrecy capacity is attained for the Gaussian channels with the Gaussian input $x_s \sim \mathcal{N}(0,P_s)$, where $\mathbf{E}[x_s^2] = P_s$,. Therefore, for a given network-wide scaling vector $\bm{\beta} = (\beta_{li})_{1 \le l \le L, 1 \le i \le n_l}$, the optimal secure ANC rate for the channels in [\[eqn:destchnlmod\]](#eqn:destchnlmod){reference-type="eqref" reference="eqn:destchnlmod"} and [\[eqn:evechnlmod\]](#eqn:evechnlmod){reference-type="eqref" reference="eqn:evechnlmod"} can be written as the following optimization problem. where $SNR_t(P_s,\bm{\beta})$, the signal-to-noise ratio at the destination $t$ is: \[\label{eqn:snrt} SNR_t(P_s,\bm{\beta}) = \frac{P_s}{\sigma^2}\frac{h_{st}^2}{1 + \sum_{l=1}^{L} \sum_{j=1}^{n_l} h_{lj,t}^2}\] and similarly, $SNR_e(P_s,\bm{\beta})$ is \[\label{eqn:snre} SNR_e(P_s,\bm{\beta}) = \frac{P_s}{\sigma^2}\frac{h_{se}^2}{1 + \sum_{l=1}^{E} \sum_{j=1}^{n_l} h_{lj,e}^2}\] Given the monotonicity of the $\log(\cdot)$ function, we have \[\begin{aligned} \bm{\beta}_{opt} &= \operatornamewithlimits{argmax}_{\bm{\beta}} \left[R_t(P_s,\bm{\beta})-R_e(P_s,\boldsymbol{\beta})\right] \nonumber\\ &= \operatornamewithlimits{argmax}_{\bm{\beta}} \frac{1+SNR_t(P_s,\bm{\beta})}{1+SNR_e(P_s,\bm{\beta})} \label{eqn:eqProb} \end{aligned}\] # The Optimal Secure ANC Rate Analysis {#sec:OptBeta} In this section, we analyze the optimal secure ANC rate problem in [\[eq:optSecrate\]](#eq:optSecrate){reference-type="eqref" reference="eq:optSecrate"} or [\[eqn:eqProb\]](#eqn:eqProb){reference-type="eqref" reference="eqn:eqProb"} first for diamond networks and then for ECGAL networks. ## Symmetric Diamond Networks {#subsec:diamondNetOptBeta} Consider a symmetric diamond with $N$ relay nodes arranged in a layer between the source and the destination as shown in Figure [\[fig:diamondNet\]](#fig:diamondNet){reference-type="ref" reference="fig:diamondNet"}. Using [\[eqn:snrt\]](#eqn:snrt){reference-type="eqref" reference="eqn:snrt"} and [\[eqn:snre\]](#eqn:snre){reference-type="eqref" reference="eqn:snre"}, the $SNR_t$ and $SNR_e$ in this case are: \[\begin{aligned} SNR_t = \frac{P_s h_s^2 }{\sigma^2} \frac{(\sum_{i=1}^{N} \beta_i )^2 h_t^2}{1 + \left(\sum_{i=1}^{N} \beta_i^2 \right) h_t^2 } &&\mbox{ and }&& SNR_e = \frac{P_s h_s^2 }{\sigma^2} \frac{(\sum_{i=1}^{N} \beta_i )^2 h_e^2}{1 + \left(\sum_{i=1}^{N} \beta_i^2 \right) h_e^2 } \end{aligned}\] Here, it is assumed that the eavesdropper chooses to snoop on all the nodes of the layer which is an optimal strategy in the symmetric layered networks for the eavesdropper as we prove later. However, in general this may not be the case. The eavesdropper can choose to snoop on fewer nodes and still get higher rate compared to the case when it snoops on all the nodes of a layer as illustrated in the following example. Although, for general layered networks it is very difficult to find which subset of relay nodes the eavesdropper chooses to snoop on so as to maximize its rate; for symmetric networks it can be easily verified that the rate at the eavesdropper is maximized when it snoops on all the nodes of a layer. For instance, consider the scenario where all the nodes transmitting at maximum power is optimum with respect to secrecy rate maximization. The SNR at the eavesdropper when it chooses to snoop on $k$ nodes is given as \[\begin{aligned} SNR_e^k & = \frac{P_s h_s^2}{\sigma^2} \frac{k^2 \beta_{max}^2 h_e^2}{1+ k \beta_{max}^2 h_e^2} \end{aligned}\] Similarly, SNR at the eavesdropper when it chooses to snoop on $k+1$ nodes is \[\begin{aligned} SNR_e^{k+1} & = \frac{P_s h_s^2}{\sigma^2} \frac{(k+1)^2 \beta_{max}^2 h_e^2}{1+ (k+1) \beta_{max}^2 h_e^2} \end{aligned}\] Clearly, \[\begin{aligned} SNR_e^{k+1}-SNR_e^k & = \frac{P_s h_s^2}{\sigma^2} \frac{{\beta}^{2} h_e^2 \left( {\beta}^{2} h_e^2 {k}^{2}+{\beta}^{2} h_e^2 k+2 k+1\right)}{\left( {\beta}^{2} h_e^2 k+1\right) \left( {\beta}^{2} h_e^2 k+{\beta}^{2} h_e^2+1\right) } \geq 0, \end{aligned}\] i.e., $SNR_e^{k+1} \geq SNR_e^k$. Thus, for such scenarios eavesdropper achieves a higher rate when it chooses to snoop on more number of nodes and eventually it will snoop on maximum possible number of nodes to maximize its rate. Figure [\[fig:3relayDiamondNetRePlot\]](#fig:3relayDiamondNetRePlot){reference-type="ref" reference="fig:3relayDiamondNetRePlot"} shows the rates achievable at the eavesdropper $(R_{e,i})$ when it chooses to snoop on $i,\ i \in \{1,2,3\}$ number of nodes of a 3 relay symmetric diamond network of the specified parameters with all the nodes transmitting at their corresponding optimum values so as to maximize the secrecy rate in each case. Here, again it can be seen that rate at the eavesdropper is increases when it snoops on all the nodes. ## ECGAL Layered Networks {#subsec:lyrNetOptBeta} In this subsection, we consider the optimal secure ANC rate problem in [\[eq:optSecrate\]](#eq:optSecrate){reference-type="eqref" reference="eq:optSecrate"} for ECGAL networks where the source communicates with the destination via $L$ intermediate relay layers with all channel gains between two adjacent layers being equal. For the sake of ease of representation, let there be $N$ relays in each layer. The eavesdropper overhears the transmission from the nodes in relay layer $M, 1 \le M \le L$. An instance of such a network is given in Figure [\[fig:layrdNetExa\]](#fig:layrdNetExa){reference-type="ref" reference="fig:layrdNetExa"}. Using [\[eqn:snrt\]](#eqn:snrt){reference-type="eqref" reference="eqn:snrt"} and [\[eqn:snre\]](#eqn:snre){reference-type="eqref" reference="eqn:snre"}, the $SNR_t$ and $SNR_e$ in this case are: \[\begin{aligned} SNR_t &= \frac{P_s}{\sigma^2} \frac{h_s^2 H_{1,M-1}^2 \left(\sum_{n=1}^N \beta_{M,n}\right)^2 h_M^2 H_{M+1,L}^2}{\left[\left(\sum_{i=1}^{M-1} G_{i,M-1}^2\right)\left(\sum_{n=1}^N \beta_{M,n}\right)^2 h_M^2 +\left(\sum_{n=1}^N \beta_{M,n}^2\right)h_M^2\right] H_{M+1,L}^2 +\sum_{i=M+1}^L G_{i,L}^2+1}\\ SNR_e &= \frac{P_s}{\sigma^2} \frac{h_s^2 H_{1,M-1}^2 \left(\sum_{n=1}^N \beta_{M,n}\right)^2 h_e^2}{\left(\sum_{i=1}^{M-1} G_{i,M-1}^2\right)\left(\sum_{n=1}^N \beta_{M,n}\right)^2 h_e^2+\left(\sum_{n=1}^N \beta_{M,n}^2\right)\,h_e^2+1} \end{aligned}\] where \[\begin{aligned} H_{i,j}^2 &= \prod_{k=i}^j \left(\sum_{n=1}^N \beta_{k,n}\right)^2 h_k^2\\ G_{i,j}^2 &= \left(\sum_{n=1}^N \beta_{i,n}^2\right) h_i^2 \prod_{k=i+1}^{j} \left(\sum_{n=1}^N \beta_{k,n}\right)^2 h_k^2 \end{aligned}\] Introduce the following parameters \[\begin{aligned} E &= h_s^2 H_{1,M-1}^2 \\ F &= \sum_{i=1}^{M-1} G_{i,M-1}^2 \\ \alpha &= \prod_{i=M+1}^{L}\left(\sum_{n=1}^N\sqrt{P_{i,n}}\right)^2 h_i^2 \\ \lambda &= \sum_{i=M+1}^L \lambda_{i} \left(\sum_{n=1}^N P_{i+1,n}\right) \prod_{j=i+2}^{L}\left(\sum_{n=1}^N\sqrt{P_{i,n}}\right)^2 h_j^2 \\ \mu &= \prod_{i=M+1}^{L}\left(\sum_{n=1}^N\sqrt{P_{i,n}}\right)^2 h_i^2 + \sum_{i=M+1}^L \mu_{i} \left(\sum_{n=1}^N P_{i+1,n}\right) \prod_{j=i+2}^{L}\left(\sum_{n=1}^N\sqrt{P_{i,n}}\right)^2 h_j^2\\ &= \alpha + \sum_{i=M+1}^L \mu_{i} \left(\sum_{n=1}^N P_{i+1,n}\right) \prod_{j=i+2}^{L}\left(\sum_{n=1}^N\sqrt{P_{i,n}}\right)^2 h_j^2 \\ \nu &= \sum_{i=M}^L \nu_{i} \left(\sum_{n=1}^N P_{i+1,n}\right) \prod_{j=i+2}^{L}\left(\sum_{n=1}^N\sqrt{P_{i,n}}\right)^2 h_j^2, \quad \nu_M = 1, \end{aligned}\] where \[\begin{aligned} \lambda_{M+1} &= P_s (s_{M+1} + \sigma^2 n_{M+1}), \quad s_{M+1} = 1, n_{M+1} = 0 &\\ \lambda_{i} &= P_s\left[s_{i-1} \left(\left(\sum_{n=1}^N\!\!\sqrt{P_{i-1,n}}\right)^2\!\! h_{i-1}^2\!\!+\!\sigma^2\right)\!\! +\! \sigma^2 n_{i-1} \left(\left(\sum_{n=1}^N P_{i-1,n}\right) h_{i-1}^2 \!\!+\!\sigma^2 \right)\right], & i \in \{M+2, \ldots, L\}\\ \mu_{M+1} &= \sigma^2 (s_{M+1} + \sigma^2 n_{M+1}), \quad s_{M+1} = 1, n_{M+1} = 0 &\\ \mu_{i} &= \sigma^2 \left[s_{i-1} \left(\left(\sum_{n=1}^N\!\!\sqrt{P_{i-1,n}}\right)^2\!\! h_{i-1}^2\!\!+\!\sigma^2\right)\!\! +\! \sigma^2 n_{i-1} \left(\left(\sum_{n=1}^N P_{i-1,n}\right) h_{i-1}^2 \!\!+\!\sigma^2 \right)\right], & i \in \{M+2, \ldots, L\}\\ \nu_{M+1} &= (s_{M+1} + \sigma^2 n_{M+1}), \quad s_{M+1} = 0, n_{M+1} = 1&\\ \nu_{i} &= \sigma^2 \left[s_{i-1} \left(\left(\sum_{n=1}^N\!\!\sqrt{P_{i-1,n}}\right)^2\!\! h_{i-1}^2\!\!+\!\sigma^2\right)\!\! +\! \sigma^2 n_{i-1} \left(\left(\sum_{n=1}^N P_{i-1,n}\right) h_{i-1}^2 \!\!+\!\sigma^2 \right)\right], & i \in \{M+2, \ldots, L\} \end{aligned}\] Using the preceding lemma and the above parameters, the problem \[\bm{\beta}_{opt} = \operatornamewithlimits{argmax}_{\bm{\beta}} \frac{1+SNR_t}{1+SNR_e}\] is reduced to the following subproblem \[\label{eqn:secRate4lemma2} (\bm{\beta}_{1,opt}, \ldots, \bm{\beta}_{M,opt}) = \operatornamewithlimits{argmax}_{(\bm{\beta}_{1}, \ldots, \bm{\beta}_{M})} \frac{1+SNR_t\|_{\bm{\beta}_{M+1:L,max}}}{1+SNR_e},\] where for a given network-wide vector of scaling factors $(\bm{\beta}_1, \ldots, \bm{\beta}_M, \bm{\beta}_{M+1,max}, \ldots, \bm{\beta}_{L,max})$, the received SNRs at the destination and the eavesdropper are, respectively \[\begin{aligned} SNR_t\|_{\bm{\beta}_{M+1:L,max}} &= \frac{P_s}{\sigma^2} \frac{A \left(\sum_{n=1}^N \beta_{M,n}\right)^2 h_M^2}{B \left(\sum_{n=1}^N \beta_{M,n}\right)^2 h_M^2 + C\left(\sum_{n=1}^N \beta_{M,n}^2\right) h_M^2 + D}\\ SNR_e &= \frac{P_s}{\sigma^2} \frac{E \left(\sum_{n=1}^N \beta_{M,n}\right)^2 h_e^2}{F \left(\sum_{n=1}^N \beta_{M,n}\right)^2 h_M^2 + \left(\sum_{n=1}^N \beta_{M,n}^2\right) h_M^2 + 1} \end{aligned}\] with \[\begin{aligned} A &= \alpha E\\ B &= \lambda E + \mu F\\ C &= \mu\\ D &= \nu \end{aligned}\] In short, Lemma [\[lemma:lyrdNetReducedBeta1\]](#lemma:lyrdNetReducedBeta1){reference-type="ref" reference="lemma:lyrdNetReducedBeta1"}-[\[lemma:lyrdNetReducedBeta3\]](#lemma:lyrdNetReducedBeta3){reference-type="ref" reference="lemma:lyrdNetReducedBeta3"} together establish that for the ECGAL layered networks with $L$ relays, the optimal vector of the scaling factors that maximizes the secure ANC rate is $\bm{\beta}_{opt} = (\bm{\beta}_{1,max}, \ldots, \bm{\beta}_{M,opt}, \bm{\beta}_{M+1,max}, \ldots, \bm{\beta}_{L,max})$, where $\beta_{M,opt}$ is given by Lemma [\[lemma:lyrdNetReducedBeta2\]](#lemma:lyrdNetReducedBeta2){reference-type="ref" reference="lemma:lyrdNetReducedBeta2"}. # High-SNR Analysis of Achievable ANC Secrecy Rate in ECGAL Networks {#sec:highSNRanalysis} We define a wireless layered network to be in high SNR regime if \[\min_{i \in \{1, \ldots, L\}} SNR_i \ge \frac{1}{\delta}\] for some small $\delta \ge 0$. Here, $SNR_i$ is the signal-to-noise ratio at the input of any of the relay nodes in the $i^{th}$ layer. Assume that each relay node in layer $i$ uses the amplification factor \[\beta_i^2 = \frac{P}{(1+\delta) P_{R_i,max}}, i \in \{1, \ldots, L\},\] where $P_{R_i,max}$ is the maximum received signal power at any relay node in the $i^{th}$ layer which in this case is equal to $N^2 P h_{i-1}^2$ as each relay node receives the transmissions from $N$ relay nodes in the previous layer with maximum transmit power constrained by $P$. It should be noted that $\beta_i$ is such that the maximum power constraint [\[eqn:pwrConstraint\]](#eqn:pwrConstraint){reference-type="eqref" reference="eqn:pwrConstraint"} is satisfied at each node as \[\begin{aligned} \beta_{i,max}^2 = \frac{P}{P_{R_i}+P_{z_i}+\sigma^2} = \frac{P}{\left(1+\frac{1}{SNR_i}\right)P_{R_i}} \geq \frac{P}{(1+\delta)P_{R_i,max}} \quad [\mbox{Since }1/SNR_i \leq \delta, P_{R_i} \leq P_{R_i,max}] \end{aligned}\] Here, $P_{R_i}$ and $P_{z_i}$ are the received signal and noise powers at the input of the node $i$, respectively. Note that as $\delta \rightarrow 0, \ \beta_i^2 \rightarrow \beta_{i,max}^2$. We now analyze the secrecy rate achievable with these scaling factors. However, before discussing the achievable secrecy rate, we discuss an upper bound on the secrecy capacity of such a network. For the source-destination(eavesdropper) path, an upper bound on the capacity is given by the capacity of the Gaussian multiple-access channel between the relays in the $L^{th}$ layer and the destination (the eavesdropper). This results in the following upper bound on the secrecy capacity of such networks: \[C_{cut} = \frac{1}{2}\log\left[\frac{1 + P_t/\sigma^2}{1 + P_e/\sigma^2}\right],\] where $P_t = N^2 P_L h_t^2$ and $P_e = N^2 P_L h_e^2$. The power of the source signal reaching the destination $t$ is: \[\begin{aligned} P_{s,t} &= P_s h_s^2 \left(\prod_{i=1}^{L-1} N^2\beta_i^2 h_i^2\right) N^2 \beta_L^2 h_t^2 = \frac{N^2 P_L h_t^2}{(1+\delta)^L} \label{eqn:source_power_at_dstntn} \end{aligned}\] The total power of noise reaching the destination $t$ from all relay nodes: \[P_{z,t} = \sum_{i=1}^L P_{z,t}^i = \sigma^2\left(\sum\limits_{i=1}^{L-1}N \beta_i^2 h_i^2 \left(\prod\limits_{j=i+1}^{L-1}(N \beta_j h_j)^2\right)N^2 \beta_L^2 h_t^2 + N \beta_L^2 h_t^2\right) \label{eqn:noise_power_at_dstntn}\] where $P_{z,t}^i$ is the noise power reaching the destination form nodes in $i^{th}$ layer. Now, \[\begin{aligned} &&P_{z,t}^1 &= \sigma^2 N \beta_1^2 h_1^2 \left(\prod\limits_{j=2}^{L-1}N^2\beta_j^2 h_j^2\right) N^2 \beta_L^2 h_t^2 = \frac{\sigma^2}{P_s h_s^2} \frac{N P h_t^2}{(1+\delta)^L} \leq \frac{\delta}{(1+\delta)^L}N P h_t^2 &&\\ &&P_{z,t}^2 &= \sigma^2 N \beta_2^2 h_1^2 \left(\prod\limits_{j=3}^{L-1}N^2\beta_j^2 h_j^2\right) N^2 \beta_L^2 h_t^2 = \frac{\sigma^2}{N^2 P h_1^2} \frac{N P h_t^2}{(1+\delta)^{L-1}} \leq \frac{\delta}{(1+\delta)^{L-1}}N P h_t^2 &&\\ && \vdots & &&\\ &&P_{z,t}^i &= \sigma^2 N \beta_i^2 h_i^2 \left(\prod\limits_{j=i+1}^{L-1}N^2\beta_j^2 h_j^2\right) N^2 \beta_L^2 h_t^2 = \frac{\sigma^2}{N^2 P h_{i-1}^2} \frac{N P h_t^2}{(1+\delta)^{L-i+1}} \leq \frac{\delta}{(1+\delta)^{L-1}}N P h_t^2 &&\\ && \vdots & &&\\ &&P_{z,t}^L &= \sigma^2 N \beta_L^2 h_t^2 = \frac{\sigma^2}{N^2 P h_{L-1}^2} \frac{N P h_t^2}{(1+\delta)} \leq \frac{\delta}{(1+\delta)}N P h_t^2 && \end{aligned}\] Therefore, \[\begin{aligned} P_{z,t} &= \sum_{i=1}^L P_{z,t}^i \leq \sum_{i=1}^L \frac{\delta}{(1+\delta)^{L-i+1} NP h_t^2} \end{aligned}\] or \[\begin{aligned} P_{z,t} & \leq NP h_t^2\left[1-\frac{1}{(1+\delta)^L}\right] \label{eqn:uppr_bnd_noise} \end{aligned}\] The results in [\[eqn:source_power_at_dstntn\]](#eqn:source_power_at_dstntn){reference-type="eqref" reference="eqn:source_power_at_dstntn"} and [\[eqn:noise_power_at_dstntn\]](#eqn:noise_power_at_dstntn){reference-type="eqref" reference="eqn:noise_power_at_dstntn"} imply that we have the following for the achievable rate at the destination \[\begin{aligned} R_t &= \frac{1}{2} \log\left[1 + \frac{1}{(1+\delta)^L} \frac{N^2 P_L h_t^2}{P_{z,t} + \sigma^2} \right] \end{aligned}\] Similarly, the source and noise power reaching the eavesdropper $e$ respectively are: \[\begin{aligned} P_{s,e} &= P_s h_s^2 \left(\prod_{i=1}^{L-1} N^2\beta_i^2 h_i^2\right) N^2 \beta_L^2 h_e^2= \frac{N^2 P_L h_e^2}{(1+\delta)^L}\\ P_{z,e} &= \sum_{i=1}^L P_{z,t}^i = \sum\limits_{i=1}^{L-1}N \beta_i^2 h_i^2 \left(\prod\limits_{j=i+1}^{L-1}(N \beta_j h_j)^2\right)N^2 \beta_L^2 h_e^2 + N \beta_L^2 h_e^2 = P_{z,t}\ h_e^2/h_t^2 \end{aligned}\] and the achievable rate at the eavesdropper is: \[\begin{aligned} R_e &= \frac{1}{2} \log\left[1 + \frac{1}{(1+\delta)^L} \frac{N^2 P_L h_e^2}{P_{z,t} h_e^2/h_t^2 + \sigma^2} \right] \end{aligned}\] Therefore, the achievable secrecy rate is \[\label{eqn:hghSNRsecRate} R_s = R_t-R_e = \frac{1}{2}\log\left[\frac{1 + \frac{1}{(1+\delta)^L} \frac{N^2 P_L h_t^2}{P_{z,t} + \sigma^2}}{1 + \frac{1}{(1+\delta)^L} \frac{N^2 P_L h_e^2}{P_{z,t} h_e^2/h_t^2 + \sigma^2}}\right]\] Thus, we have the following for the gap between the cutset upper-bound and the achievable secrecy rate \[\begin{aligned} C_{cut}-R_s &= \frac{1}{2}\log\left[\frac{1 + P_t/\sigma^2}{1 + P_e/\sigma^2}\right]-\frac{1}{2}\log\left[\frac{1 + \frac{1}{(1+\delta)^L} \frac{N^2 P_L h_t^2}{P_{z,t} + \sigma^2}}{1 + \frac{1}{(1+\delta)^L} \frac{N^2 P_L h_e^2}{P_{z,t} h_e^2/h_t^2 + \sigma^2}}\right] \label{eqn:actual_gap} \end{aligned}\] Here, R.H.S. is an increasing function of $P_{z,t}$. Thus, from [\[eqn:uppr_bnd_noise\]](#eqn:uppr_bnd_noise){reference-type="eqref" reference="eqn:uppr_bnd_noise"} and [\[eqn:actual_gap\]](#eqn:actual_gap){reference-type="eqref" reference="eqn:actual_gap"}, we have \[\begin{aligned} C_{cut}-R_s &\leq \frac{1}{2}\log\left[\frac{\left(1+\left(1-\frac{1}{(1+\delta)^L}\right) \frac{NP h_t^2}{\sigma^2}\right) \left(1+\frac{N^2 P h_t^2}{\sigma^2}\right) \left(1+\left(1-\frac{1}{(1+\delta)^L}\right)\frac{N P h_e^2}{\sigma^2} + \frac{N^2 P h_e^2}{(1+\delta)^L \sigma^2}\right)}{\left(1+\left(1-\frac{1}{(1+\delta)^L}\right) \frac{NP h_e^2}{\sigma^2}\right) \left(1+\frac{N^2 P h_e^2}{\sigma^2}\right) \left(1+\left(1-\frac{1}{(1+\delta)^L}\right)\frac{N P h_t^2}{\sigma^2} + \frac{N^2 P h_t^2}{(1+\delta)^L \sigma^2}\right)}\right] \nonumber\\ & \leq \frac{1}{2}\log\left[\!\frac{\left(1+L \delta\frac{N P h_t^2}{\sigma^2}\right)\left(1+\frac{N^2 P h_t^2}{\sigma^2}\right)\left(1+ L \delta\frac{N P h_e^2}{\sigma^2} + (1-L\delta) \frac{N^2 P h_e^2}{\sigma^2}\right)}{\left(1+L \delta\frac{N P h_e^2}{\sigma^2}\right)\left(1+\frac{N^2 P h_e^2}{\sigma^2}\right)\left(1+ L \delta\frac{N P h_t^2}{\sigma^2} + (1-L\delta) \frac{N^2 P h_t^2}{\sigma^2}\right)}\right] \quad\left[\mbox{as }{1}/{(1+\delta)^L} \ge 1-L \delta\right]\nonumber\\ &= \frac{1}{2}\log\left[\left.\frac{1+{N^2 P h_t^2}/{\sigma^2}}{1+\frac{(1-L\delta) NPh_t^2/\sigma^2}{1+L\delta P h_t^2/\sigma^2}}\middle/ \right. \frac{1+{N^2 P h_e^2}/{\sigma^2}}{1+\frac{(1-L\delta) NPh_e^2/\sigma^2}{1+L\delta P h_e^2/\sigma^2}}\right]\nonumber\\ &\leq \frac{1}{2}\log\left[\left(\left. \frac{1+L\delta NP h_t^2/\sigma^2}{(1-L\delta)}\right)\middle/ \right.\left(1+L\delta \frac{NPh_e^2}{\sigma^2}\right)\right]\nonumber\\ &=\frac{1}{2}\log\left[\frac{1}{(1-L\delta)} \frac{1+L\delta NP h_t^2/\sigma^2}{1+L\delta NP h_e^2/\sigma^2}\right] \label{eqn:gap_bnd} \end{aligned}\] Since, $R_{s,opt} \geq R_s, \ C_{cut}-R_{s,opt} \leq C_{cut}-R_{s}$. Note that as $\delta \rightarrow 0, \ C_{cut}-R_s \rightarrow 0$, i.e., secrecy rate approaches the cut-set bound. # Numerical Simulations {#sec:numSim} In this section, we present numerical results to evaluate the performance of the proposed high SNR approximation scheme. We consider a 2-layer network with two nodes in each layer and eavesdropper snooping on the transmissions of the nodes in the last layer. In Figure [\[fig:2lyrdNetSecRateHighSNRplot\]](#fig:2lyrdNetSecRateHighSNRplot){reference-type="ref" reference="fig:2lyrdNetSecRateHighSNRplot"}, we plot the achievable secrecy rate when all the nodes transmit at their maximum power along with the corresponding cut-set as a function of source power for the specified system parameters. From the figure, it can be observed that at high SNR the achievable rate lies within a constant gap from the cut-set bound. In Figure [\[fig:2lyrdNetSecRateHighSNRplot1\]](#fig:2lyrdNetSecRateHighSNRplot1){reference-type="ref" reference="fig:2lyrdNetSecRateHighSNRplot1"}, the actual gap between the cut-set bound and achievable rate is $0.05$ bits/sec/Hz while the upper bound on this gap as given by [\[eqn:gap_bnd\]](#eqn:gap_bnd){reference-type="eqref" reference="eqn:gap_bnd"} is $0.25$ bits/sec/Hz. Similarly, in Figure [\[fig:2lyrdNetSecRateHighSNRplot2\]](#fig:2lyrdNetSecRateHighSNRplot2){reference-type="ref" reference="fig:2lyrdNetSecRateHighSNRplot2"}, the actual gap is $0.03$ bits/sec/Hz while the upper bound on this gap is $0.09$ bits/sec/Hz. Thus, it can be seen that [\[eqn:gap_bnd\]](#eqn:gap_bnd){reference-type="eqref" reference="eqn:gap_bnd"} tightly approximates the gap between achievable secrecy rate and the cut-set bound. # Conclusion and Future Work {#sec:cnclsn} We consider the problem of secure ANC rate maximization over a class of Gaussian layered networks where a source communicates with a destination through $L$ intermediate relay layers with $N$ nodes in each layer in the presence of a single eavesdropper which can overhear the transmissions of the nodes in any one layer. The key contribution of is the computation of the globally optimal set of scaling factors for the nodes that maximizes the end-to-end secrecy rate for a class of layered networks. We also show that in the high-SNR regime, ANC achieves secrecy rates within a constant gap of the cutset upper bound on the secrecy capacity and numerically validate this. In future, we plan to extend this work for more general layered networks.	{'timestamp': '2016-07-18T02:07:36', 'yymm': '1607', 'arxiv_id': '1607.04525', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04525'}
# Introduction The process of particle creation from quantum vacuum because of moving boundaries or time-dependent properties of materials, commonly referred as the dynamical Casimir effect (DCE), has been investigated since the pioneering works of Moore in 1970, who showed that photons would be created in a Fabry-Perot cavity if one of the ends of the cavity walls moved periodically,. The dynamical Casimir effect is frequently used nowadays for phenomena connected with the photon creation from vacuum due to fast changes of the geometry or material properties of the medium. Moving bodies experience quantum friction and so energy damping and decoherence due to the scattering of vacuum field fluctuations. The damping is accompanied by the emission of photons, thus conserving the total energy of the combined system. An explicit connection between quantum fluctuations and the motion of boundaries was made in, where the name non-stationary Casimir effect was introduced, and in, where the names Mirror Induced Radiation and Motion-Induced Radiation (with the same abbreviation MIR) were proposed. The frequency of created Photons in a mechanically moving boundary are bounded by the mechanical frequency of the moving body and to observe a detectable number of created photons the oscillatory frequency must be of the order of GHz which arise technical problems. Therefore, recent experimental schemes focus on simulating moving boundaries by considering material bodies with time-dependent electromagnetic properties. In this scheme, for example for two semi-infinite dielectrics, the boundary is not moving mechanically but its moving is simulated or modelled by changing the electromagnetic properties of one of the dielectrics in a small slab periodically. An important factor in detecting the created photons is keeping the sample at a low temperature of $\sim$ 100 mK to suppress the number of thermal black body photons to less than unity. Particularly, the problem has been considered with mirrors (single mirror and cavities), where the input field reflected completely from the surface. Recently the Robin boundary condition (RBC) has been used as a helpful approach to consider the dynamical boundary condition for this kind of problem. The well known Drichlet and Neuwmann boundary conditions can be obtained as the limiting cases of Robin boundary condition. The aim of the present work is to use a perturbative approach to study the effect of transition trough the interface on the spectral distribution of created photons. The interface between two semi-infinite dielectrics is modelled to simulate the oscillatory motion of the moving boundary. For this purpose, the electromagnetic field quantization in the presence of a dielectric medium is reviewed briefly then a general approach to investigate the dynamical Casimir effect for simulated motion of some part of a dielectric medium is introduced and finally, the spectral distribution of created photons are derived and the effect of small transitions trough the interface has been discussed. # The electromagnetic field quantization in absorbing dielectrics In this section we review briefly the electromagnetic field quantization in the presence of two adjacent semi-infinite absorbing media with different homogeneous and isotropic dielectric functions. Therefore, The dielectric function is defined by \[\label{e} \varepsilon(x,\omega)=\left\{ \begin{array}{ll} \varepsilon_1(\omega)=n_1^2(\omega)=[\eta_1(\omega)+i\kappa_1(\omega)]^2, & x<0 \\ \varepsilon_2(\omega)=n_2^2(\omega)=[\eta_2(\omega)+i\kappa_2(\omega)]^2, & x>0 \end{array} \right.\] where the subscript indices 1 and 2 correspond to the regions $x>0$ and $x<0$, respectively. The inhomogeneous nature of the problem requires the imposition of boundary conditions on the spatial mode functions on the interface. The vector potential in frequency space satisfies the familiar equation \[\label{A} (\frac{\partial^2}{{\partial x}^2}+\varepsilon(x,\omega)\frac{\omega ^2}{c^2})\,A(x,\omega)=-\frac{1}{\varepsilon_0 c^2 S}\,J(x,\omega).\] We can decompose any field to its positive and negative frequency parts then the positive frequency part of the vector potential is given by \[\label{a} \hat{A}^+(x,\omega)=S\int_{-\infty}^\infty dx' G(x,x',\omega)\hat{J}^+(x',\omega),\] where $S$ is the interface area and the Green's function fulfills the equation \[\label{g} (\frac{\partial^2}{{\partial x}^2}+\varepsilon(x,\omega)\frac{\omega ^2}{c^2})G(x,x',\omega)=-\frac{1}{\varepsilon_0 c^2 S}\delta(x-x').\] The Green's function is obtained explicitly as \[\begin{aligned} \label{g1} G(x,x',\omega)=\frac{i}{2\varepsilon_0c\omega n_1(\omega)S}\{R_L(\omega) \exp(\frac{i\omega n_1(\omega)(x+x')}{c})\qquad\qquad\qquad \\ \nonumber +\exp(\frac{i\omega n_1(\omega)\|x-x'\|}{c})\}, \qquad\qquad\qquad\qquad\qquad\qquad x>0,\,\,x'>0 \\ \nonumber=\frac{i}{2\varepsilon_0c\omega n_2(\omega)S}T_R(\omega)\exp(\frac{i\omega[n_1(\omega)x-n_2(\omega)x']}{c}), \qquad x>0,\,\,x'<0 \end{aligned}\] \[\begin{aligned} \label{g2} G(x,x',\omega)=\frac{i}{2\varepsilon_0c\omega n_2(\omega)S}\{R_R(\omega) \exp(-\frac{i\omega n2(\omega)(x+x')}{c})\qquad\qquad\quad\\ \nonumber+\exp(\frac{i\omega n_2(\omega)\|x+x'\|}{c})\}, \qquad\qquad\qquad\qquad\qquad\qquad x<0,\,\,x'<0 \\ \nonumber =\frac{i}{2\varepsilon_0c\omega n_1(\omega)S}T_L(\omega)\exp(-\frac{i\omega[n_2(\omega)x-n_1(\omega)x']}{c}), \quad x<0,\,\,x'>0 \end{aligned}\] where $T(\omega)$ and $R(\omega)$ are the usual transmission and reflection coefficients respectively, and the subscript indices $R$ and $L$ refer to the light incident on the interface from the right or left. These coefficients are given by \[\label{R} R_L(\omega)=-R_R(\omega)=\frac{n_1(\omega)-n_2(\omega)}{n_1(\omega)+n_2(\omega)}\] \[\label{T} \frac{T_L(\omega)}{n_1(\omega)}=\frac{T_R(\omega)}{n_2(\omega)}=\frac{2}{n_1(\omega)+n_2(\omega)}\] One can show that the vector potential in space-time can be written as \[\begin{aligned} \label{A1} \hat{A}^+(x,t)=\int_0^{+\infty}d\omega(\frac{\hbar\eta(\omega)}{4\pi\varepsilon_0c\omega\varepsilon(\omega)S})^\frac{1}{2} \{\hat{c}_R(x,\omega)+\hat{c}_L(x,\omega)\}e^{-i\omega t}. \end{aligned}\] The complete expressions for the operators $\hat{c}_{1R}(x,\omega)$ and $\hat{c}_{1L}(x,\omega)$ in the positive $x$ domain are determined using ([\[a\]](#a){reference-type="ref" reference="a"}), ([\[A1\]](#A1){reference-type="ref" reference="A1"}) and insertion of ([\[g1\]](#g1){reference-type="ref" reference="g1"}), ([\[g2\]](#g2){reference-type="ref" reference="g2"}) into ([\[a\]](#a){reference-type="ref" reference="a"}) as \[\begin{aligned} \label{aaa1} \hat{c}_{1L}(x,\omega)=i(\frac{2\omega\kappa_1(\omega)}{c})^{1/2} \int_x^{+\infty}dx' \exp(-\frac{i\omega n_1(\omega)}{c}(x-x')) \hat{f}(x',\omega) \end{aligned}\] \[\begin{aligned} \label{aaa2} && \hat{c}_{1R}(x,\omega)=\exp(\frac{i\omega n_1(\omega)x}{c})\times\nonumber\\ && \bigg\{ i(\frac{2\eta_2(\omega)\omega\kappa_1(\omega)}{\eta_1(\omega)c})^{1/2}\frac{n_1(\omega)}{n_2(\omega)}T_R(\omega) \int_{-\infty}^0 dx'\, \exp(-\frac{i\omega n_2(\omega)x'}{c}) \hat{f}(x',\omega)\nonumber\\ && +i(\frac{2\omega\kappa(\omega)}{c})^{1/2}\bigg[R_L(\omega)\int_0^{+\infty}dx'\,\exp(\frac{i\omega n_1(\omega)x'}{c})\hat{f}(x',\omega)\nonumber\\ && +\int_0^xdx'\exp(-\frac{i\omega n_1(\omega)x'}{c})\hat{f}(x',\omega)\bigg]\bigg\} \end{aligned}\] where $\hat{f}(x,\omega)=\hat{J}^+(x,\omega)\sqrt{S/2\varepsilon_0\hbar\omega^2\varepsilon_i(\omega)}$. The expressions for $\hat{c}_{1L}(x,\omega)$ and $\hat{c}_{1R}(x,\omega)$ contain $e^{(\frac{-i\omega n_1(\omega)x}{c})}$ and $e^{(\frac{i\omega n_1(\omega)x}{c})}$ respectively, which shows the direction of propagation of the field operators. This property can be used to determine the terms in ([\[a\]](#a){reference-type="ref" reference="a"}) which correspond to $\hat{c}_{1L}(x,\omega)$ and $\hat{c}_{1R}(x,\omega)$ easily. # Simulating the moving boundary Motivated by experiments in which moving boundaries are simulated by time dependent properties of static systems including, changing the effective inductance of the SQUID by a time-dependent magnetic flux or MIR experiment and also ,,,, we discuss here a model to change the dielectric function of a slab dielectric with thickness $\delta q$ which is placed at the interface of semi infinite absorbing dielectrics and its dielectric function oscillates between $\varepsilon_1(\omega)$ and $\varepsilon_2(\omega)$ with the frequencies $\omega_0$. This consumption equals to the oscillation of boundary with the mechanical frequency $\omega_0$. (see figure[\[ex\]](#ex){reference-type="ref" reference="ex"}) To solve the problem through a perturbative approach, we consider the dielectric function as: \[\label{e2} \varepsilon(x,t,t')=\varepsilon(x,t-t')+F(x,t,t')\] Where \[\label{f} \lim_{\delta q\rightarrow 0}F(x,t,t')=0\] $F(x,t,t')$ simulates the motion of boundary and is taken into account as the perturbation term, which is given by \[\begin{aligned} \label{f1} \nonumber F(x,t,t')=\{\varepsilon_1(t-t')-\varepsilon_2(t-t')\}f(x,t') \end{aligned}\] \[\label{e1} f(x,t') = \left\{ \begin{array}{cc} 0 \quad\quad\quad\quad\quad\quad\quad\quad\quad x<0 \\ \sin^2(\frac{\omega_0t'}{2})e^{\frac{-x}{\delta q}}\quad\quad\quad\quad x>0 \end{array} \right.\] where $e^{\frac{-x}{\delta q}}$ limits the thickness of slab to $\delta q$. We start from inhomogeneous Helmholtz differential equation \[\label{a2} \frac{\partial^2\hat{A}(x,t)}{{\partial x}^2}-\frac{1}{c^2}\frac{\partial}{\partial t}\int dt' \varepsilon (t-t') \frac{\partial}{\partial t'}\hat{A}(x,t')=\frac{1}{\varepsilon_0c^2}\hat{J}(x,t)\] where the transverse operator $\hat{J}(x,t)$ plays the role of a Langevin force associated with the noise reservoir. The field operators are separated into positive and negative frequency components in usual way, \[\label{a+} \hat{A}(x,t)=\hat{A}^+(x,t)+\hat{A}^-(x,t)\] and the frequency space Fourier transform operators are defined according to \[\label{a+w} \hat{A}^+(x,t)=\frac{1}{\sqrt{2\pi}}\int_0^\infty d\omega \hat{A}^+(x,\omega)e^{-i\omega t}\] With similar separations and transforms for noise current operators. The negative frequency component are provided bye hermitian conjugates of the positive frequency operators. We consider the effect of motion as a small perturbation \[\label{del a} \hat{A}(x,t)=\hat{A}_0(x,t)+\delta\hat{A}(x,t)\] where the unperturbed field $\hat{A}_0(x,t)$ corresponds to a solution with a static boundary at $x=0$. The first order field $\delta\hat{A}(x,t)$ then satisfies the following equation \[\begin{aligned} \label{del a1} \frac{\partial^2\delta\hat{A}(x,t)}{{\partial x}^2}-\frac{1}{c^2}\frac{\partial}{\partial t}\int dt' \varepsilon (x,t-t') \frac{\partial}{\partial t'}\delta\hat{A}(x,t')\\ \nonumber =\frac{1}{c^2}\frac{\partial}{\partial t}\int dt' F(x,t,t') \frac{\partial}{\partial t'}\hat{A}_0(x,t') \end{aligned}\] After transforming the above equation to Fourier space and ([\[e1\]](#e1){reference-type="ref" reference="e1"}), we find \[\begin{aligned} \label{del a2} \nonumber \frac{\partial^2\delta\hat{A}(x,\omega)}{{\partial x}^2}+\frac{\omega^2}{c^2}\varepsilon(x,\omega)\delta\hat{A}(x,\omega) =-\int_{-\infty}^\infty d\omega' \omega \\ \times(\omega-\omega')f(x,\omega')\{\varepsilon_1(\omega)-\varepsilon_2(\omega)\}\hat{A}_0(x,\omega-\omega') \end{aligned}\] where $f(x,\omega)$ is the Fourier transform of $f(x,t)$ To solve ([\[del a2\]](#del a2){reference-type="ref" reference="del a2"}) for $\delta\hat{A}(x,\omega)$ in terms of $\hat{A}_0(x,\omega)$, we consider the right hand side of that as a source and use ([\[g1\]](#g1){reference-type="ref" reference="g1"}) and ([\[g2\]](#g2){reference-type="ref" reference="g2"}), then \[\begin{aligned} \label{del a3} \nonumber \delta \hat{A}(x,\omega)=-\int_{-\infty}^\infty dx'' G(x,x'',\omega)\int_{-\infty}^\infty d\omega'\omega\qquad \\ \times(\omega-\omega')f(x'',\omega')\{\varepsilon_1(\omega)-\varepsilon_2(\omega)\}\hat{A}_0(x'',\omega-\omega') \end{aligned}\] from ([\[del a3\]](#del a3){reference-type="ref" reference="del a3"}) and ([\[e1\]](#e1){reference-type="ref" reference="e1"}) we find \[\begin{aligned} \label{del a4} \nonumber \delta\hat{A}(x,\omega)=-\int_0^{\delta q}dx''G(x,x'',\omega)[\varepsilon_1(\omega)-\varepsilon_2(\omega)]\\ \nonumber \times \{\omega^2 \hat{A}_0(x'',\omega)\quad\qquad\qquad\\ \nonumber-\frac{1}{2}\omega(\omega-\omega_0)\hat{A_0}(\omega-\omega_0)\quad\\ -\frac{1}{2}\omega(\omega+\omega_0)\hat{A_0}(\omega+\omega_0)\} \end{aligned}\] From ([\[del a\]](#del a){reference-type="ref" reference="del a"}) $\delta\hat{A}(x,\omega)$ is the first order of field correction and we can separate that for negative and positive frequencies. If in ([\[del a4\]](#del a4){reference-type="ref" reference="del a4"}) we consider $\omega>0$ or positive frequencies, which correspond to annihilation operators, the final field $\hat{A}^+(x,\omega)$ contains the negative frequencies, because of $\hat{A_0}(\omega-\omega_0)$ term which contains the creation operators for $0<\omega<\omega_0$ (negative frequencies) and we easily can show, the vacuum state for static field $\hat{A}_0(x,\omega)$ is not a vacuum state with respect to dynamical field $\hat{A}(x,\omega)$ with moving boundary condition. In the other word particles are created here by frequency $\omega$ which is less than the mechanical frequency $\omega_0$. \[\begin{aligned} \label{del a4} \delta\hat{A}_1(x,\omega)=-\int_0^{\delta q} dx''\frac{i(\varepsilon_1(\omega)-\varepsilon_2(\omega))}{2\varepsilon_0c\omega n_1(\omega)S} \{R_L \exp{(\frac{i\omega n_1(\omega)(x+x'')}{c})} \\ \nonumber +\exp{(\frac{i\omega n_1(\omega)(x-x'')}{c})}\} \{\omega^2 \hat{A}_0(x'',\omega)-\frac{1}{2}\omega(\omega-\omega_0) \hat{A}_0(x'',\omega-\omega_0)\\ \nonumber-\frac{1}{2}\omega(\omega+\omega_0)\hat{A}_0(x'',\omega+\omega_0)\} \end{aligned}\] Where we have $\hat{A}_0$ from equation ([\[A1\]](#A1){reference-type="ref" reference="A1"}). We calculate the perturbation of the field for positive $x$ domain. Further physical insight is gained if we drive the perturbation term of creation and annihilation operators. Obviously in ([\[del a4\]](#del a4){reference-type="ref" reference="del a4"}) $\delta\hat{A}_1(x,\omega)$ just contain a perturbation on the rightward operator $\hat{c}_{1R}$, because of $e^{(\frac{i\omega n_1(\omega)x}{c})}$ term, which shows the right ward propagation. We expected this kind of operators correction. \[\begin{aligned} \label{c1} \delta \hat{c}_{1R}(x,\omega)=-\int_0^{\delta q} dx''\frac{i(\varepsilon_1(\omega)-\varepsilon_2(\omega))}{2\varepsilon_0c\omega n_1(\omega)S} \{R_L \exp{(\frac{i\omega n_1(\omega)(x+x'')}{c})} \\ \nonumber +\exp{(\frac{i\omega n_1(\omega)(x-x'')}{c})}\} \times \{\omega^2[\hat{c_0}_{1R}(x'',\omega)+\hat{c_0}_{1L}(x'',\omega)]\\ \nonumber-\frac{1}{2}(\omega-\omega_0)\omega\Theta(\omega-\omega_0) [\hat{c_0}_{1R}(x'',\omega-\omega_0)+\hat{c_0}_{1L}(x'',\omega-\omega_0)] \\ \nonumber-\frac{1}{2}(\omega_0-\omega)\omega\Theta(\omega_0-\omega) [\hat{c_0}_{1R}^\dag(x'',\omega_0-\omega)+\hat{c_0}_{1L}^\dag(x'',\omega_0-\omega)]\\ \nonumber-\frac{1}{2}(\omega+\omega_0)\omega [\hat{c_0}_{1R}(x'',\omega+\omega_0)+\hat{c_0}_{1L}(x'',\omega+\omega_0)]\} \end{aligned}\] Where $\hat{c_0}_{1R}$ and $\hat{c_0}_{1L}$ are the unperturbed operators which are calculated in. Easily we can drive $\delta \hat{c}_{1R}^\dag(x,\omega)$ by complex conjugating ([\[c1\]](#c1){reference-type="ref" reference="c1"}) or by using ([\[del a4\]](#del a4){reference-type="ref" reference="del a4"}) and the negative frequency domain. Both give us the same result. \[\label{&c} \hat{c}=\hat{c_0}+\delta\hat{c}\] Now we consider the lossless dielectrics where $\kappa \rightarrow 0$. In this case the commutator of the operators $\hat{c_0}_{1R}(x,\omega)$ and $\hat{c_0}_{1R}^\dag(x,\omega)$ is obtained in \[\label{cc+} [\hat{c_0}_{1R}(x,\omega),\hat{c_0}_{1R}^\dag(x',\omega')]=\delta(\omega-\omega')\exp(\frac{i\omega n_1(\omega)(x-x')}{c})\] The commutation relations between the leftwards and rightwards annihilation and creation operators are also \[\begin{aligned} \label{cc+1} \nonumber [\hat{c_0}_{1R}(x,\omega),\hat{c_0}_{1L}^\dag(x',\omega')]=[\hat{c_0}_{1L}(x,\omega),\hat{c_0}_{1R}^\dag(x',\omega')]^*\qquad \qquad\qquad \\ = \delta(\omega-\omega')R_L(\omega)\exp(\frac{i\omega n_1(\omega)(x+x')}{c}) \end{aligned}\] For $x\geq0$ domain, we can consider only rightwards operators, because the leftwards operators are leaved unchanged by the perturbation. \[\label{c1l} \hat{c}_{1L}=\hat{c_0}_{1L}\] With \[\label{n0} <0_0\|\hat{c_0}_{1R}^\dag(\omega)\hat{c_0}_{1R}(\omega)\|0_0>=<0_0\|\hat{c}_{1L}^\dag(\omega)\hat{c}_{1L}(\omega)\|0_0>=0\] Since the rightwards annihilation operator is contaminated by leftwards and rightwards creation operators, the static vacuum state $\|0_0>$ is not a vacuum state with respect to the dynamic operators. # Frequency spectrum The number of particles created with frequencies between $\omega$ and $\omega+d\omega$ $(\omega\geq 0)$ is \[\label{n} \frac{dN}{d\omega}(\omega)d\omega=<0_0\|\hat{c}_{1R}^\dag(x,\omega)\hat{c}_{1R}(x,\omega)\|0_0>\frac{d\omega}{2\pi}\] The spectrum is obtained by inserting ([\[c1\]](#c1){reference-type="ref" reference="c1"}), ([\[cc+\]](#cc+){reference-type="ref" reference="cc+"}) and ([\[cc+1\]](#cc+1){reference-type="ref" reference="cc+1"}) into ([\[n\]](#n){reference-type="ref" reference="n"}) \[\begin{aligned} \label{n1} \frac{dN}{d\omega}(\omega)=\int_0^{\delta q}dx'\int_0^{\delta q}dx'' \frac{(n_1(\omega)-n_2(\omega))^2}{2\pi(4\varepsilon_0c n_1(\omega)S)^2}(\omega-\omega_0)^2\Theta({\omega_0-\omega})\\ \nonumber\{R_L^2(\omega)\exp(\frac{-i\omega n_1(\omega)(x''-x')}{c})+2R_L(\omega)\cos[\frac{\omega n_1(\omega)}{c}(x''+x')]\\ \nonumber+\exp(\frac{i\omega n_1(\omega)(x''-x')}{c})\}\times2\{\cos[\frac{\omega n_1}{c}(x''-x')]\qquad\qquad\quad\\ \nonumber+R_L(\omega)\cos[\frac{\omega n_1}{c}(x''+x')]\}\qquad\qquad\qquad\qquad\qquad\qquad\qquad \end{aligned}\] We define dimensionless parameter $a=\frac{x'\omega_0}{c}$ and $a'=\frac{x''\omega_0}{c}$ and also $y=\frac{\omega}{\omega_0}$ which is always smaller than unity and rewrite ([\[n1\]](#n1){reference-type="ref" reference="n1"}) again. \[\begin{aligned} \label{n11} \frac{dN}{dy}(y)=\int_0^{\frac{\delta q\omega_0}{c}}da\int_0^{\frac{\delta q\omega_0}{c}}da' \frac{(n_1(y)-n_2(y))^2}{2\pi(4\varepsilon_0 n_1(y)S)^2}(y-1)^2\Theta({1-y})\\ \nonumber\{R_L^2(y)\exp(-iy n_1(y)(a'-a))+2R_L(y)\cos[y n_1(y)(a'+a)]\\ \nonumber+\exp(iy n_1(y)(a'-a))\}\times2\{\cos[y n_1(y)(a'-a)]\qquad\qquad\quad\\ \nonumber+R_L(y)\cos[y n_1(y)(a'+a)]\}\qquad\qquad\qquad\qquad\qquad\qquad\qquad \end{aligned}\] Now we are going to plot the spectrum as a function of $y$. In this paper we work with the non relativistic approximation and as the previous work on the simulation of motion of the bound, the mechanical speed of bound can be considered about $\%10$ of the speed of light. In this limit $\frac{\delta q\omega_0}{c}\backsimeq 0.1\pi$, so it is not small enough to expand ([\[n11\]](#n11){reference-type="ref" reference="n11"}) in the first order of $\frac{\delta q\omega_0}{c}$. Figure ([\[nem\]](#nem){reference-type="ref" reference="nem"}) shows the spectrum in this case. This spectrum doesn't contain symmetry around $y=\frac{1}{2}(\omega=\frac{\omega_0}{2})$ and doesn't vanish too fast with respect to $R_L$ less than unity. So we have valuable content for spectrum even in case of about $\%4$ transition of the incidental fields. Another meaningful choice for the mechanical speed of bound would be about $\%0.1$ of the speed of light where in this case $\frac{\delta q\omega_0}{c}\backsimeq 0.001\pi$ and so it would be small enough to expand ([\[n11\]](#n11){reference-type="ref" reference="n11"}) and we find \[\begin{aligned} \label{n2} \frac{dN}{dy}(y)= \frac{(n_2(y)-n_1(y))^2}{2\pi(4\varepsilon_0S)^2}(y-1)^2\Theta({1-y})\{(1-R_L)^3-2R_L^2(\frac{\delta q\omega_0 y}{c})^2\} \end{aligned}\] If we consider $R_L\rightarrow-1$, which represent the case of complete reflection of the leftward field from the bound, we find \[\label{n3} \frac{dN}{dy}(y)=-2 \frac{(n_2(y)-n_1(y))^2(\delta q\omega_0)^2}{2\pi(4\varepsilon_0 cS)^2}(y-1)^2y^2\Theta({1-y})\] Figure ([\[nem1\]](#nem1){reference-type="ref" reference="nem1"}) shows the spectrum with these considerations. As we see from figure ([\[nem\]](#nem){reference-type="ref" reference="nem"}) and ([\[nem1\]](#nem1){reference-type="ref" reference="nem1"}) the spectrum vanishes for $y\leq 1$ or in the other word $\omega \leq \omega_0$ and so no particle is created with frequency greater than the mechanical frequency of the bound. But here the spectrum ([\[n3\]](#n3){reference-type="ref" reference="n3"}) is the symmetry around $y=1/2$ where the spectrum has a peak over there (figure ([\[nem1\]](#nem1){reference-type="ref" reference="nem1"})), and in this case the spectrum is valuable just for $R_L$ too close to unity or in case of complete reflection. # Conclusion As a result of the figure ([\[nem\]](#nem){reference-type="ref" reference="nem"}) and ([\[nem1\]](#nem1){reference-type="ref" reference="nem1"}), spectrum decrease rapidly by the decrease in value of $R_L$ and actually for a small variation from $-1$, it vanishes. But in figure ([\[nem\]](#nem){reference-type="ref" reference="nem"}) the decrease in spectrum with respect to $R_L$ is less than the figure ([\[nem1\]](#nem1){reference-type="ref" reference="nem1"}). So we would have valuable content of spectrum, even in case of a little transition of the incidental fields. But at all, if we are going to detect the created particles, we would increase our chance by considering one of the medium as a conductor. In the case $R_L\rightarrow-1$ and $\delta q\rightarrow 0$ the spectrum was the same as the spectrum of dynamical casimir effect which has been studied by a variety of methods ,,,, such as Robin boundary condition.	{'timestamp': '2016-07-19T02:05:10', 'yymm': '1607', 'arxiv_id': '1607.04782', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04782'}
null	null
null	null
null	null